A rolling bearing fault diagnosis method based on L-ICGAN

Abstract

The present invention discloses a rolling bearing fault diagnosis method based on L-ICGAN, which belongs to the field of fault diagnosis technology, and includes the following steps: step S1, data preprocessing, collecting vibration acceleration signals under different states, converting the one-dimensional time domain signal into a two-dimensional logarithmic spectrum diagram through logarithmic spectrum transformation, and dividing the preprocessed data into a training set and a test set; step S2, network training, initializing the L-ICGAN model through the training set, and training the L-ICGAN model; step S3, fault diagnosis, introducing the divided test set into the trained L-ICGAN model. The present invention adopts the above-mentioned rolling bearing fault diagnosis method based on L-ICGAN, which can not only effectively remove redundant information and show stronger frequency characteristics, but also increase the generation probability of minority category samples and alleviate the problem of data imbalance.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, and in particular to a rolling bearing fault diagnosis method based on L-ICGAN. Background Art

[0002] In actual engineering scenarios, rolling bearings operate in a healthy state most of the time, but the data on the normal state of the bearings monitored is often much more than the data on the fault state, resulting in data imbalance. With the rapid development of intelligent technology, data-driven fault diagnosis methods provide many effective solutions to such problems. However, such methods have certain limitations in feature learning and extraction of abnormal data.

[0003] In order to deal with this limitation, researchers have proposed many data enhancement methods, such as balancing data from the perspective of sampling, the most representative of which is oversampling the minority class or undersampling the majority class. Undersampling will lose some valuable information that may be useful for data classification. On the other hand, oversampling randomly copies training samples, resulting in overfitting problems in the classifier.

[0004] At the same time, data generation models represented by generative adversarial networks (GANs) have been deeply studied and have achieved good results in data enhancement. Although GANs provide an excellent solution to the problem of data imbalance, there is still a lot of room for research on how to improve data generation capabilities when training data is insufficient. Summary of the invention

[0005] The purpose of the present invention is to provide a rolling bearing fault diagnosis method based on L-ICGAN, which can not only effectively remove redundant information and show stronger frequency characteristics, but also improve the generation probability of minority category samples and alleviate the data imbalance problem.

[0006] To achieve the above object, the present invention provides a rolling bearing fault diagnosis method based on L-ICGAN, comprising the following steps:

[0007] Step S1, data preprocessing, collecting vibration acceleration signals under different states as signal samples, converting the one-dimensional time domain signal into a two-dimensional logarithmic spectrum diagram through logarithmic spectrum transformation, and dividing the preprocessed data into a training set and a test set;

[0008] Step S2, network training, initializing the L-ICGAN model through the training set, and training the L-ICGAN model;

[0009] Step S3, fault diagnosis, introduces the divided test set into the trained L-ICGAN model for subsequent probability calculation and decision making.

[0010] Preferably, in step S1, the window parameters are adjusted by experiments and observation of spectrum analysis results, and the best spectrum analysis results are obtained by dynamic parameter adjustment, and the optimized spectrum analysis results are input into the L-ICGAN model as mixed batch samples, wherein the window function selects the Hanning window, the window length is 256 mm, and the overlap length of adjacent windows is 250 mm.

[0011] Preferably, the L-ICGAN model in step S2 includes an improved generator G And the improved classifier C , the improved generator G The network layer consists of five layers, all of which are deconvolution layers. The first three layers use the ReLU activation function, and the last two layers use the Tanh activation function. At the same time, residual connection blocks are added to the middle three layers.

[0012] Preferably, the improved classifier C It contains five layers, four convolutional layers and a self-attention module. The output vector dimension is the same as the total number of categories. J Consistent.

[0013] Preferably, based on the CGAN output of samples with category identification, the final result of setting the unsupervised learning task is to obtain a classifier for a single real sample , the classifier gives the classification result, assuming Q is the total sample size, then the expected estimate of the conditional entropy of the input sample is defined as:

[0014] (1);

[0015] in, Represents the expected estimate of the conditional entropy of the input sample; Represents the real sample The probability distribution of j express J a single subcategory of; represents Shannon information entropy; Represents the input sample and parameters Under the condition of Belong to j Probability of class; express Z Will Classify into categories j The conditional probability of Represents a nonlinear transformation of the model conditional probability;

[0016] For batch data , the classifier must satisfy the expectation that the number of samples in each category is the same for all categories. To meet this condition, all samples must be evenly divided into each category, then it is necessary to maximize the entropy of the edited category distribution of real samples and fake samples generated by the generator:

[0017] (2).

[0018] Preferably, the loss functions of the two parts of the classification generative adversarial network CGAN are obtained from (1) and (2):

[0019] (3);

[0020] (4);

[0021] in, represents the loss function of the classifier; represents the loss function of the generator; represents the expected value of the conditional entropy of the generated data distribution; represents the maximum value of the objective function, Z is a variable; express The minimum value of .

[0022] Preferably, the training process of the L-ICGAN model in step S2 is:

[0023] S201, initialize model parameters, number of iterations, batch size, and , ;

[0024] S202, load and generate samples: obey the noise prior distribution Z~N Load from (0,1) Q Noise samples ;

[0025] S203, load real samples: load from the samples after log spectrum transformation Q Real samples ;

[0026] S204, sample input: the mixed batch samples are input into the model;

[0027] S205, update classifier: use Adam algorithm to improve the stochastic gradient of classifier and update parameters;

[0028] S206, Secondary loading to generate samples: from noise prior distribution Reload Q Noise samples ;

[0029] S207, Update Generator: Freeze the classifier parameters and use the same algorithm to reduce the generator G Stochastic gradient updates parameters;

[0030] S208. Calculate losses: Calculate losses of each part and ;

[0031] S209, model iteration: repeat steps S202-S208 until the preset number of iterations is met, and finally complete the training and output.

[0032] Therefore, the present invention adopts the above-mentioned rolling bearing fault diagnosis method based on L-ICGAN, and the technical effects are as follows:

[0033] The present invention is based on logarithmic spectrum transformation, which enables the diagnostic method proposed in the present invention to effectively extract and process non-steady-state signals and time-varying signals, and convert one-dimensional vibration signals into two-dimensional time-frequency diagrams as input of the model, which helps to remove redundant information to show stronger frequency characteristics and enhance the visualization effect of the low-amplitude part to a certain extent.

[0034] The present invention uses a classification generative adversarial network (CGAN), which enables the diagnostic method proposed in the present invention to alternately optimize the model through an adversarial process between a generator and a classifier, and generate samples close to real samples to complete data expansion.

[0035] The present invention adds an improved generator G And the improved classifier C This allows the diagnostic method proposed in this invention to better balance the importance of each category and increase the probability of generating minority category samples, thereby alleviating the problem of data imbalance.

[0036] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is an overall diagram of an embodiment of a rolling bearing fault diagnosis method based on L-ICGAN of the present invention;

[0038] Figure 2 It is an improved generator in an embodiment of a rolling bearing fault diagnosis method based on L-ICGAN of the present invention. G Detailed composition structure diagram;

[0039] Figure 3 It is an improved classifier in an embodiment of a rolling bearing fault diagnosis method based on L-ICGAN of the present invention. C Detailed composition structure diagram. DETAILED DESCRIPTION

[0040] The technical solution of the present invention is further described below through the accompanying drawings and embodiments.

[0041] Unless otherwise defined, technical or scientific terms used in the present invention shall have the common meanings understood by one having ordinary skills in the field to which the present invention belongs.

[0042] Embodiment 1

[0043] like Figure 1 As shown, the present invention provides a rolling bearing fault diagnosis method based on L-ICGAN, comprising the following steps:

[0044] Step S1, data preprocessing, collecting rolling bearing vibration acceleration signals under different states as signal samples, and converting the one-dimensional time domain signal into a two-dimensional logarithmic spectrum diagram through logarithmic spectrum transformation, and dividing the preprocessed data into a training set and a test set; wherein the window parameters are adjusted through experiments and observation of spectrum analysis results, and the best spectrum analysis results are obtained through dynamic parameter adjustment, wherein the window function selects Hanning window, the window length is 256mm, and the overlap length of adjacent windows is 250mm.

[0045] Step S2, network training, initializing the L-ICGAN model through the training set, and training the L-ICGAN model;

[0046] The L-ICGAN model includes an improved generator G And the improved classifier C ,Generator: Responsible for generating samples of the specified category that meet the requirements based on a certain distribution of random noise vectors and category labels. Classifier: Used to classify input data, and also assist the generator to generate data with the characteristics of the specified category and improve the quality and diversity of the generated samples.

[0047] Category Generative Adversarial Network (CGAN) is a variant of GAN that aims to generate samples related to a specified category by introducing category information. In CGAN, the architecture of the generator and classifier is similar to the standard GAN, but the input vector contains a certain distribution of random noise. In addition, it also contains a category label or conditional vector. This category label can be a discrete category identifier or a continuous conditional vector, which is used to specify the category or feature to which the generated sample belongs. Thanks to CGAN's ability to output samples with category identifiers, it has a good application effect in unsupervised learning.

[0048] Assume that the ultimate goal of the unsupervised learning task is to obtain a classifier for a single real sample , the classifier gives the classification result, assuming J is the total amount of the category, Q is the total sample size, then the expected estimate of the conditional entropy of the input sample is defined as:

[0049] (1);

[0050] in, Represents the expected estimate of the conditional entropy of the input sample; Represents the real sample The probability distribution of j express J a single subcategory of; represents Shannon information entropy; Represents the input sample and parameters Under the condition of Belong to j Probability of class; express Z Will Classify into categories j The conditional probability of Represents a nonlinear transformation of the model conditional probability.

[0051] The conditional entropy of real samples is low to obtain a certain category distribution. In addition, the classifier C The fake samples generated by the corresponding generator are uncertain, which means that the conditional entropy should be maximized.

[0052] For batch data , the classifier must satisfy the expectation that the number of samples in each category is the same for all categories. To meet this condition, all samples must be evenly divided into each category, then it is necessary to maximize the entropy of the edited category distribution of real samples and fake samples generated by the generator:

[0053] (2).

[0054] From (1) and (2), we can conclude that the loss functions of the two parts of the classification generative adversarial network CGAN are:

[0055] (3);

[0056] (4);

[0057] in, Representation classifier C The loss function of Representation GeneratorG The loss function of represents the expected value of the conditional entropy of the generated data distribution; represents the maximum value of the objective function, Z is a variable; express The minimum value of .

[0058] The generator network layer consists of five layers, and the input data is 200-dimensional random noise, obeying Gaussian distribution of . Output fake samples of size 64×64. Improved generator G The detailed structure of Figure 2 shown.

[0059] 3×3 in the first layer: represents the size of the convolution kernel; 200: represents the number of channels of the feature map generated by this layer, indicating that after the deconvolution operation of the first layer, the feature map has 200 channels.

[0060] 7×7 in the second layer: indicates that after deconvolution, the spatial size of the feature map is 7×7 (width and height are both 7 pixels); 64: indicates that the number of feature map channels of this layer is reduced to 64.

[0061] 15×15 in the third layer: indicates that the deconvolution operation of this layer increases the spatial resolution of the feature map to 15×15; 32: indicates that the number of channels of the feature map of this layer is 32.

[0062] 31×31 in the fourth layer: indicates that after deconvolution of this layer, the spatial size of the feature map increases to 31×31; 16: indicates that the number of channels of the feature map is further reduced to 16.

[0063] In the fifth layer, 64×64: indicates that the spatial size of the final sample generated is a two-dimensional image of 64×64; 1: indicates that the generated sample is a single channel (grayscale image).

[0064] ×2 indicates that the operation in the second part is performed twice, which makes the generator extract fault features more effectively.

[0065] The 3×3 on the left represents the size of the convolution kernel, which is used to extract fault features more finely and helps to preserve local details.

[0066] The 4×4 on the left represents the size of the convolution kernel, which is used to significantly enlarge the spatial resolution of the feature map.

[0067] Adding residual connection blocks to the middle three layers establishes a direct relationship between input and output, promotes information flow and reduces information loss. The first three layers and the last two layers use two different activation functions. The first three layers use the nonlinear characteristics of the ReLU activation function and the characteristic that the derivative is always 1, making the neural network easier to converge and faster to train; the last two layers use the Tanh activation function, which is more The output range is between [-1,1], which helps the network learn richer feature representations. In addition, batch normalization (BN) technology is used to speed up the convergence of the network and ensure the stability of model training.

[0068] The self-attention mechanism (self-attention mechanism: contains two parts of operations, namely calculating the attention distribution and weighted average in the input sequence. The self-attention model usually adopts the "query-key-value" mode to improve network performance) is introduced into the ICGAN classifier to improve the model's ability to understand global information. The classifier mainly consists of a five-layer structure, namely four convolutional layers and a self-attention module. The dimension of its output vector is consistent with the total number of categories. The LeakyReLU function is used as the activation function to suppress the vanishing gradient and ensure rapid convergence of the model. Both BN processing and Dropout processing have certain regularization effects, reduce the risk of overfitting, and improve the generalization ability of the model. In addition, Dropout processing also improves the robustness of the model and reduces the complexity of calculation by reducing the dependency between neurons. The self-attention module is used to improve the feature learning ability of the classifier, enhance the model fitting ability and ensure the stability of the network. Improved classifier C The detailed structure of Figure 3 shown.

[0069] In the first layer, 64×64×1 represents a data tensor with a width of 64, a height of 64, and 1 channel.

[0070] In the second layer, 32×32×16 is represented as a tensor with a width of 32, a height of 32, and 16 channels (feature maps).

[0071] The 16x16×32 in the third layer is represented as a feature map with a size of 16x16 pixels and 32 channels.

[0072] The 8×8×64 in the fourth layer is represented as a feature map with a size of 8x8 pixels and 64 channels.

[0073] The 4×4×128 in the fifth layer is represented as a feature map with a size of 4x4 pixels and 128 channels.

[0074] In the self-attention mechanism, 4×128 indicates that the output of the previous convolutional layer is a 4×4 feature map, and 128 indicates the position of the output channel in the previous convolutional layer.

[0075] The 5×5 on the right represents the size of the convolution kernel, which is used to extract local features within a larger range.

[0076] The goal of the generator is to generate samples that conform to the given category label, while the classifier needs to evaluate whether the input sample is real and classify it according to the category information. The generator and classifier improve the quality of generated samples and the discrimination of real samples through alternating adversarial training. The training process of ICGAN is similar to that of ordinary GAN, which is optimized by minimizing the loss function between the generator and the classifier.

[0077] The training process of the L-ICGAN model is:

[0078] S201, initialize model parameters, number of iterations, batch size, and , ;

[0079] S202, load and generate samples: obey the noise prior distribution Z~N Load from (0,1) Q Noise samples ;

[0080] S203, load real samples: load from the samples after log spectrum transformation Q Real samples ;

[0081] S204, sample input: the mixed batch samples are input into the model;

[0082] S205, update classifier C : Using Adam algorithm to improve the classifier C Stochastic gradient, update parameters; Adam algorithm: combines the characteristics of the adaptive gradient algorithm (Adagrad) and the RMSprop algorithm, can dynamically adjust the learning rate at different stages and parameters, and has efficient convergence speed and good stability.

[0083] S206, Secondary loading to generate samples: from noise prior distribution Reload Q Noise samples ;

[0084] S207. Update Generator G : Freeze the classifier C Parameters, the same algorithm is used to reduce the stochastic gradient update parameters of the generator G;

[0085] S208. Calculate losses: Calculate losses of each part and ;

[0086] S209, model iteration: repeat steps S202-S208 until the preset number of iterations is met, and finally complete the training and output.

[0087] Step S3, fault diagnosis, introduce the divided test set into the trained L-ICGAN model, and then the softmax function is used in the output layer. (The softmax function ensures the non-negativity and sum of the output value is 1 through the exponential function and normalization processing, which satisfies the basic properties of probability) Through the normalized output, the classification decision of each sample is realized, so as to perform subsequent probability calculation and decision.

[0088] The application scenario of the present invention is rolling bearing fault diagnosis under data imbalance, especially the bearing fault diagnosis problem under unbalanced monitoring sample data. Through this method, users can check whether the application objects are similar, whether the purposes are the same, and whether the core circuit adopts the same ideas. For example: rolling bearing fault diagnosis method based on generative adversarial network, rolling bearing fault detection method under data imbalance, etc.

[0089] The diagnostic method proposed in the present invention, through the logarithmic spectrum and classification generative adversarial network, can generate characteristics that meet the specified category samples according to category information, generate samples that are relatively insufficient in number in the experiment, thereby achieving the expansion of the data set.

[0090] It is worth noting that the contents not elaborated in detail in the present invention are all prior art and are well known to those skilled in the art.

[0091] Therefore, the present invention adopts the above-mentioned rolling bearing fault diagnosis method based on L-ICGAN, which can not only effectively remove redundant information and show stronger frequency characteristics, but also increase the generation probability of minority category samples and alleviate the data imbalance problem.

[0092] Finally, it should be noted that 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, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A rolling bearing fault diagnosis method based on L-ICGAN, characterized in that: The steps include: Step S1, data preprocessing, collecting rolling bearing vibration acceleration signals under different states as signal samples, and normalizing the collected signal samples through logarithmic spectrum transformation and dividing them into a training set and a test set; Step S2, network training, initializing the L-ICGAN model through the training set, and training the L-ICGAN model; Based on the CGAN output of samples with category identification, the final result of setting the unsupervised learning task is to obtain a classifier for a single real sample x , the classifier gives the classification result, Q is the total sample size, then the expected estimate of the conditional entropy of the input sample is defined as: （1）； in, Represents the expected estimate of the conditional entropy of the input sample; Represents the real sample x The probability distribution of j Indicates the total amount of the category J a single subcategory of; represents Shannon information entropy; Represents the input sample and parameters Under the condition of Belong to j Probability of class; express Z Will Classify into categories j The conditional probability of Represents a nonlinear transformation of the model conditional probability; For batch data , the classifier must satisfy the expectation that the number of samples in each category is the same for all categories. To meet this condition, all samples must be evenly divided into each category, then it is necessary to maximize the entropy of the edited category distribution of real samples and fake samples generated by the generator: （2）； Step S3: Fault diagnosis: Introduce the divided test set into the trained L-ICGAN model.

2. The rolling bearing fault diagnosis method based on L-ICGAN according to claim 1 is characterized in that: In step S1, the window parameters are adjusted by performing logarithmic spectrum transformation on the rolling bearing vibration acceleration signal under different states and observing the spectrum analysis results, and the optimized spectrum analysis results are obtained by dynamic parameter adjustment. The optimized spectrum analysis results are input into the L-ICGAN model as mixed batch samples, where the window function selects the Hanning window, the window length is 256 mm, and the overlap length of adjacent windows is 250 mm.

3. The rolling bearing fault diagnosis method based on L-ICGAN according to claim 1 is characterized in that: The L-ICGAN model in step S2 includes the improved generator G And the improved classifier C , the improved generator G The network layer consists of five layers, all of which are deconvolution layers. The first three layers use the ReLU activation function, and the last two layers use the Tanh activation function. At the same time, residual connection blocks are added to the middle three layers.

4. The rolling bearing fault diagnosis method based on L-ICGAN according to claim 3 is characterized in that: Improved classifier C It contains five layers, four convolutional layers and a self-attention module. The output vector dimension is the same as the total number of categories. J Consistent.

5. The rolling bearing fault diagnosis method based on L-ICGAN according to claim 4 is characterized in that: From (1) and (2), we can conclude that the loss functions of the two parts of the classification generative adversarial network CGAN are: （3）； （4）； in, represents the loss function of the classifier; represents the loss function of the generator; represents the expected value of the conditional entropy of the generated data distribution; represents the maximum value of the objective function, Z is a variable; express The minimum value of .

6. The rolling bearing fault diagnosis method based on L-ICGAN according to claim 5 is characterized in that: The training process of the L-ICGAN model in step S2 is: S201, initialize model parameters, number of iterations, batch size, and , ; S202, load and generate samples: obey the noise prior distribution Z~N Load from (0,1) Q Noise samples ; S203, load real samples: load from the samples after log spectrum transformation Q Real samples ; S204, sample input: the mixed batch samples are input into the model; S205, update classifier: use Adam algorithm to improve the stochastic gradient of classifier and update parameters; S206, Secondary loading to generate samples: from noise prior distribution Reload Q Noise samples ; S207, Update Generator: Freeze the classifier parameters and use the same algorithm to reduce the generator G Stochastic gradient updates parameters; S208. Calculate losses: Calculate losses of each part and ; S209, model iteration: repeat steps S202-S208 until the preset number of iterations is met, and finally complete the training and output.

Images