Improved meta-learning fault diagnosis method based on ordered time-frequency information

Abstract

The invention provides an improved meta-learning fault diagnosis method based on ordered time-frequency information, and relates to the technical field of fault diagnosis. According to the method, time-frequency information during data preprocessing is orderly arranged to form a feature matrix, a fault diagnosis model is constructed based on an improved meta-learning network, and equipment fault diagnosis is performed by using the trained and debugged model; the improved meta learning network model comprises a space attention module, an improved multi-scale convolution module, a convolution module and a full connection layer which are connected in sequence. According to the fault diagnosis method, feature information contained in original signal data can be extracted from multiple dimensions, meanwhile, feature strengthening is conducted on the original signal data, the influence of noise signals can be resisted to a certain degree, and compared with a traditional fault diagnosis technology, the precision and robustness of fault diagnosis are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, and in particular to an improved meta-learning fault diagnosis method based on ordered time-frequency information. Background Art

[0002] In the context of intelligent big data, the precision requirements for equipment are getting higher and higher. The application forms of equipment are becoming more and more diverse, and data acquisition is becoming more and more difficult. With the promotion of artificial intelligence technology, the performance requirements for equipment are becoming more and more urgent, especially in the power source, bearing group, gearbox and other parts of the equipment. Once a failure occurs, it will cause significant economic losses and even endanger the safety of industry personnel, causing immeasurable consequences.

[0003] At present, fault diagnosis generally includes three stages: acquisition and preprocessing of raw signals, feature extraction of processed data, and identification and classification using trained diagnostic models to achieve fault diagnosis. Among them, data preprocessing and feature extraction are two important steps for the model to fully learn feature information, while identification and classification affect the accuracy of diagnosis.

[0004] Traditional fault diagnosis technology is generally based on the expert's experience and knowledge for classification and labeling, which is time-consuming and laborious. In addition, this process has certain subjective elements, which can easily lead to problems such as poor classification effect and low accuracy. At present, a large number of related research works focus on feature information of a single dimension, and only extract signal features from a certain dimension, such as time domain or frequency domain. When the equipment is in normal operation, the location and cause of the fault are random, or the signals generated by faults in different locations may be coupled when collected. Therefore, extracting signal features only from a single dimension may cause the model to only focus on a certain local information and ignore the global information, resulting in the loss of important features, which seriously affects the accuracy of the model's diagnosis and classification. In addition, when the model learns features, as the network deepens, some feature information faces ambiguity or loss, which also leads to low classification accuracy when diagnosing equipment faults. Summary of the invention

[0005] In view of the shortcomings of the prior art, the present invention provides an improved meta-learning fault diagnosis method based on ordered time-frequency information, which solves the problems of incomplete feature information extraction and low diagnostic accuracy in the prior fault diagnosis technology.

[0006] The present invention provides the following technical solution: The present invention proposes an improved meta-learning fault diagnosis method based on ordered time-frequency information, the method comprising the following steps:

[0007] S1. Obtain the original signal data of the faulty device and preprocess the original signal data. The preprocessing process is as follows:

[0008] Extract the time domain information of the original signal data, perform fast Fourier transform and wavelet transform on the original signal data to obtain the corresponding frequency domain information, arrange the obtained time domain information and frequency domain information in order, and construct a feature matrix;

[0009] S2. Construct an improved meta-learning network fault diagnosis model and train and debug the model: the meta-learning network fault diagnosis model includes a spatial attention module, an improved multi-scale convolution module, a convolution module and a fully connected layer connected in sequence; the spatial attention module carries a forward transmission layer to ensure that feature information is not lost; the improved multi-scale convolution module is based on the existing multi-scale convolution module, and the convolution layer in the multi-scale convolution module is deleted to accelerate the extraction of feature information; the convolution module includes three convolution layers for reducing the dimension of features, and two fully connected layers as classification output layers;

[0010] S3. Perform fault diagnosis based on the trained and debugged meta-learning network fault diagnosis model.

[0011] Furthermore, in the preprocessing, the obtained time domain information and frequency domain information are arranged in order, and the specific process of constructing the feature matrix is ​​as follows:

[0012] The obtained time-frequency information is selected to construct a feature matrix, and the feature matrix is ​​time domain feature information, fast Fourier transform feature information and wavelet transform feature information from top to bottom.

[0013] Furthermore, the orderly discharge process is as follows:

[0014] The eigenvalues ​​with a large range of numerical changes are placed in the center of the feature matrix, and the values ​​with a small range of changes are placed at the edge of the feature matrix to construct a feature image; among them, the eigenvalues ​​with a variance greater than 1.00 are defined as eigenvalues ​​with a large range of numerical changes, and the eigenvalues ​​with a variance less than 0.01 are defined as eigenvalues ​​with a small range of numerical changes.

[0015] Furthermore, the spatial attention module is used to reduce the dimension of the input data and then perform maximum pooling and mean pooling respectively, and concatenate the results of maximum pooling and mean pooling with the input layer.

[0016] Furthermore, the improved multi-scale convolution module removes the 1*1 convolution layer and the 3*3 max pooling layer on the basis of the original Inception module, and transforms the 5*5 convolution layer into two 3*3 convolution layers.

[0017] Furthermore, the training process of the meta-learning network fault diagnosis model is as follows:

[0018] Collect relevant data from published literature, use the preprocessing method in step S1 to process the collected relevant data, use the feature matrix obtained after preprocessing as a training set, train the constructed meta-learning network fault diagnosis model, and use the original data preprocessed in S1 as a debugging set to adjust the parameters of the trained meta-learning network fault diagnosis model.

[0019] By adopting the above technology, compared with the prior art, the beneficial effects of the present invention are as follows:

[0020] 1) The present invention obtains ordered time-frequency information from signal data, constructs a fault diagnosis model based on an improved meta-learning network, and uses the trained and debugged model to perform fault diagnosis on the equipment; wherein the ordered time-frequency information includes preprocessing the original signal to avoid the loss of feature information in the process of extracting from the original signal as much as possible; the improved meta-learning network model includes a sequentially connected spatial attention module, an improved multi-scale convolution module, a convolution module, and a fully connected layer to achieve feature enhancement and feature expression and improve model learning efficiency; the model first uses a common and easy-to-collect data set as a training set to train the initial parameters of the model; then uses the debugging set in the fault data set of the equipment to be diagnosed to debug the parameters and improve the parameter accuracy, so that the model can comprehensively learn the fault feature information in the face of scarce fault samples and achieve fault diagnosis;

[0021] 2) The spatial attention module in the present invention carries a forward transfer layer, which can ensure that feature information is not lost during feature enhancement. The improved multi-scale convolution module removes part of the network, saves computing resources, and ensures the comprehensiveness of feature information extraction. These processes greatly reduce the time for subsequent model learning features, reduce training costs and time costs;

[0022] 3) The present invention can perform fault diagnosis when there are few effective labeled samples of the faulty equipment to be diagnosed, and improve the diagnostic accuracy through rapid learning of the model. Compared with the existing fault diagnosis technology, the present invention has certain performance improvements;

[0023] 4) The present invention adjusts the position of the time domain feature information, fast Fourier transform feature information and wavelet transform feature information in the feature matrix, which not only retains the comprehensiveness of the feature information, but also ensures the model's repeated learning of important features, thereby improving the model learning efficiency and diagnostic accuracy to a certain extent. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 It is a general flow chart of the improved meta-learning fault diagnosis method based on ordered time-frequency information of the present invention;

[0025] Figure 2Flow chart of an improved meta-learning fault diagnosis method based on ordered time-frequency information in an embodiment of the present invention;

[0026] Figure 3 A diagram of a method for dividing a data set in an embodiment of the present invention;

[0027] Figure 4 A structural diagram of a feature image formed by preprocessing in an embodiment of the present invention;

[0028] Figure 5 This is a structural diagram of a spatial attention module in an embodiment of the present invention;

[0029] Figure 6 This is a structural diagram of an improved multi-scale convolution module in an embodiment of the present invention;

[0030] Figure 7 Schematic diagram of the overall structure of the improved meta-learning network fault diagnosis model in an embodiment of the present invention. DETAILED DESCRIPTION

[0031] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments of the specification. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0032] On the contrary, the present invention covers any substitution, modification, equivalent method and scheme made on the essence and scope of the present invention as defined by the claims. Further, in order to make the public have a better understanding of the present invention, some specific details are described in detail in the detailed description of the present invention below. Those skilled in the art can fully understand the present invention without the description of these details.

[0033] The present invention provides an improved meta-learning fault diagnosis method based on ordered time-frequency information, which solves the problems of incomplete feature information extraction and low diagnostic accuracy in existing fault diagnosis technologies, and realizes high-precision diagnosis of equipment faults, thereby reducing maintenance costs. The overall idea is as follows:

[0034] In order to solve the problem that the existing fault diagnosis technology lacks fault feature information and has low model diagnosis accuracy due to various reasons, the present invention proposes an improved meta-learning network fault diagnosis model based on ordered time-frequency information based on deep learning. The ordered time-frequency information mainly includes preprocessing the original signal data, extracting the time domain feature information of the signal data, obtaining the frequency domain feature information through fast Fourier transform and wavelet transform, and arranging the obtained feature information in order to form a feature matrix. Then the acquired feature matrix is ​​input into the model for deep learning. The model network mainly includes a spatial attention module, an improved multi-scale convolution module, a convolution module and a fully connected layer. Among them, the spatial attention module is used to enhance the features of the input feature image; the improved multi-scale convolution module is used to extract feature information of different scales from the enhanced feature image, including both shallow and deep information; the convolution module is used to learn the extracted feature information, and finally the fully connected layer is used to realize fault classification. The technical solution of the present invention can effectively improve the model learning efficiency and the accuracy of diagnosis, and at the same time has a certain robustness.

[0035] The technical solution of the present invention includes but is not limited to the fault diagnosis of the motor, bearing, gearbox and other components of the equipment, and is also applicable to the fault diagnosis of other high-end equipment. For ease of explanation, in the following embodiments of the present application, the fault diagnosis of the bearing in the motor is taken as an example for explanation.

[0036] Embodiment 1:

[0037] This embodiment proposes an improved meta-learning fault diagnosis method based on ordered time-frequency information. Figure 1-2 , the method comprises the following steps:

[0038] S1. Obtain the required original signal data and perform orderly time-frequency preprocessing on the signal data;

[0039] S2. Construct an improved meta-learning network fault diagnosis model; the model includes a sequentially connected spatial attention module, an improved multi-scale convolution module, a convolution module, and a fully connected layer; the spatial attention module carries a forward transmission layer to ensure that feature information is not lost; the improved multi-scale convolution module deletes some convolution layers to accelerate the extraction of feature information; the convolution module contains three convolution layers for feature dimensionality reduction, and two fully connected layers as classification output layers;

[0040] S3. Implement fault diagnosis based on the trained and debugged fault diagnosis model.

[0041] This embodiment first obtains ordered time-frequency information from signal data, then builds a fault diagnosis model based on an improved meta-learning network, and then performs model training. After the trained model is debugged, the equipment fault diagnosis is finally performed. The fault diagnosis method of this embodiment can extract signal feature information from different domains, and orderly compose feature images according to the change amplitude of its information value so that the model can learn quickly; the network model includes a sequentially connected spatial attention module, an improved multi-scale convolution module, a convolution module and a fully connected layer, first realizes the feature enhancement of the feature image, and then performs multi-scale information learning on the enhanced features, so that the model can learn the fault features more accurately, realize high-precision fault diagnosis, and have a certain robustness.

[0042] The following is combined with Figure 1-7 , and explanations of specific steps S1-S3 are provided to describe in detail the implementation process of an embodiment of the present invention.

[0043] S1. Obtain the required original signal data and perform orderly time-frequency preprocessing on the signal data.

[0044] Signal data acquisition:

[0045] When the equipment is running, the power equipment such as the motor will generate signals due to vibration. The sensors collect the signal data of various positions, such as the inner ring, outer ring, rolling element and other positions of the bearing that are prone to damage. The corresponding vibration signals of these positions are collected and sorted.

[0046] Data preprocessing:

[0047] Common preprocessing methods include Fourier transform, wavelet transform, empirical mode decomposition, etc., or directly using time domain information. However, these methods only obtain feature information from a single dimension, which may easily miss some important information. In this regard, this embodiment collects the time domain feature information of the signal and the frequency domain information after fast Fourier transform and wavelet transform, and arranges these information in order from top to bottom, such as Figure 4 As shown in , the combination constitutes a characteristic matrix. Among them, the characteristic elements in the matrix include information about changes in characteristic values ​​when a fault occurs, such as mean, maximum value, variance, standard deviation, etc. in time domain information; amplitude variance, amplitude root mean square, amplitude kurtosis, etc. in frequency domain information of fast Fourier transform; energy mean, energy minimum, pulse factor, etc. in low-frequency information of wavelet transform. Then make some adjustments to this information, such as moving the position of some characteristic elements with a large range of eigenvalue changes to the center of the matrix, and moving some characteristic elements with a small range of changes to the edge of the matrix. This not only retains the comprehensiveness of the characteristic information, but also ensures the repeated learning of important features by the model, which has a certain improvement in model learning efficiency and diagnostic accuracy.

[0048] Divide the dataset:

[0049] Generally, fault diagnosis requires a large number of labeled samples to train the model, but this presents some difficulties. In this embodiment, the model can first learn the features of a simple and easy-to-obtain labeled data set to optimize the initial parameters of the model. Then, the labeled data set of the device that needs to be fault diagnosed is used for debugging, so that the model can quickly learn the fault features. In this embodiment, the data set is divided into an easily accessible training set and a test set of the faulty device. The training set can be divided into a support set and a query set in proportion; the test set is divided into a debugging set and a detection set, such as Figure 3 As shown in the figure, the division ratio can be 7:3 or 8:2, which is determined according to the actual situation.

[0050] S2. Build an improved meta-learning network fault diagnosis model:

[0051] The model includes a sequentially connected spatial attention module, an improved multi-scale convolution module, a convolution module, and a fully connected layer; the spatial attention module carries a forward transmission layer to ensure that feature information is not lost; the improved multi-scale convolution module deletes some convolution layers to accelerate the extraction of feature information; the convolution module contains three convolution layers for feature dimensionality reduction, and two fully connected layers as classification output layers. That is, the preprocessed feature image is first input into the spatial attention module, then to the improved multi-scale convolution module, then to the convolution module, and finally to the fully connected layer to achieve fault classification. Among them:

[0052] (1) Spatial Attention Module

[0053] The spatial attention module first reduces the dimension of the input image itself, then obtains the results of maximum pooling and mean pooling respectively, fuses them and performs convolution operations, and finally fuses the output with the original feature image. The network structure is as follows: Figure 5 As shown in the figure, the effect of enhancing important features of the image is achieved. The spatial attention module can strengthen the features in the feature image space, and then fuse it with the original image to enhance the feature expression of the image, which is helpful for model learning efficiency and improving diagnostic accuracy.

[0054] (2) Improve the multi-scale convolution module

[0055] The improved multi-scale convolution module deletes the 1x1 convolution layer and 3x3 maximum pooling layer of the Inception module, and transforms the 5x5 convolution layer into two 3x3 convolution layers. The network structure is as follows: Figure 6 As shown in the figure, the input features are convolved once and twice with 3x3 convolution to extract shallow and deep feature information, and then the extracted feature information is superimposed and output to the convolution module for feature processing. This can extract comprehensive feature information and ensure the invariance of feature information, while also reducing certain computing resources and accelerating model training.

[0056] (3) Convolutional modules and fully connected layers

[0057] The convolution module consists of three 3x3 convolutional networks. The three-layer convolutional network is used to reduce the dimension of the input features and update the corresponding parameters of the model. Finally, the feature information is input into the fully connected layer, which acts as a classifier to distinguish the fault types. The overall network structure is as follows: Figure 7 As shown, SA is the spatial attention module.

[0058] S3. Implement fault diagnosis based on the trained and debugged fault diagnosis model.

[0059] After the above fault diagnosis model is constructed, in order to further ensure the accuracy of the diagnosis results, the above constructed model is first trained and optimized. Specifically:

[0060] (1) Parameter setting of fault diagnosis model

[0061] The experiment was conducted in Jupyter Notebook software, using PY torch code programming and GPU accelerated training. The loss threshold was set to 0.025, the model optimizer used Adam, the training learning rate was set to 0.05, the debugging learning rate was set to 0.005, the training batch was set to 140, and the loss function used the cross entropy loss function (Cross EntropyLoss). Each group of experiments was repeated 20 times to avoid experimental contingency.

[0062] (2) Model training and debugging

[0063] First, the support set divided and processed in S1 is input into the model. The model processes the input feature data in batches according to the batch size, updates the model parameter gradient according to the cross entropy loss function, and finds the optimal parameters. When the training loss value is less than the loss threshold, stop training and keep the current parameters. Input the query set into the trained model to judge the classification accuracy. When the accuracy meets the standard, save the optimal model at this time. Then use the debugging set to debug the model parameters so that the model can quickly learn the fault characteristics of the equipment that needs to be detected.

[0064] (3) Fault diagnosis classification

[0065] The test set data is input into the debugged model, and the classification effect is measured by different test standards such as classification accuracy, confusion matrix, F1 scores, etc. When the effect reaches the expected result, a fully usable bearing fault diagnosis model is obtained. The overall training process is as follows: Figure 2 shown.

[0066] In order to verify the accuracy of the improved meta-learning fault diagnosis method based on ordered time-frequency information in this embodiment when performing fault diagnosis, this embodiment selects part of the data in the CWRU data set of Case Western Reserve University as the training set (as shown in Table 1), and selects part of the data in the bearing data set of Paderborn University in Germany (as shown in Table 2) as the test set for experimental verification, and selects seven types of faults in different locations and types (including the six types of fault data listed in Table 1 and Table 2, and healthy bearing data signals) for classification tasks. The improved meta-learning fault diagnosis model based on ordered time-frequency information proposed in this embodiment is used for diagnostic classification, and the comprehensiveness of feature information extraction, model diagnostic accuracy and training time are all improved to a certain extent. Even when there is only one labeled fault data in the fault set of the equipment to be detected, the diagnostic accuracy rate can reach 76.12%.

[0067] In addition, the ordered time-frequency information preprocessing method (T-FW) proposed in this embodiment is compared with the traditional preprocessing methods (wavelet transform (WT), time-frequency (TF), disordered time-frequency (T-FW disorder)) in terms of feature information extraction, and the comparison results are shown in Table 3. The improved meta-learning fault diagnosis model (SMCNN) proposed in this embodiment is compared with the classical network (such as convolutional neural network (CNN), model agnostic meta-learning network (MAML), prototype network (PN)), and the comparison results are shown in Table 4. NK (7-5) refers to seven types of bearing faults, each of which is diagnosed with only five labeled fault data.

[0068] Table 1 Data selected from the CWRU dataset

[0069]

[0070] Table 2 Data selected from the bearing dataset of the University of Paderborn, Germany

[0071]

[0072] Table 3 Comparison of feature information extraction

[0073]

[0074] Table 4 Diagnostic accuracy of different models

[0075] Model CNN MAML PN SMCNN NK(7-5) 71.07％ 77.21％ 86.47％ 90.11％

[0076] As shown in Tables 3-4, the ordered time-frequency information preprocessing method of this embodiment can effectively improve the amount of feature information extracted from signal data, and help the model learn feature information quickly and comprehensively. The improved meta-learning network model based on ordered time-frequency information in this embodiment can effectively improve the classification accuracy of diagnosis, reduce training costs, and save resources and time costs.

[0077] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. An improved meta-learning fault diagnosis method based on ordered time-frequency information, characterized in that: The method comprises the following steps: S1. Obtain the original signal data of the faulty device and preprocess the original signal data. The preprocessing process is as follows: Extract the time domain information of the original signal data, perform fast Fourier transform and wavelet transform on the original signal data to obtain the corresponding frequency domain information, arrange the obtained time domain information and frequency domain information in order, and construct a feature matrix; S2. Construct an improved meta-learning network fault diagnosis model and train and debug the model: the meta-learning network fault diagnosis model includes a spatial attention module, an improved multi-scale convolution module, a convolution module and a fully connected layer connected in sequence; the spatial attention module carries a forward transmission layer to ensure that feature information is not lost; the improved multi-scale convolution module is based on the existing multi-scale convolution module, and the convolution layer in the multi-scale convolution module is deleted to accelerate the extraction of feature information; the convolution module includes three convolution layers for reducing the dimension of features, and two fully connected layers as classification output layers; S3. Perform fault diagnosis based on the trained and debugged meta-learning network fault diagnosis model.

2. The improved meta-learning fault diagnosis method based on ordered time-frequency information as claimed in claim 1, characterized in that: In the preprocessing, the obtained time domain information and frequency domain information are arranged in order, and the specific process of constructing the feature matrix is ​​as follows: The obtained time-frequency information is selected to construct a feature matrix, and the feature matrix is ​​time domain feature information, fast Fourier transform feature information and wavelet transform feature information from top to bottom.

3. The improved meta-learning fault diagnosis method based on ordered time-frequency information as claimed in claim 2, characterized in that: The orderly discharge process is as follows: The eigenvalues ​​with a large range of numerical changes are placed in the center of the feature matrix, and the values ​​with a small range of changes are placed at the edge of the feature matrix to construct a feature image; among them, the eigenvalues ​​with a variance greater than 1.00 are defined as eigenvalues ​​with a large range of numerical changes, and the eigenvalues ​​with a variance less than 0.01 are defined as eigenvalues ​​with a small range of numerical changes.

4. The improved meta-learning fault diagnosis method based on ordered time-frequency information as claimed in claim 1, characterized in that: The spatial attention module is used to reduce the dimension of the input data and then perform maximum pooling and mean pooling respectively, and concatenate the results of maximum pooling and mean pooling with the input layer.

5. The improved meta-learning fault diagnosis method based on ordered time-frequency information as claimed in claim 1, characterized in that: The improved multi-scale convolution module removes the 1*1 convolution layer and the 3*3 maxpooling pooling layer on the basis of the original Inception module, and transforms the 5*5 convolution layer into two 3*3 convolution layers.

6. The improved meta-learning fault diagnosis method based on ordered time-frequency information as claimed in claim 3, characterized in that: The training process of the meta-learning network fault diagnosis model is as follows: Collect relevant data from published literature, use the preprocessing method in step S1 to process the collected relevant data, use the feature matrix obtained after preprocessing as a training set, train the constructed meta-learning network fault diagnosis model, and use the original data preprocessed in S1 as a debugging set to adjust the parameters of the trained meta-learning network fault diagnosis model.

Images