Conditional variation auto-encoder generative adversarial network fault diagnosis method

Abstract

The invention discloses a conditional variation auto-encoder generative adversarial network fault diagnosis method, and belongs to the technical field of mechanical equipment fault diagnosis. The method comprises the following steps: S1, data preprocessing: converting a time domain vibration signal into an amplitude-frequency signal through Fourier transform, and dividing a data set into a training set and a test set; s2, sample generation: constructing a conditional variation auto-encoder generative adversarial network model to amplify a fault sample; s3, data expansion: inputting the signal samples in the training set into a conditional variation auto-encoder generative adversarial network model for training, after training is completed, performing sample generation on various faults by using label information, generating a set number of samples, and mixing the generated samples to a training sample; and S4, sample classification: inputting the training samples after data expansion into a first-layer wide convolution kernel deep convolutional neural network classification model for training, then carrying out fault classification on the signal samples in the test set, and identifying various fault signal samples by the first-layer wide convolution kernel deep convolutional neural network. According to the conditional variation auto-encoder generative adversarial network fault diagnosis method provided by the invention, on the basis of a conditional variation auto-encoder generative model, an adversarial learning framework of a conditional generative adversarial network is utilized to further improve the trueness and diversity of samples; and samples with higher quality and more similar and diversified fault sample space dimensions can be generated.

Description

Technical Field

[0001] The present invention belongs to the technical field of mechanical equipment fault diagnosis, and in particular relates to a conditional variational autoencoder generating adversarial network fault diagnosis method. Background Art

[0002] Rolling bearings are key precision components in rotating equipment and are widely used in various types of mechanical equipment such as rail vehicles, power equipment, CNC machine tools, precision instruments, and aircraft engines. Their operating status is directly related to the performance and safety of the equipment and even the entire system. Once a rolling bearing fails, the impact should not be underestimated. Minor failures may lead to a decline in production quality, which in turn causes economic losses; serious failures may cause the destruction of the machine and the death of people, resulting in major safety accidents. Therefore, in engineering production, accurate diagnosis of the health status of rolling bearings is extremely important for real-time monitoring of the operating status of mechanical equipment and timely detection of safety hazards.

[0003] In the field of rolling bearing fault diagnosis, fault data generation technology based on deep learning has become a research hotspot. Among them, the improved conditional variational autoencoder and improved generative adversarial network have achieved certain results in data expansion. However, these technologies still have obvious limitations.

[0004] When generating samples, conditional variational autoencoders have problems such as poor authenticity and lack of details. This makes the generated samples deviate from reflecting the real fault characteristics of rolling bearings and cannot provide sufficiently accurate information for the diagnosis model. In addition, in actual applications, generative adversarial networks are prone to gradient explosion and training instability. This not only causes the generated samples to differ too much from the original samples, making it difficult to effectively simulate real fault data, but also increases the difficulty and complexity of model training, reducing training efficiency and reliability. These problems have largely limited the further development and widespread application of fault data generation technology based on deep learning in the field of rolling bearing fault diagnosis, and need to be solved urgently. Summary of the invention

[0005] In view of the deficiencies existing in the related art, the object of the present invention is to provide a conditional variational autoencoder generative adversarial network fault diagnosis method to solve the problems raised in the above background technology.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] A conditional variational autoencoder generates an adversarial network fault diagnosis method, comprising the following steps:

[0008] S1. Data preprocessing

[0009] The time domain vibration signal is converted into amplitude-frequency signal through Fourier transform, and the data set is divided into training set and test set;

[0010] S2. Sample generation

[0011] Construct a conditional variational autoencoder to generate an adversarial network model to amplify fault samples;

[0012] S3, Data Expansion

[0013] The signal samples in the training set are input into the conditional variational autoencoder to generate the adversarial network model for training. After the training is completed, the label information is used to generate samples for each type of fault, generate a set number of samples, and mix the generated samples into the training samples;

[0014] S4. Sample Classification

[0015] The training samples after data expansion are input into the first layer of wide convolution kernel deep convolution neural network classification model for training, and then the signal samples in the test set are classified into fault categories. The first layer of wide convolution kernel deep convolution neural network identifies various types of fault signal samples.

[0016] In some of the embodiments, the conditional variational autoencoder generative adversarial network includes an encoder, a generator, and a discriminator, wherein the encoder is composed of multiple convolution modules, and the generator is composed of multiple deconvolution modules to reconstruct samples.

[0017] In some embodiments, step S1 data preprocessing includes:

[0018] S11, preprocessing the original data;

[0019] S12, inputting the preprocessed data into the encoder in batches;

[0020] S13, mean μ and variance σ of encoder output 2 Synthesize into hidden variable z and input into the generator;

[0021] S14. Minimize the loss function L G To update the encoder and generator parameters, minimize the loss function L D To update the discriminator parameters;

[0022] In some embodiments, step S1 of data preprocessing further includes:

[0023] S15, repeat steps S11-S14 until a preset number of training rounds is reached;

[0024] S16. Use the trained generator model to generate fault samples with the same size as the fault samples.

[0025] In some embodiments, in step S14, the loss function L G for:

[0026]

[0027] Among them, x is the real data, To generate data, D KL is the Kullback-Leibler divergence, G(z) is the output of the generator, D(·) is the authenticity of the sample evaluated by the discriminator, z is the implicit variable, c is the label information, and Q(·) is the class prediction result of the discriminator.

[0028] In some embodiments, in step S14, the loss function L D for:

[0029]

[0030] Among them, P r is the real data distribution, is the expected value of a random variable x sampled from the true data distribution, P z is the latent variable distribution, is the expected value of a random variable z sampled from the latent variable distribution.

[0031] In some embodiments, in step S13, the encoder maps the input data x to a Gaussian distribution N(μ,σ) in the latent space. 2 ), and sample the latent variable z from this distribution.

[0032] In some embodiments, step S13 specifically includes:

[0033] S131, sampling a random noise vector ε from the standard normal distribution N(0,1);

[0034] S132. Synthesize the hidden variable z through the following formula:

[0035] z=μ+σ·ε

[0036] Among them, μ is the mean vector output by the encoder, and σ is the standard deviation vector output by the encoder.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] 1. The conditional variational autoencoder generative adversarial network fault diagnosis method provided by the present invention combines the conditional generation capability of the conditional variational autoencoder and the advantages of the generative adversarial network. On the basis of the conditional variational autoencoder generation model, the adversarial learning framework of the conditional generative adversarial network is used to further improve the authenticity and diversity of the samples. It can generate samples with higher quality, more similar and more diverse fault sample space dimensions, and its training model has good stability and high accuracy.

[0039] 2. The conditional variational autoencoder generative adversarial network fault diagnosis method provided by the present invention effectively improves the quality of generated samples and enhances the diversity of generated data in terms of data expansion. It also provides a more stable training process by optimizing the learning speed of the discriminator, thereby accelerating the convergence of the model and enabling fault diagnosis to be performed even when fault samples are difficult to obtain. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0041] Figure 1 A flowchart of an embodiment of a method for diagnosing faults in an adversarial network generated by a conditional variational autoencoder of the present invention;

[0042] Figure 2 A network structure diagram of a conditional variational autoencoder for generating an adversarial network fault diagnosis method according to an embodiment of the present invention;

[0043] Figure 3 A conditional generative adversarial network structure diagram of an embodiment of a conditional variational autoencoder generative adversarial network fault diagnosis method of the present invention;

[0044] Figure 4 A conditional variational autoencoder generates an adversarial network structure diagram of an embodiment of a conditional variational autoencoder generation adversarial network fault diagnosis method of the present invention;

[0045] Figure 5 A network structure diagram of a first-layer wide convolution kernel deep convolutional neural network of an embodiment of a conditional variational autoencoder generating an adversarial network fault diagnosis method of the present invention. DETAILED DESCRIPTION

[0046] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments. 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 work are within the scope of protection of the present invention.

[0047] In the description of the present invention, it should be understood that the terms "center", "lateral", "longitudinal", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside" and the like indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as a limitation on the present invention.

[0048] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a direct connection, or an indirect connection through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0049] See attached Figures 1 to 5 , an illustrative embodiment of the conditional variational autoencoder generation adversarial network fault diagnosis method proposed in the present invention is given, and the conditional variational autoencoder generation adversarial network fault diagnosis method comprises the following steps:

[0050] S1. Data preprocessing

[0051] The time domain vibration signal is converted into amplitude-frequency signal through Fourier transform, and the data set is divided into training set and test set;

[0052] S2. Sample generation

[0053] Construct a conditional variational autoencoder to generate an adversarial network model to amplify fault samples;

[0054] S3, Data Expansion

[0055] The signal samples in the training set are input into the conditional variational autoencoder to generate the adversarial network model for training. After the training is completed, the label information is used to generate samples for each type of fault, generate a set number of samples, and mix the generated samples into the training samples;

[0056] S4. Sample Classification

[0057] The training samples after data expansion are input into the first layer of wide convolution kernel deep convolution neural network classification model for training, and then the signal samples in the test set are classified into fault categories. The first layer of wide convolution kernel deep convolution neural network identifies various types of fault signal samples.

[0058] In step S1, taking rolling bearings as an example, since the original data of rolling bearings is a one-dimensional time domain vibration signal, there are problems such as dense sampling points, unclear fault characteristics, and difficult to distinguish noise information. Therefore, it is necessary to convert the time domain vibration signal into an amplitude-frequency signal through Fourier transform.

[0059] In step S3, the set number of samples generated is 300 samples. Taking rolling bearing fault diagnosis as an example, the characteristic performances of different types of faults such as inner ring fault, outer ring fault, and rolling element fault are different. Generating 300 samples for each fault allows the model to be exposed to more diverse fault characteristics and avoid overfitting due to insufficient data. The generated samples are then mixed into the training samples to achieve data enhancement.

[0060] The conditional variational autoencoder generative adversarial network includes an encoder, a generator and a discriminator, wherein the encoder is composed of multiple convolution modules, and the generator is composed of multiple deconvolution modules to reconstruct samples.

[0061] In this embodiment, step S1 data preprocessing includes:

[0062] S11, preprocessing the original data;

[0063] S12, inputting the preprocessed data into the encoder in batches;

[0064] S13, mean μ and variance σ of encoder output 2 Synthesize into hidden variable z and input into the generator;

[0065] S14. Minimize the loss function L G To update the encoder and generator parameters, minimize the loss function L D To update the discriminator parameters;

[0066] Step S1 data preprocessing also includes:

[0067] S15, repeat steps S11-S14 until a preset number of training rounds is reached;

[0068] S16. Use the trained generator model to generate fault samples with the same size as the fault samples.

[0069] In step S13, the encoder maps the input data x to the Gaussian distribution N(μ,σ 2 ), sampling from this distribution to obtain a hidden variable z, step S13 specifically includes:

[0070] S131, sampling a random noise vector ε from the standard normal distribution N(0,1);

[0071] S132. Synthesize the hidden variable z through the following formula:

[0072] z=μ+σ·ε

[0073] Among them, μ is the mean vector output by the encoder, and σ is the standard deviation vector output by the encoder (given by the variance σ 2 square root to obtain).

[0074] The mean μ determines the central position of the latent variable z, and the variance σ 2 Determines the distribution discreteness of the latent variable z. Different input data x will correspond to different μ and σ 2 , combined with random noise ε, the generated latent variable z can capture the characteristic information of the input data (through μ and σ 2 ), and has a certain degree of randomness and diversity (through ε). This feature helps the model learn the potential distribution of data and improve the quality and diversity of generated data.

[0075] The encoder maps the input data to the latent space, and the mean μ and standard deviation σ of the variational autoencoder output 2 It is used to generate potential representations, which are then used by the generator to generate samples. By minimizing the loss function L G , the encoder can learn a better way to map the input data to the latent space and provide a more effective latent representation for the generator. In step S14, the loss function L G for:

[0076]

[0077] Among them, x is the real data, To generate data, D KL is the Kullback-Leibler divergence, G(z) is the output of the generator, D(·) is the authenticity of the sample evaluated by the discriminator, z is the implicit variable, c is the label information, and Q(·) is the class prediction result of the discriminator.

[0078] The discriminator extracts features and reduces the dimension of the input real data and the data generated by the generator. Its optimization goal is to distinguish whether the sample is a real sample. In addition, the discriminator also needs to output the category of the sample to guide the model to generate samples of a specific category.

[0079] The discriminator is used to distinguish real samples from fake samples generated by the generator. The loss function L D Used to measure the discriminator's ability to distinguish between true and false samples. Minimize the loss function L D This enables the discriminator to better identify the difference between real samples and fake samples. In step S14, the loss function L D for:

[0080]

[0081] Among them, P r is the real data distribution, is the expected value of a random variable x sampled from the true data distribution, P z is the latent variable distribution, is the expected value of a random variable z sampled from the latent variable distribution.

[0082] The generator and encoder and the discriminator optimize their respective loss functions alternately (minimizing L G and L D ) for adversarial training. This adversarial training method enables the generator to continuously improve the quality of generated samples, and the discriminator to continuously improve its ability to distinguish true and false samples, and finally reach a relatively stable state, where the generator can generate high-quality samples that are close to the real data distribution.

[0083] In order to verify the effectiveness of the conditional variational autoencoder generative adversarial network method in fault diagnosis of rolling bearings under small sample conditions, the bearing dataset of Paderborn University in Germany was used for experimental verification.

[0084] The test bench consists of the following modules: motor, torque measurement shaft, rolling bearing test module, flywheel and load motor. The experimental data are collected by installing rolling bearings with different damage types in the bearing test module. The collected data set consists of healthy bearing data and faulty bearing data. The bearing model of the PU data set is 6203 deep groove ball bearing, the sampling frequency is 64kHz, and the collected data includes vibration signal data and motor current signal data. The selected data speed is N = 1500r / min, the load torque is M = 0.7Nm, and the bearing is subjected to radial force F = 400N.

[0085] In this embodiment, the bearing fault types collected in the test include three types: outer ring fault, inner ring fault and normal condition, and contain five fault types with different damage levels.

[0086] The network parameter settings are as follows: the programming environment is python3.8, and the Pytorch1.11.0 deep learning framework. The experiment is divided into two parts: conditional variational autoencoder generation adversarial network data expansion and fault diagnosis. When training the conditional variational autoencoder generation adversarial network framework, the batch size is set to 64, the Adam optimizer is used, the learning rate is set to 0.0001, and the iteration is 1500 times. The dimension of the implicit variable z is set to 100 dimensions. When training, the first layer of the wide convolution kernel deep convolutional neural network, the batch size is set to 64, the Adam optimizer is used, the learning rate is set to 0.001, and the iteration is 100 times.

[0087] In order to verify whether the fault diagnosis method of conditional variational autoencoder generative adversarial network can overcome the problem of lack of fault samples in fault diagnosis, the bearing data set collected by the bearing fault simulation test bench of Paderborn University was tested. Three groups of experiments were set up respectively. T1 represents the training set with 10 samples in each category, T2 represents the training set with 20 samples in each category, and T3 represents the training set with 50 samples in each category. Normal is a normal sample, OF1 represents the fault type when the damage depth of the outer ring fault bearing is D∈(0,2]mm, OF2 represents the fault type when the damage depth of the outer ring fault bearing is D∈(2,4.5]mm, IF1 represents the fault type when the damage depth of the inner ring fault bearing is D∈(0,2]mm, and IF2 represents the fault type when the damage depth of the inner ring fault bearing is D∈(2,4.5]mm. The number of each category of the training set in the small sample experiment is shown in the following table. There are 300 samples in each category of the test set, for a total of 1500 samples.

[0088]

[0089] In order to verify the effectiveness of the conditional variational autoencoder generative adversarial network fault diagnosis method for small sample fault diagnosis, the conditional variational autoencoder generative adversarial network will be combined with the conditional variational autoencoder neural network and the conditional generative adversarial network two conditional generation models and the rolling bearing fault samples of Paderborn data for data expansion, and combined with the first layer wide convolution kernel deep convolutional neural network model for fault classification. The results of the fault diagnosis accuracy of the above method on the small sample data set are shown in the following table.

[0090]

[0091] As can be seen from the above table, when there are only ten fault types (data set T1), the fault diagnosis accuracy of the data generated by the conditional variational autoencoder generative adversarial network is 89%, which is much higher than the other three comparison methods. As the number of real samples increases, after the conditional variational autoencoder generative adversarial network expands the data, the average accuracy of fault diagnosis is improved to varying degrees compared with other methods. When the number of real fault samples is 50, the average fault diagnosis accuracy after data expansion gradually increases, and the conditional variational autoencoder generative adversarial network method provided in this embodiment reaches the highest 96.84%. The experimental results show that the conditional variational autoencoder generative adversarial network plus the first layer of wide convolution kernel deep convolutional neural network model can effectively solve the small sample problem in fault diagnosis.

[0092] This embodiment aims at the problem of rolling bearing fault classification under the condition of only a small amount of sample data, and proposes a model framework based on a conditional variational autoencoder generative adversarial network and a first-layer wide convolution kernel deep convolutional neural network. First, the one-dimensional original signal is converted into a frequency signal through Fourier transform. Then, the conditional variational autoencoder generative adversarial network model is used to generate samples to achieve the expansion of normal samples and fault samples. Finally, the generated data and the original data are mixed and input into the first-layer wide convolution kernel deep convolutional neural network model for feature extraction and fault identification.

[0093] In the above-mentioned illustrative embodiment, the conditional variational autoencoder generative adversarial network fault diagnosis method combines the conditional generation capability of the conditional variational autoencoder and the advantages of the generative adversarial network. The adversarial learning framework of the conditional generative adversarial network is used on the basis of the conditional variational autoencoder generation model to further improve the authenticity and diversity of the samples. It can generate higher quality samples with more similar and diverse fault sample space dimensions, and its training model has good stability and high accuracy.

[0094] Finally, it should be noted that: the various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0095] The above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, ordinary technicians in the field should understand that the specific implementation methods of the present invention can still be modified or some technical features can be replaced by equivalents without departing from the spirit of the technical solution of the present invention, which should be included in the scope of the technical solution for protection of the present invention.

Claims

1. A conditional variational autoencoder generative adversarial network fault diagnosis method, characterized in that: The steps include: S1. Data preprocessing The time domain vibration signal is converted into amplitude-frequency signal through Fourier transform, and the data set is divided into training set and test set; S2. Sample generation Construct a conditional variational autoencoder to generate an adversarial network model to amplify fault samples; S3, Data Expansion The signal samples in the training set are input into the conditional variational autoencoder to generate the adversarial network model for training. After the training is completed, the label information is used to generate samples for each type of fault, generate a set number of samples, and mix the generated samples into the training samples; S4. Sample Classification The training samples after data expansion are input into the first layer of wide convolution kernel deep convolution neural network classification model for training, and then the signal samples in the test set are classified into fault categories. The first layer of wide convolution kernel deep convolution neural network identifies various types of fault signal samples.

2. The conditional variational autoencoder generative adversarial network fault diagnosis method according to claim 1, characterized in that: The conditional variational autoencoder generative adversarial network includes an encoder, a generator and a discriminator, wherein the encoder is composed of multiple convolution modules, and the generator is composed of multiple deconvolution modules to reconstruct samples.

3. The conditional variational autoencoder generative adversarial network fault diagnosis method according to claim 2, characterized in that: Step S1 data preprocessing includes: S11, preprocessing the original data; S12, inputting the preprocessed data into the encoder in batches; S13, mean μ and variance σ of encoder output 2 Synthesize into hidden variable z and input into the generator; S14. Minimize the loss function L G To update the encoder and generator parameters, minimize the loss function L D to update the discriminator parameters.

4. The conditional variational autoencoder generative adversarial network fault diagnosis method according to claim 3, characterized in that: Step S1 data preprocessing also includes: S15, repeat steps S11-S14 until a preset number of training rounds is reached; S16. Use the trained generator model to generate fault samples with the same size as the fault samples.

5. The conditional variational autoencoder generative adversarial network fault diagnosis method according to claim 3, characterized in that: In step S14, the loss function L G for: Among them, x is the real data, To generate data, D KL is the Kullback-Leibler divergence, G(z) is the output of the generator, D(·) is the authenticity of the sample evaluated by the discriminator, z is the implicit variable, c is the label information, and Q(·) is the class prediction result of the discriminator.

6. The conditional variational autoencoder generative adversarial network fault diagnosis method according to claim 5, characterized in that: In step S14, the loss function L D for: Among them, P r is the real data distribution, is the expected value of a random variable x sampled from the true data distribution, P z is the latent variable distribution, is the expected value of a random variable z sampled from the latent variable distribution.

7. The conditional variational autoencoder generative adversarial network fault diagnosis method according to claim 3, characterized in that: In step S13, the encoder maps the input data x to the Gaussian distribution N(μ,σ 2 ), and sample the latent variable z from this distribution.

8. The conditional variational autoencoder generative adversarial network fault diagnosis method according to claim 7, characterized in that: Step S13 specifically includes: S131, sampling a random noise vector ε from the standard normal distribution N(0,1); S132. Synthesize the hidden variable z through the following formula: z=μ+σ·ε Among them, μ is the mean vector output by the encoder, and σ is the standard deviation vector output by the encoder.

Images