Bearing fault diagnosis method for enhancing interpretability of neural network based on spectral correlation

Abstract

The invention discloses a bearing fault diagnosis method for enhancing the interpretability of a neural network based on spectral correlation. The method comprises the following steps: firstly, carrying out spectral correlation calculation based on short-time Fourier transform on a signal; secondly, screening out a spectrum correlation row vector carrying most fault information according to a fault frequency amplitude center data value, and taking the spectrum correlation row vector as an input value of a neural network; designing an ejaculation network structure which comprises at least one convolution and pooling layer, a full connection layer and a Dropout mechanism, and finally realizing multi-category fault state output through Softmax; and finally, training the neural network, and obtaining a trained network model after iterative convergence. According to the method, fault features are enhanced based on a spectral correlation method, an amplitude center data value at a fault frequency is used as an evaluation parameter, fault information is embedded in a training sample in the form, a spectral correlation matrix is screened, the model interpretability is improved, the robustness of the model is optimized, and the accuracy of the model is improved. And fault diagnosis is realized based on the one-dimensional convolutional neural network.

Description

Technical Field

[0001] The present invention belongs to the technical field of fault diagnosis, and in particular relates to a bearing fault diagnosis method based on spectral correlation enhanced neural network interpretability. Background Art

[0002] In all kinds of modern rotating equipment, rolling bearings are widely used in the supporting mechanisms of motors and mechanical systems. They are key components for maintaining transmission smoothness and load-bearing capacity. Such bearings are usually responsible for achieving stable guidance of rotational motion and controlling friction during operation. Although modern bearings fully consider their strength and service life during the design process, they are often affected by the coupling of high loads, high speeds and complex working conditions during long-term operation, which can easily induce fatigue damage, microcrack propagation and structural failures. If they cannot be detected and diagnosed in time, it may cause a decrease in efficiency and increased energy consumption at the least, and may even lead to equipment shutdown and even systemic safety accidents, resulting in non-negligible economic and engineering risks. Therefore, the development of a fault diagnosis method with high recognition rate and good generalization ability has important research and application value. Summary of the Invention

[0003] In order to overcome the shortcomings of the existing technology, the present invention provides a bearing fault diagnosis method based on spectral correlation to enhance the interpretability of neural networks. First, the signal is subjected to spectral correlation calculation based on short-time Fourier transform; secondly, the spectral correlation row vector carrying the most fault information is screened out according to the fault frequency amplitude center data value as the input value of the neural network; next, the ejaculation network structure is designed to include at least one convolution and pooling layer, a fully connected layer and a Dropout mechanism, and finally multi-category fault state output is achieved through Softmax; finally, the neural network is trained, and a trained network model is obtained after iterative convergence. The method of the present invention is based on the spectral correlation method to enhance fault characteristics, and uses the amplitude center data value at the fault frequency as an evaluation parameter. In this way, fault information is embedded in the training samples, the spectral correlation matrix is ​​screened, the model interpretability is improved, and the robustness of the model is optimized, and fault diagnosis is achieved based on a one-dimensional convolutional neural network.

[0004] The technical solutions adopted by the present invention to solve the technical problems are as follows:

[0005] Step 1: Spectral-related feature extraction;

[0006] Step 1-1: Calculate the spectral correlation matrix of the original vibration signal to obtain the correlation expression;

[0007] Step 1-2: Based on the potential bearing fault frequency, locate the corresponding spectrum column in the spectrum correlation matrix;

[0008] Step 1-3: Extract the amplitude center data value at the frequency of each row spectrum correlation vector as the evaluation index;

[0009] Step 1-4: Determine whether to retain the row vector based on the evaluation value, and filter out the spectral correlation rows containing high fault information;

[0010] Step 1-5: Use the filtered spectral correlation rows as feature samples and repeat steps 1-1 to 1-4 for multiple vibration signals to construct a complete training dataset;

[0011] Step 2: Fault diagnosis;

[0012] The training data set generated by steps 1-5 is input into a one-dimensional convolutional neural network model to implement fault classification; the one-dimensional convolutional neural network model includes:

[0013] ① At least one convolutional layer and pooling layer to extract local features of different scales;

[0014] ② At least one fully connected layer for feature integration and representation;

[0015] ③Dropout mechanism to suppress overfitting and improve generalization ability;

[0016] ④The output layer uses the Softmax function to predict multi-category fault probabilities.

[0017] Preferably, the step 1 is specifically:

[0018] Step 1: Calculate the signal spectrum correlation density S. The calculation steps are:

[0019] Calculate the short-time Fourier transform, the formula is:

[0020]

[0021] Where τ is the time delay, x(t) is the vibration signal, w(t-τ) is the window function, and X(τ,f) is the short-time Fourier transform coefficient of the signal at time delay τ and frequency f;

[0022] The calculation formula of spectral correlation density S is as follows:

[0023] S(f,α0,τ)=X(f+α0,τ)·X * (f,τ)

[0024]

[0025] Where N T is the number of time frames, ||ω|| 2 is the window function energy, I ω is the center index of the window function, α0 is the current cycle frequency, X(f,τ) is the short-time Fourier transform coefficient spectrum correlation matrix of the signal at frequency f and time τ; Fs is the sampling frequency, X * (f,τ) is the complex conjugate of X(f,τ), represents the Fourier transform that converts time t into the cyclic frequency α domain;

[0026] Step 2: Calculate the evaluation parameters of the spectral correlation matrix, that is, the central data value R[n] of the amplitude at the fault frequency. The calculation formula is:

[0027] R[n]=E[N f ]-u

[0028] Where E[N f ] is the amplitude corresponding to the fault frequency in each mode envelope spectrum, μ is the mean of the row vector;

[0029] Step 3: Calculate the row vector with the largest central data value as the output result, that is, the enhanced fault feature.

[0030] A computer program enables a computer to execute the bearing fault diagnosis method.

[0031] An electronic device comprises: a processor and a memory; the memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory, so that the electronic device performs the above-mentioned bearing fault diagnosis method.

[0032] A computer-readable storage medium stores a computer program, which implements the above-mentioned bearing fault diagnosis method when executed by a processor.

[0033] A chip includes: a processor for calling and running a computer program from a memory, so that a device equipped with the chip executes the above-mentioned bearing fault diagnosis method.

[0034] A computer program product includes a computer storage medium storing a computer program, wherein the computer program includes instructions executable by at least one processor, and when the instructions are executed by the at least one processor, the above-mentioned bearing fault diagnosis method is implemented.

[0035] The beneficial effects of the present invention are as follows:

[0036] This invention combines a physics-driven feature enhancement strategy with a data-driven deep learning model, effectively improving the diagnostic accuracy and stability of the fault diagnosis system under low signal-to-noise ratio, complex background and multiple fault modes, and has good engineering promotion and application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a flow chart of the method of the present invention;

[0038] Figure 2 It is a signal with fault information according to an embodiment of the present invention;

[0039] Figure 3 is the spectral correlation matrix of an embodiment of the present invention;

[0040] Figure 4 is the amplitude center data value at the row vector fault frequency in an embodiment of the present invention;

[0041] Figure 5 is the feature extraction output result of the embodiment of the present invention;

[0042] Figure 6 This is the confusion matrix result of the verification set of the embodiment of the present invention. DETAILED DESCRIPTION

[0043] The present invention will be further described below with reference to the accompanying drawings and examples.

[0044] This paper provides a bearing diagnosis method based on fault information to enhance the interpretability of neural networks. This method aims to address the shortcomings of existing technologies, such as insufficient bearing fault feature extraction capabilities, unstable classification performance, and poor model interpretability in environments with high interference and low signal-to-noise ratios. By introducing a fault frequency-guided spectral correlation row filtering mechanism, it effectively enhances the explicit representation of key fault information in the input data. Combined with the deep modeling capabilities of convolutional neural networks, this method achieves high-precision identification of rolling bearing faults.

[0045] This invention effectively enhances the interpretability and recognition accuracy of neural network models through a spectral correlation and physical feature guidance mechanism. The method first performs spectral correlation calculations on the vibration signal to construct a spectral correlation matrix of joint features. Subsequently, at the spectral location of the typical bearing fault frequency, the amplitude center data value of each spectral correlation row vector at that frequency point is extracted as an evaluation metric to determine whether the row vector contains critical fault information. By quantifying the distribution of the amplitude center data values, informative spectral correlation rows are selected, forming a highly representative training sample with clear physical meaning.

[0046] The training data generated by the extraction process is then fed into a simple one-dimensional convolutional neural network. This neural network consists of at least one set of convolutional and pooling layers, a fully connected layer, and a dropout module to mitigate overfitting and enhance feature extraction robustness. Finally, a softmax activation function is used to achieve multi-classification output for different fault types. During the training phase, an iterative optimization strategy is used to continuously adjust network weights until the model converges, completing the construction of the fault recognition model.

[0047] This approach, combining physics-guided feature extraction with a deep learning architecture, significantly improves the model's ability to identify early, subtle faults. It is particularly well-suited for online condition monitoring and intelligent diagnosis of rolling bearings in industrial environments. The proposed solution not only enhances the credibility and interpretability of diagnostic results but also strengthens the model's stability and practicality in complex noise environments.

[0048] The present invention designs the following technical solutions:

[0049] (1) Spectral correlation feature extraction: used to extract the most representative fault feature row vector from the vibration signal. The specific operation is as follows:

[0050] ① Calculate the spectral correlation matrix of the original vibration signal to obtain the correlation expression;

[0051] ② Based on the potential bearing fault frequency, locate the corresponding spectrum column in the spectrum correlation matrix;

[0052] ③ Extract the amplitude center data value at the frequency of each row of spectrum correlation vector as an evaluation indicator;

[0053] ④Determine whether to retain the row vector based on the evaluation value, and filter out the spectral correlation rows containing high fault information;

[0054] ⑤ Use the filtered spectrum-related rows as feature samples and repeat the above steps to construct a complete training dataset.

[0055] (2) Fault diagnosis:

[0056] The enhanced feature signal is input into a one-dimensional convolutional neural network model to achieve fault classification. The one-dimensional convolutional neural network model includes:

[0057] ① At least one convolutional layer and pooling layer to extract local features of different scales;

[0058] ② At least one fully connected layer for feature integration and representation;

[0059] ③Dropout mechanism to suppress overfitting and improve generalization ability;

[0060] ④The output layer uses the Softmax function to predict multi-category fault probabilities.

[0061] The network parameters of the one-dimensional convolutional neural network model, such as the convolution kernel size, the number of convolution kernels, the pooling parameters, the number of neurons in the fully connected layer, and the Dropout discarding rate, can all be flexibly adjusted according to different signal dimensions, fault mode types, and training requirements, without affecting the basic technical effects of the method of the present invention, and should all fall within the scope of protection of the present invention.

[0062] Example:

[0063] Figure 1 This is a flow chart of a bearing diagnosis method that enhances the interpretability of a neural network based on fault information. In this embodiment, the method is implemented using the Matlab software platform. The method is demonstrated and explained based on experimental signals.

[0064] Figure 2 is the experimental signal with fault, Figure 3 is the fault signal spectrum correlation matrix, Figure 4 Evaluation parameters for each row vector, Figure 5 is the output result of feature extraction, Figure 6 Confusion matrix results for the fault diagnosis module validation set.

[0065] Based on this, in this embodiment, the method includes the following steps:

[0066] (1) Signal decomposition module based on marginal spectrum extreme value distribution:

[0067] The first step is to calculate the signal spectrum correlation density S. The calculation steps are:

[0068] Calculate the short-time Fourier transform, the formula is:

[0069]

[0070] like Figure 2 As shown, the window size of the short-time Fourier transform is set to 100 and the sampling frequency f s Set to 25600. The calculation formula of spectral correlation density S is as follows:

[0071] S(f,α0,τ)=X(f+α0,τ)·X * (f,τ)

[0072]

[0073] Where N T is the number of time frames, ||ω|| 2 is the window function energy, I ω is the center index of the window function, α0 is the current cycle frequency, and X(f,τ) is the short-time Fourier transform coefficient spectrum correlation matrix of the signal at frequency f and time τ. Figure 3 shown.

[0074] The second step is to calculate the evaluation parameters of the spectral correlation matrix, that is, the central data value R[n] of the amplitude at the fault frequency. The calculation formula is:

[0075] R[n]=E[N f ]-u

[0076] Where E[N f] is the amplitude corresponding to the fault frequency in each mode envelope spectrum, and μ is the mean value of the row vector. Figure 4 shown.

[0077] The third step is to calculate the row vector with the largest central data value as the output result, that is, the enhanced fault feature, as shown in the following example: Figure 5 shown.

[0078] (3) Fault diagnosis module, that is, fault diagnosis is realized based on one-dimensional convolutional neural network. The confusion matrix of the validation set is as follows: Figure 6 As shown in the figure, 0 represents "healthy state", 1 represents "inner ring fault", and 2 represents "outer ring fault".

Claims

1. A bearing fault diagnosis method based on spectral correlation enhanced neural network interpretability, characterized in that: The steps include: Step 1: Spectral-related feature extraction; Step 1-1: Calculate the spectral correlation matrix of the original vibration signal to obtain the correlation expression; Step 1-2: Based on the potential bearing fault frequency, locate the corresponding spectrum column in the spectrum correlation matrix; Step 1-3: Extract the amplitude center data value at the frequency of each row spectrum correlation vector as the evaluation index; Step 1-4: Determine whether to retain the row vector based on the evaluation value, and filter out the spectral correlation rows containing high fault information; Step 1-5: Use the filtered spectral correlation rows as feature samples and repeat steps 1-1 to 1-4 for multiple vibration signals to construct a complete training dataset; Step 2: Fault diagnosis; The training data set generated by steps 1-5 is input into a one-dimensional convolutional neural network model to implement fault classification; the one-dimensional convolutional neural network model includes: ① At least one convolutional layer and pooling layer to extract local features of different scales; ② At least one fully connected layer for feature integration and representation; ③Dropout mechanism to suppress overfitting and improve generalization ability; ④The output layer uses the Softmax function to predict multi-category fault probabilities.

2. The bearing fault diagnosis method based on spectrum correlation enhanced neural network interpretability according to claim 1 is characterized in that: The step 1 is specifically as follows: Step 1: Calculate the signal spectrum correlation density S. The calculation steps are: Calculate the short-time Fourier transform, the formula is: Where τ is the time delay, x(t) is the vibration signal, w(t-τ) is the window function, and X(τ,f) is the short-time Fourier transform coefficient of the signal at time delay τ and frequency f; The calculation formula of spectral correlation density S is as follows: S(f,α0,τ)=X(f+α0,τ)·X * (f,t) Where N T is the number of time frames, ||ω|| 2 is the window function energy, I ω is the center index of the window function, α0 is the current cycle frequency, X(f,τ) is the short-time Fourier transform coefficient spectrum correlation matrix of the signal at frequency f and time τ; F s is the sampling frequency, X * (f,τ) is the complex conjugate of X(f,τ), represents the Fourier transform that converts time t into the cyclic frequency α domain; Step 2: Calculate the evaluation parameters of the spectral correlation matrix, that is, the central data value R[n] of the amplitude at the fault frequency. The calculation formula is: R[n]=E[N f ]-u Where E[N f ] is the amplitude corresponding to the fault frequency in each mode envelope spectrum, μ is the mean of the row vector; Step 3: Calculate the row vector with the largest central data value as the output result, that is, the enhanced fault feature.

3. A computer program, characterized in that The computer program enables a computer to execute the method according to any one of claims 1 to 2.

4. An electronic device, characterized in that: include: processor and memory; The memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory, so that the electronic device performs the method according to any one of claims 1 to 2.

5. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 2 is implemented.

6. A chip, characterized in that: include: A processor, configured to call and run a computer program from a memory, so that a device equipped with the chip executes the method according to any one of claims 1 to 2.

7. A computer program product, characterized in that The computer program product comprises a computer storage medium storing a computer program, wherein the computer program comprises instructions executable by at least one processor, and when the instructions are executed by the at least one processor, the method according to any one of claims 1 to 2 is implemented.

Images