TWI587294B - Detection method of abnormal sound of apparatus and detection device - Google Patents

Description

Device abnormal sound detecting method and detecting device

The invention relates to a method and a device for detecting abnormal sound of a machine, in particular to a method and a device for detecting an abnormal sound based on SVM linear classification and cluster analysis.

Automated testing of abnormal sounds has been a problem in factory testing for the past few years. The noise emitted by the product during normal operation is normal noise or abnormal noise (alien sound). Past production tests are judged by skilled workers, including fans, transformers, power products, and projections. Sound testing by equipment, motors and other equipment. The sound test must rely on the judgment of the operator, and the human factor is the most difficult factor to control in the detection system, which restricts the realization of the zero defect goal of product quality. At the same time, the necessary personnel work also limits the realization of production automation.

There are several ways to test the current sound: using a sound power meter. This method can measure the total power of the volume or the total volume. If the total volume is too high, it can be detected as bad. However, the abnormal sound is usually an abnormality from a certain component, and the power meter cannot distinguish the components of the sound. If the component of the anomaly is not the main power component, it is usually not detected.

Use a spectrum analyzer to measure. The spectrum analyzer can view the component values of each frequency, especially for the detection of narrowband signals. For example, a single tone from a buzzer can be easily transmitted. The spectrum analyzer is detected. Spectrum analysis is also the main tool used by engineering units to make abnormal sound decisions. However, most of the abnormal sound modes are broadband noise, and the components at each single frequency are not significant. Therefore, it is difficult for the spectrum analyzer to distinguish the form of broadband noise. This method has a significant false positive rate and cannot provide high detection. Out of effectiveness.

Use the sound power spectrum. This measurement method combines the measurement of sound power and spectrum, and is measured using the weighting method of various octave power spectra. This measurement is closer to the human ear's response to sound, but its basic limitations are the same as the first two. Normal noise and abnormal noise have little difference in the performance of the power spectrum, and the range of normal noise may exceed this small gap, making the detection capability very limited.

Use sound quality measurements. This includes measurements with sound quality factors such as Loudness, roughness, and sharpness. The weighting of these sound qualities simulates the human response, transforming the physical quantity into the human ear, and is closer to the person's perception of loudness, roughness, and sharpness. However, these sound quality factors are actually unable to effectively characterize the abnormality of noise. It is essentially weighted on the basis of the spectrum or power spectrum, and its detection limit factor is the same as the spectrum or power spectrum.

Resonance demodulation technology using an accelerometer. This is a method of replacing the abnormal sound test with vibration. At the resonant frequency of the accelerometer, if a continuous excitation is added, the resonance of the accelerometer is excited. The source of this continuous incentive is the abnormal sound. This method has proven to be effective in the measurement of abnormal sounds in rotating machinery, especially in the case of Palin abnormalities. A Palin anomaly triggers a series of broadband noises whose frequency components cover the resonant frequency of the accelerometer and excite resonance. The excitation frequency signal can be demodulated to obtain the excitation source characteristic frequency. This method is considered effective in some tests, but it is not very versatile. Effective detection is not possible for most exception modes.

Measurements were made using standard non-sounding chambers. This measurement method requires a standard non-sound chamber and a high-level sound measurement system. The advantage is that the background noise in the non-sound chamber is suppressed to a low level, and the measurement signal can be measured with a low-noise sound acquisition system. The background noise is minimized, so that the ability to discriminate the wideband signal of the abnormal sound can be mentioned to the maximum. This method is expensive only for equipment and is suitable for use in research and analysis, and is difficult to apply to manufacturing.

However, regardless of which of the above test methods is used, the existing test method must have a predetermined standard value on the obtained measurement value. For example, spectrum measurements can yield response values for each frequency. The general concept is to study the analysis of normal and abnormal products and try to find out the spectral specifications on the sound measurement. However, the essence of the abnormal sound is not the single-tone, but the various broadband components are generated at the same time, so that the components of different frequencies do not have independence, and it is theoretically impossible to set the specification value of each frequency component. This feature explains why in the past attempts to measure and build specifications in a variety of ways, but it has never been possible to find an effective way to judge.

Based on the above, due to the dependence between the frequency components, it is necessary to establish a reasonable specification value, and it is not possible to analyze each frequency component, and at least the linear combination of the components should be analyzed. Therefore, the existing linear classification algorithm with reference statistics is used for classification analysis. The algorithm of the linear classifier needs to first establish a good sample and a defective sample. Using a known good product feature and a known defective product feature, a classification surface is calculated, and the sample to be tested is classified and determined. There are many statistical classification algorithms that can be used, including linear regression, logistic regression, generative algorithms, Bayesian algorithms, and SVM classification algorithms. In this project, SVM (Support Vector Machine) classification algorithm is used, and SVM is considered to have greater robustness in many papers. The statistical classification algorithm does not calculate the specifications for each feature individually, but classifies and determines the linear combination of features. Unvoiced sound is the basic principle of multi-spectral nature. After preliminary analysis, the SVM can effectively determine the sample sound of a given database.

Actually, when used, it is found that some significant abnormal sounds are misjudged in the classification algorithm. One possible reason is that SVM is a teaching learning algorithm that needs to learn the construction of the classification surface from the teaching of the database, so adding these wrong judgments to the database may solve the problem. However, in fact, these significant abnormal sounds are added to the sample database, but they confuse the classification surface and reduce the effectiveness of the SVM classification. To explore this phenomenon, the possible reason is to use the statistical linear classifier, which is based on the linear separability of the hypothetical classification object. However, the linear separability cannot be 100% satisfied. When the condition of linear separability is not met, the SVM will Correct classification judgment cannot be made. Significant abnormal sounds are significantly outliers, and the distribution of the main sound categories in the feature space is far away. The linear separable assumption is no longer valid, so it is added to the database to facilitate the construction of the classifier.

In view of the problems existing in the prior art, an object of the present invention is to provide a method and apparatus for automatically detecting abnormal sound of a device.

The method for detecting abnormal sound of a device of the present invention comprises the following steps: collecting a sound signal when the device is running; preprocessing the collected sound signal to obtain a sound processing signal; and extracting a plurality of characteristic parameters from the sound processing signal Performing cluster analysis and SVM linear classification on the extracted plurality of feature parameters and samples in the database; The sound signal is predicted to be an abnormal sound according to the cluster analysis and the SVM linear classification result.

Further, pre-processing the collected sound signal comprises: converting the collected sound signal into a standard sound pressure value in a frequency domain according to a sensitivity of the microphone; and performing calibration processing on the standard sound pressure value by using a weighting factor.

Further, the clustering analysis comprises: performing cluster analysis according to the extracted Euclidean distances between the plurality of feature parameters and the samples in the database; and determining, according to the result of the cluster analysis, the plurality of feature parameters Whether the sound signal and the samples in the database are out of range.

Further, the clustering analysis further comprises: calculating an Euclidean distance between the plurality of feature parameters and a sample in the database in the feature space; and when the calculated Euclidean distance is greater than or equal to a threshold, the clustering The result of the analysis is outliers; when the calculated Euclidean distance is less than the threshold, the result of the cluster analysis is not outlier.

Further, the samples in the database include normal sound samples and abnormal sound samples.

Further, the detection method further includes: A normal sound sample and an abnormal sound sample of the device are established; and a classification surface in the feature space is calculated according to the feature parameter in the normal sound sample and the feature parameter in the abnormal sound sample.

Further, the SVM linear classification comprises: performing SVM linear classification according to the extracted positional relationship between the plurality of characteristic parameters and the classification surface; determining, according to a result of the SVM linear classification, that the sound signal corresponding to the plurality of characteristic parameters is located The side of the abnormal sound sample of the classification surface or the side of the normal sound sample of the classification surface.

Further, the linear classification of the SVM further includes: calculating a relative positional relationship between the plurality of feature parameters and the classification surface; and when the plurality of feature parameters are located on an side of the abnormal sound sample of the classification surface, corresponding to the plurality of characteristic parameters This sound signal is classified as an abnormal sound.

Further, when the result of the cluster analysis is an outlier, the sound signal is predicted to be an abnormal sound according to the result of the cluster analysis; when the result of the cluster analysis is not outlier, the result is predicted according to the result of the SVM linear classification. Whether the sound signal is abnormal.

The apparatus for detecting abnormal sound of the device of the present invention comprises: an acquisition unit, configured to collect a sound signal when the device is in operation; and a pre-processing unit, configured to pre-process the collected sound signal to obtain a sound processing signal; An extracting unit, extracting a plurality of feature parameters from the sound processing signal; an analyzing unit, configured to perform cluster analysis on the extracted plurality of feature parameters and samples in the database; and a classifying unit configured to extract the extracted The feature parameters and the samples in the database are subjected to SVM linear classification; and the prediction unit is configured to predict whether the sound signal is an abnormal sound according to the cluster analysis and the SVM linear classification result.

Further, the pre-processing unit includes: a conversion unit that converts the collected sound signal into a standard sound pressure value in a frequency domain according to a sensitivity of the microphone; and a calibration unit that performs a calibration process on the standard sound pressure value by using a weighting factor.

Further, the analyzing unit includes: a first calculating unit, configured to calculate an Euclidean distance between the plurality of feature parameters and a sample in the database in the feature space; the first result output unit, when the calculated Euclidean distance When the threshold is greater than or equal to a threshold, the result of the cluster analysis is outlier.

Further, the samples in the database include normal sound samples and abnormal sound samples.

Further, the method further includes: a database construction unit that establishes the normal sound sample of the device and the abnormal sound sample; The second calculating unit calculates a classification surface in the feature space according to the feature parameter in the normal sound sample and the feature parameter in the abnormal sound sample.

Further, the classifying unit includes: a second calculating unit, configured to calculate a relative positional relationship between the plurality of feature parameters and the classifying surface; and a second result output unit, wherein the plurality of feature parameters are located in the abnormal sound sample of the classifying surface On the side, the sound signal corresponding to the plurality of feature parameters is classified as an abnormal sound.

The invention detects the abnormal sound of the device by the combination of cluster analysis and SVM linear classification, so that the nonlinear problem of the sound sample in the feature space can be excluded by cluster analysis, and the small range satisfies the linear hypothesis. The sound samples are classified using the SVM linear classification. Compared with the prior art, it has the advantages of simple operation and high prediction accuracy.

S11~S12‧‧‧Steps

S21~S25‧‧‧Steps

S221~S222‧‧‧Steps

31‧‧‧Building unit

32‧‧‧Computation unit

4‧‧‧Sound detection device

41‧‧‧ acquisition unit

42‧‧‧Pretreatment unit

43‧‧‧Extraction unit

44‧‧‧Analysis unit

45‧‧‧Classification unit

46‧‧‧ Forecasting unit

5‧‧‧Sound detection device

1 is a schematic flowchart of a method for detecting an abnormal sound of a device according to an embodiment of the present invention; FIG. 2 is a schematic flowchart of a method for detecting an abnormal sound of a device according to another embodiment of the present invention; FIG. 3 is a schematic diagram of a device different according to an embodiment of the present invention; FIG. 4 is a schematic flowchart of clustering analysis in a device abnormal sound detecting method according to an embodiment of the present invention; FIG. 5 is a schematic diagram of device abnormal sound detecting according to an embodiment of the present invention; FIG. 6 is a schematic structural diagram of a device for detecting abnormal sound of a device according to an embodiment of the present invention; FIG. FIG. 7 is a schematic structural diagram of a device for detecting abnormal sound of a device according to another embodiment of the present invention.

As shown in FIG. 1 , a method for detecting abnormal sound of a device according to an embodiment of the present invention includes the following steps:

Step S21: collecting a sound signal when the device is running. The sound signal of the collecting device during operation can be obtained by an acoustic sensor installed on the device.

Step S22: pre-processing the collected sound signal to obtain a sound processing signal;

Step S23: extracting a plurality of feature parameters from the sound processing signal;

Step S24: Perform cluster analysis and SVM linear classification on the extracted plurality of feature parameters and samples in the database, wherein the samples in the database are divided into normal sound samples and abnormal sound samples, and the normal sound samples and the abnormal sound samples are included. Each includes multiple feature parameters.

Step S25: predicting whether the sound signal is an abnormal sound according to the cluster analysis and the SVM linear classification result.

As shown in FIG. 3, the pre-processing of the collected sound signal in step S22 specifically includes: step S221: converting the collected sound signal into a standard sound pressure value in the frequency domain according to the sensitivity of the microphone, and simply detecting the sound pressure in the time domain. In terms of value, the consistency of the repeated test is not good, and the spectrum or power spectrum calculation can improve the repeatability, and the developed spectrum can be from 0 Hz to 22 kHz, but the invention is not limited thereto; step S222: using the weighting factor pair The standard sound pressure value is calibrated in order to The perceived value of the human can correspond to the high value of the spectrum. For example, the A-weighting can filter out the indistinguishable frequency of the human ear and provide assistance for screening out unwanted signals. The weighting method of the present invention has four options, A-weighting, Direct map, Interpolate weighting, and No weighting. A-weighting is A-weighting on waveforms. The main function of this function is not used to do A weighting weighting, but is used to import the microphone response function after microphone calibration. Direct map is a waveform that is subjected to Fourier transform and directly multiplied by the map array to output the weighted waveform. The direct map needs to confirm the length and frequency interval of the spectrum according to the duration and sampling frequency of the input waveform, and generate the frequency point completely. The corresponding map array is directly multiplied to output the weighted waveform.

As shown in FIG. 4, the clustering analysis in step S25 specifically includes: step S251: calculating an Euclidean distance between the plurality of feature parameters and a sample in the database in the feature space; for example, two n-dimensional vectors a ( Euclidean distance d12 between X11, X12, ... X1n) and b(X21, X22...X2n): Step S252: determining whether the calculated Euclidean distance is greater than a threshold value; Step S253: When the calculated Euclidean distance is greater than Or equal to a threshold, the result of the cluster analysis is outlier; step S254: when the calculated Euclidean distance is less than the threshold, the result of the cluster analysis is not outlier.

As shown in FIG. 2, a method for detecting abnormal sound of a device according to another embodiment of the present invention. its Compared with the detection method of the device abnormal sound shown in FIG. 1, the method further includes:

Step S11: Establish a normal sound sample and an abnormal sound sample of the device. Firstly, the sound signal of the device is normally operated for a certain period of time, and after the pre-processing and feature extraction, a normal sound sample is obtained. Then, the collecting device abnormally runs the sound signal for a certain period of time, and after the above preprocessing and feature extraction, an abnormal sound sample is obtained. However, the normal sound sample and the abnormal sound sample are not necessarily established by the follow-up as in the present embodiment, and the normal sound sample and the abnormal sound sample may also be preset beforehand.

Step S12: Calculate a classification surface in the feature space according to the feature parameter in the normal sound sample and the feature parameter in the abnormal sound sample. The calculation of the classification surface is calculated using the existing linear discriminant function.

As shown in FIG. 5, the SVM linear classification in step S25 specifically includes: step S255: calculating a relative position relationship between the plurality of feature parameters and the classification surface, that is, calculating a plurality of characteristic parameters and a normal sound sample of the classification surface of the Euclidean The distance, and the Euclidean distance of the plurality of characteristic parameters and the abnormal sound sample of the classification surface, are judged to be located on one side of the normal sound sample or on the side of the abnormal sound sample according to the length of the Euclidean distance of the plurality of characteristic parameters; Step S256: Comparing the feature The Euclidean distance between the parameter and the normal sound sample, and the Euclidean distance between the feature parameter and the abnormal sound sample; Step S257: When the Euclidean distance between the feature parameter and the abnormal sound sample is less than the normal sound sample When the Euclidean distance is different, the plurality of characteristic parameters are closer to the abnormal sound sample of the classification surface, and the sound signal corresponding to the plurality of characteristic parameters is classified as abnormal. sound; Step S258: when the Euclidean distance between the feature parameter and the normal sound sample is smaller than the Euclidean distance between the abnormal sound samples, the plurality of feature parameters are closer to the normal sound sample of the classification surface, and are located at the classification surface. On one side of the normal sound sample, the sound signal corresponding to the plurality of feature parameters is classified as a normal sound.

The acquired multiple feature parameters may be nonlinear due to their possible linearity with samples in the database. The SVM linear classification method cannot accurately predict the nonlinear part. Therefore, the non-linear part of the plurality of characteristic parameters is excluded according to the result of the cluster analysis in the embodiment of the present invention, that is, when the result of the cluster analysis in the above step S25 is an outlier, according to the cluster analysis The result directly predicts that the sound signal is an abnormal sound. When the result of the cluster analysis is not outlier, it is proved to be linear with the sample in the database, so that the sound signal can be predicted to be an abnormal sound according to the result of the SVM linear classification, that is, when the plurality of characteristic parameters are located in the classification On the side of the abnormal sound sample of the surface, the sound signal corresponding to the plurality of characteristic parameters is classified as an abnormal sound, and when the plurality of characteristic parameters are located on the side of the normal sound sample of the classified surface, the plurality of characteristic parameters are The corresponding sound signal is classified as a normal sound.

As shown in FIG. 6, the apparatus for detecting abnormal sound of a device according to an embodiment of the present invention includes: an acquisition unit 41 for collecting a sound signal when the device is in operation; and a pre-processing unit 42 for collecting the The sound signal is pre-processed to obtain a sound processing signal; the extracting unit 43 extracts a plurality of characteristic parameters from the sound processing signal; and the analyzing unit 44 is configured to aggregate the extracted plurality of characteristic parameters and the samples in the database Class analysis a classification unit 45, configured to perform SVM linear classification on the extracted plurality of feature parameters and samples in the database; and a prediction unit 46, configured to predict, according to the cluster analysis and the SVM linear classification result, whether the sound signal is an abnormal sound .

The pre-processing unit 42 in this embodiment specifically includes: a conversion unit that converts the collected sound signal into a standard sound pressure value in a frequency domain according to sensitivity of the microphone; and a calibration unit that calibrates the standard sound pressure value by using a weighting factor . The processing method of the conversion unit and the calibration unit has been described in the foregoing abnormal sound detection method, and details are not described herein again. In addition, the structure of the pre-processing unit of the present invention is not limited thereto, and the conversion unit or the calibration unit may be omitted.

The analyzing unit 44 in this embodiment includes: a first calculating unit, configured to calculate an Euclidean distance between the plurality of feature parameters and a sample in the database in the feature space; the first result output unit, when the calculated When the Euclidean distance is greater than or equal to a threshold, the result of the cluster analysis is outlier.

As shown in FIG. 7, a device for detecting abnormal sound of a device according to another embodiment of the present invention. Compared with the device for detecting abnormal sounds shown in FIG. 3, the method further includes: a database construction unit 31, which establishes the normal sound sample of the device and the abnormal sound sample; and a second calculating unit 32, according to the feature in the normal sound sample Parameters and the abnormal sound The feature parameters in the tone samples are calculated to be a classification surface in the feature space.

The classification unit 45 in this embodiment includes: a second calculation unit, configured to calculate a relative positional relationship between the plurality of feature parameters and a normal sound sample and an abnormal sound sample of the classification surface; and a second result output unit, when the plurality of features The parameter is located on the side of the abnormal sound sample of the classification surface, and the sound signal corresponding to the plurality of characteristic parameters is classified as an abnormal sound.

The exemplary embodiments of the present invention have been particularly shown and described above. It is to be understood that the invention is not limited to the disclosed embodiments, but the invention is intended to cover various modifications and equivalents.

S21~S25‧‧‧Steps

Claims (11)

A method for detecting abnormal sound of a device includes the following steps: collecting a sound signal when the device is running; preprocessing the collected sound signal to obtain a sound processing signal; extracting a plurality of characteristic parameters from the sound processing signal; Extracting the plurality of feature parameters and samples in the database for cluster analysis and SVM linear classification, wherein the cluster analysis comprises: according to the extracted plurality of feature parameters and the feature space in the database between the samples Performing the cluster analysis on the Euclidean distance; and determining, according to the result of the cluster analysis, whether the sound signal corresponding to the plurality of feature parameters and the sample in the database are out of group, and the SVM linear classification comprises: according to the extracted Performing the SVM linear classification on the positional relationship between the plurality of characteristic parameters and a classification surface; and determining, according to the result of the SVM linear classification, that the sound signal corresponding to the plurality of characteristic parameters is located on the side of the abnormal sound sample of the classification surface Or the normal sound sample side of the classification surface; and predicting the sound based on the cluster analysis and the SVM linear classification result Whether the number of abnormal noise. The detecting method of claim 1, wherein the pre-processing the collected sound signal comprises: converting the collected sound signal into a frequency domain standard according to sensitivity of the microphone. Sound pressure value; the standard sound pressure value is calibrated using a weighting factor. The detection method of claim 1, wherein the clustering analysis further comprises: calculating an Euclidean distance between the plurality of characteristic parameters and a sample in the database in the feature space; when the calculated Euclidean When the distance is greater than or equal to a threshold, the result of the cluster analysis is outlier; when the calculated Euclidean distance is less than the threshold, the result of the cluster analysis is not outlier. The detection method of claim 1, wherein the sample in the database comprises a normal sound sample and an abnormal sound sample. The detecting method of claim 4, wherein the detecting method further comprises: establishing a normal sound sample and an abnormal sound sample of the device; calculating according to the characteristic parameter in the normal sound sample and the characteristic parameter in the abnormal sound sample; The classification face in the feature space is obtained. The detection method of claim 1, wherein the SVM linear classification further comprises: calculating a relative positional relationship between the plurality of characteristic parameters and the classification surface; and when the plurality of characteristic parameters are located in the abnormal sound sample of the classification surface On one side, the sound signal corresponding to the plurality of feature parameters is classified as an abnormal sound. The detection method of claim 1, wherein: when the result of the cluster analysis is an outlier, predicting the sound signal as an abnormal sound according to the result of the cluster analysis; When the result of the cluster analysis is not outlier, the sound signal is predicted to be an abnormal sound according to the result of the SVM linear classification. A device for detecting abnormal sound of a device, comprising: an acquisition unit, configured to collect a sound signal when the device is in operation; and a pre-processing unit, configured to perform pre-processing on the collected sound signal to obtain a sound processing signal; a plurality of feature parameters are extracted from the sound processing signal; the analyzing unit is configured to perform cluster analysis on the extracted plurality of feature parameters and samples in the database, wherein the analyzing unit comprises: a first calculating unit, configured to calculate the The Euclidean distance between the plurality of feature parameters and the sample in the database in the feature space; and the first result output unit, when the calculated Euclidean distance is greater than or equal to a threshold, the result of the cluster analysis is An out-of-group; a classification unit, configured to perform SVM linear classification on the extracted plurality of feature parameters and samples in the database, wherein the classification unit comprises: a second calculation unit, configured to calculate the plurality of feature parameters and a classification surface a relative positional relationship between the normal sound sample and the abnormal sound sample; and a second result output unit when the plurality of feature parameters are located in the minute The sound signal corresponding to the plurality of feature parameters is classified as an abnormal sound on the side of the abnormal sound sample of the face; and the predicting unit is configured to predict whether the sound signal is different according to the cluster analysis and the SVM linear classification result sound. The detecting device of claim 8, wherein the pre-processing unit comprises: a converting unit, converting the collected sound signal according to the sensitivity of the microphone to The standard sound pressure value in the frequency domain; the calibration unit calibrates the standard sound pressure value by using a weighting factor. The detecting device of claim 8, wherein the sample in the database comprises a normal sound sample and an abnormal sound sample. The detecting device of claim 10, further comprising: a database construction unit that establishes the normal sound sample of the device and the abnormal sound sample; and a second calculating unit, according to the feature parameter in the normal sound sample and the The feature parameters in the abnormal sound sample are calculated to obtain the classification face in the feature space.