JP2006292734A - Judgment model creation support device, inspection device, durability test device and durability test method for inspection device - Google Patents

Abstract

An inspection apparatus suitable for automatically performing an endurance test is provided.
A sensor is attached to a work (engine) (S1), and an endurance test is started (S2). After a certain period of time, the durability test is temporarily stopped (S3). And the acquired sensing data is passed to the determination algorithm preparation means in the test | inspection apparatus 10 (S4). If it is in the initial state, it is estimated as a normal state, and the judgment algorithm creating means creates a reference space and a judgment model that are digitized as normal waveform data and defines a normal region, and registers it in the judgment algorithm (S5). Is set to a threshold value for determining an abnormality (S6). Next, the endurance test is restarted (S7), and the engine is kept rotating for a long time. Sensing data obtained from each sensor during the endurance test is given to the determination algorithm in the inspection device, state determination is performed in real time (S8), and if the threshold is exceeded, an instantaneous stop command is output to the endurance tester ( S9).
[Selection] Figure 2

Description

  The present invention relates to a determination model creation support device for an inspection device, an inspection device, an abnormality detection device for an endurance test device, and an endurance test method. More specifically, the present invention extracts feature quantities from an input waveform signal. And determining the state based on the extracted feature amount.

  There are product inspections and facility diagnosis for inspecting whether the inspection object is normal or abnormal based on sound generated from the inspection object. Equipment diagnosis is based on the vibration and sound generated by machine tools and production facilities themselves, and diagnoses whether the machine tools and production facilities themselves are operating normally, and whether maintenance such as maintenance and adjustment is almost necessary. It is. Specific examples include equipment diagnosis such as NC processing machines, semiconductor plants, and food plants. In the product inspection, whether a product is a normal product or a defective product is inspected based on vibration or sound generated from the product. Both are common in that they are inspected based on vibration or sound. Product inspection is mainly explained. Some products manufactured by production facilities and production systems have a built-in sound source or vibration source. There are also products that generate sound or vibrate depending on the operation of the product. For example, in refrigerators, air conditioners, and washing machines for home appliances, components such as a motor are incorporated, and when operated, sound and vibration are generated as the motor rotates. For example, in automobiles, there are sound sources or vibration sources all over the engine, power steering, power seat, mission and so on.

  Some of the sounds and vibrations related to such products are inevitably generated with normal operation, and others are generated with defects. Abnormal noise and vibration associated with the failure are caused by abnormal contact inside the motor, bearing (bearing) abnormality in the rotation mechanism, abnormal contact inside the rotation mechanism, imbalance in the rotation mechanism, contamination, etc. More specifically, examples of abnormal sounds generated by the operation of the mechanism include abnormal sounds in which the rotating part and the fixed part inside the motor rub against each other for a moment during rotation. The abnormal sound of the rotating mechanism includes an abnormal sound caused by a gear chip occurring once per rotation of the rotating gear, a foreign object biting into the gear, a spot flaw on the bearing, and the like. An example of a sound that a person feels unpleasant is a “kick” sound that is mixed for a moment in a certain operation sound. Assuming that a normal product can only hear a certain operating sound, a product that produces a “click” sound can be regarded as a defective product.

  In addition, a ceramic product or a product made by combining resin parts does not have a sound source or a vibration source, but may be inspected for cracks or the like. In the inspection of such a product, the pottery or resin part to be inspected is hit with a tool such as a hammer and a sound is generated. When there is no crack in the object, a sound that resonates with a high tone is produced, but when there is a crack, a dull sound is produced.

  Note that “sound” in the specification includes sound and vibration. In the specification, abnormal sounds and abnormal vibrations are collectively referred to as “abnormal sounds” or “abnormal sounds”. In addition, “vibration” is used to include vibration and sound.

  The sound accompanying such an abnormality or defect is not only unpleasant for humans, but may also cause a failure in the product itself. Products that produce such sounds need to be inspected during the production process to classify non-defective products. Therefore, for the purpose of quality assurance for each product manufactured by the production equipment and production system, the production factory usually performs "sensory inspection" that relies on the five senses such as hearing and touch by the inspector. Judging whether there is any abnormal sound. Specifically, humans listen to sounds with their ears or humans touch them with their hands to check vibrations. The sensory test is defined by the sensory test term JIS (Japanese Industrial Standard) Z8144.

  By the way, the sensory test that relies on the five senses of the inspector requires not only a skilled technique but also a large variation due to individual differences in determination results. Furthermore, there is a problem that it is difficult to make data and numerical values of judgment results of sensory tests difficult to manage. Therefore, in order to solve such a problem, there is an abnormal sound inspection apparatus for the purpose of inspection based on a quantitative and clear standard.

  This abnormal noise inspection apparatus is an apparatus for the purpose of automating the “sensory inspection” process. It measures vibration and sound of a product drive unit with a sensor and analyzes and inspects an analog signal captured by the sensor. .

  As an abnormal sound inspection apparatus that automatically performs an inspection (so-called abnormal sound inspection) for determining normality / abnormality from a vibration waveform obtained from an inspection object as described above, there is one disclosed in Patent Document 1 in the past. The invention disclosed in Patent Document 1 comprehensively determines normality / abnormality of an inspection object using a feature amount obtained from a time axis waveform and a feature amount obtained from a frequency waveform.

  Further, as in the invention disclosed in Patent Document 2, a normal area where a normal product exists is formed only from normal data obtained from a non-defective product (normal product), and if the detected value is within the normal region, it is determined to be normal. However, there is a technique for determining an abnormality if the detected value is outside the normal region. In the invention disclosed in Patent Document 2, using a plurality of input information, a normal region in which a normal state is allowed is set as a multidimensional vector, and if the detected value is within the normal region, it is determined as normal, If it is out of the area, it is judged as abnormal.

Japanese Patent No. 3484665 Japanese Patent No. 3103193

  By the way, in the development of engines, transmissions, tires, etc., trial operation is continued at the prototype stage, and it is often confirmed by endurance tests that there are no weak parts (parts damaged during the trial operation). In this endurance test, test workpieces (the aforementioned prototype engine, transmission, tire, etc.) are recorded in real time with a tachometer, thermometer, vibration meter, or the like. In this test, if one place breakage occurs, if it does not stop immediately, the damaged part will cause another part breakage, secondary damage, tertiary damage, ... Even if the test is stopped after the enlargement and the test workpiece is disassembled and confirmed, it is extremely difficult to identify the weakest primary breakage site, and even if it can be identified, it takes a lot of time and labor.

  On the other hand, a part that is primarily damaged is not suddenly damaged, and some change occurs as a sign of the damage. For example, subtle abnormalities occur due to wear or deformation before normal breakage, and by continuing the durability test in that state, the primary breakage of parts that cannot be endured due to abnormal enlargement eventually occurs. appear. And when such a subtle abnormality occurs or at least a primary breakage occurs, this may be a change in workpiece vibration due to a change in workpiece component movement, a change in workpiece drive load, a change in workpiece temperature for each part, Then, a change in the work housing distortion occurs, and accordingly, a change in the work drive sound due to a change in each part occurs. That is, the plurality of complex information is emitted as sound.

  There is a difference in the timbre generated from the test object being tested for durability under abnormal conditions (before and after primary damage). Therefore, as in the case of normal abnormal noise inspection, if an inspector who is familiar with the test work is stationed near the test work during the endurance test and hears the sound, a change in the timbre can be detected, and the primary damage will occur at that stage. It can be detected that it has occurred. In other words, humans do not look at specific information (only specific frequencies, only specific vibration modes, and only specific part temperatures), but rather capture and judge the pattern (changes) of the timbre in all frequency bands. You will notice the difference between normal and abnormal states. Therefore, as in the inspection system used in the conventional endurance test, the abnormality is determined based on specific information (only a specific frequency, only a specific vibration mode, and only a specific part temperature). As long as it is impossible to detect an abnormality at the stage of primary breakage or a predictive stage at which primary breakage is likely to occur, it can be detected only after the damage has expanded.

  Thus, although it is true that abnormalities are often noticed at the stage of primary damage or a sign stage where primary damage is likely to occur if the auditory sense is used, the feasibility of detecting abnormalities by such inspectors is low. . This is because the endurance test may last for 24 hours to a week if it is long, and it is practically impossible to monitor whether there is any abnormality due to human beings. In addition, it is normal for a person to find a primary breakage with a higher sensitivity from time to time, or to notice it quickly after a secondary or tertiary breakage. It is the same as sound inspection.

  Further, in the case of an inspection apparatus for judging pass / fail (abnormality) based on conventional waveform signals such as sound and vibration disclosed in patent documents, etc., a test work is driven and sensing is performed during that time, and a fixed time such as 5 to 30 seconds. The waveform data for the number of seconds is captured, and the normality of the captured waveform data is determined. That is, pass / fail judgment is performed on the entire waveform data of a fixed number of seconds obtained by driving for a fixed time.

  On the other hand, in the case of the durability test, as described above, very long waveform data is continuously input compared to the continuous waveform targeted by the conventional abnormal sound inspection apparatus of 24 hours or more, and the input is performed. Since the determination is performed on the waveform data in real time, the conventional inspection device (abnormal sound inspection device) cannot be applied as it is.

  In addition, considering the implementation in the endurance test, the time allowed to interrupt and stop the test when damage occurs is short, so normal / abnormal for waveform data of a short time such as 0.1 to 1 second. Judging from this point, it is difficult to apply the conventional pass / fail judgment (abnormality detection) algorithm based on waveform data for several seconds.

  The present invention can perform pass / fail determination / abnormality determination even for short-time waveform data, and can be applied to waveform data obtained continuously from an inspection object that is continuously driven for a relatively long time. Another object of the present invention is to provide a determination model creation support apparatus and inspection apparatus for an inspection apparatus, an abnormality detection apparatus for an endurance test apparatus, and an endurance test method that can detect the abnormality in a short time when an abnormality occurs.

  (1) A determination model creation support apparatus for an inspection apparatus according to the present invention is for an inspection apparatus that extracts a feature amount from waveform data acquired from an inspection target product and determines the state of the inspection target product based on the extracted feature amount. A determination model creation support apparatus for dividing waveform data creation means for dividing the acquired waveform data into data for a set unit time, and a plurality of unit times for the unit time created by the divided waveform data creation means Based on the divided waveform data, means for calculating the feature amount in each divided waveform data unit, and for determining the state of the inspection target product for each unit time based on the feature amount calculated in the divided waveform data unit A model creation means for creating a judgment model is provided. In this way, even if the unit time (N seconds) is short, since it is formed from different divided waveform data, an accurate determination model can be created.

  (2) The divided waveform data creating means may have a function of dividing the waveform data every unit time from a predetermined position. (3) Further, the divided waveform data creating means may have a function of randomly determining a division start position when dividing every unit time. (4) Furthermore, the divided waveform data creating means may have a function of randomly determining a unit time that becomes a reference length when the division processing is performed. In either case, since a large number of divided waveform data can be collected, a highly accurate determination model can be created. In the former, the divided waveform data can be obtained evenly, and in the latter, it is randomly selected, so a judgment model suitable for reliably detecting abnormalities that do not know at which point in actual inspection / judgment is created. can do.

  (5) Further, for an inspection apparatus that extracts a feature amount from a waveform signal acquired from an inspection target product whose operation repeatedly varies in a predetermined pattern, and determines the state of the inspection target product based on the extracted feature amount A determination model creation support device, a divided waveform data creation unit that divides acquired waveform data into data for a set unit time, and a plurality of divisions for a unit time created by the divided waveform data creation unit Based on the waveform data, means for calculating the feature amount in each divided waveform data unit, and on the basis of the feature amount calculated in the divided waveform data unit, each of the patterns in the pattern divided by the divided waveform data creating means A model creation means for creating a determination model for a region can also be provided. In this way, for example, when the motion (motion profile) of the inspection object is periodically and repeatedly executed, a determination model suitable for each divided region can be manufactured.

  (6) In the above (1) to (5), the means for obtaining the feature amount performs a frame division process for further dividing the divided waveform data into frames of a predetermined size, and also based on the divided frames. And a function of randomly determining a predetermined size. (7) In this case, it is preferable that the means for obtaining the feature amount performs frequency analysis on each frame obtained by executing the frame division process and obtains the feature amount in the entire frequency band. In this way, since a large number of frames can be collected, a highly accurate determination model can be created. In particular, even when a determination model for the region is created according to one divided waveform data, a highly accurate determination model can be created based on a large amount of information.

  (8) The inspection apparatus according to the present invention is configured to perform state determination on the acquired waveform data to be inspected based on the determination model created by the determination model creation support apparatus of (1) to (7) described above. Means are provided.

  (9) Further, the inspection apparatus is for an apparatus whose operation is repeatedly fluctuated in a predetermined pattern, and the determination model creation support apparatus according to (5) is based on waveform data corresponding to one pattern portion that fluctuates repeatedly. When the created determination model for each region is acquired and the state of the inspection target is determined, it is preferable to include a determination unit that performs determination processing using the determination model of the region corresponding to each of the acquired waveform data portions.

  (10) Further, the durability test apparatus according to the present invention is an endurance test apparatus for testing the durability of a product to be inspected, and waveform data acquired during a period when the operation after the start of the durability test is stable is normal data. Using the determination model created as follows, it can be configured to include determination means for determining whether there is an abnormality based on the waveform data acquired during the execution of the durability test.

  (11) In this case, the determination model creation function may be configured by the above-described determination model creation support devices (1) to (7).

  (12) In the durability test method according to the present invention, the driving body to be subjected to the durability test is rotated for a predetermined time, and the waveform data acquired at that time is generated as a determination model according to any one of (1) to (7) A processing step to be applied to the support device, a processing model in which the determination model creation support device creates a determination model for abnormality detection using the given waveform data as normal data, and then executes an endurance test. Based on the waveform data acquired during execution, a processing step for determining the presence / absence of an abnormality using the determination model created by the determination model creation support device, and a processing step for outputting an abnormality detection signal when an abnormality is detected I tried to run. With the output of the abnormality detection signal, the rotation of the driving body can be stopped.

  In the present invention, quality determination / abnormality determination can be performed even with short-time waveform data. As a result, even for waveform data that is continuously acquired from an inspection object that has been continuously driven for a relatively long time, the abnormality is detected in a short time when an abnormality occurs, and the durability test is stopped when the abnormality occurs. It becomes possible to do. Of course, it is possible to make a determination in a shorter time by acquiring waveform data for a few seconds, which has been generally performed conventionally, and applying it to an inspection apparatus for determining normality / abnormality as well as the durability test.

  FIG. 1 shows a schematic configuration of an endurance test system which is one preferred embodiment of the present invention. As shown in FIG. 1, a driving device 2 such as a motoring bench is connected to a rotating shaft of an engine 1 that is a test object, and the driving device 2 is driven to rotate to forcibly rotate the engine 1. Can be done. The drive device 2 includes a servo motor and a servo driver that controls the rotation of the servo motor, and the built-in servo motor rotates at a predetermined rotation speed in accordance with a control command from the PLC 3. Since it is a servo motor, it can be rotated at a constant speed at an arbitrary rotation speed, and the speed increase / decrease control can be performed with high accuracy. As a result, the engine 1 can be continuously driven at a desired rotational speed for a relatively long time (one day to two weeks or more).

  A sensor (microphone 4) for detecting sound generated from the engine 1 during the test is arranged around the engine 1 to be subjected to the durability test, and the vibration sensor 5 for detecting the vibration of the engine 1 is disposed on the engine 1. It arrange | positions in the state contacted to the predetermined position. In the present embodiment, the microphones 4 are installed at two locations above and below the engine 1 to collect sound information from different directions, but the number of installations and installation positions are arbitrary. The same applies to the vibration sensor 5. Further, as shown in the drawing, it is not always necessary to install different types of sensors such as the microphone 4 for detecting sound and the vibration sensor 5 for detecting vibration. In short, it is only necessary to be able to detect waveform data as a basis for detecting an abnormality.

  Since analog waveform data is output from the various sensors of the microphone 4 and the vibration sensor 5, each waveform data is changed to digital data by the AD converter 6 and then given to the inspection device 10. Yes. Further, the operation timing is given from the PLC 3 to the inspection device 10, and the waveform data from the monitoring sensor (encoder, torque meter, etc.) installed in the driving device 2 is AD converted from the driving device 2. It is given through the vessel 6.

  The inspection apparatus 10 includes a database 7 that stores waveform data and data such as a determination model and a determination rule for detecting an abnormality, and is acquired from the A / D converter 6 in accordance with the determination model acquired from the database 7. The presence / absence of an abnormality in the waveform data is determined, and when an abnormality is detected, an abnormality occurrence notification is sent to the PLC 3 as a determination result. When the PLC 3 receives the abnormality occurrence notification, the PLC 3 outputs an instantaneous stop command to the drive device 2 and instantaneously stops the rotation of the servo motor of the drive device 2. Since it is a servo motor, it can be stopped relatively instantaneously upon receiving an instantaneous stop command. Thereby, the rotation of the engine 1 during the durability test is also stopped.

  As will be described later, the abnormality can be detected based on waveform data of, for example, about 1 second, so that the abnormality can be detected almost simultaneously with the occurrence of the abnormality. The abnormality detected here includes not only the occurrence of primary damage of parts of the engine 1, but also abnormally a state where there is a high possibility of primary damage if the driving is continued as it is. Since the determination can be made quickly and accurately according to the waveform data for 1 second in this way, the rotation result of the drive device 2 is stopped and the rotation of the engine 1 is also stopped by transmitting the detection result to the PLC 3. To do. Thereby, it is possible to perform in a short period (a little over 1 second) from the abnormality detection of the primary damage (or its fear) to the actual stop of the engine 1 that is rotationally driven.

  FIG. 2 is a diagram for explaining the overall flow of the durability test performed in this embodiment. First, a sensor is attached to the workpiece (S1). The workpiece referred to here is the engine 1 that is the subject of the durability test. That is, as shown in FIG. 1, after the engine 1 to be subjected to the durability test is attached to a driving device 2 such as a motoring bender, the sensors 4 and 5 are attached to predetermined positions of the engine 1. The attachment method is appropriately set according to each sensor, and there are some attachment methods in contact with each other, and some attachments at a predetermined distance. The method for installing and attaching these sensors is the same as that used in the conventional abnormal noise inspection.

  If the sensor installation is completed, the durability test is started (S2). At this time, various operation profiles are repeatedly performed such as rotating at a constant speed at a certain speed for a certain time, accelerating, decelerating, or repeatedly accelerating, decelerating, or constant speed in a predetermined pattern. Then, during the execution of each operation pull file, sensing data (waveform data) acquired from various sensors (microphone 4, vibration sensor 5, monitoring sensor mounted on the driving device 2, etc.) is converted into an A / D converter 6. Then, the data is converted into digital data and taken into the inspection apparatus 10. The captured waveform data is stored in the database 7 or stored in a primary storage unit or the like in the inspection apparatus 10. Basically, since the engine 1 subjected to the durability test is a non-defective product, the waveform data which is sensing data obtained immediately after the start of the durability test can be said to be normal waveform data based on the non-defective product.

  If a certain time has elapsed after the start of the durability test, the durability test is temporarily stopped (S3). Then, the sensing data acquired so far is transferred to the determination algorithm creating means (determination model creation support device) mounted in the inspection apparatus 10 as normal state data (S4). Then, the determination algorithm creating means digitizes the given normal waveform data, creates a reference space and a judgment model that defines the normal region, and registers them in the judgment algorithm (S5). This determination algorithm is one of the abduction programs installed in the inspection apparatus 10 and realizes an abnormality detection means (function) that performs state determination (determination of whether an abnormality has occurred) during an endurance test. It is. Actually, the registration here is registered in the database 7, for example. Note that a work (engine) that requires a leveling operation acquires sensing data several times as an initial normal state after the leveling operation ends, and creates a determination algorithm (determination model) based on the acquired sensing data.

  Strictly speaking, the waveform data of the non-defective product is not only different depending on the engine 1, but is different depending on the combination of the driving device 2 and the engine 1, the connection state when the engine 1 is attached to the driving device 2, and the surrounding environment. Therefore, each time an endurance test is performed on each engine 1, normal waveform data that takes into consideration the environment of the endurance test that is performed for a certain period of time each time before the actual continuous operation for a long time is performed. Was obtained, and a reference space / judgment model was created.

  When the above-described generation process of the monitoring algorithm of S4 and S5 is completed, a threshold value (a value for specifying / detecting an area away from the reference space by a certain distance) determined from the determination model is set (the value for specifying / detecting an area). S6). This threshold value is registered in a determination algorithm (function for executing the algorithm) in the inspection apparatus 10.

  And it shifts to operation of an actual endurance test. That is, the durability test is resumed (S7), and the engine 1 is continuously rotated for a long time according to the required operation profile. Then, sensing data (waveform data) obtained from each sensor during the execution of the durability test is given to the determination algorithm in the inspection apparatus 10, and the state is determined in real time (S8). If the threshold is exceeded, an instantaneous stop command is issued. Is output to the durability tester (drive device 2) (S9).

  That is, as will be described later, the inspection apparatus 10 can determine the presence / absence of an abnormality based on the waveform data for each unit time. Therefore, if an abnormality is detected exceeding the threshold, an abnormality detection notification is sent to the PLC 3 as a determination result. (The bit of the determination result is changed from “0” to “1”, etc.). The PLC 3 that has received this abnormality detection notification outputs an instantaneous stop command to the drive device 2. As a result, the servo motor in the drive device 2 stops urgently, so that the rotation of the engine 1 also stops instantaneously. Therefore, abnormality detection can be performed in real time (for example, about 1 second after the occurrence of abnormality), and the end process of the durability test can be performed without delay, so the engine can be stopped before secondary damage occurs. it can. Therefore, when the primary breakage occurs, the damaged part can be easily identified, and even when it stops before the primary breakage, identification of the weakened (probable breakage) part Is relatively easy to do. As a result, the time required for the durability test and the analysis based thereon can be greatly shortened.

  FIG. 3 shows a schematic image of determination algorithm generation and determination processing based thereon according to the present embodiment. FIG. 3A is a schematic image diagram of an (abnormal sound) inspection apparatus used in a conventional in-line inspection machine. In the case of an in-line inspection machine, a determination model is created based on a timbre in a normal state for a certain length of time (6 seconds in the figure) as a determination criterion, and pass / fail determination (abnormality determination) is performed according to the determination model.

  If the determination algorithm of this conventional in-line inspection machine is applied as it is, for example, a determination model is created from a waveform for 6 seconds. Therefore, if there is an abnormality somewhere in this 6 seconds, it can be determined that there is an abnormality. However, the inspection machine cannot be stopped at the timing when an abnormality occurs. In the case of conventional in-line inspection machines, it is only necessary to determine whether the test target is normal or abnormal, so specify the part in the series of waveform data where the abnormality occurred and it was determined to be defective. This is because it is not necessary (it is not necessary to stop at the timing of occurrence of an abnormality).

  Therefore, in this embodiment, when waveform data of a certain length of normal state is acquired, a determination model is not created from the entire waveform data, but as shown in FIG. In this example, a determination model is generated on the basis of the timbre of the normal state at that time, and a determination algorithm based on the determination model is executed. Thus, in the case of FIG. 3B, it can be detected whether there is an abnormality every second, and an abnormality can be detected without waiting for 6 seconds as in the case of FIG. 3A, and can be stopped there. Actually, there is a slight delay until the stop command is generated because of the determination process.

  In FIG. 3B, the waveform data for 6 seconds is divided every second, and “determination model 1”, “determination model 2”, “determination model 3”, “determination model 4”, Although the determination model is created separately from “determination model 5” and “determination model 6”, for example, the content of the endurance test is to examine the state change when rotating at a constant speed for a long time at a certain rotation speed. In the case of a thing, if other external conditions do not change, a more accurate determination model can be created by actually creating a determination model for one second based on each determination model 1 to 6. . Also, in the case of an endurance test where the speed is repeatedly changed in a predetermined pattern such as acceleration or deceleration (a constant speed state may be entered midway), the model can be selected according to different external conditions such as each speed. Create and make decisions based on it. In short, instead of creating a judgment model for the entire waveform data acquired in a lump, the judgment model is created by dividing it into shorter times. The unit to be monitored is not fixed and may be 0.5 seconds or 1 second or more. This is set according to the specification (allowable time) required from the occurrence of abnormality (detection) to the stop.

  Next, a specific internal structure of the inspection apparatus 10 for implementing the various embodiments described above will be described. FIG. 4 shows functional blocks for creating a determination algorithm. The inspection device 10 acquires waveform data from various measurement devices (such as the microphone 4 and the vibration sensor 5) disposed in the engine 1 and the drive device 2. In the present embodiment, since the waveform data is acquired from a plurality of types and a plurality of locations in an endurance test for one engine, the inspection apparatus 10 handles a plurality of waveform data.

  The waveform data obtained by executing S2 and S3 in FIG. 2 is given to the waveform data dividing means 11, the dummy NG generating means 12, and the recorded waveform display means 13. The waveform data dividing means 11 divides the acquired waveform data into a plurality of areas. A feature amount is obtained for each divided waveform data divided by the waveform data dividing means 11, and a final determination model is created.

  That is, the waveform data dividing means 11 divides the acquired continuous waveform data by unit time (N seconds) as shown in FIG. The N seconds use at least an instantaneous stop time, that is, a time required for an emergency stop from the occurrence of an abnormality requested by the user (time allowed for abnormality detection). In the present embodiment, in response to a request to detect abnormality within one second, the waveform data is divided in units of one second with N = 1. FIG. 5 shows a cut-out display of 0 to 6 seconds during the continuous test at constant speed rotation and constant torque.

  As will be described later, by using the determination model created based on the divided waveform data divided by N = 1, an abnormality can be detected within one second, and the instantaneous stop time associated with the occurrence of the abnormality is about 1 second. Then, as shown in FIG. 5, as indicated by a solid double-directional arrow, first, (1) division is successively performed in units of N seconds (1 second) from the beginning. That is, it is divided regularly every second. Thereby, in the illustrated example, six pieces of divided waveform data are extracted. As will be described later, feature amount extraction is performed for each divided waveform data. Therefore, since all of the waveform data existing in 6 seconds is present in any one of the divided waveform data, it is possible to influence somewhere when generating the feature amount. Furthermore, in the present embodiment, as the processing of (2), as shown by the double-pointed arrows in FIG. 5, the division start position is randomized and a predetermined number of waveform data is extracted. The divided waveform data extracted by the process (2) does not belong to any of the divided waveform data extracted by randomly dividing the division start position out of the waveform data for 6 seconds, if there is a partial overlap. There is also a waveform portion. Thus, even if there is a waveform portion that does not belong to any divided waveform data, it belongs to any of the divided waveform data based on the processing of (1) and is therefore reflected in the feature amount extraction. Absent.

  For example, “0.7 second to 1.7 second” region, “0.75 second to 1.75 second” region, “0.9 second to 1.9 second” The divided waveform data is extracted for each area, such as the area of “0.5 second to 1.5 second”. However, the length of the divided waveform data extracted in any case is the same (1 second).

  Further, N can take an arbitrary value. For example, when N = 2, it is as shown in FIG. In this case as well, three waveform data (solid double-directional arrows in the figure) can be obtained sequentially and sequentially from the beginning in units of 2 seconds, and the division start position is randomly set as indicated by the broken-line double arrows in FIG. Then, a predetermined number of waveform data is extracted. The maximum value of N is the longest time of the acquired waveform data (6 seconds in FIG. 5).

  In this way, even if the time required for an emergency stop from the occurrence of an abnormality requested by the user is 1 second, depending on the cause of the abnormality, it cannot be detected from the divided waveform data for 1 second, and a longer length (for example, Some of them can be detected for the first time by the feature amount based on the divided waveform data (for 2 seconds). In such a case, it takes about 2 seconds to stop due to the occurrence of an abnormality, which is longer than 1 second requested by the user, but still can be stopped earlier than the conventional 6 seconds, and the location of the abnormality can be easily identified. . In addition, although 1 second and 2 seconds were demonstrated as an example for convenience, actually, the value of N is also set at random. Further, N may be an integer or not limited to an integer.

  That is, if it is desired to stop after one second, basically, a determination model generated from a waveform cut every second is the main determination. However, in order to make use of information having a trend of a span of 1 second or longer, it is preferable to use a determination model created from waveform data for this time. Of course, N seconds may be any number, and a plurality of databases and determination models may be prepared.

  In this way, the determination rule (determination model) created by setting N to a value longer than 1 does not contribute to instantaneous interruption within 1 second, but has an advantage that information on a long time trend can be incorporated into the determination. If there is a surplus in computing capability, a judgment model created from various time-dependent waveform data may be incorporated.

  The dummy NG generation means 12 deforms normal waveform data obtained by executing S2 and S3. The dummy NG data created by the dummy NG generating means 12 is given to the waveform data dividing means 11. In the case of the present embodiment, sample data for a normal product can be obtained, but sample data (NG data) for a defective product (abnormal) cannot be obtained. This is because in the endurance test, the data obtained at the initial stage as a premise is normal data, and the purpose is to detect the occurrence of abnormality by continuous operation for a long time. This is because the sample data cannot be obtained.

  Therefore, sample data (waveform data) of a normal product is taken in by the dummy NG generation means 12, and pseudo NG data is created based on the normal waveform data. The dummy NG data created in this way can be used for evaluation of the created determination algorithm. In other words, the accuracy of the created recognition algorithm can be estimated based on whether or not the dummy NG generation unit 12 provides dummy NG data to the subsequent units, and based on the determination, whether or not the determination can be made correctly. That is, it can be used as data for verifying whether or not the determination model can correctly detect the change from the initial state.

  FIG. 7 shows an example of the dummy NG generation means 12. The dummy NG generation unit 12 includes a waveform deformation parameter setting unit 12a on the input side, and sets parameters for deformation for the input waveform data.

  Examples of parameters set by the waveform deformation parameter setting means 12a include: (1) Abnormal mode waveform synthesis (synthesis of eccentric abnormal waveform, synthesis of impact waveform) in waveform library, (2) nth-order amplification of drive condition specific frequency (Rotation frequency, 1st to 4th order amplitude of meshing frequency is multiplied by 1.5), (3) Specific or random frequency amplification (Amplitude of frequency 500 to 1000 Hz is multiplied by 1.2) (4) FM modulation, AM modulation, (5) phase shift (the phase of the original waveform is slightly shifted and synthesized with the original waveform), etc.

  Here, in the waveform synthesis of (1), by synthesizing an abnormal waveform that does not occur in the normal waveform, the synthesized waveform data is overlapped with the waveform data different from the normal data due to the effect of the synthesized abnormal waveform portion. Appears and becomes abnormal data. In (2), for example, assuming a gear or the like, sound or vibration generated at the time of abnormality due to the meshing frequency appears at a specific frequency. This meshing frequency can be calculated from the number of teeth and the rotational frequency. Therefore, by increasing the n-th order amplitude of each frequency (by increasing the power of a normal product), abnormal data different from normal data is obtained. Although detailed explanation is omitted, in other cases, waveform data that cannot be obtained normally can be generated.

  The set parameters are given to the next-stage modification specification setting means 12b together with the waveform data. The deformation specification setting means 12b sets a specification for deforming the waveform deformation parameter based on the experiment plan direct table. For example, the waveform synthesis is “two levels of on / off” and “three levels of frequency for a specific frequency”. "With respect to amplification, there are" three levels of 1.2 times, 1.5 times and 2 times ". In this way, the value of each parameter when the waveform is deformed is set.

  Then, the waveform data selection means 12c randomly selects the waveform data by the number of experiments in the orthogonal table. And the waveform deformation | transformation means 12d changes the waveform data with respect to the selected waveform data according to the deformation | transformation amount selected by the deformation | transformation specification setting means based on an orthogonal table. As a result, a dummy NG waveform is finally generated and added to normal waveform data as waveform data. That is, even if it can be determined based on this waveform data, the reliability of the inspection algorithm of the inspection apparatus 10 can be improved if it is determined that the product is defective.

  The recorded waveform display means 13 displays the waveform data fetched from each measuring device or the recorded waveform data stored in the recording means. Since the inspection apparatus 10 can be constituted by a general-purpose personal computer or the like, the recorded waveform display means 13 can be realized by a display monitor provided in the personal computer. In the case of an endurance test, continuous waveform data is taken in and abnormality determination is performed in real time. Therefore, in a normal operation, the recorded waveform display means 13 displays the waveform data input in real time in a through state. It will be. In addition, since the waveform data continuously captured in this way is stored in the database 7 or other storage means, the waveform data stored at the time of the analysis in the case of performing the analysis after the stop due to the occurrence of abnormality is stored. May be read and displayed. That is, the “recording waveform” in the recording waveform display means 13 has both the meanings of “recording waveform” and “recorded waveform”.

  The divided waveform data divided by the waveform data dividing means is given to the waveform data digitizing means 14 at the next stage. The waveform data digitizing means 14 digitizes the divided waveform data given as will be described later and converts it into a feature value.

  The waveform data digitizing means 14 is connected to a digitizing means adjusting means 9 and a waveform data history recording means (database) 7. The digitizing means adjusting means 9 performs parameter adjustment of the feature quantity when the waveform data digitizing means 14 extracts the feature quantity, and has a function of giving an instruction for parameter adjustment to the waveform data digitizing means 14. Have. Band division setting, frame division setting, frame variation number setting, etc. are also performed.

  The waveform data history recording means 7 records waveform data such as sound generated when the engine 1 is forcibly rotated using the driving device 2. Normal / abnormal judgment results may be stored in association with each other. In addition, the recorded waveform data may be reproduced for such a determination result, and a person may determine and correct the determination result.

  The waveform data digitizing means 14 extracts a predetermined feature amount from the given divided waveform data. As the feature quantity to be extracted, various values can be used in addition to the average value of the vibration level and the RMS (Route Mean Square Value) indicating the magnitude. The feature amount is generated for each condition, for example, for each unit time (N) when the divided waveform data is created. Further, although not shown in the figure, it is created for each operation profile such as ambient temperature, rotation speed, acceleration / deceleration states, and the like. The waveform data digitizing means 14 can output and display the waveform data digitizing results, that is, the obtained characteristic quantities to the digitized result display means 13 ″.

  Here, when rotating at a constant speed, first, a feature amount is obtained in divided waveform data units in which N is the same. For example, when N = 1, feature amounts are obtained for each of the divided waveform data extracted every second by the methods (1) and (2). The determination result (normal) and the feature quantity obtained by the waveform data digitizing means 14 are associated with each other and stored in the feature quantity / history database 15. At this time, the feature amount and history are recorded in the record No. Are stored in association with each other.

  An example of the data structure of the feature quantity / history database 15 is shown in FIG. In the history column (normal / abnormal (with abnormality type)), data (history type) given from the waveform data history recording means 7 is stored, and from the waveform data digitizing means 14 in the subsequent feature quantity column. Each given feature is stored. When a model (determination algorithm) is created, all the history columns are “normal”.

  In FIG. 8, Fx1, Fx2, Fx3,. This is the feature amount for the divided waveform data specified by For example, record No. 1 is a feature amount for the first divided waveform data, record No. 2 is a feature amount for the next divided waveform data, and record No. 6 is a feature amount for the last divided waveform data. It is. Further, although not specifically shown, the feature amount for each divided waveform data in which the start position of (2) is set at random is recorded.

  The feature amount extraction process described above is created based on the divided waveform data (time axis system waveform shape) divided on the time axis, but in this embodiment, the frequency axis obtained by frequency decomposition such as FFT or order conversion is further used. System features are also extracted. The target waveform data to be subjected to the FFT processing is also divided waveform data. In this embodiment, as shown in FIG. 9, each divided waveform data is divided into frames with a predetermined frame size. In the example of FIG. 9, it is divided into m pieces. Because of frame division, the frame size is set to N seconds or less (in this example, 1 second or less) and is determined randomly. However, it is set to a power of 2 in relation to digital processing.

  Then, FFT is performed for each frame, and frequency components of each frequency band (F1 to Fn) are obtained. As a result, as shown in FIG. 10, the frequency component is specified by the matrix of each frequency band (F1 to Fn) and each frame (T1 to Tm). Therefore, the average value and standard deviation of power are calculated in the time direction for each frequency band. Then, the obtained average value and standard deviation are rearranged in a horizontal row to obtain the feature quantity (1) (see FIG. 11).

  Thereafter, the frame size is re-determined at random, and the same processing as described above is performed to obtain the feature amount (2). Thereafter, the designated number of times is performed n times, and all feature quantities (1) to (n) are arranged in a horizontal row to obtain a final feature quantity (see FIG. 11). In FIG. 11, the drawing is divided into two stages for convenience of illustration, but in actuality, the images are stored in one row (consecutive data group). In this example, (1) to (n) are arranged in a horizontal row and used as feature amounts. However, a logarithmic conversion of these may be used as a final feature amount.

  The graph shown in FIG. 12 represents the timbre during driving of the test work as a frequency pattern (shows an example in which an arbitrary frequency band is divided into 40). Unlike a measurement device, human hearing is not detected by a specific frequency alone, but is considered to recognize the entire frequency band of the audible region as a certain pattern. Therefore, it can be recognized that “the sound has changed” regardless of the power of any frequency band. If a specific breakage is detected as in an inspection device in a conventional in-line inspection, since it may be possible to discriminate only at a specific frequency, a filtering process for extracting the specific frequency as a pre-processing is usually performed. Execute. However, in the case of the durability test to which the present embodiment is applied, it is not known in which part the primary damage occurs, and therefore it is not known in which frequency band the sound change that occurs due to the one-time damage occurs. Therefore, in the present embodiment, a feature amount is obtained with the entire frequency band as the inspection target range, and a determination model is created to recognize the entire frequency band as a pattern like human hearing.

  That is, in the present embodiment, based on the above-described concept, not only a specific frequency is seen, but the entire frequency band (F1 to Fn) is seen, and the frequency pattern is expressed numerically by an arbitrary method, so that human hearing I tried to imitate. Although this principle is explained based on the change of sound, it goes without saying that this principle can be applied even if the sound is generalized as a waveform. For example, if vibration is sensed by a vibration sensor (acceleration sensor) and the waveform is reproduced by a speaker, a person can sense a change in state similar to that sensed by a microphone. The same can be said even if the acceleration sensor is changed to a temperature sensor or a strain gauge.

  FIG. 13 shows an example of the internal structure of the waveform data digitizing means 14. As shown in FIG. 13, the waveform data digitizing means 14 stores a plurality of divided waveform data acquired from the waveform data dividing means 11 for the time axis system waveform shape template group storage unit 14a for storing each divided waveform data. Each of which has a frame dividing unit 14f for performing frame dividing processing. As described with reference to FIG. 9, the frame dividing unit 14 f has a function of first randomly determining a frame size to be divided and dividing each divided waveform data into frames according to the determined frame size. Then, the frame division process is executed a predetermined number of times while changing the frame size. The waveform data generated by the frame dividing unit 14f is given to the frequency resolution processing unit 14b. The frequency decomposition processing unit 14b performs frequency decomposition such as FFT and order conversion on the given waveform unit waveform data. That is, the process described with reference to FIG. 10 is executed. Then, the obtained feature amount data is stored in the template group storage unit 14c having the frequency axis system waveform shape. The data registered in the template group storage units 14a and 14c is obtained by extracting a plurality of types of feature data from one divided waveform data. Various types of feature data are obtained by specific functions.

  By the way, items to be obtained by the waveform data digitizing means 14 using a specific function are predetermined such as an average value and a maximum value (may be added if necessary). The (arithmetic expression) includes adjustable parameters (coefficients and constants), and the accuracy of pass / fail judgment (abnormality detection) is improved by appropriately setting the parameters. In other words, if the adjustment is not appropriate, the accuracy of the pass / fail judgment (abnormality detection) decreases. Conventionally, the setting of such parameters is finally determined by a skilled inspector who adjusts by trial and error. Of course, in the present invention, the parameter may be set by manual operation in the same manner as in the past, but in this embodiment, the inspection apparatus 10 automatically optimizes the parameter and sets the specific function set to the optimum value. The waveform data is digitized (feature value extraction).

  Specifically, the waveform data digitizing means 14 includes a specific function optimizing unit 14d that adjusts and optimizes a specific function used when digitizing the waveform data. The specific function optimizing unit 14d changes various parameters of the specific function according to instructions from the digitizing means adjusting means 9. Specifically, it has a function of executing the flowchart shown in FIG.

  That is, the specific function optimizing unit 14d first examines the combination effect of the coefficients and constants of the specific function, and sets the investigation condition based on the experiment plan direct table (S11). That is, according to the instruction from the digitizing means adjusting means 9, a plurality of combinations of coefficients and constants are set, and an orthogonal table associated with the investigation conditions (survey specifications) is created. Next, the specific function optimizing unit 14d calculates dynamic characteristics (sn ratio) for each investigation specification (experiment NO.) Using normality and abnormality as signal factors and the number of data as error factors (S12). In other words, a plurality of waveform data is digitized using a specific function set with parameters (coefficients and constants) defined in each survey specification (features are obtained), and OK (normal / good) The distance between the group of feature quantities (numerical values) indicating NG and the group of feature quantities (numerical values) indicating NG (abnormal / bad) is determined.

  Then, the specific function optimization unit 14d calculates a process average of the coefficients and constants of the specific function (S13), and selects a value with a high sn ratio for each count and constant of the specific function (S14). Based on the selected value, the parameters (coefficients and constants) of the specific function are determined, and the specific function using the coefficients and constants is set to be optimal (S15). The above-described parameter evaluation / setting of the specific function is performed for each specific function, that is, for each feature amount.

  Further, the specific function optimizing unit 14d gives the optimized specific function to the numerical processing unit 14e. Then, the digitization processing unit 14e digitizes the waveform data using each set optimum specific function and outputs the obtained feature amount. The feature quantity / history storage means 7 stores the output feature quantity.

  The data stored in the feature quantity / history database 15 can be displayed on the registered content display means 13 ′ or can be changed by operating the edit deletion means 16. The registered content display means 13 'here and the recording waveform display means 13 described above can be physically realized by the same monitor. The data stored in the feature quantity / history database 15 is stored in the time division database 18 every divided time (N seconds). The data stored in the time division database 18 is output to the registered content display means 13 'and displayed. The editing / deleting means 16 can delete / change data stored in the time division database 18.

  Each data recorded in the time division database 18 is given to the determination model generation means (for each time) 19 in the next stage for each division time. The determination model generation unit 19 generates a determination formula for determining whether or not the waveform data (feature amount) to be inspected matches each history type based on the feature amount for each time. That is, the determination model generation means 19 creates a determination formula for determining normality based on normal data whose history type is normal.

  Note that the state determination formula can take various methods such as Mahalanobis distance method, Euclidean distance method, normal-abnormal contrast method, neural network method, and fuzzy method using a membership function. Then, as will be described later, it can be created automatically, or can be created by a person as in the conventional case.

  In the case of a test at a constant speed operation, one type of judgment model is created. However, the judgment model is created from data for several seconds. That is, a determination model is generated from feature amounts of a plurality of waveform data divided every second. The determination model generated here specifies a normal range (for example, within ± 3σ from the average) from the average and standard deviation of each feature quantity, and determines a normal / abnormality based on the degree of deviation from the normal range ( (Judgment model). As a determination model, a normal range is obtained for each feature quantity, each feature quantity is obtained from waveform data acquired every second, and each feature quantity is within the above normal range (a certain degree of divergence). If there is even one feature quantity that is abnormal (or a predetermined number), it may be judged that an abnormality has occurred, or each feature quantity is preliminarily determined as one multidimensional vector. It is also possible to determine the presence or absence of abnormality based on the difference between the multidimensional vector and the multidimensional vector obtained from the waveform data to be inspected.

  In the case of the present invention, since the determination execution means 21 initially determines based only on normal knowledge, only those whose history is “normal” are stored in the feature quantity / history database 15. Also in the time division database 18, only data with the history type “normal” shown in FIG. 8 is created and stored. For this reason, the determination model generation unit 19 also generates a determination formula for determining normality and sets it in the determination execution unit 21.

Here, the generation of the judgment model is
Normal = f (feature value 1, feature value 2,... Feature value n)
As described above, the feature value is a column vector. Therefore, normal (label) is normalized and handled as 1 (of course, other than 1 may be used). As a result, when each feature value is different from the initial state immediately after the start of the durability test, f (feature value 1, feature value 2,... Feature value n) is a value close to 1. It will not come out. In other words, if f (feature value 1, feature value 2,... Feature value n) is seen, it can be determined whether it is close to normal or not close.

Thus, normal = f (feature value 1, feature value 2,... Feature value n)
Various algorithms such as Euclidean distance, Mahalanobis distance, and neural network model can be used as an algorithm for calculating the difference from the normal product using the expression.

  Each data recorded in the time division database 18 is also given to the instantaneous stop threshold setting means 20. The instantaneous stop threshold value setting means 20 determines whether the result obtained by calculating the feature amount obtained from the waveform data obtained from the engine 1 using the determination formula matches the history type. The threshold for discriminating whether or not is determined. The determined threshold value is given to the determination execution means 21.

  As a result, a determination algorithm (determination model) is generated, and the determination execution unit 21 determines pass / fail (abnormality determination) based on the waveform data (features) to be inspected using the set determination formula and threshold value. To do. The determination result is output via the display means 23 and the output means 24 and stored in the result storage means 25. The result storage means 25 may store not only the state determination result but also determination (history) made by a person, waveform data, feature amount, and the like in association with each other. The display means 23 is physically the same as the recording waveform table means 13 and the like.

  As described above, the determination model generation means 19 can take various methods. As an example, the determination model generation means 19 can have the internal structure shown in FIG. This illustrated is an example for realizing the Euclidean distance method. First, the feature amount data for each history type stored in the normal database 18a and the abnormal database 18b (created by the dummy NG generating unit 12) of the time division database 18 are respectively calculated and stored in the corresponding Euclidean distance calculating / accumulating units 19a and 19b. To give. The normal data Euclidean distance calculation / accumulation means 19a calculates and accumulates the Euclidean distance by calculating the square root of the sum of the squares of the feature quantities based on the acquired normal data (feature quantities). Also, the Euclidean distance calculation / accumulation means 19b for abnormal data is obtained and stored in the same manner as the Euclidean distance for normal data, except that the data to be processed is abnormal data. Further, when the abnormal data is divided for each abnormality type, the Euclidean distance is obtained and stored for each abnormality type.

  The Euclidean distance of each data calculated and accumulated as described above is given to the corresponding statistic calculation means 19c and 19d. The normal statistic calculating means 19c for calculating the statistic of the normal data Euclidean distance calculates a statistic such as the maximum value, average value, standard deviation, etc. of the given Euclidean distance of the plurality of normal data. Similarly, the anomaly statistic calculating means 19d for calculating the anomaly data Euclidean distance statistic calculates the statistic such as the maximum value, average value, standard deviation, etc. of the Euclidean distance of the given plural anomaly data. Also in this case, it calculates | requires for every abnormality kind.

The statistic obtained by the respective statistic calculating means 19c, 19d is given to the determination formula determining unit 19e at the next stage. The determination formula determination unit 19e includes the maximum normal statistic value (normal maximum value) obtained by the normal statistic calculation unit 19c and the minimum abnormal statistic value (abnormal minimum value) obtained by the abnormal statistic calculation unit 19d. Compare
It is determined whether or not normal maximum value <abnormal minimum value. If the conditional expression is satisfied, it is determined that the expression for obtaining the Euclidean distance of the feature amount set at that time is correct, and is set in the determination execution means 21. Thereby, the determination execution means 21 calculates the Euclidean distance by obtaining the square root of the sum of the squares of the given feature values.

  If the conditional expression is not satisfied, the setting of the specific function for calculating the feature amount is inappropriate. Therefore, the determination formula determining unit 19e instructs the quantifying means adjusting means 9 to determine its parameter (specific function Request change of coefficients and constants. Receiving this, the digitizing means adjusting means 9 sets the coefficient and constant of the specific function set last time to a value other than the previously set value. As a result, the specific function is changed, so that the feature quantity quantified based on the changed specific function is changed, and the statistical value is also changed. By repeatedly executing this process, the determination formula determination unit 19e generates a condition that satisfies the conditions.

  When the determination model generation means 19 is obtained by the Euclidean distance method as described above, the determination model generation means 19 calculates the normal data Euclidean distance to the determination execution means 21 (the sum of the squares of the feature values). Specifies the square root function.

  FIG. 16 shows another configuration of the determination model generation means 19. In this example, the normal-abnormal contrast method is realized. That is, the feature amount data for each history type stored in the normal database 18a and the abnormality database 18b of the time division database 18 is given to the feature amount calculation means 19f. The feature quantity calculating means 19f selects an arbitrary specific function and calculates the feature quantity of the data stored in the database using the selected specific function. The feature amount of normal data is stored in the feature amount data storage unit 19g of normal data, and the feature amount of abnormal data is stored in the feature amount data storage unit 19h of abnormal data.

  The feature amount of each data calculated and accumulated as described above is given to the corresponding statistic calculation means 19i, 19j. The normal statistic calculating means 19i for calculating the statistic of the normal data feature quantity calculates the statistic quantity such as the maximum value, average value, standard deviation, etc. of the given normal data feature quantity. Similarly, the abnormal statistic calculation means 19j for calculating the statistical amount of the abnormal data feature amount calculates the statistical amount such as the maximum value, the average value, and the standard deviation of the given abnormal data. Also in this case, it calculates | requires for every abnormality kind.

Then, the statistic obtained by each statistic calculation means 19i, 19j is given to the determination formula determination unit 19k in the next stage. The determination formula determination unit 19k is configured to use the maximum normal statistic value (normal maximum value) obtained by the normal statistic calculation unit 19i and the minimum abnormal statistic value (abnormal minimum value) obtained by the abnormal statistic calculation unit 19j. And compare
It is determined whether or not normal maximum value <abnormal minimum value. If the conditional expression is satisfied, it is determined that the specific function selected by the feature amount calculation unit 19f is correct and is defined in the state determination unit 21. Thereby, the determination execution means 21 calculates the Euclidean distance by obtaining the square root of the sum of the squares of the given feature values. Then, the state is determined based on whether or not the Euclidean distance is equal to or greater than a threshold value.

  If the conditional expression is not satisfied, it can be determined that the selected specific function is inappropriate. Therefore, the feature amount calculation unit 19f is requested to change the specific function to be used. Receiving this, the feature amount calculation means 19f selects a specific function different from the previous one, and calculates the feature amount again. As a result, the specific function is changed, so that the feature quantity quantified based on the changed specific function is changed, and the statistical value is also changed. By repeatedly executing this process, the determination formula determination unit 19k selects a condition having a condition.

  The conditions in the determination formula determination unit 19k are not limited to those described above. For example, the maximum value may be normal average + 3σ, and the minimum value may be abnormal average −3σ, and various changes can be made. It is.

  When the determination model generation means 19 is obtained by the normal-abnormal contrast method as described above, the determination model generation means 19 defines a specific function selected for the determination execution means 21.

  FIG. 17 shows still another configuration of the determination model generation means 19. In this example, a neural network system is realized. That is, the feature amount data for each history type stored in the normal database 18a and the abnormal database 18b of the time division database 18 is given to the screening means 19m. This screening means 19m calculates an outlier of each database and deletes the data. Outliers are, for example, (1) those that deviate from the average ± 3σ, (2) the data from the maximum value to the third and the data from the minimum value to the third, average, standard without using a total of 6 points Deviations can be calculated and deviated from the mean ± 3σ.

  The data screened by the screening means 19m is given to the database integration means 19n to integrate the data. Then, the learning means 19p inputs the feature data stored in the integrated database, and learns (constructs a clustering model) a neural network model that outputs the history level. Various methods used in the neural network can be used for the learning process. When the learning is completed, the neural network model of the learning result is defined as a state determination unit and set in the determination execution unit 21.

  FIG. 18 shows still another configuration of the determination model generation means 19. In this example, a plurality of feature quantities are integrated based only on normal data, and a model for discrimination using the Mahalanobis distance is created. The feature amount of normal data stored in the normal database 18a is given to the statistical processing means 19q, and the statistic amount of all feature amounts is calculated there. The calculated statistics are the average and standard deviation. That is, an average value and a standard deviation for each feature amount are obtained, and an average vector in which the average values of the feature amounts are grouped together and a standard deviation vector in which the standard deviations are grouped together are obtained.

  The obtained statistics are given to the normal database standardization means 19r, and the normal database standardization means 19r standardizes the data stored in the normal database 18a with the average vector and the standard deviation vector. This is because normalization and standardization are attempted because the numerical values of the feature quantities vary. Further, a correlation matrix for each feature amount is obtained by means 19s for obtaining a correlation matrix, and the obtained correlation matrix is passed to means 19t for obtaining an inverse matrix to obtain an inverse matrix of the correlation matrix.

  The average vector, standard deviation vector, and inverse matrix obtained by the above means are stored in the storage means 19u. Each data stored in the storage means is given to the Mahalanobis distance calculation formula generating means 19v to obtain a Mahalanobis distance calculation formula.

  That is, the Mahalanobis distance D ^ 2 is an average m1 for each feature quantity when the data of n inspection target workpieces is measured with the number of feature quantities k as X1, X2,. , M2,..., Mk and standard deviations σ1, σ2,. If the component of the inverse matrix of the correlation matrix at this time is aij, the Mahalanobis distance is defined by the following equation. The calculation formula generation unit 31 f generates such a formula and sets it in the determination execution unit 21.

  The normal data obtained in the initial stage from the start of the durability test is similar in pattern to the ideal normal data, and therefore is plotted at a position close to the evaluation standard, and the Mahalanobis distance is a value near 1. On the other hand, the abnormal data generated due to the occurrence of loss or the like in the primary is plotted far away from the evaluation standard in accordance with the pattern difference from the normal data, and the Mahalanobis distance becomes a large value. Therefore, it is possible to easily determine whether the Mahalanobis distance is close to 1 or not. Note that a function of evaluating the reliability contribution rate of the Mahalanobis distance of a specific function to be used and deleting a low contribution rate may be added.

  In conventional inspection equipment, data for the time required for calculation is cut out from the sampled waveform data for the entire time, and a set of data obtained by dividing the cut-out data by a fixed number of data is obtained. One frame is extracted, and a plurality of types of feature amounts are extracted for each frame. Then, for each feature quantity obtained from all the frames, a representative feature quantity calculation value is obtained for each of the same kind of feature quantity by an average or other various methods. Therefore, the number of representative feature value calculation values is calculated by the number corresponding to the type of feature value. The pass / fail determination (abnormality determination) is performed using all of the plurality of representative feature amount calculation values or a predetermined number selected from them. Regardless of the number of representative feature value calculation values used, comparison is made in units of feature values (representative feature value calculation values) (scalar amounts).

  On the other hand, in the present embodiment, the obtained plural types of feature quantities are collected and converted into one numerical value (multiple waveform evaluation: vector quantity). As a matter of course, the model for determining pass / fail (abnormality determination) is also a vector amount generated by obtaining a plurality of feature amounts. Therefore, the pass / fail judgment is made by comparing the model amount with the vector quantity based on the waveform data to be inspected. If the distance between the two vectors is within a certain range, the model is matched. Judged that it is different. In other words, in the case of only normal determination, the reference model needs to have at least one numerical value (multiple waveform evaluation: vector amount) obtained by collectively collecting a plurality of types of feature amounts, and obtaining the distance to it. The pass / fail judgment (abnormality judgment) can be performed. In other words, after calculating a vector quantity that summarizes each feature quantity (actually, each representative feature quantity computation value obtained based on a plurality of frames), it is possible to make a pass / fail judgment with only one computation process for obtaining the distance. it can. The distance between the two vector quantities may be calculated based on the Mahalanobis distance as in the present embodiment, or may be calculated using various methods such as the Euclidean distance.

  This unification of feature quantities can also be used to unify feature quantities of a plurality of waveform data. Of course, for waveform data from different sensors and measurement locations, it is not necessary to unify the features as described above, but to make a judgment based on the waveforms obtained for each measurement source (sensor) of each waveform data. It is also possible to make a comprehensive judgment (for example, temporary damage or the like if any one of them is abnormal). In other words, when there are multiple sensing data (sound, temperature, vibration, etc.), the decision algorithm is created using the methods described above for each waveform, and finally the decision results are integrated by the decision algorithm. be able to.

  The instantaneous stop threshold value setting means 20 sets a threshold value by manual operation. That is, as shown in FIG. 19, the instantaneous stop threshold value setting means 20 gives the feature amount data for the normal data stored in the time division database 18 to the normal distribution confirmation means 20 a, and the abnormality stored in the time division database 18. The feature amount data for the data (dummy NG data) is given to the dummy NG distribution confirmation means 20b. When an abnormality type is set, it is given for each abnormality type. Here, normal data and abnormal data may be separated based on history information, and normal and abnormal may be separated based on a person's determination result.

Each distribution confirmation means 20a, 20b calculates the distribution status of the acquired feature quantity for each history type. For example, the average value, median value, standard deviation, quartile, n × σ (n = 1, 2) , ...) is calculated. Then, each calculated value is given to the normal distribution-abnormal type distribution positional relationship calculating means 20c. This normal distribution-abnormal type distribution positional relationship calculating means 20c calculates the positional relationship TX between the normal distribution and one abnormal type distribution. For example, the positional relationship TA between the normal distribution and the abnormal type A, the positional relationship TB between the normal distribution and the abnormal type B, and so on are obtained.
Here, the positional relationship TX (X = A, B, C,...) Is a numerical value on the feature amount.
TX = normal (average + 3σ) −abnormality type X (average−3σ)
It can ask for. The average can be changed to the median, 3σ can be changed to the quartile, or nXσ (n = 1, 2,...) Can be changed.

  The positional relationship TX with each abnormality type obtained by the normal distribution-abnormal distribution positional relationship calculating unit 20c is given to the threshold value determining unit 20d. The threshold value determination unit 20d confirms the sign of TX and obtains ΔX according to the following rules.

  When TX is negative, the normal distribution and the abnormal type distribution are partially overlapped, so the intermediate position of TX is set to ΔX. Specifically, it calculates | requires by the following formula.

ΔX = 1/2 {normal (average + 3σ) + abnormality type X (average−3σ)}
When TX is 0 and positive, since both distributions do not overlap, ΔX is set on the distribution side of the abnormal type X. Specifically, it calculates | requires by the following formula.

ΔX = abnormality type X (average −3σ)
Then, for each abnormality type, each ΔX is obtained and the minimum is set to a threshold value Δ (Δ = min (ΔX)). In the above formula, the average can be changed to the median value, or 3σ can be changed to the quartile or n × σ (n = 1, 2,...).

  FIG. 20 shows another configuration of the instantaneous stop threshold value setting means 20. In the example shown in FIG. 19, a normal distribution and an abnormal (dummy NG) distribution are required. In this example, the distribution can be set based on one of the distributions. Specifically, the data stored in the time division database 18 is collected, and the values of all the feature values of the engine 1 applied to the inspection device 10 are totaled. Specifically, the standard deviation σ of the feature value is obtained.

  The instantaneous stop threshold setting means 20 includes various registration means 20f. Specifically, (1) discard cost A0 of the engine 1, (2) threshold value Δ0 (arbitrary feature amount level) for discard determination, and (3) rework cost A of the engine 1 are registered. Here, the disposal cost is a cost associated with disposal processing when it is determined to be abnormal (defective). For example, there are costs for manufacturing and costs for disposal. The rework cost is a cost required for replacing the parts of the engine 1 determined to be abnormal (defective) and making it a non-defective product.

  The feature amount collected by the feature amount totaling unit 20e and the registration information input by the registration unit 20f are given to the loss function calculation unit 20g in the next stage. This loss function calculation means 20g calculates the loss function L based on the following formula.

L = (A0 / Δ0 ^ 2) × σ ^ 2
Then, the loss function L obtained in this way is given to the next-stage threshold value calculation means 20h, where the threshold value Δ is calculated based on the following equation.

Δ = (A / A0) ^ (1/2) × Δ0

  Here, the evaluation function L will be described. The change in characteristics is a change in quality, but by indicating this as a loss amount, quality is made to function as a management index. That is, management is performed with a threshold value that balances the cost of maintaining the quality in the process of interest and the sum of the amount of quality damage occurring in the subsequent process to the minimum.

  In other words, when quality is considered as the cost for functional stability, quality can be defined by economic loss, and the management target (threshold value) can be determined by the amount. For example, if the function is not stable, it will be complained and damaged. This is a user's damage. On the other hand, if the stability of the function is examined and revised or discarded there, it is an economic loss of the process of interest. Further, if a product is shipped without causing a loss in the process, the loss increases at the user's stage. Conversely, if the loss is completely absorbed by repairing the process, the loss at the process stage is maximized. What is important is to balance and minimize these two losses.

  The above two loss balances can be expressed as shown in FIG. That is, the economic loss increases as the quality of a certain feature amount Y deteriorates. And generally, when quality (functionality) deteriorates, loss rises rapidly. Then, when the threshold value Δ is obtained from the above-described equation, the area of the OK side region α and the area of the NG side region β are equal, and the two losses can be balanced and the loss can be minimized.

  Further, according to this method, the threshold can be set only with one history type such as only normal data or only abnormal data. Therefore, it is preferable to determine the threshold value by such a method during a period in which only normality determination is performed.

  As described above, in the present embodiment, the determination model is created based on normal state data as shown in S4 and S5 in FIG. 2, but in order to create a more accurate determination rule model, Data in an abnormal state (dummy NG data) may be used.

  FIG. 22 shows the internal configuration of the inspection apparatus 10 during inspection operation. During the inspection operation, basically, among the algorithm creation functions shown in FIG. 4, functions necessary only for creating a judgment model and adjusting a judgment model are not required. Specifically, the digitizing means adjusting means 9, the dummy NG generating means 12, and the determination model generating means 19 are not necessary. If only the abnormality detection is taken into consideration, the editing / deleting means 16 and the like are not required, but are provided to enable manual correction when the judgment of the inspection apparatus is wrong.

  Then, when focusing on the inspection operation, since waveform data is continuously input from various sensors during the durability test of the engine 1, the waveform data dividing means 11 divides the waveform in unit time (N seconds). Thus, a divided waveform is generated and passed to the waveform data digitizing means 14 at the next stage. In this division process, the division is performed sequentially (regularly) in order from the top. However, when a model is created (N is set randomly at the time of algorithm creation, a plurality of types of unit times are provided, and divided waveform data is generated based on the unit time, divided waveform data is created for each unit time. Each divided waveform data composed of a plurality of types of unit times (for example, 1 second, 2 seconds, 3 seconds,...) Is created.

  The waveform data digitizing means 14 digitizes the divided waveform data created based on the sensing data acquired during the endurance test of the engine 1 to obtain a feature amount, and stores the feature amount in the feature amount / history database 15. Store. In the case where determination by a person is simultaneously performed on the same target, the determination result of the person may be stored in the feature quantity / history database 15 as history information. However, in the endurance test, unlike the normal in-line inspection, the engine 1 is continuously rotated for a long time, so a person (inspector) is stationed and listens to the sound generated from the engine 1 at all times. There is no such thing as inputting judgment continuously. Therefore, it is useful to determine the accuracy of the determination model of the analysis / inspection apparatus 10 after that by making an auxiliary determination at an appropriate timing (for example, every scheduled time) and registering the determination result as auxiliary data.

  The feature amount data stored in the feature amount / history database 15 is given to the determination execution means 21, where state determination (abnormality determination) is performed. The obtained state determination result is displayed on the display unit 23, output to the output unit 24, or stored in the result storage unit 25.

  Next, the effect at the time of test operation is demonstrated, giving a specific example. After generating the algorithm (judgment model), the test profile is repeatedly executed as the actual endurance test. Abnormalities are detected in real time by evaluating the degree of normality for each unit time (N seconds) from the beginning (0 second) to the end (X seconds) of the profile. In this embodiment, since the basic for instantaneous stop is set to 1 second, the result (feature value) obtained by quantification from a waveform having a length of 1 second is obtained. The feature quantity performed at this time is in the form of row vector data as shown in FIG. This row vector data is generated continuously during the test in 1 second increments.

  In the case of FIG. 23, since it is rotating at a constant speed, the same determination model (“determination model 1” in the illustrated example) is used (see FIG. 23B). Then, the normality degree is calculated for each divided waveform data (see FIG. 23C). As described above, it is set to 1 if it is close to normal data for the sake of convenience, and as the degree of abnormality increases, the degree of divergence increases and the determination model is created so as to depart from 1 (take a larger value). In the case of the durability test, as shown in FIG. 23, there are many values near 1 at the beginning of the test.

  Further, in this embodiment, since the models of 3 seconds and 6 seconds are set in addition to 1 second required for instantaneous stop as the unit time (N seconds), the normality degree is calculated based on the respective divided waveform data. (See FIGS. 23D and 23E).

  On the other hand, as shown in FIG. 24, for example, it fluctuated in the vicinity of 1 until the 4th second, but the normality at the 5th second becomes 2.3, becomes 3.4 at the 6th second, and gradually increases from 1. I'm away. If the abnormality determination threshold is less than 2.3, a test machine stop command is output at the determination time point in the fifth second. As a matter of course, in such a case, since the durability test of the engine 1 is stopped, the waveform data at the 6th second is not input, and therefore the determination process is not performed (in FIG. 24, the threshold value is less than 2.3). Since it was not, it is judged until the 6th second).

  Although the specific example described above shows a case where the determination model used for each divided waveform data is common because of constant speed rotation, the present invention is not limited to this, and the operation changes in speed. It can also be applied to profiles. That is, for example, for the convenience of explanation, as shown in FIG. 25, the rotational speed of the engine 1 is accelerated in the first second, and is rotated at a constant speed in the second and third seconds. → Assume that there is an operation profile that repeatedly executes an operation of deceleration → acceleration → deceleration in the sixth second.

  In order to create a determination model (algorithm) in such a case, first, the waveform data dividing unit 11 divides the normal waveform data acquired in the initial stage every second. When the speed (number of rotations) changes in a short time as described above, the division start time is not changed randomly. However, for example, when the same operation continues for a certain period, a method of randomly changing the division start position may be added. Furthermore, in the example of FIG. 25, the divided waveform data in units of one second are in the same operation state, but this is not necessarily the case in the present invention. In other words, in the case of accelerating for 0.5 seconds during one second and then rotating at a constant speed for 0.5 seconds, the operation changes in one divided waveform data in this way. Since it can be extracted as a feature, there is no problem if the quality is determined at the same time.

  Then, the waveform data digitizing means divides the acquired divided waveform data into frames and obtains feature quantities. By adopting a method of randomizing the frame size when dividing the frame, a highly accurate determination model can be created even if only waveform data for one second can be acquired. Furthermore, in practice, in order to generate a determination model, a test profile as shown in FIG. 25 can be repeated several times to collect waveform data at the same timing. Thereby, a more accurate algorithm (determination model) can be created. As an example, in FIG. 26, the waveform data for eight times is collected for each divided waveform data, and the number of data sufficient for model creation can be obtained by repeating further. Then, determination models 1 to 6 are created for the first to sixth seconds, respectively.

If the respective models 1 to 6 from the first second to the sixth second are created by the above-described method, the process proceeds to an actual determination process in the durability test. Judgment is performed using a judgment model corresponding to each state of the motion profile. For this timing, for example, each divided waveform data can be cut out accurately by acquiring a timing signal from the PLC 3. When the determination result as shown in FIG. 27 is obtained, since it is a value around 1 for any determination model, it can be determined that it is operating normally.

  On the other hand, if the duration time of the durability test is increased as in FIG. 24, the determination result based on the determination model 5 at the 5th second becomes 2.3 as shown in FIG. The determination result based on is 3.4. In this case, since the judgment models are different, it is unclear whether it is gradually moving away from 1 or whether different parts are about to be primary damaged by acceleration (5th second) and deceleration (6th second). But anyway, there is no doubt that weak parts are about to break or may be damaged.

  Then, if the abnormality determination threshold for all determination models is less than 2.3, a test machine stop command is output at the determination time of the determination model 5. As a matter of course, in such a case, since the durability test of the engine 1 is stopped, the waveform data at the 6th second is not input, and therefore the determination process is not performed (in FIG. 28, the threshold value is less than 2.3). Since it was not, it is judged until the 6th second). The threshold value can be set for each determination model.

  FIGS. 29 and 30 show display examples of changes in normality, that is, usage methods as durability monitoring equipment. In the illustrated example, a certain moment is set to 0 second, and after 2 seconds, some minor abnormality (the judgment value is increased to 2.1) is detected (the tone is changed for a moment), and then returns to normal. Then, an example is shown in which the judgment value suddenly deteriorates to 10 or more (the timbre changes rapidly) after the minor abnormality is detected. In such a case, whether or not a minor abnormality with a determination value of 2.1 is a formal abnormality for instantaneous stop is determined by the user's way of thinking (in the case of 2.1, a sign of abnormality) In other cases, it may be due to fluctuations in normal operation and other errors, and may not always be a precursor to primary damage, or vice versa.) Therefore, for example, the threshold can be determined by calculating the base of the normality distribution (A) at the normal time (initial state). For example, when the normal range is an average value ± 3 × standard deviation, 99.7% corresponds to the normal range, and a range exceeding the normal range can be regarded as abnormal.

  In the application example (endurance test system) shown in FIG. 1, the waveform data of the analog signal is output from each sensor, and is supplied to the inspection apparatus 10 that has been A / D converted by the A / D converter 6. However, if the information processing device is built in the sensor itself and is output as a digital signal, the waveform data (digital) may be supplied to the inspection device 10 as it is or via a D / D converter. good. Furthermore, depending on the capabilities of the information devices, some functions of the inspection device 10 may be executed by the information processing device.

  In FIG. 1, for the sake of convenience of explanation, there is one endurance test target composed of a combination of the drive device 2 and the engine 1 connected to the inspection device 10 and monitored for abnormalities. In some cases, an endurance test for the test object is simultaneously performed. In such a case, an inspection apparatus may be connected to each inspection object. However, it is inefficient and it is not preferable to install a plurality of inspection apparatuses 10 because a lot of cost and installation place are required. Therefore, a single inspection apparatus 10 may monitor inspections for a plurality of durability test objects. In this case, if the engine 1 and the drive device 2 are both of the same type, it may be possible to share the judgment model, but preferably normal data obtained immediately after the start of the durability test for each of them. Based on the above, a decision model algorithm is created, and each is individually judged. This means that even if the model is the same, if the installation location is different, the surrounding environment will be different, so it is necessary to consider the influence of the surrounding sound, etc. Even if the model is the same, it is expected that there will be some variation, so it is more accurate to create a decision model algorithm suitable for each and perform the decision process based on it .

  Furthermore, in the above-described embodiments and modifications, the inspection apparatus is installed at the site where the durability test is performed, but the present invention is not limited to this, and the inspection apparatus itself is a development site. Is connected to a computer (information gathering device) installed at a separate remote location using a high-speed dedicated line or other communication network, and the inspection device 10 acquires waveform data from the development site in real time. It is also possible to make a pass / fail determination (abnormality determination) and notify the result (especially an emergency stop due to an abnormality) using the same or different communication network. In particular, when manual setting is required for setting or adjusting a determination model, according to the present embodiment, it is possible to perform concentrated processing at a monitoring center or the like without going to the site.

  In addition, information is collected in an information collection device such as a data logger, and judgments are made. The judgment results are sent together with other collected data to a computer in an external analysis center via a communication network, where abnormalities occur. The cause may be identified and returned to the site.

As described above, in this embodiment, in order to create a model, three random processing units are provided so that an abnormality can be detected accurately in a short time.
(1) a processing unit that randomly determines where to extract unit time (N seconds) data from the waveform data;
(2) a processing unit that randomly determines a plurality of patterns of time length of a frame when further dividing the waveform as a frame when digitizing the extracted waveform;
(3) Processing for cutting out a waveform having a length of N seconds separately from a waveform cut out every second when it is necessary to capture a swell tendency over one second or more even in a determination every second There is a department. A process of cutting out a plurality of N seconds at random is performed.

  Although the inspection apparatus 10 of the above-described embodiment has been described as an example used for a durability test, the present invention is not limited to this, and the conventional inspection field for abnormal noise, assembly errors, and output characteristics has been used. Applicable. It can also be applied to offline production lines, such as mass production lines, where prototypes are inspected separately from mass production. More specifically, the inspection apparatus 10 according to the present embodiment includes, for example, an inspection device for an automobile drive module such as an automobile engine (sound) and a transmission (vibration), an electric door mirror, an electric power seat, and an electric column. It can be used as an inspection machine for motor actuator modules of automobiles such as (positioning of steering wheel), an evaluation device for abnormal noise, assembly error, output characteristics in the above development, and an evaluation device for a prototype under development.

  It can also be applied as an inspection machine for motor-driven home appliances such as refrigerators, air conditioner indoor / outdoor units, washing machines, vacuum cleaners, printers, etc., and as an evaluation device for abnormal noise, assembly errors and output characteristics in the above development.

  Furthermore, the present invention can also be applied as equipment diagnosis equipment that performs equipment state determination (abnormal state / normal state) such as NC processing machines, semiconductor plants, and food plants. This is because, in facility diagnosis, the creation of a judgment formula (judgment rule) for the presence / absence of an abnormality based on the sample data at the time of an abnormality is based on the default fact / fixed idea. The idea is to try to determine whether it is abnormal. Immediately after the installation of equipment, it is usually used while adjusting the equipment (or adjusting / changing the operating parameter settings), so an “abnormal condition” occurs in an unstable manner. This can be eliminated by performing maintenance and adjusting the equipment.

  In other words, some of the abnormal conditions can be prevented from occurring after the solution equipment is in a stable operation period. This means that some of the “abnormal conditions” of the equipment are not generated and some of the “defective products” of the inspection object are not generated. This means that the present invention can be applied as an equipment diagnosis apparatus for performing equipment state discrimination (abnormal state / normal state). When applied to the equipment diagnosis apparatus, the “initial state” corresponds to a stage before the equipment is stably operated. In addition, with regard to abnormality type knowledge, since the operation of the equipment is stabilized, the parts of the equipment that require periodic maintenance adjustments are identified due to changes over time in the equipment itself. What is necessary is just to specify a state (two with abnormality and abnormality type), and to produce | generate abnormality determination knowledge based on the data for every abnormality type. If a solution is applied to the abnormality determination knowledge and the problem does not occur, the abnormality type knowledge of the abnormality type may be deleted, and the determination process may be performed in the deleted state.

  Further, the equipment is not limited to a plant or the like, but can be applied as a diagnostic device that determines the state of various articles including vehicles such as cars and airplanes. For example, taking a vehicle as an example, normal knowledge is generated based on only normal state data about the engine state at the prototype stage. Naturally, an abnormal state occurs at the time of the trial production, but some of the abnormal states are not generated by the trial production improvement. Therefore, at the initial stage of prototyping, a decision rule is created only from normal data, and the improvement of the prototyping is progressed to resolve some abnormal conditions so that they do not occur. Identify and generate abnormality type knowledge from the data of the abnormal state. By doing so, it becomes possible to determine a normal state and a specific abnormal state. In this way, data and knowledge are accumulated from the prototype stage, and using the normal knowledge and abnormality type knowledge, it is possible to create a diagnostic device that determines whether it is normal and which type of abnormality, and that diagnostic device is a finished product. As a result, it can be mounted on vehicles and airplanes on the market and diagnosed as normal or abnormal based on engine vibration.

It is a figure showing the schematic structure of the endurance test system which is one of the suitable one embodiments of the present invention. It is a figure explaining the whole flow of the endurance test performed in this embodiment. It is a figure which shows the schematic image of the determination algorithm production | generation of this embodiment, and the determination process based on it. It is a figure which shows the functional block for producing a determination algorithm. It is a figure explaining the function of a waveform data division | segmentation means. It is a figure explaining the function of a waveform data division | segmentation means. It is a figure which shows the internal structure of a dummy NG production | generation means. 3 is a diagram illustrating an example of a data structure of a feature quantity / history database 15. FIG. It is a figure explaining the function (frame division | segmentation) of a numerical processing part. It is a figure explaining the function (frame division | segmentation) of a numerical processing part. It is a figure explaining the function (frame division | segmentation) of a numerical processing part. It is a figure explaining the necessity of the function (frame division | segmentation) of a numerical processing part. It is a figure which shows the internal structure of a waveform data digitization part. It is a flowchart which shows a part of function of a waveform data digitization part. It is a figure which shows the internal structure of a determination model production | generation means. It is a figure which shows the internal structure of a determination model production | generation means. It is a figure which shows the internal structure of a determination model production | generation means. It is a figure which shows the internal structure of a determination model production | generation means. It is a block diagram which shows an example of the internal structure of an instantaneous stop threshold value setting means. It is a block diagram which shows an example of the internal structure of an instantaneous stop threshold value setting means. It is a figure explaining the principle of operation of the instantaneous power failure threshold setting means shown in FIG. It is a block diagram which shows suitable one Embodiment of the inspection apparatus (at the time of test | inspection operation) which concerns on this invention. It is a figure explaining the function of one suitable embodiment of the inspection device (at the time of inspection operation) concerning the present invention. It is a figure explaining the function of one suitable embodiment of the inspection device (at the time of inspection operation) concerning the present invention. It is a figure explaining the judgment model creation function based on another operation profile. It is a figure explaining the judgment model creation function based on another operation profile. It is a figure explaining another embodiment at the time of the test operation based on another operation profile. It is a figure explaining another embodiment at the time of the test operation based on another operation profile. It is a figure explaining the principle of threshold value setting. It is a figure explaining the principle of threshold value setting.

Explanation of symbols

9 Digitizing means adjusting means 10 Inspection device 11 Waveform data digitizing means 12 Dummy NG generating means 13 Recorded waveform display means 13 'Registered content display means 13 "Digitization result display means 14 Digitizing means adjusting means 14 Record history of waveform data Means 15 Feature quantity / history database 16 Editing / deleting means 17 History type classification means 18 Time division database 18a Normal database 18b Abnormal database 19 Determination model generation means 20 Instantaneous power failure threshold setting means 21 State determination means 22 State determination expression update determination means 23 Display Means 24 Output means 25 Result storage means

Claims (12)

It is a determination model creation support device for an inspection device that extracts a feature amount from waveform data acquired from an inspection target product and determines the state of the inspection target product based on the extracted feature amount,
Divided waveform data creating means for dividing the acquired waveform data into data for a set unit time; and
Based on a plurality of divided waveform data for a unit time created by the divided waveform data creating means, means for calculating a feature amount in each divided waveform data unit;
Model creation means for creating a determination model for determining the state of the product to be inspected every unit time based on the feature amount calculated in the divided waveform data unit;
A determination model creation support apparatus for an inspection apparatus, comprising:   2. The determination model creation support apparatus for an inspection apparatus according to claim 1, wherein the divided waveform data creation means has a function of dividing waveform data from the predetermined position every unit time.   3. The determination model creation for an inspection apparatus according to claim 1, wherein the divided waveform data creation means has a function of randomly determining a division start position at the time of division for each unit time. Support device.   4. The inspection apparatus according to claim 1, wherein the divided waveform data creating unit has a function of randomly determining a unit time serving as a reference length for performing the dividing process. 5. Judgment model creation support device. Judgment model creation support device for an inspection device that extracts a feature amount from a waveform signal acquired from an inspection target product whose operation repeatedly varies in a predetermined pattern and determines the state of the inspection target product based on the extracted feature amount Because
Divided waveform data creating means for dividing the acquired waveform data into data for a set unit time; and
Based on a plurality of divided waveform data for a unit time created by the divided waveform data creating means, means for calculating a feature amount in each divided waveform data unit;
An inspection characterized by comprising model creation means for creating a determination model for each region in the pattern divided by the divided waveform data creation means based on the feature amount calculated in units of the divided waveform data Judgment model creation support device for a device.   The means for obtaining the feature amount includes a function of performing frame division processing for further dividing the divided waveform data into frames of a predetermined size, and extracting a feature amount based on the divided frames, The determination model creation support apparatus for an inspection apparatus according to any one of claims 1 to 5, further comprising a function of determining at random.   The means for obtaining the feature amount performs frequency analysis on each frame obtained by executing the frame division processing, and obtains the feature amount in all frequency bands. A determination model creation support device for the described inspection device. Based on the determination model created by the determination model creation support device according to any one of claims 1 to 7,
An inspection apparatus comprising determination means for performing state determination on acquired waveform data to be inspected. An inspection apparatus for a device whose operation repeatedly changes in a predetermined pattern,
The determination model creation support device according to claim 5 acquires a determination model for each region created based on waveform data corresponding to one pattern portion that repeatedly changes,
An inspection apparatus comprising determination means for performing determination processing using a determination model of an area corresponding to each acquired waveform data portion when determining the state of an inspection object. An endurance test device for testing the durability of an inspection object,
Using a determination model created as normal data of waveform data acquired during the period when the operation after the endurance test is stable, it is determined whether there is an abnormality based on the waveform data acquired during the execution of the endurance test. An endurance test apparatus comprising a determination means.   The durability test apparatus according to claim 10, wherein the determination model is created by the determination model creation support apparatus according to claim 1. A processing step of rotating a driving body to be subjected to an endurance test for a predetermined time, and supplying waveform data acquired at that time to the determination model creation support device according to any one of claims 1 to 7,
The determination model creation support device creates a determination model for abnormality detection using the given waveform data as normal data, and
Thereafter, a process step of performing an endurance test and determining whether there is an abnormality using the determination model created by the determination model creation support device based on the waveform data acquired during the execution of the endurance test;
An endurance test method for executing a processing step of outputting an abnormality detection signal when an abnormality is detected.

Images