Industrial robot anomaly detection method based on recursive feature elimination and autoencoder

Abstract

The application discloses an industrial robot anomaly detection method based on recursive feature elimination and a self-encoder, and comprises the following steps: obtaining a first preset number of normal sample signals; constructing a one-dimensional convolution self-encoder model, which is used for inputting a training set and a verification set of the normal sample signals into the one-dimensional convolution self-encoder model to perform training and verification, and obtaining a reconstructed signal of a robot motion track; constructing a feature set 1 according to extracted time domain features; introducing a recursive feature elimination method to perform effective feature screening; obtaining a percentage of effective feature weights according to weights corresponding to the effective features; performing a difference between weight features of a to-be-detected acceleration signal and weight features of the reconstructed signal to obtain feature errors of the to-be-detected acceleration signal; and judging a state of the industrial robot according to the feature errors of the to-be-detected acceleration signal. The application does not require an operator to master a large amount of professional knowledge of the industrial robot, and does not need to establish a complex dynamic model, and can directly perform detection based on original acceleration signals collected by a sensor, so that the process of the industrial robot anomaly detection is simplified.

Description

Technical Field

[0001] This invention relates to an industrial robot anomaly detection method based on recursive feature elimination and autoencoder, belonging to the field of industrial robot anomaly detection based on end-effector vibration signals. Background Technology

[0002] Abnormalities during robot operation can lead to a decline in product quality. Therefore, detecting abnormal robot conditions is a necessary measure to ensure robot lifespan and product quality. Timely detection and resolution of abnormal conditions are crucial for ensuring the continuous operation of industrial robots, helping managers determine maintenance timing based on changes in their operating status, and developing appropriate maintenance strategies, thereby improving robot reliability and safety.

[0003] In the field of anomaly detection in industrial robots, current research methods mainly fall into three categories: deep learning-based methods, data distribution-based methods, and distance-based methods. Deep learning methods offer high anomaly detection accuracy but suffer from the need for large amounts of data and sensitivity to parameters; data distribution-based methods are fast but have low detection accuracy; and distance-based methods are computationally simple but overly sensitive to the choice of distance metric. Among these methods, the autoencoder reconstruction method within deep learning exhibits significant advantages in automatic feature extraction, nonlinear relationship processing, and large-scale data adaptation. Chen T et al. addressed the lack of sufficient industrial robot anomaly data for data-driven technologies by proposing an unsupervised anomaly detection method for industrial robots—a sliding window convolutional variational autoencoder. This method achieves real-time anomaly detection in both space and time by processing multivariate time-series data. Zhong ZD et al. addressed the data scarcity issue in industrial robot anomaly detection, which limits the accuracy of data-driven anomaly detection, by proposing a sliding window one-dimensional convolutional autoencoder to determine whether the industrial robot's operating state is normal.

[0004] In robot anomaly detection based on reconstruction methods, the goal is to distinguish between normal and abnormal signals. Therefore, accurately extracting the features of the original and reconstructed signals is crucial for achieving accurate anomaly detection. Summary of the Invention

[0005] This invention provides an anomaly detection method for industrial robots based on recursive feature elimination and autoencoder, which can detect whether anomalies occur during robot movement when anomalies are present.

[0006] The technical solution of this invention is:

[0007] According to a first aspect of the present invention, an anomaly detection method for industrial robots based on recursive feature elimination and autoencoder is provided, the method comprising the following steps:

[0008] Step 1: Based on the selected motion mode, collect the end-effector acceleration signal of the industrial robot under normal motion through the attitude sensor assembled at the end of the industrial robot, and divide the acceleration signal according to the motion period to obtain a first preset number of normal sample signals, and divide the training set and validation set.

[0009] Step 2: Construct a one-dimensional convolutional autoencoder model, which is used to input the training set and validation set of normal sample signals into the one-dimensional convolutional autoencoder model for training and validation, and obtain the reconstructed signal of the robot motion trajectory.

[0010] Step 3: Extract M time-domain features from the second preset number of normal sample signals; construct feature set 1 based on the extracted M time-domain features; introduce recursive feature elimination method to screen effective features and obtain a third preset number of N effective features;

[0011] Step 4: Based on the weights corresponding to the obtained effective features, obtain the percentage of effective feature weights;

[0012] Step 5: Calculate the weighted features of the acceleration signal to be detected based on the effective features and percentage of effective feature weights of the acceleration signal to be detected; calculate the weighted features of the reconstructed signal based on the effective features and percentage of effective feature weights of the reconstructed signal obtained in Step 2; subtract the weighted features of the acceleration signal to be detected from the weighted features of the reconstructed signal to obtain the feature error of the acceleration signal to be detected.

[0013] Step 6: Based on the characteristic error of the acceleration signal to be detected, the state of the industrial robot is judged to realize the anomaly detection of the industrial robot.

[0014] Furthermore, the one-dimensional convolutional autoencoder model is divided into two parts: an encoder and a decoder. The encoder includes a convolutional layer, a ReLU activation function layer, a dropout layer, and a pooling layer; the decoder includes a transposed convolutional layer, a ReLU activation function layer, a dropout layer, and a pooling layer.

[0015] Furthermore, the recursive feature elimination method is specifically as follows: feature set 1 is used as the initial input of the support vector machine model, and the support vector machine model is used for MN iterations. In each iteration, the time-domain feature with the smallest weight is removed by using the weight information provided by the support vector machine model. For the time-domain features retained after the s-th iteration, feature set s+1 is constructed as the input of the support vector machine model for the s+1-th iteration, until the required number of iterations is reached. The weight information provided by the support vector machine model in the MN-th iteration is sorted, and the time-domain features represented by the first N weight information are retained as effective features.

[0016] Furthermore, the percentage formula for the effective feature weights is as follows: Where, |ω j | is the absolute value of the weight of the j-th effective feature. It is the sum of the absolute values ​​of all valid feature weights.

[0017] Furthermore, the weighted features of the acceleration signal to be detected, the weighted features of the reconstructed signal, and the feature error of the acceleration signal to be detected are expressed as follows:

[0018]

[0019] Among them, W j f represents the weight of the j-th effective feature. j For the j-th effective feature of the acceleration signal to be detected, F j H represents the j-th effective feature of the reconstructed signal. F H represents the characteristic error of the acceleration signal to be detected. o H is the weighted feature of the acceleration signal to be detected. r The weighted features are used to reconstruct the signal.

[0020] Furthermore, the determination of the industrial robot's state based on the characteristic error of the detected acceleration signal is specifically expressed as follows:

[0021]

[0022] Where UL is the upper boundary, H F This represents the characteristic error of the acceleration signal to be detected.

[0023] According to a second aspect of the present invention, an industrial robot anomaly detection device based on recursive feature elimination and autoencoder is provided, comprising a module for performing the industrial robot anomaly detection method based on recursive feature elimination and autoencoder as described above.

[0024] According to a third aspect of the present invention, a terminal device is provided, including a memory, a processor, and a program stored in the memory and executable by the processor, wherein the processor, when executing the program, implements the industrial robot anomaly detection method based on recursive feature elimination and autoencoder as described in any one of the preceding claims.

[0025] The beneficial effects of this invention are as follows: This invention does not require operators to possess extensive professional knowledge of industrial robots, nor does it require the establishment of complex dynamic models. It can directly detect anomalies based on the raw acceleration signals collected by sensors, simplifying the process of anomaly detection in industrial robots. Specifically, this invention can reconstruct sample signals through a designed one-dimensional convolutional autoencoder model and perform effective feature filtering using a recursive feature elimination method, reducing computational complexity and noise interference in detection. Furthermore, based on the characteristic error and threshold of the acceleration signal to be detected, the state of the signal to be detected can be effectively distinguished, thereby improving the detection effect of anomalies in industrial robots. Compared with other methods, this invention not only provides a new theoretical perspective but also has significant advantages in practical applications, offering a more effective and economical solution for health monitoring and anomaly detection of industrial robot joints, and has important research significance and application value. Attached Figure Description

[0026] Figure 1 This is a flowchart of the present invention;

[0027] Figure 2 It is an industrial robot sample signal acquisition experimental platform;

[0028] Figure 3 This is a diagram of the structure of a one-dimensional convolutional autoencoder model;

[0029] Figure 4 This is a classification result diagram of the detection of the acceleration signal to be detected from the industrial robot. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this application can be arbitrarily combined with each other.

[0031] Example 1: As Figures 1-4 As shown, according to a first aspect of the present invention, an anomaly detection method for industrial robots based on recursive feature elimination and autoencoder is provided, the method steps of which are as follows:

[0032] Step 1: Based on the selected motion mode, collect the end-effector acceleration signal of the industrial robot under normal motion through the attitude sensor assembled at the end of the industrial robot, and divide the acceleration signal according to the motion period to obtain a first preset number of normal sample signals, and divide the training set and validation set.

[0033] Step 2: Construct a one-dimensional convolutional autoencoder model, which is used to input the training set and validation set of normal sample signals into the one-dimensional convolutional autoencoder model for training and validation, and obtain the reconstructed signal of the robot motion trajectory.

[0034] Step 3: Extract M time-domain features from the second preset number of normal sample signals; construct feature set 1 based on the extracted M time-domain features; introduce recursive feature elimination method to screen effective features and obtain a third preset number of N effective features;

[0035] Step 4: Based on the weights corresponding to the obtained effective features, obtain the percentage of effective feature weights;

[0036] Step 5: Calculate the weighted features of the acceleration signal to be detected based on the effective features and percentage of effective feature weights of the acceleration signal to be detected; calculate the weighted features of the reconstructed signal based on the effective features and percentage of effective feature weights of the reconstructed signal obtained in Step 2; subtract the weighted features of the acceleration signal to be detected from the weighted features of the reconstructed signal to obtain the feature error of the acceleration signal to be detected.

[0037] Step 6: Based on the characteristic error of the acceleration signal to be detected, the state of the industrial robot is judged to realize the anomaly detection of the industrial robot.

[0038] Furthermore, the one-dimensional convolutional autoencoder model is divided into two parts: an encoder and a decoder. The encoder includes a convolutional layer, a ReLU activation function layer, a dropout layer, and a pooling layer; the decoder includes a transposed convolutional layer, a ReLU activation function layer, a dropout layer, and a pooling layer.

[0039] Furthermore, the recursive feature elimination method is specifically as follows: feature set 1 is used as the initial input of the support vector machine model, and the support vector machine model is used for MN iterations. In each iteration, the time-domain feature with the smallest weight is removed by using the weight information provided by the support vector machine model. For the time-domain features retained after the s-th iteration, feature set s+1 is constructed as the input of the support vector machine model for the s+1-th iteration, until the required number of iterations is reached. The weight information provided by the support vector machine model in the MN-th iteration is sorted, and the time-domain features represented by the first N weight information are retained as effective features.

[0040] Furthermore, the percentage formula for the effective feature weights is as follows: Where, |ω j | is the absolute value of the weight of the j-th effective feature. It is the sum of the absolute values ​​of all valid feature weights.

[0041] Furthermore, the weighted features of the acceleration signal to be detected, the weighted features of the reconstructed signal, and the feature error of the acceleration signal to be detected are expressed as follows:

[0042]

[0043] Among them, W j f represents the weight of the j-th effective feature. j For the j-th effective feature of the acceleration signal to be detected, F j H represents the j-th effective feature of the reconstructed signal. F H represents the characteristic error of the acceleration signal to be detected. o H is the weighted feature of the acceleration signal to be detected. r The weighted features are used to reconstruct the signal.

[0044] Furthermore, the determination of the industrial robot's state based on the characteristic error of the detected acceleration signal is specifically expressed as follows:

[0045]

[0046] Where UL is the upper boundary, H F This represents the characteristic error of the acceleration signal to be detected.

[0047] According to a second aspect of the present invention, an industrial robot anomaly detection device based on recursive feature elimination and autoencoder is provided, comprising a module for performing the industrial robot anomaly detection method based on recursive feature elimination and autoencoder as described above. Specifically, it includes: a first module for performing step 1: acquiring end-effector acceleration signals of the industrial robot under normal motion using an attitude sensor mounted on the end effector of the industrial robot according to a selected motion mode, and segmenting the acceleration signals according to the motion period to obtain a first preset number of normal sample signals, and dividing them into a training set and a validation set; a second module for performing step 2: constructing a one-dimensional convolutional autoencoder model, for inputting the training set and validation set of the normal sample signals into the one-dimensional convolutional autoencoder model for training and validation, and obtaining a reconstructed signal of the robot motion trajectory; a third module for performing step 3: extracting M temporal features from the second preset number of normal sample signals; constructing feature set 1 based on the extracted M temporal features; and introducing a recursive feature elimination method for effective... The system employs a feature selection process to obtain a third set of N effective features. A fourth module executes step 4: based on the weights corresponding to the obtained effective features, it calculates the percentage of effective feature weights. A fifth module executes step 5: based on the effective features of the acceleration signal to be detected and the percentage of effective feature weights, it calculates the weighted features of the reconstructed signal; based on the effective features of the reconstructed signal obtained in step 2 and the percentage of effective feature weights, it calculates the weighted features of the reconstructed signal; the difference between the weighted features of the acceleration signal to be detected and the weighted features of the reconstructed signal is calculated to obtain the feature error of the acceleration signal to be detected. A sixth module executes step 6: based on the feature error of the acceleration signal to be detected, it judges the state of the industrial robot to achieve anomaly detection of the industrial robot.

[0048] According to a third aspect of the present invention, a terminal device is provided, including a memory, a processor, and a program stored in the memory and executable by the processor, wherein the processor executes the program to implement the industrial robot anomaly detection method based on recursive feature elimination and autoencoder as described in any one of the above embodiments.

[0049] Example 2: Reference Figures 1-4 As shown, an optional embodiment of the present invention will be described in detail below, and the specific steps of the method are as follows:

[0050] Step 1: According to the test path published in the Industrial Robot Dynamic Stability Test Method, select the robot's motion as a rectangular trajectory motion in the same plane; based on the selected rectangular trajectory motion mode, collect the end-effector acceleration signal of the industrial robot under normal motion through the attitude sensor (MTI-630) assembled at the end of the industrial robot, and divide the acceleration signal according to the motion period to obtain a first preset number of normal sample signals, and divide it into training set and validation set, with 90% as training set and 10% as validation set.

[0051] Specifically, the experiment was conducted on a QJR6-1 six-axis industrial robot. An MTI-630 attitude sensor (i.e., the accelerometer shown in the diagram) was installed at the end flange of the industrial robot. Figure 2 As shown. Following the test path selection in the industrial robot dynamic stability test method, the robot's motion is a rectangular trajectory within the same plane. The rectangle's side length is 500mm, and the motion speed is 250mm / s. Acceleration signals generated by the industrial robot during its motion are collected, with the attitude sensor sampling frequency set to 400Hz. To avoid the influence of temperature on signal acquisition, signal acquisition is performed after 30 minutes of normal robot movement. The signal in the y-direction of each complete cycle is saved as a sample signal, totaling 5120 points. A total of 300 sample signals are collected for model training and verification. The specific industrial robot dynamic stability test method mentioned above is: Chongqing Robotics Society Group Standard, Industrial Robot Dynamic Stability Test Method. [T / CRSS 0007-2023]. ICS25.040.30.CCS J28.

[0052] Industrial robot anomalies include short-term and long-term anomalies. Short-term anomalies manifest as deviations from normal signal patterns for a short period, such as transient impacts. An anomaly simulation dataset is constructed using collected normal sample signals. From 300 sample signals, 100 sample signals are randomly selected as the base samples for constructing the anomaly dataset. Anomaly signals with random impact amplitudes (range 2-3 m / s) are inserted at points 4000-4100 of each base sample. 2 This process creates a short-term anomaly sample dataset. Long-term anomalies manifest as continuous abnormal vibrations over extended periods. An eccentric motor is mounted on the robot's arm, and its activation is controlled via the robot's I / O. The eccentric motor starts synchronously as the robotic arm moves, and 100 sets of long-term anomaly signals are collected as a long-term anomaly sample dataset.

[0053] Step 2: Construct a one-dimensional convolutional autoencoder model, which is used to input the training set and validation set of normal sample signals into the one-dimensional convolutional autoencoder model for training and validation, and obtain the reconstructed signal of the robot's rectangular motion trajectory;

[0054] The one-dimensional convolutional autoencoder model is divided into two parts: an encoder and a decoder, and its structure is as follows: Figure 3 As shown, the encoder consists of a convolutional layer, a ReLU activation function layer, a dropout layer, and a pooling layer. The convolutional and pooling layers encode the data, achieving downsampling of the sample data. The decoder consists of a transposed convolutional layer, a ReLU activation function layer, a dropout layer, and a pooling layer, achieving upsampling of the sample data. After constructing the one-dimensional convolutional neural network autoencoder model, the model parameters are determined. The kernel size of the one-dimensional convolutional layer is 7, the stride is 2, and the number of channels is 32; the kernel size of the transposed convolutional layer is 7, the stride is 2, and the number of channels is 16.

[0055] Specifically, when training a one-dimensional convolutional autoencoder model, the training set of normal sample signals is used as the input to the encoder, and the output of the encoder is H. n =σ0(W0X) n +b0), Among them, X n H is the original sample signal input. n The hidden features extracted by the encoder, The output vector reconstructed by the encoder is given by W1 and W2, where W1 and W2 are weights, b0 and b1 are biases, and σ0 and σ1 are ReLU activation functions. The output of the pooling layer is j = max(y1, y2, y3, ..., y n ), where n is the width of the pooling region, y n This is the nth data point in the pooling region.

[0056] The encoder's output is used as the decoder's input. The transposed convolutional layer in the decoder is the reverse computation of the convolutional layer. The output of the i-th channel of the transposed convolutional layer is: D i =ReLU(∑X*ω) i +b i ), where ω i It is a transposed convolution kernel, * indicates transposed convolution calculation, ReLU is the activation layer, b i As a bias; the final output of the decoder is: Where, j k This indicates the location of the maximum value recorded by the decoder pooling layer.

[0057] When the validation set of normal sample signals is input into the trained one-dimensional convolutional autoencoder model for validation, the mean squared error loss (MSE) function is used. The MSE function measures the error by calculating the difference between the predicted value and the true value. The specific calculation formula is as follows: Where C represents the number of samples, X i This represents the true value of a normal sample signal. This represents the predicted value obtained through a one-dimensional convolutional autoencoder model. When the value of the loss function approaches 0, it indicates a better fit between the model and the signals in the validation set, meaning the difference between the predicted and actual values ​​is smaller. At this point, training stops, and the final robot reconstruction signal is obtained.

[0058] Step 3: Extract M time-domain features from the second preset number of normal sample signals; construct feature set 1 based on the extracted time-domain features; introduce recursive feature elimination method to screen effective features and obtain a third preset number of N effective features;

[0059] The recursive feature elimination method is specifically as follows:

[0060] Feature set 1 is used as the initial input to the support vector machine (SVM) model. The SVM model undergoes MN iterations. In each iteration, the time-domain feature with the smallest weight is removed using the weight information provided by the SVM model. For the time-domain features retained after the s-th iteration, feature set s+1 is constructed and used as the input to the SVM model for the s+1-th iteration. This process continues until the required number of iterations is reached. The weight information provided by the SVM model in the MN-th iteration is sorted, and the time-domain features represented by the top N weight information are retained as valid features. For example, feature set 1 is constructed from 10 time-domain features. The third preset quantity is 5. After 5 iterations using the SVM model, feature set 1 is input into the SVM model for the first iteration, resulting in feature set 2 constructed from 9 time-domain features. Feature set 2 is input into the SVM model for the second iteration, resulting in feature set 3 constructed from 8 time-domain features. This process continues until, after 5 iterations, feature set 6 is finally obtained from 5 time-domain features.

[0061] The M time-domain features include mean, variance, standard deviation, maximum, minimum, root mean square (RMS), range, median, kurtosis, and skewness. After effective feature selection using a recursive feature elimination method, mean, variance, standard deviation, range, and RMS are selected as effective features.

[0062] In this embodiment of the invention, both the first preset quantity and the second preset quantity are 300; if there are enough normal samples collected initially, the first preset quantity and the second preset quantity can be increased as needed.

[0063] Step 4: Based on the weights corresponding to the obtained effective features, obtain the percentage of effective feature weights;

[0064] The percentage of effective feature weights provides a better indication of the proportion of each feature in the classification process. The formula for the percentage of effective feature weights is as follows: Where, |ω j | is the absolute value of the weight of the j-th effective feature. It is the sum of the absolute values ​​of all valid feature weights.

[0065] For example, Table 1 below shows the weight comparison of each effective feature in 5 iterations. The symbol 0 indicates that the feature is the one that needs to be removed in the current iteration. Taking column Iter1 as an example, after the first iteration, the time domain feature Min with the smallest weight is removed, and feature set 2 is constructed as the input for the second iteration. The same applies to the others. After 5 iterations, the percentage of effective feature weights is obtained according to the weights corresponding to the effective features, as shown in Table 2.

[0066] Table 1 Feature weights during the iterative process

[0067] feature Iter1 Iter2 Iter3 Iter4 Iter5 Mean 0.0246 0.23909 0.2561 0.0228 0.0246 Variance 0.0315 0.3562 0.0378 0.0441 0.0427 Std 0.037 0.041 0.0436 0.4917 0.0483 Kurtosis <![CDATA[5×10 -5 ]]> 0 0 0 0 Skewnss <![CDATA[3×10 -5 ]]> <![CDATA[7×10 -5 ]]> 0.0001 0 0 Max <![CDATA[8×10 -5 ]]> 0.0001 0 0 0 Min 0 0 0 0 0 Range <![CDATA[5×10 -5 ]]> 0.0002 0.0004 <![CDATA[8×10 -5 ]]> <![CDATA[8×10 -5 ]]> Median 0.0001 0.0003 0.0002 <![CDATA[1×10 -5 ]]> 0 RMS 0.0245 0.0239 0.0256 0.2262 0.0244

[0068] Table 2 Percentage of Effective Feature Weights

[0069] Feature Name weight percentage Mean 17.38％ Variance 30.77％ Std 34.69％ Range 0.02％ RMS 17.14％

[0070] Step 5: Calculate the weighted features of the acceleration signal to be detected based on the effective features and percentage of effective feature weights of the acceleration signal to be detected; calculate the weighted features of the reconstructed signal based on the effective features and percentage of effective feature weights of the reconstructed signal obtained in Step 2; subtract the weighted features of the acceleration signal to be detected from the weighted features of the reconstructed signal to obtain the feature error of the acceleration signal to be detected.

[0071] The weighted features of the acceleration signal to be detected, the difference between the weighted features of the reconstructed signal, and the feature error of the acceleration signal to be detected are expressed as follows:

[0072]

[0073] Among them, W j f represents the weight of the j-th effective feature. j For the j-th effective feature of the acceleration signal to be detected, F j H represents the j-th effective feature of the reconstructed signal. F H represents the characteristic error of the acceleration signal to be detected. o H is the weighted feature of the acceleration signal to be detected. r The weighted features are used to reconstruct the signal.

[0074] Step 6: Perform anomaly detection on the acquired acceleration signal of the industrial robot. The judgment expression is as follows:

[0075]

[0076] Where UL is the upper boundary, H F This represents the characteristic error of the acceleration signal to be detected.

[0077] That is, after judging the acceleration signal to be detected, if the characteristic error of the acceleration signal to be detected is less than or equal to the upper boundary, the robot's current state is considered to be normal; otherwise, the robot's current state is considered to be abnormal.

[0078] The upper boundary is expressed as follows:

[0079] UL = Q3 + 1.5IQR

[0080] Where Q3 is the upper quartile of the weighted features of the fourth preset number of normal acceleration signals to be detected, and IQR is the quartile range of the weighted features of the fourth preset number of normal acceleration signals to be detected.

[0081] One hundred normal sample signals and 50 short-term abnormal samples were selected as the acceleration signals to be detected. Based on the upper quartile of the weighted features of the 100 normal acceleration signals and the quartile range of the weighted features of the acceleration signal to be detected, UL = 0.2 was obtained. The acceleration signals to be detected were then tested using the method of this invention, and the detection results are as follows: Figure 4 As shown in (1), through Figure 4 (1) The method of the present invention demonstrated for detecting anomalies in normal sample signals and short-term abnormal signals showed an overall accuracy of 87.3%.

[0082] One hundred normal sample signals and 50 long-term abnormal samples were selected as the acceleration signals to be detected. Based on the upper quartile of the weighted features of the 100 normal acceleration signals and the quartile range of the weighted features of the acceleration signal to be detected, UL = 0.2 was obtained. The acceleration signals to be detected were then tested using the method of this invention, and the detection results are as follows: Figure 4 As shown in (2), through Figure 4 (2) The method of the present invention demonstrated for detecting anomalies in normal sample signals and long-term abnormal signals shows an overall accuracy of 90%.

[0083] The specific embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. An industrial robot anomaly detection method based on recursive feature elimination and autoencoder, characterized in that, The method steps are as follows: Step 1: Based on the selected motion mode, collect the end-effector acceleration signal of the industrial robot under normal motion through the attitude sensor assembled at the end of the industrial robot, and divide the acceleration signal according to the motion period to obtain a first preset number of normal sample signals, and divide the training set and validation set. Step 2: Construct a one-dimensional convolutional autoencoder model, which is used to input the training set and validation set of normal sample signals into the one-dimensional convolutional autoencoder model for training and validation, and obtain the reconstructed signal of the robot motion trajectory. Step 3: Extract M time-domain features from the second preset number of normal sample signals; construct feature set 1 based on the extracted M time-domain features; introduce recursive feature elimination method to screen effective features and obtain a third preset number of N effective features; Step 4: Based on the weights corresponding to the obtained effective features, obtain the percentage of effective feature weights; Step 5: Calculate the weighted features of the acceleration signal to be detected based on the effective features and percentage of effective feature weights of the acceleration signal to be detected; calculate the weighted features of the reconstructed signal based on the effective features and percentage of effective feature weights of the reconstructed signal obtained in Step 2; subtract the weighted features of the acceleration signal to be detected from the weighted features of the reconstructed signal to obtain the feature error of the acceleration signal to be detected. Step 6: Based on the characteristic error of the acceleration signal to be detected, the state of the industrial robot is judged to realize the anomaly detection of the industrial robot; The recursive feature elimination method is specifically as follows: Feature set 1 is used as the initial input to the support vector machine (SVM) model. The SVM model is iterated MN times. In each iteration, the time-domain feature with the smallest weight is removed using the weight information provided by the SVM model. For the time-domain features retained after the s-th iteration, feature set s+1 is constructed as the input to the SVM model for the s+1-th iteration. This process continues until the required number of iterations is reached. The weight information provided by the SVM model in the MN-th iteration is sorted, and the time-domain features represented by the top N weight information are retained as valid features.

2. The industrial robot anomaly detection method based on recursive feature elimination and autoencoder according to claim 1, characterized in that, The one-dimensional convolutional autoencoder model is divided into two parts: an encoder and a decoder. The encoder contains a convolutional layer, a ReLU activation function layer, a dropout layer, and a pooling layer; the decoder contains a transposed convolutional layer, a ReLU activation function layer, a dropout layer, and a pooling layer.

3. The industrial robot anomaly detection method based on recursive feature elimination and autoencoder according to claim 1, characterized in that, The percentage formula for the effective feature weights is: ,in, It is the absolute value of the weight of the j-th effective feature. It is the sum of the absolute values ​​of all valid feature weights.

4. The industrial robot anomaly detection method based on recursive feature elimination and autoencoder according to claim 1, characterized in that, The weighted features of the acceleration signal to be detected, the weighted features of the reconstructed signal, and the feature error of the acceleration signal to be detected are expressed as follows: in, This represents the weight of the j-th effective feature. Let j be the j-th effective feature of the acceleration signal to be detected. This represents the j-th effective feature of the reconstructed signal; The characteristic error of the acceleration signal to be detected, The weighted features of the acceleration signal to be detected The weighted features are used to reconstruct the signal.

5. The industrial robot anomaly detection method based on recursive feature elimination and autoencoder according to claim 1, characterized in that, The state of the industrial robot is determined based on the characteristic error of the acceleration signal to be detected, and the specific expression is as follows: in, The upper boundary, This represents the characteristic error of the acceleration signal to be detected.

6. An industrial robot anomaly detection device based on recursive feature elimination and autoencoder, characterized in that, Includes a module for performing the industrial robot anomaly detection method based on recursive feature elimination and autoencoder as described in any one of claims 1-5.

7. A terminal device, characterized in that: The system includes a memory, a processor, and a program stored in the memory and executable by the processor. When the processor executes the program, it implements the industrial robot anomaly detection method based on recursive feature elimination and autoencoder as described in any one of claims 1-5.

Images