Industrial robot joint failure detection method and system

Abstract

A kind of industrial robot joint fault detection method and system, the method comprises: the modeling and simulation of industrial robot are carried out;Set its joint fault mode, carry out fault simulation and construct fault mode dataset;Joint fault detection is carried out by fault detection network;Fault detection network includes the fault detector established for each joint, and the fault detector includes feature extractor and anomaly detector;Feature extractor includes the multilayer perception based on fully connected layer, using multidimensional scale scaling loss function, the original data corresponding to joint state is reduced dimension, and deep feature is extracted;Anomaly detector uses auto-encoder to carry out feature data reconstruction to the feature of original data after dimension reduction, and whether the error between the feature extracted by feature extractor and reconstructed feature data is used to judge joint state data normal or not.The present application can effectively reduce the noise interference in signal, improve fault detection precision, and guarantee the generalization of robot fault detection.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of mechanical arms and fault detection, in particular to an industrial robot arm joint fault detection method and system. BACKGROUND

[0002] Industrial robots, as an important device to improve productivity, quality and safety in the era of Industry 4.0, have been widely used in manufacturing processes, hazardous waste, management and space exploration. With the wide application of industrial robots, it is increasingly important to monitor the health status of robots in a timely manner. Robot system failure not only causes economic losses in industrial production, but also poses a threat to safety. For example, in the automotive industry, a robot failure will disrupt the operation of the entire production line, causing significant unplanned downtime and economic losses. Therefore, timely monitoring of the health status of industrial robots and reporting possible fault locations are key technologies to ensure the reliable and safe operation of industrial robots.

[0003] According to different strategies, the data-driven fault detection and isolation methods of industrial robots are mainly divided into three categories: (1) Nominal model-based fault diagnosis method: a mathematical model is used to describe the system behavior between control input and state output as a reference. When a fault occurs, the measured output will deviate from the expected value, and the residual signal will exceed a certain detection threshold, thereby detecting the anomaly. (2) Statistical distribution difference analysis of monitoring signals: in the specific working scenario of the repetitive work of industrial robots, sample the monitoring signals, and design appropriate health indicators that are sensitive to changes in monitoring signals. If a fault occurs, the monitoring signal will change.(3) Combination of feature extraction and pattern recognition: the fault of industrial robots is attributed to mechanical problems. Monitor the vibration signal, first perform signal processing, then perform feature extraction, and finally use machine learning algorithms to automatically learn the relationship between the extracted features and the health patterns.

[0004] In the existing literature "FAULT DETECTION AND ISOLATION IN ROBOTIC MANIPULATOR VIA HYBRID NEURAL NETWORKS" (Fault Detection and Isolation in Robotic Manipulator via Hybrid Neural Networks), a recurrent neural network (RNN) is proposed, and the network training algorithm is derived. The model output is compared with the sensor output to obtain the residual error; the fault diagnosis rule is derived from the residual error to detect the thruster fault. However, in the task of mechanical arm fault detection, the generation of residual data is relied on, and in more general scenarios, facing complex environmental conditions, especially in the case of small mechanical arm joint bearing torque and large noise interference, there may be a problem of insufficient generalization.

[0005] It is to be understood that the information disclosed in the Background section is merely for the purpose of enhancing the understanding of the present application, and thus can include information that is not prior art to those skilled in the art. SUMMARY

[0006] The purpose of the present application is to provide an industrial robot joint fault detection method and system to reduce noise interference in the original detection signal, improve fault detection accuracy, and ensure the generalization of robot fault detection.

[0007] To achieve the above purpose, the present application adopts the following technical solutions:

[0008] An industrial robot joint fault detection method, comprising the following steps:

[0009] S1, modeling and simulation of the industrial robot are performed;

[0010] S2, the joint fault mode of the industrial robot is set, fault simulation is performed, and a fault mode data set is constructed;

[0011] S3, joint fault detection of the industrial robot is performed through a trained fault detection network; wherein the fault detection network comprises a fault detector established for each joint, the fault detector comprises a feature extractor and an anomaly detector; the feature extractor comprises a multilayer perceptron based on a full connection layer, uses a multidimensional scale shrinkage loss function, and reduces the dimension of the original data corresponding to the joint state to extract deep features; the anomaly detector uses an autoencoder to reconstruct feature data from the features of the original data after dimension reduction, and uses the error between the features extracted by the feature extractor and the reconstructed feature data to determine whether the joint state data is normal.

[0012] In some embodiments, the last layer of the feature extractor uses a multidimensional scale shrinkage loss function to make the feature space after data dimension reduction retain the spatial distribution information of the original data.

[0013] In some embodiments, the multidimensional scale shrinkage loss function is:

[0014]

[0015] wherein L m is the loss function, |||2 represents the 2-norm, η s is a scale factor, X i and X j are the i-th and j-th samples in the data set, F i and F j are the features of the samples X i and X jIn the low-dimensional feature space, S represents the number of samples in the data set; the defined loss is the Euclidean distance between the original data, which is proportional to the Euclidean distance in the low-dimensional feature space,

[0016] Wherein, the Euclidean distance between the original data and the Euclidean distance between the low-dimensional feature space is subtracted, and the absolute value is obtained, and the average of all absolute values is obtained, and the obtained loss is used to maintain the distance structure between any two samples in the original space when learning the low-dimensional space features of the data samples.

[0017] In some embodiments, the error between the features extracted by the feature extractor and the feature reconstruction data is used to determine whether the data is normal or not, including:

[0018] The mean and standard deviation of the reconstruction error are used to construct a fault detection boundary, according to the following formula:

[0019] T up =μ+βδ

[0020] T down =μ-βδ

[0021] Wherein μ is the mean of the reconstruction error, δ is the standard deviation, β is the parameter of the standard deviation, T up is the upper limit of the reconstruction error, T down is the lower limit of the reconstruction error, and β is a hyperparameter.

[0022] When the reconstruction error is within the fault monitoring boundary, it is determined that the joint state data is normal, otherwise it is determined to be abnormal.

[0023] In some embodiments, when training the fault detection network, the feature extractor is first trained to extract low-dimensional features, and then the training of the anomaly detector is performed after the training of the feature extractor is completed, and the loss function of the anomaly detector is the reconstruction error.

[0024] In some embodiments, the input of the feature extractor is the torque and angle signal sample corresponding to the joint, and the hyperparameters in each feature extractor are set to be consistent.

[0025] In some embodiments, the feature extractor has four layers, the first three layers use the Relu activation function, and the last layer uses the Sigmoid activation function.

[0026] In some embodiments, the feature extractor is provided with an Adam optimizer.

[0027] The application also provides an industrial robot arm joint fault detection system, comprising a trained fault detection network, the fault detection network comprising a fault detector established for each joint, the fault detector comprising a feature extractor and an anomaly detector; the feature extractor comprising a multilayer perceptron based on a full connection layer, using a multidimensional scaling loss function to reduce the dimensionality of original data corresponding to a joint state and extract deep features; the anomaly detector using an autoencoder to reconstruct feature data from the features of the original data after dimensionality reduction, and using the error between the features extracted by the feature extractor and the reconstructed feature data to determine whether the joint state data is normal.

[0028] The application also provides a computer readable storage medium storing a computer program, the computer program being executed by a processor to implement the method according to any one of the above embodiments.

[0029] The application has the following advantages:

[0030] The application constructs a fault detection network comprising a fault detector for each joint, each fault detector comprising a feature extractor and an anomaly detector, wherein the feature extractor uses a multilayer perceptron based on a full connection layer to take advantage of a deep neural network, and uses a multidimensional scaling function to reduce the dimensionality of original data corresponding to a joint state, so that deep features can be extracted from original data with a lot of noise interference; the anomaly detector uses an autoencoder to reconstruct features after dimensionality reduction of original data, and uses the error between the extracted features and the reconstructed feature data to determine whether the joint state data is abnormal, effectively reducing the influence of noise on data, achieving high detection accuracy, realizing reliable monitoring of the health state of an industrial robot arm, and having good generalization when applied to robot arm fault detection.

[0031] Specifically, the advantages of the embodiments of the application include:

[0032] 1. High detection accuracy is achieved. The application can be used for health state monitoring of an industrial robot arm, helps to discover abnormal states of an industrial robot arm in time, and reduces losses caused by robot arm failure.

[0033] 2. Deep features of normal data are learned. Due to noise interference, the accuracy is limited when using existing anomaly monitoring methods for industrial robot arm fault monitoring, and the use of a feature extractor to extract deep features of normal data can effectively reduce noise interference in signals.

[0034] 3. A new fault detection framework is proposed. A train of thought is provided for fault detection of multi-channel signals, and a fault detector can be established for each channel signal in combination with data characteristics.

[0035] 4. Consider the mechanical arm failure form comprehensively. The mechanical arm failure form is considered from a more general and comprehensive perspective, ensuring the generalization of the proposed method when applied to mechanical arm failure detection.

[0036] The comparative experiment results show that the accuracy, precision and recall of the application are all better than those of the existing failure detection methods.

[0037] In the preferred embodiment, the designed multi-dimensional scale scaling loss function is used in the feature extractor, the effective features are learned through the neural network, the distribution distance between the learned low-dimensional space features is ensured to be relatively close to the distribution distance between the original data, and thus the purpose of realizing the dimension reduction of the original data while preserving the spatial distribution information of the original data can be ensured. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 is a flowchart of the industrial robot arm joint failure detection method of the embodiment of the application.

[0039] Figure 2 is a schematic diagram of the modeling and simulation of the industrial robot arm of the embodiment of the application.

[0040] Figure 3 is a schematic diagram of the feedforward-feedback control of the UR5 robot arm of the embodiment of the application.

[0041] Figure 4 is a schematic diagram of the failure detection network of the industrial robot arm of the embodiment of the application.

[0042] Figure 5 is a schematic diagram of the establishment of the failure detector for the six joints of the industrial robot arm of the embodiment of the application. DETAILED DESCRIPTION

[0043] The application will be further described below with reference to the drawings and in combination with the preferred embodiments. It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0044] Key term definition:

[0045] 1. Autoencoder:

[0046] An autoencoder is a kind of neural network, which can use an encoder to encode and compress input data into meaningful low-dimensional feature representation, and then decode the data compressed by the encoder through a decoder to reconstruct the original input data and make the output as similar as possible to the original input, in an unsupervised manner to reduce the dimension of data, learn the information representation of data, and belongs to a kind of unsupervised machine learning model.

[0047] Specifically, assuming that the sample data set is X, the original data feature dimension number is m, the sample number is n, the input data is mapped to f(X) by the encoder, the decoder decodes f(X), reconstructs the input data and outputs X', and the encoder encoding formula is as follows:

[0048] Y = f(X) = s f (WX + b X )

[0049] Wherein, s f is an activation function, W and b are the network weight matrix and bias vector of the encoder respectively. g(Y) is the decoder mapping output, Sg is the decoder activation function, W' and b Y are the network weight matrix and bias vector of the decoder respectively. The decoder decoding formula is as follows:

[0050] x' = g(Y) = s g (W'Y + b Y )

[0051] 2. Multidimensional scaling loss function:

[0052] Multidimensional scaling is often used for manifold learning and data dimensionality reduction. When performing data dimensionality reduction, multidimensional scaling can realize the function of keeping the spatial distribution distance of data samples in the low-dimensional feature space close to the mutual distance of the original data itself in the high-dimensional space.

[0053] The embodiment of the application provides a kind of industrial robot arm joint fault detection method, comprising the following steps:

[0054] S1, modeling and simulation of industrial robot arm are carried out;

[0055] S2, set the joint fault mode of industrial robot arm, carry out fault simulation and construct fault mode data set;

[0056] S3, the joint fault detection of industrial robot arm is carried out by the trained fault detection network;Wherein, the fault detection network includes the fault detector established for each joint, and the fault detector includes a feature extractor and an anomaly detector;The feature extractor includes a multilayer perception based on full connection layer, uses multidimensional scaling loss function, and carries out dimensionality reduction to the original data corresponding to joint state, extracts deep feature;The anomaly detector uses autoencoder to reconstruct feature data for the feature after dimensionality reduction of original data, and judges whether joint state data is normal or not using the error between the feature extracted by the feature extractor and the reconstructed feature data.

[0057] In some embodiments, the last layer of the feature extractor uses the multidimensional scaling loss function, so that the feature space after data dimensionality reduction retains the spatial distribution information of the original data.

[0058] In a preferred embodiment, the present application is based on the idea of multi-dimensional scaling, and a multi-dimensional scaling loss function is designed as follows:

[0059]

[0060] where L m is the loss function, |||2 represents the 2-norm, η s is a scale factor, X i and X j are the i-th and j-th samples in the data set, F i and F j are the representations of the samples X i and X j in the low-dimensional feature space, and S represents the number of samples in the data set. The defined loss is proportional to the Euclidean distance between the original data and the Euclidean distance in the low-dimensional feature space. The loss is obtained by taking the absolute value of the difference between the Euclidean distance between the original data and the Euclidean distance in the low-dimensional feature space, and then taking the average of all the absolute values. The purpose of this is to maintain the distance structure between any two samples in the original space when learning the low-dimensional feature of the data samples.

[0061] The feature extractor uses the designed multi-dimensional scaling loss function to learn effective features through a neural network, ensuring that the distribution distance between the learned low-dimensional features is close to the distribution distance between the original data. This can ensure that the original data is reduced in dimension while preserving the distribution information of the original data space.

[0062] In a preferred embodiment, the error between the features extracted by the feature extractor and the feature reconstruction data is used to determine whether the data is normal or not, including:

[0063] The mean and standard deviation of the reconstruction error are used to construct a fault detection boundary, according to the following formula:

[0064] T up = μ + βδ

[0065] T down = μ - βδ

[0066] where μ is the mean of the reconstruction error, δ is the standard deviation, β is a parameter of the standard deviation, T up is the upper limit of the reconstruction error, T down is the lower limit of the reconstruction error, and β is a hyperparameter; if the reconstruction error is within the fault monitoring boundary, the joint state data is determined to be normal, otherwise it is determined to be abnormal.

[0067] In a preferred embodiment, when training the fault detection network, the feature extractor is first trained to extract low-dimensional features, and after the training of the feature extractor is completed, the training of the anomaly detector is performed, and the loss function of the anomaly detector is the reconstruction error.

[0068] The input of the feature extractor is the torque and angle signal sample corresponding to the joint, and the hyperparameters in each feature extractor are set consistently.

[0069] The feature extractor has four layers, the first three layers use the Relu activation function, and the last layer uses the Sigmoid activation function.

[0070] The Adam optimizer is set in the feature extractor.

[0071] The embodiment of the application also proposes an industrial robot arm joint fault detection system, which comprises a trained fault detection network, the fault detection network comprises a fault detector established for each joint, the fault detector comprises a feature extractor and an anomaly detector; the feature extractor comprises a multilayer perceptron based on a fully connected layer, uses a multidimensional scaling loss function, and performs dimension reduction on original data corresponding to the joint state to extract deep features; the anomaly detector uses an autoencoder to reconstruct feature data of the features after dimension reduction of the original data, and uses the error between the features extracted by the feature extractor and the reconstructed feature data to determine whether the joint state data is normal.

[0072] The embodiment of the application also proposes a computer readable storage medium storing a computer program, and the computer program is executed by a processor to implement the industrial robot arm joint fault detection method as described above.

[0073] The embodiment of the application uses industrial robot arm joint torque, angle and other state data to monitor the health status of the industrial robot arm. In the design of the fault detection network framework, a fault detector is established for each joint, and if all the fault detectors are normal, the robot arm is considered to be normal, otherwise, it is considered to be faulty. A single fault detector comprises a feature extractor and an anomaly detector. The feature extractor is composed of a fully connected layer, which takes advantage of the deep neural network and extracts deep features from simulation data severely disturbed by noise with the help of a multidimensional scaling function. The anomaly detector uses an autoencoder to construct a fault detection boundary using the mean and standard deviation of the reconstruction error, and achieves high detection accuracy.

[0074] The specific embodiments of the industrial robot arm joint fault detection method are further described below.

[0075] As Figure 1As shown, the industrial robot joint fault detection method of the embodiment includes the following three steps: industrial robot modeling and simulation, fault simulation and data set construction, and industrial robot joint fault detection experiment. Since there is no publicly available data set for different fault modes of the industrial robot at present, in order to complete the industrial robot joint fault detection under a single working condition, the data set must be constructed first, so the first step of the embodiment is to complete the industrial robot modeling and simulation; then, different fault modes of the industrial robot are set, fault simulation is performed and data is collected; finally, according to the characteristics of the data in the data set, a network framework is designed to realize the detection of the fault joint of the industrial robot.

[0076] I. Industrial robot modeling and simulation

[0077] As shown in Figure 2 , a typical industrial robot UR5 is taken as a specific object, and CoppeliaSim (a dynamic robot simulator) and MATLAB (a mathematical software) are used for joint simulation of dynamics simulation and control. The specific idea is as follows:

[0078] 1. Construct the dynamics model of the UR5 robot in the CoppeliaSim environment;

[0079] 2. Use MATLAB for path planning and realize feedforward-feedback control;

[0080] Joint simulation of the two realizes the motion of the UR5 robot under the desired trajectory. To achieve this purpose, three problems need to be solved, namely, the UR5 modeling problem, the path planning problem, and the control strategy design problem.

[0081] Modeling problem: the simulation model is constructed by joint simulation of CoppeliaSim and Matlab and is controlled. The URDF (Unified Robot Description Format) file of UR5 is imported into CoppeliaSim and Matlab to establish the model. In MATLAB, the importrobot function in the Robotics System Toolbox can be used to directly import the model; after importing UR5 in CoppeliaSim, joint control mode, environment simulation step, physical engine simulation step, control mode, etc. need to be designed. Matlab, as the control end of the model, reads the joint angle of the robot from CoppeliaSim at each calculation step, and outputs the joint control torque to the model in CoppeliaSim to achieve the purpose of controlling the motion of the robot.

[0082] Path planning problem: The path planning problem mainly involves inverse kinematics and quintic polynomial interpolation. Through the UR5 inverse kinematics model, the angles of each joint at the target position of the UR5 robot arm can be obtained according to the position and attitude of the end of the robot arm. Then through the quintic polynomial interpolation method, the complete motion path equation of each joint is obtained by substituting the initial position and target position angles, angular velocity, and angular acceleration of each joint, where θ(t) is the angle of each joint, is the angular velocity of each joint, is the angular acceleration of each joint, a0, a1, a2, a3, a4, a5 are parameters of the quintic polynomial planning equation, and t is time. The quintic polynomial interpolation formula is as follows:

[0083]

[0084] The DH parameters of the UR5 robot arm are shown in Table 1. The DH parameters are important information for modeling the robot arm, and the inverse kinematics equation of the six-degree-of-freedom robot arm can be derived using the DH parameters. Among them θ θi is the joint angle, ai is the link length, di is the link offset, α θti is the link twist angle, and i is the i-th joint.

[0085] Table 1 DH parameters of UR5 robot arm

[0086]

[0087]

[0088] The end target position and rotation matrix of the experimental trajectory are shown in Table 2. By setting the end target position and rotation matrix of the robot arm, the angles of each joint at the target state can be obtained by substituting the inverse kinematics equation of the UR5 robot arm, and then the quintic polynomial planning is brought in to obtain the relationship between the angle, angular velocity, and angular acceleration of each joint with time, thereby achieving the purpose of trajectory planning. Among them, trajectories l1, l2, and l3 are the set simulation trajectories of the robot arm.

[0089] Table 2 End target position and rotation matrix of experimental trajectory

[0090]

[0091] Control strategy problem: As Figure 3As shown, the feedforward-feedback control strategy is adopted to control the UR5 in CoppeliaSim. When calculating the feedforward input torque, the inverseDynamics function is used to obtain the expected torque of the UR5 robot arm under the expected trajectory; when designing the feedback input torque, the PD (Proportional-Derivative) control method is used, and the PD controller differential term KD and integral term coefficient KP should satisfy the Routh criterion, so that when the motion trajectory of the robot arm deviates, the error can quickly converge and reach a stable state.

[0092] K D 、K P are respectively set to K D = K P = 10I 6*6 , where I 6*6 is a 6*6 unit matrix. In the simulation of the robot arm, white noise with a mean of 0 and a standard deviation of 0.1 is added to each joint.

[0093] II. Fault simulation and data set construction

[0094] In the design of robot arm joint fault simulation, the main concern is the actuator failure. Five different faults are designed for each joint of the robot arm, which are constant deviation failure, ramp failure, time-varying complex deviation failure, constant partial failure, and time-varying partial failure.

[0095] The design of different fault degrees of each joint of the robot arm is shown in Table 3, where τ denotes the expected torque of the joint, t denotes the time, and τ R denotes the actual torque of the joint.

[0096] Table 3 Different fault degree design of each joint of the robot arm

[0097]

[0098]

[0099] In the construction of the fault simulation data set, no multi-joint fault is designed, and multiple fault modes are not designed for the same joint.

[0100] In constructing the simulation fault data set, data within one cycle is extracted as a sample. A cycle refers to the movement of the robot arm from the initial position to the target position and then back to the initial position. Sampling is performed every 50 milliseconds, and there are 400 data points in each cycle. In order to avoid the continuous accumulation of faults causing serious trajectory deviation, the simulation model is restarted every 5 consecutive cycles. Under each experimental trajectory, each joint fault is simulated, and 100 samples are obtained for each fault type of a single joint. A total of 500 fault data for each joint and 3000 samples for six joints are obtained.

[0101] III. Industrial robot joint fault detection

[0102] The purpose of the industrial robot health state monitoring under single working condition is to find the industrial robot fault in time and reduce the threat to industrial production and the safety of employees caused by the industrial robot fault. In the detection, three set industrial robot running tracks are taken as the objects, the state data of the industrial robot in normal operation is taken as the training set data, and the test set data includes normal data and running data under the fault state of the industrial robot.

[0103] 1. Network structure

[0104] Due to the existence of much noise interference in the original signal, the traditional anomaly detection method is difficult to achieve the purpose of effective industrial robot health state monitoring. In order to achieve the purpose of industrial robot anomaly detection under single working condition, it is extremely important to reduce the dimension of the original data and extract effective feature representation.

[0105] Combined with this idea, the embodiment proposes an industrial robot health state monitoring network under single working condition, which includes two parts. The first part is a feature extractor, and the second part is an anomaly detector. When performing the anomaly detection task, only the joint input torque and output angle obtained when the industrial robot is in normal state are taken as the input data.

[0106] The feature extractor is composed of multiple layers of perception mechanism. A multi-dimensional scale shrinkage loss function is used in the last layer of the feature extractor to ensure that the spatial distribution information of the original data is still retained in the feature space after dimension reduction. The anomaly detector adopts an autoencoder. The features after dimension reduction of the original data are used as the input data of the autoencoder. The feature data is reconstructed by using the autoencoder. The mean and standard deviation of the reconstruction error are used to construct the fault detection boundary, so as to achieve the purpose of industrial robot health state monitoring under single working condition. Assuming that the mean of the reconstruction error is μ, the standard deviation is δ, β is the parameter of the standard deviation, the upper limit of the reconstruction error is T up , and the lower limit of the reconstruction error is T down , which can be expressed by the following formula:

[0107] T up =μ+βδ

[0108] T down =μ-βδ

[0109] Where β is a hyperparameter. When the reconstruction error is within the fault monitoring boundary, it means that the data is normal, otherwise it means that the data is an abnormal data.

[0110] The industrial robot health state monitoring network under single working condition establishes a fault detector for each joint,Figure 4 The network structure of a single fault detector is shown, which is composed of two parts of feature extractor and anomaly detector. In training, the feature extractor (composed of fully connected layers, and the loss function is multi-dimensional scaling loss) is first trained to extract low-dimensional features, and after the training of the feature extractor is completed, the training of the anomaly detector begins. The anomaly detector adopts the design of autoencoder, and the loss function is reconstruction error.

[0111] In the detection process of one specific embodiment, as shown in Figure 4 , for an industrial robot arm UR5 with 6 joints, a total of six feature extractors are established, respectively corresponding to the 6 joints of the industrial robot arm UR5, and the input of each feature extractor is the torque and angle signal sample of the corresponding joint, and the dimension of each single input sample is (400*2). The hyperparameter settings (such as the number of neural network hidden layers, the type of optimizer, etc.) in each feature extractor are consistent. It should be understood that the number of feature extractors and corresponding anomaly detectors is not limited by the present application.

[0112] The data used for training the model is the joint torque and angle of the industrial robot under normal operating conditions in three trajectories, and the number of normal data in the training set is 200. The test set includes 2800 industrial robot normal operating state data and 3000 fault data. The feature extractor has a total of four layers, the first three layers use the Relu activation function, and the last layer uses the Sigmoid activation function, and the neural network dimension of each layer is [400-256-128-32-16]. In training, a health state monitoring model is established for each of the six joints of the UR5. Only when all joints are determined to be normal by the anomaly detector, the joint signal of the robot is considered to be normal.

[0113] The mechanical arm health state monitoring network framework is shown in Figure 5 , which establishes a fault detector for each joint. The y1-y6 indicates the sample label (normal or abnormal) predicted by each fault detector. Only when y1-y6 are normal, the sample is considered normal, that is, y indicates a normal label, and the robot is in a healthy state. As long as one label shows abnormal, the robot is considered to have failed, that is, y is an abnormal label, and the robot is in a fault state. In addition, the feature extractor is further described for this framework, which reduces the dimension of the data of the corresponding joint state, extracts the features, and then takes the extracted features as the input of the anomaly detector (autoencoder) to obtain the feature reconstruction data. Finally, the error between the features extracted by the feature extractor and the feature reconstruction data is used to determine whether the sample data is normal or not.

[0114] The optimizer used in this embodiment is Adam, the learning rate is set to 0.0005, the iteration is 5 times, and the fault detection boundary β is set to 3. The accuracy (the proportion of correct prediction), precision (the number of correctly predicted normal samples, the proportion of predicted normal samples (samples predicted as normal, which may actually be abnormal samples) in the predicted normal sample number), and recall rate (the proportion of normal samples predicted) are used as the performance detection indicators of the model. Table 4 is the industrial robot fault detection result of this embodiment, which shows the results of using the joint state data (torque, angle) at this time to detect faults when the robot runs on trajectory l1, trajectory l2 and trajectory l3, for example, the average precision of five experiments of trajectory l1 is 1.

[0115] Table 4 Industrial robot fault detection results

[0116]

[0117] In order to verify the effectiveness of the method of this embodiment, it is particularly important to use the commonly used anomaly detection model to do comparative experiments. The long short-term memory (LSTM), the long short-term memory-based autoencoder (LSTM-AE) and the autoencoder (AE) are all relatively common anomaly detection models, therefore, the LSTM, the LSTM-AE and the AE are used to do comparative experiments, and the results of the comparative experiments are compared and analyzed with the results of the method of this embodiment, so as to show the superiority of the method of this embodiment.

[0118] 200 normal data are used as the training set, and there are 3100 data in the test set, of which 100 are normal data, and 500 are abnormal data under each joint, that is, 100 data of each type of fault under each joint. The robot anomaly detection results based on LSTM, LSTM-AE and AE are shown in Tables 5 to 7.

[0119] Table 5 Robot fault detection results based on LSTM

[0120]

[0121]

[0122] Table 6 Robot fault detection results based on LSTM-AE

[0123] Experimental task Accuracy Precision Recall trajectory l1 84.9±0.36 17.6±0.37 99.8±0.4 trajectory l2 91.4±7.26 33.8±10.8 99.6±0.80 trajectory l3 86.3±0.75 19.1±0.83 99.8±0.4

[0124] Table 7 Robot fault detection results based on AE

[0125] Experimental task Accuracy Precision Recall trajectory l1 88.8±2.06 22.6±2.91 99.0±1.10 trajectory l2 86.0±0.77 18.8±0.85 100±0 trajectory l3 77.9±1.11 12.8±0.55 100±0

[0126] As can be seen from the experimental results shown in Tables 5 to 7, the accuracy, precision and recall of the existing fault detection method are lower than the accuracy of the method of the embodiment of the present application, so the industrial robot joint fault detection method of the embodiment of the present application is superior to the existing fault detection method.

[0127] The industrial robot joint fault detection method of the embodiment of the present application reduces the influence of noise on data by designing an industrial robot joint fault detection framework to establish a fault detector for each of the six joints of the robot arm. Furthermore, by designing the network structure of the fault detector, model training is performed in stages and modules, that is, the training of the feature extractor is performed first, a multidimensional scaling loss function is used to preserve the original spatial distribution of normal data samples, and then the training of the abnormality detection network structure is performed, and a fault detection boundary is constructed.

[0128] The present application has the following effects:

[0129] 1. High detection accuracy is achieved. The method can be used for industrial robot health state monitoring, which helps to discover abnormal states of the industrial robot in a timely manner and reduce losses caused by robot faults.

[0130] 2. Deep features of normal data are learned. Due to noise interference, the accuracy of using existing abnormality monitoring methods for industrial robot fault monitoring is limited, and using a feature extractor to extract deep features of normal data can effectively reduce noise interference in signals.

[0131] 3. A new fault detection framework is proposed. It provides a way for fault detection of multi-channel signals, and can establish a fault detector for each channel signal according to the characteristics of the data.

[0132] 4. Comprehensive consideration of robot fault forms. The method considers robot fault forms from a more general and comprehensive perspective, ensuring the generalization of the method when applied to robot fault detection.

[0133] The above is a further detailed description of the present application in combination with a specific preferred embodiment, and the specific implementation of the present application should not be limited to these descriptions. For those skilled in the art to which the present application belongs, without departing from the concept of the present application, a number of equivalent alternatives or obvious modifications can be made, and the same performance or use should be considered as belonging to the protection scope of the present application.

Claims

1. A method for detecting a failure of an industrial robot joint, characterized by, The method comprises the following steps: S1, modeling and simulation of an industrial robot arm are performed; S2, a joint fault mode of the industrial robot arm is set, fault simulation is performed, and a fault mode dataset is constructed; S3, joint fault detection of the industrial robot arm is performed through a trained fault detection network; the fault detection network comprises a fault detector established for each joint, and the fault detector comprises a feature extractor and an anomaly detector; the feature extractor comprises a multilayer perceptron based on a full connection layer, uses a multidimensional scaling loss function, performs dimension reduction on original data corresponding to a joint state, and extracts deep features; the anomaly detector uses an autoencoder to perform feature data reconstruction on the features after dimension reduction of the original data, and judges whether the joint state data is normal or not by using an error between the features extracted by the feature extractor and the reconstructed feature data; the multidimensional scaling loss function is as follows: where L m is the loss function, || ||2 denotes the 2-norm, η s is a scaling factor, X i and X j are the i-th and j-th samples in the dataset, F i and F j are the representations of the samples X i and X j in the low-dimensional feature space, and S denotes the number of samples in the dataset; the defined loss is proportional to the Euclidean distance between the original data, wherein a difference between a Euclidean distance between original data and a Euclidean distance between low-dimensional feature spaces of the original data is obtained, absolute values of all the differences are obtained, an average of the absolute values is obtained, and the average is taken as a loss obtained, so that a distance structure between any two samples in an original space is maintained when low-dimensional space features of learning data samples are learned; the judgment of whether the data is normal or not by using the error between the features extracted by the feature extractor and the feature reconstruction data comprises: a fault detection boundary is constructed by using a mean and a standard deviation of the reconstruction error, and the following formula is used: T up = μ + βδ T down = μ - βδ where μ is the mean of the reconstruction error, δ is the standard deviation, β is a parameter of the standard deviation, T up is an upper bound of the reconstruction error, T down is a lower bound of the reconstruction error, and β is a hyperparameter. when the reconstruction error is within the fault detection boundary, it is judged that the joint state data is normal, otherwise, it is judged to be abnormal.

2. The industrial robotic arm joint fault detection method of claim 1, wherein, The last layer of the feature extractor uses the multidimensional scaling loss function, so that the feature space after dimension reduction of the data retains the spatial distribution information of the original data.

3. The industrial robot arm joint fault detection method according to any one of claims 1 to 2, characterized in that, When the fault detection network is trained, the feature extractor is first trained to extract low-dimensional features, and then the training of the anomaly detector is performed after the training of the feature extractor is completed, and the loss function of the anomaly detector is a reconstruction error.

4. The industrial robot joint fault detection method according to any one of claims 1 to 2, characterized in that, The input of the feature extractor is torque and angle signal samples of the corresponding joint, and the hyperparameters in each feature extractor are set to be consistent.

5. The industrial robot joint fault detection method according to any one of claims 1 to 2, characterized in that, The feature extractor has four layers, the first three layers use a Relu activation function, and the last layer uses a Sigmoid activation function.

6. The industrial robot joint fault detection method according to any one of claims 1 to 2, characterized in that, An Adam optimizer is arranged in the feature extractor.

7. An industrial robot arm joint fault detection system characterized by, The method comprises the following steps: S1, modeling and simulation of an industrial robot arm are performed; S2, a joint fault mode of the industrial robot arm is set, fault simulation is performed, and a fault mode dataset is constructed; S3, joint fault detection of the industrial robot arm is performed through a trained fault detection network; the fault detection network comprises a fault detector established for each joint, and the fault detector comprises a feature extractor and an anomaly detector; the feature extractor comprises a multilayer perceptron based on a full connection layer, uses a multidimensional scaling loss function, performs dimension reduction on original data corresponding to a joint state, and extracts deep features; the anomaly detector uses an autoencoder to perform feature data reconstruction on the features after dimension reduction of the original data, and judges whether the joint state data is normal or not by using an error between the features extracted by the feature extractor and the reconstructed feature data; the multidimensional scaling loss function is as follows: where L m is a loss function, || ||2 denotes the 2-norm, η s is a scaling factor, X i and X j are the i-th and j-th samples in the dataset, F i and F j are the representations of the samples X i and X j in the low-dimensional feature space, and S denotes the number of samples in the dataset; the defined loss is proportional to the Euclidean distance between the original data and the Euclidean distance in the low-dimensional feature space, The difference between the Euclidean distance between the original data and the Euclidean distance between the low-dimensional feature space of the original data is taken, and the absolute value is obtained, and the average of all the absolute values is taken as the loss obtained, so as to maintain the distance structure between any two samples in the original space when learning the low-dimensional space features of the data samples; The error between the feature extracted by the feature extractor and the feature reconstruction data is used to determine whether the data is normal, and the error between the feature extracted by the feature extractor and the feature reconstruction data is used to determine whether the data is normal. The mean and standard deviation of the reconstruction error are used to construct a fault detection boundary, according to the following formula: T up = μ + βδ T down = μ - βδ where μ is the mean of the reconstruction error, δ is the standard deviation, β is a parameter of the standard deviation, T up is an upper bound of the reconstruction error, T down is a lower bound of the reconstruction error, β is a hyperparameter; When the reconstruction error is within the fault monitoring boundary, it is determined that the joint state data is normal, otherwise it is determined to be abnormal.

8. A computer readable storage medium storing a computer program, characterized in that, The computer program is executed by the processor to realize the method of any one of claims 1 to 6.

Images