Abrasive particle signal identification method based on fusion attention module and parallel convolutional network
By integrating the attention module with the parallel convolution network, the problem of signal interference in the inductive oil abrasive sensor is solved, and the accurate identification and feature extraction of abrasive signals are realized, which improves the accuracy of abrasive detection and Practicality.
Patent Information
- Application Number
- CN202411009764.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-26
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2044-07-26
AI Technical Summary
In actual application, the existing induction oil abrasive sensor is affected by external electromagnetic interference and mechanical vibration, resulting in a large number of interference components in the induced voltage signal, which affects the accurate identification and quantitative analysis of the abrasive signal. It is difficult for traditional signal processing methods to be confident. A balance between noise ratio enhancement and abrasive feature protection is achieved, and the feature extraction process is complex and requires professional knowledge.
Adopting abrasive signal identification method based on the fusion attention module and the parallel convolution network, harmonic interference is suppressed through windowed discrete Fourier transform, iterative interpolation discrete Fourier transform algorithm and frequency domain compensation method, multi-scale features are extracted and Deep learning technology is used to classify abrasive signal to realize signal background noise suppression and feature extraction.
While avoiding the distortion of target signal characteristics, it realizes fast and accurate identification and classification of abrasive grain characteristics, reduces the requirements for signal processing expertise, and improves the practicality and accuracy of abrasive grain detection.
Smart Images

Figure CN119004200A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of oil wear particle monitoring, and specifically relates to a wear particle signal recognition method based on a fusion attention module and a parallel convolutional network. Background Art
[0002] Online oil wear particle detection uses non-destructive sensors installed in lubricating oil pipelines to detect and analyze wear particles in the oil in real time, so as to grasp the wear of key components of the target machinery, and plays an important role in the estimation of the operating status and life prediction of complex equipment. With the update and iteration of mechanical technology, many oil wear particle detection technologies with different principles have been developed and continuously innovated. Inductive debris sensor (IDS) detects debris by sensing the non-periodic regular voltage signal generated when particles pass through. With its high sensitivity to ferromagnetic wear particles, easy installation and low cost, and sensor performance is not easily affected by oil impurities, it is very popular among manufacturers and researchers of aircraft engines, marine engines and generator gearboxes. However, in actual application scenarios, affected by external electromagnetic interference and mechanical vibration, the induced voltage signal collected by IDS will inevitably contain a large number of interference components, which will have a great impact on the accurate identification and extraction of wear particle signal characteristics and subsequent quantitative analysis.
[0003] In order to solve this problem, many literatures have been published to improve the wear particle detection performance of IDS by using signal processing and feature extraction methods. Signal processing technology mainly relies on modal decomposition, adaptive filtering, frequency band separation and other methods to improve the signal-to-noise ratio of the detection signal, so as to suppress the influence of non-wear particle signals on the accurate identification of wear particles. Feature extraction mainly includes binary identification methods based on the geometric morphological characteristics of wear particle signals and the threshold of the conversion domain. The principle is that the wear particle induction signal has stable local nonlinear characteristics with high discrimination from noise. The main wear particle signal extraction process is segmentation-eigenvalue calculation-threshold classification. To achieve the evaluation of the mechanical operation status, it is necessary to accurately count ferromagnetic wear particles of different sizes, so as to provide a basis for the subsequent mechanical wear area analysis and wear degree evaluation. Although these two main wear particle signal processing methods have good results in scenarios with sufficient prior knowledge, they are overly dependent on the signal-to-noise suppression effect, resulting in the change of the local morphological characteristics of the wear particle induction voltage signal due to residual interference components or excessive filtering. The protection of wear particle features still cannot meet the requirements of accurate extraction and analysis of wear particle features. At the same time, methods that rely on simple nonlinear features will affect the wear particle recognition effect of the algorithm due to the rationality of feature threshold selection, the high similarity between noise morphology and wear particle features, etc., which brings challenges to the use and promotion of the technology. On the other hand, manual feature extraction and index discrimination is an extremely complex process, which often requires a wealth of professional knowledge accumulation and a large number of field experiments, resulting in the low versatility of the above traditional methods, which in turn affects the practicality and actual use effect of IDS. Summary of the invention
[0004] In order to solve the above technical problems, the present invention proposes a wear particle signal recognition method based on a fusion attention module and a parallel convolutional network, comprising:
[0005] S1: The lubricating oil wear particle monitoring system constructed by using the inductive wear particle detection sensor collects data of the lubricating oil liquid containing ferromagnetic wear particles to obtain the original signal to be processed;
[0006] S2: Perform a windowed discrete Fourier transform on the collected original signal to obtain the amplitude frequency domain spectrum of the entire signal, and calculate the amplitude, frequency and phase of each low-frequency harmonic interference through iterative interpolation discrete Fourier transform algorithm and frequency domain compensation method to construct the anti-phase component for harmonic interference suppression and generate simulation samples;
[0007] S3: Calculate the optimal decomposition layer number J of the stationary wavelet according to the center frequency of the wear particles, perform equal-length segmentation on the simulation samples according to the optimal decomposition layer number J of the stationary wavelet, extract the features of the wear particle signal and the simulation samples after equal-length segmentation to form a multi-scale feature training set, and complete the labeling process;
[0008] S4: Use the parallel convolutional network combined with the fusion attention module to encode and reconstruct the J multi-scale feature vectors of a single sample in the training set, and perform position encoding and vector embedding on the output J reconstructed high-dimensional vectors. Use the energy self-attention model to establish the dependency relationship between the global and local information vectors, obtain the final classification vector, and use Softmax to transform it into the predicted label probability;
[0009] S5: Construct a cross entropy loss function based on the predicted label probability, use the cross entropy loss function to fit the model, and obtain the model with the best classification accuracy after multiple trainings as the wear particle signal identifier. Then, obtain a limited number of real wear particle signal samples with the proposed signal processing parameters, and enhance the actual detection performance of the identifier through feature transfer learning.
[0010] S6: Perform signal segmentation and multi-scale feature decomposition on the raw data collected in real time, and input the migrated network model to obtain the wear particle signal identification result;
[0011] S7: Wear particle signal is reconstructed based on the accurate model of wear particle signal and key features of samples to obtain noise-free wear particle detection signal, and finally wear particle feature extraction is realized.
[0012] Beneficial effects of the present invention:
[0013] The present invention realizes the extraction of wear particle features based on parallel convolutional network operations after suppressing the background noise of the collected signal, while protecting the quantitative characteristics of the wear particle signal to the maximum extent; currently, the traditional algorithms based on signal decomposition and adaptive filtering in this field are difficult to strike a balance between the protection of wear particle features and the improvement of the signal-to-noise ratio. The most prominent problem is that it will change the amplitude of the wear particle voltage signal, thereby affecting the judgment of the health status of the machinery and failing to provide a reliable basis for the subsequent quantitative analysis of the wear of the machinery. The wear particle signal identification method based on the fusion attention module and the parallel convolutional network proposed in the present invention can combine deep learning technology to quickly and accurately identify and classify wear particle features while avoiding the distortion of the target signal features. In addition, the use of convolutional neural networks to realize wear particle feature extraction avoids the problems of difficulty in selecting the threshold values of each characteristic value required by the use of traditional feature extraction algorithms, and reduces the requirements for users' professional knowledge in the field of signal processing. BRIEF DESCRIPTION OF THE DRAWINGS
[0014] Figure 1 It is a flow chart of a wear particle signal recognition method based on a fusion attention module and a parallel convolutional network of the present invention;
[0015] Figure 2 It is a schematic diagram of the lubricating oil wear particle monitoring system of the present invention;
[0016] Figure 3 A schematic diagram of the raw signal to be processed output by the inductive wear particle sensor;
[0017] Figure 4 It is a schematic diagram of the structure of the fusion attention module of the present invention;
[0018] Figure 5 Schematic diagram of the structure of the parallel convolutional network of the present invention;
[0019] Figure 6 A schematic diagram of the original signal to be processed and the signal after harmonic elimination extracted by the wear particle monitoring system of the present invention;
[0020] Figure 7 A schematic diagram of a wear induced voltage signal identified by a parallel convolutional network of the present invention;
[0021] Figure 8 It is a schematic diagram of the reconstructed abrasive induced voltage signal of the present invention. DETAILED DESCRIPTION
[0022] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0023] The present invention provides a wear particle signal recognition method based on a fusion attention module and a parallel convolutional network, such as Figure 1 As shown, the following steps are included:
[0024] S1: The lubricating oil wear particle monitoring system constructed by using the inductive wear particle detection sensor collects data of the lubricating oil liquid containing ferromagnetic wear particles to obtain the original signal to be processed;
[0025] S2: Perform a windowed discrete Fourier transform on the collected original signal to obtain the amplitude frequency domain spectrum of the entire signal, and calculate the amplitude, frequency and phase of each low-frequency harmonic interference through iterative interpolation discrete Fourier transform algorithm and frequency domain compensation method to construct the anti-phase component for harmonic interference suppression and generate simulation samples;
[0026] S3: Calculate the optimal decomposition layer number J of the stationary wavelet according to the center frequency of the wear particles, perform equal-length segmentation on the simulation samples according to the optimal decomposition layer number J of the stationary wavelet, extract the features of the wear particle signal and the simulation samples after equal-length segmentation to form a multi-scale feature training set, and complete the labeling process;
[0027] S4: Use the parallel convolutional network combined with the fusion attention module to encode and reconstruct the J multi-scale feature vectors of a single sample in the training set, and perform position encoding and vector embedding on the output J reconstructed high-dimensional vectors. Use the energy self-attention model to establish the dependency relationship between the global and local information vectors, obtain the final classification vector, and use Softmax to transform it into the predicted label probability;
[0028] S5: Construct a cross entropy loss function based on the predicted label probability, use the cross entropy loss function to fit the model, and obtain the model with the best classification accuracy after multiple trainings as the wear particle signal identifier. Then, obtain a limited number of real wear particle signal samples with the proposed signal processing parameters, and enhance the actual detection performance of the identifier through feature transfer learning.
[0029] S6: Perform signal segmentation and multi-scale feature decomposition on the raw data collected in real time, and input the migrated network model to obtain the wear particle signal identification result;
[0030] S7: Wear particle signal is reconstructed based on the accurate model of wear particle signal and key features of samples to obtain noise-free wear particle detection signal, and finally wear particle feature extraction is realized.
[0031] Specifically, the process of collecting raw signals by the lubricating oil wear particle monitoring system includes:
[0032] Build as Figure 2 The platform diagram of the lubricating oil wear particle monitoring system is shown in FIG. 1 , in which the acquisition card model is NI-9219, and the sampling frequency is set to f s =5000, the sampling time is 60s, and after removing the subsequent redundant signals, the wear particle detection signal with 280000 sampling points is obtained. The inductive wear particle detection sensor uses DC excitation, the driving current I=0.5A, the amplifier multiple of the preamplifier is 2000, and the peristaltic pump flow rate is set to 510ml / min. The abrasive particle sizes (equivalent diameter of the sphere) are 220μm, 73μm, 113μm, 65μm, 82μm, and three copper particles with an average of 400μm passing through the inductive wear particle detection sensor in turn. The sensor will generate 8 corresponding induced voltage signals similar to a single-cycle sine waveform. The collected original signal to be processed is as follows Figure 3 As shown, it can be seen that the raw signal to be processed collected in real time contains not only the wear induced voltage signal, but also the harmonic interference caused by structural vibration and random noise caused by electromagnetic waves, etc., so it is impossible to quickly and accurately extract the morphological characteristics of the tiny wear induced voltage signal.
[0033] S2. Estimate the key features of the wear particle signal based on the sensor structure parameters and the dynamic parameters of the deployment environment, generate simulation data with adjustable feature parameters based on the accurate model of the output signal of the inductive oil wear particle sensor, and pre-process the original data and generated data;
[0034] Specifically, the collected original signal is first subjected to a windowed discrete Fourier transform to obtain the amplitude frequency domain spectrum of the entire signal; the amplitude, frequency and phase of each low-frequency harmonic interference are calculated through the iterative interpolation discrete Fourier transform algorithm and the frequency domain compensation method to construct the anti-phase component for harmonic interference suppression. The generated data is simulated based on the estimated wear particle velocity, size and sensor structural state parameters. Among them, the length of the simulation sample N = 192, the wear particle velocity is regarded as approximately equal to the oil velocity, and according to the frequency and velocity conversion formula, the wear particle frequency is approximately f d =106Hz. The specific process includes:
[0035] S21. Measure the length of the sensor detection area and the oil velocity to approximate the center frequency of the wear particles. The calculation formula is:
[0036] f d =v d / l, where v d ≈v l
[0037] In the formula, f d is the center frequency of the abrasive particle, v d is the abrasive velocity, v l is the oil flow velocity, l is the length of the sensor detection area.
[0038] S22. Performing a windowed Fourier transform on the original signal y(n) to be processed to obtain an amplitude spectrum of the original signal;
[0039] S23. Extract the amplitudes corresponding to all frequency points in the entire frequency band, perform discrete spectrum correction and normalization, extract the spectrum lines greater than 0.1 as harmonic signals, extract their amplitudes and phases to construct an anti-phase continuous sinusoidal signal and superimpose it with the original signal to achieve harmonic interference suppression;
[0040] S24. Based on the estimated wear induced voltage signal frequency f d and the estimated amplitudes {A1,A2,...,A Q} model, and use the wear particle signal theoretical model to generate Q-type simulated wear particle samples with different amplitudes. The wear particle signal model is expressed as:
[0041]
[0042] Among them, α is an adjustable parameter set according to the sensor coil structure parameters, Φ ris a simplified sensor structure parameter, and τ is the time delay of the center point of a single wear particle signal.
[0043] S25. Collect sufficient non-abrasive signals using sensors, and perform preprocessing such as harmonic suppression using the same parameters to form noise samples; add Gaussian white noise to the generated data with a varying signal-to-noise ratio to perform data enhancement, and finally complete data preprocessing.
[0044] S3. Estimate the center frequency of the wear particles and calculate the optimal number of decomposition layers of the stationary wavelet, segment the simulation signal segments of equal length and extract the J-layer decomposition approximate coefficients of equal wear particles and noise samples to form a multi-scale feature training vector set, and complete the labeling process;
[0045] Specifically, the steps of multi-scale feature extraction and training set construction include:
[0046] S31. According to the discrete wavelet transform (DWT) decomposition principle, the number of sample stationary wavelet transform (SWT) decomposition layers is calculated by the following formula:
[0047] J = floor((lnf s -lnαf d ) / ln2)-1
[0048] In the formula, f s is the sampling frequency, α is the expansion coefficient to prevent the wear particle characteristics from being destroyed in the decomposition. According to the wear particle estimation frequency, J is taken as 4 and α is taken as 2.
[0049] S32. For the u-th sample x u ∈R 1×N , where u = 1, 2, ..., U, and its j-th layer feature is expressed as:
[0050] η u,j =Ψ a (x u ; j)
[0051] In the formula, Ψ a represents the SWT approximation coefficient operator. Further, the original sample x u The multi-scale feature vector of can be expressed as:
[0052] η u =[η u,1 ,η u,2 ,...η u,j ,...,η u,J ]
[0053] S33. After labeling the training samples, the training set can be obtained:
[0054] Y train ={η u |yu}
[0055] Among them, y u is the label of the uth sample. When the sample contains wear particle features, y = 0; when the sample is an interference signal sample, y = 1.
[0056] S4. Use the parallel convolutional network combined with the fusion attention module to encode and reconstruct the four multi-scale feature vectors of the u-th sample respectively, and perform position encoding and initial classification vector insertion on the output four reconstructed high-dimensional vectors, use the self-attention model to establish the dependency relationship between vectors, and combine global and local information to obtain the final classification vector;
[0057] like Figure 4 As shown, the process of fusing the attention module to obtain global and local information is as follows:
[0058] S41: Output feature map F = [f1,f2,...,f i ,...,f I ], where f I For the I-th feature in the feature map, first calculate its channel global mean and global energy to obtain the fused channel attention. Its global mean and global energy are calculated by the following formula:
[0059]
[0060] Among them, z avg is the global mean, z pow is the global energy, M is the total number of feature maps, and f i (m) is the vector f i The mth instance in , I is the total number of channels;
[0061] Through vector concatenation and linear fusion to fuse channel attention, we get the global attention weight:
[0062]
[0063] Among them, g c is the global attention weight, δ is the Sigmoid activation function, Wc is the learnable linear transformation matrix, C(·) is the convolutional layer, is vector concatenation, k is the convolution kernel size, β c is the error term of the fully connected layer;
[0064] Based on the global attention weight, the feature map channels are weighted to obtain the global information vector:
[0065] F'=F·g c +F
[0066] Among them, F' is the global information vector;
[0067] S42. Use average pooling and energy pooling to obtain the local mean and local energy v of the feature map from the vector dimension respectively avg ∈R 1×M and v pow ∈R 1×M , and further use the convolution operation with a kernel length of 7 to obtain its fused local attention weight: g s =δ(C s ([v avg ;v pow ],7)), where g s is the local attention weight, C s (,7) is the local convolution operation with kernel length 7;
[0068] Based on the local attention weight, the feature map channel is weighted to obtain the local information vector:
[0069] F"=[f(i)·g s +f(i),i=1,2,...,I]
[0070] Where F'' is the local information vector, f(i) is the i-th feature in the feature map, and · represents the Hadamard product.
[0071] like Figure 5 As shown, the parallel convolutional network combined with the fusion attention module consists of 5 layers of weight-shared stacked convolution blocks, a flattening layer, a pooling layer, a fully connected layer, and a one-dimensional Vision Transformer (ViT) self-attention classification layer, wherein the weight-shared stacked convolution blocks are one-dimensional convolution + layer normalization + ReLU activation + fusion attention module;
[0072] For the input sample {η u ;y u}, first use weight-shared stacked convolution blocks to extract their multi-scale high-dimensional feature maps: H u =[h u,1 ,h u,2 ,...,h u,j ,...,h u,J ], where η u For sample x u The multi-scale feature vector, y u is the label of the u-th sample, H u Represents sample x u The multi-scale high-dimensional feature map, h u,J For sample x u The J-th layer of multi-scale high-dimensional features;
[0073] For H uEach instance in is flattened separately, and the average and maximum pooling with a kernel length of 4 and a step size of 4 are used to reduce the amount of data for each instance, and a shared weight fully connected layer is further used to obtain the reconstructed multi-scale feature vector: in, For sample x u The reconstructed multi-scale feature vector of For sample x u The Jth layer reconstructs multi-scale features;
[0074] The reconstructed multi-scale feature vector is embedded in the classification vector and positionally encoded, and then sent to the ViT model to model the global and local information dependency relationship of the multi-scale features, and finally the classification is completed through a multi-layer perceptron based on the fitted classification vector.
[0075] Specifically, the final classification vector length of any sample is 2, and Softmax is used to transform it into predicted label probability.
[0076] The structural parameters of the parallel convolutional network convolution module are shown in Table 1:
[0077] Table 1 Parallel convolutional network convolution module structure parameters
[0078]
[0079] The cross entropy loss function is used to fit the model, and the model with the best classification accuracy after multiple trainings is obtained as the wear particle signal identifier. A limited number of real wear particle signal samples are obtained with the proposed signal processing parameters, and the actual detection performance of the identifier is enhanced through feature transfer learning.
[0080] Specifically, the cross entropy function used in the example is defined as;
[0081]
[0082] Among them, θ u is the true label of the u-th sample, is the predicted label of the u-th sample.
[0083] In this example, the real wear particle samples are first pre-processed by low-pass filtering and harmonic elimination with the same preset parameters. After that, the key parameters such as the characteristic signal width and amplitude are estimated based on the estimated wear particle parameters, and a sliding signal window is constructed. The sliding similarity calculation of the real signal in this window is performed to obtain the similarity feature domain. According to the environmental parameters and structural parameters, 0.9 is selected as the similarity threshold and the serial number greater than the threshold is extracted to locate the real wear particle signal. Finally, under the premise of complete wear particle features, a real wear particle signal sample with a length of N = 192 is extracted, and multi-scale feature extraction is performed after labeling, and the model is further fitted based on feature transfer learning.
[0084] S6. Perform signal segmentation and multi-scale feature decomposition on the raw data collected in real time, and input the migrated network model to obtain the wear particle signal identification result;
[0085] Furthermore, the real-time data described in step S6 is first preprocessed with the same parameters, and then a window with a length of N and a step size of N / 2 is used for sliding segmentation to obtain the sample to be identified, and the sample prediction is completed through the fitted parallel convolutional network, the non-wear particle signal fragments are set to zero, and finally the wear particle identification is realized.
[0086] Specifically, the length of the segmented signal is N = 192, and the number of multi-scale feature decomposition layers is J = 4. In this example, a detection signal with 250,000 sampling points and 8 wear particles is used to verify the algorithm. Figure 6 The identification results are shown in Figure 7 It can be seen that a total of 8 wear particle induced voltage signals are accurately extracted, and the edge features of the wear particle signal are not destroyed.
[0087] S7. Reconstruct the wear signal based on the accurate model of the wear signal and the key features of the sample to obtain the noise-free wear detection signal, and finally realize the wear feature extraction.
[0088] Furthermore, in step S7, the wear signal reconstruction first obtains the numerical value and serial number of the maximum and minimum values of the sample, and calculates the amplitude parameter With the width parameter The feature reconstruction is realized based on the mathematical model of wear induced voltage. The specific calculation process of the parameters is expressed as:
[0089]
[0090] The reconstructed wear particle signal is as follows: Figure 8 As shown in the figure, it can be seen that the amplitude characteristics of the wear particle signal are proportional to the size of the wear particles passing through the sensor in sequence, and the width characteristics remain stable when the flow rate remains unchanged. Furthermore, by classifying and counting the wear particles in different amplitude ranges, it is convenient to analyze the wear particles separately in the future, so as to infer the wear state of the mechanical equipment.
[0091] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.
Claims
1. A wear particle signal recognition method based on fusion attention module and parallel convolutional network, characterized in that: include: S1: The lubricating oil wear particle monitoring system constructed by using the inductive wear particle detection sensor collects data of the lubricating oil liquid containing ferromagnetic wear particles to obtain the original signal to be processed; S2: Perform a windowed discrete Fourier transform on the collected original signal to obtain the amplitude frequency domain spectrum of the entire signal, and calculate the amplitude, frequency and phase of each low-frequency harmonic interference through iterative interpolation discrete Fourier transform algorithm and frequency domain compensation method to construct the anti-phase component for harmonic interference suppression and generate simulation samples; S3: Calculate the optimal decomposition layer number J of the stationary wavelet according to the center frequency of the wear particles, perform equal-length segmentation on the simulation samples according to the optimal decomposition layer number J of the stationary wavelet, extract the features of the wear particle signal and the simulation samples after equal-length segmentation to form a multi-scale feature training set, and complete the labeling process; S4: Use the parallel convolutional network combined with the fusion attention module to encode and reconstruct the J multi-scale feature vectors of a single sample in the training set, and perform position encoding and vector embedding on the output J reconstructed high-dimensional vectors. Use the energy self-attention model to establish the dependency relationship between the global and local information vectors, obtain the final classification vector, and use Softmax to transform it into the predicted label probability; S5: Construct a cross entropy loss function based on the predicted label probability, use the cross entropy loss function to fit the model, and obtain the model with the best classification accuracy after multiple trainings as the wear particle signal identifier. Then, obtain a limited number of real wear particle signal samples with the proposed signal processing parameters, and enhance the actual detection performance of the identifier through feature transfer learning. S6: Perform signal segmentation and multi-scale feature decomposition on the raw data collected in real time, and input the migrated network model to obtain the wear particle signal identification result; S7: Wear particle signal is reconstructed based on the accurate model of wear particle signal and key features of samples to obtain noise-free wear particle detection signal, and finally wear particle feature extraction is realized.
2. According to claim 1, a wear particle signal identification method based on fusion attention module and parallel convolutional network is characterized in that: The lubricating oil wear particle monitoring system constructed by the inductive wear particle detection sensor is composed of a wear particle sensor, a signal acquisition card, an oil tank, an ultrasonic flow velocity sensor, a peristaltic pump, a DC power supply, a wear particle receiving device and an experimental base. The signal acquisition card model is NI-9219, and the sampling frequency is set to f s =5000, the sampling time is 60s, and after removing the subsequent redundant signals, the wear particle detection signal with 280000 sampling points is obtained; the inductive wear particle detection sensor adopts DC excitation, the driving current I=0.5A, the amplifier multiple of the preamplifier is 2000, and the peristaltic pump flow rate is set to 510ml / min; the spherical equivalent diameters of the abrasive particles are 220μm, 73μm, 113μm, 65μm, 82μm, and three copper particles with an average of 400μm passing through the inductive oil wear particle detection sensor in turn, generating corresponding 8 induced voltage signals similar to a single-cycle sine waveform.
3. According to claim 1, a wear particle signal identification method based on fusion attention module and parallel convolutional network is characterized in that: The collected original signal is subjected to windowed discrete Fourier transform to obtain the amplitude frequency domain spectrum of the entire signal. The amplitude, frequency and phase of each low-frequency harmonic interference are calculated by iterative interpolation discrete Fourier transform algorithm and frequency domain compensation method to construct the anti-phase component for harmonic interference suppression and generate simulation samples, including: S21: Measure the length of the sensor detection area and the oil velocity to approximate the center frequency of the wear particle: f d =v d / l, where f d is the center frequency of the abrasive particle, v d is the abrasive velocity, v d ≈v l , v l is the oil flow velocity, l is the length of the sensor detection area; S22: Performing a windowed Fourier transform on the original signal y(n) to be processed to obtain an amplitude spectrum of the original signal; S23: extract the amplitudes corresponding to all frequency points in the entire frequency band, perform discrete spectrum correction and normalization processing, extract the spectrum lines greater than 0.1 as harmonic signals, extract their amplitudes and phases to construct anti-phase continuous sinusoidal signals and superimpose them with the original signals, so as to achieve harmonic interference suppression and generate simulated wear particle samples; S24: Based on the estimated wear induced voltage signal frequency f d and the estimated amplitudes {A1,A2,...,A Q } model, using the wear particle signal theoretical model to generate Q-type simulated wear particle samples with different amplitudes; S25: Collect sufficient non-abrasive signals using sensors, perform harmonic suppression with the same parameters to generate simulated non-abrasive samples; add Gaussian white noise to the generated data with a varying signal-to-noise ratio to perform data enhancement, and finally complete data preprocessing.
4. According to claim 3, a wear particle signal identification method based on fusion attention module and parallel convolutional network is characterized in that: The wear particle signal model comprises: Among them, ψ(n; A q ,Φ r ) is the wear particle signal model, α is an adjustable parameter set according to the sensor coil structure parameters, Φ r To simplify the sensor structure parameters, τ is the time delay of the center point of a single wear particle signal, n is the number of signal points, and A q is the amplitude of the qth wear-induced voltage signal, f s is the sampling frequency.
5. According to claim 1, a wear particle signal identification method based on fusion attention module and parallel convolutional network is characterized in that: The optimal decomposition level J of the stationary wavelet is calculated according to the center frequency of the wear particles, including: J=floor((lnf s -lnβf d ) / ln 2)-1 Among them, f s is the sampling frequency, β is the expansion coefficient to prevent the wear particle characteristics from being destroyed during decomposition, and according to the wear particle estimation frequency, the maximum decomposition layer J of the wavelet is 4, the frequency band expansion coefficient β is 2, and f d is the center frequency of the abrasive particles.
6. The wear particle signal identification method based on the fusion of attention module and parallel convolutional network according to claim 1 is characterized in that: The features of the wear particle signal and the simulated samples after equal length segmentation are extracted to form a multi-scale feature training set, and the labeling process is completed, including: The same amount of wear particle signals and simulation sample fragments are combined into a sample set. For the u-th sample x in the sample set, u , where u = 1, 2, ..., U, and its j-th layer feature is represented as: η u,j =Ψ a (x u ; j), its multi-scale feature vector is expressed as: η u =[η u,1 ,η u,2 ,...η u,j ,...,η u,J ], where U is the number of samples, η u,j For sample x u The j-th layer feature, Ψ a is the SWT approximation coefficient operator, η u For sample x u The multi-scale feature vector, η u,J For sample x u The J-th layer features of Perform sample labeling: Y train ={η u |y u }, where Y train is a labeled sample, y u is the label of the u-th sample. When the sample contains wear particle features, y=0; when the sample is an interference signal sample, y=1. All samples in the sample set are labeled to obtain the final training set.
7. The wear particle signal identification method based on the fusion of attention module and parallel convolutional network according to claim 1 is characterized in that: The parallel convolutional network combined with the fusion attention module consists of 5 layers of weight-shared stacked convolution blocks, flattening layers, pooling layers, fully connected layers, and a one-dimensional Vision The Transformer (ViT) self-attention classification layer consists of a weight-sharing stacked convolution block consisting of a one-dimensional convolution + layer normalization + ReLU activation + fused attention module. For the input sample {η u ;y u }, first use weight-shared stacked convolution blocks to extract their multi-scale high-dimensional feature maps: H u =[h u,1 ,h u,2 ,...,h u,j ,...,h u,J ], where η u For sample x u The multi-scale feature vector, y u is the label of the u-th sample, H u Represents sample x u The multi-scale high-dimensional feature map, h u,J For sample x u The J-th layer of multi-scale high-dimensional features; For H u Each instance in is flattened separately, and the average and maximum pooling with a kernel length of 4 and a step size of 4 are used to reduce the amount of data for each instance, and a shared weight fully connected layer is further used to obtain the reconstructed multi-scale feature vector: in, For sample x u The reconstructed multi-scale feature vector of For sample x u The Jth layer reconstructs multi-scale features; The reconstructed multi-scale feature vector is embedded in the classification vector and positionally encoded, and then sent to the ViT model to model the global and local information dependency relationship of the multi-scale features, and finally the classification is completed through a multi-layer perceptron based on the fitted classification vector.
8. The wear particle signal identification method based on the fusion of attention module and parallel convolutional network according to claim 8 is characterized in that: The reconstructed multi-scale feature vector is classified and embedded in the vector and positionally encoded, and then sent to the ViT model to model the global and local information dependency relationship of the multi-scale features, including: Among them, ζ is the output feature map of the ViT module, Te(·) is the Transformer layer operator in the ViT model, For sample x u The reconstructed multi-scale feature vector, λ u To insert the classification encoding vector, E pos is a classification vector.
9. According to the wear particle signal identification method based on fusion attention module and parallel convolution network according to claim 1, it is characterized in that: The global and local information vectors include: S41: Output feature map F = [f1,f2,...,f i ,...,f I ], where f I For the I-th feature in the feature map, first calculate its channel global mean and global energy to obtain the fused channel attention. Its global mean and global energy are calculated by the following formula: Among them, z avg is the global mean, z pow is the global energy, M is the total number of feature maps, and f i (m) is the vector f i The mth instance in , I is the total number of channels; Through vector concatenation and linear fusion to fuse channel attention, we get the global attention weight: Among them, g c is the global attention weight, δ is the Sigmoid activation function, W c is a learnable linear transformation matrix, C(·) is a convolutional layer, is vector concatenation, k is the convolution kernel size, β c is the error term of the fully connected layer; Based on the global attention weight, the feature map channels are weighted to obtain the global information vector: F'=F·g c +F Among them, F' is the global information vector; S42. Use average pooling and energy pooling to obtain the local mean and local energy v of the feature map from the vector dimension respectively avg ∈R 1×M and v pow ∈R 1×M , and further use the convolution operation with a kernel length of 7 to obtain its fused local attention weight: g s =δ(C s ([v avg ;v pow ],7)), where g s is the local attention weight, C s (,7) is the local convolution operation with kernel length 7; Based on the local attention weight, the feature map channel is weighted to obtain the local information vector: F”=[f(i)·g s +f(i),i=1,2,...,I] Where F'' is the local information vector, f(i) is the i-th feature in the feature map, and · represents the Hadamard product.
10. The wear particle signal identification method based on fusion attention module and parallel convolutional network according to claim 1 is characterized in that: The cross entropy function comprises: in, is the cross entropy function, U is the number of samples, θ u is the true label of the u-th sample, is the predicted label of the u-th sample.
Citation Information
Patent Citations
Protein secondary structure prediction method based on multi-scale convolution attention neural network
CN112767997A
El nino extreme weather early warning method and device based on incremental learning
US20230393304A1
Cited By
Wear detection method and system, terminal and storage medium
CN121980482A
