Intelligent interpretable mechanical fault diagnosis method based on Gabor wavelet book

Abstract

The invention discloses an interpretable mechanical fault intelligent diagnosis method based on a Gabor wavelet book, and the method comprises the steps: constructing 100 Gabor wavelets containing different frequency components, and enabling the Gabor wavelets to form an initial weight matrix of an automatic encoder, so as to guide a model to extract multi-scale and multi-direction fault features; only 10% of gear variable speed samples are selected as training samples, and a training sample set is established in a random segment taking mode; then, the training sample set is directly input into a Gabor wavelet initialized automatic encoder for training, and the network iteration step number is optimized to only 10 times; next, the trained weight matrix is utilized to map the original sample into learning features, and the learning features are combined with category labels to jointly train a Softmax regression classifier; and taking the rest 90% of gear variable speed samples as a test sample set to verify and diagnose fault categories. According to the method, the interpretability research of the network weight is realized, and the accuracy and the calculation efficiency of fault diagnosis of the gear under the condition of variable rotating speed are obviously improved.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis of rotating machinery vibration signals, and in particular to an interpretable intelligent mechanical fault diagnosis method based on Gabor wavelet transform. Background Art

[0002] In modern industry, mechanical intelligent fault diagnosis technology plays a vital role in ensuring the stable operation of the entire system. Therefore, fault diagnosis has always been the focus of attention in industry and academia. Traditional fault diagnosis methods, such as vibration analysis, acoustic detection, and temperature monitoring, can provide effective diagnostic information in some cases, but their diagnostic capabilities are often limited when dealing with complex fault modes or early faults. With the rapid development of artificial intelligence and deep learning, data-driven fault diagnosis methods have gradually shown their unique advantages, but deep learning models often face the problem of poor interpretability. This means that the internal working mechanism of the model is not transparent to users, which limits their widespread application in fault diagnosis. Therefore, developing an intelligent diagnosis model that has both high performance and strong interpretability has become an important research direction in the current field of fault diagnosis.

[0003] As a commonly used tool in image processing and feature extraction, Gabor wavelets perform well in image processing, texture analysis, and target detection. Gabor wavelets are sensitive to image edges and can provide excellent direction and scale selection characteristics. Compared with the traditional Fourier transform, Gabor wavelet transform has good time-frequency localization characteristics. This means that the direction, bandwidth, and center frequency of the Gabor filter can be easily adjusted to achieve the best resolution of the signal in the spatial domain and frequency domain. Gabor wavelet transform also has multi-resolution characteristics, that is, scaling ability. By applying a set of Gabor wavelets with different time-frequency characteristics to the image transform, each channel can capture certain local features of the input image, so that the image can be analyzed at different granularity levels as needed. The kernels used in the Gabor transform are very similar to the two-dimensional receptive field profiles of simple cells in the mammalian visual cortex, with excellent spatial locality and direction selectivity. They are able to capture spatial frequency (scale) and local structural features in multiple directions within the local area of ​​the image. Therefore, Gabor decomposition can be regarded as a directional microscope that is sensitive to direction and scale. At the same time, the two-dimensional Gabor function is also similar to the basic features of the enhanced image, such as edges, peaks, valleys, and ridge profiles.

[0004] In summary, developing an interpretable mechanical fault intelligent diagnosis method based on Gabor wavelet transform to enhance the interpretability and performance of intelligent networks in mechanical fault diagnosis is an urgent and challenging technology. Summary of the invention

[0005] To solve the above technical problems, the present invention discloses an interpretable intelligent mechanical fault diagnosis method based on Gabor wavelet, which uses a set of designed Gabor wavelets to initialize the weight vector of the neural network to guide the model to extract multi-scale and multi-directional fault features.

[0006] To achieve the above object, the present invention adopts the following technical solutions:

[0007] An interpretable mechanical fault intelligent diagnosis method based on Gabor wavelet transform includes the following steps:

[0008] A. Design a set of Gabor wavelet filter banks that consider all different frequency bands and phase directions:

[0009] A1. This set of filters includes 100 Gabor wavelet functions. Each wavelet has a different frequency characteristic distribution. All wavelets occupy the entire frequency range from low to high, ensuring comprehensive coverage of each frequency band.

[0010] A2. The dimension of the designed Gabor wavelet is 100, which exhibits narrow bandwidth characteristics in the frequency domain and can extract features from a fixed information frequency band;

[0011] A3. Arrange 100 Gabor wavelet functions into a 100×100 matrix by row as the initial weight matrix of the autoencoder;

[0012] A4. Select several gears with different health conditions as training samples and label them with corresponding clear labels. Then, the vibration acceleration sensors installed on the gear base are used to collect the vibration acceleration signals generated by these gears in different health conditions during the variable speed rotation process.

[0013] A5. Perform multiple overlapping random segmentation processing on the signal samples to construct a training sample set;

[0014] A6. Use a simplified autoencoder model iteration number and training sample set size, directly input the training sample set into the autoencoder model for network training, and thus efficiently obtain the trained weight matrix W;

[0015] A7. Use the trained weight matrix W to map the sample set, and use the ReLU function as the activation function of the network to accurately capture the local feature information of the sample:

[0016] );

[0017] A8. Obtain the final feature f of the sample by averaging all local features i :

[0018] ;

[0019] A9. Extract all sample features f i The corresponding labels are input into the Softmax regression classifier to obtain the training accuracy of the diagnostic model;

[0020] B. Fault diagnosis of the test sample set under gear variable speed:

[0021] B1. Perform average segmentation processing on the test signal samples to construct a test sample set;

[0022] B2. According to steps A7 to A8, the trained autoencoder model is used to extract local features from the segment information in the test sample set, and the extracted feature values ​​are averaged;

[0023] B3. Calculate the classification probability of the test sample through the Softmax regression model, and select the healthy category label closest to the calculated probability through comparison as the healthy category judgment of the test sample.

[0024] Optionally, in step A1, different health states of the gears include normal, sun gear pitting, sun gear dents, sun gear wear, planet gear pitting, planet gear dents, and planet gear wear.

[0025] Optionally, in steps A1 and B1, the gear speed varies between 2000-2500 r / min.

[0026] Optionally, in steps A1 and B1, the data sampling frequency of the gear is 25.6 kHz, 100 samples are collected for each gear health condition, and the data dimension of each sample is 1200.

[0027] Optionally, in step A2, a Z-segment signal is extracted from the training sample by a random segmentation method to form a training set. , where the jth segment contains N in The data points are represented as , N in Represents the input dimension of the autoencoder, N out Represents the output dimension.

[0028] Optionally, in step A5, 100 is selected as the input and output dimensions of the autoencoder, the random segment size is set to 100, and the number of random segments is set to 100.

[0029] Optionally, in step A6, the number of iterations of the autoencoder model is 10, the ratio of the training sample set is 10%, and the step of inputting the data set into the autoencoder model to train the weight matrix W includes:

[0030] A61, encoding process:

[0031]

[0032] in, Indicates from The calculated hidden layer coding vector; f is the coding function, the present invention uses the sigmoid function, W is the coding weight matrix, is the bias vector.

[0033] A62, decoding process:

[0034]

[0035] Among them, is the decoding function, The decoded weight matrix, is the bias vector.

[0036] The beneficial effect of the present invention is that the present invention uses 100 Gabor wavelets containing different frequency components to form the initial weight matrix of the automatic encoder to guide the model to extract multi-scale and multi-directional fault features; realizes the interpretability research of the network weights, and improves the accuracy and computational efficiency of gear fault detection under variable speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 10 time domain and frequency domain schematic diagrams of the Gabor wavelet generated by the present invention.

[0038] Figure 2 The present invention is a flow chart of an interpretable intelligent mechanical fault diagnosis method based on Gabor wavelet. DETAILED DESCRIPTION

[0039] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0040] An interpretable mechanical fault intelligent diagnosis method based on Gabor wavelet transform includes the following steps:

[0041] A. Design a set of Gabor wavelet filter banks that consider all different frequency bands and phase directions:

[0042] A1. This set of filters includes 100 Gabor wavelet functions. Each wavelet has a different frequency characteristic distribution. All wavelets occupy the entire frequency range from low to high, ensuring comprehensive coverage of each frequency band.

[0043] A2. The dimension of the designed Gabor wavelet is 100, which exhibits narrow bandwidth characteristics in the frequency domain and can extract features from a fixed information frequency band;

[0044] A3. Arrange 100 Gabor wavelet functions into a 100×100 matrix by row as the initial weight matrix of the autoencoder;

[0045] A4. Select several gears with different health conditions as training samples and label them with corresponding clear labels. Then, the vibration acceleration sensors installed on the gear base are used to collect the vibration acceleration signals generated by these gears in different health conditions during the variable speed rotation process.

[0046] A5. Perform multiple overlapping random segmentation processing on the signal samples to construct a training sample set;

[0047] A6. Use a simplified autoencoder model iteration number and training sample set size, directly input the training sample set into the autoencoder model for network training, and thus efficiently obtain the trained weight matrix W;

[0048] A7. Use the trained weight matrix W to map the sample set, and use the ReLU function as the activation function of the network to accurately capture the local feature information of the sample:

[0049] );

[0050] A8. Obtain the final feature f of the sample by averaging all local features i :

[0051] ;

[0052] A9. Extract all sample features f i The corresponding labels are input into the Softmax regression classifier to obtain the training accuracy of the diagnostic model;

[0053] B. Fault diagnosis of the test sample set under gear variable speed:

[0054] B1. Perform average segmentation processing on the test signal samples to construct a test sample set;

[0055] B2. According to steps A7 to A8, the trained autoencoder model is used to extract local features from the segment information in the test sample set, and the extracted feature values ​​are averaged;

[0056] B3. Calculate the classification probability of the test sample through the Softmax regression model, and select the healthy category label closest to the calculated probability through comparison as the healthy category judgment of the test sample.

[0057] Optionally, in step A1, different health states of the gears include normal, sun gear pitting, sun gear dents, sun gear wear, planet gear pitting, planet gear dents, and planet gear wear.

[0058] Optionally, in steps A1 and B1, the gear speed varies between 2000-2500 r / min.

[0059] Optionally, in steps A1 and B1, the data sampling frequency of the gear is 25.6 kHz, 100 samples are collected for each gear health condition, and the data dimension of each sample is 1200.

[0060] Optionally, in step A2, a Z-segment signal is extracted from the training sample by a random segmentation method to form a training set. , where the jth segment contains N in The data points are represented as , N in Represents the input dimension of the autoencoder, N out Represents the output dimension.

[0061] Optionally, in step A5, 100 is selected as the input and output dimensions of the autoencoder, the random segment size is set to 100, and the number of random segments is set to 100.

[0062] Optionally, in step A6, the number of iterations of the autoencoder model is 10, the ratio of the training sample set is 10%, and the step of inputting the data set into the autoencoder model to train the weight matrix W includes:

[0063] A61, encoding process:

[0064]

[0065] in, Indicates from The calculated hidden layer coding vector; f is the coding function, the present invention uses the sigmoid function, W is the coding weight matrix, is the bias vector.

[0066] A62, decoding process:

[0067]

[0068] Among them, is the decoding function, The decoded weight matrix, is the bias vector.

[0069] Using the method of the present invention, a training accuracy of 100% and a test accuracy of 98.38% can be achieved. In order to verify the effectiveness of the method of the present invention, several other method models are used for comparison, and the comparison results are shown in the following table:

[0070] Table 1

[0071] From the comparison results in the table, it can be seen that the test accuracy of the autoencoder method is 88.01%; the test accuracy of the parallel sparse filtering is 92.77%; the test accuracy of the batch-normalized autoencoder is 95.04%; all of which are lower than the diagnostic accuracy of the method of the present invention.

[0072] The present invention adopts 100 Gabor wavelet lists containing different frequency components to form the initial weight matrix of the automatic encoder to guide the model to extract multi-scale and multi-directional fault features; realizes the interpretability research of the network weights, improves the accuracy and computational efficiency of gear fault detection under variable speed, and shows good effect in solving the problem of gear fault diagnosis.

[0073] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by technicians in this technical field within the essential scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. An interpretable mechanical fault intelligent diagnosis method based on Gabor wavelet transform, characterized in that: The following steps are involved: A. Design a set of Gabor wavelet filter banks that consider all different frequency bands and phase directions: A1. This set of filters includes 100 Gabor wavelet functions. Each wavelet has a different frequency characteristic distribution. All wavelets occupy the entire frequency range from low to high, ensuring comprehensive coverage of each frequency band. A2. The dimension of the designed Gabor wavelet is 100, which exhibits narrow bandwidth characteristics in the frequency domain and can extract features from a fixed information frequency band; A3. Arrange 100 Gabor wavelet functions into a 100×100 matrix by row as the initial weight matrix of the autoencoder; A4. Select several gears with different health conditions as training samples and label them with corresponding clear labels. Then, the vibration acceleration sensors installed on the gear base are used to collect the vibration acceleration signals generated by these gears in different health conditions during the variable speed rotation process. A5. Perform multiple overlapping random segmentation processing on the signal samples to construct a training sample set; A6. Use a simplified autoencoder model iteration number and training sample set size, directly input the training sample set into the autoencoder model for network training, and thus efficiently obtain the trained weight matrix W; A7. Use the trained weight matrix W to map the sample set, and use the ReLU function as the activation function of the network to accurately capture the local feature information of the sample: )； A8. Obtain the final feature f of the sample by averaging all local features i : ； A9. Extract all sample features f i The corresponding labels are input into the Softmax regression classifier to obtain the training accuracy of the diagnostic model; B. Fault diagnosis of the test sample set under gear variable speed: B1. Perform average segmentation processing on the test signal samples to construct a test sample set; B2. According to steps A7 to A8, the trained autoencoder model is used to extract local features from the segment information in the test sample set, and the extracted feature values ​​are averaged; B3. Calculate the classification probability of the test sample through the Softmax regression model, and select the healthy category label closest to the calculated probability through comparison as the healthy category judgment of the test sample.

2. The interpretable mechanical fault intelligent diagnosis method based on Gabor wavelet transform as claimed in claim 1, characterized in that: In step A1, different health states of the gears include normal, sun gear pitting, sun gear denting, sun gear wear, planet gear pitting, planet gear denting, and planet gear wear.

3. The interpretable mechanical fault intelligent diagnosis method based on Gabor wavelet transform as claimed in claim 1, characterized in that: In steps A1 and B1, the gear speed varies between 2000-2500 r / min.

4. The interpretable mechanical fault intelligent diagnosis method based on Gabor wavelet transform as claimed in claim 1, characterized in that: In steps A1 and B1, the data sampling frequency of the gear is 25.6kHz, 100 samples are collected for each gear health condition, and the data dimension of each sample is 1200.

5. The interpretable mechanical fault intelligent diagnosis method based on Gabor wavelet transform as claimed in claim 1, characterized in that: In step A2, a random segment method is used to extract Z segment signals from the training samples to form a training set. , where the jth segment contains N in The data points are represented as , N in Represents the input dimension of the autoencoder, N out Represents the output dimension.

6. The interpretable mechanical fault intelligent diagnosis method based on Gabor wavelet transform as claimed in claim 1, characterized in that: In step A5, 100 is selected as the input and output dimensions of the autoencoder, the random segment size is set to 100, and the number of random segments is set to 100.

7. The interpretable mechanical fault intelligent diagnosis method based on Gabor wavelet transform as claimed in claim 1, characterized in that: In step A6, the number of iterations of the autoencoder model is 10, the ratio of the training sample set is 10%, and the step of inputting the data set into the autoencoder model to train the weight matrix W includes: A61, encoding process: in, Indicates from The calculated hidden layer coding vector; f is the coding function, the present invention uses the sigmoid function, W is the coding weight matrix, is the bias vector; A62, decoding process: Among them, is the decoding function, The decoded weight matrix, is the bias vector.

Images