Multi-node deep neural network bearing fault feature extraction method based on knowledge guidance

Abstract

The invention relates to bearing fault feature extraction, belongs to the technical field of fault prediction and health management (PHM), and particularly relates to a multi-node deep neural network bearing fault feature extraction method based on knowledge guidance, which comprises the following steps: S1, bearing fault data acquisition; s2, data preprocessing; and S3, multi-node deep neural network structure training and fault feature extraction are carried out, and bearing state identification is carried out. According to the invention, fault feature extraction can be well carried out.

Description

TECHNICAL FIELD

[0001] The present application relates to bearing fault feature extraction, belonging to the technical field of fault prediction and health management (PHM), and in particular to a kind of bearing fault feature extraction method based on knowledge guide multi-node deep neural network. BACKGROUND

[0002] Fault feature extraction has important role in distinguishing train abnormal condition, guaranteeing train safety.But the objective function of traditional deconvolution method is mostly constructed on time domain, sensitive to random pulse, and the iteration initial value set by experience cannot guarantee its applicability in low signal-to-noise ratio scene.Deconvolution algorithm can self-adaptively adjust filter coefficient to realize the extraction enhancement of fault feature, and is considered as the most effective tool to detect and extract repeated transient impact representing fault.In addition, deconvolution can be further extended to unsupervised feature learning, but it still has the following problems: (1) many deconvolution detection algorithms do not consider the global optimal solution of non-convex objective function, and local optimal solution may not extract fault feature in low signal-to-noise ratio scene.(2) the performance of many bearing fault detection algorithms depends on prior knowledge, such as accurate fault feature frequency, but in engineering practice, accurate fault feature frequency is not easy to obtain.(3) some health indicators representing bearing fault only consider impulsiveness, without considering periodicity, resulting in sensitivity to random pulse. SUMMARY

[0003] The present application provides a kind of bearing fault feature extraction method based on knowledge guide multi-node deep neural network, which can overcome some or some defects of prior art.

[0004] According to the bearing fault feature extraction method based on knowledge guide multi-node deep neural network of the present application, the following steps are included: S1, bearing fault data acquisition; S2, data preprocessing; S3, multi-node deep neural network structure training and fault feature extraction, and bearing state identification.

[0005] Preferably, in S2, the collected signal is processed by sliding window interception, and the length of data collected per second is used as the window length;When the collected signal is variable speed data, angle domain resampling processing is carried out.

[0006] Preferably, in S3, the multi-node deep neural network structure includes an input layer, an output layer and two hidden layers; The connection weight coefficient between the input layer and the first hidden layer is the filter coefficient corresponding to the 1 / 3 binary tree division subband, wherein the number of nodes of the first hidden layer is consistent with the number of subbands; After the signal is processed in the input layer, m initial filter signals are obtained; Then, the weight coefficients between the two hidden layers are randomly initialized to further extract fault features; After the signal passes through the two hidden layers, m filter signals are still obtained, and the filter coefficients are continuously updated using the gradient descent algorithm until the gradient is 0 or the set number of iterations is reached; Finally, the optimal filter signal is selected in the output layer for envelope demodulation analysis to identify the bearing state.

[0007] As a preferred embodiment, the specific steps of S3 are as follows: S3.1, divide the frequency band according to the 1 / 3 binary tree structure to obtain the filter coefficients g corresponding to each sub-band; S3.2, construct a Hankel matrix to obtain the initial filter signal according to the filter coefficients; S3.3, initialize the number of iterations; S3.4, the initial filter signal is subjected to Hilbert transform, square envelope transform and FFT transform to obtain the square envelope spectrum; S3.5, construct the objective function; S3.6, update the filter coefficients according to the gradient descent method until the partial derivative of the filter with respect to the objective function is 0 or the maximum number of iterations is reached; S3.7, obtain the optimal filter coefficients and the filtered signal; S3.8, perform envelope demodulation analysis on the filtered signal to determine the fault type.

[0008] As a preferred embodiment, the mathematical expression of the objective function is as follows: (1) In the formula, denotes the time-domain mean kurtosis, denotes the square envelope spectrum L2 / L1 norm, denotes the filter signal, denotes the weight coefficient; is a sign term used to convert the maximum optimization problem into a minimum optimization problem; The mean kurtosis is calculated as follows: (2) In the formula, denotes the number of sub-signals divided by the sliding window method, denotes the kurtosis value of each sub-signal, which is calculated by the following formula: (3) In the formula, denotes the length of each sub-signal; represents the sample mean; i represents the sample mean; wherein, is mathematically expressed as follows: (4) wherein, represents the square envelope spectrum of the filtered signal, represents the length of the square envelope spectrum; The bearing fault feature extraction is converted into a target function solving problem, and the mathematical expression of the target function to be optimized is: (5) wherein, represents the real part of the Fourier matrix, represents the imaginary part of the Fourier matrix, represents the Hankel matrix, is the filter coefficient.

[0009] As a preferred, the target function optimization solving method is: The original signal is filtered by constructing a feature matrix, and the filtering process is represented as: (6) wherein, represents the Hankel matrix, represents the collected vibration signal, is the length of the collected signal, represents the filter length, represents the filter coefficient to be solved; The Hilbert transform (HT transform) of the filtered signal is represented as: (7) (8) wherein, represents the analytical signal after HT transform, represents the convolution matrix, is a circulant matrix constructed according to the filter represents the transposed matrix, represents the length of the filter; The square envelope of the analytical signal is represented as: (9) The square envelope spectrum of the analytical signal is represented as: (10)​​ In the formula, , The Fourier matrix is ​​expressed mathematically as follows: (11) In the formula: This indicates the maximum frequency of the signal, which is determined based on the signal's sampling frequency. Based on the above derivation, the squared envelope spectrum of the signal is obtained through step-by-step calculations, and its mathematical expression is: (12) The objective function is solved using the gradient descent method, and partial derivatives are obtained through the chain rule. Specifically, the partial derivative of the objective function with respect to the filter is calculated using the following formula: (13) In the formula: , , , ,

[0010] , , , Represents the identity matrix.

[0011] The beneficial effects of this invention are as follows: First, this invention proposes a novel knowledge-guided multi-node deep neural network structure to extract fault features. The node settings and weight coefficients of the network have practical physical meaning, which increases the interpretability of fault feature extraction.

[0012] Secondly, this invention provides an iterative initial value setting strategy close to the resonance band, which can effectively improve the applicability of the proposed method in low signal-to-noise ratio scenarios.

[0013] Furthermore, the construction of the objective function takes into account the properties of fault characteristics in both the time domain and the envelope spectrum, and the fault characteristics are more comprehensively characterized through weighted combination.

[0014] Finally, the applicability of the proposed method was tested in constant speed / variable speed fault data and early fault data. The results show that the proposed method still has good robustness even in low signal-to-noise ratio acquired data containing random pulses. Attached Figure Description

[0015] Figure 1This is a flowchart of a knowledge-guided multi-node deep neural network-based bearing fault feature extraction method in an embodiment. Figure 2 This is a schematic diagram illustrating the setting of initial values ​​for iteration in the embodiment. Detailed Implementation

[0016] To further understand the content of this invention, a detailed description of the invention will be provided in conjunction with the accompanying drawings and embodiments. It should be understood that the embodiments are merely illustrative and not limiting of the invention.

[0017] Example like Figure 1 As shown in the figure, this embodiment provides a knowledge-guided multi-node deep neural network-based bearing fault feature extraction method, which includes the following steps: S1. Bearing fault data acquisition; Simulation data from bearing fault models (different operating conditions, bearing bench tests, different fault types), high-speed train bench tests, and publicly available bearing fault data provide data support for the verification of the proposed method. Specifically, this includes bearing vibration signals. y Maximum number of iterations k Filter length l .

[0018] S2, Data Preprocessing; The acquired signals are processed using a sliding window capture method, with the length of data acquired per second as the window length; when the acquired signal is variable speed data, angle domain resampling processing is performed.

[0019] S3. Training of multi-node deep neural network structure and extraction of fault features, and identification of bearing status.

[0020] In S3, the multi-node deep neural network structure includes one input layer, one output layer, and two hidden layers; The connection weight coefficients between the input layer and the first hidden layer are the filter coefficients corresponding to the subband division of the 1 / 3 binary tree, where the number of nodes in the first hidden layer is the same as the number of subbands; After the signal is processed in the input layer, m initial filtered signals are obtained; Then, the weight coefficients between the two hidden layers are randomly initialized to further extract fault features; After the signal passes through two hidden layers, m filtered signals are still obtained. The gradient descent algorithm is used to continuously update the filter coefficients until the gradient is 0 or the set number of iterations is reached. Finally, the optimal filtered signal is selected at the output layer for envelope demodulation analysis to identify the bearing condition.

[0021] The specific steps for S3 are as follows: S3.1 Divide the frequency bands according to a 1 / 3 binary tree structure and obtain the filter coefficients g corresponding to each sub-band; the filter coefficients can be obtained using the "fir1" function in Matlab; S3.2 Construct the Hankel matrix and obtain the initial filtered signal based on the filter coefficients; sequentially use the sub-band filter coefficients as g in Formula 6 to obtain the initial filtered signal; S3.3, Initialize the number of iterations; S3.4 The initial filtered signal is transformed by Hilbert transform, square envelope transform and FFT transform to obtain the square envelope spectrum; S3.5 Construct the objective function; S3.6 Update the filter coefficients according to the gradient descent method (Formulas 6 to 13) until the partial derivative of the objective function with respect to the filter is 0 (Formula 13 equals 0) or the maximum number of iterations is reached; S3.7. Obtain the optimal filter coefficients according to Formula 5, and obtain the filtered signal according to Formula 6. S3.8 Perform envelope demodulation analysis on the filtered signal to determine the fault type.

[0022] Iteration initial value setting The initial value of the iteration affects the performance of deconvolution. The initial value of the iteration is close to the resonant frequency band, which can obtain better fault extraction results, but the resonant frequency band cannot be obtained in advance.

[0023] Therefore, this embodiment utilizes the sub-bands divided by the filter bank according to the 1 / 3 binary tree structure to obtain the corresponding filter coefficients, and sets the initial values ​​for iteration accordingly. The corresponding sub-band numbers are as follows: Figure 2 As shown in part (a), the number of layers is set to 3, which yields 24 sub-bands, and the corresponding filter coefficients are used as the initial values ​​for iteration.

[0024] Figure 2 Part (b) is a schematic diagram of the set filter bank. For ease of illustration, some overlapping filter banks are not drawn, but the set filter bank can cover the resonant frequency band, thereby increasing the possibility of the BD algorithm obtaining the global optimal solution.

[0025] Since the frequency band division structure based on the 1 / 3 binary tree can obtain the approximate location of the resonance band (RFB), there exists an initial value for iteration close to the RFB.

[0026] objective function The mathematical expression is: (1) In the formula, Indicates the time-domain mean kurtosis. Denotes the L2 / L1 norm of the squared envelope spectrum. Indicates the filtered signal. This represents the weighting coefficient, set to 0.4 or 0.6. It is a symbolic term used to transform a maximum optimization problem into a minimum optimization problem; The mean kurtosis is calculated as follows: (2) In the formula, This indicates the number of sub-signals segmented using the sliding window method. The kurtosis value of each sub-signal is calculated using the following formula: (3) In the formula: Indicates the length of each sub-signal; Indicates the sample number i One data point, This represents the sample mean; in, The mathematical expression is as follows: (4) In the formula, This represents the squared envelope spectrum of the filtered signal. Indicates the length of the squared envelope spectrum; The problem of extracting bearing fault features is transformed into solving an objective function. The mathematical expression of the objective function to be optimized is as follows: (5) In the formula: Denotes the real part of the Fourier matrix. Denotes the imaginary part of the Fourier matrix. Represents the Hankel matrix. These are the filter coefficients.

[0027] The method for optimizing the objective function is as follows: Because the "FFT" and "Hilbert" functions encapsulated in MATLAB involve complex number operations, they cannot be used directly when iterating solutions using the gradient descent strategy.

[0028] Therefore, a time-domain synthesis method is used to perform Hilbert transform (HT) on the signal.

[0029] The SES can then be obtained by multiplying the Fourier matrix and the analytic signal envelope matrix.

[0030] Since all the matrices or vectors obtained by this transformation belong to the real number field, it becomes very convenient to solve the objective function using the gradient descent method.

[0031] The original signal is filtered by constructing a feature matrix. The filtering process is represented as follows: (6) In the formula: Represents the Hankel matrix. This indicates the collected vibration signals. The length of the acquired signal, Indicates the filter length. This represents the filter coefficients that need to be solved; The Hilbert transform (HT transform) of the filtered signal is expressed as: (7) (8) In the formula: This represents the analytic signal after the HT transformation. Represents the convolution matrix. It is based on the filter Constructed circular matrix, Represents the transpose matrix. Indicates the length of the filter; Square envelope of analytic signal Represented as: (9) Analyzing the square envelope spectrum of a signal Represented as: (10) In the formula, , The Fourier matrix is ​​expressed mathematically as follows: (11) In the formula: This indicates the maximum frequency of the signal, which is determined based on the signal's sampling frequency. Based on the above derivation, the squared envelope spectrum of the signal is obtained through step-by-step calculations, and its mathematical expression is: (12) It should be noted that the differentiation based on the chain rule is based on the real number field, while in the above formula... , Both are real-field matrices. The gradient descent algorithm can be used to solve for the filter coefficients, and then the original signal can be filtered to achieve decoupling and separation of fault components.

[0032] The objective function is solved using the gradient descent method, and partial derivatives are obtained through the chain rule. Specifically, the partial derivative of the objective function with respect to the filter is calculated using the following formula: (13) In the formula: , , , ,

[0033] , , , Represents the identity matrix.

[0034] This embodiment mainly involves defining an objective function that considers both the time domain of the filtered signal and the fault characteristics in the ES.

[0035] Subsequently, the gradient descent algorithm is used to solve the objective function to determine the optimal filter coefficients and achieve the separation of fault components from the original signal.

[0036] Ultimately, by analyzing the SES of the filtered signal, the fault type can be identified.

[0037] This embodiment introduces multiple initial values ​​covering the resonance zone to form a multi-node network structure for learning fault features. The innovation of this embodiment lies primarily in proposing a novel strategy for obtaining near-global optimal solutions using a multi-node network structure. This method can adaptively select the initial iteration values ​​and the number of nodes required to extract fault features, without relying on prior knowledge. It expands the generalization ability of unsupervised sparse feature learning in low signal-to-noise ratio scenarios, making it more suitable for actual high-speed train operating conditions. The main contributions are reflected in the following aspects: (1) A new strategy for obtaining an approximate estimate of the global optimal solution is provided. The initial value of the iteration and the number of nodes can be adaptively set according to the properties of the acquired signal itself, avoiding the difficulties of pre-setting and effectively expanding its applicability to extract fault features in low signal-to-noise ratio scenarios.

[0038] (2)(2) The adaptive setting of the initial value of the iteration and the number of nodes has physical significance. The multi-node network structure ensures that the initial value of the iteration in the node covers the bearing resonance zone.

[0039] (3)(3) Without relying on precise prior fault characteristic frequency or fault cycle detection technology (which does not increase the computational burden), it is possible to accurately extract fault characteristics from noise pollution signals (including random pulses).

[0040] The present invention and its embodiments have been described above illustratively. This description is not restrictive, and the figures shown are only one embodiment of the present invention; the actual structure is not limited thereto. Therefore, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the present invention, such designs should fall within the protection scope of the present invention.

Claims

1. A knowledge-guided multi-node deep neural network-based bearing fault feature extraction method, characterized in that: Includes the following steps: S1. Bearing fault data acquisition; S2, Data Preprocessing; S3. Training of multi-node deep neural network structure and extraction of fault features, and identification of bearing status.

2. The knowledge-guided multi-node deep neural network-based bearing fault feature extraction method according to claim 1, characterized in that: In S2, a sliding window method is used to process the acquired signal, with the length of data acquired per second as the window length; when the acquired signal is variable speed data, angle domain resampling processing is performed.

3. The knowledge-guided multi-node deep neural network-based bearing fault feature extraction method according to claim 2, characterized in that: In S3, the multi-node deep neural network structure includes one input layer, one output layer, and two hidden layers; The connection weight coefficients between the input layer and the first hidden layer are the filter coefficients corresponding to the subband division of the 1 / 3 binary tree, where the number of nodes in the first hidden layer is the same as the number of subbands; After the signal is processed in the input layer, m initial filtered signals are obtained; Then, the weight coefficients between the two hidden layers are randomly initialized to further extract fault features; After the signal passes through two hidden layers, m filtered signals are still obtained. The gradient descent algorithm is used to continuously update the filter coefficients until the gradient is 0 or the set number of iterations is reached. Finally, the optimal filtered signal is selected at the output layer for envelope demodulation analysis to identify the bearing condition.

4. The knowledge-guided multi-node deep neural network-based bearing fault feature extraction method according to claim 3, characterized in that: The specific steps for S3 are as follows: S3.

1. Divide the frequency bands according to a 1 / 3 binary tree structure and obtain the filter coefficients g corresponding to each sub-band; S3.2 Construct the Hankel matrix and obtain the initial filtered signal based on the filter coefficients; S3.3, Initialize the number of iterations; S3.4 The initial filtered signal is transformed by Hilbert transform, square envelope transform and FFT transform to obtain the square envelope spectrum; S3.5 Construct the objective function; S3.6 Update the filter coefficients using the gradient descent method until the partial derivative of the objective function with respect to the filter is 0 or the maximum number of iterations is reached. S3.

7. Obtain the optimal filter coefficients and the filtered signal; S3.8 Perform envelope demodulation analysis on the filtered signal to determine the fault type.

5. The knowledge-guided multi-node deep neural network-based bearing fault feature extraction method according to claim 4, characterized in that: objective function The mathematical expression is: (1) In the formula, Indicates the time-domain mean kurtosis. Denotes the L2 / L1 norm of the squared envelope spectrum. Indicates the filtered signal. Indicates the weighting coefficient; It is a symbolic term used to transform a maximum optimization problem into a minimum optimization problem; The mean kurtosis is calculated as follows: (2) In the formula, This indicates the number of sub-signals segmented using the sliding window method. The kurtosis value of each sub-signal is calculated using the following formula: (3) In the formula: Indicates the length of each sub-signal; Indicates the sample number i One data point, This represents the sample mean; in, The mathematical expression is as follows: (4) In the formula, This represents the squared envelope spectrum of the filtered signal. Indicates the length of the squared envelope spectrum; The problem of extracting bearing fault features is transformed into solving an objective function. The mathematical expression of the objective function to be optimized is as follows: (5) In the formula: Denotes the real part of the Fourier matrix. Denotes the imaginary part of the Fourier matrix. Represents the Hankel matrix. These are the filter coefficients.

6. The knowledge-guided multi-node deep neural network-based bearing fault feature extraction method according to claim 5, characterized in that: The method for optimizing the objective function is as follows: The original signal is filtered by constructing a feature matrix. The filtering process is represented as follows: (6) In the formula: Represents the Hankel matrix. This indicates the collected vibration signals. The length of the acquired signal, Indicates the filter length. This represents the filter coefficients that need to be solved; The HT transform of the filtered signal is expressed as: (7) (8) In the formula: This represents the analytic signal after the HT transformation. Represents the convolution matrix, It is based on the filter Constructed circular matrix, Represents the transpose matrix. Indicates the length of the filter; Square envelope of analytic signal Represented as: (9) Square envelope spectrum of analytical signal Represented as: (10) In the formula, , The Fourier matrix is ​​expressed mathematically as follows: (11) In the formula: This indicates the maximum frequency of the signal, which is determined based on the signal's sampling frequency. Based on the above derivation, the squared envelope spectrum of the signal is obtained through step-by-step calculations, and its mathematical expression is: (12) The objective function is solved using the gradient descent method, and partial derivatives are obtained through the chain rule. Specifically, the partial derivative of the objective function with respect to the filter is calculated using the following formula: (13) In the formula: , , , , ； , , , Represents the identity matrix.

Images