Intelligent detection method of abnormal data of multi-dimensional signals in electromechanical servo system based on TimeGAN

Abstract

This invention discloses a TimeGAN-based intelligent detection method for abnormal data in multidimensional signals of electromechanical servo systems. The method includes collecting time series data from electromechanical servo system sensors and performing data preprocessing; inputting the data into a TimeGAN model for training to generate a synthetic data set of a certain size, which is then merged with the original samples into an extended data set; initializing an anomaly detection model based on LSTM-Deep SVDD; inputting the extended data set into an anomaly detection model based on LSTM-Deep SVDD for training to optimize the hypersphere radius; inputting test data into the trained anomaly detection model to calculate a judgment criterion; and inputting the data to be detected into the trained anomaly detection model to calculate a judgment threshold and determine whether the data is abnormal. The method can be used for detecting abnormal data in multidimensional signals when the original data of the electromechanical system is insufficient, and can improve the accuracy and precision of the detection model.

Description

Technical Field

[0001] The present invention belongs to the technical field of condition monitoring of electromechanical servo systems, and particularly relates to an intelligent detection method for abnormal multi-dimensional signal data of electromechanical servo systems based on TimeGAN. Background Art

[0002] Electromechanical servo systems, which use electric motors as their primary drive elements, are widely used in key areas such as industrial production and national defense science and technology. As a critical component of control systems, the accuracy and reliability of servo systems are closely linked to their performance and operating status. During the operation of electromechanical servo systems, sensors are inevitably affected by component wear, changes in the operating environment, and other interference signals, resulting in anomalies in their output. During the sensor data acquisition process, factors such as measurement accuracy, measurement methods, and data variation can also contribute to the generation of anomalies. Under the influence of anomalies, the stability and accuracy of electromechanical servo systems can be degraded to a certain extent, resulting in resource waste, economic losses, and labor consumption. Monitoring the abnormal data output by electromechanical servo systems can timely determine the system's operating status, enabling early warning of faults. This is crucial for ensuring the stability and safety of electromechanical systems.

[0003] Currently, mainstream anomaly detection methods for electromechanical servo systems can be categorized into statistical and classification-based methods. Statistical methods, such as principal component analysis and partial least squares, offer relatively stable detection capabilities, but detection thresholds often require pre-setting and manual adjustment when dealing with complex data. Classification-based anomaly detection methods, on the other hand, can be further categorized into traditional methods and deep learning methods. Traditional methods, such as Bayesian networks and support vector machines, offer good generalization capabilities but suffer from lower detection accuracy. As engineering requirements for systems continue to rise, electromechanical servo systems are becoming increasingly complex, making the causes of system anomalies more complex, varied, and difficult to locate. The resulting data is complex and has low correlations between different data types. Relying solely on traditional data processing methods can lead to problems such as difficulty finding valid data and missing key data features. Therefore, the introduction of deep learning into anomaly data detection has become a major development trend in recent years. In the field of anomaly detection in time series data, deep learning models can fully exploit the internal correlations between data, thereby learning effective data features and enabling anomaly detection in time series data.

[0004] However, deep learning-based anomaly data detection methods typically require a large number of raw samples to support model training. When the size and quality of the dataset are not compatible with the computing power of the deep learning model, this can lead to model overfitting or training failure. Complex electromechanical systems often face challenges such as data collection difficulties, high data collection costs, and complex test experiment design. This results in a small number of raw samples, making it difficult to provide sufficient data support for deep learning network training. Summary of the Invention

[0005] The purpose of the present invention is to provide an intelligent detection method for anomaly data of multidimensional signals in electromechanical servo systems based on TimeGAN. This method can be used for anomaly data detection of multidimensional signals in electromechanical servo systems when the original data of the electromechanical system is insufficient. When the training sample size is small, the data set size can be expanded through data augmentation, providing sufficient data for the anomaly detection model based on deep learning, thereby improving the accuracy and precision of the detection model.

[0006] The technical solutions of the present invention are as follows:

[0007] A TimeGAN-based intelligent detection method for abnormal multi-dimensional signal data in an electromechanical servo system includes the following steps:

[0008] S1: Collecting time series data of multiple sensors of the same type of electromechanical servo system and performing data preprocessing on the collected time series data;

[0009] S2: Input the above preprocessed time series data into the TimeGAN model for training. Use the trained TimeGAN model to generate a synthetic dataset of a certain size, which is merged with the original samples to form an expanded dataset.

[0010] S3: Initialize the anomaly detection model based on LSTM-Deep SVDD;

[0011] S4: Input the expanded dataset into the LSTM-Deep SVDD-based anomaly detection model for training and optimize the hypersphere radius so that the original data is shrunk as much as possible within the hypersphere with c as the center and R as the radius;

[0012] S5: Input the test data into the trained anomaly detection model and calculate the judgment criteria;

[0013] S6: Input the data to be detected into the trained anomaly detection model and calculate the judgment threshold s to determine whether the data is abnormal.

[0014] Furthermore, the preprocessing methods in S1 include missing value filling, data normalization, filtering, and dataset dimension reconstruction.

[0015] Furthermore, in S1, multiple time series data of the same type of electromechanical servo system sensors are collected, and the method for preprocessing the collected time series data is as follows:

[0016] Missing values ​​are filled using the mode of the values ​​collected by the sensor to which the missing value belongs;

[0017] Data normalization uses linear normalization, and its formula is:

[0018] X std =(XX min ) / (X max -X min )

[0019] Among them, X represents the time series data collected by the sensor, X max 、X min Indicates the maximum and minimum values ​​of time series data, X std Represents the data after normalization;

[0020] The filtering is implemented using a sliding average algorithm, and its formula is:

[0021]

[0022]

[0023] Among them, l represents the radius of the sliding window, x t For X std Time series data, t is x t The corresponding time point, T represents the total number of time series, i is the data sequence number; p t Represents x t The result after sliding average, X ave Represents the data set after filtering;

[0024] The dataset dimension reconstruction is as follows:

[0025] The collected multi-dimensional sensor time series data belongs to the two-dimensional table type data, with a total of m rows of collection time, c n Column sensor data, read into the matrix form The TimeGAN model requires three-dimensional data input and uses a sliding window to intercept data to add the third dimension: take a sliding window of length l, move it along the direction of increasing data set acquisition time according to a certain time step, and perform such movement k times in total. Then, the two-dimensional data sets obtained by each movement are superimposed in sequence to obtain the required three-dimensional data set.

[0026] Furthermore, the method for generating the synthetic dataset in S2 is:

[0027] Build a data augmentation model based on autoencoder networks and generative adversarial networks;

[0028] Train the autoencoder network to learn the static and dynamic features of the samples;

[0029] Jointly train the autoencoder network and the generative adversarial network so that the latent features in the generative adversarial network are synchronized with the latent features in the autoencoder network;

[0030] Generate synthetic data using a trained data augmentation model.

[0031] Furthermore, the method for initializing the LSTM-Deep SVDD-based anomaly detection model in S3 is:

[0032] Construct a two-layer LSTM prediction network, and use the last set of values ​​of each time series of the expanded data set as the predicted true value y. The output of the prediction network includes the prediction results and network parameters hn2; y and The mean square error is used as the loss function for predicting network training, specifically:

[0033]

[0034] Where N1 is the total number of data in the training set; the expanded data set is input into the LSTM network for pre-training, and the network parameters hn2 output by each batch are summed, and the average value is used as the center of the hypersphere c;

[0035] Initialize the hypersphere radius.

[0036] Furthermore, in S4, the expanded dataset is input into the anomaly detection model based on LSTM-Deep SVDD for training. The method for optimizing the hypersphere radius is as follows:

[0037] The calculation formula for the distance between the training data point and the center of the hypersphere is:

[0038] Dist(x i ,W)=||hn2(x i ,W)-c|| p / 4

[0039] The loss function of the network is:

[0040]

[0041] The training dataset of the model is x i is a set of time series data in the training set, N1 is the total number of samples in the training set; W={W 1 ,...,W L} are network parameters, W′ is the weight of the hidden layer l∈{1,...,L}, the second term of the loss function is the weight decay regularization of the training network parameters, the hyperparameter λ>0, L is the total number of hidden layers; in the optimization process, the alternating optimization strategy is used to optimize R and the number of training times respectively.

[0042] Furthermore, in S5, the test data is input into the trained anomaly detection model, and the method for calculating the judgment standard is as follows:

[0043] Input the test set into the trained anomaly data detection model, calculate the distance of each set of time series data in the test set relative to the center of the hypersphere and save it; calculate the absolute value of the difference between the saved distance and the radius R of the hypersphere, select the distance values ​​of the data points close to the surface of the hypersphere, and calculate the mean. The absolute value of the difference between this data and the radius R is recorded as E. The judgment criteria for abnormal data are:

[0044] Dist<R+E.

[0045] Furthermore, in S6, the data to be detected is input into the trained anomaly detection model, and the judgment threshold s is calculated. The method for judging whether the data is abnormal is as follows:

[0046] Select Dist as the judgment threshold s, which is the following calculation formula:

[0047] s(x i )=||hc2(x i ,W)-c|| p / 4

[0048] When s is less than the judgment standard, the data is judged to be normal data, otherwise it is abnormal data.

[0049] Compared with the existing technology, the advantage of the present invention is that it can continue to use the deep learning model to mine data features when the original samples are insufficient, improve the problems of the inability to train the deep learning model and unsatisfactory training results caused by insufficient data volume, and improve the accuracy and stability of the anomaly detection model. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 This is a flow chart of the intelligent detection method for abnormal multi-dimensional signal data of electromechanical servo systems based on TimeGAN;

[0051] Figure 2 Schematic diagram of the data augmentation model;

[0052] Figure 3 Schematic diagram of the abnormal data detection model. DETAILED DESCRIPTION

[0053] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0054] This paper designs a TimeGAN-based intelligent detection method for abnormal data in multidimensional signals of electromechanical servo systems. The method can be divided into three parts: data preprocessing, data enhancement model construction, and anomaly detection model construction.

[0055] The multidimensional signal used in this embodiment is a data set of similar sensor signals distributed at different locations in the electromechanical system. The obtained two-dimensional data set is represented as Among them, X is the time series data collected by different sensors, c n =1,...,M is the sensor number, and M is the total number of sensors.

[0056] A TimeGAN-based intelligent detection method for abnormal multi-dimensional signal data in an electromechanical servo system includes the following steps:

[0057] S1: Collecting time series data of multiple sensors of the same type of electromechanical servo system and performing data preprocessing on the collected time series data;

[0058] Data preprocessing includes three steps, as follows:

[0059] Step 1: Fill in missing values

[0060] Due to the influence of the test environment and the performance limitations of the sensors themselves, the collected electromechanical system sensor signals may contain some missing values. To prevent this from affecting subsequent data processing and analysis, we use mode filling to fill the missing values. This means that for each sensor's collected time series data X, we calculate its mode and use it to fill in the missing values.

[0061] Step 2: Normalization

[0062] Normalization uses the linear normalization method, and its formula is:

[0063] X std =(XX min ) / (X max -X min )

[0064] Among them, X represents the time series data collected by the sensor, X max 、Xmin Indicates the maximum and minimum values ​​of time series data, X std Represents the normalized data.

[0065] Step 3: Filtering

[0066] The filtering uses a sliding average algorithm for milestones, and its formula is:

[0067]

[0068]

[0069] Among them, l represents the radius of the sliding window, x t For X std Time series data, t is x t The corresponding time point, T represents the total number of time series, i is the data sequence number; p t Represents x t The result after sliding average, X ave Represents the data set after filtering;

[0070] Step 4: Dimension Reconstruction

[0071] In some embodiments, the collected multi-dimensional sensor time series data belongs to a two-dimensional table type data, with a total of m rows of collection time, c n Column sensor data, read into the matrix form Since the TimeGAN model requires three-dimensional data input, a sliding window is used here to intercept the data to add the third dimension. The specific content is as follows:

[0072] X ave Take a sliding window of length l, move it along the direction of increasing data set acquisition time at a time step of 1, and perform such movement k times in total. Then, superimpose the two-dimensional data sets obtained by each movement in sequence to obtain the required three-dimensional data set.

[0073] S2: Input the above preprocessed time series data into the TimeGAN model for training. Use the trained TimeGAN model to generate a synthetic dataset of a certain size, which is merged with the original samples to form an expanded dataset.

[0074] The method of constructing the data enhancement model TimeGAN is as follows:

[0075] The TimeGAN model is primarily composed of four networks: an embedding network, a reconstruction network, a sequence generator, and a sequence discriminator. During training, the autoencoder network composed of the first two networks is trained together with the generative adversarial network composed of the second two networks. The adversarial network operates in the latent space provided by the embedded network, and the latent features of real samples and synthetic data are synchronized using a supervised loss function. The specific contents are as follows:

[0076] 3D dataset after data preprocessing Participate in the training of data augmentation model as a training set. The training set can be expressed as Where N is the total number of training samples, n∈{1,...,N}, s is the static feature of the training sample, x t is the dynamic feature of the training sample.

[0077] The training process consists of three steps, as follows:

[0078] Step 1: Train the autoencoder network

[0079] The internal network model of the autoencoder network is constructed using the autoencoder. r First, we use the embedding network to reduce the dimension and get where h s is the static feature of the data after dimensionality reduction of the training set data, h t is the dynamic feature of the data after dimensionality reduction, N is the total number of training samples, n∈{1,...,N}. Then the dimensionality-reduced data is input into the reconstruction network to restore it to the dimension of the training sample, and the output is recorded as in To reconstruct the static characteristics of the data, To reconstruct the dynamic characteristics of the data. The loss function is used to calculate the and T r The difference is reduced during the iteration process and T r The loss function for calculating sequence reconstruction is:

[0080]

[0081] in, is the maximum likelihood estimate.

[0082] Step 2: Joint training of the autoencoder network and the generative adversarial network

[0083] In the generative adversarial network, the sequence generator consists of a static feature generator g S and dynamic feature generator g X The sequence discriminator includes the static feature discriminator d Sand dynamic feature discriminator d X .

[0084] The training set and random time series noise are represented as {(s,x t )} and {(z s ,z t )}, where z s and z t The static and dynamic features of the input noise are inputted respectively, and both are inputted into the trained embedding network and sequence generator at the same time, and the output of the embedding network is {(h s ,h t )}, the output of the sequence generator is Use the loss function L S The outputs of the two are synchronized so that the sequence generator can learn the dynamic characteristics of the training set. The loss function L of the generator is S as follows:

[0085]

[0086] The output of the sequence generator and the output of the embedding network are mixed and input into the sequence discriminator. The sequence discriminator outputs the discrimination result Use the loss function L U For training, the loss function L of the discriminator U as follows:

[0087]

[0088] Step 3: Generate synthetic data using the trained data augmentation model. In this embodiment, a synthetic data set equal to the training set is generated and combined with the training set to form an augmented data set.

[0089] S3: Initialize the anomaly detection model based on LSTM-Deep SVDD;

[0090] S4: Input the expanded dataset into the LSTM-Deep SVDD-based anomaly detection model for training and optimize the hypersphere radius so that the original data is shrunk as much as possible within the hypersphere with c as the center and R as the radius;

[0091] S5: Input the test data into the trained anomaly detection model and calculate the judgment criteria;

[0092] S6: Input the data to be detected into the trained anomaly detection model and calculate the judgment threshold s to determine whether the data is abnormal.

[0093] Anomaly detection model construction

[0094] In this paper, a model with LSTM as the core and Deep SVDD as the main architecture is used for anomaly detection. The core idea of ​​the algorithm is to use LSTM to map samples from the input space to the output space, so that the samples in the output space are as much as possible contracted within the hypersphere with c as the center and R as the radius. Divide into training set and test set N1 and N2 are the number of samples in the training set and test set, respectively, which serve as the input of the anomaly detection model. The specific content of the anomaly detection model construction is:

[0095] Step 1: Detection model construction and initialization

[0096] The internal network structure of the anomaly detection model includes two LSTM layers, a dropout layer, two fully connected layers, and a Relu activation function layer. The first part of the network is the LSTM layer and the dropout layer, which map samples from the input space to the output space. In this embodiment, the output dimension of the first LSTM layer is 50, the output dimension of the second LSTM layer is 100, and the output hn2 of the second LSTM layer is retained as the subsequent input. The second part of the network is the fully connected layer and the activation function layer, which restore the samples from the output space to the original dimension and form a prediction network together with the first part. The output of the fully connected layer is the prediction result. Take the training set n1=1,K,each set of time series data in N1 The last set of values ​​A l As the predicted true label y, the predicted training set is y and The mean square error is used as the loss function for predicting network training, specifically:

[0097]

[0098] Where N1 is the total number of data in the training set; the expanded data set is input into the LSTM network for pre-training, and the network parameters hn2 output by each batch are summed, and the average value is used as the center c of the hypersphere.

[0099] Initialize the hypersphere radius, which is set to 0 in the embodiment. This value has little effect on the calculation results of the network.

[0100] Step 2: Anomaly detection model training

[0101] Retain the pre-trained network weights and input the training set TR for training to optimize the radius of the hypersphere. Since the training set contains the majority of normal data, a single-class Deep SVDD framework is used for optimization. The distance between the training data point and the center of the hypersphere is calculated as:

[0102] Dist(x i ,W)=||hn2(x i ,W)-c|| p / 4

[0103] The loss function of the network is:

[0104]

[0105] The training dataset of the model is x i is a set of time series data in the training set, N1 is the total number of samples in the training set; W={W 1 ,...,W L} are network parameters, W′ is the weight of the hidden layer l∈{1,...,L}, the second term of the loss function is the weight decay regularization of the training network parameters, the hyperparameter λ>0, L is the total number of hidden layers; in the optimization process, the alternating optimization strategy is used to optimize R and the number of training times respectively.

[0106] Step 3: Detection Threshold

[0107] The test set is input into the trained anomaly data detection model. The distance of each set of time series data in the test set relative to the center of the hypersphere is calculated and saved. The absolute value of the difference between the saved distance data and the radius R of the hypersphere is calculated. In this embodiment, data with a difference less than 0.0002 is considered to be close to the surface of the hypersphere. The mean of these data is calculated, and the absolute value of the difference between the data and the radius R is recorded as E. The judgment criteria for abnormal data are as follows:

[0108] Dist<R+E

[0109] When performing anomaly detection, the calculation formula for the input data judgment threshold s is:

[0110] s(x i )=||hc2(x i ,W)-c|| p / 4

[0111] When s is less than the judgment standard, the data is judged to be normal data, otherwise it is abnormal data.

[0112] The present invention provides a method for intelligent detection of abnormal data of multi-dimensional signals of electromechanical servo system based on TimeGAN. After obtaining multiple sensor signals of electromechanical system, the multi-sensor signals are first pre-processed, and then the processed data is sent to the time series data enhancement network TimeGAN (Time-series Generative Adversarial Networks) to generate synthetic data with a distribution similar to that of the original samples, and merged with the original data to achieve data enhancement to obtain an expanded data set. After that, the expanded data set is input as a training sample into an abnormal data detection algorithm (LSTM-Deep SVDD) with a long short-term memory neural network LSTM (Long Short-term memory) as the core and a deep support vector description Deep SVDD (Deep Support Vector Data Description) as the main architecture for training to obtain a multi-dimensional abnormal data intelligent detection model. Finally, based on the detection model, it is automatically determined whether there is abnormal data in the multi-dimensional test data. Compared with the prior art, the advantage of the present invention is that it can continue to use the deep learning model to mine data features when the original sample is insufficient, improve the problems such as the inability of the deep learning model to be trained and the unsatisfactory training effect caused by insufficient data volume, and improve the accuracy and stability of the abnormality detection model.

[0113] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A TimeGAN-based intelligent detection method for abnormal multi-dimensional signal data in electromechanical servo systems, characterized by: The following steps are involved: S1: Collecting time series data of multiple sensors of the same type of electromechanical servo system and performing data preprocessing on the collected time series data; S2: Input the above preprocessed time series data into the TimeGAN model for training. Use the trained TimeGAN model to generate a synthetic dataset of a certain size, which is merged with the original samples to form an expanded dataset. S3: Initialize the anomaly detection model based on LSTM-Deep SVDD; S4: Input the expanded dataset into the LSTM-Deep SVDD-based anomaly detection model for training and optimize the hypersphere radius so that the original data is shrunk as much as possible within the hypersphere with c as the center and R as the radius; S5: Input the test data into the trained anomaly detection model and calculate the judgment criteria; S6: Input the data to be detected into the trained anomaly detection model and calculate the judgment threshold s to determine whether the data is abnormal; The method for initializing the LSTM-Deep SVDD-based anomaly detection model in S3 is: Construct a two-layer LSTM prediction network, and use the last set of values ​​of each time series of the expanded data set as the predicted true value y. The output of the prediction network includes the prediction results and network parameters hn2; y and The mean square error is used as the loss function for predicting network training, specifically: Where N1 is the total number of data in the training set; the expanded data set is input into the LSTM network for pre-training, and the network parameters hn2 output by each batch are summed, and the average value is used as the center of the hypersphere c; Initialize the hypersphere radius.

2. The TimeGAN-based intelligent detection method for abnormal multi-dimensional signal data in electromechanical servo systems according to claim 1 is characterized in that: The preprocessing methods in S1 include missing value filling, data normalization, filtering, and dataset dimension reconstruction.

3. The TimeGAN-based intelligent detection method for abnormal multi-dimensional signal data in electromechanical servo systems according to claim 2 is characterized in that: In S1, multiple time series data of the same type of electromechanical servo system sensors are collected, and the method for preprocessing the collected time series data is as follows: Missing values ​​are filled using the mode of the values ​​collected by the sensor to which the missing value belongs; Data normalization uses linear normalization, and its formula is: X std =(X-X min ) / (X max -X min ) Among them, X represents the time series data collected by the sensor, X max 、X min Indicates the maximum and minimum values ​​of time series data, X std Represents the data after normalization; The filtering is implemented using a sliding average algorithm, and its formula is: Among them, l represents the radius of the sliding window, x t For X std Time series data, t is x t The corresponding time point, T represents the total number of time series, i is the data sequence number; p t Represents x t The result after sliding average, X ave Represents the data set after filtering; The dataset dimension reconstruction is as follows: The collected multi-dimensional sensor time series data belongs to the two-dimensional table type data, with a total of m rows of collection time, c n Column sensor data, read into the matrix form The TimeGAN model requires three-dimensional data input and uses a sliding window to intercept data to add the third dimension: take a sliding window of length l, move it along the direction of increasing data set acquisition time according to a certain time step, and perform such movement k times in total. Then, the two-dimensional data sets obtained by each movement are superimposed in sequence to obtain the required three-dimensional data set.

4. The TimeGAN-based intelligent detection method for abnormal multi-dimensional signal data in electromechanical servo systems according to claim 3 is characterized in that: The method for generating the synthetic dataset in S2 is: Build a data augmentation model based on autoencoder networks and generative adversarial networks; Train the autoencoder network to learn the static and dynamic features of the samples; Jointly train the autoencoder network and the generative adversarial network so that the latent features in the generative adversarial network are synchronized with the latent features in the autoencoder network; Generate synthetic data using a trained data augmentation model.

5. The TimeGAN-based intelligent detection method for abnormal multi-dimensional signal data in electromechanical servo systems according to claim 4 is characterized in that: In S4, the expanded dataset is input into the LSTM-Deep SVDD-based anomaly detection model for training. The method for optimizing the hypersphere radius is as follows: The calculation formula for the distance between the training data point and the center of the hypersphere is: Dist(x i ,W)=||hn2(x i ,W)-c|| p / 4 The loss function of the network is: The training dataset of the model is x i is a set of time series data in the training set, N1 is the total number of samples in the training set; W={W 1 ,...,W L } is the network parameter, W l The second term of the weight loss function for the hidden layer l∈{1,...,L} is the weight decay regularization of the training network parameters, the hyperparameter λ>0, and L is the total number of hidden layers; during the optimization process, an alternating optimization strategy is used to optimize R and the number of training times respectively.

6. The TimeGAN-based intelligent detection method for abnormal multi-dimensional signal data in electromechanical servo systems according to claim 5 is characterized in that: In S5, the test data is input into the trained anomaly detection model, and the method for calculating the judgment standard is as follows: Input the test set into the trained anomaly data detection model, calculate the distance of each set of time series data in the test set relative to the center of the hypersphere and save it; calculate the absolute value of the difference between the saved distance and the radius R of the hypersphere, select the distance values ​​of the data points close to the surface of the hypersphere, and calculate the mean. The absolute value of the difference between this data and the radius R is recorded as E. The judgment criteria for abnormal data are: Dist<R+E.

7. The TimeGAN-based intelligent detection method for abnormal multi-dimensional signal data in electromechanical servo systems according to claim 6 is characterized in that: In S6, the data to be detected is input into the trained anomaly detection model, and the judgment threshold s is calculated. The method for determining whether the data is abnormal is as follows: Select Dist as the judgment threshold s, which is the following calculation formula: s(x i )=||hc2(x i ,W)-c|| p / 4 When s is less than the judgment standard, the data is judged to be normal data, otherwise it is abnormal data.

Images