A lifelong learning method for incremental fault diagnosis of bearings based on generative feature replay

By using the method of generating feature replay, a multi-stage bearing fault diagnosis model is constructed, and the feature distillation and replay alignment loss functions are used to prevent model forgetting, which solves the catastrophic forgetting problems caused by changing working conditions and complex fault types in bearing fault diagnosis. , achieving high-precision and robust fault diagnosis effect.

CN116108346BActive Publication Date: 2024-11-22SUZHOU UNIV
View PDF 6 Cites 0 Cited by

Patent Information

Application Number
CN202310126044.8
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-02-17
Publication Date
2024-11-22
Estimated Expiration
2043-02-17

AI Technical Summary

Technical Problem

The catastrophic forgetting problems caused by changing operating conditions and complex and variable fault types in the actual operation process of the prior art have led to poor diagnostic effects of deep learning models in bearing fault diagnosis.

Method used

Adopting the lifelong learning method of incremental bearing fault diagnosis based on generated feature playback, by collecting bearing vibration acceleration time domain signals of different fault types, performing preprocessing and feature extraction, building a multi-stage fault diagnosis model, and using feature distillation and replay Alignment loss functions prevent model forgetting, update classifiers and generators, ensuring similarity between features and generated features.

Benefits of technology

It realizes high-precision and robust bearing fault diagnosis, overcomes catastrophic forgetting problems, and is suitable for fault diagnosis of complex working conditions and variable fault types.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN116108346B_ABST
    Figure CN116108346B_ABST
Patent Text Reader

Abstract

The present invention relates to a lifelong learning method for bearing incremental fault diagnosis based on generated feature replay, comprising: dividing a bearing state data set into a plurality of different diagnosis stages; learning grayscale image samples in an initial stage, training a first feature extractor and a first classifier; using the features extracted by the first feature extractor, alternately training using an adversarial generative network, and obtaining a first feature generator; constructing an original fault diagnosis model, setting the number of neurons in its fully connected layer to the number of fault types in the initial stage; in the incremental learning stage, using the original fault diagnosis model in the n-1 stage, training and updating the fault diagnosis model in the n-stage, using a feature distillation loss function to reduce the difference between the features extracted by the n-th feature extractor and the n-1-th feature extractor, using a replay alignment loss to constrain the features generated by the n-th feature generator and the n-1-th feature generator to be similar, updating the number of neurons in the fully connected layer to the sum of the number of fault types from the initial stage to the n-th stage, and obtaining a final fault diagnosis model.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of mechanical fault diagnosis and machine learning, and in particular to a lifelong learning method and system for bearing incremental fault diagnosis based on generated feature replay. Background Art

[0002] Driven by the technological level, modern industry is developing in the direction of informatization and intelligence. As an important carrier of modern industry, the failure of industrial equipment will not only cause huge economic losses, but also bring huge safety risks. Bearings are an indispensable part of industrial equipment and need to work for a long time in harsh environments such as high temperature and high pressure. The healthy and stable operation of bearings under harsh conditions is very important for industrial production. Therefore, it is very important to monitor the health status of bearings.

[0003] Traditional machine learning often requires a lot of expertise and complex feature engineering, which means manual processing by professionals. First, perform in-depth exploratory data analysis on the data set, then use PCn or other methods to reduce the dimension of the data, and finally extract features. The most troublesome step is the manual step. There is not much overlap in expertise in different fields. When studying problems in different fields, it often requires multiple professionals in the field to perform manual feature extraction, which increases the research cost.

[0004] The development of deep learning has made it possible to automatically detect and identify faults in machine parts using artificial intelligence. For example, convolutional neural networks (CNNs), autoencoders (NEs), recurrent neural networks (RNNs), deep belief networks (DBNs), and multilayer perceptrons have been introduced into the field of fault diagnosis. However, these diagnostic models need to be based on two assumptions: the training data and the test data are distributed in the same way and the training data is sufficient; that is, sufficient fault information under different working conditions should be collected for each fault type to ensure high-precision diagnosis. However, due to the variable working conditions of the machine during actual operation and the complex and variable fault types, the two assumptions are difficult to achieve under actual working conditions, resulting in poor diagnostic results for the model. This problem is called catastrophic forgetting. Solving catastrophic forgetting and establishing a stable, reliable, and sustainable intelligent diagnostic model are important issues in the field of deep learning. Summary of the invention

[0005] Therefore, the technical problem to be solved by the present invention is to overcome the catastrophic forgetting problem caused by the variable working conditions and complex and variable fault types in the actual operation process of the prior art.

[0006] In order to solve the above technical problems, the present invention provides a lifelong learning method for bearing incremental fault diagnosis based on generated feature replay, comprising:

[0007] S1: Collect the time domain signals of bearing vibration acceleration of different fault types, perform preprocessing, obtain grayscale image samples, add fault type labels, build a bearing status data set, and divide it into multiple data sets of different diagnosis stages;

[0008] S2: Construct an original fault diagnosis model, set the number of neurons in the fully connected layer of its classifier to the number of fault types in the initial stage; train a first feature extractor and a first classifier based on grayscale image samples of the fault types in the initial stage; use the features extracted by the first feature extractor to alternately train the discriminator and the generator to obtain a first feature generator;

[0009] S3: In the nth stage, the nth feature extractor and the nth classifier are initialized through the n-1th feature extractor and the n-1th classifier in the n-1th stage; the nth feature extractor is updated by training with the grayscale image samples of the fault type in the nth stage, and the feature distillation loss function is used to reduce the difference between the features extracted by the nth feature extractor and the n-1th feature extractor;

[0010] Wherein, n = 1, 2, 3, ..., N; when n is 1, it is the initial stage; N is the number of diagnostic stages;

[0011] S4: Use the features extracted by the nth feature extractor and the features generated by the n-1th feature generator to jointly train and update the nth classifier, and use the classification loss function to prevent the nth classifier from forgetting; update the number of neurons in the fully connected layer of the nth classifier to the sum of the number of fault types from the initial stage to the nth stage;

[0012] S5: Initialize the nth feature generator of the nth stage through the n-1th feature generator, and then jointly train the features generated by the n-1th feature generator and the features extracted by the nth feature extractor to obtain the nth feature generator; at the same time, use the replay alignment loss function to constrain the features generated by the nth feature generator to be similar to the features generated by the n-1th feature generator to prevent the nth feature generator from forgetting;

[0013] S6: Repeat steps S3 to S5 until all diagnosis stages are trained to obtain a final fault diagnosis model when n=N;

[0014] S7: Use the final fault diagnosis model to diagnose the grayscale image samples in the untrained test set of all learned tasks to obtain the bearing fault diagnosis results.

[0015] In one embodiment of the present invention, the step S1 specifically includes:

[0016] Use acceleration sensors to collect bearing vibration acceleration time domain signals of different fault types, add fault type labels, build bearing status data sets, and divide fault diagnosis tasks into different stages;

[0017] Divide N+1 fault diagnosis tasks into N+1 learning stages; the number of learning stages increases with the number of diagnosis tasks. The datasets D0, D1, K, D n are learned sequentially; the data set in stage n Be learned, D n The i-th sample data, is the fault type label corresponding to the i-th sample data, K n is the number of samples in the dataset at stage n; P n Represents all fault types T that have been learned 0:n-1 ={T0,T1,K,T n-1}, Q n Indicates the fault type T learned in the current stage n The number of n+1 =P n +Q n , so Expressed as

[0018] In one embodiment of the present invention, the preprocessing of the bearing vibration acceleration time domain signal to obtain a grayscale image sample comprises:

[0019] Performing a fast Fourier transform on the bearing vibration acceleration time domain signal to transform the bearing vibration acceleration time domain signal into a bearing vibration acceleration frequency domain signal while retaining a shape of 1×1024;

[0020] The bearing vibration acceleration frequency domain signal is reshaped from 1×1024 to 1×32×32 and reconstructed in the new dimension to obtain a 1×32×32 grayscale image sample.

[0021] In one embodiment of the present invention, the training of the first feature extractor and the first classifier based on the grayscale image samples of the fault type in the initial stage includes:

[0022] The grayscale image samples in the initial stage of Resent-18 network learning are used to train the first feature extractor F0 and the first classifier C0, and the original fault diagnosis model M0 is obtained. The number of neurons in its fully connected layer is set to the number of fault types in the initial stage.

[0023] The classification loss in the initial stage is the cross entropy loss function:

[0024]

[0025] in, represents the bearing vibration acceleration frequency domain signal in the initial stage, γ represents the true label, is the classification prediction result of the original fault diagnosis model M0, K n is the number of grayscale image samples in the nth stage, and Q0 represents the number of fault types T0 learned in the initial stage.

[0026] In one embodiment of the present invention, the first feature extracted by the first feature extractor is alternately trained using a generative adversarial network to obtain a first feature generator, including:

[0027] Using the first feature extracted by the first feature extractor F0, Wasserstein GAN is used to alternately update the discriminator D0 and the generator to obtain the first feature generator G0:

[0028]

[0029]

[0030] The first feature generator is adapted to feature generation and continuous learning through the discriminator loss function and the generator loss function:

[0031] Discriminator loss function

[0032] Generator loss function

[0033] in, represents the data set of the nth stage, t is the required fault type, z is the random noise obeying the Gaussian distribution, represents the replay alignment loss function.

[0034] In one embodiment of the present invention, the replay alignment loss function It is expressed as:

[0035]

[0036] Among them, G n (t,z) represents the features generated by the nth feature generator, G n-1 (t,z) represents the features generated by the n-1th feature generator. When conditioned on a given prior fault type t and a given latent vector z, the generated features of the nth stage are constrained to be similar to those of the n-1th stage.

[0037] In one embodiment of the present invention, the feature distillation loss function is:

[0038]

[0039] Among them, F n-1(x) represents the feature extracted by the n-1th feature extractor, F n-1 (x) represents the feature extracted by the nth feature extractor.

[0040] In one embodiment of the present invention, the classification loss in the incremental learning stage includes the new classification task loss and the old classification task loss:

[0041]

[0042] The new classification task loss is the cross entropy loss of the fault type learned in the nth stage:

[0043]

[0044] The old classification task loss is the cross entropy loss of the fault types learned in all stages before the nth stage:

[0045]

[0046] Among them, P n Represents all fault types T that have been learned 0:n-1 ={T0,T1,K,T n-1}, Q n Indicates the fault type T learned in stage n n The number of n+1 =P n +Q n , γ represents the true label, is the classification prediction result of the original fault diagnosis model M0.

[0047] In one embodiment of the present invention, after the grayscale image samples of all learned tasks are diagnosed using the final fault diagnosis model to obtain the bearing fault diagnosis result, the method further includes:

[0048] All fault types in the bearing condition dataset are tested using the diagnostic tasks of all learned fault types to verify that the final fault diagnosis model has the ability to overcome catastrophic forgetting.

[0049] The embodiment of the present invention further provides a lifelong learning system for bearing incremental fault diagnosis based on generated feature replay, comprising:

[0050] The data acquisition and preprocessing module collects the bearing vibration acceleration time domain signals of different fault types, constructs the bearing status data set, and divides it into multiple bearing fault diagnosis tasks at different diagnosis stages; preprocesses the bearing vibration acceleration time domain signals to obtain grayscale image samples;

[0051] The initial stage training module is used to train the first feature extractor and the first classifier by learning the grayscale image samples in the initial stage; the first feature extracted by the first feature extractor is used to perform alternating training using the adversarial generative network to obtain the first feature generator; the original fault diagnosis model is constructed, and the number of neurons in its fully connected layer is set to the number of fault types in the initial stage;

[0052] The incremental phase training module trains the nth feature extractor, nth classifier and nth feature generator of the nth phase according to the n-1th feature extractor, n-1th classifier and n-1th feature generator of the n-1th phase; uses the feature distillation loss function to reduce the difference between the nth feature and the n-1th feature, uses the replay alignment loss function to constrain the nth generated feature generated by the nth feature generator to be similar to the n-1th generated feature, updates the number of neurons in the fully connected layer of the classifier to the sum of the number of fault types from the initial phase to the nth phase, and obtains the final fault diagnosis model;

[0053] The fault diagnosis module uses the final fault diagnosis model to diagnose the grayscale image samples in the untrained test set of all learned tasks to obtain the bearing fault diagnosis results.

[0054] The above technical solution of the present invention has the following advantages compared with the prior art:

[0055] The lifelong learning method for bearing incremental fault diagnosis based on generated feature replay described in the present invention trains a feature extractor, a classifier and a feature generator through learning grayscale image samples in the initial stage to construct an original fault diagnosis model; in the incremental learning stage, a feature extractor is used to extract features of the fault category of the nth stage, and a feature distillation loss function is used to reduce the difference between the nth feature and the n-1th feature to prevent forgetting of the feature extractor; a classifier is used to predict the target label, and a classification loss function is used to prevent forgetting of the classifier; a feature generator is used to generate an accurate old class feature distribution to facilitate sharing of parameters between different classes and tasks, and a replay alignment loss function is used to ensure that the feature generator correctly generates features from the old type by making the generated features of the i-th stage similar to the generated features of the n-1th stage, so that the final fault diagnosis model provided by the present invention solves the problem of catastrophic forgetting, has high diagnostic accuracy, strong robustness, and is suitable for fault diagnosis of incremental fault types. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to make the content of the present invention more clearly understood, the present invention is further described in detail below according to specific embodiments of the present invention in conjunction with the accompanying drawings, wherein

[0057] Figure 1 It is a flowchart of the steps of a lifelong learning method for bearing incremental fault diagnosis based on feature replay provided by an embodiment of the present invention;

[0058] Figure 2 It is a principle block diagram of a lifelong learning method for bearing incremental fault diagnosis based on feature replay provided by an embodiment of the present invention;

[0059] Figure 3 is a schematic diagram of a rolling bearing data generation test bench provided in an embodiment of the present invention;

[0060] Figure 4 is a schematic diagram comparing the fault diagnosis accuracy of the fault diagnosis model provided by the embodiment of the present invention at different stages with other methods;

[0061] Figure 5 is a schematic diagram of feature visualization at different stages provided by an embodiment of the present invention;

[0062] Figure 6 It is a visual schematic diagram of real features and generated features provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0063] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it, but the embodiments are not intended to limit the present invention.

[0064] Reference Figure 1 As shown, the steps of the lifelong learning method for bearing incremental fault diagnosis based on feature replay according to the present invention include:

[0065] S1: Collect the time domain signals of bearing vibration acceleration of different fault types, perform preprocessing, obtain grayscale image samples, add fault type labels, build a bearing status data set, and divide it into multiple data sets of different diagnosis stages;

[0066] Use acceleration sensors to collect bearing vibration acceleration time domain signals of different fault types, build bearing status data sets, and divide fault diagnosis tasks into different stages;

[0067] Divide N+1 fault diagnosis tasks into N+1 learning stages; the number of learning stages increases with the number of diagnosis tasks. The datasets D0, D1, K, D n are learned sequentially; the data set in stage n Be learned, D n The i-th sample data, is the fault type label corresponding to the i-th sample data, K n is the number of samples in the dataset at stage n; P n Represents all fault types T that have been learned 0:n-1 ={T0,T1,K,Tn-1}, Q n Indicates the fault type T learned in the current stage n The number of n+1 =P n +Q n ,so It can also be expressed as

[0068] Perform fast Fourier transform (FFT) on the bearing vibration acceleration time domain signal to transform the bearing vibration acceleration time domain signal from a time domain signal to a frequency domain signal, while retaining the shape of 1×1024 to help the feature learner extract features more easily;

[0069] The bearing vibration acceleration frequency domain signal is reshaped from 1×1024 to 1×32×32. The 1024 points on each channel are divided into 32 parts, each with 32 points, and reconstructed in the new dimension to obtain a grayscale image sample with an ideal shape of 1×32×32.

[0070] S2: Reference Figure 2 As shown in FIG. 1 , an original fault diagnosis model is constructed, and the number of neurons in the fully connected layer of the classifier is set to the number of fault types in the initial stage; the first feature extractor and the first classifier are trained according to the grayscale image samples of the fault types in the initial stage; the discriminator and the generator are alternately trained using the features extracted by the first feature extractor to obtain the first feature generator

[0071] Using Resnet-18 to learn the bearing vibration acceleration frequency domain signal in the initial stage The original fault diagnosis model M0 including the first feature extractor F0 and the first classifier C0 is obtained by training, and is used to identify the fault type T0, and the number of neurons in its fully connected layer is set to the number of fault types in the initial stage;

[0072] The classification loss in the initial stage is the cross entropy loss: Where γ is the true label, is the classification prediction result of model M0;

[0073] The first feature extractor F0 and the first classifier C0 are frozen after training, and the first feature u0 extracted by the first feature extractor F0 is used to train the first feature generator G0. WassersteinGAN is used and adapted to feature generation and continuous learning through the following loss:

[0074] Initial stage discriminator loss:

[0075]

[0076] Generator loss in the initial stage;

[0077]

[0078] Where t is the desired fault type and z is random noise that follows a Gaussian distribution.

[0079] S3: In the nth stage, the nth feature extractor and the nth classifier are initialized through the n-1th feature extractor and the n-1th classifier in the n-1th stage; the nth feature extractor is updated by training with the grayscale image samples of the fault type in the nth stage, and the feature distillation loss function is used to reduce the difference between the features extracted by the nth feature extractor and the n-1th feature extractor;

[0080] Wherein, n = 1, 2, 3, ..., N; when n is 1, it is the initial stage; N is the number of diagnostic stages;

[0081] Feature distillation loss function:

[0082]

[0083] Among them, F n-1 (x) represents the feature extracted by the n-1th feature extractor, F n-1 (x) represents the feature extracted by the nth feature extractor.

[0084] S4: Use the features extracted by the nth feature extractor and the features generated by the n-1th feature generator to jointly train and update the nth classifier, and use the classification loss function to prevent the nth classifier from forgetting; update the number of neurons in the fully connected layer of the nth classifier to the sum of the number of fault types from the initial stage to the nth stage;

[0085] Alternating updates of the discriminator D in the adversarial generative network n and the generator G n , the calculation formula is as follows:

[0086]

[0087]

[0088] is the loss of the discriminator in Wasserstein GAN, is the loss of the generator in Wasserstein GAN. is the replay alignment loss, which constrains the features generated in the current stage to be similar to the features generated in the past stage when conditioned on a given prior fault type t and a given latent vector z; It can also be considered as a distillation, calculated as follows:

[0089]

[0090]

[0091]

[0092] G n (t,z) represents the features generated by the nth feature generator;

[0093] S5: Initialize the nth feature generator of the nth stage through the n-1th feature generator, and then jointly train the features generated by the n-1th feature generator and the features extracted by the nth feature extractor to obtain the nth feature generator; at the same time, use the replay alignment loss function to constrain the features generated by the nth feature generator to be similar to the features generated by the n-1th feature generator to prevent the nth feature generator from forgetting;

[0094] The classification loss in the incremental learning stage includes the new classification task loss and the old classification task loss:

[0095]

[0096] The new classification task loss is the cross entropy loss of the fault type learned in the nth stage:

[0097]

[0098] The old classification task loss is the cross entropy loss of the fault types learned in all stages before the nth stage:

[0099]

[0100] When the nth feature extractor and nth classifier of the nth stage are trained, they are frozen and the training of the nth feature extractor is started. The nth feature generator of the nth stage is initialized by the n-1th feature generator of the n-1th stage;

[0101] The nth feature extracted by the nth feature extractor in the nth stage and the n-1th feature generated by the n-1th feature generator trained in the n-1th stage are used to jointly train the nth feature generator in the nth stage for generating features for joint training in the next stage. Since unexpected fault types are rare and complex in fault diagnosis, GnN is applied to generate increasingly accurate feature distributions to promote parameter sharing between different categories and tasks;

[0102] S6: Repeat steps S3 to S5 until all diagnosis stages are trained to obtain a final fault diagnosis model when n=N;

[0103] S7: Use the final fault diagnosis model to diagnose the grayscale image samples in the test set of all learned tasks to obtain the bearing fault diagnosis results.

[0104] Use the bearing vibration acceleration time domain signals of all learned tasks and the grayscale image samples generated after preprocessing to test the performance of the bearing fault model in new and old tasks and verify the model's ability to not forget learning. When the model training is completed, it must be able to complete the diagnosis tasks of all learned fault types and test all fault types T contained in the bearing status dataset x. 0:n , used to demonstrate that the model has the ability to overcome catastrophic forgetting.

[0105] Based on the above embodiment, the embodiment of the present invention further provides a lifelong learning system for bearing incremental fault diagnosis based on generated feature replay, comprising:

[0106] The data acquisition and preprocessing module collects the bearing vibration acceleration time domain signals of different fault types, constructs the bearing status data set, and divides it into multiple bearing fault diagnosis tasks at different diagnosis stages; preprocesses the bearing vibration acceleration time domain signals to obtain grayscale image samples;

[0107] The initial stage training module is used to train the first feature extractor and the first classifier by learning the grayscale image samples in the initial stage; the first feature extracted by the first feature extractor is used to perform alternating training using the adversarial generative network to obtain the first feature generator; the original fault diagnosis model is constructed, and the number of neurons in its fully connected layer is set to the number of fault types in the initial stage;

[0108] The incremental phase training module trains the nth feature extractor, nth classifier and nth feature generator of the nth phase according to the n-1th feature extractor, n-1th classifier and n-1th feature generator of the n-1th phase; uses the feature distillation loss function to reduce the difference between the nth feature and the n-1th feature, uses the replay alignment loss function to constrain the nth generated feature generated by the nth feature generator to be similar to the n-1th generated feature, updates the number of neurons in the fully connected layer of the classifier to the sum of the number of fault types from the initial phase to the nth phase, and obtains the final fault diagnosis model;

[0109] The fault diagnosis module uses the final fault diagnosis model to diagnose the grayscale image samples in the untrained test set of all learned tasks to obtain the bearing fault diagnosis results.

[0110] Specifically, in one embodiment of the present invention, the bearing vibration data used is collected from the fault diagnosis laboratory of Soochow University. Figure 3As shown in the figure, the test system includes a motor with a rated speed of 1450RMP and a rated frequency of 50Hz, a plum coupling and a spindle with a normal bearing, and a loading device connected to the test bearing. In addition, the test system also has an acceleration sensor model PCBM621B40 with a sensitivity of 10mv / g, a signal acquisition instrument model INV3018C and a computer.

[0111] A bearing fault data set containing sixteen health states was constructed using the signals collected by the above-mentioned test bench; the 16 health states include 0.2mm inner ring fault, 0.2mm roller fault, 0.2mm outer ring fault, 0.3mm inner ring fault, 0.3mm roller fault, 0.3mm outer ring fault, 0.4mm inner ring fault, 0.4mm roller fault, 0.4mm outer ring fault, 0.5mm inner ring fault, 0.5mm roller fault, 0.5mm outer ring fault, 0.6mm inner ring fault, 0.6mm roller fault, 0.6mm fault and normal bearing.

[0112] When collecting actual data, the speed was set to 896.1r / min and the sampling frequency was set to 10.6KHz. By adjusting the nut, the load size became 2KN.

[0113] Because the fault information collected for the first time when the bearing is used is comprehensive, there are few unexpected fault types that occur later. Therefore, as shown in Table 1, we design the number of fault types in the initial stage, that is, stage 0, to be larger, while the number of fault types in other tasks is smaller to simulate the actual use of our bearings. Each fault type contains 100 training samples and 100 test samples, and each sample contains 1024 data points.

[0114] Table 1 Bearing health status and task settings

[0115]

[0116]

[0117] Referring to Table 2, the structural parameters of the Resnet-18 network are:

[0118] Table 2. ResNet-18 backbone network structure parameters

[0119]

[0120] Specifically, based on the above description, the 16 collected bearing health states are divided into 4 diagnosis stages, and the first feature extractor and the first classifier are trained according to the grayscale image samples in the initial stage 0. The first feature generator is trained based on the first feature using the adversarial generative network; the original fault diagnosis model consisting of the first feature extractor, the first classifier and the first feature generator is obtained; the incremental stage of learning is continued, and the second feature extractor, the first classifier and the first feature generator of the initial stage 0 are used to train the second feature extractor, the second classifier and the second feature generator of the second stage 1 until the fourth feature extractor, the fourth classifier and the fourth feature generator are trained to obtain the final bearing fault diagnosis model; the bearing vibration acceleration time domain signals of all learned tasks are input into the final bearing fault diagnosis model to obtain the bearing fault diagnosis result.

[0121] Reference Figure 4 As shown, it is a schematic diagram comparing the fault diagnosis accuracy of the fault diagnosis model of the present invention at different stages with five existing methods. Comparison method 1 is Fine-tuning of nll network pnrnmeters, which is a method for fine-tuning the entire network in transfer learning; comparison method 2 is EWC, which is a classic regularization-based method in lifelong learning; comparison method 3 is LWF, which introduces knowledge distillation into the field of lifelong learning for the first time; comparison method 4 is iCnRL, which introduces samples into lifelong learning for the first time; comparison method 5 is LUCIR, which combines samples and uses a loss based on cosine similarity instead of softmnx to solve the problems of imbalance and catastrophic forgetting; Figure 4 It can be seen that compared with comparison method 1, comparison method 2, comparison method 3, comparison method 4 and comparison method 5, the fault diagnosis model provided by the present invention has the highest accuracy in each stage. Figure 5 As shown in the figure, the characteristic visualization diagram of the fault diagnosis model of the present invention at different stages shows that when the number of fault types increases, the model can also achieve good clustering effect on new fault types; at the same time, the knowledge of old types will not be forgotten, and the clustering effect of old types is also very good. Figure 6 As shown, it is a visualized schematic diagram of real features and generated features. It can be seen that the generated features of the present invention have a high overlap rate with the real features, that is, the fault diagnosis accuracy is high.

[0122] The lifelong learning method for bearing incremental fault diagnosis based on generated feature replay described in the present invention trains a feature extractor, a classifier and a feature generator through learning grayscale image samples in the initial stage to construct an original fault diagnosis model; in the incremental learning stage, a feature extractor is used to extract features of the fault category of the nth stage, and a feature distillation loss function is used to prevent the feature extractor from forgetting; a classifier is used to predict the target label, and a classification loss function is used to prevent the classifier from forgetting; a feature generator is used to generate an accurate old class feature distribution to facilitate sharing of parameters between different classes and tasks, and a replay alignment loss function is used to ensure that the feature generator correctly generates features from the old type by making the generated features of the nth stage similar to the generated features of the n-1th stage, so that the final fault diagnosis model provided by the present invention has high diagnostic accuracy and strong robustness, is suitable for fault diagnosis of incremental fault types, and can be widely used in incremental fault diagnosis tasks of complex systems such as machinery, electricity, chemical industry, and aviation.

[0123] Obviously, the above embodiments are merely examples for the purpose of clear explanation and are not intended to limit the implementation methods. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived therefrom are still within the scope of protection of the present invention.

Claims

1. A lifelong learning method for bearing incremental fault diagnosis based on generative feature replay, characterized in that: include: S1: Collect the time domain signals of bearing vibration acceleration of different fault types, perform preprocessing, obtain grayscale image samples, add fault type labels, build a bearing status data set, and divide it into multiple data sets of different diagnosis stages; S2: Build the original fault diagnosis model and set the number of neurons in its classifier fully connected layer to the number of fault types in the initial stage; According to the grayscale image samples of the fault type in the initial stage, a first feature extractor and a first classifier are trained; using the features extracted by the first feature extractor, a discriminator and a generator are alternately trained to obtain a first feature generator; S3: In the nth stage, the nth feature extractor and the nth classifier are initialized through the n-1th feature extractor and the n-1th classifier in the n-1th stage; the nth feature extractor is updated by training with the grayscale image samples of the fault type in the nth stage, and the feature distillation loss function is used to reduce the difference between the features extracted by the nth feature extractor and the n-1th feature extractor; The feature distillation loss function is: Wherein, n=1,2,3,…,N; when n is 1, it is the initial stage; N is the number of diagnostic stages; Expressing expectation, F n-1 (x) represents the feature extracted by the n-1th feature extractor, F n (x) represents the feature extracted by the nth feature extractor; S4: Use the features extracted by the nth feature extractor and the features generated by the n-1th feature generator to jointly train and update the nth classifier, and use the classification loss function to prevent the nth classifier from forgetting; update the number of neurons in the fully connected layer of the nth classifier to the sum of the number of fault types from the initial stage to the nth stage; S5: Initialize the nth feature generator of the nth stage through the n-1th feature generator, and then jointly train the features generated by the n-1th feature generator and the features extracted by the nth feature extractor to obtain the nth feature generator; at the same time, use the replay alignment loss function to constrain the features generated by the nth feature generator to be similar to the features generated by the n-1th feature generator to prevent the nth feature generator from forgetting; The replay alignment loss function It is expressed as: in, Expressing expectation, G n (t,z) represents the features generated by the nth feature generator, G n-1 (t,z) represents the features generated by the n-1th feature generator. When conditioned on a given prior fault type t and a given latent vector z, the generated features of the nth stage are constrained to be similar to those of the n-1th stage; S6: Repeat steps S3 to S5 until all diagnosis stages are trained to obtain a final fault diagnosis model when n=N; S7: Use the final fault diagnosis model to diagnose the grayscale image samples in the untrained test set of all learned tasks to obtain the bearing fault diagnosis results.

2. The lifelong learning method for bearing incremental fault diagnosis based on generated feature replay according to claim 1 is characterized in that: The step S1 specifically includes: Use acceleration sensors to collect bearing vibration acceleration time domain signals of different fault types, add fault type labels, build bearing status data sets, and divide fault diagnosis tasks into different stages; Divide N+1 fault diagnosis tasks into N+1 learning stages; the number of learning stages increases with the number of diagnosis tasks. The datasets D0, D1, K, D N are learned sequentially; the data set in stage n Be learned, D n The i-th sample data, is the fault type label corresponding to the i-th sample data, K n is the number of samples in the dataset at stage n; P n represents the number of all learned fault types T0:n-1={T0,T1,K,Tn-1}, Q n Indicates the fault type T learned in the current stage n The number of n+1 =P n +Q n , so Expressed as 3. The lifelong learning method for bearing incremental fault diagnosis based on generated feature replay according to claim 2 is characterized in that: The preprocessing of the bearing vibration acceleration time domain signal to obtain a grayscale image sample comprises: Performing a fast Fourier transform on the bearing vibration acceleration time domain signal to transform the bearing vibration acceleration time domain signal into a bearing vibration acceleration frequency domain signal while retaining a shape of 1×1024; The bearing vibration acceleration frequency domain signal is reshaped from 1×1024 to 1×32×32 and reconstructed in the new dimension to obtain a 1×32×32 grayscale image sample.

4. The lifelong learning method for bearing incremental fault diagnosis based on generated feature replay according to claim 3 is characterized in that: The training of a first feature extractor and a first classifier based on the grayscale image samples of the fault type in the initial stage includes: The grayscale image samples in the initial stage of Resent-18 network learning are used to train the first feature extractor F0 and the first classifier C0, and the original fault diagnosis model M0 is obtained. The number of neurons in its fully connected layer is set to the number of fault types in the initial stage. The classification loss in the initial stage is the cross entropy loss function: in, represents the bearing vibration acceleration frequency domain signal in the initial stage, γ represents the true label, is the classification prediction result of the original fault diagnosis model M0, K n is the number of grayscale image samples in the nth stage, and Q0 represents the number of fault types T0 learned in the initial stage.

5. The lifelong learning method for bearing incremental fault diagnosis based on generated feature replay according to claim 1 is characterized in that: Using the first feature extracted by the first feature extractor, alternately training with the adversarial generative network to obtain a first feature generator, including: Using the first feature extracted by the first feature extractor F0, Wasserstein GAN is used to alternately update the discriminator D0 and the generator to obtain the first feature generator G0: The first feature generator is adapted to feature generation and continuous learning through the discriminator loss function and the generator loss function: Discriminator loss function Generator loss function in, represents the data set of the nth stage, t is the required fault type, z is the random noise obeying the Gaussian distribution, represents the replay alignment loss function.

6. The lifelong learning method for bearing incremental fault diagnosis based on generated feature replay according to claim 1 is characterized in that: The classification loss function in the incremental learning phase includes the new classification task loss and the old classification task loss: The new classification task loss is the cross entropy loss of the fault type learned in the nth stage: The old classification task loss is the cross entropy loss of the fault types learned in all stages before the nth stage: Among them, P n represents the number of all learned fault types T0:n-1={T0,T1,K,Tn-1}, Q n Indicates the fault type T learned in stage n n The number of n+1 =P n +Q n , γ represents the true label, is the classification prediction result of the original fault diagnosis model M0.

7. The lifelong learning method for bearing incremental fault diagnosis based on generated feature replay according to claim 1 is characterized in that: After the final fault diagnosis model is used to diagnose the bearing vibration acceleration time domain signals of all learned tasks and the bearing fault diagnosis results are obtained, the method further includes: All fault types in the bearing condition dataset are tested using the diagnostic tasks of all learned fault types to verify that the final fault diagnosis model has the ability to overcome catastrophic forgetting.

8. A lifelong learning system for bearing incremental fault diagnosis based on generated feature replay, characterized in that: include: The data acquisition and preprocessing module collects the bearing vibration acceleration time domain signals of different fault types, constructs the bearing status data set, and divides it into multiple bearing fault diagnosis tasks at different diagnosis stages; preprocesses the bearing vibration acceleration time domain signals to obtain grayscale image samples; The initial stage training module is used to train a first feature extractor and a first classifier by learning grayscale image samples in the initial stage; the first feature extracted by the first feature extractor is used to perform alternating training using a generative adversarial network to obtain a first feature generator; Construct the original fault diagnosis model and set the number of neurons in its fully connected layer of the classifier to the number of fault types in the initial stage; The incremental phase training module trains the nth feature extractor, nth classifier and nth feature generator of the nth phase according to the n-1th feature extractor, n-1th classifier and n-1th feature generator of the n-1th phase; uses the feature distillation loss function to reduce the difference between the nth feature and the n-1th feature, uses the replay alignment loss function to constrain the nth generated feature generated by the nth feature generator to be similar to the n-1th generated feature, updates the number of neurons in the fully connected layer of the classifier to the sum of the number of fault types from the initial stage to the nth stage, and obtains the final fault diagnosis model; the feature distillation loss function is: Wherein, n=1,2,3,…,N; when n is 1, it is the initial stage; N is the number of diagnostic stages; Expressing expectation, F n-1 (x) represents the feature extracted by the n-1th feature extractor, F n-1 (x) represents the feature extracted by the nth feature extractor; the replay alignment loss function It is expressed as: in, Expressing expectation, G n (t,z) represents the features generated by the nth feature generator, G n-1 (t,z) represents the features generated by the n-1th feature generator. When conditioned on a given prior fault type t and a given latent vector z, the generated features of the nth stage are constrained to be similar to those of the n-1th stage; The fault diagnosis module uses the final fault diagnosis model to diagnose the grayscale image samples in the untrained test set of all learned tasks to obtain the bearing fault diagnosis results.

Citation Information

Patent Citations

  • Incremental equipment fault diagnosis method based on knowledge distillation and hidden layer sharing

    CN110162018A

  • Incremental learning method based on generative adversarial network knowledge distillation

    CN111967534A

  • Online intelligent fault diagnosis method and system for rotating machinery for lifelong learning

    CN113962289A

  • Gearbox incremental fault diagnosis method and system based on lifelong learning

    CN114429153A

  • Fault diagnosis method and device based on multi-granularity regularization heavy balance incremental learning

    CN114609994A