Oil abrasive particle feature recognition method based on convolutional neural network and signal matrix

Abstract

The invention discloses an oil abrasive particle feature recognition method based on a convolutional neural network and signal matrix, and the method comprises the steps: firstly carrying out the real-time data collection of lubricating oil through an induction type abrasive particle detection sensor, obtaining a to-be-processed original signal, carrying out the harmonic parameter estimation and the abrasive particle center frequency estimation of a detection signal, and carrying out the recognition of the abrasive particle center frequency; simulating the detection signal model according to the sensor output signal mathematical model, generating a plurality of groups of simulation abrasive particle signals by using the simulation signal model, setting parameters according to the abrasive particle center frequency, and performing low-pass filtering and harmonic interference suppression; compared with the traditional algorithm, the algorithm provided by the invention can further improve the signal-to-noise ratio of the abrasive particle signal, explores the application of the deep learning technology in the aspects of signal processing and feature extraction of the inductive abrasive particle sensor, and proposes a reliable data processing strategy and a complete novel abrasive particle detection operation framework. Effective abrasive particle voltage feature extraction is achieved, and the actual operation requirement of the induction type abrasive particle sensor is met.

Description

Technical Field

[0001] The technical field of oil wear particle monitoring specifically relates to an oil wear particle feature recognition method based on convolutional neural network and signal matrix. Background Art

[0002] Oil wear particle detection technology is an important branch of mechanical lubricating oil state detection technology, and plays a very important role in mechanical operation state estimation and life prediction. With the update and iteration of mechanical technology, many oil wear particle detection technologies with different principles have been developed and continuously innovated. Inductive oil wear particle sensors have become a type of sensor that has attracted much attention in the field of oil wear particle detection due to their high sensitivity to ferromagnetic wear particles, easy installation and low cost, and sensor performance is not easily affected by oil impurities. However, in actual mechanical and engineering applications, the induced electrical signals collected by the inductive oil wear particle sensors will inevitably contain interference components due to the unavoidable influence of external electromagnetic interference and mechanical vibration. If these interference components are not filtered out, the detection and quantitative analysis of wear particle signals are almost impossible to achieve.

[0003] In order to solve the above problems, many literatures have been used to improve the wear particle detection performance of inductive sensors by using methods such as signal processing and data feature engineering. Signal processing technology mainly relies on adaptive filtering, frequency band separation and other methods to improve the signal-to-noise ratio of wear particle electrical signals, so as to directly observe the wear particle voltage signal from the processed collected signal. Data feature engineering mainly includes steps such as value selection and eigenvalue calculation. Its principle is that the wear particle induction signal has a stable eigenvalue 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 achieved good results, the wear particle induction voltage identification relying on signal processing technology will inevitably change the morphological characteristics of the wear particles due to reasons such as the filtering threshold, the same noise frequency band and the wear particle frequency band, and the protection of the wear particle characteristics still cannot meet the requirements of quantitative feature extraction and analysis of the wear particles. The performance of methods that rely on data features will be affected by the rationality of feature value selection, the high similarity between noise morphology and wear particle features, etc., which will affect the wear particle identification effect of the algorithm, bringing challenges to the use and advancement of the technology. In addition, manual feature extraction is often an extremely complex process that requires a wealth of professional knowledge accumulation and a lot of mathematical derivation. These problems may limit the practical application of wear particle sensors. Summary of the invention

[0004] In order to solve the above technical problems, the present invention provides an oil wear particle feature recognition method based on convolutional neural network and signal matrix, comprising the following steps:

[0005] S1: The lubricating oil wear particle monitoring system constructed by using the inductive wear particle detection sensor collects real-time data of the lubricating oil containing ferromagnetic wear particles to obtain the original signal to be processed;

[0006] S2: Accurately estimate the harmonic parameters of the original signal based on spectrum correction, and estimate the center frequency of the oil wear particle signal according to the oil flow velocity;

[0007] S3: According to the obtained signal parameters, a mathematical model based on the sensor output signal is constructed to generate multiple groups of simulated wear particle detection signals with different wear particle signal amplitudes, indexable positions, and random noises;

[0008] S4: low-pass filtering and harmonic elimination are performed on the simulated wear particle detection signal, and each group of simulated detection signals is segmented using the semi-overlapping segmentation method according to the wear particle signal center frequency setting parameters, and the simulated wear particle signal segments are extracted according to the wear particle signal position index value, and the corresponding number of noise segments are extracted and labeled;

[0009] S5: Each extracted simulated wear signal segment and noise segment is decomposed by J-level using stationary wavelet transform, and the approximate components after the four-level stationary wavelet transform decomposition are extracted to form the characteristic scale vector F = [a J-3 ,a J-2 ,a J-1 ,a J ] and use the Gram Angular Field (GADF) to transform each scale vector in the vector F to obtain a two-dimensional image. The four component images are concatenated into one image and then the maximum pooling is performed to generate the network training data set;

[0010] S6: Use the generated network training data set to train the convolutional neural network model to obtain a wear particle signal identifier with high recognition accuracy;

[0011] S7: The original signal to be processed is subjected to low-pass filtering, harmonic elimination, segmentation and signal matrixing according to the same hyperparameters, and the trained network is used to identify wear particles, eliminate noise segments, and complete repetitive elimination;

[0012] S8: According to the recognition result, the corresponding one-dimensional wear particle signal is returned and de-duplication operation is performed, and a sliding filter based on a Gaussian window function is used to perform a signal smoothing operation, and then the ideal wear particle signal based on the wear particle mathematical model is used to locate the wear particle, further eliminate noise interference and complete high-precision wear particle quantitative feature extraction;

[0013] S9: Reconstruct the detection signal based on the wear particle fragments and the corresponding index positions to obtain a noise-free wear particle detection signal, and finally complete the accurate counting of the wear particles.

[0014] Compared with the prior art, the present invention has the following beneficial effects:

[0015] After achieving background noise suppression of the collected signal, the present invention can realize wear feature extraction based on deep learning technology while maximally protecting the quantitative characteristics of the wear signal without relying on additional filtering algorithms and feature extraction methods. At present, traditional algorithms in this field will inevitably cause distortion of target signal characteristics. The most prominent problem is that it will change the amplitude of the wear voltage signal, thereby affecting the judgment of the health status of the machinery and failing to provide a reliable basis for the subsequent wear evolution analysis of the machinery. The oil wear feature recognition method based on convolutional neural networks and signal matrixing proposed in the present invention can use deep learning technology to quickly and accurately identify and classify wear features without damaging the wear features and retaining the original features of the signal to the maximum extent, thereby realizing quantitative analysis of wear features; in addition, the use of convolutional neural networks to realize wear feature extraction avoids the problems of large amount of calculation and long calculation time caused by the use of complex traditional feature extraction algorithms, and also reduces the requirements for users' professional knowledge in the field of signal processing;

[0016] The present invention can generate reliable model training data according to the detection signal parameters, effectively reducing the demand for detection data with wear particles and noise labels, further alleviating the problem of difficulty in obtaining labeled training data due to the difficulty of sensor experiments, improving the efficiency of algorithm use, lowering the usage threshold, and broadening the application scenarios; at the same time, compared with traditional algorithms, the present invention can further improve the signal-to-noise ratio of wear particle signals, explore the application of deep learning technology in signal processing and feature extraction of inductive wear particle sensors, and propose a reliable data processing strategy and a complete new wear particle detection operation framework, to achieve effective wear particle voltage feature extraction, and meet the actual operation needs of inductive wear particle sensors. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 A flow chart of a method for identifying oil wear particle characteristics based on a convolutional neural network and signal matrixing according to the present invention;

[0018] Figure 2 This is a schematic diagram of the lubricating oil wear particle monitoring system;

[0019] Figure 3 Schematic diagram of the original signal to be processed and the signal after harmonic elimination extracted by the wear monitoring system;

[0020] Figure 4A schematic diagram of a simulated detection signal generated according to parameter estimation and a simulated detection signal after harmonic elimination;

[0021] Figure 5 Schematic diagram of the reconstructed detection signal after sliding filtering and wear particle positioning. 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 method for identifying oil wear particle characteristics based on convolutional neural network and signal matrix. Figure 1 As shown, the following steps are included:

[0024] S1. A lubricating oil wear particle monitoring system constructed by using an inductive wear particle detection sensor performs real-time data acquisition on the lubricating oil liquid containing ferromagnetic wear particles to obtain the original signal to be processed;

[0025] The process of collecting raw signals of lubricating oil wear particle monitoring system is as follows: 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, sampling time is 70s, and after removing the subsequent redundant signals, the wear particle detection signal with sampling points N=350000 is obtained. The inductive wear particle detection sensor uses DC excitation, 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 wear particle sizes (equivalent diameter of the sphere) are 220mm, 80mm, 50mm, 150mm, 60mm, and 210mm respectively. The sensor will generate 6 corresponding induced voltage signals similar to a single-cycle sine waveform.

[0026] The collected raw signal to be processed is as follows Figure 3 As shown in the figure, 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.

[0027] S2. Accurately estimate the harmonic parameters of the original signal based on spectrum correction, and estimate the center frequency of the oil wear particle signal according to the oil flow velocity;

[0028] 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 by an iterative interpolation discrete Fourier transform algorithm and a frequency domain compensation method; the amplitude spectrum is normalized, and the spectrum lines with normalized amplitudes greater than 0.1 are extracted as effective harmonic components and the corresponding phase values ​​in the phase spectrum are extracted to obtain the parameters γ(a j ,f j ,θ j ), where j is the serial number of each harmonic component. The wear particle velocity is considered to be approximately equal to the oil velocity. According to the frequency and velocity conversion formula, the wear particle frequency is approximately f d =67Hz:

[0029] Measure the length of the sensor detection area and the oil velocity to approximate the center frequency of the wear particle. The calculation formula is: fd = vd / l, where vd ≈ vl; where fd is the center frequency of the wear particle, vd is the wear particle velocity, vl is the oil flow velocity, and l is the length of the sensor detection area;

[0030] Perform a windowed Fourier transform on the original signal y(n) to be processed to obtain the original signal amplitude spectrum;

[0031] Extract the amplitudes corresponding to all frequency points in the entire frequency band, perform discrete spectrum correction and normalization, and extract values ​​greater than 0.2 as harmonic signals, and values ​​less than 0.05 as noise;

[0032] Then the original signal is subjected to phase spectrum change to obtain the phase spectrum, and the corresponding phase is extracted according to the frequency corresponding to the amplitude of each harmonic signal in the amplitude spectrum, thereby forming multiple sets of harmonic signal parameters γ(a j ,f j ,θ j ), where j is the index of the extracted harmonic;

[0033] S3. Building a mathematical model based on the sensor output signal according to the obtained signal parameters to generate multiple groups of simulated wear particle detection signals with different wear particle signal amplitudes, indexable positions, and random noises;

[0034] According to the sensor output signal model, the generated simulation detection signal includes the ideal wear particle signal, the harmonic component formed by the linear superposition of multiple sinusoidal signals with different frequencies composed of the harmonic parameters measured by the harmonic parameter estimation method, and the random noise produced according to the preset signal-to-noise ratio. The wear particle signal and the harmonic wave are expressed by the following formulas, where the harmonic classification is multiplied by the correction coefficient α=4;

[0035] For wear particle signal:

[0036] Where nd is the number of wear particle signals, T i is the position of the i-th wear particle in the sequence, η is a set random variable, and its value range is [0.9,1.1]. w(n) is a rectangular window function, and its value satisfies:

[0037] For harmonics: Where n h is the number of harmonic classifications; the length of the signal is 50000 points;

[0038] The generated wear particle detection signal is as follows: Figure 4 As shown in the figure, it can be seen that after parameter estimation, the harmonic interference in the simulation signal has restored the harmonics in the detection signal to a great extent. After harmonic elimination, the simulated harmonic interference is destroyed, and the residual noise component is close to the real detection signal after harmonic elimination;

[0039] S4. Perform low-pass filtering and harmonic elimination on the generated simulation signal, segment each group of simulation detection signals using the semi-overlap segmentation method according to the center frequency setting parameters of the wear particle signal, extract the simulated wear particle signal segment according to the wear particle signal position index value, and then extract the corresponding number of noise segments and label them;

[0040] Perform a low-pass filter on the signal, where the cutoff frequency f c =2f d To protect the wear particle features from damage; the fixed window length semi-overlap sliding segmentation method is used for sample segmentation, where the window length is L w , the sliding step length is L w / 2;

[0041] The calculation formula for window length is:

[0042] Where ceil(·) means rounding up, s d ∈[2.5,3] is the expansion coefficient set to ensure the integrity of the abrasive features;

[0043] After calculation, Lw = 224. After completing the signal segmentation, for the generated sequence with a length of 50,000, the number of samples after segmentation is: Where floor(·) is rounded down to obtain M signal samples;

[0044] By dividing, we can obtain a collection of equal-length signal segments X = {X 1 ,X 2 ,...,X m}, where m = 1, 2, ..., 444. For a single vector X = (x 1 ,x2 ,...,x p ), where p = 1, 2, ..., 224;

[0045] S5. For each extracted simulated wear signal segment and noise segment, the stationary wavelet transform is used to perform J-level decomposition, and the approximate components after the four-level stationary wavelet transform decomposition are extracted to form a characteristic scale vector F = [a J-3 ,a J-2 ,a J-1 ,a J ] and use the Gram Angular Field (GADF) to transform each scale vector in the vector F to obtain a two-dimensional image. The four component images are concatenated into one image and then the maximum pooling is performed to generate the network training data set;

[0046] Use stationary wavelet transform to perform J-level decomposition on any signal segment, and convert the approximate component a obtained from the stationary wavelet transform into J-3 、a J-2 、a J-1 and a J All extracted to form a multi-scale vector can be expressed as:

[0047] F=[a J-3 ,a J-2 ,a J-1 ,a J ]

[0048] Four two-dimensional images A are obtained by performing Gram angle field transformation on each component of vector F J-3 ,

[0049] A J-2 , A J-1 and A J ;

[0050] The calculation formula of the Gram angle field is as follows:

[0051] First, normalize the vector:

[0052] To transform the coordinate system:

[0053] Perform Gram angular field matrix operations:

[0054] Since the length of the approximate signal after stationary wavelet transform is the same as that of the original signal, the side length of the transformed two-dimensional image is also T. The four images are spliced ​​together to obtain an image G with a side length of 2T, which can be expressed as:

[0055]

[0056] The concatenated image G is further pooled using a 2×2 convolution kernel to generate a T×T image. 1 ,x 2 ,...,x p}, where p = 1, 2, ..., L w , the corresponding multi-scale feature matrix set as the data for training the neural network is represented as G = {G 1 ,G 2 ,...,G p};

[0057] The wear particle segments and the same number of random noise segments are extracted to perform the above operations to complete the sample set preparation, where the wear particle segments are marked as 0 and the noise segments are marked as 1.

[0058] S6. Using the generated network training data set to train the convolutional neural network model, a wear particle signal identifier with high recognition accuracy is obtained;

[0059] Specifically, the generated network training data set is used to train the convolutional neural network model to obtain a wear particle signal identifier with high recognition accuracy. A deep convolutional neural network such as ResNet50 is used to extract higher-dimensional hidden features of the sample, enhance the distinction between the wear particle signal and the noise, and achieve the effect of improving the accuracy of the identifier.

[0060] S7. The original signal to be processed is subjected to low-pass filtering, harmonic elimination, segmentation and signal matrixing according to the same hyperparameters, and the trained network is used to identify wear particles, eliminate noise segments, and complete repetitive elimination;

[0061] Specifically, when performing step S7, the low-pass filter cutoff frequency is set to f c =134, segment partition window length L w = 224. After completing the identification and deduplication operation, 6 wear particle fragments were identified;

[0062] S8. According to the recognition result, the corresponding one-dimensional wear particle signal is returned and deduplication operation is performed, a sliding filter based on a Gaussian window function is used to perform a signal smoothing operation, and then an ideal wear particle signal based on a wear particle mathematical model is used to locate the wear particle, further eliminate noise interference and complete high-precision wear particle quantitative feature extraction;

[0063] The specific steps of abrasive particle positioning are:

[0064] First, the segments with continuous indexes are spliced, and the overlapping parts are multiplied by a coefficient of 0.5 to ensure that the wear particle signal is not distorted;

[0065] The signal is smoothed by sliding filtering, and the frequency f is estimated based on the wear particles. d Generate a sequence with an amplitude of 1 and a length of The ideal wear particle signal;

[0066] This signal is used as a window to slide on the one-dimensional signal sample with a step size of 1, and the normalized cosine similarity calculation with the equal-length segments is performed to obtain the wear particle position weight. 1 With η 1 , the formula for calculating normalized cosine similarity is:

[0067] in, and is a normalized equal-length vector. Its normalization process can be expressed as:

[0068] For any sample vector X, its wear particle position weight can be expressed as

[0069] The number g corresponding to the maximum weight value max(P) is taken as the position of the wear particle in the vector. The weight vector is obtained by constructing a window function and the non-zero interval is extended by 1 / 10 to both sides to complete the wear particle positioning.

[0070] Among them, the weight vector is expressed as:

[0071]

[0072] By dividing the vector X by w p Multiplication achieves further denoising while protecting wear particle features;

[0073] S9. Reconstruct the detection signal based on the wear particle fragments and the corresponding index positions to obtain a noise-free wear particle detection signal, and finally complete the wear particle accurate counting. The reconstructed signal is as follows Figure 5 Place;

[0074] Specifically, in the reconstruction of the detection signal of the wear particle fragments and the corresponding index positions, the signal is constructed into a sequence with the same length as the detection signal and accurate wear particle position by zero padding, and the signal reconstruction is completed by linear superposition.

[0075] 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 method for oil wear particle feature recognition based on convolutional neural network and signal matrix, characterized in that: The following steps are involved: S1: The lubricating oil wear particle monitoring system constructed by using the inductive wear particle detection sensor collects real-time data of the lubricating oil containing ferromagnetic wear particles to obtain the original signal to be processed; S2: Accurately estimate the harmonic parameters of the original signal based on spectrum correction, and estimate the center frequency of the oil wear particle signal according to the oil flow velocity; S3: According to the obtained signal parameters, a mathematical model based on the sensor output signal is constructed to generate multiple groups of simulated wear particle detection signals with different wear particle signal amplitudes, indexable positions, and random noises; S4: low-pass filtering and harmonic elimination are performed on the simulated wear particle detection signal, and each group of simulated detection signals is segmented using the semi-overlapping segmentation method according to the wear particle signal center frequency setting parameters, and the simulated wear particle signal segments are extracted according to the wear particle signal position index value, and the corresponding number of noise segments are extracted and labeled; S5: Each extracted simulated wear signal segment and noise segment is decomposed by J-level using stationary wavelet transform, and the approximate components after the four-level stationary wavelet transform decomposition are extracted to form the characteristic scale vector F = [a J-3 ,a J-2 ,a J-1 ,a J ] and use the Gram Angular Field (GADF) to transform each scale vector in the vector F to obtain a two-dimensional image. The four component images are concatenated into one image and then the maximum pooling is performed to generate the network training data set; S6: Use the generated network training data set to train the convolutional neural network model to obtain a wear particle signal identifier with high recognition accuracy; S7: The original signal to be processed is subjected to low-pass filtering, harmonic elimination, segmentation and signal matrixing according to the same hyperparameters, and the trained network is used to identify wear particles, eliminate noise segments, and complete repetitive elimination; S8: According to the recognition result, the corresponding one-dimensional wear particle signal is returned and de-duplication operation is performed, and a sliding filter based on a Gaussian window function is used to perform a signal smoothing operation, and then the ideal wear particle signal based on the wear particle mathematical model is used to locate the wear particle, further eliminate noise interference and complete high-precision wear particle quantitative feature extraction; S9: Reconstruct the detection signal based on the wear particle fragments and the corresponding index positions to obtain a noise-free wear particle detection signal, and finally complete the accurate counting of the wear particles.

2. The oil wear particle feature recognition method based on convolutional neural network and signal matrix according to claim 1 is characterized in that: In step S2: The length of the sensor detection area and the oil velocity are measured to approximate the center frequency of the wear particle. The calculation formula is: f d =v d / l, where v d ≈v l Where, 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; Perform windowed Fourier transform on the original signal y(n) to be processed to obtain the amplitude spectrum of the original signal; Extract the amplitudes corresponding to all frequency points in the entire frequency band, perform discrete spectrum correction and normalization, and extract values ​​greater than 0.2 as harmonic signals, and values ​​less than 0.05 as noise; Then the original signal is subjected to phase spectrum change to obtain the phase spectrum, and the corresponding phase is extracted according to the frequency corresponding to the amplitude of each harmonic signal in the amplitude spectrum, thereby forming multiple sets of harmonic signal parameters γ(a j ,f j ,θ j ), where j is the number of the extracted harmonics, a j is the harmonic component amplitude factor, f j is the harmonic component frequency, θ j is the phase factor of the harmonic component.

3. The oil wear particle feature recognition method based on convolutional neural network and signal matrix according to claim 1 is characterized in that: In step S3: The simulation detection signal generated according to the mathematical model of the sensor output signal includes an ideal wear particle signal, a harmonic component formed by linear superposition of multiple sinusoidal signals with different frequencies composed of harmonic parameters measured based on the harmonic parameter estimation method, and random noise produced according to a preset signal-to-noise ratio. The wear particle signal and harmonics are expressed by the following formulas, where the harmonic classification is multiplied by the correction coefficient α=4; Wear particle signal: Among them, f s is the sampling frequency, n d is the number of wear particle signals, T i is the position of the i-th wear particle in the sequence, η is a set random variable with a value range of [0.9, 1.1]; w(n) is a rectangular window function whose value satisfies: harmonic: Among them, n h is the number of harmonic classifications; the length of the signal is 50,000 points, a j is the harmonic component amplitude factor, f j is the harmonic component frequency, θ j is the phase factor of the harmonic component.

4. The oil wear particle feature recognition method based on convolutional neural network and signal matrix according to claim 1 is characterized in that: In step S4: Perform a low-pass filter on the signal, where the cutoff frequency f c =2f d To protect the abrasive features from damage, fd is the center frequency of the abrasive; The sample segmentation is performed using the semi-overlapping sliding segmentation method with a fixed window length, where the window length is The sliding step length is L w / 2, where f s is the sampling frequency, ceil(·) means rounding up, s d ∈[2.5,3] represents the expansion coefficient set to ensure the integrity of the wear particle features. After completing the signal segmentation, for the generated sequence of length N, the number of samples after segmentation is: Where M is the number of samples after segmentation, and floor(·) is rounded down; By dividing, we can obtain a set of equal-length signal segments X = {x1, x2, ..., x m }, where x m is the mth signal segment, m = 1, 2, ..., M, for a single vector x = (y1, y2, ..., y p ), where y p is the pth instance element in the vector, p=1,2,...,L w .

5. The oil wear particle feature recognition method based on convolutional neural network and signal matrix according to claim 1 is characterized in that: In step S5: Use stationary wavelet transform to perform J-level decomposition on any signal segment, and obtain four approximate components a from the stationary wavelet transform J-3 、a J-2 、a J-1 and a J Composed of multi-scale vector F = [a J-3 ,a J-2 ,a J-1 ,a J ]; Four two-dimensional images are obtained by performing Gram angle field transformation on each component of the multi-scale vector F A J-3 , A J-2 , A J-1 Japanese A J ; Since the approximate signal after stationary wavelet transform has the same length as the original signal, the side length of the transformed two-dimensional image is also T. The four images are spliced ​​to obtain an image with a side length of 2T. Further, a 2×2 convolution kernel is used to perform maximum pooling on the spliced ​​image G to generate a T×T image; The equal-length signal segment collection X = {x1, x2, ..., x m } as the data for training the neural network, and obtain the corresponding multi-scale feature matrix set θ = {G1, G2, ..., G M }, where x m is the mth signal segment, G M is the Mth instance element of the multi-scale feature matrix set; Abrasive grain fragments and an equal number of random noise fragments are extracted to perform the above operation to complete the sample set production, and the abrasive grain fragments are marked as 0 and the noise fragments are marked as 1.

6. The oil wear particle feature recognition method based on convolutional neural network and signal matrix according to claim 5 is characterized in that: The calculation formula of the Gram angle field is as follows: First, normalize the vector: in, is the normalized vector, x m is the original vector; To transform the coordinate system: Among them, θ p is the angle in the angular coordinate system, r p is the radius in the angular coordinate system, P is the length of the fragment sequence, t p The timestamp corresponding to the current element; Perform Gram angular field matrix operations: Among them, m is the serial number, I is the unit vector, is the normalized vector, for The transpose of .

7. The oil wear particle feature recognition method based on convolutional neural network and signal matrix according to claim 1 is characterized in that: In step S7, the cutoff frequency of the low-pass filter is set to f c =2f d The harmonic elimination method is to perform frequency domain parameter estimation and spectrum correction on the entire signal, construct an inverted signal and superimpose it with the original signal to suppress harmonics. The window length of the segment division is L w , the step length is L w / 2, where f d is the center frequency of the abrasive particles.

8. The oil wear particle feature recognition method based on convolutional neural network and signal matrix according to claim 1 is characterized in that: The specific steps of positioning the abrasive particles in step S8 are: First, the segments with continuous indexes are spliced, and the overlapping parts are multiplied by a coefficient of 0.5 to ensure that the wear particle signal is not distorted; The signal is smoothed by sliding filtering, and the frequency f is estimated based on the wear particles. d Generate a sequence with an amplitude of 1 and a length of The ideal wear particle signal is: s is the sampling frequency, f d is the center frequency of the abrasive particles; The signal is used as a window to slide on the one-dimensional signal sample with a step size of 1, and the normalized cosine similarity calculation with the equal-length segment is performed to obtain the wear particle position weight. For equal-length vectors η1 and η2, the formula for the wear particle position weight calculated by normalized cosine similarity is: Among them, p(η1,η2) is the wear particle position weight for equal length vectors η1 and η2, is the normalized vector η1, is the normalized vector η2; For any vector The wear particle position weight can be expressed as Take the number g corresponding to the maximum weight value max(Q) as the position of the wear particle in the vector, construct the window function to obtain the weight vector and extend the non-zero interval to both sides by 1 / 10 to complete the wear particle positioning, where the weight vector w m The Boolean value of the pth instance in can be calculated by the following formula: Among them, l w is the window length for segment division, l s is the length of the identification window interval, q g is a single weight value, is the last weight value in the current sample; By adding the vector With the weight vector w m A Hadamard product operation is performed to achieve further denoising while preserving the wear particle features.

9. The oil wear particle feature recognition method based on convolutional neural network and signal matrix according to claim 1 is characterized in that: In step S9, in the detection signal reconstruction based on the wear particle fragments and the corresponding index positions, the signal is constructed into a sequence with the same length as the detection signal and accurate wear particle positions by zero padding, and the signal reconstruction is completed by linear superposition.

Images