A motor bearing fault diagnosis method based on a reconstruction generative adversarial network

Abstract

The application discloses a motor bearing fault diagnosis method and system based on a reconstruction generative adversarial network. The method comprises the following steps: collecting majority class health samples and minority class fault samples under an unbalanced data set, and labeling them as real samples; constructing a generator network and a discriminator network, wherein the generator network comprises an input layer, four convolutional layers, and an output layer, and the discriminator network comprises an input layer, two convolutional layers, and a fully connected layer; constructing a reconstruction layer network, wherein the reconstruction layer network is composed of two sub-networks, namely an encoder and a decoder; training the generator network based on a random noise signal, and outputting a signal labeled as a generated sample; training the discriminator network based on the real samples and the generated samples; training the reconstructor based on the generated sample, obtaining a reconstructed generated sample, calculating the Kendall correlation between the reconstructed generated sample and the real sample, and feeding back the Kendall correlation; mixing and expanding the trained generated sample and the real sample to a balanced sample data set, and inputting the balanced sample data set into a neural network module to realize fault diagnosis under the unbalanced data set. The application can generate diversified sample data based on the minority class fault data, expand the unbalanced data set to an ideal balanced state, and train a fault diagnosis model without collecting fault sample data in an experimental environment for a long time.

Description

A method for motor bearing fault diagnosis based on reconstructed generative adversarial networks Technical Field

[0001] This invention relates to the field of motor bearing fault diagnosis, and specifically to a method and system for diagnosing motor bearing imbalance datasets based on reconstructed generative adversarial networks. Background Technology

[0002] With the widespread application of smart sensors in industry, a massive amount of process data from machine operation has been accumulated, providing a foundation for data-driven fault diagnosis. As machine systems become increasingly complex, traditional fault diagnosis methods are no longer sufficient to meet the requirements of fault detection. In recent years, deep learning, due to its network structure's ability to automatically learn data features, has seen rapid development and application in the field of machine fault diagnosis.

[0003] While deep learning has achieved significant results in machine fault diagnosis, it's important to note that its success relies on balanced datasets. However, in engineering practice, the collected samples are typically imbalanced. Machines operate in a fault-free state for extended periods, with uptime far exceeding the time spent in operation. It's also not advisable to operate machines in a faulty state to collect data samples, as this will shorten their lifespan. Traditional deep learning models, when trained on imbalanced data, force the classifier to sacrifice diagnostic accuracy for minority class samples, favoring predictions for majority class samples, ultimately leading to misclassification and poor fault diagnosis results.

[0004] To address the challenges posed by imbalanced datasets in fault diagnosis, various methods have been proposed. These methods can be categorized into two types based on the fault diagnosis process: data-based methods and model-based methods. Synthetic minority oversampling technique (SMOTE) and generative adversarial networks (GANs) represent data-based and model-based methods, respectively. SMOTE synthesizes new minority class samples through random linear interpolation between minority class samples, while GANs generate usable sample data by learning the features of existing data samples. However, in SMOTE, samples existing at the classification margin may increase the difficulty of classification, and the newly synthesized samples contain only limited fault information. Furthermore, GAN data augmentation models typically only include a generator and a discriminator, requiring overcoming the mode collapse and gradient vanishing problems. Therefore, how to effectively diagnose motor bearing faults using imbalanced datasets is a pressing issue that needs to be addressed. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention provides a fault diagnosis method for motor bearings under unbalanced datasets. This method introduces a reconstruction module and an autoencoder function, which can quickly and accurately learn the feature signals of minority class fault samples, improve the quality of generated samples of the model, and solve the problems mentioned in the background art.

[0006] To achieve the above objectives, this invention provides a method for diagnosing motor bearing faults based on reconstructed generative adversarial networks, specifically including the following steps:

[0007] 1) When it is necessary to perform fault detection on motor bearings under an imbalanced dataset, it is necessary to obtain the feature datasets corresponding to the majority class healthy samples and the minority class fault samples. According to the imbalance ratio, the majority class normal sample signals and the minority class fault sample signals are collected respectively. The minority class fault data under the imbalanced dataset is labeled as real samples.

[0008] 2) Construct generator and discriminator networks based on neural network models;

[0009] The random module generates a set of random signal data as the input to the generator network. The output of the generator network is labeled as generated samples and trained together with real samples to train the discriminator network. The specific structure of the generator network and the discriminator network is as follows: after the input layer, the generator network uses four convolutional layers as feature extractors and uses ELU as the activation function. Each convolutional layer is connected to a normalization layer and a dropout layer respectively.

[0010] The output of the convolutional layer of the discriminator network is connected to the fully connected layer. The fully connected layer uses the sigmoid function as the activation function to achieve the classification effect and to determine whether the generated samples are real or fake.

[0011] 3) The generated samples output by the generator network are passed to the reconstruction layer network, where the encoder converts the data into a low-dimensional latent space representation, and the input data is reconstructed by the decoder using the latent factors it has learned;

[0012] 4) After constructing three networks—generator network, discriminator network, and reconstruction layer network—a set of random noise signals of the same dimension as the real sample data is generated and used as the input to the generator network. The output of the generator network after training is labeled as the generated sample. The real sample and the generated sample are used together as the input to the discriminator network. The discriminator network distinguishes between real and fake generated samples and calculates the loss function L1 between generated samples and real samples based on the cross loss function. The loss function L1 is then passed back to the generator network. The generated sample is used as the input sample of the reconstructor. After the reconstruction network extracts the intrinsic feature samples, the sample reconstruction signal is obtained. The sample reconstruction signal and the real sample are processed by Kennard coefficient calculation. The resulting loss function L2 is then passed back to the generator network to supervise the generation direction and quality of the samples.

[0013] The formula for calculating the cross-loss function is as follows:

[0014]

[0015] In the formula, L represents the cross-loss function, and y represents the true label value. Represents the predicted label;

[0016] 5) Based on the reconstruction of the generative adversarial network training, high-quality generated samples are obtained. The trained generated samples are mixed with the minority class real samples according to the same fault mode to expand to a balanced sample dataset. The balanced dataset is divided into training set data and test set data. The training set data and test set data are respectively sent to the convolutional neural network module for training and testing to obtain bearing fault diagnosis results.

[0017] In a specific embodiment of the present invention, in step 1), the imbalance ratio is 50:1.

[0018] In one embodiment of the present invention, in step 3), the encoder converts the original time series data X into two parameters in the latent space: mean and standard deviation; these parameters represent the posterior distribution learned by the neural network; latent factors are obtained by using variational inference reparameterization techniques, and the decoder reconstructs new samples from the latent factors.

[0019] In another embodiment of the present invention, in step 4), the generated sample is used as the input sample of the reconstructor. After the reconstruction network extracts the intrinsic feature samples, the sample reconstruction signal is obtained. The sample reconstruction signal and the real sample are processed by Kennard coefficient calculation, and the resulting loss function L2 is back-propagated to the generator network to supervise the generation direction and quality of the samples. The specific implementation is as follows:

[0020] Design the loss function for the generator network (G), discriminator network (D), and reconstruction layer network (EN) model. The formula for the loss function of the generative adversarial network is as follows:

[0021]

[0022] In the formula, Let U(G,EN,D) represent the maximum likelihood function, G,EN,D represent the generator, encoder, and discriminator, respectively, U(G,EN,D) represent the generative adversarial network loss function, p(x,y) represent the discriminator's conditional distribution, and (x,y) represent the real sample and the generated sample, respectively. p(x,y) Let p(x,y) represent the expected value of the term, and D(x,y) represent the sample data after processing by the discriminator. p ENThe expected value of the (x,y) terms, p EN (x,y) represents the feature probability distribution of the reconstructor, p g (x,y) represents the probability distribution of generator features. Represents (x,y)-p g (x,y) and G(z,y) represent the sample data generated by the generator, and β1 and β2 are both hyperparameters;

[0023] A loss function based on the Kennard coefficient is introduced into the reconstructor to measure the difference between the generated samples and various faulty samples; the Kennard loss function formula is as follows:

[0024]

[0025] In the formula, L k x represents the Kennard loss function value. i m x represents the i-th sample of category m. m ave represents the mean of all m samples in all categories, and K represents the Kennard coefficient;

[0026] The final loss function of the reconstructed generative adversarial network, designed jointly by the generator network, discriminator network, and reconstruction layer network, is as follows:

[0027]

[0028] In the formula, U * (G,EN,D) represents the loss function of the generative adversarial network after the reconstruction layer network.

[0029] In another specific embodiment of the present invention, in step 5), the balanced dataset is divided into training set data and test set data in, for example, a 7:3 ratio.

[0030] Compared with existing technologies, the advantages of this invention are as follows: Based on the generator and discriminator networks included in GANs, a reconstruction layer network composed of autoencoders is added. After encoding and decoding the generated samples, their latent feature signals are further extracted. Based on the Kennard correlation coefficient criterion, the difference between the reconstructed signal and the real sample signal is calculated and fed back to the generator network. The introduction of the reconstruction module can accelerate the rapid convergence of the generative adversarial network and supervise the direction of the generated samples, improving the quality of the generated data and avoiding the pattern collapse problem in the training network. The dataset expanded with generated samples can significantly improve the accuracy of fault diagnosis of motor bearings. Attached Figure Description

[0031] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0032] Figure 1 is a flowchart of a motor bearing fault diagnosis method based on reconstructed generative adversarial network disclosed in this invention;

[0033] Figure 2 is a diagram of the reconstructed network structure selected in this invention;

[0034] Figure 3 is a schematic diagram of the experimental setup for the CWRU dataset selected in this invention;

[0035] Figure 4 is a schematic diagram of the reconstructed generative adversarial network structure disclosed in this invention. Detailed Implementation

[0036] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0037] As shown in Figure 1, this invention discloses a flowchart of a motor bearing fault diagnosis method based on reconstructed generative adversarial networks. The method includes the following steps:

[0038] 1) When it is necessary to perform fault detection on motor bearings under an imbalanced dataset, it is first necessary to obtain the feature datasets corresponding to the majority class healthy samples and the minority class fault samples. According to the imbalance ratio, for example, 50:1, the majority class normal sample signals and the minority class fault sample signals are collected respectively. The minority class fault data under the imbalanced dataset is labeled as real samples.

[0039] 2) Construct generator and discriminator networks based on neural network models.

[0040] The generator network generates a set of random signal data using the random module as input. The generator network's output is labeled as generated samples and used to train the discriminator network along with real samples. Specifically, the generator and discriminator networks are structured as follows: after the input layer, the generator network uses four convolutional layers as feature extractors, employing ELU as the activation function (a technique well-known to those skilled in the art). To improve the model's generalization ability and stability, each convolutional layer is followed by a normalization layer and a dropout layer.

[0041] The discriminator network has a similar structure and function to the generator network. The difference is that the output of the convolutional layer of the discriminator network is connected to the fully connected layer. The fully connected layer uses the sigmoid function as the activation function to achieve the classification effect and to determine whether the generated samples are real or fake.

[0042] 3) Figure 2 illustrates the data reconstruction network process. The generated samples output by the generator network are passed to the reconstruction layer network, where the encoder converts the data into a low-dimensional latent space representation. The input data is then reconstructed by the decoder using its learned latent factors (this technique is well-known to those skilled in the art). In one embodiment of the invention, the encoder converts the original time-series data X into two parameters in the latent space: the mean and the standard deviation. These parameters represent the posterior distribution learned by the neural network. The decoder reconstructs new samples from these latent factors using reparameterization techniques in variational inference (variational inference techniques are known to those skilled in the art).

[0043] 4) After constructing the generator network, discriminator network, and reconstruction layer network, a set of random noise signals with the same dimension as the real sample data is generated and used as the input to the generator network. The output of the generator network after training is labeled as the generated samples. Both real and generated samples are used as input to the discriminator network. The discriminator network distinguishes between real and fake generated samples and calculates the loss function L1 between generated and real samples based on the cross-loss function (formula given later). The loss function L1 is then passed back to the generator network. The generated samples are used as input to the reconstructor network. After the reconstruction network extracts intrinsic feature samples, a sample reconstruction signal is obtained. The sample reconstruction signal and the real samples are processed using Kennard coefficients (formula given later). The resulting loss function L2 is then passed back to the generator network to supervise the generation direction and quality of the samples. The cross-loss function calculation formula is as follows:

[0044]

[0045] In the formula, L represents the cross-loss function, and y represents the true label value. This represents the predicted label.

[0046] 5) Based on the reconstructed generative adversarial network training, high-quality generated samples are obtained. The trained generated samples are then mixed with minority class real samples (directly collected in step 1) according to the same fault mode to augment the balanced sample dataset (the mixing and augmentation mentioned here only requires mixing according to the described same mode; no special implementation process is required, and this technique is known to those skilled in the art). The balanced dataset is divided into training and testing sets in, for example, a 7:3 ratio. The training and testing sets are then fed into the convolutional neural network module for training and testing, respectively, to obtain bearing fault diagnosis results.

[0047] The CWRU dataset experimental platform, shown in Figure 3, uses vibration sensors to collect vibration signal data from the drive-end accelerator. The platform uses a 16-channel DAT recorder to collect vibration signals corresponding to four different states: normal, outer race damage, inner race damage, and rolling element damage. The motor load for each state ranges from 0 to 3 HP, and each fault state is set with three fault levels: 7 mils, 14 mils, and 21 mils. The motor speeds include 1797 rpm, 1772 rpm, 1750 rpm, and 1730 rpm. This example uses data corresponding to a sampling frequency of 48 kHz, a load of 1 HP, and a motor speed of 1772 rpm.

[0048] Figure 4 shows a schematic diagram of the reconstructed generative adversarial network structure disclosed in this invention. To avoid the problems of mode collapse and gradient vanishing during the training process of generative adversarial networks, this example redesigns the loss functions of the generator network (G), discriminator network (D), and reconstruction layer network (EN). The formula for the generative adversarial network loss function is as follows:

[0049]

[0050] In the formula, Let U(G,EN,D) represent the maximum likelihood function, G,EN,D represent the generator, encoder, and discriminator, respectively, U(G,EN,D) represent the generative adversarial network loss function, p(x,y) represent the discriminator's conditional distribution, and (x,y) represent the real sample and the generated sample, respectively. p(x,y) Let p(x,y) represent the expected value of the term, and D(x,y) represent the sample data after processing by the discriminator. p EN The expected value of the (x,y) terms, p EN (x,y) represents the feature probability distribution of the reconstructor, p g (x,y) represents the probability distribution of generator features. Represents (x,y)-p g (x,y) and G(z,y) represent the sample data generated by the generator, and β1 and β2 both represent hyperparameters.

[0051] A Kennard coefficient-based loss function metric is introduced into the reconstructor to measure the difference between generated samples and various fault samples, keeping in mind that the purpose of generating samples is to have characteristics comparable to fault samples of the corresponding types. The Kennard loss function formula is as follows:

[0052]

[0053] In the formula, L k x represents the Kennard loss function value. i mx represents the i-th sample of category m. m ave represents the mean of all m samples in all categories, K represents the Kennard coefficient (K specifically represents the coefficient calculated for the two samples in parentheses), and N represents the total number of samples.

[0054] The final loss function of the reconstructed generative adversarial network, designed jointly by the generator network, discriminator network, and reconstruction layer network, is as follows:

[0055]

[0056] In the formula, U * (G,EN,D) represents the loss function of the generative adversarial network after the reconstruction layer network.

[0057] In summary, the generative adversarial network framework of this invention mainly consists of three parts: a generator network, a discriminator network, and a reconstruction layer network. By redesigning the loss function of the generative adversarial network and incorporating the constraint of the Kennard loss function of the reconstruction layer network, the generative adversarial network can achieve higher speeds.

[0058] It converges, and the quality of its generated samples is also improved, making it highly versatile for expanding different types of imbalanced datasets.

Claims

1. A method for diagnosing motor bearing faults based on reconstructed generative adversarial networks, characterized in that, Specifically, the following steps are included: 1) When fault detection of motor bearings in an imbalanced dataset is required, it is necessary to obtain feature datasets corresponding to the majority class healthy samples and the minority class fault samples. According to the imbalance ratio, the signals of the majority class normal samples and the minority class fault samples are collected separately, and the minority class fault data in the imbalanced dataset is labeled as real samples; 2) A generator network and a discriminator network are constructed based on a neural network model; a set of random signal data is generated using the random module as the input to the generator network, and the output of the generator network is labeled as generated samples, which are then used to train the discriminator network together with real samples; the generator network... The discriminator network structure is as follows: After the input layer, the generator network uses four convolutional layers as feature extractors and ELU as the activation function; each convolutional layer is connected to a normalization layer and a dropout layer respectively; the output of the convolutional layer of the discriminator network is connected to a fully connected layer, which uses the sigmoid function as the activation function to achieve classification and determine the authenticity of the generated samples; 3) The generated samples output by the generator network are passed to the reconstruction layer network, where the encoder converts the data into a low-dimensional latent space representation, and the input data is reconstructed by the decoder using the latent factors it has learned; 4) After constructing the generator network, discriminator network, and reconstruction layer network, a set of random noise signals with the same dimension as the real sample data is generated and used as the input of the generator network. The output of the generator network after training is labeled as the generated sample. Both real and generated samples are used as input to the discriminator network. The discriminator network distinguishes between real and fake generated samples and calculates the loss function L1 between generated and real samples based on the cross-loss function. The loss function L1 is then passed back to the generator network. The generated samples serve as input to the reconstructor network. After the reconstructor network extracts intrinsic feature samples, it obtains the sample reconstruction signal. The sample reconstruction signal and the real samples are processed using Kennard coefficients, and the resulting loss function L2 is passed back to the generator network to supervise the generation direction and quality of the samples. The formula for calculating the cross-loss function is as follows: In the formula, This represents the cross-loss function, and y represents the true label value. Represents the predicted label; 5) Based on the reconstruction of the generative adversarial network training, high-quality generated samples are obtained. The trained generated samples are mixed with the minority class real samples according to the same fault mode to expand to a balanced sample dataset. The balanced dataset is divided into training set data and test set data. The training set data and test set data are respectively sent to the convolutional neural network module for training and testing to obtain bearing fault diagnosis results.

2. The motor bearing fault diagnosis method based on reconstructed generative adversarial networks as described in claim 1, characterized in that, In step 1), the imbalance ratio is 50:

1.

3. The motor bearing fault diagnosis method based on reconstructed generative adversarial networks as described in claim 1, characterized in that, In step 3), the encoder transforms the original time series data X into two parameters in the latent space: the mean and the standard deviation; these parameters represent the posterior distribution learned by the neural network; latent factors are obtained by using variational inference reparameterization techniques, and the decoder reconstructs new samples from the latent factors.

4. The motor bearing fault diagnosis method based on reconstructed generative adversarial networks as described in claim 1, characterized in that, In step 4), the generated samples serve as input samples for the reconstructor. After the reconstruction network extracts intrinsic feature samples, the sample reconstruction signal is obtained. The sample reconstruction signal and the real samples are processed using Kennard coefficients. The resulting loss function L2 is then passed back to the generator network to supervise the generation direction and quality of the samples. The specific implementation is as follows: The loss functions of the generator network (G), discriminator network (D), and reconstruction layer network (EN) are designed. The formula for the generative adversarial network loss function is as follows: In the formula, Represents the maximum likelihood function. These represent the generator, encoder, and discriminator, respectively. This represents the loss function of the generative adversarial network. The discriminator represents the conditional distribution. Representing real samples and generated samples, express The mathematical expectation of the term, This represents the sample data after processing by the discriminator. express mathematical expectation, Represents the probability distribution of features of the reconstructor. Represents the probability distribution of generator features. express ， This indicates that the generator produces sample data. All represent hyperparameters; a loss function based on the Kennard coefficient is introduced into the reconstructor to measure the difference between the generated samples and various faulty samples; the Kennard loss function formula is as follows: In the formula, Represents the Kennard loss function value. Representative category The One sample, Represents all categories The sample mean, Represents the Kennard coefficient. The total number of samples is represented by ; the final loss function of the reconstructed generative adversarial network, designed jointly by the generator network, discriminator network, and reconstruction layer network, is as follows: In the formula, This represents the loss function of the generative adversarial network after the reconstruction layer network.

5. The motor bearing fault diagnosis method based on reconstructed generative adversarial networks as described in claim 1, characterized in that, In step 5), the balanced dataset is divided into training data and test data in a ratio of, for example, 7:3.

Images