Cross-domain rolling bearing intelligent fault diagnosis method based on domain adaptation

Abstract

The invention provides a cross-domain rolling bearing intelligent fault diagnosis method based on domain adaptation. The method comprises the following steps: (1) carrying out fast Fourier transform on an original time sequence signal, extracting frequency domain information, normalizing the frequency domain information, and reshaping the frequency domain information into two-dimensional grayscale image data; (2) constructing an SCG-TNet network model, including initial dimensionality reduction convolution, lightweight reverse residual convolution blocks based on channel shuffling and two-layer feature enhancement convolution, and realizing global feature extraction; a graph data structure and a graph convolution module of the high-dimensional feature graph are constructed based on a K-nearest neighbor algorithm, and global feature fusion is realized; (3) designing a CGWloss function, reducing the data distribution difference between a source domain and a target domain, and realizing cross-domain feature alignment; (4) balancing classification loss and cross-domain alignment loss by adopting a dynamic loss weight adjustment strategy, and optimizing a training process; and (5) using a back propagation mechanism and an Adam optimizer to minimize a loss function value, and training the model. The method is good in performance under different working conditions and can effectively adapt to complex working conditions.

Description

Technical Field

[0001] The present invention relates to a rolling bearing fault diagnosis method, namely a cross-domain intelligent rolling bearing fault diagnosis method based on domain adaptation, which is used to train a model based on the labeled data in the source domain when there are no labels for the target domain samples, and improve the fault classification accuracy of the target domain. The focus is on the cross-domain diagnosis in the case where the feature distributions of the source domain and the target domain are different but the label distributions are the same, and it belongs to the field of fault diagnosis. Background Art

[0002] The intelligent fault diagnosis technology of industrial equipment has become an important part of intelligent manufacturing.

[0003] The vibration signals generated by rolling bearings during the operation of equipment are an important data source for fault diagnosis. However, traditional fault diagnosis methods rely on human experience for feature extraction, which not only has poor adaptability and needs to be frequently adjusted under different working conditions, but also has limited generalization ability, and it is difficult to obtain data labels under some working conditions. In recent years, convolutional neural networks have been applied in fault diagnosis and can automatically extract features, but there are still problems such as easy overfitting and insufficient adaptability to complex tasks.

[0004] To solve these problems, the present invention proposes a cross-domain intelligent rolling bearing fault diagnosis method based on domain adaptation to achieve intelligent fault diagnosis of rolling bearings. Its core lies in the network model SCG-TNet (Shuffle Convolution&Graph Transfer Network) and the loss function CGWloss (Correlation Alignment, Geometric-aware Maximum Mean Discrepancy, and Local Wasserstein Distance Loss). By combining a lightweight convolutional neural network for feature extraction and using a two-layer graph convolutional network for global feature modeling, and at the same time combining the standard classification loss with the cross-domain alignment loss to optimize the model weight matrix, bias term value, and normalization coefficient. This method can effectively adapt to the operating environment of rolling bearings across domains and improve the diagnosis accuracy. Summary of the Invention

[0005] The object of the present invention is to achieve fault diagnosis through a domain adaptation scheme, that is, when realizing cross-domain diagnosis, only using the source domain sample features, source domain sample labels, and target domain sample features, without using the target domain sample labels. Specifically, it includes data preprocessing, local feature extraction by a lightweight convolutional module, construction of a graph data structure, global feature modeling and cross-domain feature alignment based on a graph convolutional module, so as to improve the fault diagnosis accuracy of rolling bearings in different domains.

[0006] Specifically, it is manifested as obtaining the original fault data, performing fast Fourier transform and normalization operations, reshaping them into two-dimensional image samples, using a lightweight convolution module to extract fault features, and avoiding overfitting through the channel shuffle strategy. Based on the high-dimensional feature maps extracted by the convolution module, the K-nearest neighbor algorithm is used to construct a graph data structure, and two layers of graph convolution are introduced for global feature fusion. In order to enhance the cross-domain adaptation ability of the model, the present invention combines the standard cross-entropy classification loss and the cross-domain alignment loss CGWloss to optimize the network weight matrix, bias term value, and normalization coefficient, where CGWloss is composed of the correlation alignment loss CORAL (Correlation Alignment), geometric-aware maximum mean discrepancy GA-MMD (Geometric-aware Maximum Mean Discrepancy), and local Wasserstein distance (Local Wasserstein) to reduce the feature distribution difference between the source domain and the target domain. During the 300-round iterative training process, the model minimizes the classification loss and the cross-domain alignment loss, and combines the optimizer Adam to finally obtain a rolling bearing fault diagnosis model that is robust in the cross-domain case. In the verification of the public dataset used in the present invention, cross-domain fault diagnosis between tasks of 4 different working conditions, each of which contains 13 fault types, is realized. The following is the application process of the entire invention:

[0007] S1 Acquisition of original fault data

[0008] The present invention uses the publicly available bearing dataset of the University of Paderborn, Germany for verification. The detailed information and usage data of this dataset are freely available at the address https: / / groups.uni-paderborn.de / kat / BearingDataCenter / 。The bearing model used is the 6203 deep groove ball bearing. It contains 4 bearing fault signals under different working conditions, and the true labels of the signals are known. The bearing faults are real damages obtained by the test bench through accelerated life tests. In the present invention, real damage data of 13 different fault types under each of the 4 working conditions are used for verification. The 4 working conditions are as follows: (1) Working condition 1 (N15_M07_F10): The rotational speed is set at 1500 rpm, the load torque is 0.7 Nm, and the radial force is 1000 N, serving as the reference working condition; (2) Working condition 2 (N09_M07_F10): The rotational speed is reduced to 900 rpm, while the load torque remains 0.7 Nm and the radial force remains 1000 N; (3) Working condition 3 (N15_M01_F10): Based on the reference working condition, the load torque is reduced to 0.1 Nm, the rotational speed remains 1500 rpm, and the radial force remains 1000 N; (4) Working condition 4 (N15_M07_F04): Based on the reference working condition, the radial force is reduced to 400 N, the rotational speed remains 1500 rpm, and the load torque remains 0.7 Nm. The 13 different fault types include: (1) Single-point fatigue pitting on the outer ring, that is, small fatigue spalls occur at local positions on the outer ring surface due to long-term operation, forming isolated pitting; (2) Repetitive fatigue pitting on the outer ring, that is, multiple similar fatigue pittings are distributed on the outer ring surface and appear repeatedly at a certain interval; (3) Multiple fatigue pittings on the outer ring, that is, there are different types or degrees of fatigue pittings on the outer ring surface, including a combination of single-point and repetitive pittings; (4) Plastic deformation of the outer ring, that is, permanent indentations or depressions appear on the outer ring surface due to hard particle contamination or external impact; (5) Single-point fatigue pitting on the inner ring, that is, fatigue spalling occurs at a local position on the inner ring surface, forming a single pitting; (6) Repetitive fatigue pitting on the inner ring, that is, multiple fatigue pittings are arranged regularly on the inner ring surface, forming repetitive damage; (7) Multiple fatigue pittings on the inner ring, that is, a combined form of single-point pitting and repetitive pitting exists on the inner ring surface at the same time; (8) Plastic deformation of the inner ring, that is, permanent deformation is formed on the inner ring surface due to foreign matter inclusion or high-load impact; (9) Single-point fatigue pitting occurs simultaneously on the inner and outer rings, that is, single fatigue pittings appear at local positions on the inner and outer rings respectively; (10) Repetitive fatigue pitting occurs simultaneously on the inner and outer rings, that is, multiple spaced pittings appear on both the inner and outer ring surfaces; (11) Multiple fatigue pittings occur simultaneously on the inner and outer rings, that is, different forms or severities of pitting combinations appear on both the inner and outer ring surfaces; (12) Plastic deformation occurs simultaneously on the inner and outer rings, that is, both form permanent deformations due to the action of hard particles or impact loads; (13) Healthy bearing, that is, a bearing without damage.

[0009] The present invention is a general method for improving the fault diagnosis ability of the target domain by using source domain samples with known labels when the labels of the target domain samples are unknown. Although only the German-Paderborn University bearing dataset, which is more difficult in the diagnosis task of the public dataset, is used for verification here, it is also applicable to other bearings under the conditions of changing load, speed, and torque.

[0010] S2 Data Processing and Sample Construction

[0011] Aiming at the problem of the distribution difference of vibration signals under different working conditions, the fast Fourier transform is used to extract frequency domain features, and the standard normalization method is used to process the frequency domain signal data after the fast Fourier transform. The formula is as follows:

[0012]

[0013] where p represents the frequency domain signal data after the fast Fourier transform, and p * is the normalized data, where p max is 1, and p min is 0. After the normalization process, it is reshaped into a two-dimensional grayscale image format as the input of the model. In this experiment, a single sample contains 1024 consecutive sampling points in the original data. Overlapping sampling is adopted with an overlapping rate of 50%. The training set and the validation set are divided into 1:1. Finally, a grayscale image dataset with a single size of 32*32 is obtained. Each type of fault has 400 samples, and the number of samples under each working condition is 5200.

[0014] S3 Cross-Domain Rolling Bearing Fault Classification Based on SCG-TNet

[0015] SCG-TNet consists of four parts: a lightweight convolution module for feature extraction, two-layer graph convolution modules for global modeling, cross-domain loss for optimizing feature alignment, and a classification module. Its working process is as follows:

[0016] (1) Feature Extraction Module:

[0017] First, for the input grayscale image dataset in the constructed network, the source domain contains labels, and the target domain does not use labels. It is dimensionally reduced through an initial convolution layer with a 4*4 convolution kernel and a stride of 2 to extract preliminary features. Subsequently, feature extraction is sequentially performed through two reverse residual blocks. Each residual block consists of: a 1*1 convolution kernel, a dimensionality reduction convolution with a stride of 1, a 3*3 convolution kernel, a depth convolution with a stride of 2, and a 1*1 convolution kernel, an upsampling convolution with a stride of 1. In order to enhance the information interaction between different channels of the same sample and avoid overfitting, the present invention adopts a channel shuffle strategy, and its process is as follows:

[0018] X' = Reshape(X, (N, G, V / G, H, W)) (2)

[0019] Among them, X is the original sample, X' is the sample after channel shuffling, Reshape(x, y) is the reshaping function that reshapes the parameters in the input sample x according to the new specification set in y. N is the batch size, which is 128, G is the number of channel groups, which is 2, V is the channel dimension, which is 48 in the present invention, V / G is the number of remaining channels of the sample after channel division, and the core of shuffling is to randomly shuffle the order of these remaining channels. H and W are the feature map sizes, which are the height and width of the feature map respectively. After channel shuffling, two convolutional modules with 3×3 convolutional kernels and a stride of 2 are connected to achieve downsampling.

[0020] (2) Graph data structure modeling module:

[0021] Based on the extracted high-dimensional feature maps, SCG-TNet calculates the feature similarity between samples through the K-nearest neighbor algorithm and constructs an adjacency matrix. The nodes are the pixel points in each high-dimensional feature map. In the present invention, the value of K is selected as 8, and two-layer graph convolution is used for global feature fusion:

[0022] The input dimension of the first layer of graph convolutional layer is 96, and the hidden layer is 64, which is used to learn local associations.

[0023] The input dimension of the second layer of graph convolutional layer is 64, and the output dimension is 32, which enhances global feature interaction.

[0024] (3) Cross-domain alignment optimization strategy:

[0025] In order to align the data distributions of the source domain and the target domain under different working conditions, the present invention designs a loss function CGWloss for cross-domain feature alignment, which is also the core of domain adaptation, Loss CGW represents the total cross-domain alignment loss, Loss CORAL represents the related alignment loss CORAL, Loss GA-MMD represents the geometric-aware maximum mean discrepancy GA-MMD, Loss LW represents the local Wasserstein distance, and the calculation formula is as follows:

[0026] Loss CGW =αLoss CORAL +βLoss GA-MMD +γLoss LW (3)

[0027] Among them, α, β, and λ are hyperparameters that can be artificially defined to balance the relationships between various loss functions. When their values are taken as 1, 0.9, and 0.8 respectively, the optimal performance can be obtained.

[0028] The related alignment loss CORAL is used to align the covariance matrices of the source domain and the target domain features, and the loss calculation formula is:

[0029]

[0030] Among them, d is the feature dimension, which is the final input feature dimension of the classifier. In the present invention, the value is 96, which is the final input feature dimension of the classifier, ||g|| F Calculate the square root of the sum of the squares of all elements in the matrix. C S , C T are the covariance matrices of the source domain and the target domain respectively, and the calculation formula is:

[0031]

[0032] Among them, X S , X T represent the feature matrices of the source domain and the target domain respectively, n S , n T are the number of samples in the source domain and the target domain.

[0033] Secondly, the geometric-aware maximum mean discrepancy GA-MMD is an improvement based on the MMD loss, which introduces a geometric-sensitive kernel function to enhance the ability to perceive the shape of the cross-domain feature distribution. Its calculation formula is as follows:

[0034]

[0035] Among them, in Loss GA-MMD in and are the transposed vectors of the feature vectors of two different points in the source domain; and are the transposed vectors of the feature vectors of two different points in the target domain; P(x, y) is the final kernel function of GA-MMD, which is innovated based on the MMD kernel function; in MMD, the Gaussian kernel function is x and y respectively represent the sample feature vectors of the specified domain, which can be replaced by and The variable meanings in the following formula are the same as above; σ is the kernel bandwidth, and the function exp(L)=e L is used for exponential calculation, e is the base of the natural logarithm, and its value is approximately 2.718281828459. The main purpose of explaining this function is to explain the calculation process of k(x, y). α ij , β ij , γ ijis an adaptive weight used to adjust the influence degree of kernel calculation and is continuously adjusted by the loss value. Specifically, after each calculation of the total loss, the network calculates the gradients of these weights based on the measurement result of the GA-MMD on the feature distribution difference between the source domain and the target domain, and then adjusts their values during parameter update. As the training iteration progresses, these weights tend to continuously reduce the GA-MMD loss, thereby adaptively adjusting the influence degree of the kernel function when aligning the feature distributions of different domains. In addition, based on the standard MMD loss, GA-MMD adds a geometric sensitivity factor φ(x,y), which, when multiplied by the Gaussian kernel function k(x,y), enhances the perception ability of the cross-domain feature distribution shape. The formula is as follows:

[0036] φ(x,y) = exp(-λ geo ||x - y||2) (8)

[0037] where λ geo is used to amplify the influence between samples with high similarity, and its value reaches the optimal when the GA-MMD loss value is minimized. The formula is:

[0038]

[0039] Therefore, the final Gaussian kernel function P(x,y) in Loss GA-MMD is:

[0040]

[0041] where the Gaussian kernel function in MMD is x and y respectively represent the sample feature vectors of the specified domain and can be replaced in GA-MMD with and

[0042] Finally, the local Wasserstein distance is used to measure the local distribution difference between cross-domain samples and improve the local region alignment effect. The calculation is as follows:

[0043]

[0044] where M is the number of local regions, which refers to the number of regions selected during local region alignment and should be selected according to the number of fault classifications. υ(p,q) is the set of optimal transport probability distributions that satisfy the marginal constraints, which is obtained through the optimal transport theory and is used to map the samples in the source domain to the target domain to ensure the minimum cost of transport. The min() function takes the minimum value within the interval range, and π ij is the transport matrix, which represents the transport weight from the source domain samples to the target domain samples, that is, the probability of transporting from the source domain samples to the target samples in the optimal transport scheme. is the feature distance between the source domain and target domain samples.

[0045] (4) Classification module: It consists of a flattening layer, a global average pooling layer, a flattening layer, a fully connected layer, with an input channel of 96, an output channel of 13, and a Softmax activation function. The prediction probabilities of each category are calculated through the feature vectors output by the above network. The category with the largest prediction probability is the fault type predicted by the model, realizing the fault classification of rolling bearings.

[0046] S4 Training and optimization process

[0047] SCG-TNet adopts the Adam optimizer, with a batch size set to 128, an initial learning rate of 0.002, and a maximum number of iterations of 300. In addition, in order to balance the classification loss and the cross-domain loss, the present invention proposes a dynamic loss weight adjustment strategy:

[0048] λ CGW = λ max (1 - e -kp ) (12)

[0049] where λ max is the maximum loss weight, which is set to 2 in the present invention, that is, the ratio of the standard cross-entropy loss function to the custom cross-domain alignment loss function CGWloss is at least 1:2. k is the growth rate. From the start of the first iteration to the end of the last iteration, the value of k grows from zero to positive infinity. p represents the current training progress ratio, and its value is the ratio of the current iteration number to the total number of iterations. For example, if the number of training rounds in the present invention is 300 rounds, then the value of p in the first round of training is This strategy ensures that in the initial stage of training, the focus is mainly on classification performance, and cross-domain alignment is gradually strengthened as the training progresses, thereby improving the generalization ability. Finally, the loss function expression of this model is:

[0050] Loss = Loss CE + λ CGW Loss CGW (13)

[0051] Loss CE is the standard cross-entropy loss.

[0052] Compared with the prior art, the creativity of the present invention is mainly reflected in:

[0053] 1. Based on the domain adaptation strategy, integrating lightweight convolution and graph convolution, the SCG-TNet model is proposed to achieve cross-domain fault diagnosis of rolling bearings, specifically including: The lightweight convolution module is responsible for extracting local fault features and solving the overfitting problem through channel shuffling. The two-layer graph convolution module models the global relationship between samples by constructing a graph data structure, enhancing the model's adaptability to cross-domain data. Classification of target domain samples can be achieved only by using the feature of the target domain sample, without using the label of the target domain sample.

[0054] 2. The cross-domain alignment loss function CGWloss is designed: This loss function combines the relevant alignment loss CORAL, the innovatively designed geometric-aware maximum mean discrepancy GA-MMD, and the improved local Wasserstein distance to align the feature distributions of the source domain and the target domain from different perspectives. At the same time, a dynamic loss weight adjustment strategy is introduced to enable the model to adaptively focus on the classification task and the alignment task at different stages, accelerating the convergence speed. Brief Description of the Drawings

[0055] Figure 1 It is the flowchart of the method for realizing cross-domain intelligent fault diagnosis of rolling bearings by the neural network model combining lightweight convolution and graph convolution based on domain adaptation of the present invention.

[0056] Figure 2 It is the flow diagram of collecting vibration signal data.

[0057] Figure 3 It is the neural network structure diagram of the lightweight convolution combined with graph convolution of the present invention.

[0058] Figure 4 It is the flow diagram of calculating the hybrid loss function in the present invention.

[0059] Figure 5 It is the confusion matrix of the proposed method under different cross-domain tasks.

[0060] Figure 6 It is the T-SNE visualization of the proposed method under different cross-domain tasks. Detailed Embodiment

[0061] The following will be described in detail based on the drawings and the specific process.

[0062] According to Figure 1 in the flowchart of the method for realizing cross-domain intelligent fault diagnosis of rolling bearings by the neural network model combining lightweight convolution and graph convolution based on domain adaptation, its main process is: (1) Collect data of 6203 model rolling bearings; (2) Perform fast Fourier transform on the collected time series signals, and after normalization, reshape them into two-dimensional grayscale images and make labels; (3) Input the grayscale images under different working conditions into Figure 3The shown network model structure only contains source domain sample features and labels, and target domain sample features. First, initial convolution is performed through one layer to achieve preliminary feature mapping, then high-dimensional mapping is achieved through two inverse residual convolution blocks, and feature enhancement is achieved through two layers of high-dimensional convolution; (4) In the high-dimensional feature map, a graph data structure is constructed based on the K-nearest neighbor algorithm, and global information interaction is achieved through two layers of graph convolution layers. (5) Finally, the loss value is calculated according to the standard cross-entropy loss function and the designed cross-domain alignment loss, and backpropagation is performed. The Adam optimizer is used to update the model weights, bias terms, and normalization coefficient parameters to minimize the loss value. The prediction probabilities of the model for different classes of samples are output through the classification model, and the class with the maximum prediction probability is taken as the prediction class of this sample to achieve rolling bearing fault diagnosis. The cross-domain rolling bearing intelligent fault diagnosis method structure diagram based on domain adaptation lightweight convolution combined with graph convolution network is described in detail below in combination with the diagnostic method flowchart.

[0063] As Figure 2 shown, it is a simplified diagram of the process from data acquisition to storage in the computer. The test bench consists of a bearing housing and a motor. The motor provides power for the shafts of four 6203-type test bearings in the bearing housing. The test bearings rotate under the radial load applied by the spring-screw mechanism. The present invention involves a total of 4 different working conditions, and each working condition contains 13 different health state situations, which is a 13-classification. The 4 different working conditions are respectively:

[0064] (1) PU_0: Rotation speed of 1500 rpm, load torque of 0.7 Nm, radial force of 1000 N

[0065] (2) PU_1: Rotation speed of 900 rpm, load torque of 0.7 Nm, radial force of 1000 N

[0066] (3) PU_2: Rotation speed of 1500 rpm, load torque of 0.1 Nm, radial force of 1000 N

[0067] (4) PU_3: Rotation speed of 1500 rpm, load torque of 0.7 Nm, radial force of 400 N

[0068] For the data under each working condition, samples are intercepted with a sample length of 1024, and the intercepted samples are converted into grayscale image samples of 32*32. There are 5200 grayscale images under each working condition, and the corresponding sample labels are also included.

[0069] In step 2), the collected timing signal is subjected to fast Fourier transform and normalization processing, and the process is as shown in formula (1); after normalization, an image conversion program is edited using the Python language under the Pytorch framework to implement the data conversion from the format of 1024*1 to the two-dimensional grayscale image format of 32*32.

[0070] The neural network structure combining lightweight convolution and graph convolution established in step 3) is as Figure 3 shown.

[0071] SCG-TNet uses a lightweight convolution module to extract local features. First, the input grayscale image is subjected to dimensionality reduction through an initial convolution layer with a 4*4 convolution kernel and a stride of 2 to extract preliminary features, and the Swish activation function is used to enhance the non-linear expression ability.

[0072] Subsequently, the input data is sequentially passed through two reverse residual blocks for feature extraction. First, channel division is performed. In the present invention, average division is adopted, that is, the number of channels in each part is half of the number of channels of the original sample. One part is used for residual connection, and the other part is used for feature extraction and channel shuffling and enters two reverse residual blocks. Each reverse residual block consists of a 1*1 dimensionality reduction convolution for reducing the channel dimension, a 3*3 depth convolution for extracting local spatial features, and a 1*1 dimensionality increase convolution for restoring the number of channels. The stride of these convolution operations is 1 to keep the dimension of the feature map unchanged. At the same time, batch normalization is adopted after each convolution layer to stabilize the training, and the ReLU6 activation function is used. The specific calculation process is as follows:

[0073] F res = ReLU6(BN(Conv 1x1 (BN(Conv 3x3 (ReLU6(BN(Conv 1x1 (F in )))))))(9)

[0074] Among them, ReLU6 (Rectified Linear Unit 6) is the activation function, BN (Batch Normalization) is the batch normalization operation, Conv txt is the convolution operation, t is the size of the convolution kernel, 1 and 3 are selected in the present invention, and F in is the input feature of the previous layer.

[0075] In addition, the present invention introduces a channel shuffling strategy, and the calculation is shown in formula (2). First, let the number of channels of the image sample be V, and it is divided into left and right branches, and the batch size, image width, and image height remain unchanged. The number of channels in both branches of the present invention is V / 2, where the right branch is used as the residual connection, and the left branch is subjected to a series of the above convolution operations and then after randomly shuffling the channel order, channel fusion is performed with the residual connection part, and finally the number of sample channels is restored to the original number V to achieve channel shuffling.

[0076] After channel shuffling, two 3*3 convolution layers with a stride of 2 are input for downsampling. At this time, the original sample feature map is mapped to a high-dimensional feature space and becomes a high-dimensional feature map, which is convenient for the creation of subsequent graph data structures.

[0077] Based on the high-dimensional feature map, SCG-TNet constructs an adjacency matrix based on the K-nearest neighbor algorithm. In the present invention, K = 8 is selected to calculate the feature similarity of pixel points in the high-dimensional feature map, and the corresponding adjacency matrix is formed to record the direct connection relationships of each pixel point. In order to model global information, SCG-TNet adopts a two-layer graph convolutional network for global feature interaction.

[0078] Finally, the classification module adopts a flattening layer + global average pooling layer + flattening layer + fully connected layer. After the fully connected layer outputs unnormalized scores, the Softmax activation function is used to convert them into probabilities, and the category with the maximum probability is selected as the predicted category to achieve classification and obtain the experimental conclusion.

[0079] The dataset adopted in the present invention includes four different working conditions. Each type of dataset contains 5,200 samples, and the sample size is 32 * 32. The division ratio of training samples to validation samples is 1:1. Each working condition dataset contains samples of 13 different health states, and the number of samples in 13 categories is balanced. During the training process, only all the information of source domain samples and the features of target domain samples are read, the labels of target domain samples are not read, the batch size is set to 128, the optimizer Adam is selected, the initial learning rate is set to 0.002, and the maximum number of iterations is set to 300.

[0080] The classification effect of the model proposed by this method is represented by a confusion matrix, as Figure 5 shown. The abscissa of the confusion matrix is the predicted label of the model, and the ordinate is the true label of the sample. It can be obtained from the image that when working condition PU_0 is the training set and working condition PU_1 is the validation set, the accuracy rate is 52.8%; when working condition PU_0 is the training set and working condition PU_2 is the validation set, the accuracy rate is 100%; when working condition PU_0 is the training set and working condition PU_3 is the validation set, the accuracy rate is 98.3%; when working condition PU_1 is the training set and working condition PU_2 is the validation set, the accuracy rate is 75.1%; when working condition PU_1 is the training set and working condition PU_3 is the validation set, the accuracy rate is 75.7%; when working condition PU_2 is the training set and working condition PU_3 is the validation set, the accuracy rate is 98.7%;

[0081] In order to verify the classification effect of the proposed method, the results are visually displayed by T-SNE, and the results are as Figure 6As shown. It can be seen from the results that, consistent with the conclusion drawn from the confusion matrix, when performing cross-domain diagnosis of 13 different types of faults, for the three cases where the working condition PU_0 is the training set and the working condition PU_2 is the validation set; the working condition PU_0 is the training set and the working condition PU_3 is the validation set; the working condition PU_2 is the training set and the working condition PU_3 is the validation set, good clustering can be achieved. For the two cases where the working condition PU_1 is the training set and the working condition PU_2 is the validation set; the working condition PU_1 is the training set and the working condition PU_3 is the validation set, the clustering effect is average. The clustering effect for the case where the working condition PU_0 is the training set and the working condition PU_1 is the validation set needs to be improved. The above conclusions comprehensively prove the effectiveness of the invention.

Claims

1. A cross-domain rolling bearing intelligent fault diagnosis method based on domain adaptation, characterized in that: The following steps are involved: First, the time domain vibration signal of the rolling bearing is collected and fast Fourier transformed to obtain frequency domain information. After the fast Fourier transformed signal is normalized, it is reshaped into a two-dimensional grayscale image sample and labeled according to the actual fault type. It is verified using a public bearing dataset, which contains bearing fault signals under multiple different working conditions. The true label of the signal is known. The bearing fault is the real damage obtained by the test bench through accelerated life testing. Each working condition contains 13 types of real damage data of different fault types for verification. The 13 different types of faults include: (1) single-point fatigue pitting of the outer ring, i.e., tiny fatigue peeling occurs at a local position on the outer ring surface due to long-term operation, forming isolated pitting; (2) repetitive fatigue pitting of the outer ring, i.e., multiple similar fatigue pitting is distributed on the outer ring surface and recurs at a certain interval; (3) multiple fatigue pitting of the outer ring, i.e., there are fatigue pitting of different types or degrees on the outer ring surface, including a combination of single-point and repetitive pitting; (4) plastic deformation of the outer ring, i.e., permanent indentations or depressions appear on the outer ring surface due to hard particle contamination or external impact; (5) single-point fatigue pitting of the inner ring, i.e., fatigue peeling occurs at a local position on the inner ring surface, forming a single pitting; (6) repetitive fatigue pitting of the inner ring, i.e., multiple fatigue pitting is arranged in a certain pattern on the inner ring surface, forming repetitive damage; (7) Multiple fatigue pitting on the inner ring, i.e., a combination of single pitting and repetitive pitting on the inner ring surface; (8) Plastic deformation of the inner ring, i.e., permanent deformation of the inner ring surface due to foreign matter inclusion or high load impact; (9) Single fatigue pitting on both the inner and outer rings, i.e., single fatigue pitting on the local positions of the inner and outer rings respectively; (10) Repetitive fatigue pitting on both the inner and outer rings, i.e., multiple pitting arranged at intervals on the surfaces of both the inner and outer rings; (11) Multiple fatigue pitting on both the inner and outer rings, i.e., a combination of pitting of different forms or severity on the surfaces of both the inner and outer rings; (12) Plastic deformation on both the inner and outer rings, i.e., permanent deformation on both due to hard particles or impact loads; (13) Healthy bearings, i.e., bearings that have not been damaged; Secondly, a network model SCG-TNet is built, including initial dimensionality reduction convolution, a lightweight reverse residual block based on channel shuffling, and two layers of feature enhancement convolution; a graph data structure of a high-dimensional feature graph is constructed based on the K nearest neighbor algorithm, a graph convolution module, and a classification module; the lightweight reverse residual block based on channel shuffling divides the channels of the input two-dimensional graph samples, one part is used for residual connection, and the other part is extracted through a lightweight convolution module composed of three depth-separable convolution layers with different convolution kernel sizes and different step sizes, and then the channels are randomly sorted and fused with the residual part to restore the dimension; the graph convolution module consists of two layers of graph convolution layers, and the classification module consists of a fully connected layer, a global average pooling layer, a flattening layer, and an activation function; Input labeled source domain samples and unlabeled target domain samples into the network, and output the classification results of the target domain samples; The image sample is input into the network model SCG-TNet, and the classification result is output, which specifically includes the steps of: The lightweight convolution feature extraction module is used to extract features from grayscale images, and the channel shuffling strategy is combined to avoid overfitting. The graph data structure of high-dimensional feature graph is constructed through K-nearest neighbor algorithm, and two layers of graph convolution are used to achieve global feature fusion; The feature distribution of the source domain and the target domain are aligned through the loss function CGWloss, which includes the correlation alignment loss CORAL, the geometrically-aware maximum mean difference GA-MMD, and the local Wasserstein distance; After each round of iteration, the model updates the model parameters through back propagation of the loss value, including the network weight matrix, bias term and normalization layer parameters; after 300 rounds of iteration, the model obtains the optimal model by minimizing the classification loss and cross-domain alignment loss. At this time, the model is saved and verified on the target domain. The classification module identifies rolling bearing faults according to the probability of different faults corresponding to each sample in the target domain. The category with the highest fault probability is the predicted fault category of the model for the sample, and finally a robust fault diagnosis model is obtained.

2. The cross-domain rolling bearing intelligent fault diagnosis method based on domain adaptation according to claim 1 is characterized in that: The structure of the network model SCG-TNet; the model is implemented under the Pytorch framework and mainly includes a feature extraction module, a graph data structure modeling module, and a classification module; The feature extraction module uses lightweight convolution for local feature extraction. First, a grayscale image sample with a size of 32*32 is input, which contains labeled source domain samples and unlabeled target domain samples. The initial feature extraction is performed through a convolution layer with a 4*4 convolution kernel and a step size of 2. Batch normalization is used, and the Swish activation function is used to achieve preliminary feature mapping. Subsequently, two reverse residual blocks are entered, each of which consists of a 1*1 dimensionality reduction convolution for reducing the channel dimension, a 3*3 depth convolution for extracting local spatial features, and a 1*1 dimensionality increase convolution for restoring the number of channels. The step size of these convolution operations is 1 to keep the dimension of the feature map unchanged. At the same time, batch normalization is used after each convolution layer to stabilize the training, and the ReLU6 activation function is used. In addition, a channel shuffling strategy is adopted in the reverse residual block. First, the channel dimension of the image sample is set to V, and it is divided into two branches, left and right. The batch size, image width, and image height remain unchanged. The number of channels in both branches is V / 2. The right branch is used as a residual connection, and the left branch is subjected to a series of convolution operations and randomly shuffled channel order, and then channel fusion is performed with the residual connection part. Finally, the number of sample channels is restored to the original number V, realizing channel shuffling. In the graph data structure modeling module, the feature adjacency matrix is ​​constructed based on the K nearest neighbor algorithm. The similarity is determined according to the distance between the corresponding channel features of each pixel in the high-dimensional image. The closer the distance, the higher the similarity. The K nearest neighbors of the pixel in each feature map are determined according to the similarity, and the adjacency matrix is ​​constructed to form a graph structure. The K value is 8. A two-layer graph convolution network is used on the constructed graph data structure. The first layer of graph convolution is used to capture local geometric structure information, with an input dimension of 96 and a hidden layer dimension of 64. The second layer of graph convolution is used to further enhance global feature interaction, with an input dimension of 64 and an output dimension of 32. Each layer of graph convolution uses the ReLU activation function. The classification module consists of a flattening layer, a global average pooling layer, a flattening layer, and a fully connected layer. Finally, the output layer uses the Softmax activation function to calculate the probability of each fault category corresponding to the sample. The fault with the highest probability is the sample fault predicted by the model, thus achieving fault classification. In terms of model parameters, the initial number of channels of the lightweight convolutional neural network is set to 48, and the final number of extracted channels is 96; the hidden layer dimension of the graph convolutional network is set to 64, and the output dimension is set to 32; during training, the batch size is set to 128, the optimizer is Adam, the initial learning rate is set to 0.002, and the number of iterations is 300.

3. The cross-domain rolling bearing intelligent fault diagnosis method based on domain adaptation according to claim 1 is characterized in that Design of hybrid loss function; During the training process of SCG-TNet, the standard cross entropy loss function and the custom cross-domain alignment loss function CGWloss are used to optimize the classification performance and cross-domain feature alignment capabilities at the same time; the overall loss function is defined as follows: Loss=Loss CE +λ CGW Loss CGW (1) Among them, Loss CE is the standard cross entropy loss, Loss CGW is the cross-domain feature alignment loss, λ CGW is the dynamic loss weight, which is used to adjust the ratio of the two losses during training. Its definition is shown in formula (12); standard cross entropy loss Loss CE , by inputting the true label and predicted label of the sample, calculating the loss value, the model optimizes the classification accuracy by minimizing the loss; Loss CGW It is the cross-domain alignment loss CGWloss, which consists of three parts: correlation alignment loss CORAL, geometrically aware maximum mean difference GA-MMD, and local Wasserstein distance; the expression is as follows: Loss CGW =αLoss CORAL +βLoss GA-MMD +γLoss Wa (3) Among them, α, β, and λ are hyperparameters used to balance the relationship between the loss functions. When the values ​​are 1, 0.9, and 0.8, respectively, the optimal performance is obtained; First, the correlation alignment loss CORAL is used to align the covariance matrix of the source domain and target domain features, which is defined as follows: Among them, d is the feature dimension, is the final input feature dimension of the classifier, and ||g|| F Calculate the square root of all elements in the matrix; C S , C T are the covariance matrices of the source domain and the target domain respectively, and the calculation formula is: Among them, X S ,X T Represent the feature matrices of the source domain and the target domain respectively, n S ,n T is the number of samples in the source domain and the target domain; Secondly, the geometrically perceived maximum mean difference GA-MMD is based on the improvement of MMD loss and introduces a geometrically sensitive kernel function to enhance the perception of the cross-domain feature distribution shape; its calculation formula is as follows: Among them, in Los GA-MMD middle and is the transposed vector of the eigenvectors of two different points in the source domain; and is the transposed vector of the feature vectors of two different points in the target domain; P(x,y) is the final kernel function of GA-MMD, which is based on the innovation of MMD kernel function; in MMD, the Gaussian kernel function is x and y represent the sample feature vectors of the specified domain, which can be replaced by and The variables in the following formula have the same meaning as above; σ is the kernel bandwidth, and the function exp(L) = e L Used for exponential calculations, e is the base of the natural logarithm. The explanation of this function is mainly to explain the calculation process of k(x,y); α ij , β ij , γ ij To be specific, each time after calculating the total loss, the network will calculate the gradient of these weights based on the GA-MMD measurement of the difference in feature distribution between the source domain and the target domain, and then adjust their values ​​when updating the parameters; as the training iterations proceed, these weights will tend to reduce the GA-MMD loss; in addition, on the basis of the standard MMD loss, GA-MMD adds a geometric sensitivity factor φ(x,y), which is multiplied by the Gaussian kernel function k(x,y) to enhance the perception of the shape of cross-domain feature distribution. The formula is as follows: φ(x,y)=exp(-λ geo ||xy||2) (8) Among them, λ geo It is used to amplify the influence between pairs of samples with high similarity. Its value is optimal when the GA-MMD loss value is minimum. The formula is: Therefore, Loss GA-MMD The final Gaussian kernel function P(x,y) is: Among them, the Gaussian kernel function in MMD is x and y represent the sample feature vectors of the specified domain, which can be replaced by and Finally, the local Wasserstein distance is used to measure the local distribution differences between cross-domain samples and improve the local region alignment effect. It is calculated as follows: Among them, M is the number of local regions, which refers to the number of regions selected when aligning local regions, and should be selected according to the number of fault categories; υ(p,q) is the optimal transmission probability distribution set that satisfies the edge constraint, which is obtained through the optimal transmission theory and is used to map samples in the source domain to the target domain to ensure the minimum transmission cost; the min() function takes the minimum value within the interval range, π ij is the transmission matrix, which represents the transmission weight from the source domain sample to the target domain sample, that is, the probability of transmission from the source domain sample to the target sample in the optimal transmission scheme; is the feature distance between source domain and target domain samples; During the training process, SCG-TNet uses dynamic loss weights to adjust the cross-domain loss weights to ensure that the classification task dominates in the early stage of training, and gradually enhances the cross-domain alignment capability in the later stage; its calculation formula is: l CGW =λ max (1-e -kp ) (12) Among them, λ max is the maximum loss weight, which is set to 2, that is, the minimum ratio of the standard cross entropy loss function to the custom cross-domain alignment loss function CGWloss is 1:2; k is the growth rate, from the first round of iteration to the end of the last round of iteration, the k value increases from zero to positive infinity; p represents the current training progress ratio, and its value is the ratio of the current iteration round to the total number of rounds.

Images