Signal generation and quality evaluation method based on cross-domain migration of generative adversarial network

Abstract

The application discloses a cross-domain migration generative adversarial network signal generation and quality evaluation method. First, the collected mechanical equipment monitoring data under different working conditions are intercepted and the like, and after time-frequency domain transformation, a training set and a test set under each working condition are constructed respectively; then, an auxiliary classification generative adversarial network model is established based on a migration network adaptive regularization term and a network weight complexity regularization term constraint, the network hidden layer unit number and the hyperparameters of the network model are determined through a Bayesian optimization method, and the data under each working condition is used for model training; finally, the test set after sample expansion is input into the model, the diagnostic precision is calculated, meanwhile, the equidistance feature mapping dimension reduction and the JS divergence similarity measurement method are used to evaluate the closeness of the data probability distribution between the generated sample and the real sample features, so that the data generation quality evaluation is realized.

Description

Technical Field

[0001] This invention belongs to the field of industrial machinery monitoring data analysis and enhancement, and in particular to a method for cross-domain migration generative adversarial network signal generation and quality evaluation. Background Technology

[0002] In the field of data-driven intelligent manufacturing technology, high-quality data is a decisive factor in ensuring the accuracy of CNC machining processes and the monitoring and fault diagnosis of industrial equipment operation status. However, in the long-term, high-frequency condition monitoring of equipment such as CNC machine tools and industrial robots, there are many uncontrollable factors such as strong electromagnetic fields, environmental noise, and equipment aging, which lead to unstable signal transmission and data distortion, resulting in a decline in data quality. On the other hand, the operating conditions of industrial equipment are complex and variable. The monitoring data of core components such as industrial robot reducers, CNC machine tool tools, lead screw guides, and bearings show significant cross-domain differences, diverse feature degradation trends, and limited sample data due to their variable operating conditions, process parameters, external interference, and dynamic environment. Current mainstream data modeling methods and fault diagnosis methods under constant operating conditions cannot accurately assess and predict the health status of equipment under variable operating conditions. Therefore, how to expand data augmentation based on existing small sample data to achieve data-driven fault diagnosis modeling under variable operating conditions and across domains while ensuring data generation quality is the key technology for solving the problem of high-quality generation and modeling of cross-domain data.

[0003] Chinese patent CN114298267A discloses a fault diagnosis method and application based on a bidirectional attention generative adversarial network (GAN). This method constructs a bidirectional attention GAN model, generates additional fault samples, and inputs them into a trained fault classification model to obtain diagnostic results. However, this method can only expand and model samples for a single working condition, ignoring the changes in data distribution caused by complex and varied processing conditions. Chinese patent CN115796238A discloses a small-sample mechanical fault diagnosis method and device based on implicit data augmentation. This method is based on a backbone diagnostic network of residual modules and introduces an implicit data augmentation network. However, this method only directly judges the quality of generated data by improving the network prediction accuracy, lacking an evaluation of the quality of generated data. Chinese patent CN114997211A discloses a cross-working-condition fault diagnosis method based on an improved adversarial network and attention mechanism. It achieves cross-domain diagnosis of fault signals by training an adversarial network on the collected fault vibration signals from different working conditions. However, this method relies on sufficient data from different fields. In the field of industrial equipment condition monitoring, where fault data is scarce, the lack of data expansion and enhancement will prevent the model from being effectively trained and applied. Chinese patent CN115017946A discloses a small-sample class-imbalanced intelligent fault diagnosis method for bearings. This method combines an autoencoder network and a dual discriminator to build a generative adversarial network for data expansion. However, this method does not consider the quality evaluation of the generated data and only improves the diagnostic performance by increasing the prediction accuracy of the network, lacking good interpretability.

[0004] In summary, current generative data augmentation methods lack intuitive and effective methods for evaluating the quality of generated signal data. Existing data augmentation and generation methods primarily use fault classification accuracy as an indirect indicator to determine the reliability of generated data quality, and fail to consider the impact of data distribution under different operating conditions on signal generation quality. This limits the further development and application of generative data augmentation methods in the field of mechanical fault diagnosis. Summary of the Invention

[0005] The purpose of this invention is to provide a method for signal generation and quality evaluation based on cross-domain migration generative adversarial networks, so as to complete fault classification under different operating conditions and evaluate the quality of data generation in the process of generating fault data of small samples of mechanical equipment under different operating conditions.

[0006] The technical solution for achieving the objective of this invention is as follows:

[0007] The raw data were collected from vibration samples of normal, slightly abnormal, and seriously faulty mechanical equipment during processing under multiple different working conditions.

[0008] Data preprocessing involves dividing the signal into segments and then using singular value decomposition and Gaussian mixture clustering model. The cluster labels are filtered by differential sequence entropy to remove unloaded or empty intervals in the signal.

[0009] First, the signal is divided into segments of length L. To facilitate the Fast Fourier Transform, let L = 2. i (i = 8, 9, 10, 11, 12); the signal is converted into an m×n matrix through vector reconstruction and singular value decomposition is performed; the singular value vectors of dimension m are obtained after decomposition, and the first k vectors are selected as input to a Gaussian mixture clustering model; the clustering model outputs cluster labels t after calculation, and samples of the class with smaller signal amplitudes corresponding to the labels are removed; the clustering results of the singular value vectors at k = 1…m are calculated in sequence, and the difference sequence entropy of the cluster labels t is calculated. The k value that minimizes the sequence entropy is selected as k. min The value is shown in the following formula:

[0010]

[0011] In the formula, This represents the sample label obtained by clustering the j-th sequence using the first k columns of the singular value matrix as input to the Gaussian mixture clustering model.

[0012] Based on the auxiliary classifier generative adversarial network, the following improvements are made: (1) The number of simultaneous modeling conditions in the generative model part of the network is set to d, and the network output unit is modified to L×d, where d is the number of simultaneous modeling conditions; (2) The model input of the discriminative model part of the network is set to L×d; (3) The model loss function of the discriminative model part of the network is modified by adding a domain feature covariance difference regularization term L. CORAL and network weight complexity constraints The domain feature covariance difference regularization term is shown in the following formula:

[0013]

[0014] Among them, covariance W t To determine the model network weights; (4) regularize the covariance difference term L of the domain features. CORAL and network weight complexity constraints By incorporating hyperparameters λ1 and λ2 into the discriminative model network, the loss function is shown in the following equation:

[0015]

[0016] The hyperparameters λ1 and λ2 were determined using Bayesian optimization.

[0017] To evaluate the quality of the generated data, the generated signal and the real signal are first subjected to min-max normalization, and the normalization formula is shown in the following formula.

[0018]

[0019] The signal is dimensionally reduced using isometric feature mapping, and the first two dimensions of the reduced data are selected as features. The number of intervals between the maximum and minimum feature values ​​is set to N, and the number of values ​​in each interval is counted to obtain the histogram frequency vector v. The JS divergence between the generated sample frequency vector p1 and the original sample frequency vector p2 is calculated to obtain the distance similarity index s, as shown in the following formula:

[0020]

[0021] The similarity index s ranges from [0,1]. The smaller the value and the closer it is to 0, the greater the similarity in data distribution between the generated data and the original data, and the higher the quality of the generated data. The larger the value and the closer it is to 1, the smaller the similarity in data distribution between the generated data and the original data, and the lower the quality of the generated data.

[0022] The significant advantages of this invention compared to existing technologies are:

[0023] (1) The generative adversarial network signal generation method based on cross-domain migration proposed in this invention, compared with the existing multi-stage and step-by-step modeling process in data generation and cross-domain modeling, can simultaneously incorporate data generation and cross-domain modeling into a whole network model, thereby taking into account both the accuracy of modeling diagnosis and the quality of data generation, and having a better global optimization effect.

[0024] (2) Traditional data generation methods based on adversarial networks indirectly evaluate the quality of generated data by the accuracy of model classification. The signal quality evaluation method of generative adversarial networks based on cross-domain migration proposed in this invention directly calculates the feature distribution index between generated signals and real signals and gives data quality evaluation criteria, thereby intuitively and effectively evaluating the quality of generated signals and having a simpler and more reliable evaluation process. Attached Figure Description

[0025] Figure 1 The flowchart of this invention

[0026] Figure 2 Generate cross-domain feature distribution map for cross-domain transfer auxiliary classification using adversarial networks

[0027] Figure 3 Generative adversarial networks (GANs) generate probability distribution maps of samples for cross-domain transfer learning-aided classification, comparing them with other methods and the original signal. Detailed Implementation

[0028] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0029] This invention discloses a cross-domain transfer-based generative adversarial network (GAN) signal generation and quality evaluation method. Applicable to milling cutters, gears, bearings, etc., the method involves collecting vibration signals from faults under different working conditions of the equipment through experiments, inputting them into a cross-domain GAN model for feature learning and data generation augmentation. The method constructs a GAN model including sample filtering, a cross-domain transfer-based auxiliary classification GAN, and a data quality evaluation method to clean and extract vibration signals from mechanical equipment faults. It also completes the construction of an auxiliary classification GAN model constrained by transfer network adaptation regularization and network model complexity regularization. Simultaneously, an attention mechanism layer is added to the discriminator of the GAN to increase the weight of important features. Finally, using linear and nonlinear dimensionality reduction methods and similarity measurement methods, the method evaluates the similarity of the data probability distribution between the generated samples and the real samples, thus achieving data generation quality evaluation. The specific steps include:

[0030] Step 1: Raw data were collected from vibration samples of mechanical equipment during processing under multiple different working conditions, including normal, slightly abnormal, and serious fault samples, and then labeled.

[0031] Step 2: Taking milling tool sample data as an example, the workpiece used in the experiment was a No. 45 steel part, and all milling experiments were performed on the same workpiece. The machine tools used included three types of flat-end mills: normal, worn, and broken. Vibration signals for each tool were collected at spindle speeds of 2000, 2600, and 3200 RPM during workpiece machining. The vibration signals were collected using a triaxial accelerometer 1A314E fixed to the outside of the spindle, with a data sampling frequency of 12kHz, covering the X, Y, and Z directions of the spindle. For each signal sample, a signal length of 512 was selected, and 2000 samples were constructed for each fault type. 70% of the samples were randomly selected as the training set under the three operating conditions, and the remaining 30% were used as the test set.

[0032] Table 1 Definition of Cross-Domain Wear Identification

[0033]

[0034] Step 3: The experimentally collected milling process includes three parts: climb milling, conventional milling, and idle rotation. The signals collected during the idle rotation cannot reflect the tool status because the tool and workpiece are not in contact or being machined.

[0035] First, the data is preprocessed. Based on the singular value decomposition and Gaussian mixture clustering model, the signal is divided into segments of length 1024. This is then converted into a 32*32 signal matrix through vector reconstruction and subjected to singular value decomposition. Second, after decomposition, a 32-dimensional singular value vector is obtained, and the top k elements of the vector are selected. min Input a Gaussian mixture clustering model; calculate the clustering results of the singular value vectors from k=1 to 32 in sequence, and calculate the difference sequence entropy of the cluster labels t. Select the k value that minimizes the sequence entropy as k. min The value is shown in the following formula:

[0036]

[0037] Calculations show that the entropy of the difference sequence is minimized when k=5. The singular values ​​of the first 5 dimensions are selected as the input to the final Gaussian mixture clustering model. After calculation, the clustering model outputs cluster labels t, and samples of the class with smaller signal amplitudes corresponding to the labels are removed to obtain the signal with the empty part removed.

[0038] Step 4: Input samples into the cross-domain generative adversarial network. The construction method of the cross-domain generative adversarial network is as follows: The neural network construction model of the generator and discriminator is shown in the table below.

[0039] Table 2 Generator Network Model of Cross-Domain Generative Adversarial Networks

[0040]

[0041] Table 3 Discriminator Network Model for Cross-Domain Generative Adversarial Networks

[0042]

[0043] Step 5: Construct the discriminator loss function of the generative adversarial network using hyperparameters of the transfer regularization term and the network's complex constraint terms, where the regularization constraints are as follows;

[0044]

[0045] Step 6: The attention mechanism is added to the generator and discriminator in the auxiliary classification generative adversarial network. The transfer regularization term uses the maximum mean difference term to constrain the discriminator loss function. The generator loss function is as follows:

[0046] L G =-E z logD(G(z|y))-E z logP(c|G(z|y))

[0047] The discriminator network loss function is as follows:

[0048]

[0049] Step 7: Train the model using training sets under different operating conditions; expand the test set using samples generated from the conditional adversarial generative network for each different operating condition, and directly obtain the test results through the auxiliary classification generative adversarial network to obtain the fault diagnosis accuracy. The comparison of model accuracy before and after cross-domain testing is shown in Table 4.

[0050] Table 4. Training loss and prediction accuracy of comparative methods based on deep neural networks

[0051]

[0052] The above results indicate that the generative network for cross-domain transfer-assisted classification achieves better predictive performance than the general transfer generative adversarial network.

[0053] Step 8:

[0054] The generated and actual spectra are subjected to dimensionality reduction. The dimensionality-reduced data samples are then normalized to eliminate the dimensions of the distribution across different dimensions. Next, a data distribution histogram is calculated for all sample points, distributed within an interval of 100. The histogram count is then tallied, and the data distribution index is calculated. The data generation quality assessment method comprises two stages: data dimensionality reduction and similarity measurement.

[0055] Data dimensionality reduction methods utilize nonlinear manifold learning and isometric feature mapping to obtain the first two dimensions of features after dimensionality reduction. After histogram interval statistics, similarity metrics are used for measurement. Comparison methods include Euclidean distance, Bhattacharyya distance, KL divergence, and JS divergence. The distribution of the generated data after dimensionality reduction using isometric feature mapping is shown below compared to the actual data distribution. Figure 3 As shown in Table 5.

[0056] Table 5 shows the differences in data distribution between the generated and real signals obtained by various data generation models using the equidistant feature mapping method.

[0057]

[0058] Since the JS divergence interval is within the range of [0,1], this value can be used to determine the quality of data generation. Comparison with other similarity indicators shows that the samples generated by the cross-domain transfer-based generative adversarial network have the smallest distance and the highest similarity to the real samples, reflecting the effectiveness of the proposed data generation method.

[0059] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for cross-domain migration generative adversarial network signal generation and quality evaluation, characterized in that... The steps are as follows: Step 1: Collect monitoring data of mechanical equipment parts during operation under different working conditions. The monitoring data includes one or more of the monitoring data collected by vibration sensors, temperature sensors, acoustic emission sensors, and current measurement modules, which are used for subsequent model training. Step 2, preprocessing the above monitoring data: After dividing the signal into segments of length L, use singular value decomposition and Gaussian mixture clustering model, filter the clustering labels by differential sequence entropy, and remove unloaded or empty intervals of the signal. Step 3: Obtain the spectrum of the filtered samples using Fast Fourier Transform, and divide the model into training and testing sets; after constructing a generative adversarial network model for cross-domain transfer learning, train the model parameters using the training dataset, and use the test set as input to the trained model to obtain data generation and monitoring data fault prediction results; the adversarial network model is improved based on the auxiliary classifier generative adversarial network in the following ways, with the following steps: Step 3.1: Modify the network output unit of the generative model part contained in the network to L×d, where d is the number of working conditions modeled simultaneously; Step 3.2: Set the model input of the discriminant model part contained in the network to L×d; Step 3.3: Add a domain feature covariance difference regularization term L to the model loss function of the discriminative model part included in the network. CORAL and network weight complexity constraints The domain feature covariance difference regularization term is shown in the following formula: Among them, covariance , To determine the network weight matrix of the discriminant model; Step 3.4, the domain feature covariance difference regularization term L is... CORAL and network weight complexity constraints Combining hyperparameters and The loss function of the network incorporated into the discriminative model is shown in the following equation: in and Hyperparameters were determined using Bayesian optimization. Step 4: After normalizing the generated and real signals output by the data generation model, use the equidistant feature mapping and JS divergence similarity measurement methods to evaluate the data distribution differences between the generated samples and the real samples; calculate the quality evaluation index, and use the value of the index as the basis for judging the data generation quality.

2. The method for cross-domain migration generative adversarial network signal generation and quality evaluation according to claim 1, characterized in that, Step 2 includes: Step 2.1: Divide the signal into segments of length L, and let L=2. i The values ​​of i = 8, 9, 10, 11, 12 are converted into an m×n signal matrix through vector reconstruction and then subjected to singular value decomposition, where m×n = 2. i =L; Step 2.2: After decomposition, obtain a singular value vector of dimension m, and select the first k vectors. min One input Gaussian mixture clustering model; Step 2.3: After calculation, the clustering model outputs cluster labels t, and the samples with smaller signal amplitudes corresponding to the labels are removed.

3. The method for cross-domain migration generative adversarial network signal generation and quality evaluation according to claim 2, characterized in that, In step 2.2, the clustering results of the singular value vectors at k=1…m are calculated sequentially, and the difference sequence entropy of the cluster label t is calculated. The k value that minimizes the sequence entropy is selected as k. min As shown in the following formula: 。 4. The method for cross-domain migration generative adversarial network signal generation and quality evaluation according to claim 1, characterized in that, The calculation of quality evaluation indicators in step 4 includes: Step 4.1: Perform max-min normalization on the generated signal and the real signal. The normalization formula is shown below: Step 4.2: Use the equidistant feature mapping method to reduce the dimensionality of the signal. Select the first two dimensions of the reduced data and convert them into a one-dimensional vector as the feature vector v. f ; Step 4.3: Sort the eigenvalues ​​in order of magnitude, and also sort them by the maximum eigenvalue. and minimum value Set the feature partitioning interval, and set the interval length to N; use the frequency of the statistical feature value in each interval as the frequency vector element to finally form the statistical frequency vector v; Step 4.4: Calculate the JS divergence between the generated sample statistical frequency vector p1 and the original sample statistical frequency vector p2 to obtain the distance similarity index s, as shown in the following formula: After obtaining the distance similarity index s, the value of the index is used as the basis for judging the quality of data generation.

5. The method for cross-domain migration generative adversarial network signal generation and quality evaluation according to claim 4, characterized in that, In the data generation quality evaluation in step 4.4, the smaller the value and the closer it is to 0, the greater the similarity of data distribution between the generated data and the original data, and the higher the data generation quality. The larger the value and the closer it is to 1, the less similar the data distribution is between the generated data and the original data, and the lower the quality of the generated data.

Images