JP2023018765A - Anomaly Detection System for Detecting Abnormality in Robot, Method for Detecting Abnormality in Robot, and Computer Program - Google Patents

Abstract

Kind Code: A1 A technology capable of accurately detecting an abnormality immediately after shipment of a robot or after a change in operation is provided. An anomaly detection system of the present disclosure includes a state input acquisition unit that acquires a state input representing a state of a robot; an anomaly score calculation unit that receives the state input and calculates an anomaly score that indicates the degree of an anomaly; a determination unit that determines whether or not an abnormality has occurred by comparing the abnormality score with a threshold. The anomaly score calculation unit consists of a general-purpose anomaly detection model that calculates a general-purpose score that indicates the degree of anomaly independently of the robot's usage environment, and a correction that corrects the general-purpose score by specializing in the robot's usage environment. a specialized anomaly detection model that calculates a value; and a correction unit that calculates an anomaly score by correcting the general score with a correction value. [Selection drawing] Fig. 3

Description

The present disclosure relates to an anomaly detection system for detecting an anomaly in a robot, a method for detecting an anomaly in a robot, and a computer program.

Patent Literature 1 discloses a failure prediction system using a machine learning model for learning failure conditions. This failure prediction system learns failure conditions using data obtained when industrial robots are operated.

JP 2017-120649 A

However, in the above conventional technology, in order for the failure prediction system to exhibit sufficient accuracy, it is necessary to repeat the learning cycle until the machine learning model has learned sufficiently immediately after shipment or after changing the robot operation. There was a problem that the burden on the

A first aspect of the present disclosure provides an anomaly detection system that detects an anomaly in a robot. This anomaly detection system includes a state input acquisition unit that acquires a state input representing the state of the robot, an anomaly score calculation unit that receives the state input and calculates an anomaly score indicating the degree of the anomaly, and the anomaly score. with a threshold to determine whether or not the abnormality has occurred. The anomaly score calculation unit includes a general-purpose anomaly detection model that calculates a general-purpose score indicating the extent of the anomaly independently of the environment in which the robot is used, and a general-purpose score that is specific to the environment in which the robot is used. and a correction unit that calculates the abnormality score by correcting the general-purpose score with the correction value.

According to a second aspect of the present disclosure, a method of detecting an abnormality in a robot is provided. This method comprises the steps of: (a) obtaining a state input representing the state of the robot; (b) receiving the state input and calculating an abnormality score indicating the extent of the abnormality; and determining whether or not the abnormality occurs by comparing the score with a threshold. The step (c) includes (c1) calculating a general score indicating the degree of abnormality without depending on the usage environment of the robot; and (c2) calculating the abnormality score by correcting the general score with the correction value.

According to a third aspect of the present disclosure, there is provided a computer program that causes a processor to execute processing for detecting an abnormality in a robot. This computer program includes (a) a process of acquiring a state input representing the state of the robot, (b) a process of receiving the state input and calculating an abnormality score indicating the degree of the abnormality, and (c) the and determining whether or not the abnormality has occurred by comparing the abnormality score with a threshold value. The process (c) includes (c1) a process of calculating a general score indicating the degree of abnormality without depending on the usage environment of the robot; and (c2) calculating the abnormality score by correcting the general score with the correction value.

Explanatory drawing which shows the structure of the robot system in embodiment. 1 is a functional block diagram of an information processing device according to an embodiment; FIG. FIG. 2 is a block diagram showing the internal configuration of an anomaly detector; FIG. 3 is a block diagram showing an example of the internal configuration of a general-purpose anomaly detection model; 4 is a flowchart showing the procedure of learning processing for a general-purpose anomaly detection model; 4 is a flowchart showing the procedure of abnormality detection processing in a specific environment; Explanatory drawing which shows the example of a change of an abnormality score.

FIG. 1 is an explanatory diagram showing an example of a robot system according to an embodiment. This robot system includes a robot 100 , a control device 200 that controls the robot 100 , and an information processing device 300 . The robot 100 is installed on a pedestal 500 . The information processing device 300 is, for example, a personal computer.

The robot 100 has a base 110 and a robot arm 120 . An end effector 150 is attached to the tip of the robot arm 120 . A vibration sensor 410 and a force sensor 420 are installed at the tip of the robot arm 120 . In the present disclosure, “hand end of robot arm 120 ” means a portion from the tip of robot arm 120 to end effector 150 .

The robot arm 120 is sequentially connected by four joints J1 to J4. A "joint" is also called an "axis". A TCP (Tool Center Point) as a control point of the robot 100 is set near the tip of the robot arm 120 . A “control point” is a reference point for controlling the robot arm 120 . TCP can be set at any position. Controlling the robot 100 means controlling the position and orientation of this TCP. In this embodiment, a four-axis robot in which the robot arm 120 has four joints J1 to J4 is exemplified, but a robot having an arbitrary arm mechanism having one or more joints can be used. Further, the robot 100 of this embodiment is a horizontal articulated robot, but a vertical articulated robot may be used.

Vibration sensor 410 is a sensor capable of detecting vibration of robot 100 . In general, vibration can be detected by measuring time-series changes in any of position, velocity, and acceleration. As the vibration sensor 410, for example, a position sensor, a gyro sensor, an acceleration sensor, an inertial measurement unit, etc. can be used. Note that the vibration sensor 410 may be installed at a position other than the tip of the robot arm 120 . For example, a vibration sensor may be installed on the gantry 500 when the vibration of the gantry 500 is a constraint condition for the optimization process. Note that the vibration sensor 410 can be omitted if vibration is not detected as one type of abnormality.

The force sensor 420 is a sensor that detects force applied to the end effector 150 . As the force sensor 420, it is possible to use a load cell capable of detecting force in a single axial direction, or a force sensor or torque sensor capable of detecting force components in multiple axial directions. In this embodiment, a 6-axis force sensor is used as the force sensor 420 . A six-axis force sensor detects the magnitude of force parallel to three mutually orthogonal detection axes and the magnitude of torque around the three detection axes in a unique sensor coordinate system. Note that the force sensor 420 may be provided at a position other than the tip of the robot arm 120, and may be provided at one or more of the joints J1 to J4, for example.

In this embodiment, the robot 100 performs the work of assembling the two works WK1 and WK2 by fitting the first work WK1 into the hole HL of the second work WK2. In this work, force control using force detected by the force sensor 420 is executed. Note that the force sensor 420 can be omitted if force control is not used. This work is a specific work that is repeatedly performed while the robot 100 is installed in a specific environment such as a factory. As will be described later, the anomaly detection system of the present embodiment is configured to correctly calculate an anomaly score indicating the degree of anomaly of the robot 100 when the robot 100 performs this specific work. As the specific work, various other works can be executed.

FIG. 2 is a block diagram showing functions of the information processing apparatus 300. As shown in FIG. The information processing apparatus 300 has a processor 310 , a memory 320 , an interface circuit 330 , an input device 340 and a display section 350 connected to the interface circuit 330 . The interface circuit 330 is also connected to the controller 200 . Measurement results of the sensors 400 of the robot 100 are supplied to the information processing device 300 via the control device 200 .

The sensors 400 include, in addition to the vibration sensor 410 and the force sensor 420 described with reference to FIG. and encoder 440 . At least some of the sensors 400 may be utilized to detect state inputs representing the state of the robot 100 .

The processor 310 functions as an anomaly detection section 600 that detects an anomaly of the robot 100 . The function of anomaly detection unit 600 is implemented by processor 310 executing a computer program stored in memory 320 . However, part or all of the functions of the abnormality detection unit 600 may be realized by a hardware circuit.

The memory 320 stores an abnormality score history AH, a threshold history TH, and a robot control program RP. The anomaly score history AH is a history of anomaly scores calculated by the anomaly detection unit 600 . The threshold history TH is a history of thresholds used when the anomaly detection unit 600 determines whether there is an anomaly based on the anomaly score. The robot control program RP is composed of a plurality of instructions for operating the robot 100 .

FIG. 3 is a block diagram showing the internal configuration of the abnormality detection section 600. As shown in FIG. The abnormality detection unit 600 includes a state input acquisition unit 610 that acquires the state input Si, an abnormality score calculation unit 620 that calculates the abnormality score Sc, and determines whether there is an abnormality by comparing the abnormality score Sc with a threshold value Th. a threshold determination unit 640 that determines a threshold Th; a history registration unit 650 that registers a history such as an abnormality score history SH and a threshold history TH in the memory 320; and a learning execution unit 660 .

The abnormality score calculator 620 has a function of receiving a state input Si and calculating an abnormality score Sc indicating the degree of abnormality. Abnormalities of the robot 100 include excessive torque, vibration, abnormal noise, overheating, and the like, which are caused by aged deterioration or breakage of parts such as reduction gears, motors, timing belts, etc. of each joint. The anomaly score calculator 620 is configured to calculate an anomaly score Sc for one or more of these anomalies. The anomaly score calculator 620 includes a general-purpose anomaly detection model 621 , a specialized anomaly detection model 622 , and a corrector 623 . The general-purpose anomaly detection model 621 calculates a general-purpose score Sg indicating the degree of anomaly according to the state input Si without depending on the usage environment of the robot 100 . The specialized anomaly detection model 622 is specialized for a specific usage environment of the robot 100 as illustrated in FIG. 1, and calculates a correction value ΔS for correcting the general score Sg according to the state input Si. The correction unit 623 calculates the abnormality score Sc by correcting the general score Sg with the correction value ΔS.

As the state input Si, for example, one or more of the following items can be used.
(1) Information on robot models:
Model name, arm length, design value or measured value of motor output.
(2) Load information:
Measured hand weight, workpiece weight, and torque applied to each joint.
(3) Information on robot movement:
The position, velocity, or acceleration of each joint, the position, velocity, or acceleration of the robot's hand, the measured value of vibration, and the measured value of temperature.
However, the state input Si preferably contains at least part of the information regarding the robot's motion. Some of these state inputs Si may be input by the user.

The general-purpose anomaly detection model 621 is configured to output, for example, any of the following as the general-purpose score Sg.
(G1) Time-series data of measured values such as torque, vibration, noise, or temperature measured by the sensors 400 and time-series data reconfigured by the general-purpose anomaly detection model 621 in response to the input Reconstruction error, which is the error.
(G2) Probability of abnormal state.
When the reconstruction error in (G1) above is calculated as the general-purpose score Sg, the general-purpose anomaly detection model 621 is configured as follows, for example.

FIG. 4 is a block diagram showing an example of the internal configuration of the general-purpose anomaly detection model 621 that calculates the reconstruction error as the general-purpose score Sg. In this example, time-series data including vibration Bm, position Pm of control point TCP, velocity Vm and acceleration Am, and joint torque Tm is input to abnormality score calculator 620 as state input Si. Vibration Bm is a value measured by vibration sensor 410 . The position Pm, velocity Vm, and acceleration Am of the control point TCP may be values calculated from the motion command to the robot 100 or may be measured values measured by the sensors 400 . The joint torque Tm is a value calculated from the measured motor current at each joint.

In the example of FIG. 4 , general-purpose anomaly detection model 621 includes reconstruction model 625 and score calculator 626 . The reconstruction model 625 receives input of the time-series data of the vibration Bm measured by the vibration sensor 410, and outputs reconstructed time-series data of the vibration Br. This reconstructed model 625 is composed of a machine learning model such as AutoEncoder. The reconstruction model 625 can be realized using, for example, a neural network such as RNN (Recurrent Neural Network) or CNN (Convolutional Neural Network) used in time series prediction. Alternatively, other types of neural networks such as seq2seq or Transformer Neural Networks may be used.

The score calculator 626 calculates the error between the input time-series data of the vibration Bm and the reconstructed time-series data of the vibration Br as the general score Sg. As the error, a general error calculation method such as mean squared error (MSE) or mean absolute error (MAE) can be used. FIG. 4 shows an example of calculating the mean squared error as the general score Sg.

The specialized anomaly detection model 622 is configured as a regression model for estimating the correction value ΔS from the position Pm, velocity Vm and acceleration Am of the control point TCP, and time-series data of the joint torque Tm. A method of learning the correction value ΔS will be described later. The state input to be input to the specialized anomaly detection model 622 is not limited to this example, and can be arbitrarily selected from the various state inputs described above, for example. The correction unit 623 calculates the abnormality score Sc by subtracting the correction value ΔS from the general score Sg.

When the general-purpose anomaly detection model 621 outputs the probability of being in an abnormal state as the general-purpose score Sg, the general-purpose anomaly detection model 621 is configured as a classification model that classifies whether it is abnormal or normal, and learning is performed. be. In this case, the probability of the anomaly class output from the general-purpose anomaly detection model 621 can be used as it is as the general-purpose score Sg, so the score calculator 626 in FIG. 4 is unnecessary.

It is preferable that at least part of the input data to the general-purpose anomaly detection model 621 and the input data to the specialized anomaly detection model 622 are different. Further, it is preferable that the computational algorithm in the general-purpose anomaly detection model 621 and the computational algorithm in the specialized anomaly detection model 622 are also different. As for the computation algorithm, in the example of FIG. 4, the algorithm of the general-purpose anomaly detection model 621 is reconstruction of input data, and the algorithm of the specialized anomaly detection model 622 is regression. In this way, it is preferable that the general-purpose anomaly detection model 621 and the specialized anomaly detection model 622 differ in at least one of the input data and the arithmetic algorithm. The reason for this is that if both the input and calculation algorithms are the same, the outputs of the general-purpose anomaly detection model 621 and the specialized anomaly detection model 622 may always be almost the same. .

FIG. 5 is a flow chart showing the procedure of learning processing of the general-purpose anomaly detection model 621. As shown in FIG. Since this learning is normally performed in a normal state before shipment of the robot 100, the usage environment differs from the actual usage environment shown in FIG.

In step S110, the learning execution unit 660 randomly determines a motion and causes the robot 100 to move. In step S<b>120 , the learning execution unit 660 records various states during operation of the robot 100 in the memory 320 . In step S130, it is determined whether or not a preset number of operations have been completed. If not completed, the process returns to step S110 and steps S110 and S120 are repeated. When the preset number of operations has been completed, the learning execution unit 660 proceeds to step S<b>140 and executes learning of the general-purpose anomaly detection model 621 . When the learning ends, the general-purpose score Sg output from the general-purpose anomaly detection model 621 shown in FIGS. 3 and 4 indicates almost zero. In other words, learning of the general-purpose anomaly detection model 621 is performed so that the general-purpose score Sg becomes zero. The reason for this is that the robot 100 does not have an abnormality when the general-purpose abnormality detection model 621 is learned. In step S<b>150 , the trained general-purpose anomaly detection model 621 is stored in the memory 320 . At this point, the specialized anomaly detection model 622 is in an unlearned state. In the processing procedure of FIG. 5, various motions are executed so that the general-purpose score Sg becomes zero. It is configured to calculate the score Sg.

FIG. 6 is a flow chart showing the procedure of abnormality detection processing in a specific environment. This processing is processing performed when the robot 100 performs a specific work in a specific actual use environment as illustrated in FIG.

In step S210, the abnormality detection unit 600 causes the robot 100 to operate in a specified work in the actual use environment. In step S<b>220 , the abnormality detection unit 600 records various states during operation of the robot 100 in the memory 320 . In step S230, the general-purpose anomaly detection model 621 calculates a general-purpose score Sg. In step S240, it is determined whether learning of the specialized anomaly detection model 622 has been completed. If the specialized anomaly detection model 622 is not trained, the general score Sg is used as is as the anomaly score Sc, and the process proceeds to step S260, which will be described later. In this case, the correction unit 623 shown in FIGS. 3 and 4 outputs the general score Sg as it is as the abnormality score Sc without correcting the general score Sg. If the learning of the specialized anomaly detection model 622 has been completed, the process proceeds to step S250, the specialized anomaly detection model 622 obtains the correction value ΔS, and the correction unit 623 corrects the general score Sg with the correction value ΔS. By doing so, the abnormality score Sc is calculated.

In step S260, the threshold determining unit 640 determines a threshold Th for distinguishing whether the abnormality score Sc indicates abnormal or normal. As a method for determining the threshold value Th, any one of the following methods can be adopted.
(1) Use a threshold value Th set in advance before shipment.
(2) Use a threshold Th determined by the user from the history of anomaly scores Sc.
(3) Obtain the distribution of the abnormality scores Sc from the history of the abnormality scores Sc, and determine the threshold value Th as a value that deviates from the distribution.
(4) An optimum threshold value Th is determined by machine learning from the history of state inputs Si input to the abnormality score calculator 620, the history of abnormality scores Sc, the history of threshold values Th, and the like.
However, in the following description, it is assumed that the threshold Th is constant.

In step S270, the determination unit 630 determines whether or not there is an abnormality depending on whether or not the abnormality score Sc is equal to or greater than the threshold Th. That is, when the abnormality score Sc is equal to or greater than the threshold Th, it is determined that there is an abnormality, and when the abnormality score Sc is less than the threshold Th, it is determined that there is no abnormality. If there is an abnormality, the process proceeds to step S280, and the abnormality detection unit 600 warns the user that an abnormality has occurred. This warning is given by displaying an alarm using the display unit 350 or by an audio warning using a speaker. Further, in step S280, the abnormality detection unit 600 may force the robot 100 to stop. If there is no abnormality, the process proceeds from step S270 to step S290.

In step S290, the learning execution unit 660 determines whether or not the specialized anomaly detection model 622 can be learned. This determination is made according to, for example, whether or not a specific work has been performed a sufficient number of times in a specific usage environment. When the specialized anomaly detection model 622 can be learned, the learning executing unit 660 executes the learning of the specialized anomaly detection model 622 in step S300.

FIG. 7 is an explanatory diagram showing an example of change in the anomaly score Sc according to learning of the specialized anomaly detection model 622. As shown in FIG. During the period from time t0 to t1, the robot 100 is in a general-purpose environment before shipment, there is no abnormality in the robot 100, and learning of the general-purpose abnormality detection model 621 is completed. During this period, the general-purpose score Sg output from the general-purpose anomaly detection model 621 is almost zero, and this general-purpose score Sg becomes the anomaly score Sc as it is. Actually, the value of the general-purpose score Sg shows various values close to zero according to the state input Si, but the drawing is simplified in FIG.

After time t1, the robot 100 is installed in a specific actual usage environment as illustrated in FIG. 1 and performs a specific task. During the period from time t1 to t2, no abnormality has occurred in the robot 100, but the general-purpose score Sg output from the general-purpose abnormality detection model 621 indicates a value greater than zero. The reason for this is that since the general-purpose anomaly detection model 621 has not learned about the specific work, the general-purpose score Sg may take a non-zero value when the specific work is executed even if the robot 100 does not have any anomalies. It is from. During the period from time t1 to t2, learning of the specialized anomaly detection model 622 is not executed, and the general-purpose score Sg output from the general-purpose anomaly detection model 621 is used as is as the anomaly score Sc.

At time t2, learning of the specialized anomaly detection model 622 is executed. This learning is performed so that the correction value ΔS, which is the output of the specialized anomaly detection model 622, matches the general score Sg during the period from time t1 to t2. As a result, after time t2, the abnormality score Sc calculated by the abnormality score calculation unit 620 becomes almost zero unless an abnormality has occurred in the robot 100 . Thereafter, as time passes and an abnormality or malfunction occurs in the robot 100, the abnormality score Sc gradually increases. Then, when the abnormality score Sc becomes equal to or greater than the threshold Th, the occurrence of abnormality is detected. Note that the dashed line indicates the change in the general score Sg after time t2.

In this way, when performing specific work in an actual use environment, it is possible to obtain a more accurate value as the abnormality score Sc by executing the learning of the specialized abnormality detection model 622 . Therefore, anomalies can be detected with higher accuracy than when only the general-purpose anomaly detection model 621 is used. In addition, even immediately after shipment of the robot 100 or after a specific action is changed, it is not necessary to learn one anomaly detection model from scratch as in the conventional art, and only the specialized anomaly detection model 622 needs to be learned. Able to quickly adapt to specific movements. In addition, since the general-purpose anomaly detection model 621 does not need to be re-learned after the shipment of the robot 100, anomaly detection can be started immediately after shipment.

As described above, in the above embodiment, the general-purpose score Sg calculated by the general-purpose anomaly detection model 621 is calculated with the correction value ΔS calculated by the specialized anomaly detection model 622 to calculate the anomaly score Sc. Therefore, anomalies can be detected with higher accuracy than when only the general-purpose anomaly detection model 621 is used. Further, even if the usage environment or specific work of the robot 100 is changed, it is possible to accurately detect anomalies simply by learning the specialized anomaly detection model 622 .

・Other forms:
The present disclosure is not limited to the embodiments described above, and can be implemented in various forms without departing from the scope of the present disclosure. For example, the present disclosure can also be implemented in the following aspects. The technical features in the above embodiments corresponding to the technical features in each form described below are used to solve some or all of the problems of the present disclosure, or to achieve some or all of the effects of the present disclosure. In order to achieve the above, it is possible to appropriately replace or combine them. Also, if the technical features are not described as essential in this specification, they can be deleted as appropriate.

(1) According to a first aspect of the present disclosure, there is provided an anomaly detection system that detects an anomaly in a robot. This anomaly detection system includes a state input acquisition unit that acquires a state input representing the state of the robot, an anomaly score calculation unit that receives the state input and calculates an anomaly score indicating the degree of the anomaly, and the anomaly score. with a threshold to determine whether or not the abnormality has occurred. The anomaly score calculation unit includes a general-purpose anomaly detection model that calculates a general-purpose score indicating the extent of the anomaly independently of the environment in which the robot is used, and a general-purpose score that is specific to the environment in which the robot is used. and a correction unit that calculates the abnormality score by correcting the general-purpose score with the correction value.
According to this anomaly detection system, the anomaly score is calculated by calculating the general-purpose score calculated by the general-purpose anomaly detection model with the correction value calculated by the specialized anomaly detection model. Abnormality can be detected with higher accuracy than when only using In addition, even if the usage environment of the robot is changed, it is possible to detect anomalies simply by learning a specialized anomaly detection model.

(2) The anomaly detection system further includes a learning execution unit that executes learning of the specialized anomaly detection model in accordance with a change in the usage environment of the robot, wherein the learning execution unit performs The learning of the general-purpose anomaly detection model may not be performed regardless of whether or not the usage environment is changed.
According to this anomaly detection system, only the specialized anomaly detection model needs to be learned according to changes in the environment in which the robot is used, so the anomaly detection model can be learned efficiently. In addition, since the general-purpose anomaly detection model does not need to be re-learned according to changes in the usage environment, users can immediately start anomaly detection immediately after shipment of the robot or after a change in the usage environment.

(3) In the above anomaly detection system, the specialized anomaly detection model is a model that calculates the correction value by specializing an action included in a specific work performed by the robot in the usage environment. The type anomaly detection model may be a model that calculates the general-purpose score for a plurality of general-purpose actions without specializing the actions included in the specific work.
According to this anomaly detection system, an anomaly score can be obtained correctly for a specific work performed in the environment in which the robot is used.

(4) According to a second aspect of the present disclosure, a robot system is provided. A robot system is provided for this robot system. This robot system includes a robot, sensors that detect a specific performance index value and a specific constraint evaluation value in the work controlled by the robot, and a parameter setting unit that executes processing for setting control parameters of the robot. , provided. The parameter setting unit performs (a) a process of receiving settings of an objective function and constraint conditions used in the optimization process of the control parameters, and (b) causing the robot to perform a task using the candidate values of the control parameters. , a process of measuring the performance index value for obtaining the objective function and the constraint evaluation value related to the constraint condition during the work; and (c) the value of the objective function obtained from the performance index value. (d) a plurality of candidates by repeating the process (b) and the process (c), by performing the optimization process using (e) a processing result including a correlation diagram showing the objective function value and the constraint evaluation value for each of the plurality of candidate values; and a process of displaying the .

(5) According to a third aspect of the present disclosure, there is provided a computer program that causes a processor to execute processing for detecting an abnormality in a robot. This computer program includes (a) a process of acquiring a state input representing the state of the robot, (b) a process of receiving the state input and calculating an abnormality score indicating the degree of the abnormality, and (c) the and determining whether or not the abnormality has occurred by comparing the abnormality score with a threshold value. The process (c) includes (c1) a process of calculating a general score indicating the degree of abnormality without depending on the usage environment of the robot; and (c2) calculating the abnormality score by correcting the general score with the correction value.

The present disclosure can also be implemented in various forms other than those described above. For example, realization in the form of a robot system comprising a robot and a robot control device, a computer program for realizing the functions of the robot control device, a non-transitory storage medium on which the computer program is recorded, etc. can do.

DESCRIPTION OF SYMBOLS 100... Robot, 110... Base, 120... Robot arm, 150... End effector, 200... Control device, 300... Information processing device, 310... Processor, 320... Memory, 330... Interface circuit, 340... Input device, 350... Display unit 400 Sensors 410 Vibration sensor 420 Force sensor 430 Current sensor 440 Joint encoder 500 Base 600 Abnormality detection unit 610 State input acquisition unit 620 Abnormality score calculation Part 621 General-purpose anomaly detection model 622 Specialized anomaly detection model 623 Correction unit 625 Reconfiguration model 626 Score calculation unit 630 Judgment unit 640 Threshold determination unit 650 History Registration unit 660...Learning execution unit

Claims (5)

An anomaly detection system for detecting an anomaly of a robot,
a state input acquisition unit that acquires a state input representing the state of the robot;
an anomaly score calculation unit that receives the state input and calculates an anomaly score indicating the degree of the anomaly;
a determination unit that determines whether or not the abnormality has occurred by comparing the abnormality score with a threshold;
with
The abnormality score calculation unit,
a general-purpose anomaly detection model that calculates a general-purpose score indicating the extent of the anomaly without depending on the usage environment of the robot;
a specialized anomaly detection model that calculates a correction value for correcting the general-purpose score, specialized for the usage environment of the robot;
a correction unit that calculates the abnormality score by correcting the general-purpose score with the correction value;
Anomaly detection system. The anomaly detection system according to claim 1, further comprising:
a learning execution unit that executes learning of the specialized anomaly detection model according to a change in the usage environment of the robot;
The anomaly detection system, wherein the learning execution unit does not learn the general-purpose anomaly detection model regardless of whether or not the usage environment of the robot is changed. The anomaly detection system according to claim 1 or 2,
The specialized anomaly detection model is a model that calculates the correction value by specializing an action included in a specific work performed by the robot in the usage environment,
The anomaly detection system, wherein the general-purpose anomaly detection model is a model that calculates the general-purpose score for a plurality of general-purpose actions without specializing the actions included in the specific work. A method for detecting an abnormality in a robot, comprising:
(a) obtaining a state input representing a state of the robot;
(b) receiving the state input and calculating an anomaly score indicating the extent of the anomaly;
(c) determining whether the abnormality has occurred by comparing the abnormality score with a threshold;
including
The step (c) is
(c1) calculating a general-purpose score indicating the degree of abnormality without depending on the operating environment of the robot;
(c2) a step of calculating a correction value for correcting the general-purpose score, specifically for the usage environment of the robot;
(c2) calculating the abnormality score by correcting the universal score with the correction value;
method including. A computer program that causes a processor to execute processing for detecting an abnormality in a robot,
(a) acquiring a state input representing the state of the robot;
(b) a process of receiving the state input and calculating an anomaly score indicating the degree of the anomaly;
(c) determining whether or not the abnormality has occurred by comparing the abnormality score with a threshold;
a computer program that causes the processor to execute
The processing (c) is
(c1) a process of calculating a general-purpose score indicating the extent of the abnormality without depending on the usage environment of the robot;
(c2) a process of calculating a correction value for correcting the general-purpose score, specifically for the usage environment of the robot;
(c2) a process of calculating the abnormality score by correcting the general score with the correction value;
computer programs, including

Images