A rolling bearing cross-domain fault diagnosis algorithm based on a residual network

Abstract

The application discloses a rolling bearing cross-domain fault diagnosis algorithm based on a residual network, considers that vibration signal feature distributions are inconsistent under different working conditions of a rolling bearing and that sample labels to be diagnosed are difficult to obtain, adopts a multi-scale feature extraction module to extract multi-scale features of the vibration signal in a general feature extraction stage of a source domain and a target domain, adopts a residual network structure to avoid model performance degradation, considers differences between edge distributions and conditional distributions of different domains in a domain self-adaption stage, introduces a joint maximum mean difference criterion, the measurement criterion matches joint distribution distances by calculating target domain pseudo labels, thereby reducing differences between sub-class fault feature distributions under different working conditions, and realizes domain self-adaption; the algorithm can utilize diagnosis knowledge learned from the source domain to realize fault diagnosis of the target domain without labeled data, and has a good application prospect.

Description

TECHNICAL FIELD

[0001] The application relates to the fault diagnosis problem of a rotating mechanical system under different working conditions, and designs a rolling bearing cross-domain fault diagnosis algorithm based on a residual network, and belongs to the technical field of fault diagnosis. BACKGROUND

[0002] Rolling bearings are the most frequently used key components in rotating machinery, and are widely used in modern industrial production. The health status of the rolling bearings has an important influence on the smooth operation of the entire mechanical equipment. Fault diagnosis of the rolling bearings can effectively ensure production safety and life safety in industrial production, and has great engineering practical significance.

[0003] With the rapid development of artificial intelligence, deep learning is widely applied to the field of fault diagnosis due to its strong learning ability and feature self-extraction ability, and has become a research hotspot. Deep learning can automatically extract deep fault information, effectively avoiding the limitations of manual selection of fault features.

[0004] However, the fault diagnosis algorithm based on deep learning needs to be established on the basis of independent and identically distributed training set and test set. In actual industry, due to the changes of temperature, speed, load and other factors, the distribution of the training set and the test set usually has a drift phenomenon, which cannot meet the same distribution assumption. This leads to the difficulty of applying the fault diagnosis algorithm trained under the original working condition to the new working condition, greatly reducing the generalization performance of the fault diagnosis algorithm. Therefore, how to solve the problem that the change of working conditions causes the change of data distribution and reduces the applicability of the model, that is, how to realize cross-domain fault diagnosis has important practical significance.

[0005] In view of the above cross-domain fault diagnosis problem, the most direct way is to re-collect and label the training data under the new working condition, and then train the model again. However, this method needs to spend more manpower, material resources and time cost. Therefore, relevant scholars apply the transfer learning theory to the field of fault diagnosis. Transfer learning can solve related target domain problems by known source domain knowledge: learning in the source domain, and migrating the knowledge learned from the source domain to different but related target domains (in the target domain, thereby solving the new related task in the target domain, and improving the generalization ability of the classification model. SUMMARY

[0006] The application aims at the above research background, and proposes a rolling bearing cross-domain fault diagnosis algorithm based on a residual network on the basis of the prior art method.

[0007] Technical solution: In order to achieve the above-mentioned purpose of the application, the application provides a rolling bearing cross-domain fault diagnosis algorithm based on a residual network, characterized by: a multi-scale feature extraction module is used to extract multi-scale features of a vibration signal, a residual network structure is used to avoid model performance degradation, and differences in edge distribution and conditional distribution of different fields are considered, a joint maximum mean difference criterion is introduced, a target domain pseudo label is calculated to match the joint distribution distance to reduce the difference in sub-class fault feature distribution under different working conditions, and field adaptation is realized, including the following specific steps:

[0008] Step 1) On a bearing test bench, through a vibration signal data acquisition system, collect acceleration vibration signals of rolling bearings under different faults and different rotating speeds to obtain an original vibration signal data set;

[0009] Step 2) In order to make full use of data and extract more effective features, data enhancement is realized, the algorithm uses an equal-interval sliding window data enhancement method to overlap sample the original data to obtain a sample data set, each sample contains 2048 sample data points, and the sample is denoted as Data = [(x1, y1), …, (x i , y i ), …, (x n , y n )] T , Data is a single sample after segmentation, x i is a single vibration signal sample data point, and y i is a fault class label of the sample data;

[0010] Step 3) Set a rolling bearing transfer learning task, take data under a certain working condition as a source domain and data under another working condition as a target domain, wherein the labels of the source domain data are known, the labels of the target domain data are unknown, set all source domain data and 60% of the target domain data as a training set, and the remaining 40% of the target domain data as a validation set;

[0011] Step 4) Set a new rolling bearing cross-domain fault diagnosis model under Pytorch, and use the training set to train the fault diagnosis model, the fault diagnosis model uses the MSGD optimization algorithm, the batch size is set to 64, the iteration number (Epoch) is set to 100 times, and the initial value of the learning rate is set to 1e-4;

[0012] In order to reduce the training time and speed up the model convergence, in the feature extraction part, a pre-trained multi-scale feature fusion residual network model is adopted, and in the field adaptation part, JMMD measurement is adopted for field adaptation; finally, a Softmax classifier is used for bearing fault diagnosis of the unlabeled target domain, and the specific steps are as follows:

[0013] Step 4.1) The feature extraction part of the model is composed of a first layer of multi-scale feature extraction and fusion layer, three residual modules, a full connection layer and a Softmax classifier, and finally outputs the fault classification result through the Softmax classifier;

[0014] Step 4.2) In the first layer, a multi-scale feature extraction module is built through parallel learning, a three-channel structure is used for multi-scale feature extraction, different size convolution kernels are used to extract multi-scale fault feature frequencies from the signal, the convolution kernel sizes of each branch structure are 64*1, 128*1 and 256*1 respectively, by using different size convolution kernels to extract features of different scales, both the details of the shallow layer and the deep layer information are retained, and at the same time, in order to further optimize the fault data and improve the network diagnosis effect, batch normalization and activation function are added after each convolution layer; finally, the feature information of different scales is stacked and spliced together through the interlayer concatenation layer (Concat) to generate new output features;

[0015] Step 4.3) In the remaining layers, the idea of residual network is adopted, the output result of the previous layer neural network is directly input into the network of the next layer through cross-layer linking, which is composed of a series of residual blocks, and each residual module includes a convolution layer, a nonlinear activation function layer and a cross-layer linking;

[0016] Step 4.4) The above structure is combined to extract the domain-invariant features between the source domain and the target domain of the rolling bearing, and the model parameters are as follows:

[0017]

[0018] Step 4.5) The domain adaptation part of the model uses joint maximum mean difference to reduce the displacement of the source domain and the target domain in the conditional distribution and the edge distribution in different domains, the difference between the source domain and the target domain is mainly reflected in the full connection layer, the domain adaptation part uses the full connection layer to learn the transfer knowledge, adopts joint maximum mean difference measurement to calculate the distribution difference and minimizes it, realizes the distribution alignment of the source domain and the target domain, and realizes the transfer learning of the rolling bearing under different working conditions;

[0019] Step 5) Input the verification set into the trained deep residual shrinkage network model to identify the fault type of the test sample;

[0020] Beneficial effects: the rolling bearing cross-domain fault diagnosis algorithm based on the residual network provided by the application can use the source domain data with similar fault categories to classify and identify the unknown target domain data set, and realize the fault diagnosis of the target domain. And this algorithm does not need any signal processing, reduces the dependence on signal processing and target domain label data, and has the following specific advantages:

[0021] (1) The method of overlapping sampling is introduced in the application, the original vibration signal is overlapped and intercepted when the sample set is constructed, and the problem of less original vibration signal can be effectively avoided;

[0022] (2) In the fault feature extraction part, considering that the vibration signal has a multi-scale characteristic, a multi-scale feature extraction and fusion layer is used in the first layer, which accelerates the training speed of the model, and can make the model extract more comprehensive multi-scale fault features, and the depth of the model is deepened through the residual module, and the degradation problem of the deep neural network is solved;

[0023] (3) In the field adaptation part, the joint maximum mean difference is introduced to calculate the distribution difference, and the minimum is minimized to realize the field adaptation. The traditional maximum mean difference only considers the edge distribution difference between the source domain and the target domain, and the joint maximum mean difference adds the edge difference and the conditional difference to measure the difference between the domains. Therefore, the algorithm can more accurately measure the difference between different domains when performing field adaptation;

[0024] The method provided by the application is a rolling bearing cross-domain fault diagnosis algorithm based on a residual network, has certain practical application value, is easy to implement, high in accuracy, and can be widely applied to cross-domain fault diagnosis of rolling bearings. BRIEF DESCRIPTION OF DRAWINGS

[0025] Figure 1 The flowchart of the rolling bearing cross-domain fault diagnosis algorithm based on the residual network of the application is shown in the figure;

[0026] Figure 2 The basic residual module structure diagram of the residual network of the application is shown in the figure;

[0027] Figure 3 The residual network model structure diagram based on multi-scale feature fusion of the application is shown in the figure;

[0028] Figure 4 The rolling bearing cross-domain fault diagnosis model structure diagram based on the residual network of the application is shown in the figure; DETAILED DESCRIPTION

[0029] The application will be further explained in combination with the drawings.

[0030] Figure 1 The fault diagnosis flowchart of the rolling bearing cross-domain fault diagnosis algorithm based on the residual network is shown in the figure, and includes the following specific steps:

[0031] Step 1) On the bearing test bench, the acceleration vibration signal of the rolling bearing under different faults and different rotating speeds is collected through the vibration signal data acquisition system, and the original vibration signal data set is obtained;

[0032] Step 2) In order to fully utilize the data and extract more effective features, data augmentation is implemented, and the algorithm uses an equal-interval sliding window data augmentation method to overlap sample the original data to obtain a sample data set, each sample containing 2048 sample data points, and the sample is denoted as Data = [(x1, y1), …, (x i , y i ), …, (x n , y n )] T , Data is a single sample after segmentation processing, x i is a single vibration signal sample data point, and y i is the fault category label of the sample data;

[0033] Step 3) Set the transfer learning task of the rolling bearing, take the data under a certain working condition as the source domain and the data under another working condition as the target domain, wherein the labels of the source domain data are known and the labels of the target domain data are unknown, set all the source domain data and 60% of the target domain data as the training set, and the remaining 40% of the target domain data as the validation set;

[0034] Step 4) Set a new rolling bearing cross-domain fault diagnosis model under Pytorch, and use the training set to train the fault diagnosis model, the fault diagnosis model uses the MSGD optimization algorithm, the batch size is set to 64, the iteration number (Epoch) is set to 100 times, and the initial value of the learning rate is set to 1e-4;

[0035] In order to reduce the training time and speed up the convergence of the model, in the feature extraction part, a pre-trained multi-scale feature fusion residual network model is used, and in the domain adaptation part, JMMD measurement is used for domain adaptation; Finally, the Softmax classifier is used for bearing fault diagnosis of the unlabeled target domain, and the specific steps are as follows:

[0036] Step 4.1) The feature extraction part of the model is composed of the first multi-scale feature extraction and fusion layer, three residual modules, a fully connected layer, and a Softmax classifier, and finally the fault classification result is output through the Softmax classifier;

[0037] Step 4.2) In the first layer, a multi-scale feature extraction module is built by parallel learning, a parallel three-channel structure is used for multi-scale feature extraction, different sizes of convolution kernels are used to extract multi-scale fault feature frequencies from the signal, the convolution kernel sizes of each branch structure are 64x1, 128x1 and 256x1 respectively, by using different sizes of convolution kernels to extract features of different scales, both the detailed information of the shallow layer and the deep information are retained, at the same time, in order to further optimize the fault data and improve the network diagnosis effect, batch normalization and activation function are added after each convolution layer; finally, the features of different scales are stacked together by the interlayer concatenation layer (Concat) to generate new output features;

[0038] Step 4.3) In the remaining layers, the idea of residual network is adopted, the output results of the previous layer neural network are directly input into the network of the next layer through cross-layer linking, which is composed of a series of residual blocks, each residual module includes a convolution layer, a nonlinear activation function layer and a cross-layer linking;

[0039] The basic residual module structure of the residual network is shown in Figure 2 As can be seen, the residual module directly adds the input data of the module to the output data through the jump layer connection, because of the existence of the addition term, the gradient component can also be directly transmitted to the next layer when calculating the gradient, which improves the training speed of the model and avoids the performance degradation problem caused by gradient disappearance and gradient explosion;

[0040] Step 4.4) The above multi-scale feature extraction module and residual network structure are combined together to establish a residual network model based on multi-scale feature fusion, which is used to extract domain-invariant features between the source domain and the target domain of the rolling bearing, the structure of the residual network model based on multi-scale feature fusion is shown in Figure 3 In the model, multiple convolution layers with different convolution kernel sizes are used to extract more abstract features, in order to enhance the fault diagnosis ability of the model, after the second layer, the convolution kernel dimension of all convolution layers is 3x1, and the number of convolution kernels of each layer is twice the number of convolution kernels of the previous layer, the model parameters are as follows:

[0041]

[0042] Step 4.5) The domain adaptation part of the model uses joint maximum mean difference to reduce the displacement of the source domain and the target domain in the conditional distribution and the edge distribution in different domains, the difference between the source domain and the target domain is mainly reflected in the fully connected layer, the domain adaptation part uses the fully connected layer to learn the transfer knowledge, uses joint maximum mean difference measurement to calculate and minimize the distribution difference, realizes the distribution alignment of the source domain and the target domain, and realizes the transfer learning of rolling bearings under different working conditions;

[0043] Let the source domain data be... Target domain data is Where n s n is the source domain sample number. t The target domain sample is numbered, and the source domain X is... s and target domain X t The difference in distribution between them is denoted as d(X). s X t );

[0044] Marginal distribution differences between the two domains for

[0045]

[0046] In the formula, H k Represents the regenerating kernel Hilbert space, where k is the kernel. The function representing the nonlinear mapping from the original space to the regenerated Hilbert space, where k uses a Gaussian kernel function: In the formula, σ is the kernel width;

[0047] Let there be L types of category labels. Then, for any category l ∈ {1, ..., L}, the Q in label l is... s (x s |y s =l) and Q t (x t |y t Differences in conditional distribution between =l) for

[0048]

[0049] In the formula, and These represent the source and target domain data labeled 'l', respectively. and These represent the number of source domain data and the number of target domain data corresponding to the label 'l', respectively.

[0050] Then the joint maximum average difference D H (J s J t )for

[0051]

[0052] In the formula, J s and j t These are the joint probability distributions of the source domain data and the target domain data, respectively.

[0053] Step 4.6) The structure of the cross-domain fault diagnosis model for rolling bearings based on multi-scale feature fusion residual networks is as follows: Figure 4As shown, this algorithm trains a cross-domain fault diagnosis model with the goal of minimizing the loss function, where the loss function L is...

[0054] L = L C +γL J (4)

[0055] In the formula, L C L is the multi-class classification loss function between the predicted labels and the true labels of the source domain samples. J To represent the domain distribution differences learned from cross-domain datasets, γ is a tradeoff parameter greater than 0, controlling L... J To what extent do we participate in training the network?

[0056] Step 5) Input the validation set into the trained cross-domain fault diagnosis model to identify the fault type of the test sample;

[0057] The residual network-based cross-domain fault diagnosis algorithm for rolling bearings of this invention has a high fault identification accuracy. Rolling bearing tests under different working conditions and different faults are set up on a rolling bearing test bench. The experiment considers four bearing health states: normal state, inner ring fault, outer ring fault, and rolling element fault. The bearing fault size includes 0.5mm and 0.8mm. In addition, according to the change of rotation speed, fault simulation experiments with speeds of 3000rpm, 4000rpm, and 5000rpm are designed to simulate the operating conditions under different working conditions. Each working condition includes one health state and six faults.

[0058] Based on four different operating conditions at different speeds, six sets of cross-condition fault diagnosis tasks were set up for comprehensive analysis: A→B, A→C, B→A, B→C, C→A, and C→B. Each task uses a fixed operating condition as the source domain and the other different operating conditions as the target domain. The target tasks not only cover different operating conditions but also include different fault severity levels. During the migration process, the source domain data is labeled, while the target domain data is unlabeled. Labeled data in the source domain and unlabeled training data in the target domain are used for training, while the remaining unlabeled samples in the target domain are used for testing to verify the fault diagnosis algorithm. The algorithm of this invention was verified in the above-mentioned different cross-domain diagnosis tasks. The accuracy and average accuracy of the algorithm in the six sets of cross-condition fault diagnosis tasks are shown below. It can be seen that the algorithm finally achieved an average accuracy of over 97%, achieving good cross-domain fault diagnosis results.

[0059]

[0060] It can be seen that the residual network-based cross-domain fault diagnosis algorithm for rolling bearings of the present invention can effectively identify various faults of rolling bearings when the operating conditions of rolling bearings change and the target domain samples are unlabeled, and the diagnosis effect is relatively accurate.

Claims

1. A cross-domain fault diagnosis algorithm for rolling bearings based on residual networks, comprising the following specific steps: Step 1) On the bearing test bench, the rolling bearing acceleration vibration signal under different faults and different speeds is collected by the vibration signal data acquisition system to obtain the original vibration signal dataset; Step 2) To fully utilize the data, extract more effective features, and achieve data augmentation, this algorithm uses an equal-interval sliding window data augmentation method to perform overlapping sampling on the original data to obtain a sample dataset. Each sample contains 2048 sampling data points, denoted as Data = [(x1, y1), ..., (x... i y i ), ..., (x n y n )] T Data is a single sample after segmentation, x i It is a single vibration signal sampling data point, y i These are the fault category labels for the sample data; Step 3) Set up the transfer learning task for rolling bearings, taking data under one working condition as the source domain and data under another working condition as the target domain, where, The labels of the source domain data are known, while the labels of the target domain data are unknown. Let all the source domain data and 60% of the target domain data be the training set, and the remaining 40% of the target domain data be the validation set. Step 4) Set up a new cross-domain fault diagnosis model for rolling bearings in PyTorch and train the fault diagnosis model using the training set. The fault diagnosis model uses the MSGD optimization algorithm, with the batch size set to 64, the number of epochs set to 100, and the initial learning rate set to 1e-4. To reduce training time and accelerate model convergence, the algorithm employs a pre-trained multi-scale feature fusion residual network model for general feature extraction and uses the JMMD metric for domain adaptation. Finally, a Softmax classifier is used for bearing fault diagnosis in the unlabeled target domain. The feature extraction part of the model consists of a first-layer multi-scale feature extraction and fusion layer, three residual modules, a fully connected layer, and a Softmax classifier. Finally, the fault classification result is output through the Softmax classifier. Its structural parameters are as follows: The domain adaptation part of the model uses the joint maximum mean difference to reduce the displacement between the source domain and the target domain in terms of conditional distribution and marginal distribution. The difference between the source domain and the target domain is mainly reflected in the fully connected layer. The domain adaptation part uses the fully connected layer to learn transfer knowledge and uses the joint maximum mean difference metric to calculate and minimize the distribution difference, thereby realizing the transfer learning of rolling bearings under different working conditions. Step 5) Input the validation set into the trained deep residual shrinkage network model to identify the fault type of the test sample.

Images