Gear reducer motor fault diagnosis method based on acoustic emission

Abstract

The invention provides a gear reducer motor fault diagnosis method based on acoustic emission. The method is applied to the technical field of motor fault diagnosis, and comprises the following steps: acquiring an original acoustic emission signal when a gear reducer motor operates; performing multi-dimensional preprocessing on the acoustic emission signals, extracting signal features, and constructing an input feature matrix according to the extracted signal features and the original acoustic emission signals; constructing a machine learning network model comprising a gating circulation unit and an attention mechanism, inputting the input feature matrix into the trained machine learning network model, and outputting a classification result; calculating a Lorentz curve and a Gini coefficient G of the amplitude of the original acoustic emission signal; and carrying out weighted fusion on a classification result and the Gini coefficient G, and outputting a final fault decision value. According to the invention, the fault diagnosis precision and prediction capability of the gear reducer motor are significantly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of motor fault diagnosis, and particularly to a fault diagnosis method for a gear reduction motor based on acoustic emission. Background Art

[0002] Gear reduction motors are widely used in industrial equipment, automation systems and other fields. However, during long-term operation, key components such as gears and bearings may fail due to factors such as wear, fatigue, and impact, which will affect the operation efficiency and safety of the equipment. Traditional fault diagnosis methods (such as vibration analysis and temperature monitoring) are limited by low-frequency response and installation limitations, and it is difficult to capture early weak fault characteristics. In recent years, acoustic emission technology has received extensive attention in order to achieve efficient and accurate fault diagnosis. Acoustic emission technology can collect high-frequency elastic wave signals during the operation of the equipment in real time, and identify abnormal states inside the equipment through signal analysis. However, traditional acoustic emission signal processing methods often have problems such as insufficient feature extraction, large noise interference, and low fault recognition accuracy when facing complex working conditions and non-stationary signals. The specific manifestations are as follows: Insufficient feature extraction: Traditional time-frequency analysis methods (such as short-time Fourier transform) have low resolution for non-stationary signals and it is difficult to separate fault impact characteristics from background noise.

[0003] Serious noise interference: Electromagnetic interference and mechanical friction noise in the industrial environment result in a low signal-to-noise ratio of the signal, and traditional threshold denoising is likely to cause distortion of effective signals.

[0004] Low fault recognition accuracy: The acoustic emission signals during the operation of gear reduction motors have non-stationary and non-linear characteristics, and traditional signal processing methods are difficult to effectively extract their characteristic information, resulting in low fault recognition accuracy.

[0005] Therefore, there is an urgent need for a fault diagnosis method for gear reduction motors that can fully extract the characteristic information in acoustic emission signals, effectively suppress noise interference, improve the accuracy and robustness of fault recognition, and can specifically identify low-frequency fault modes. Summary of the Invention

[0006] The present invention provides a fault diagnosis method for a gear reduction motor based on acoustic emission, which combines time-frequency analysis, modal decomposition, signal reconstruction, fractal theory, cost-sensitive learning and statistical methods, solves the problems of difficult signal feature extraction, serious noise interference, and low fault mode recognition accuracy existing in traditional gear reduction motor fault detection, and provides a new solution for the health monitoring of gear reduction motors.

[0007] To achieve the above object, it is realized through the following technical solutions: The present invention provides a fault diagnosis method for a gear reduction motor based on acoustic emission, comprising the following steps: Collect the original acoustic emission signal during the operation of the gear reduction motor; Perform multi-dimensional preprocessing on the acoustic emission signal, extract signal features, and construct an input feature matrix based on the extracted signal features and the original acoustic emission signal; Construct a machine learning network model including a gated recurrent unit and an attention mechanism, and input the input feature matrix into the trained machine learning network model to output a classification result; Calculate the Lorentz curve and Gini coefficient of the amplitude of the original acoustic emission signal; Fusion the classification result and the Gini coefficient with weights to output the final fault decision value.

[0008] Further, among them, performing multi-dimensional preprocessing on the acoustic emission signal, extracting signal features, and constructing an input feature matrix based on the extracted signal features and the original acoustic emission signal includes: Decompose the original acoustic emission signal by wavelet packet transform to obtain a time-frequency signal, and calculate the signal time-frequency coherence and time-frequency energy change rate according to the time-frequency signal; Decompose the original acoustic emission signal into subspaces with different resolutions based on multi-resolution analysis to extract signal components in different frequency ranges; Perform Hilbert dynamic mode decomposition on the time-frequency signal to extract the active mode and obtain the dynamic mode decomposition result; Perform adaptive denoising and reconstruct the signal on the time-frequency signal using Stein's unbiased risk estimate to obtain a reconstructed acoustic emission signal; Perform Henon mapping and Sierpinski triangle mapping on the reconstructed acoustic emission signal in sequence, and then perform Weierstrass transform to generate fractal feature data; Integrate the original occurrence signal, the time-frequency signal, the signal time-frequency coherence, the time-frequency energy change rate, the signal components in different frequency ranges, the dynamic mode, the reconstructed acoustic emission signal, and the fractal feature data to construct an input feature matrix for classification.

[0009] Further, among them, the step of decomposing the original acoustic emission signal by wavelet packet transform to obtain a time-frequency signal, and calculating the signal time-frequency coherence and time-frequency energy change rate according to the time-frequency signal includes: Decompose the original acoustic emission signal by wavelet packet transform, and recursively decompose the original acoustic emission signal to obtain sub-signals in different frequency bands to extract signal components in different frequency ranges, and obtain a time-frequency signal representation; Based on the time-frequency signal representation, obtain the time-frequency signal matrix after wavelet packet transform; Calculate the signal time-frequency coherence and the time-frequency energy change rate from the time-frequency signal matrix.

[0010] Further, wherein decomposing the original acoustic emission signal into subspaces with different resolutions based on multi-resolution analysis to extract signal components in different frequency ranges includes: Select a wavelet basis function and a decomposition scale; Perform recursive decomposition on the original acoustic emission signal, and gradually extract the time-frequency local features of the original acoustic emission signal, thereby obtaining the approximation signal and the detail signal at each scale; Reconstruct the signal components in different frequency ranges according to the approximation signal and the detail signal.

[0011] Further, wherein perform Hilbert dynamic mode decomposition on the time-frequency signal, extract the active mode, and obtain the dynamic mode decomposition result, including: Construct a Hilbert matrix according to the time-frequency signal data ; According to the Hilbert matrix Construct a state transition matrix, including: divide the Hilbert matrix into a forward data matrix and a backward data matrix , and satisfy the state transition relationship: , where is the state transition matrix, representing the dynamic evolution characteristics of the signal; According to the forward data matrix and the backward data matrix , through eigenvalue decomposition, obtain the eigenvalue matrix and the dynamic mode matrix; Combine the dynamic mode matrix and the eigenvalue matrix to generate the Hilbert dynamic mode decomposition result.

[0012] Further, wherein according to the forward data matrix and the backward data matrix , through eigenvalue decomposition, obtain the eigenvalue matrix and the dynamic mode matrix, including: Calculate the covariance matrix according to the forward data matrix ; ; Perform singular value decomposition on the forward data matrix , and obtain , where is the left singular vector matrix, representing the mode of the data; is a diagonal matrix containing singular values that measure the importance of each mode in the data; is a right singular vector matrix for linear transformation of data; According to the backward data matrix calculate the dimensionality reduction state transition matrix ; Calculate the eigen - decomposition: , where is the dynamic mode matrix; is the eigenvalue matrix corresponding to the mode.

[0013] Furthermore, among them, the time - frequency signal matrix is adaptively denoised and reconstructed by Stein unbiased risk estimation to obtain the reconstructed acoustic emission signal , including: Based on the time - frequency signal matrix obtained by wavelet packet transform, determine the optimal denoising threshold through Stein unbiased risk estimation; According to the optimal denoising threshold, perform soft - threshold denoising on the time - frequency signal matrix to generate a coefficient matrix after soft - threshold processing; Based on the coefficient matrix after soft - threshold processing, reconstruct the acoustic emission signal through inverse wavelet packet transform .

[0014] Furthermore, among them, for the reconstructed acoustic emission signal perform Henon mapping and Sierpinski triangle mapping in sequence, and then perform Weierstrass transform to generate fractal feature data , including: Take the denoised and reconstructed acoustic emission signal as the initial input, and generate a chaotic signal component through Henon mapping ; For the chaotic signal component , perform fractal transformation through Sierpinski triangle mapping to generate an intermediate fractal signal ; For the intermediate fractal signal construct acoustic emission test data with fractal features through Weierstrass cosine function mapping, that is, fractal feature data .

[0015] Furthermore, among them, the machine learning network model includes a backbone network; The backbone network uses a gated recurrent unit structure, including an update gate and a reset gate; Introduce an attention mechanism on the basis of the backbone network, and the context vector enhanced by attention The hidden state update formula incorporated into the backbone network enables the model to dynamically focus on key time steps.

[0016] Furthermore, the loss function of the machine learning network model is composed of cost-sensitive weighted cross entropy loss, smooth L1 loss and log hyperbolic cosine loss.

[0017] Compared with the prior art, the present invention achieves the following beneficial effects: 1. This paper proposes a time-frequency energy change rate analysis method that uses wavelet packet transform to perform time-frequency decomposition of acoustic emission signals, construct a time-frequency signal matrix, and calculate the signal's time-frequency energy change rate. This method can effectively capture the transient impact characteristics of sudden faults and enable early fault identification by detecting dramatic fluctuations in the energy change rate.

[0018] 2. This paper proposes Hilbert dynamic mode decomposition (DMD). By constructing a Hilbert matrix and introducing a DMD method, the method decomposes the signal's time series features and extracts its primary dynamic modes. This method can analyze the non-stationary signal characteristics during gear reduction motor operation and enhance the ability to identify fault modes under complex operating conditions.

[0019] 3. This paper addresses potential noise interference in acoustic emission signals by using Stein's unbiased risk estimation method for adaptive signal denoising and smooth reconstruction. This method effectively reduces noise while preserving the key signal features, improving the accuracy of subsequent feature extraction and classification.

[0020] 4. To more accurately simulate the complex characteristics of gear reduction motor fault signals, this paper introduces the Weierstrass cosine function to generate test data with fractal characteristics. This method can better reflect the multi-scale characteristics of the signal and provide data samples that are closer to actual operating conditions for training fault diagnosis systems.

[0021] 5. Because different fault categories have different probabilities of occurrence, directly using traditional cross-entropy loss may result in a low recognition rate for some categories. This paper proposes a cost-sensitive weighted cross-entropy loss function, which assigns different weights to different categories to improve the recognition of low-frequency fault modes and reduce the false positive rate.

[0022] 6. We propose using the Lorentz curve and Gini coefficient for fault determination. We calculate the Lorentz curve of the acoustic emission signal amplitude and combine it with the Gini coefficient to assess the unevenness of the signal distribution. This method can effectively measure the amplitude concentration of the fault signal and provide a new criterion for fault severity assessment.

[0023] It should be understood that the content described in the Summary of the Invention section is not intended to limit the key or important features of the embodiments of the present invention, nor to limit the scope of the present invention. Other features of the present invention will become readily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In combination with the accompanying drawings and with reference to the following detailed description, the above and other features, advantages, and aspects of the embodiments of the present invention will become more apparent. The drawings are used to better understand the solution and do not constitute a limitation to the present invention. In the drawings, the same or similar reference numerals represent the same or similar elements, where: Figure 1 FIG. shows a schematic diagram of the specific steps of a method for fault diagnosis of a gear reduction motor based on acoustic emission according to an embodiment of the present invention; Figure 2 FIG. shows a flowchart of a method for fault diagnosis of a gear reduction motor based on acoustic emission according to an embodiment of the present invention; Figure 3 FIG. is a performance comparison diagram of different models. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Apparently, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art without making creative efforts based on the embodiments of the present invention fall within the scope of protection of the present invention.

[0026] In addition, the term "and / or" in this document is merely a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist simultaneously, and B exists alone. In addition, the character " / " in this document generally represents an "or" relationship between the associated objects before and after.

[0027] Figure 1 FIG. shows a schematic diagram of the specific steps of a method for fault diagnosis of a gear reduction motor based on acoustic emission; Figure 2 FIG. shows a flowchart of a method for fault diagnosis of a gear reduction motor based on acoustic emission. As Figure 1 and Figure 2 shown, a method 100 for fault diagnosis of a gear reduction motor based on acoustic emission includes the following steps: S110: Collect the original acoustic emission signals during the operation of the gear reduction motor; In this step S110, a high-sensitivity acoustic emission sensor is installed at the key parts of the gear reduction motor to collect the acoustic emission signals generated during the operation in real time.

[0028] S120: Perform multi-dimensional preprocessing on the acoustic emission signal, extract signal features, and construct an input feature matrix based on the extracted signal features and the original acoustic emission signal; In step S120, the acoustic emission signal is preprocessed in multiple dimensions to extract signal features to the greatest extent. Step S120 specifically includes the following steps: S121: Decompose the original acoustic emission signal using wavelet packet transform to obtain a time-frequency signal, and calculate the signal time-frequency coherence and the time-frequency energy change rate ; This step S121 specifically includes: S1211: Decompose the original acoustic emission signal using wavelet packet transform, and recursively decompose the original acoustic emission signal to obtain sub-signals in different frequency bands to extract signal components in different frequency ranges and obtain a time-frequency signal representation; Specifically, wavelet packet transform is used to decompose the acoustic emission signal, and the signal is recursively decomposed to obtain sub-signals in different frequency bands, which is suitable for analyzing non-stationary signals to extract signal components in different frequency ranges and obtain a time-frequency representation. The recursive decomposition formula of wavelet packet transform is as follows:

[0029] where represents the wavelet packet coefficient of the th layer and the th sub-band at time , that is, the time-frequency signal, which is the time-frequency representation of each sub-band in the wavelet packet decomposition process; represents the time-frequency signal of the rd layer and the th sub-band calculated by passing the low-pass filter; represents the time-frequency signal of the th sub-band of the rd layer; : represents the time-frequency signal of the th sub-band of the high-pass filter; represents the time-frequency signal of the th sub-band of the rd layer; represents the level of wavelet packet decomposition (e.g., j = 0, 1, …, J, where J is the maximum decomposition level); represents the sub - band index (frequency channel); represents the time index (discrete sampling points, which is the signal length), used to determine the specific time position of the wavelet packet coefficient to be calculated or described currently, and is the identifier of the time dimension in the final result. The value range is determined according to the specific time series length and the concerned time interval.

[0030] S1212: Obtain the time - frequency signal matrix after wavelet packet transform according to the said time - frequency signal representation; In this step S1212, obtain the time - frequency signal matrix after wavelet packet transform according to the said time - frequency signal representation :

[0031] where, represents the wavelet packet coefficient at time and frequency ;

[0032] S1213: Calculate the signal time - frequency coherence and time - frequency energy change rate from the time - frequency signal matrix.

[0033] The time - frequency coherence is used to measure the similarity between different signal components. Suppose two acoustic emission signals (such as dual - channel or multi - channel sensor signals) are respectively and , and their time - frequency representations obtained by the wavelet packet transform in step S1212 are:

[0034]

[0035] Then their time - frequency coherence is:

[0036] where, represents the time - frequency coherence of the two acoustic emission signals at time and frequency When , it means that the two signals have strong coherence (i.e., strong correlation) at this time - frequency point; when , it means that the two signals are relatively independent (i.e., uncorrelated) at this time - frequency point; represents the time - frequency cross - spectral density; represents the time - frequency power spectral density of signal represents the Time-frequency power spectral density; is the expected value, usually estimated by smoothing through a local time window (such as moving average); is the complex conjugate of.

[0037] The time-frequency energy change rate is used to measure the change of the energy of the acoustic emission signal over time to identify sudden faults or abnormalities.

[0038] Known signal The time-frequency energy after wavelet packet transform of the known signal is expressed as:

[0039] Then its time-frequency energy change rate is:

[0040] where represents the signal energy at time and frequency ; represents the signal energy at time and frequency ; is the time-frequency energy change rate at time and frequency . When is large, it indicates that the energy at this time-frequency point changes violently, which may correspond to the impact signal when the fault occurs; when is small, it indicates that the signal is stable and the system is in normal working condition.

[0041] S122: Decompose the original acoustic emission signal into subspaces with different resolutions based on multi-resolution analysis to extract signal components in different frequency ranges ; Multi-resolution analysis is a hierarchical signal decomposition method that decomposes a signal into subspaces of different scales (resolutions) through wavelet transform. Each scale corresponds to a specific frequency range. Low scales (high resolutions) capture high-frequency details, and high scales (low resolutions) capture low-frequency approximations. Its core idea is to decompose the signal step by step to form a hierarchical frequency band structure. In this embodiment, based on multi-resolution analysis, the acoustic emission signal is decomposed into subspaces with different resolutions. It projects the acoustic emission signal onto subspaces at different scales (resolutions) to extract signal components in different frequency ranges. Decomposing the acoustic emission signal into different subbands according to frequency facilitates the analysis of fault characteristics in specific frequency bands. For example: Bearing faults often manifest as high-frequency resonances, while gear wear may show energy changes in the mid-low frequency band.

[0042] Step S122 specifically includes: S1221: Select a wavelet basis function (such as Daubechies 4) and a decomposition scale (i.e., the maximum decomposition scale J, determined according to the signal sampling rate); S1222: Perform recursive decomposition on the original acoustic emission signal, extract the time-frequency local features of the original acoustic emission signal step by step, and then obtain the approximation signals at each scale and detail signals ; In this step S1222, filter and downsample to generate the approximation signal and detail signals . Specifically, it includes: Initialization: Let the original signal be .

[0043] Step-by-step decomposition: At scale j, decompose through the low-pass filter h and the high-pass filter g : ,

[0044] where , .

[0045] Recursively to the maximum scale J (such as J = 5).

[0046] S1221: Reconstruct different frequency range signal components based on the approximation signal and detail signals . .

[0047] This step S1221 is signal reconstruction: Specifically, at scale j, is used to retain the low-frequency trend of the signal (such as the steady-state component of normal gear operation), is used to capture high-frequency transient features (such as impacts or noises caused by faults), and then integrate to obtain :

[0048] : Represents the complete reconstruction of the signal at scale j, including: : The low-frequency approximation at scale j (constituted by the scaling function and the smoothing coefficient ).

[0049] , that is : The high-frequency details from scale 1 to j (constituted by the wavelet function and the detail coefficients Composition).

[0050] Finally, output the reconstructed multi - resolution signal , as well as the approximation coefficients at each scale and detail coefficients .

[0051] Through hierarchical decomposition, multi - resolution analysis forms components at different scales, enabling the energy of the acoustic emission signal to be distributed according to frequency components. The reconstructed multi - resolution signal obtained through hierarchical decomposition can reveal the characteristics of the signal in different frequency ranges, providing a multi - scale analysis basis for fault feature extraction and classification. Among them, and at different scales can be used as input features to improve the model's ability to identify complex fault patterns.

[0052] Band - energy calculation is used to calculate the energy at each scale to quantify fault features: ,

[0053] For example, when detecting tooth root cracks in a gear - reduction motor in a certain scenario, the sampling rate fs = 5 kHz and the decomposition is up to J = 4. Through multi - resolution decomposition, we get : 0 - 1.56 kHz (normal meshing frequency), : 12.5 – 25 kHz (high - frequency resonance caused by cracks); then perform feature extraction: calculate the time - frequency energy change rate of , and detect the points of sudden energy increase; finally, perform fault judgment: if > threshold, it is judged that a fault has occurred and a crack alarm is triggered.

[0054] S123: Perform Hilbert dynamic mode decomposition on the time - frequency signal, extract the active mode, and obtain the dynamic mode decomposition result ; Step S123 specifically includes: S1231: Construct a Hilbert matrix according to the time - frequency signal data ; In order to perform Hilbert dynamic mode decomposition, first, it is necessary to construct a Hilbert matrix from the time - frequency signal data obtained in step S1212:

[0055] Among them, is the Hilbert matrix, constructed from the time - frequency signal data ; Time-frequency signal data; : The number of rows and columns of the Hilbert matrix determines the time window length of the signal.

[0056] S1232: Construct a state transition matrix according to the Hilbert matrix including: dividing the Hilbert matrix into a forward data matrix and a backward data matrix ; In this step S1232, construct two data matrices:

[0057] They satisfy the following state transition relationship:

[0058] where and represent the Hilbert matrices of the previous and next time steps respectively, and are defined as the forward data matrix and the backward data matrix respectively; is the state transition matrix, representing the dynamic evolution characteristics of the signal.

[0059] S1233: According to the forward data matrix and the backward data matrix , through eigenvalue decomposition, obtain the eigenvalue matrix and the dynamic mode matrix; This step S1233 performs dynamic mode decomposition and extracts the active mode of the signal through the following steps, including: S12331: Calculate the covariance matrix Q according to the forward data matrix :

[0060] S12332: Perform singular value decomposition (SVD) on the forward data matrix to obtain:

[0061] where is the left singular vector matrix, representing the mode of the data (mode); is the diagonal matrix, containing singular values, measuring the importance of each mode in the data; is the right singular vector matrix, used for linear transformation of the data; S12333: Calculate the reduced-dimensional state transition matrix according to the backward data matrix :

[0062] Among them, is the dimensionality reduction state transition matrix, describing the evolution relationship of the signal in the low-dimensional space; is the transpose of the left singular matrix obtained by singular value decomposition, used to project the data into the modal space; is the orthogonal transformation in the singular value decomposition, used to calculate the state transition matrix.

[0063] S12334: Calculate the eigen decomposition:

[0064] Among them, is the dynamic modal matrix; is the eigenvalue matrix corresponding to the mode.

[0065] S1234: Combine the dynamic modal matrix with the eigenvalue matrix to generate the Hilbert dynamic modal decomposition result.

[0066] Finally, through this step S1234, the dynamic mode generated by the Hilbert dynamic modal decomposition can be used to analyze the dominant dynamic characteristics of the acoustic emission signal, and the generated Hilbert dynamic modal decomposition result is expressed as:

[0067] This result can be used for subsequent feature extraction and classification.

[0068] In this step S123, the time-frequency signal matrix is converted into a time-sequence related Hankel structure to capture the dynamic evolution characteristics of the signal, and the analytical ability for non-stationary signals is enhanced through time-frequency domain embedding. The dynamic relationship of the signal in the time dimension is established through state transition modeling, and the state transition matrix A is extracted, which can solve the problem of insufficient modeling of the time-sequence dependence by traditional methods. The dominant modes are extracted through singular value decomposition (SVD) and eigen decomposition, the dynamic modes related to faults are separated, and the covariance analysis and dimensionality reduction techniques are combined to improve the calculation efficiency and modal purity. The physical interpretable dynamic mode is generated through modal reconstruction, providing key features for fault classification, and the fault sensitivity under complex working conditions is enhanced by fusing time-frequency analysis and dynamic mode decomposition.

[0069] S124: Perform adaptive denoising and reconstruct the signal on the said time-frequency signal using Stein's unbiased risk estimate to obtain the reconstructed acoustic emission signal ; This step S124 uses Stein's unbiased risk estimate to reconstruct the smoothed acoustic emission signal, specifically including: S1241: Based on the time-frequency signal matrix obtained by wavelet packet transform, determine the optimal denoising threshold through Stein's unbiased risk estimate;

[0070] Among them, is the optimal threshold obtained by Stein's unbiased risk estimate method; is the wavelet coefficient after wavelet packet transform, representing the energy distribution of the acoustic emission signal in a specific time-frequency domain (a certain scale, frequency band). Larger absolute values correspond to the effective components of the signal, and smaller absolute values may be noise; is the variance of noise estimation, quantifying the intensity of background noise, usually calculated through the high-frequency subband coefficients (such as the finest scale) of wavelet transform. For example: , and the median() function is used to robustly estimate the dispersion degree of high-frequency noise; is the Heaviside function (taking 1 when the input is greater than 0, otherwise taking 0). is the denoising threshold, used to distinguish the effective components of the signal from noise. When , it is considered that this coefficient contains the effective signal, otherwise it is regarded as noise and set to zero or shrunk.

[0071] Minimize this formula through Stein's unbiased risk estimate, realizing the estimation of the mean square error after denoising without the true signal, thereby optimizing the threshold λ, being able to achieve the denoising goal, and minimizing the total risk by balancing signal retention (the first term) and noise suppression (the second term).

[0072] S1242: Perform soft threshold denoising on the time-frequency signal matrix according to the optimal denoising threshold to generate a coefficient matrix after soft threshold processing; In this S1242, after using the threshold determined by Stein's unbiased risk estimate, perform soft threshold denoising on the time-frequency signal matrix to generate a corrected wavelet coefficient matrix , and its expression is:

[0073] Among them, is the wavelet coefficient after soft threshold processing; is the sign function, indicating the positive or negative of the wavelet coefficient.

[0074] S1243: Reconstruct the acoustic emission signal through inverse wavelet packet transform based on the coefficient matrix after soft threshold processing .

[0075] After soft thresholding, in step S1243, an inverse wavelet transform is performed to reconstruct the smoothed signal:

[0076] where, is the reconstructed signal after denoising; are the wavelet coefficients after thresholding (i.e., the version after soft thresholding by Stein's unbiased risk estimator); is the wavelet basis function.

[0077] Through adaptive threshold calculation, soft thresholding, and inverse transform reconstruction, step S124 can achieve automatic adaptation of the threshold to the signal-to-noise ratio of the signal. Moreover, compared with the hard threshold, soft thresholding reduces the pseudo-Gibbs oscillations during signal reconstruction, improves the smoothness, and finally maps the denoised wavelet coefficients back to the time domain through inverse transform reconstruction to generate a high-fidelity smoothed signal. Combining with the multi-resolution characteristics of the wavelet packet, it can ensure the integrity of the local features of the reconstructed signal in the time-frequency domain.

[0078] S125: Perform Henon mapping and Sierpinski triangle mapping on the reconstructed acoustic emission signal in sequence, and then perform the Weierstrass transform to generate fractal feature data ; In step S125, the reconstructed acoustic emission signal data obtained by denoising in step S124 is used to generate data with a fractal structure using Henon mapping, Sierpinski triangle mapping, and the Weierstrass cosine function, specifically including: S1251: Use the denoised and reconstructed acoustic emission signal as the initial input to generate a chaotic signal component through Henon mapping; The Henon mapping is a typical chaotic system. We use the acoustic emission reconstruction signal data obtained by denoising in step S124 as the initial value

[0079] where, : The signal component after Henon mapping, which is calculated based on the current and the previous and represents a part of the result after the Henon mapping transformation. In the time series, it will be updated according to the formula as the difference advances, reflecting the mapping of the original signal at different times. Transformation effect : It is also the signal component after the Henon mapping. It is calculated based on the at the previous moment and is mutually related and jointly constitutes the output result of the Henon mapping. At each time step, it interacts with and participates in the calculation of and at the next moment : Henon mapping parameter, satisfying , b ∈ [0.2, 0.4], usually taking .

[0080] S1252: For the chaotic signal component , perform a fractal transformation through the Sierpinski triangle mapping to generate an intermediate fractal signal ; The Sierpinski triangle mapping is a typical fractal generation method. For the chaotic signal component generated in step S1251, perform a fractal transformation through a modified Sierpinski triangle mapping to generate an intermediate fractal signal , and its expression is:

[0081] where is the signal component after the Sierpinski triangle mapping, that is, the intermediate fractal signal; is the exponential factor controlling the frequency change, ensuring the self - similarity of the fractal structure.

[0082] 1253: For the intermediate fractal signal , construct acoustic emission test data with fractal characteristics through the Weierstrass cosine function mapping, that is, fractal characteristic data .

[0083] The Weierstrass cosine function mapping is a fractal function. In this embodiment, for the intermediate fractal signal after the Sierpinski triangle mapping obtained in step S1252, construct acoustic emission test data with fractal characteristics through the Weierstrass cosine function mapping, and its formula is:

[0084] where is the final acoustic emission signal data with a fractal structure, that is, fractal characteristic data; : Parameter for controlling amplitude attenuation; : Parameter for controlling frequency growth, satisfying to ensure the convergence of the fractal dimension.

[0085] In this step S125, the denoised acoustic emission signal is mapped to the chaotic space through the Henon map to generate a signal component with non-linear dynamic characteristics, simulating the mixed characteristics of noise and fault signals in the actual working condition. Through the Sierpinski map, the terms are used to generate a fractal basis, endowing the signal with multi-scale self-similar characteristics. Combining chaotic signals and fractal geometry, transitional data with both randomness and structure is generated . Weierstrass map, where the parameter co-design ( ) ensures that the fractal dimension is controllable and adapts to the data requirements of different fault modes.

[0086] S126: For the original occurrence signal and the time-frequency signal obtained in steps S121 - S125 , signal time-frequency coherence , time-frequency energy change rate , the signal components in the different frequency ranges , dynamic mode , reconstructed acoustic emission signal , the fractal feature data are integrated to construct an input feature matrix for classification .

[0087] This step S126 constructs an input feature matrix for classification :

[0088] Finally, is used as the input of the machine learning network model. In addition, if image data fusion is considered, the amplitude curve of the acoustic emission signal data can be converted into a grayscale Figure 2 value matrix and put into a neural network suitable for image processing for classification.

[0089] S130: Construct a machine learning network model including a gated recurrent unit and an attention mechanism, and input the input feature matrix into the trained machine learning network model to output a classification result; Among them, the machine learning network model includes a backbone network; the backbone network uses a gated recurrent unit structure, including an update gate and a reset gate;

[0090] Among them, : The input features at the current time step; : The hidden state at the previous time step, used to transmit historical information; : Input to the weight matrix of the update gate; : Hidden state to the weight matrix of the update gate; : The bias term of the update gate; : The update gate, used to control the degree of combination of new and old information; : The weight matrix and bias term of the reset gate; : The reset gate, used to control the degree of forgetting of historical information; : Calculate the weights and biases of the candidate hidden state; : The reset gate acts on the old hidden state to control the degree of information retention; : The candidate hidden state, representing the new information at the current time step; : The hidden state at the current time step; : Retain part of the historical information; : Introduce new information; : The Sigmoid activation function, which maps the input to the interval [0,1], used to convert the calculation result of the update gate into a probability value; : Element-wise multiplication operation. It means multiplying the output of the reset gate element-wise with the hidden state of the previous time step, and determining how much historical information to retain according to the state of the reset gate; : The hyperbolic tangent activation function.

[0091] This structure can effectively capture the short-term and long-term dependencies of audio signals and avoid the vanishing gradient problem.

[0092] Based on the backbone network, an attention mechanism is introduced, and the attention-enhanced context vector is incorporated into the hidden state update formula of the backbone network, enabling the model to dynamically focus on key time steps.

[0093]

[0094] Among them, : Trainable parameters; : The normalized attention weight, which represents the importance of the hidden state of the current time step among the hidden states of all time steps; : The context vector, representing global information, which is obtained by applying the attention weights to the hidden states of all time steps It is obtained by weighted summation, which can highlight the information of the time steps considered important by the model and help the model focus more on the key information. : The total number of time steps.

[0095] Finally, the context vector enhanced by attention is incorporated into the hidden state update formula of the backbone network, enabling the model to dynamically focus on the key time steps.

[0096] After that, dimensionality reduction is performed through a fully connected layer. The Dropout layer is used to randomly discard neurons to prevent overfitting. Finally, the probability of each category is generated through the Softmax output layer to determine the classification result :

[0097] Among them, : The weight matrix of the fully connected layer, used to transform the context vector to the hidden layer space; : The bias term of the fully connected layer; : The activation function, used to introduce non-linearity and enable the model to have stronger expressive power; : The weight matrix of the output layer, used to map to the final category probability space; : The bias term of the output layer; : The normalization function, used to calculate the probability of each category; : Randomly discard neurons, used to prevent the model from overfitting and improve the generalization ability; : The category probability distribution predicted by the model; Take the category corresponding to the maximum probability as the final classification result.

[0098] The final classification result is determined by decide.

[0099] Construct the loss function: In order to combine the classification and regression losses and improve the model's attention to high-cost samples, the present invention constructs a comprehensive loss function, which is composed of a cost-sensitive weighted cross-entropy loss, a smooth L1 loss, and a logarithmic hyperbolic cosine loss:

[0100] Among them: : Hyperparameter, used to balance the importance of different loss terms; : Cost-sensitive weighted cross-entropy loss, improving the attention to high-cost categories; : Smooth L1 loss, improving the sensitivity to small errors, while suppressing the influence of large errors and enhancing the robustness to outliers; : The logarithmic hyperbolic cosine loss is numerically stable and combines the advantages of L1 and L2.

[0101] Among them, the cost-sensitive weighted cross-entropy loss is:

[0102] Among them, : The true label (one-hot encoded) of the th sample; : The Softmax output probability of the predicted class of the th sample; : The total number of samples; : The class weight, calculated as follows:

[0103] Among them is the misclassification cost of class , making high-cost misclassified samples have a greater impact on model training. Cost-sensitive weighting assigns different weights to each sample according to the sample's class and the corresponding misclassification cost. Samples with high-cost misclassifications are assigned higher weights. During model training, these samples will have a greater impact on the construction of the model, making the model more inclined to correctly classify these high-cost samples.

[0104] Smooth L1 loss is applicable to regression tasks and can be used to measure the error of predicted continuous values, making the model more sensitive to small errors and more robust to large errors:

[0105] Among them, : The predicted value output by the model; : The true value (e.g., regression target or class probability); : The threshold, controlling the transition region between L1 and L2.

[0106] Logarithmic hyperbolic cosine loss is a smoothed version of

[0108]

[0109] It combines the smoothness of mean squared error and the robustness of absolute error, and is more stable than in numerical calculations.

[0110] S140: Calculate the Lorentz curve and the Gini coefficient G of the amplitude of the original acoustic emission signal; In this step S140, the Lorentz curve and Gini coefficient analysis are introduced into the fault diagnosis of the gear reduction motor, and the maximum values of the amplitude curves of the original acoustic emission signals are sorted, and the Lorentz curve function is constructed:

[0111] where : the th acoustic emission signal value after amplitude sorting; : the total number of signal samples; : the signal sampling time interval; : represents the current cumulative time ratio, that is, the ratio of the first sampling points to the total time, indicating the proportion of the signal considered within the entire sampling time range. : the current cumulative signal amplitude ratio, which constitutes the Lorentz curve.

[0112] The Lorentz curve reflects the distribution of the signal amplitude. If the amplitude is highly concentrated (i.e., the signal intensity varies greatly), and the curve is far from the diagonal line, it indicates that the signal fluctuates violently, which may mean a more serious fault.

[0113] The Gini coefficient measures the degree of deviation of the Lorentz curve from the diagonal line, that is:

[0114] where is the area under the Lorentz curve, and the calculation method is:

[0115] The Gini coefficient ranges from : : The signal is uniform, indicating that the gear is operating normally.

[0116] : The signal is highly non-uniform, indicating that the gear may have serious faults.

[0117] S150: Weightedly fuse the classification result with the Gini coefficient G and output the final fault decision value.

[0118] Fuse the classification result obtained in step S130 and the classification result obtained in step S140 Perform weighting to obtain the final judgment result of the gear reduction motor failure:

[0119] wherein, : The judgment value of the final gear reduction motor failure; : The weighting coefficient, satisfying , and is used to balance the influence of machine learning classification and signal distribution analysis.

[0120] According to the above embodiments of the present invention, a time-frequency signal matrix is constructed by using wavelet packet transform, and combined with the analysis of the time-frequency energy change rate, the transient impact characteristics can be accurately captured. The Hilbert dynamic mode decomposition technology realizes the analysis of the time series characteristics of non-stationary signals. The Stein unbiased risk estimation (SURE) threshold denoising method can dynamically adjust the threshold according to the signal characteristics. In a strong noise environment with a signal-to-noise ratio (SNR) of 5 dB, a feature fidelity of more than 92% can still be maintained. The fractal test data generated by the Weierstrass cosine function effectively solves the problem of model overfitting caused by scarce samples. Through the combination of the above theoretical innovation and engineering practice, the present invention forms a complete technical chain from signal acquisition, feature extraction, intelligent classification to state evaluation, significantly improving the fault diagnosis accuracy and prediction ability of the gear reduction motor, and providing key technical support for intelligent manufacturing and equipment health management.

[0121] Experimental comparison: Figure 3 is a performance comparison chart of the present invention and different models. As Figure 3 shown, the accuracy reflects the proportion of samples that are actually positive among the samples predicted as positive by the model. It can be seen from the bar chart that in terms of accuracy, the proposed algorithm model performs the most prominently, reaching 0.9989, which means that there are very few false alarms in the prediction of this model. Followed by the long short-term memory network model with an accuracy of 0.9774, which also shows a relatively high accuracy level. In contrast, the accuracy of the convolutional recurrent neural network model is the lowest, only 0.7664, indicating that there may be more false alarms when this model judges the failure of the gear reduction motor.

[0122] The recall rate measures the ability of the model to correctly identify positive examples, that is, how many of the samples that are actually positive are predicted as positive by the model. In terms of recall rate, the proposed algorithm model performs excellently again, reaching 0.9938, indicating that the model can well capture the actual gear reduction motor failures and there are very few missed reports. The recall rate of the deep neural network model is also relatively high, 0.9573, showing good performance. However, the recall rate of the convolutional recurrent neural network model is relatively low, only 0.5816, indicating that this model may miss many actual gear reduction motor failures.

[0123] The F1-score is an indicator that comprehensively considers precision and recall. It is the harmonic mean of precision and recall, which can more comprehensively reflect the performance of the model. As can be seen from the figure, the F1-score of the proposed algorithm model is the highest, reaching 0.9963, which further proves that the model has excellent performance in both precision and recall, and is a model with excellent performance. The F1-score of the convolutional neural network model is 0.9531, which also has good performance. While the F1-score of the convolutional recurrent neural network model is the lowest, only 0.6613, indicating that there is a large gap in the comprehensive performance of this model compared with other models.

[0124] Through the comprehensive analysis of the three indicators, it can be found that the proposed algorithm model performs the best in the gear reduction motor fault detection task, leading other models in terms of precision, recall, and F1-score, and can efficiently and accurately detect gear reduction motor faults, reducing false alarms and missed detections. The long short-term memory network and the deep neural network model also have good performance in different indicators, indicating that they are effective in dealing with the gear reduction motor fault detection task. In contrast, the convolutional recurrent neural network model performs poorly in all three indicators, showing an obvious gap with other models, and may need further improvement or optimization. The performance of the convolutional neural network model is at a medium level, and it can be selected for use according to specific requirements and resource limitations in practical applications.

[0125] It should be understood that various forms of the processes shown above can be used, steps can be reordered, added, or deleted. For example, the steps described in the present invention can be executed in parallel, sequentially, or in a different order, as long as the desired results of the technical solutions disclosed in the present invention can be achieved, and no limitation is made herein.

[0126] The above specific embodiments do not constitute a limitation to the protection scope of the present invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A fault diagnosis method for a gear reduction motor based on acoustic emission, characterized in that, It includes the following steps: Collect the original acoustic emission signal when the gear reduction motor is running; Perform multi-dimensional preprocessing on the acoustic emission signal, extract signal features, and construct an input feature matrix based on the extracted signal features and the original acoustic emission signal; Construct a machine learning network model including a gated recurrent unit and an attention mechanism, and input the input feature matrix into the trained machine learning network model to output a classification result; Calculate the Lorentz curve and Gini coefficient of the amplitude of the original acoustic emission signal; Fusion the classification result and the Gini coefficient with weights, and output the final fault decision value.

2. The method for fault diagnosis of a gear reduction motor based on acoustic emission according to claim 1, wherein Wherein, Performing multi-dimensional preprocessing on the acoustic emission signal, extracting signal features, and constructing an input feature matrix based on the extracted signal features and the original acoustic emission signal includes: Decompose the original acoustic emission signal by wavelet packet transform to obtain a time-frequency signal, and calculate the signal time-frequency coherence and time-frequency energy change rate according to the time-frequency signal; Decompose the original acoustic emission signal into subspaces with different resolutions based on multi-resolution analysis to extract signal components in different frequency ranges; Perform Hilbert dynamic mode decomposition on the time-frequency signal, extract the active mode, and obtain the dynamic mode decomposition result; Perform adaptive denoising and reconstruction of the signal on the time-frequency signal by Stein's unbiased risk estimate to obtain a reconstructed acoustic emission signal; Perform Henon mapping and Sierpinski triangle mapping on the reconstructed acoustic emission signal in sequence, and then perform Weierstrass transform to generate fractal feature data; Integrate the original occurrence signal, the time-frequency signal, the signal time-frequency coherence, the time-frequency energy change rate, the signal components in different frequency ranges, the dynamic mode, the reconstructed acoustic emission signal, and the fractal feature data to construct an input feature matrix for classification.

3. The method for diagnosing faults of a gear reduction motor based on acoustic emission according to claim 2, wherein Wherein, The step of decomposing the original acoustic emission signal by wavelet packet transform to obtain a time-frequency signal, and calculating the signal time-frequency coherence and time-frequency energy change rate according to the time-frequency signal includes: Decompose the original acoustic emission signal by wavelet packet transform, and recursively decompose the original acoustic emission signal to obtain sub-signals in different frequency bands to extract signal components in different frequency ranges, and obtain a time-frequency signal representation; According to the time-frequency signal representation, obtain a time-frequency signal matrix after wavelet packet transform; Calculate the signal time-frequency coherence and time-frequency energy change rate from the time-frequency signal matrix.

4. The method for fault diagnosis of a gear reduction motor based on acoustic emission according to claim 2, wherein, Wherein, The step of decomposing the original acoustic emission signal into subspaces with different resolutions based on multi-resolution analysis to extract signal components in different frequency ranges includes: Select a wavelet basis function and a decomposition scale; Recursively decompose the original acoustic emission signal, and extract the time-frequency local features of the original acoustic emission signal step by step, and then obtain the approximate signal and detail signal at each scale; Reconstruct different frequency range signal components according to the approximate signal and the detail signal.

5. The method for fault diagnosis of a gear reduction motor based on acoustic emission according to claim 2, wherein Wherein, Performing Hilbert dynamic mode decomposition on the time-frequency signal, extracting the active mode, and obtaining the dynamic mode decomposition result includes: Construct a Hilbert matrix based on the time-frequency signal data ; According to the Hilbert matrix to construct a state transition matrix, including: dividing the Hilbert matrix into a forward data matrix and a backward data matrix , and satisfying the state transition relationship: , where is the state transition matrix, representing the dynamic evolution characteristics of the signal; According to the forward data matrix and the backward data matrix , through eigenvalue decomposition, an eigenvalue matrix and a dynamic mode matrix are obtained; Combine the dynamic modal matrix with the eigenvalue matrix to generate the Hilbert dynamic modal decomposition result.

6. The method for fault diagnosis of a gear reduction motor based on acoustic emission according to claim 5, wherein, Among them, According to the forward data matrix and the backward data matrix , through eigenvalue decomposition, an eigenvalue matrix and a dynamic mode matrix are obtained, including: According to the forward data matrix Calculate the covariance matrix ; Perform singular value decomposition on the forward data matrix to obtain , where is the left singular vector matrix, representing the mode of the data; is a diagonal matrix containing singular values, measuring the importance of each mode in the data; is the right singular vector matrix, used for linear transformation of the data; According to the said backward data matrix calculate the dimensionality reduction state transition matrix ; Calculate the eigen - decomposition: , where is the dynamic modal matrix; is the eigenvalue matrix corresponding to the modes.

7. The fault diagnosis method of the gear reduction motor based on acoustic emission according to claim 2, characterized in that Among them, Adopt Stein's unbiased risk estimate for the time-frequency signal matrix to adaptively denoise and reconstruct the signal, and obtain the reconstructed acoustic emission signal , including: Based on the time-frequency signal matrix obtained by wavelet packet transform, determine the optimal denoising threshold through Stein's unbiased risk estimator; Perform soft threshold denoising on the time-frequency signal matrix according to the optimal denoising threshold to generate a coefficient matrix after soft threshold processing; Based on the coefficient matrix after soft thresholding processing, the acoustic emission signal is reconstructed through inverse wavelet packet transform .

8. The method for fault diagnosis of a gear reduction motor based on acoustic emission according to claim 2 or 7, characterized in that, Among them, For the reconstructed acoustic emission signal perform Henon mapping and Sierpinski triangle mapping in sequence, and then perform Weierstrass transform to generate fractal feature data , including: The denoised and reconstructed acoustic emission signal is used as the initial input to generate chaotic signal components through the Henon map ; For the chaotic signal component , perform a fractal transformation through the Sierpinski triangle mapping to generate an intermediate fractal signal ; For the intermediate fractal signal By mapping through the Weierstrass cosine function, acoustic emission test data with fractal characteristics, i.e., fractal characteristic data, is constructed .

9. The method for fault diagnosis of a gear reduction motor based on acoustic emission according to claim 1, characterized in that, Among them, The machine learning network model includes a backbone network; The backbone network uses a gated recurrent unit structure, including an update gate and a reset gate; Introduce an attention mechanism on the basis of the backbone network, and incorporate the context vector enhanced by attention into the hidden state update formula of the backbone network, enabling the model to dynamically focus on key time steps.

10. The method for fault diagnosis of a gear reduction motor based on acoustic emission according to claim 9, wherein Among them, The loss function of the machine learning network model is composed of cost-sensitive weighted cross-entropy loss, smooth L1 loss, and logarithmic hyperbolic cosine loss.

Images