A multi-scale fusion and time-frequency confrontation bearing fault signal generation method and system

Abstract

The present disclosure belongs to the technical field of bearing fault diagnosis, and specifically provides a bearing fault signal generation method and system based on multi-scale fusion and time-frequency confrontation, wherein the method comprises the following steps: constructing a generator based on multi-scale attention dynamic weighted fusion; constructing a time domain discriminator and a frequency domain discriminator, combining the outputs of the time domain discriminator and the frequency domain discriminator to obtain a comprehensive discriminator; using Wassertein distance as a loss function and adding a gradient penalty strategy; alternately training the generator and the comprehensive discriminator to obtain a bearing fault signal generation model; and inputting bearing fault data into the bearing fault signal generation model to obtain a fault generation signal. The method overcomes the problem of too small traditional bearing fault diagnosis samples, and improves the sample fidelity. By constructing the time domain discriminator and the frequency domain discriminator, the time-frequency characteristics of the fault samples are comprehensively analyzed, so that the accuracy and robustness of the diagnosis in the later fault diagnosis process can be enhanced.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of bearing fault diagnosis, and particularly relates to a bearing fault signal generation method and system based on multi-scale fusion and time-frequency confrontation. BACKGROUND

[0002] As one of the main rotating mechanical components, bearings are widely used in various fields such as engineering machinery, rail transportation, precision machine tools, and instruments. During operation, bearings are prone to wear, fatigue spalling, corrosion, and other faults under the action of force and load. These faults will produce characteristic vibration signals during the operation of the bearing. By monitoring and analyzing these vibration signals, real-time monitoring of the bearing state can be achieved, and possible faults can be detected and prevented in a timely manner, thereby significantly improving the reliability and service life of the equipment. With the rapid development of artificial intelligence, end-to-end deep learning technology has shown great potential in rotating machinery fault diagnosis. Based on deep learning models such as convolutional neural networks (CNN), deep belief networks (DBN), and stacked autoencoders (SAE), important achievements have been made in bearing fault diagnosis research. For the training of deep learning models, obtaining sufficient fault samples is crucial. However, in actual working conditions, fault samples are often difficult to obtain or have a very high cost. Therefore, how to effectively generate bearing fault samples has become a pressing problem.

[0003] In the prior art, a commonly used bearing signal generation method is based on a physical model. This method constructs an accurate bearing dynamics model through numerical simulation techniques such as finite element analysis (FEA) and multi-body dynamics analysis (MDA), simulates the bearing operation and fault process, and then generates vibration signals from the simulation results using signal processing techniques such as Fourier transform or wavelet transform. Another commonly used method is based on chaos theory, which generates signals with complex dynamic characteristics using chaos mapping (such as Lorenz mapping, Hénon mapping, etc.), and then adjusts the signal parameters (such as frequency, amplitude, etc.) to make them similar to actual bearing fault signals. However, both of the above methods have certain limitations and problems. The physical model-based method often requires the establishment of a complex physical model, while the chaos theory-based method requires manual parameter adjustment, which is inefficient. Therefore, how to generate high-quality and diverse bearing fault signals without relying on complex physical models and manual parameter adjustment has become an important research topic. The solution to this problem will have a profound impact on the research and application of bearing fault diagnosis technology.

[0004] In the field of mechanical equipment fault diagnosis, methods based on physical models and chaos theory have certain applications in generating bearing fault signals, but still have some defects.

[0005] The physical model-based method relies on numerical simulation techniques such as finite element analysis (FEA) and multi-body dynamics analysis (MDA). Its drawbacks include the need for accurate description of bearing structure and operating environment, which is often difficult in practical applications. At the same time, due to the large amount of calculation, the simulation process takes a long time. This makes the physical model-based method limited in bearing fault signal generation. In addition, this method requires high professional knowledge of the modeler, which may limit the accuracy of the model. Although the physical model can provide more realistic signal generation, it requires high time and effort in practical applications.

[0006] The chaos theory-based method generates signals with complex dynamic characteristics using chaotic maps such as Lorenz map and Henon map, and then adjusts the signal parameters such as frequency and amplitude to make them similar to the actual bearing fault signals. Although this method can generate signals with non-linear and non-stationary characteristics, it requires high selection and parameter adjustment of chaotic maps, and the generated signals may need further adjustment and verification. In addition, this method may not cover the fault signals of bearings under various working conditions, which affects the training effect of deep learning models. Therefore, the chaos theory-based method still has certain limitations in generating bearing fault signals.

[0007] In summary, the existing bearing fault signal generation methods have certain limitations, such as dependence on professional knowledge, large amount of calculation, insufficient diversity and authenticity of generated signals, etc. These defects limit the effectiveness of these methods in bearing fault diagnosis and prediction applications. Therefore, it is of great practical significance to study a new bearing fault signal generation method to more effectively support the training of deep learning models. SUMMARY

[0008] The present disclosure aims to at least solve one of the technical problems existing in the prior art, and proposes a multi-scale fusion and time-frequency adversarial bearing fault signal generation method and system.

[0009] In a first aspect, the present disclosure provides a multi-scale fusion and time-frequency adversarial bearing fault signal generation method, comprising:

[0010] constructing a multi-scale attention dynamic weighted fusion generator;

[0011] constructing a time domain discriminator and a frequency domain discriminator, and combining the outputs of the time domain discriminator and the frequency domain discriminator to obtain a comprehensive discriminator;

[0012] using Wassertein distance as the loss function and adding a gradient penalty strategy, alternately training the generator and the comprehensive discriminator to obtain a bearing fault signal generation model;

[0013] The bearing fault data is input into the bearing fault signal generation model to obtain the fault generation signal.

[0014] Preferably, the generator for constructing multi-scale attention-based dynamic weighted fusion specifically includes:

[0015] Construct a multi-scale generator that contains multiple sub-generators with consistent network structures;

[0016] A multi-head self-attention weighted fusion module is constructed to aggregate output features from multiple sub-generators in a dynamically weighted manner.

[0017] Preferably, the construction of the multi-head self-attention weighted fusion module, which aggregates output features from multiple sub-generators in a dynamically weighted manner, specifically includes:

[0018] The first step is to obtain the query matrix Q, the key matrix K, and the value matrix V by linear transformation of the input noise vector Z;

[0019] The second step is to split Q, K, and V along the head dimension to obtain Q corresponding to the i-th head. i K i and V i Calculate the scaled dot product Attention for each head;

[0020] The third step is to calculate the weighted feature Y by combining the h-th head and the i-th generator. hi =W hi ⊙X i Among them, W hi X represents the attention weight of the h-th head to the i-th generator. i This represents the output feature of the i-th sub-generator;

[0021] The fourth step is to sum the weighted features of all heads and generators to obtain the final feature fusion result.

[0022] Preferably, the construction of the time-domain discriminator specifically includes:

[0023] The first step is to define a one-dimensional convolutional layer to extract local features from the input sequence data;

[0024] The second step is to add nonlinearity through the LeakyReLU activation function to improve the expressive power of the time-domain discriminator;

[0025] The third step is to use a Dropout layer for regularization, randomly ignoring some neurons to prevent the model from overfitting.

[0026] The fourth step involves the model reducing the dimensionality of the data using an average pooling layer while retaining key predefined features.

[0027] In the fifth step, all features are integrated through a fully connected layer to output the final result of the time domain discriminator.

[0028] Preferably, the construction of the frequency domain discriminator specifically comprises:

[0029] Firstly, the input vibration signal sample is subjected to SFTF transformation to convert the vibration signal data from the time domain to the frequency domain.

[0030] Then, the frequency domain features are extracted through the same one-dimensional convolution layer, LeakyReLU activation function, Dropout layer, and average pooling layer.

[0031] Preferably, the alternating iterative training of the generator and the comprehensive discriminator specifically comprises:

[0032] The original bearing fault signal is subjected to resampling, noise reduction, and other preprocessing operations to obtain vibration signals of different fault types with a length of 2048;

[0033] The parameters of the generator and the comprehensive discriminator are randomly initialized.

[0034] The generator and the comprehensive discriminator are cross-trained, and the parameters are optimized and updated.

[0035] Preferably, the cross-training of the generator and the comprehensive discriminator and the optimization and updating of the parameters specifically comprise:

[0036] Firstly, the generator parameters are fixed, and the discriminator is trained for 5 times.

[0037] Then, the discriminator parameters are fixed, and the generator is trained for 5 times.

[0038] Finally, the training iteration is continuously performed until the preset training number of rounds is reached.

[0039] Preferably, the training of the generator comprises:

[0040] The noise signal is input into the generator to generate a batch of generated samples X fake ;

[0041] X fake is directly input into the time domain discriminator D t to obtain the output D t (X fake ) of the time domain discriminator; meanwhile, X fake is converted to the frequency domain through a spectrum transformation function SFTF (Short-time Fourier Transform) and then input into the frequency domain discriminator D f to obtain the output D f (SFTF(X fakeThe labels of all generated samples are set to 1.

[0042] Calculate the generator loss L(G), which is equal to the sum of the outputs of the time-domain discriminator and the frequency-domain discriminator. The specific calculation formula is as follows:

[0043] L(G) = E[D t (X fake ))]+E[D f (SFTF(X fake ))]

[0044] Where E is the expected value, which is the average of all generated samples;

[0045] The Adam optimizer is used to update the generator's parameters based on the calculated loss.

[0046] Preferably, the training of the integrated discriminator includes:

[0047] Input the real samples from the time domain into the time domain discriminator D. t The true output D of the time-domain discriminator is obtained from this. t (X real ), and assign label 1 to these real samples, inputting the real samples in the frequency domain into the frequency domain discriminator D. f The true output D of the frequency domain discriminator is obtained from this. f (SFTF(X real Similarly, label 1 is assigned to these samples;

[0048] Calculate the Wasserstein loss L(D) for the real samples. real )=E[D t (X real )]+E[D f (SFTF(X real ))], where E is the calculation of the expected value;

[0049] The time-domain generated samples from generator G are input into the time-domain discriminator D. t The time-domain output D of the sample is obtained. t (G(X fake The generated samples are then labeled -1 and input into the frequency domain discriminator D. f The sample frequency domain output D is obtained from f (SFTF(G(X fake And assign a label of -1 to these samples;

[0050] Calculate the Wasserstein loss L(D) of the generated samples. fake )=E[D t (G(Xfake ))]+E[D f (SFTF(G(X fake )))];

[0051] The difference between the Wasserstein loss of the real sample and the Wasserstein loss of the generated sample is calculated, and the total loss of the discriminator is obtained:

[0052] L(D)=L(D real )-L(D fake )

[0053] Finally, the parameters of the discriminator are updated according to this loss using the Adam optimizer.

[0054] In a second aspect, the present disclosure provides a multi-scale fusion and time-frequency confrontation bearing fault signal generation system, which can be used to implement the bearing fault signal generation method based on multi-scale fusion and time-frequency confrontation described above. The system comprises:

[0055] A generator construction module configured to construct a multi-scale attention dynamic weighted fusion generator;

[0056] A discriminator construction module configured to construct a time-domain discriminator and a frequency-domain discriminator, and combine the outputs of the time-domain discriminator and the frequency-domain discriminator to obtain a comprehensive discriminator;

[0057] A training module configured to use Wassertein distance as a loss function, add a gradient penalty strategy, and alternately and iteratively train the generator and the comprehensive discriminator to obtain a bearing fault signal generation model;

[0058] An output module configured to input bearing fault data into the bearing fault signal generation model to obtain a fault generated signal. BRIEF DESCRIPTION OF DRAWINGS

[0059] Figure 1 A flowchart of a multi-scale fusion and time-frequency confrontation bearing fault signal generation method provided by an embodiment of the present disclosure;

[0060] Figure 2 A network structure diagram of a sub-generator provided by an embodiment of the present disclosure;

[0061] Figure 3 A comparison diagram of real and generated signals provided by an embodiment of the present disclosure;

[0062] Figure 4 A comparison diagram of real and generated signal probability densities provided by an embodiment of the present disclosure. DETAILED DESCRIPTION

[0063] In order for those skilled in the art to better understand the technical solutions of the present disclosure, the present disclosure is further described in detail below in conjunction with the drawings and specific embodiments.

[0064] Unless otherwise defined, technical terms or scientific terms used in the present disclosure shall have the ordinary meaning understood by one of ordinary skill in the art to which the present disclosure pertains. The terms "first", "second", and similar terms used in the present disclosure do not denote any order, quantity, or importance, but are used to distinguish different components. Similarly, the terms "one", "a", or "the" and similar terms do not denote a quantity of at least one, but are used to indicate the existence of at least one. The terms "include", "contain", and similar terms mean that the elements or objects before the terms encompass the elements or objects listed after the terms and their equivalents, and do not exclude other elements or objects. The terms "connect" or "connected" and similar terms do not mean a physical or mechanical connection, but can include an electrical connection, whether direct or indirect. The terms "up", "down", "left", "right", and the like are used only for relative positional relationships, and when the absolute position of the described object changes, the relative positional relationship can also change accordingly.

[0065] In the various drawings, the same elements are denoted by similar reference numerals. For the sake of clarity, not all parts of the drawings are drawn to scale. In addition, some parts that are well known can not be shown in the drawings.

[0066] Many specific details of the present disclosure are described below, such as the structure, materials, dimensions, processing procedures, and techniques of components, in order to more clearly understand the present disclosure. However, as those skilled in the art can understand, the present disclosure can be implemented without these specific details.

[0067] As shown in Figure 1 and Figure 2 The present disclosure provides a multi-scale fusion and time-frequency confrontation bearing fault signal generation method, comprising the following steps:

[0068] S1, constructing a multi-scale attention dynamic weighted fusion generator;

[0069] S2, constructing a time domain discriminator and a frequency domain discriminator, combining the outputs of the time domain discriminator and the frequency domain discriminator to obtain a comprehensive discriminator;

[0070] S3, using Wassertein distance as a loss function, and adding a gradient penalty strategy, alternately training the generator and the comprehensive discriminator to obtain a bearing fault signal generation model;

[0071] S4, inputting bearing fault data into the bearing fault signal generation model to obtain a fault generated signal.

[0072] It should be noted that the technical solutions of the present disclosure do not limit the sequence of the above steps, that is, the execution sequence of each step can be arranged arbitrarily.

[0073] The objectives of the present disclosure mainly focus on the following aspects:

[0074] Overcoming small sample limitations: In the field of bearing fault diagnosis, there is often a problem of limited sample size, which poses a challenge to the training and generalization performance of the fault diagnosis model. The present disclosure aims to generate more bearing fault signal samples through an improved generative adversarial network (GAN) to overcome the small sample limitation.

[0075] Improving sample fidelity: Existing bearing fault signal generation techniques often struggle to produce highly realistic samples, which can negatively impact the accuracy of fault diagnosis. The objective of the present disclosure is to improve the structure of the generator and replace traditional two-dimensional convolution with one-dimensional convolution to more effectively learn the patterns of time series data and improve the fidelity of the samples.

[0076] Enhancing the accuracy and robustness of fault diagnosis: Many existing bearing fault diagnosis methods fail to adequately consider the importance of time-frequency features, which can affect the accuracy and robustness of fault diagnosis. The objective of the present disclosure is to construct time and frequency discriminators to comprehensively analyze the time-frequency features of fault samples, thereby enhancing the accuracy and robustness of diagnosis in the later fault diagnosis process.

[0077] In summary, the main objective of the present disclosure is to develop a new bearing fault signal generation method to address the small sample problem, the lack of sample fidelity, and the lack of accuracy and robustness of deep learning-based fault diagnosis models in existing technologies.

[0078] In a specific implementation scenario, the present disclosure proposes a multi-scale fusion and time-frequency adversarial bearing fault signal generation method, i.e., a bearing fault signal generation method based on multi-scale attention dynamic weighting fusion and hybrid time-frequency adversarial generation, to address the problem of low quality of bearing fault sample generation and data imbalance. At the same time, the present disclosure uses one-dimensional convolution to replace the two-dimensional convolution of traditional GAN networks to better learn the patterns of time series data, further enhancing the fidelity of the generated samples. As Figure 1The network structure of the present disclosure is shown, which improves the generator structure of the traditional GAN, introduces a dynamic weighted multi-scale generator, and uses a multi-head attention mechanism to weight and fuse features of different scales to achieve more detailed feature expression and more accurate generation effect. Then, time domain and frequency domain discriminators are respectively constructed to evaluate the similarity of the generated samples from the time-frequency angle. By constructing the time domain and frequency domain discriminators, the time-frequency characteristics of the fault samples are comprehensively analyzed, thereby enhancing the accuracy and robustness of fault diagnosis. At the same time, a new type of adversarial loss function is designed in combination with the Wasserstein distance and gradient penalty to stabilize the training process and avoid gradient disappearance. The specific scheme is divided into the following steps:

[0079] Step 1: Construct a multi-scale attention dynamic weighted fusion generator.

[0080] The present disclosure changes the traditional single generator design to multiple sub-generators and designs a new framework of "multi-scale generator". The network structure of each sub-generator is consistent, as shown in Figure 2 The network structure of a single sub-generator includes a fully connected layer, a batch normalization (Batch Normalization) layer, a PReLU activation function, a reshape layer, a one-dimensional transpose convolution (Conv1D Transpose) layer, and a Dropout layer, etc. The input is a one-dimensional noise signal with a length of 100, which is mapped to a 128-dimensional hidden space through the fully connected layer. Then, the batch normalization layer is used to reduce the internal covariate shift, and the PReLU activation function is used to introduce nonlinearity. Then, the Reshape is used to flatten into a single channel sequence as the input of the first one-dimensional transpose convolution layer. Each transpose convolution layer uses a convolution kernel with a kernel size of 4 and a step size of 2, and uses the "same" padding mode to ensure that the feature size increases by 2 times in order. At the same time, BatchNorm standardization and PReLU activation are followed after each convolution layer. The last transpose convolution layer reduces the number of channels to 1 and expands the feature sequence length to 2048. Finally, the dimension is compressed to the shape of the target signal through the fully connected layer.

[0081] Then, a multi-head self-attention weighted fusion module is constructed to dynamically weight and aggregate the output features from multiple sub-generators. Specifically:

[0082] First, the input noise vector Z is linearly transformed to obtain the query matrix Q, the key matrix K, and the value matrix V.

[0083] Second, Q, K, and V are split in the head dimension to obtain Q i , K i , and V i corresponding to the i-th head. The scaled dot-product attention is calculated for each head:

[0084]

[0085] where d k is a scaling factor. This prevents the inner product from being too large. The scaled dot-product attention outputs for each head are concatenated to obtain the final multi-head self-attention output.

[0086] Thirdly, the obtained attention matrix is used to weight and fuse the output feature X of each sub-generator. The weighted feature is calculated for the hth head and the ith generator:

[0087] Y hi = W hi ⊙X i ,

[0088] where W hi represents the attention weight of the hth head to the ith generator, and X i represents the output feature of the ith sub-generator.

[0089] Fourthly, the weighted features of all heads and generators are summed to obtain the final feature fusion result, i.e., the generated signal Y:

[0090]

[0091] Step 2: Constructing the time domain discriminator and the frequency domain discriminator. The time domain discriminator is specifically designed to handle signal features in the time domain. Specifically, it includes:

[0092] First, a one-dimensional convolutional layer is defined, with a convolution kernel size of 3, a number of convolution kernels of 512, and a padding method of "same" for the convolution operation. This one-dimensional convolutional layer can extract local features from the input sequence data.

[0093] Then, a LeakyReLU activation function is used to increase nonlinearity and improve the model's expression ability.

[0094] Then, the model is regularized by a Dropout layer, which randomly ignores some neurons to prevent overfitting of the model.

[0095] Then, the model reduces the dimension of the data while preserving important feature information through an average pooling layer. This process is repeated three times, with the number of convolution kernels in each convolutional layer being 512, 1024, and 2048, respectively. After each convolutional layer, there is a LeakyReLU activation function, a Dropout layer, and an average pooling layer.

[0096] Then, the model converts the multi-dimensional data into one-dimensional data through a Flatten layer, making it easier for the fully connected layer to process.

[0097] Finally, the model integrates all features through a fully connected layer to output the final result.

[0098] The design of a frequency domain discriminator is the same as that of a time domain discriminator, but it deals with signal characteristics in the frequency domain.

[0099] First, the input vibration signal sample is subjected to SFTF transformation to convert the vibration signal data from the time domain to the frequency domain.

[0100] Then, frequency domain features are extracted using the same one-dimensional convolutional layer, LeakyReLU activation function, Dropout layer, and average pooling layer.

[0101] The time-frequency hybrid discriminator combines the outputs of the two discriminators to utilize signal characteristics in both the time and frequency domains for discrimination. This design allows the model to fully leverage information from the input data in both the time and frequency domains, resulting in better discrimination.

[0102] Step 3: Design the loss function.

[0103] During training, Wassertein distance is used instead of the traditional JS divergence to measure the distribution difference between generated samples and real samples. The specific formula is as follows:

[0104]

[0105] Among them, Π(P) r P g ) is P r and P g It is the set of γ values ​​of the joint distribution of marginal distributions; W(P) r P g ) is the infimum of the expectation of γ(x,y).

[0106] In the original Wassertein distance-based GAN, weight pruning of the discriminator can lead to gradient explosion. Therefore, this disclosure uses gradient penalty instead of weight pruning, as follows:

[0107]

[0108] In the formula ||·|| p Let p be the norm; The uniform sampling between real and generated samples is calculated using the following formula:

[0109]

[0110] This design effectively combines information from time-domain and frequency-domain data to update model parameters, thereby achieving a significant performance improvement.

[0111] Step 4: Model training and sample generation.

[0112] The training process of the present disclosure adopts the idea of a generative adversarial network, and continuously iteratively optimizes the generator and the discriminator to achieve higher quality of generated samples. The specific training process is as follows:

[0113] (1) Data preprocessing: Before model training, data preprocessing is first performed. Since the present disclosure is specifically for time series data, the original bearing fault signal needs to be resampled, denoised, and other preprocessing operations, and finally a vibration signal of different fault types with a length of 2048 is obtained to facilitate subsequent model learning.

[0114] (2) Randomly initialize the parameters of the generator and the discriminator. In addition, the hyperparameters of the optimizer Adam are also appropriately initialized.

[0115] (3) Generator training.

[0116] The noise signal Z ∈ R m×d is input, where m is the batch size and d is the dimension of the latent space. In the present disclosure, d = 100, and for each sub-generator G i (i = 1, 2, …, n, where n is the number of sub-generators), receives the noise signal as input and inputs it into the generator to generate a batch of generated samples X fake .

[0117] X fake is directly input into the time domain discriminator D t , to obtain the output of the time domain discriminator D t (X fake ). At the same time, X fake is converted to the frequency domain through a short-time Fourier transform function SFTF, and then input into the frequency domain discriminator D f , to obtain the output of the frequency domain discriminator D f (SFTF(X fake )). Wherein the label of the generated sample is set to 1. Then, the loss L(G) of the generator is calculated, which is equal to the sum of the output of the time domain discriminator and the output of the frequency domain discriminator. The specific calculation formula is as follows:

[0118] L(G) = E[D t (X fake ))] + E[D f (SFTF(X fake ))]

[0119] where E is the computation of the expectation, i.e. the average over all generated samples. Finally, the parameters of the generator are updated using the Adam optimizer according to the computed loss.

[0120] (4) Comprehensive discriminator training.

[0121] In the discriminator training phase, real samples in time domain are input into the time domain discriminator D t to get the real output of the time domain discriminator D t (X real ), and label 1 is assigned to these real samples. For real samples in frequency domain (converted by the spectrum transform function SFTF), they are input into the frequency domain discriminator D f to get the real output of the frequency domain discriminator D f (SFTF(X real )), and label 1 is also assigned to these samples. Then, the expectation values of these results are added to get the Wasserstein loss of real samples:

[0122] L(D real ) = E[D t (X real )] + E[D f (SFTF(X real ))

[0123] Then, generated samples in time domain (generated by the generator G) are input into the time domain discriminator D t to get the sample time domain output D t (G(X fake )), and label -1 is assigned to these generated samples. For generated samples in frequency domain (converted by SFTF on the time domain generated samples), they are also input into the frequency domain discriminator D f to get the sample frequency domain output D f (SFTF(G(X fake ))), and label -1 is assigned to these samples. The expectation values of these results are added to get the Wasserstein loss of generated samples:

[0124] L(D fake ) = E[D t (G(X fake ))]+E[D f (SFTF(G(X fake )))]

[0125] The difference between the Wasserstein loss of real samples and the Wasserstein loss of generated samples is calculated to get the total loss of the discriminator:

[0126] L(D) = L(Dreal )-L(D fake )

[0127] Finally, the parameters of the discriminator are updated according to this loss using the Adam optimizer.

[0128] (5) Alternating iterative optimization: After the start of training, first fix the generator parameters, train the discriminator for 5 times, then fix the parameters of the discriminator, and train the generator for 5 times. Finally, continue to train iterations until the preset training number is reached. This alternating iteration method can help maintain the balance between the generator and the discriminator, preventing one model from dominating the training process too early.

[0129] The invention point of the present disclosure is:

[0130] 1. Multi-scale attention dynamic weighting fusion generator: The generator is no longer a single structure, but is composed of multiple sub-generators, each of which can generate features of different scales from input noise. Then, by innovatively introducing a multi-head attention mechanism, weights are dynamically assigned to each sub-generator, and the generated features are finally fused. This design enables the model to capture more rich and fine feature information, thereby improving the quality and authenticity of the generated samples.

[0131] 2. Hybrid time-frequency discriminator: In this disclosure, the discriminator is designed in a new mode, including a time-domain discriminator and a frequency-domain discriminator. The time-domain discriminator compares the generated samples and real samples in the time sequence, while the frequency-domain discriminator compares their similarity in frequency distribution. This hybrid discriminator comprehensively evaluates the similarity of the samples from multiple angles, greatly improving the accuracy of discrimination.

[0132] 3. Improved loss function: In the training process, the Wassertein distance is used as the loss function, and a gradient penalty strategy is added. Compared with the traditional JS divergence, it can more effectively solve the mode collapse problem, thereby improving the training stability of the model and the quality of the generated samples.

[0133] 4. One-dimensional convolution operation: The model uses one-dimensional convolution operation instead of traditional two-dimensional convolution, making the model more suitable for processing time series data, better learning and extracting data patterns, and further improving the realism of the generated samples.

[0134] The integration of these key technical points makes the present disclosure have significant advantages in dealing with small sample and data imbalance problems in mechanical equipment bearing fault diagnosis, and has high practical value.

[0135] The embodiment of the present disclosure also provides a multi-scale fusion and time-frequency confrontation bearing fault signal generation system, which can be used to realize the bearing fault signal generation method based on multi-scale fusion and time-frequency confrontation.

[0136] A generator construction module is configured to construct a multi-scale attention dynamic weighted fusion generator.

[0137] A discriminator construction module is configured to construct a time domain discriminator and a frequency domain discriminator, and combine the outputs of the time domain discriminator and the frequency domain discriminator to obtain a comprehensive discriminator.

[0138] A training module is configured to use Wassertein distance as a loss function, and add a gradient penalty strategy, and alternately iteratively train the generator and the comprehensive discriminator to obtain a bearing fault signal generation model.

[0139] An output module is configured to input bearing fault data into the bearing fault signal generation model to obtain a fault generation signal.

[0140] The accuracy of the model is compared and analyzed in a specific case as follows:

[0141] The present disclosure uses a bearing fault dataset from Case Western Reserve University (CWRU) to perform a bearing vibration signal generation experiment. This dataset contains bearing vibration signals under various fault conditions, which is very suitable for bearing fault diagnosis and performance research. The experimental data is the driving end bearing data with a sampling frequency of 12 kHz and a rotating speed of 1750 revolutions per minute. The fault types tested include inner ring fault (IR 014), outer ring fault (OR 014) and rolling element fault (B 014), wherein 014 corresponds to a fault size of 0.356. The model parameter settings are as shown in Table 1:

[0142] Table 1 Model parameter settings

[0143] Parameter name Set value Attention heads (generator and discriminator) 5 Wasserstein distance weight 10 Gradient penalty weight 2 Batch size 10 Generator learning rate (Adam) 0.01 Time domain discriminator learning rate (Adam) 0.005 Frequency domain discriminator learning rate (Adam) 0.005 Number of training epochs 5000

[0144] Through the algorithm model designed by the present disclosure, three kinds of bearing fault signals including inner ring fault, outer ring fault and rolling element fault are generated. As shown in Figure 3 The comparison of the three kinds of generated fault signals and the real signals is shown in the figure. The generated fault signals have high similarity with the real signals in waveform. At the same time, in order to more intuitively evaluate the similarity of the generated signals and the real signals, further analysis is performed from the perspective of probability distribution. Figure 4The generator shows the probability distribution of the real signal and the generated signal under three different fault states. Although there are some differences between the distribution of the generated data and the distribution of the real data in some cases, these differences are relatively small, so it can be considered that the generator has successfully learned the distribution of the real fault signal.

[0145] It is proved by experiments. The present disclosure has shown significant advantages and effects in bearing fault signal generation. First, the present disclosure can generate a variety of high-quality bearing fault signals, which are highly similar to real fault signals in waveform and probability distribution, fully demonstrating the ability of the model to learn and replicate the complex characteristics of real fault signals. At the same time, this method provides a new solution to the problem of data imbalance. In many practical applications, the problem of data imbalance is a challenge that cannot be ignored. Normal bearing data is often much more than fault data. The present disclosure can generate a large number of fault signals, which helps to balance the number of normal and fault data, so as to improve the training effect and prediction accuracy of the fault diagnosis model. Finally, the present disclosure can continue to train and optimize the fault diagnosis model without new fault data, avoiding high-cost and high-risk fault simulation experiments, thereby saving experimental resources and costs. In summary, the present disclosure proposes an efficient and reliable method that can effectively improve the efficiency and accuracy of bearing fault diagnosis by generating high-quality bearing fault signals, which has important practical significance for preventing equipment failure and improving equipment maintenance efficiency.

[0146] It can be understood that the above embodiments are only exemplary embodiments adopted for illustrating the principles of the present disclosure, but the present disclosure is not limited thereto. Various modifications and improvements can be made by those skilled in the art without departing from the spirit and essence of the present disclosure, and these modifications and improvements are also considered within the protection scope of the present disclosure.

Claims

1. A method for generating bearing fault signals with multi-scale fusion and time-frequency confrontation, characterized in that, The application relates to a bearing fault signal generation method and device. The application comprises the following steps: A multi-scale attention dynamic weighting fusion generator is constructed; A time domain discriminator and a frequency domain discriminator are constructed, and the outputs of the time domain discriminator and the frequency domain discriminator are combined to obtain a comprehensive discriminator; Wassertein distance is used as a loss function, and a gradient penalty strategy is added, and the generator and the comprehensive discriminator are alternately iteratively trained to obtain a bearing fault signal generation model; Bearing fault data is input into the bearing fault signal generation model to obtain a fault generation signal; The multi-scale attention dynamic weighting fusion generator comprises the following steps: A multi-scale generator is constructed, and a plurality of sub-generators with consistent network structures are included; A multi-head self-attention weighting fusion module is constructed to dynamically aggregate the output features from the plurality of sub-generators; In the first step, the input noise vector Z is transformed linearly to obtain the query matrix Q, the key matrix K and the value matrix V K V ;​ The second step is to perform a head-dimensional analysis. Q , K , V Split into the first... i The corresponding size Q i , K i and V i Calculate the scaled dot product Attention for each head; The third step is to conduct research on the first... h Height and First i Each generator calculates weighted features. ,in, W hi Indicates the first h The head is the first i Attention weights for each generator This represents the output feature of the i-th sub-generator; The multi-head self-attention weighting fusion module dynamically aggregates the output features from the plurality of sub-generators, and comprises the following steps: In the fourth step, the weighted features of all heads and generators are summed to obtain a final feature fusion result; The time domain discriminator is constructed, and comprises the following steps: In the first step, a one-dimensional convolution layer is defined to extract local features from input sequence data; In the second step, a LeakyReLU activation function is used to increase nonlinearity to improve the expression ability of the time domain discriminator; In the third step, a Dropout layer is used for regularization, and part of the neurons are randomly ignored to prevent model overfitting; In the fourth step, the model is reduced in dimension through an average pooling layer while retaining preset important feature information; In the fifth step, a fully connected layer is used to integrate all features, and the final result of the time domain discriminator is output. The frequency domain discriminator is constructed, and comprises the following steps: Firstly, SFTF transformation is performed on the input vibration signal sample to convert the vibration signal data from the time domain to the frequency domain; 2. The multi-scale fusion and time-frequency confrontation bearing fault signal generation method according to claim 1, characterized in that, Then, the same one-dimensional convolution layer, LeakyReLU activation function, Dropout layer and average pooling layer are used to extract frequency domain features. The alternately iteratively trained generator and comprehensive discriminator comprise the following steps: The original bearing fault signal is resampled and denoised to obtain vibration signals with a length of 2048 for different fault types; The parameters of the generator and the comprehensive discriminator are randomly initialized; 3. The method of claim 2, wherein, The generator and the comprehensive discriminator are cross-trained, and the parameters are updated and optimized. The cross-training of the generator and the comprehensive discriminator and the optimization and updating of the parameters comprise the following steps: Firstly, the generator parameters are fixed, and the discriminator is trained for 5 times; Then, the discriminator parameters are fixed, and the generator is trained for 5 times; 4. The method of claim 2, wherein, Finally, the training iteration is continuously performed until a preset training number is reached. feeding a noise signal as input into the generator to generate a batch of generated samples X fake ; Will X fake directly input to the time domain discriminator D t , get the output of the time domain discriminator D t ( X fake );At the same time, input X fake to the frequency domain discriminator through a short-time Fourier transform (SFTF) function D f , get the output of the frequency domain discriminator D f (SFTF( X fake ));Wherein, the labels of the generated samples are all set to 1; Calculate the loss of the generator L ( G The loss is equal to the sum of the outputs of the time-domain discriminator and the frequency-domain discriminator, and the specific calculation formula is as follows: L ( G ) = E [ D t ( X fake ))]+ E [ D f (SFTF( X fake ))] wherein E is the calculation of the expected value, i.e. the average over all generated samples; The training of the generator comprises the following steps:

5. The method of claim 2, wherein, The parameters of the generator are updated by using an Adam optimizer according to the calculated loss. real samples in time domain are input into the time domain discriminator D t to obtain real outputs of the time domain discriminator D t ( X real ), and the real samples in frequency domain are input into the frequency domain discriminator D f to obtain real outputs of the frequency domain discriminator D f (SFTF( X real ), and the samples are also assigned with label 1; computing the Wasserstein loss for real samples L ( D real ) = E [ D t ( X real )] + E [ D f (SFTF( X real ))], wherein E is the computation of the expected value; The generator G The generated time domain generated samples are input to the time domain discriminator D t to obtain sample time domain outputs D t G X fake The generated frequency domain generated samples are input to the frequency domain discriminator D f to obtain sample frequency domain outputs D f G X fake ​​​​​ Wasserstein loss of the computed generated samples L ( D fake )= E [ D t ( G ( X fake ))] + E [ D f (SFTF( G ( X fake )))] The training of the comprehensive discriminator comprises the following steps: L ( D )= L ( D real )- L ( D fake ) The difference between the Wassertein loss of the real sample and the Wassertein loss of the generated sample is calculated to obtain the total loss of the discriminator: Finally, the parameters of the discriminator are updated by using an Adam optimizer according to the loss.

6. A multi-scale fusion and time-frequency confrontation bearing fault signal generation system, characterized in that, The system can be used to realize the multi-scale fusion and time-frequency confrontation bearing fault signal generation method in any one of claims 1-5 above, and the system comprises: a generator construction module configured to construct a multi-scale attention dynamic weighted fusion generator; a discriminator construction module configured to construct a time domain discriminator and a frequency domain discriminator, and combine outputs of the time domain discriminator and the frequency domain discriminator to obtain a comprehensive discriminator; a training module configured to use a Wassertein distance as a loss function, add a gradient penalty strategy, and alternately and iteratively train the generator and the comprehensive discriminator to obtain a bearing fault signal generation model; an output module configured to input bearing fault data into the bearing fault signal generation model to obtain a fault generation signal.

Images