Bearing cross-domain fault diagnosis method and system based on model-agnostic meta-learning

Abstract

The application provides a bearing cross-domain fault diagnosis method based on model-agnostic meta-learning, for the bearing vibration signal after preprocessing, the noise threshold of the bearing vibration signal is determined according to the adaptive threshold network trained by the source domain data, and the bearing vibration signal after removing noise is obtained; according to the bearing vibration signal after removing noise and the inner loop network and the outer loop network of the model-agnostic meta-learning bearing cross-domain fault diagnosis model trained by the source domain, the final bearing fault diagnosis result is obtained; the adaptive threshold acquisition network is used to adaptively generate a signal threshold, without expert knowledge, and the improved threshold function can directly filter out interference, combined with the model-agnostic meta-learning network, the accuracy of fault diagnosis is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fault diagnosis, and in particular to a bearing cross-domain fault diagnosis method and system based on model-agnostic meta-learning. BACKGROUND

[0002] The statements in this section merely provide background information related to the present application and do not necessarily constitute the prior art.

[0003] As a basic component of rotating machinery equipment, the normal operation of bearings is of great significance to industrial safety production. At present, fault monitoring and diagnosis driven by sensor monitoring data have made great progress. However, in actual engineering scenarios, mechanical equipment is often in variable speed and variable load working conditions. In addition, mechanical equipment works in a healthy state most of the time, making it difficult to obtain enough data to support the establishment of a deep learning model, resulting in serious performance degradation. Therefore, it is crucial to realize a cross-domain fault classifier with good generalization ability under limited data conditions.

[0004] The inventors found that cross-domain fault diagnosis refers to data used for model training and testing conforming to different distributions, i.e., the data of the source domain and the target domain are similar but not completely the same. At present, transfer learning is the mainstream method to solve this problem, and its principle is to use the developed model as the initial point and re-use it in the development process of another model; however, to optimize the various feature space differences between the source domain and the target domain, transfer learning is a difficult process, and when the distribution difference is large, negative transfer is likely to occur, resulting in an undesirable transfer classification result. SUMMARY

[0005] In order to solve the problems of the prior art, the present application provides a bearing cross-domain fault diagnosis method and system based on model-agnostic meta-learning, which uses an adaptive threshold acquisition network to generate signal thresholds adaptively, without the need for expert knowledge, and in combination with an improved threshold function to directly filter out interference, and in combination with a model-agnostic meta-learning network to improve the accuracy of fault diagnosis.

[0006] In order to achieve the above purpose, the present application adopts the following technical solutions:

[0007] The present application provides a bearing cross-domain fault diagnosis method based on model-agnostic meta-learning in a first aspect.

[0008] A bearing cross-domain fault diagnosis method based on model-agnostic meta-learning includes the following processes:

[0009] Obtain the bearing vibration signal and pre-process it;

[0010] The adaptive threshold processing module is configured to: according to the self-adaptive threshold network trained by the source domain data, determine the noise threshold of the bearing vibration signal, and obtain the bearing vibration signal after removing the noise for the pretreated bearing vibration signal.

[0011] The fault diagnosis module is configured to: according to the bearing vibration signal after removing the noise and the inner loop network and the outer loop network of the model-agnostic meta-learning bearing cross-domain fault diagnosis model trained by the source domain, obtain the final bearing fault diagnosis result.

[0012] The second aspect of the present application provides a bearing cross-domain fault diagnosis system based on model-agnostic meta-learning.

[0013] The bearing cross-domain fault diagnosis system based on model-agnostic meta-learning comprises:

[0014] The preprocessing module is configured to: acquire a bearing vibration signal and perform preprocessing.

[0015] The adaptive threshold processing module is configured to: according to the self-adaptive threshold network trained by the source domain data, determine the noise threshold of the bearing vibration signal, and obtain the bearing vibration signal after removing the noise for the pretreated bearing vibration signal.

[0016] The fault diagnosis module is configured to: according to the bearing vibration signal after removing the noise and the inner loop network and the outer loop network of the model-agnostic meta-learning bearing cross-domain fault diagnosis model trained by the source domain, obtain the final bearing fault diagnosis result.

[0017] The third aspect of the present application provides a computer readable storage medium, which stores a program, and the program is executed by a processor to realize the steps of the bearing cross-domain fault diagnosis method based on model-agnostic meta-learning according to the first aspect of the present application.

[0018] The fourth aspect of the present application provides an electronic device, which comprises a memory, a processor and a program stored in the memory and executable on the processor, and the processor realizes the steps of the bearing cross-domain fault diagnosis method based on model-agnostic meta-learning according to the first aspect of the present application when executing the program.

[0019] Compared with the prior art, the present application has the following advantages:

[0020] 1、The present application innovatively proposes a bearing cross-domain fault diagnosis method and system based on model-agnostic meta-learning, which adopts a self-adaptive threshold acquisition network to adaptively generate a signal threshold, does not require expert knowledge, directly filters out interference in combination with an improved threshold function, and improves the accuracy of cross-domain fault diagnosis in combination with a model-agnostic meta-learning network.

[0021] 2. The application innovatively provides a bearing cross-domain fault diagnosis method and system based on model-agnostic meta-learning, in actual engineering scenarios, mechanical equipment is often in variable speed and variable load working conditions, it is difficult to obtain complete and accurate vibration data, the meta-learning can help the model to learn quickly under a small amount of samples, is suitable for small sample and multi-task scenes, and can solve the problems of quick learning and quick adaptation in the case of lacking training samples for new bearing working conditions.

[0022] Advantages of the additional aspects of the application will be partially given in the following description, partially become obvious from the following description, or be known by the practice of the application. BRIEF DESCRIPTION OF DRAWINGS

[0023] The drawings accompanying the specification of this application form a part thereof, serve to provide further understanding of the application, and together with the description of the exemplary embodiments of the application given below, explain the application, and do not constitute an improper limitation of the application.

[0024] Figure 1 A bearing cross-domain fault diagnosis method based on model-agnostic meta-learning is provided for embodiment 1 of the application, and a flowchart of the whole method is shown in the figure;

[0025] Figure 2 A rolling bearing fault simulation platform HFZZ-II is provided for embodiment 1 of the application, and a schematic diagram is shown in the figure;

[0026] Figure 3 A structure diagram of an adaptive threshold network is provided for embodiment 1 of the application;

[0027] Figure 4 A structure diagram of a threshold acquisition module is provided for embodiment 1 of the application;

[0028] Figure 5 A parameter optimization process diagram of a model-agnostic meta-learning method is provided for embodiment 1 of the application;

[0029] Figure 6 A bearing vibration signal of outer ring fault under 1750 RPM rotating speed is provided for embodiment 1 of the application;

[0030] Figure 7 A two-dimensional time-frequency representation obtained by synchronous compression wavelet transform of the signal is provided for embodiment 1 of the application;

[0031] Figure 8 Experimental results under different cross-domain diagnosis tasks are provided for embodiment 1 of the application. DETAILED DESCRIPTION

[0032] The application will be further described below in combination with the drawings and embodiments.

[0033] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0034] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0035] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0036] Example 1:

[0037] like Figure 1 As shown, Embodiment 1 of the present invention provides a bearing cross-domain fault diagnosis method based on model-agnostic meta-learning, including the following process:

[0038] S1: Data Acquisition and Cross-Domain Dataset Establishment

[0039] The HFZZ-II rolling bearing fault simulation platform was built, specifically as follows: Figure 2 As shown, the platform consists of a variable frequency speed-regulating three-phase AC asynchronous motor, a motor control system, a shaft, a bearing housing (including bearings), a radial loading device, and acceleration sensors (one each in the horizontal and vertical directions). It uses piezoelectric acceleration sensors to collect bearing vibration signals of nine different fault types at speeds of 1750 RPM (L1), 2000 RPM (L2), and 2250 RPM (L3), including single or combined faults such as normal state, outer ring fault, inner ring fault, and rolling element fault.

[0040] The sensor collects data at a frequency of 12.8kHz, acquiring 30 signals under each operating condition, for a total of 810 signals under the three speed conditions. Each signal can be represented as x = (x1, x2, ..., x...). N ), x i Let represent the i-th sampling point of the vibration signal x, 1≤i≤N, where N is the total number of sampling points of the signal, set to 1024. Five categories are randomly selected from the nine fault types at each rotational speed. Each category contains one signal to form a meta-task, i.e., the 5-way 1-shot principle. This random selection is repeated 300 times to form the meta-training set. The meta-test set consists of 200 meta-tasks randomly selected at different rotational speeds from the meta-training set according to the 5-way 1-shot principle.

[0041] S2: Signal preprocessing

[0042] Compared with one-dimensional signals, two-dimensional signals can more richly and intuitively represent rolling bearing fault feature information, and are conducive to feature extraction and learning of convolution kernels. Synchronous compressive wavelet transform is a method for analyzing time-frequency characteristics of nonlinear and non-stationary signals, and can deeply mine internal information of signals. Therefore, synchronous compressive wavelet transform is selected as a signal preprocessing method.

[0043] The continuous wavelet transform of each signal x in S1 can be expressed as:

[0044]

[0045] where W x (a,b) is a complex wavelet coefficient spectrum of the signal, a and b are a scale factor and a time translation factor respectively, a -1 / 2 is a normalization constant, which can ensure that the energy before and after the transformation is unchanged, and ψ is a wavelet basis function.

[0046] According to the wavelet coefficient |W x (a,b; ψ) of an arbitrary point (a,b), the instantaneous frequency ω x (a,b) can be obtained.

[0047]

[0048] After calculation by formula (2), a mapping between a point (a,b) in the time-scale plane and a point (ω x (a,b),b) in the time-frequency plane can be established. At this time, the energy value of an arbitrary frequency ω l around the interval can be compressed to ω l , thereby establishing a new time-frequency diagram and improving the blurring phenomenon in the scale direction. The expression of synchronous compressive wavelet transform is:

[0049]

[0050] In formula (3), T x (ω l ,b) is a synchronous compressive transform spectrum, a k is a discrete scale; and k is the number of scales.

[0051] According to formula (3), a two-dimensional time-frequency representation of each signal in the meta training set and the meta test set under each fault after synchronous compressive wavelet transform processing can be obtained. The two-dimensional time-frequency diagram is converted into a gray-scale diagram with a size of 224x224, which is used as the input of the subsequent model.

[0052] S3: Building and training of bearing cross-domain fault diagnosis model based on model-agnostic meta-learning

[0053] The bearing cross-domain fault diagnosis model based on model-agnostic meta-learning is built, and the meta-training set is input into the model for training. This step is composed of three parts: adaptive threshold network building in the bearing cross-domain fault diagnosis model based on model-agnostic meta-learning, model-agnostic meta-learning model inner loop training in the bearing cross-domain fault diagnosis model based on model-agnostic meta-learning, and model-agnostic meta-learning model outer loop training in the bearing cross-domain fault diagnosis model based on model-agnostic meta-learning.

[0054] S3.1: Adaptive threshold network building

[0055] The adaptive threshold network can automatically determine the noise threshold of the signal, remove the interference of irrelevant information, and improve the accuracy of cross-domain bearing fault diagnosis. The specific structure is as shown in Figure 3 . The structure of the threshold acquisition module is as shown in Figure 4 . In the threshold acquisition module, the Sigmoid activation function limits the parameters of the previous layer to [0, 1] through equation (4):

[0056]

[0057] In the formula, w n is the feature of the nth neuron, and a n is the scaling parameter of the nth feature.

[0058] After passing through the threshold acquisition module, the final threshold is calculated as:

[0059]

[0060] In the formula, λ n is the threshold of the nth feature mapping channel; represents the average value of the ith and jth fault features in the nth feature mapping channel; i, j, and n represent the width, height, and channel of the feature, respectively. The final threshold is controlled within a reasonable range to ensure that the output feature is not all 0.

[0061] The improved threshold function expression is:

[0062]

[0063] In the formula, input is the input feature, output is the filtered output feature, and λ n is the threshold determined by the threshold acquisition module.

[0064] S3.2: Model-agnostic meta-learning model inner loop training

[0065] The inner loop structure of the model-agnostic meta-model can be any basic network model for extracting features of a specific task. The loss function is selected as the cross-entropy loss function, but due to the deepening of the network layers, overfitting phenomenon is easy to occur, therefore, L2 regularization is added on this basis to prevent network parameter overfitting, improve the robustness of the model, and the loss function of the model is:

[0066]

[0067] In the formula, x (j) and y (j) respectively represent the jth data and label; η is a regularization parameter; ω, k are the regularization parameters and the number of regularization parameters respectively, and i represents the ith task.

[0068] Suppose the model is f θ , when the model is applied to a new training task, the updating process of the parameter θ is:

[0069]

[0070] In the formula, θ i ' is the optimal parameter updated on the ith task, α is the learning rate of the inner loop, L li represents the task loss, is the loss gradient, is the added regularization term, in the bearing fault diagnosis, the optimization process can be represented by Figure 5 .

[0071] S3.3: Model-agnostic meta-learning model outer loop training

[0072] In the outer loop, the model-agnostic meta-learning model needs to quickly adapt to new tasks. Therefore, the updated parameter θ must have the smallest loss error on all meta-training tasks, and the final meta-target can be represented as:

[0073]

[0074] In the formula, p(l) is the distribution of the meta task.

[0075] Finally, according to formula (10), the model is updated by using the stochastic gradient descent method, and the parameters of the model are determined after several iterations, and the model training is completed.

[0076]

[0077] In the formula, β represents the learning rate of the outer loop of the model.

[0078] S4: Bearing cross-domain fault diagnosis

[0079] After the bearing cross-domain fault diagnosis model based on model-agnostic meta-learning is completed, the data of the meta test set is input into the trained model to complete the type diagnosis of the bearing cross-domain fault.

[0080] A specific example is provided below:

[0081] According to S1, the signal is processed to determine the specific migration task of the rolling bearing, as shown in Table 1.

[0082] Table 1: Rolling bearing migration task detail description table

[0083]

[0084] For S2, an example is taken of a signal of a bearing with outer ring fault at a speed of 1750 RPM. The time domain waveform of the signal is as shown in Figure 6 The two-dimensional time-frequency representation of the signal can be obtained by performing synchronous compression wavelet transform on the signal, as shown in Figure 7

[0085] After S3 completes the bearing cross-domain fault diagnosis model based on model-agnostic meta-learning, migration fault diagnosis testing is performed, and comparison is made with other transfer learning network algorithms under the migration conditions set in Table 1 to highlight the effectiveness and superiority of the method. The experimental results are finally averaged over 10 experiments, as shown in Figure 8 From the results, it can be seen that the method proposed in the present application has a certain degree of performance improvement compared with other transfer algorithms under various migration conditions.

[0086] Embodiment 2:

[0087] Embodiment 2 of the present application provides a bearing cross-domain fault diagnosis system based on model-agnostic meta-learning, comprising:

[0088] The preprocessing module is configured to obtain a bearing vibration signal and perform preprocessing;

[0089] The adaptive threshold processing module is configured to determine the noise threshold of the bearing vibration signal according to the adaptive threshold network trained using the source domain data, to obtain a bearing vibration signal after noise removal;

[0090] The fault diagnosis module is configured to obtain a final bearing fault diagnosis result according to the bearing vibration signal after noise removal, and the inner loop network and the outer loop network of the model-agnostic meta-learning bearing cross-domain fault diagnosis model trained using the source domain.

[0091] The working method of the system is the same as the bearing cross-domain fault diagnosis method based on model-agnostic meta-learning provided in Embodiment 1, which will not be repeated here. ​

[0092] Embodiment 3:

[0093] Embodiment 3 provides a computer readable storage medium, having stored thereon a program, which when executed by a processor implements the steps of the bearing cross-domain fault diagnosis method based on model-agnostic meta-learning as described in Embodiment 1 of the present application.

[0094] Embodiment 4:

[0095] Embodiment 4 provides an electronic device, comprising a memory, a processor, and a program stored in the memory and capable of running on the processor, wherein the processor implements the steps of the bearing cross-domain fault diagnosis method based on model-agnostic meta-learning as described in Embodiment 1 of the present application when executing the program.

[0096] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. The present application can have various modifications and changes for those skilled in the art. Any modification, equivalent replacement, improvement, etc. within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A bearing cross-domain fault diagnosis method based on model-agnostic meta-learning, characterized in that, The method comprises the following steps: obtaining a bearing vibration signal and preprocessing the same; determining a noise threshold of the bearing vibration signal according to an adaptive threshold network trained by source domain data, to obtain a bearing vibration signal after noise removal; obtaining a final bearing fault diagnosis result according to the bearing vibration signal after noise removal and an inner loop network and an outer loop network of a model-agnostic meta-learning bearing cross-domain fault diagnosis model trained by the source domain.

2. The bearing cross-domain fault diagnosis method based on model-agnostic meta-learning according to claim 1, wherein the preprocessing comprises: performing synchronous compression wavelet transform on the obtained bearing vibration signal to obtain a corresponding two-dimensional time-frequency image, and converting the two-dimensional time-frequency image into a grayscale image with a set size.

3. The bearing cross-domain fault diagnosis method based on model-agnostic meta-learning according to claim 1, wherein the step of determining the noise threshold of the bearing vibration signal comprises:

4. The bearing cross-domain fault diagnosis method based on model-agnostic meta-learning according to claim 3, wherein the step of determining the noise threshold of the bearing vibration signal comprises:

5. The bearing cross-domain fault diagnosis method based on model-agnostic meta-learning according to claim 3, comprising: No. n The noise threshold for the feature is: at the ... n Feature mapping channels The average value of the first fault characteristic and the first fault characteristic n The product of the scaling parameters of each feature map channel, where, i and j These represent the width and height of the feature, respectively. a preprocessing module configured to obtain a bearing vibration signal and preprocess the same; No. n The scaling parameters for each feature are: ,in, For the first n Features of each neuron. an adaptive threshold processing module configured to determine a noise threshold of the bearing vibration signal according to an adaptive threshold network trained by source domain data, to obtain a bearing vibration signal after noise removal, for the bearing vibration signal after preprocessing; When the absolute value of the characteristic of the input signal is greater than or equal to the noise threshold, the output characteristic is: ; otherwise, the output characteristic is zero. wherein input is an input feature, is a threshold value determined by the threshold obtaining module. 6.A bearing cross-domain fault diagnosis system based on model-agnostic meta-learning, characterized in that, a fault diagnosis module configured to obtain a final bearing fault diagnosis result according to the bearing vibration signal after noise removal and an inner loop network and an outer loop network of a model-agnostic meta-learning bearing cross-domain fault diagnosis model trained by the source domain.

7. The bearing cross-domain fault diagnosis system based on model-agnostic meta-learning according to claim 6, wherein the step of determining the noise threshold of the bearing vibration signal comprises:

8. The bearing cross-domain fault diagnosis system based on model-agnostic meta-learning according to claim 7, wherein the step of determining the noise threshold of the bearing vibration signal comprises: The program is executed by the processor to implement the steps in the bearing cross-domain fault diagnosis method based on model-agnostic meta-learning according to any one of claims 1-5. The processor executes the program to implement the steps in the bearing cross-domain fault diagnosis method based on model-agnostic meta-learning according to any one of claims 1-5. ​ The noise threshold of the first n characteristic is the product of the average value of the first n characteristic mapping channel faulty characteristics and the scaling parameter of the first n characteristic mapping channel, wherein, i and j respectively represent the width and height of the characteristic. ​ The scaling parameter of the nth feature is: wherein, is the feature of the nth neuron. Or, when the absolute value of the feature of the input signal is greater than or equal to the noise threshold, the output feature is: ; otherwise, the output feature is zero. wherein input is an input feature, is a threshold value determined by the threshold obtaining module.

9. A computer-readable storage medium having stored thereon a program, characterized in that, ​ 10. An electronic device comprising a memory, a processor, and a program stored on the memory and executable on the processor, characterized in that, ​

Images