A bearing fault signal extraction model containing a sparse-dense convolution dictionary

Abstract

The application provides a bearing fault signal extraction model containing a sparse-dense convolution dictionary, which comprises a sparse channel for extracting sparse signals in a vibration signal of a bearing, the sparse channel comprising a sparse encoder for extracting sparse feature information of the sparse signals in the vibration signal and a sparse decoder for restoring the sparse feature information into the sparse signals, the sparse encoder being solved iteratively by using a deep convolution network; the bearing fault signal extraction model comprises a dense channel in parallel with the sparse channel and used for extracting dense signals in the vibration signal of the bearing, the dense channel comprising a dense encoder for extracting dense feature information of the dense signals in the vibration signal and a dense decoder for restoring the dense feature information into the dense signals. The application can overcome the defect that a multiple dictionary model is seriously dependent on expert experience when different dictionary atoms are selected, significantly reduce the expert dependence of the model, and enhance the generalization ability and data driving characteristics of the model.

Description

Technical Field

[0001] This invention relates to the field of bearing fault identification technology, specifically to a bearing fault signal extraction model containing a sparse-dense convolution dictionary. Background Technology

[0002] In modern industry, large and complex mechanical equipment is widely used in key economic sectors such as aerospace, energy production, and high-end manufacturing. As one of the most critical rotating components in this type of equipment, the operational health of rolling bearings directly determines the safety and reliability of the entire mechanical system. However, the harsh and complex operating environment of industrial sites makes bearings highly susceptible to sudden mechanical failures, which can lead to unplanned downtime, significant economic losses, and in severe cases, even endanger the lives and property of personnel.

[0003] Currently, vibration signal analysis-based fault diagnosis is the mainstream technology for monitoring bearing health. However, due to the increasingly complex mechanical structures of modern rotating machinery systems (such as those containing gears and rotors), the acquired vibration signals are often subject to strong background noise interference from multiple sources. Therefore, accurately extracting the weak impact characteristics representing early faults from vibration signals submerged in strong noise has become a key technical challenge in the field of bearing fault diagnosis.

[0004] In existing technologies, wavelet analysis, wavelet packet analysis, spectral kurtosis, and time-frequency analysis techniques such as empirical mode decomposition (EMD) and variational mode decomposition (VMD) are used to filter the acquired vibration signals to reduce noise and enhance fault characteristics. However, when faced with multi-component noise interference, simply filtering to separate the required signal is insufficient to fully grasp all component information in the signal, leading to missed fault detections.

[0005] Sparse representation theory, due to its ability to quantify the sparsity of signals, naturally aligns with the sparse distribution characteristics of bearing fault impact signals, and is widely used in the field of weak fault feature extraction. However, traditional sparse representation methods have two major limitations: First, it is difficult to effectively capture the translation-invariant properties of the signal, and the signal needs to be manually segmented; otherwise, redundant representations will be learned, affecting the feature extraction performance. Secondly, this method requires the construction of one or more overcomplete dictionaries, and the design and selection of dictionary atoms largely depend on the prior knowledge and expert experience of professionals, which severely limits the adaptability of the algorithm and its promotion in practical engineering applications. Summary of the Invention

[0006] In view of this, the problem to be solved by the present invention is to provide a bearing fault signal extraction model containing a sparse-dense convolution dictionary, which can reduce the model's dependence on expert experience and effectively enhance the model's generalization ability and data self-driving characteristics.

[0007] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A bearing fault signal extraction model containing a sparse-dense convolution dictionary includes a sparse channel for extracting sparse signals within the vibration signal of the bearing. The sparse channel includes a sparse encoder for extracting sparse feature information of the sparse signal within the vibration signal and a sparse decoder for recovering the sparse feature information into a sparse signal. The sparse encoder is solved iteratively using a deep convolutional network, and the solution model is as follows: , in, Indicates a vibration signal. This represents the k-th dictionary atom in the sparse convolution dictionary D. This represents the convolution operation, where K represents the number of atoms in the sparse convolution dictionary D, and D represents the sparse convolution dictionary. This represents the convolutional sparse vector output by the k-th dictionary atom processing. Represents sparse parameters, This represents the set of all dictionary atoms belonging to the sparse convolution dictionary D.

[0008] Furthermore, it includes a dense channel running parallel to the sparse channel for extracting dense signals within the vibration signal of the bearing, the dense channel including a dense encoder for extracting dense feature information of dense signals within the vibration signal and a dense decoder for recovering the dense feature information into a dense signal.

[0009] Furthermore, this includes self-supervised learning strategies: reconstructing outputs based on dense and sparse channels. The error between the vibration signal X is used to construct a loss function. Based on the loss function, the dense and sparse channels are automatically optimized and updated until the error meets the requirements.

[0010] Furthermore, the vibration signal includes time-domain sparse fault impact signals generated by bearing failure, time-domain dense frequency harmonic signals generated by bearing rotation, and background noise signals generated by external vibration interference.

[0011] Furthermore, the sparse signal includes a fault impulse signal and a partially time-domain sparse background noise signal, while the dense signal includes a frequency conversion harmonic signal and a partially time-domain dense background noise signal.

[0012] Furthermore, the solution model is based on iterative calculation using the ReLU operator, which is: , in, This represents the sparse feature vector output in the (t+1)th iteration; This indicates deconvolution calculation; Let X represent the intermediate vector of the (t+1)th iteration, X represent the vibration signal, and c represent the step size factor. The intermediate vector The formula is: , , in, , Let these represent the sparse feature vectors output in the t-th and (t-1)-th iterations, respectively. and Let represent the correction step size for the t-th and t+1-th iterations, respectively.

[0013] Furthermore, the sparse encoder incorporates a residual network strategy to enhance feature extraction capabilities, the formula of which is: , , in, Let represent the intermediate vector in the t-th iteration.

[0014] The beneficial effects of this invention are: A solution model for a sparse encoder that can be iteratively solved using a deep convolutional network is established. The convolutional dictionary computation of the solution model enables the model to effectively extract features with displacement-invariant properties. Deep convolutional computation overcomes the inherent shortcomings of traditional sparse models, such as difficulty in capturing this property and easy learning of redundant representations, and effectively improves the model's accuracy in extracting sparse fault features and displacement-invariant sparse fault features.

[0015] Set to include intermediate vectors The ReLU operator, used as the activation function in deep convolutional networks, differs from the activation function in the conventional FISTA (Fast Iterative Shrinkage-Thresholding Algorithm) algorithm by using an intermediate vector. Increase the signal value of newly added sparse features during the iteration process to improve the accuracy of sparse feature extraction.

[0016] Reconstructed output based on dense and sparse channels The error between the vibration signal X is used to construct a loss function. Based on the loss function, the solution model in dense and sparse channels is trained and updated. The entire training and update process does not require human operation. The solution model can be trained and updated in an unsupervised or self-supervised manner, so that the solution model can automatically extract the multi-source sparse signals of various mechanical equipment.

[0017] By incorporating a residual network strategy (specifically using the FISTA residual learning structure) into the iterative computation of the sparse convolutional network, the addition of a residual learning structure between two iterations can effectively prevent the degradation of feature extraction capabilities caused by excessive model depth or too many convolutional layers, thereby improving network performance and stability. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of the sparse channel in a bearing fault signal extraction model containing a sparse-dense convolution dictionary according to the present invention. Figure 2 This is a flowchart of the sparse channel in a bearing fault signal extraction model containing a sparse-dense convolution dictionary according to the present invention; Figure 3 This is a schematic diagram of the residual network structure in a bearing fault signal extraction model containing a sparse-dense convolution dictionary according to the present invention. Figure 4 It is the original vibration signal waveform and its envelope spectrum; Figure 5 It is based on the time-domain waveform and envelope spectrum of the sparse signal after reconstruction of the sparse channel. Detailed Implementation

[0019] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the specification of this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed terms.

[0021] This invention provides a bearing fault signal extraction model containing a sparse-dense convolution dictionary, such as... Figure 1 As shown, it includes a sparse channel for extracting sparse signals from the vibration signals of the bearing. The sparse signals contain fault impact signals generated by bearing failure, and the type of bearing failure can be determined by analyzing the envelope spectrum of the sparse signals.

[0022] The operation process of the sparse channel is as follows: the sparse channel acquires the actual vibration signal of a bearing during operation and extracts the sparse signal from the vibration signal. The extracted sparse signal contains the fault impact signal and some time-domain discontinuous background noise signal. Since the time-domain continuous background noise signal is filtered out, the energy proportion of the background noise in the sparse signal is reduced, and its impact on the fault analysis results is small.

[0023] The vibration signal of a bearing typically includes a time-domain sparse fault impact signal caused by bearing failure, a time-domain dense rotational frequency harmonic signal caused by bearing rotation, and a background noise signal caused by external vibration interference; the background noise signal includes both time-domain dense and time-domain sparse signals. In one embodiment of this application, the sparse signal includes the fault impact signal and a partially time-domain sparse background noise signal.

[0024] The sparse channel includes a sparse encoder and a sparse decoder. The sparse encoder is used to extract sparse feature information of the sparse signal in the vibration signal, and the sparse decoder is used to restore the sparse feature information into a sparse signal. The sparse signal is separated from the vibration signal through signal reconstruction.

[0025] Specifically, the sparse encoder includes a first convolutional layer and a feature extraction layer. The first convolutional layer converts the sparse signal in the vibration signal into the original feature map based on the sparse convolution dictionary D. The feature extraction layer is used to extract the original feature map. Inner sparse features and generating sparse feature vectors The sparse decoder includes a first deconvolutional layer, which is based on a sparse convolutional dictionary D and sparse feature vectors. Deconvolution calculations can regenerate sparse signals.

[0026] To improve the sparse encoder's ability to extract sparse signals, the sparse encoder incorporates a solution model capable of performing depth convolution calculations. Through several iterations of convolution calculations, it efficiently extracts sparse fault features and displacement-invariant sparse fault features, thereby enhancing the sparse encoder's ability to extract sparse signals.

[0027] The mathematical expression for the solution model of a sparse encoder is: , in, Indicates a vibration signal. This represents the k-th dictionary atom in the sparse convolution dictionary D. This represents the convolution operation, where K represents the number of atoms in the sparse convolution dictionary D, and D represents the sparse convolution dictionary. This represents the convolutional sparse vector output by the k-th dictionary atom processing. Represents sparse parameters, This represents the set of all dictionary atoms belonging to the sparse convolution dictionary D.

[0028] The first part of the solution model represents the original vibration signal term, the second part represents the signal reconstruction term, and the third part represents the sparsity penalty term. The smaller the output value of the solution model, the closer the reconstructed term is to the original vibration signal term, and the more accurate the subsequently extracted sparse signal will be.

[0029] The sparse encoder employs the Iterative Thresholding Algorithm (FISTA) to compute the solution model, and replaces the soft thresholding operator in the traditional FISTA with a ReLU operator based on the ReLU activation function. .

[0030] Specifically: conventional soft threshold operator The expression is: , in, This represents a non-negative soft threshold activation function. This represents the negative soft threshold activation function.

[0031] The ReLU activation function is: , Non-negative soft threshold activation function : , Based on the computational similarity between the two activation functions, the expression for the ReLU operator can be derived as follows: , in, This represents the set soft threshold, and z represents the input variable.

[0032] Therefore, the expression for the ReLU operator of a sparse encoder is: , in, This represents the sparse feature vector output in the (t+1)th iteration; This indicates deconvolution calculation; Let X represent the intermediate vector of the (t+1)th iteration, X represent the vibration signal, and c represent the step size factor.

[0033] It is the result of a linear combination of the directions of two cyclic iterations, and its expression is: , , in, , Let these represent the sparse feature vectors output in the t-th and (t-1)-th iterations, respectively. and Let represent the correction step size for the t-th and t+1-th iterations, respectively.

[0034] By setting an intermediate vector within the ReLU operator By linearly correlating the results of two iterations through the iteration step size, the signal value of the newly added sparse features can be increased (by weighting and summing the sparse feature vectors extracted in the t-th iteration compared to the (t-1)-th iteration to generate an intermediate vector). This improves the ability to extract sparse features, thereby increasing the network adaptability of the ReLU operator and the solution model.

[0035] like Figure 2 As shown, to avoid the degradation of the sparse encoder's feature extraction capability during iterative computation, a residual network structure is added between two iterations. The expression for residual learning is: , , Where D represents the sparse convolution dictionary matrix; D T This represents the corresponding deconvolution matrix; Let represent the sparse feature vector output in the t-th iteration. Let represent the intermediate vector calculated in the t-th iteration; c represents the step size factor.

[0036] like Figure 3 As shown, to achieve a self-encoded sparse channel sparse dictionary and improve the adaptability of the sparse channel in different usage environments, a dense channel is set up in parallel with the sparse channel to extract the dense signal within the bearing's vibration signal. The dense signal is a vibration signal that is dense in the time domain. In one embodiment of this application, the dense signal includes a frequency conversion harmonic signal and a portion of the time-domain dense background noise signal.

[0037] The dense channel comprises a dense encoder and a dense decoder. The dense encoder extracts dense feature information from the dense signal within the vibration signal, while the dense decoder reconstructs the dense signal from the dense feature information. The dense encoder includes a second convolutional layer based on a dense convolution dictionary. A Transform dense signals within vibration signals into dense feature vectors. The dense decoder includes a second deconvolutional layer, which is based on a dense convolution dictionary.A and dense eigenvectors Row deconvolution calculations can regenerate dense signals.

[0038] Based on the constraint of "reconstruction consistency," the encoders in both dense and sparse channels are reconstructed simultaneously, improving the solution model's ability to extract and adapt to vibration signals. The constraint is as follows: , Where X represents the vibration signal of the bearing; This represents a fault impulse signal exhibiting time-domain sparsity. This represents a frequency-modulated harmonic signal with time-domain density. This represents the background noise signal.

[0039] The extraction model autoencoder process involves superimposing the sparse signal from the sparse channel output and the dense signal from the dense channel output to form the reconstructed output. The final output will be The reconstruction error between the signal X and the vibration signal X is used as the loss function, and the encoders in the sparse and dense channels are updated in a self-supervised training mode until the reconstruction error meets the requirements.

[0040] The formula for the loss function is: , Where n represents the length of the input signal sample, This represents the loss function.

[0041] The specific implementation process and beneficial effects of this application plan are as follows: Step 1: Acquire actual vibration signals Actual vibration signals were generated using a bearing fault simulation test bench, model ABLT-1A, designed by Hangzhou Bearing Test and Research Center Co., Ltd. The sampling frequency of the actual vibration signals was set. Set to 51200H, the bearing's rotational frequency The frequency was set to 50Hz. A localized outer ring fault was pre-invented in the bearing. Based on the bearing type and fault frequency, the characteristic frequency of the outer ring fault of the bearing can be calculated. It is 181 Hz.

[0042] Step 2: Analyze the original vibration signal Time-domain analysis: such as Figure 4 As shown in (a) and Figure 4(c), the time-domain waveform of the vibration signal is subjected to strong interference and noise, and it is impossible to identify any periodic impact components related to the fault.

[0043] Frequency domain analysis: such as Figure 4(b) shows the envelope spectrum of the vibration signal, where the frequency components related to the bearing rotation frequency (50Hz) dominate (in the figure). The 181 Hz bearing outer ring fault characteristic frequency and its harmonics are not prominent (see figure). Even some fault characteristic frequencies' harmonics are completely masked (lacking obvious 2×). and 3 ).

[0044] Step 3: Model Configuration and Parameter Initialization Input settings: Divide the original vibration signal into signal segments of preset length and use them as model input.

[0045] Sparse channel configuration: The sparse convolution dictionary D is initialized to contain 32 atoms, each with a length of 16.

[0046] Dense channel configuration: The dense convolution dictionary A is initialized to contain 8 atoms, each with a length of 16.

[0047] Convolutional layer parameters: The stride of both sparse and dense channels is set to 4.

[0048] Sparse solver configuration: The number of encoder iterations (i.e., encoder loop count) within the sparse channel is set to 8; sparse parameters Set it to 0.5, and set the step size factor c to 5.

[0049] Residual Networks: In the iterative computation of sparse channels, a residual learning structure is introduced to enhance the learning ability.

[0050] Step 4: Self-supervised training of the model Loss function: The reconstruction error is used as the model's loss function, that is, minimizing the difference between the input signal x and the total reconstructed output of the model. The mean square error between them. It is the sum of the sparse channel reconstructed signal and the dense channel reconstructed signal.

[0051] Optimizer: The Adam optimizer is selected to update the model parameters, and the learning rate is set to 0.01.

[0052] Training: Feed several prepared training vibration signal segments into the model for iterative training until the model converges.

[0053] Step 5: Separation and Reconstruction of Original Vibration Signals The processed signal segment from step 3 is input into the sparse channel of the model, where the sparse channel adaptively decomposes the sparse feature vector within the signal segment. (via sparse channels and ReLU operators); then decoded and reconstructed by transpose convolution of the trained sparse dictionary D to obtain the separated and reconstructed sparse signal components.

[0054] Step 6: Fault Diagnosis Temporal analysis: Sparse eigenvectors The time-domain waveform is as follows Figure 5 As shown in (a), the waveform exhibits clear impact characteristics. In areas where no fault impact occurs, the signal amplitude is close to zero, and the background noise and periodic interference in the original vibration signal are effectively suppressed.

[0055] Envelope spectral analysis: sparse eigenvectors The envelope spectrum is shown in Figure 5(b). The fault characteristic frequency of 181 Hz in the figure is... Its harmonics are the main frequency components in the spectrum. The bearing rotation frequency, which dominates the original vibration signal, has been removed.

[0056] This application's solution, by setting up a sparse-dense parallel architecture, can separate and reconstruct the sparse fault impact features in the actual vibration signal of a bearing containing strong sinusoidal interference into a sparse channel. Then, by analyzing the reconstructed signal of the sparse channel, it can accurately extract and diagnose fault features that cannot be identified in the original signal.

[0057] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention shall still fall within the scope of the technical solution of the present invention.

Claims

1. A bearing fault signal extraction model containing a sparse-dense convolution dictionary, characterized in that, The system includes a sparse channel for extracting sparse signals from the vibration signal of a bearing, and a dense channel running parallel to the sparse channel for extracting dense signals from the vibration signal of the bearing. The sparse channel includes a sparse encoder for extracting sparse feature information of the sparse signal within the vibration signal and a sparse decoder for recovering the sparse signal from the sparse feature information. The dense channel includes a dense encoder for extracting dense feature information of the dense signal within the vibration signal and a dense decoder for recovering the dense signal from the dense feature information. The system outputs a reconstruction based on the dense and sparse channels. The error between the vibration signal X is used to construct a loss function. Based on the loss function, the dense and sparse channels are automatically optimized and updated until the error meets the requirements. The dense encoder includes a second convolutional layer, which is based on a dense convolutional dictionary. A Transform dense signals within vibration signals into dense feature vectors. The dense decoder includes a second deconvolutional layer, which is based on a dense convolution dictionary. A and dense eigenvectors Row deconvolution calculations can regenerate dense signals; The sparse encoder is solved iteratively using a deep convolutional network, and the solution model is as follows: , in, Indicates a vibration signal. This represents the k-th dictionary atom in the sparse convolution dictionary D. This represents the convolution operation, where K represents the number of atoms in the sparse convolution dictionary D, and D represents the sparse convolution dictionary. This represents the convolutional sparse vector output by the k-th dictionary atom processing. Represents sparse parameters, This represents the set of all dictionary atoms belonging to the sparse convolution dictionary D; The sparse encoder employs a threshold iteration algorithm to compute the solution model, and sets up a ReLU operator based on the ReLU activation function to replace the soft threshold operator in the traditional FISTA. The solution model is based on iterative calculation using the ReLU operator, which is: , in, This represents the sparse feature vector output in the (t+1)th iteration; This indicates deconvolution calculation; Let X represent the intermediate vector of the (t+1)th iteration, X represent the vibration signal, and c represent the step size factor. The intermediate vector The formula is: ， ， in, , Let these represent the sparse feature vectors output in the t-th and (t-1)-th iterations, respectively. and Let represent the correction step size for the t-th and t+1-th iterations, respectively.

2. The bearing fault signal extraction model containing a sparse-dense convolution dictionary according to claim 1, characterized in that, The vibration signals include time-sparse fault impact signals caused by bearing failure, time-dense frequency harmonic signals caused by bearing rotation, and background noise signals caused by external vibration interference.

3. The bearing fault signal extraction model containing a sparse-dense convolution dictionary according to claim 2, characterized in that, The sparse signal includes fault impulse signals and some time-domain sparse background noise signals, while the dense signal includes frequency conversion harmonic signals and some time-domain dense background noise signals.

4. The bearing fault signal extraction model containing a sparse-dense convolution dictionary according to claim 1, characterized in that, The sparse encoder incorporates a residual network strategy to enhance feature extraction capabilities, the formula of which is: , , in, Let represent the intermediate vector in the t-th iteration.

Images