An oil particle feature recognition method based on convolutional neural network and signal matrixing

Abstract

The application discloses an oil abrasive particle feature recognition method based on a convolutional neural network and signal matrixing, which firstly uses an inductive abrasive particle detection sensor to collect real-time data of lubricating oil, obtains original signals to be processed, estimates harmonic parameters and abrasive particle center frequency of the detection signals, simulates a detection signal model according to a sensor output signal mathematical model, generates a plurality of groups of simulation abrasive particle signals using the simulation signal model, sets parameters according to the abrasive particle center frequency, and carries out low-pass filtering and harmonic interference suppression; compared with a traditional algorithm, the algorithm can further improve the signal-to-noise ratio of the abrasive particle signals, explores the application of deep learning technology in signal processing and feature extraction of the inductive abrasive particle sensor, and proposes a reliable data processing strategy and a complete new abrasive particle detection operation framework, realizes effective abrasive particle voltage feature extraction, and meets the actual operation needs of the inductive abrasive particle sensor.

Description

TECHNICAL FIELD

[0001] The technical field relates to oil wear particle monitoring, and particularly to an oil wear particle feature recognition method based on a convolutional neural network and signal matrixing. BACKGROUND

[0002] Oil wear particle detection technology is an important branch of mechanical lubricating oil condition 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 of different principles have been developed and continuously innovated. Inductive oil wear particle sensors have become a kind of sensor that attracts much attention in the field of oil wear particle detection today due to their high sensitivity to ferromagnetic particles, easy installation, low cost, and sensor performance not easily affected by oil impurities. However, in actual mechanical and engineering field applications, the inductive oil wear particle sensor will inevitably contain interference components in the inductive electric signal collected due to unavoidable external electromagnetic interference and mechanical vibration. If these interference components are not filtered out, the detection and quantitative analysis of the wear particle signal will be almost impossible.

[0003] In order to solve the above problems, many documents have used signal processing and data feature engineering methods to improve the wear particle detection performance of inductive sensors. Signal processing technology mainly relies on adaptive filtering, frequency band separation and other methods to improve the signal-to-noise ratio of the wear particle electric signal, so that the wear particle voltage signal can be directly observed from the processed collected signal. Data feature engineering mainly includes numerical selection and feature value calculation steps, and its principle is that the wear particle inductive signal has stable feature values that are highly distinguishable from noise. The main wear particle signal extraction process is mainly composed of segment division, feature value calculation and threshold classification. To evaluate the mechanical operation state, it is necessary to accurately count the ferromagnetic wear particles of different sizes, so as to provide a basis for 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 inductive voltage recognition relying on signal processing technology will inevitably change the morphology of the wear particles due to filtering threshold, noise frequency band and wear particle frequency band, etc. The protection of wear particle features still cannot meet the requirements of wear particle quantitative feature extraction and analysis. The performance of the method relying on data features will be affected by whether the feature values are selected reasonably, the similarity between noise morphology features and wear particle features, etc., which will affect the wear particle recognition effect of the algorithm and bring challenges to the use and promotion of the technology. Moreover, artificial feature extraction is often a very complex process that requires rich professional knowledge accumulation and a lot of mathematical derivation. These problems may limit the practical application of wear particle sensors. SUMMARY

[0004] To solve the above technical problems, the application provides an oil abrasive particle feature recognition method based on a convolutional neural network and signal matrixing, comprising the following steps:

[0005] S1: using a lubricating oil abrasive particle monitoring system constructed by an inductive abrasive particle detection sensor to collect real-time data of lubricating oil containing ferromagnetic wear particles, and obtaining a raw signal to be processed;

[0006] S2: performing harmonic parameter accurate estimation based on spectrum correction on the raw signal, and estimating the center frequency of the oil abrasive particle signal according to the oil flow velocity;

[0007] S3: according to the obtained signal parameters, building a simulation abrasive particle detection signal with different abrasive particle signal amplitudes, indexable positions and random noise based on a mathematical model of the sensor output signal;

[0008] S4: performing low-pass filtering and harmonic elimination on the simulation abrasive particle detection signal, using a half-overlapping segmentation method to segment each simulation detection signal according to the abrasive particle signal center frequency setting parameter, extracting the simulation abrasive particle signal segment according to the abrasive particle signal position index value, and randomly extracting a corresponding number of noise segments and performing label processing;

[0009] S5: using stationary wavelet transform to perform J-level decomposition on each extracted simulation abrasive particle signal segment and noise segment, extracting the approximate components after the last four-level stationary wavelet transform decomposition to form a feature scale vector F=[a J-3 ,a J-2 ,a J-1 ,a J ], and using a Gram angle field (GADF) to transform each scale vector in the vector F to obtain a two-dimensional image. After splicing the images of the four components into one image, maximum pooling is performed to generate a network training data set;

[0010] S6: using the generated network training data set to train a convolutional neural network model to obtain an abrasive particle signal recognizer with high recognition accuracy;

[0011] S7: performing low-pass filtering, harmonic elimination, segment division and signal matrixing on the raw signal to be processed according to the same hyperparameters, and using the trained network to recognize the abrasive particles, exclude noise segments, and complete the repeatability exclusion;

[0012] S8: returning the corresponding one-dimensional abrasive particle signal according to the recognition result and performing a de-duplication operation, using a sliding filter based on a Gaussian window function to perform signal smoothing, and using an ideal abrasive particle signal based on an abrasive particle mathematical model to perform abrasive particle positioning, further excluding noise interference and completing high-precision abrasive particle quantitative feature extraction;

[0013] S9: Based on the abrasive particle fragment and the corresponding index position, the detection signal is reconstructed to obtain the noise-free abrasive particle detection signal, and the accurate counting of the abrasive particle is finally completed.

[0014] Compared with the prior art, the beneficial effects of the present application are:

[0015] The present application can realize abrasive feature extraction based on deep learning technology while protecting the quantitative features of abrasive signals to the maximum extent after suppressing the background noise of the collected signals without relying on additional filtering algorithms and feature extraction methods. Currently, traditional algorithms in the field inevitably cause distortion of target signal features, and the most prominent problem is that the amplitude of the abrasive voltage signal is changed, thereby affecting the judgment of the mechanical health status and failing to provide reliable basis for subsequent mechanical wear evolution analysis. The oil abrasive feature recognition method based on convolutional neural network and signal matrixing can quickly and accurately identify and classify abrasive features using deep learning technology without damaging abrasive features and maximizing the preservation of original signal features, thereby realizing quantitative analysis of abrasive features. In addition, the use of convolutional neural network for abrasive feature extraction avoids the problems of large calculation amount and long calculation time caused by the use of complex traditional feature extraction algorithms, and reduces the requirement for users' professional knowledge in the field of signal processing.

[0016] The present application can generate reliable model training data according to the detection signal parameters, effectively reduce the demand for detection data with abrasive particle and noise labels, further alleviate the problem of difficulty in obtaining labeled training data caused by sensor experiments, improve the use efficiency of the algorithm, reduce the use threshold, and broaden the application scenarios. At the same time, compared with traditional algorithms, the present application can further improve the signal-to-noise ratio of abrasive signals, explore the application of deep learning technology in inductive abrasive sensor signal processing and feature extraction, and propose a reliable data processing strategy and complete new abrasive detection operation framework to realize effective abrasive voltage feature extraction and meet the actual operation needs of inductive abrasive sensors. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 The flowchart of the oil abrasive feature recognition method based on convolutional neural network and signal matrixing of the present application is shown in the figure.

[0018] Figure 2 The schematic diagram of the lubricating oil abrasive monitoring system is shown in the figure.

[0019] Figure 3 The schematic diagram of the original signal to be processed and the signal after harmonic elimination extracted by the abrasive monitoring system is shown in the figure.

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

[0021] Figure 5 A schematic diagram of a reconstructed detection signal after sliding filtering and abrasive positioning. DETAILED DESCRIPTION

[0022] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0023] The present application provides an oil particle feature recognition method based on convolutional neural network and signal matrixing, as shown in Figure 1 The method comprises the following steps:

[0024] S1. A lubricating oil particle monitoring system constructed by an inductive abrasive particle detection sensor is used to collect real-time data of lubricating oil containing ferromagnetic wear particles, and obtain a raw signal to be processed;

[0025] The process of collecting the raw signal by the lubricating oil particle monitoring system is as follows, and a platform schematic diagram of the lubricating oil particle monitoring system is as shown in Figure 2 The model of the acquisition card is NI-9219, the sampling frequency is set to f s = 5000, the sampling time is 70s, and the number of sampling points of the abrasive particle detection signal obtained after removing the subsequent redundant signal is N = 350000. The inductive abrasive particle detection sensor adopts direct current excitation, the driving current I = 0.5A, the amplifier multiple of the preamplifier is 2000, and the flow rate of the peristaltic pump is set to 510ml / min. The abrasive particle size (spherical equivalent diameter) passes through the inductive abrasive particle detection sensor in the order of 220mm, 80mm, 50mm, 150mm, 60mm and 210mm, and the sensor will generate 6 inductive voltage signals similar to a single-cycle sinusoidal waveform.

[0026] The collected raw signal to be processed is as shown in Figure 3 It can be seen that the raw signal to be processed collected in real time not only contains the abrasive particle inductive voltage signal, but also contains harmonic interference caused by structural vibration and random noise caused by electromagnetic waves and the like, so that the morphological characteristics of the tiny abrasive particle inductive voltage signal cannot be quickly and accurately extracted

[0027] S2. The raw signal is subjected to harmonic parameter accurate estimation based on spectrum correction, and the center frequency of the oil particle signal is estimated according to the oil flow velocity;

[0028] The collected original signal is first windowed and discrete Fourier transformed to obtain the amplitude frequency spectrum of the whole 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. The amplitude spectrum is normalized, and the spectral line with normalized amplitude greater than 0.1 is extracted as the effective harmonic component, and the corresponding phase value in the phase spectrum is extracted to obtain the harmonic component parameters γ(a j ,f j ,θ j ), where j is the serial number of each harmonic component. The abrasive particle speed is approximately equal to the oil speed, and the abrasive particle frequency is about f d = 67 Hz according to the frequency and speed conversion formula.

[0029] The length of the detection area of the measuring sensor is measured, and the oil speed is measured to approximately estimate the center frequency of the abrasive particles. The calculation formula is fd = vd / l, where vd ≈ vl; in the formula, fd is the center frequency of the abrasive particles, vd is the abrasive particle speed, vl is the oil flow speed, and l is the length of the detection area of the sensor.

[0030] The original signal y(n) to be processed is windowed and Fourier transformed to obtain the amplitude frequency spectrum of the original signal.

[0031] The amplitudes corresponding to all frequency points in the whole frequency band are extracted, and the discrete frequency spectrum is corrected and normalized. The values greater than 0.2 are extracted as harmonic signals, and the values less than 0.05 are regarded as noise.

[0032] The phase spectrum of the original signal is changed 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, so as to form multiple groups of harmonic signal parameters γ(a j ,f j ,θ j ), where j is the serial number of the extracted harmonic.

[0033] S3. According to the obtained signal parameters, a mathematical model based on the sensor output signal is built to generate multiple groups of simulated abrasive particle detection signals with different abrasive particle signal amplitudes, indexable positions and random noise.

[0034] According to the sensor output signal model, the generated simulated detection signal contains ideal abrasive particle signals, multiple frequency different sinusoidal signals linearly superimposed by the harmonic parameters measured based on the harmonic parameter estimation method, and random noise produced according to the preset signal to noise ratio. The abrasive particle signal and the harmonic are respectively expressed by the following formula, where the harmonic classification is multiplied by the correction coefficient α = 4.

[0035] The abrasive particle signal is:

[0036] where nd is the number of abrasive particle signals, T i is the position of the i-th abrasive 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 classification; the length of the signal is 50000 points;

[0038] The generated abrasive particle detection signal is shown in Figure 4 It can be seen that after parameter estimation, the harmonic interference in the simulation signal greatly restores the harmonics in the detection signal. After harmonic elimination, the simulation harmonic interference is destroyed, and the residual noise component is close to the real detection signal after harmonic elimination;

[0039] S4. Low-pass filtering and harmonic elimination are performed on the generated simulation signal. According to the center frequency of the abrasive particle signal, parameters are set to use a half-overlapping segmentation method to segment each group of simulation detection signals. According to the abrasive particle signal position index value, the simulation abrasive particle signal segment is extracted, and the corresponding number of noise segments are randomly extracted and labeled;

[0040] The signal is subjected to low-pass filtering operation, and the cutoff frequency f c = 2f d is used to protect the abrasive particle characteristics from being damaged; a fixed window length half-overlapping sliding segmentation method is used for sample segmentation, where the window length is L w , and the sliding step length is L w / 2;

[0041] The calculation formula of the window length is:

[0042] where ceil(·) represents rounding up, s d ∈[2.5, 3] is a set expansion coefficient to ensure the integrity of the abrasive particle characteristics;

[0043] After calculation, Lw=224, after signal segmentation, for a generated sequence with a length of 50000, the number of segmented samples is: where floor(·) is the floor function, and M signal samples are obtained;

[0044] Through division, an equal-length signal fragment collection X={X1, X2,...,X m} is obtained, where m=1, 2,..., 444. For a single vector X=(x1, x2,...,x p ), where p=1, 2,..., 224;

[0045] S5. For each extracted simulation abrasive grain signal segment and noise segment, use stationary wavelet transform for J-level decomposition, and extract the approximate components after the last four-level stationary wavelet transform decomposition to form a feature scale vector F = [a J-3 ,a J-2 ,a J-1 ,a J ] and use the Gram angle field (GADF) to transform each scale vector in the vector F to obtain a two-dimensional image. After splicing the images of the four components into one image, maximum pooling is performed to generate a network training data set;

[0046] For any signal segment, use stationary wavelet transform for J-level decomposition, and extract all the approximate components a J-3 ,a J-2 ,a J-1 and a J obtained by stationary wavelet transform to form a multi-scale vector, which can be represented as:

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

[0048] Gram angle field transformation is performed on each component in the vector F to obtain four two-dimensional images A 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] Perform coordinate system transformation:

[0053] Perform Gram angle field matrix operation:

[0054] Since the approximate signal after stationary wavelet transform has the same length as the original signal, the length of the converted two-dimensional image is also T. Four images are spliced to obtain an image G with a side length of 2T, which can be represented as:

[0055]

[0056] Further maximum pooling is performed on the spliced image G using a 2x2 convolution kernel to generate a TxT image. For input data X = {x1, x2,..., xp}, wherein p = 1, 2,..., L w , a corresponding multi-scale feature matrix set as the data for training the neural network is represented as G = {G1, G2,..., G p};

[0057] The abrasive particle segment and the same amount of random noise segment are subjected to the above operation to complete the sample set preparation. The abrasive particle segment is marked as 0, and the noise segment is marked as 1.

[0058] S6. Using the generated network training data set to train the convolutional neural network model to obtain an abrasive particle signal recognizer with high recognition accuracy;

[0059] Specifically, in the process of using the generated network training data set to train the convolutional neural network model to obtain an abrasive particle signal recognizer with high recognition accuracy, a deep convolutional neural network such as ResNet50 is used to extract higher-dimensional hidden features of the sample, strengthen the discrimination between abrasive particle signals and noise, and achieve the effect of improving the accuracy of the recognizer.

[0060] S7. The original signal to be processed is subjected to low-pass filtering, harmonic elimination, segment division, and signal matrixing according to the same hyperparameters, and the trained network is used for abrasive particle recognition to exclude noise segments and complete repetitive exclusion;

[0061] Specifically, in step S7, the low-pass filter cutoff frequency is set to f c = 134, and the segment division window length L w = 224. After completing the recognition and performing the de-duplication operation, 6 abrasive particle segments are identified.

[0062] S8. According to the recognition result, the corresponding one-dimensional abrasive particle signal is returned and de-duplication operation is performed, signal smoothing operation is performed using sliding filtering based on Gaussian window function, and abrasive particle positioning is performed using ideal abrasive particle signal based on abrasive particle mathematical model, further excluding noise interference and completing high-precision abrasive particle quantitative feature extraction;

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

[0064] First, the index-continuous segments are spliced, and the overlapping part is multiplied by a coefficient of 0.5 to ensure that the abrasive particle signal is not distorted.

[0065] Signal smoothing operation is achieved by sliding filtering, and an ideal abrasive particle signal with an amplitude of 1 and a sequence length of is generated according to the estimated frequency f d of the abrasive particle.

[0066] The signal is windowed on the one-dimensional signal sample with a step of 1, and the normalized cosine similarity calculation is performed on the equal-length segments to obtain the abrasive particle position weight. The formula for the normalized cosine similarity calculation of the equal-length vectors η1 and η1 is:

[0067] wherein, and are normalized equal-length vectors. The normalization process can be represented as:

[0068] For any sample vector X, the abrasive particle position weight can be represented as

[0069] Taking the serial number g corresponding to the maximum value max(P) of the weight as the position of the abrasive 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, and the abrasive particle positioning is completed;

[0070] wherein, the weight vector is represented as:

[0071]

[0072] Further denoising is achieved by multiplying the vector X and w p together, while protecting the abrasive particle features;

[0073] S9. Based on the abrasive particle segment and the corresponding index position, the detection signal is reconstructed to obtain a noise-free abrasive particle detection signal, and the final abrasive particle accurate counting is completed, and the reconstructed signal is as Figure 5 shown.

[0074] Specifically, in the abrasive particle segment and the corresponding index position detection signal reconstruction, the signal is constructed to have the same length as the detection signal by zero padding, and the sequence of the abrasive particle position is accurate, and the signal reconstruction is completed by linear superposition.

[0075] Although embodiments of the present application have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made therein without departing from the principles and spirit of the application, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for identifying oil abrasive particles based on convolutional neural networks and signal matrixization, characterized in that, Includes the following steps: S1: The lubricating oil wear monitoring system, constructed using an inductive wear particle detection sensor, performs real-time data acquisition on lubricating oil containing ferromagnetic wear particles to obtain the raw 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 abrasive signal based on the oil flow velocity; S3: Based on the obtained signal parameters, build a mathematical model based on the sensor output signal to generate multiple sets of simulated abrasive detection signals with different amplitudes, indexable positions, and random noise. S4: Perform low-pass filtering and harmonic elimination on the simulated abrasive detection signal. Based on the center frequency of the abrasive signal, set parameters and use the semi-overlapping segmentation method to segment each group of simulated detection signals. Extract the simulated abrasive signal segments based on the abrasive signal position index value, and simultaneously extract the corresponding number of noise segments and label them. S5: For each extracted simulated abrasive particle signal segment and noise segment, perform J-level decomposition using stationary wavelet transform. The approximate components after the four-level stationary wavelet transform decomposition constitute the feature scale vector F = [a J-3 ,a J-2 ,a J-1 ,a J The Gram angle field is used to transform each scale vector in vector F to obtain a two-dimensional image. The images of the four components are stitched together into one image and then max pooling is performed to generate the network training dataset. S6: Use the generated network training dataset to train the convolutional neural network model to obtain a wear particle signal recognizer with high recognition accuracy; S7: Perform low-pass filtering, harmonic elimination, segmentation and signal matrixing on the original signal to be processed according to the same hyperparameters, and use the trained network to identify wear particles, eliminate noise segments and complete repetitive elimination. S8: Based on the recognition results, return the corresponding one-dimensional abrasive signal and perform deduplication. Use a sliding filter based on the Gaussian window function to smooth the signal. Then use the ideal abrasive signal based on the abrasive mathematical model to locate the abrasive, further eliminate noise interference and complete the high-precision quantitative feature extraction of abrasive. S9: Based on the abrasive fragments and their corresponding index positions, the detection signal is reconstructed to obtain a noise-free abrasive detection signal, and finally, accurate abrasive counting is completed.

2. The method for identifying oil abrasive particles based on convolutional neural networks and signal matrixing according to claim 1, characterized in that, In step S2: The length of the sensor's detection area and the oil velocity are measured to approximate the abrasive grain center frequency. The calculation formula is: f d =v d / l, where v d ≈v l In the formula, f d v is the center frequency of the abrasive grains. d v is the abrasive particle velocity. l Here, l represents the oil flow velocity, and l represents the length of the sensor detection area. The amplitude spectrum of the original signal is obtained by performing a windowed Fourier transform on the original signal y(n). The amplitude values ​​corresponding to all frequency points within the entire frequency band are extracted, and discrete spectrum correction and normalization are performed. Values ​​greater than 0.2 are considered as harmonic signals, and values ​​less than 0.05 are considered as noise. The original signal is then subjected to phase spectrum transformation to obtain the phase spectrum. The corresponding phase is extracted based on the frequency corresponding to the amplitude of each harmonic signal in the amplitude spectrum, thus constructing multiple sets of harmonic signal parameters γ(a) containing amplitude, frequency, and phase. j ,f j ,θ j ), where j is the sequence number of the harmonics to be extracted, a j f is the harmonic component amplitude factor. j For the harmonic component frequency, θ j This is the phase factor for harmonic components.

3. The method for identifying oil abrasive particles based on convolutional neural networks and signal matrixing according to claim 1, characterized in that, In step S3: The simulated detection signal generated based on the mathematical model of the sensor output signal includes an ideal abrasive signal, harmonic components consisting of multiple sinusoidal signals of different frequencies linearly superimposed based on the harmonic parameters measured by the harmonic parameter estimation method, and random noise generated according to the preset signal-to-noise ratio. The abrasive signal and harmonics are expressed by the following formulas, where the harmonic classification is multiplied by the correction coefficient α = 4. Abrasive signal: Among them, f s Where n is the sampling frequency d It sets the number of abrasive grain signals, T i Let be the position of the i-th abrasive grain in the sequence, η be a random variable with a value range of [0.9, 1.1], and w(n) be a rectangular window function whose values ​​satisfy: harmonic: Where, n h This refers to the number of harmonic categories; the signal length is 50,000 points, a j f is the harmonic component amplitude factor. j For the harmonic component frequency, θ j This is the phase factor for harmonic components.

4. The method for identifying oil abrasive particles based on convolutional neural networks and signal matrixing according to claim 1, characterized in that, In step S4: The signal is low-pass filtered, where the cutoff frequency f is... c =2f d To protect the abrasive grains from damage, fd is the center frequency of the abrasive grains; Sample segmentation is performed using a semi-overlapping sliding segmentation method with a fixed window length, where the window length is... The sliding step size is L w / 2, where f s The sampling frequency is ceil(·), which represents rounding up. d ∈[2.5,3] represents the set expansion coefficient to ensure the integrity of the abrasive grain features. After signal segmentation, for a generated sequence of length N, the number of samples after segmentation is: Where M is the number of samples after segmentation, and floor(·) is the floor function; By dividing the signal into segments of equal length, a set of X = {x1, x2, ..., x...} is obtained. m }, where x m For the m-th signal segment, m = 1, 2, ..., M, for a single vector x = (y1, y2, ..., y... p ), where y p Let p be the p-th instance element in the vector, where p = 1, 2, ..., L w .

5. The method for identifying oil abrasive features based on convolutional neural networks and signal matrixing according to claim 1, characterized in that, In step S5: J-order decomposition is performed on any signal segment using stationary wavelet transform, yielding four approximate components a from the stationary wavelet transform. J-3 a J-2 a J-1 and a J Composition of multi-scale vector F = [a J-3 ,a J-2 ,a J-1 ,a J ]; Four two-dimensional images A are obtained by performing Gram angular field transformation on each component of the multi-scale vector F. J-3 A J-2 A J-1 and A J ; Since the approximate signal after the stationary wavelet transform has the same length as the original signal, the side length of the transformed two-dimensional image is also T. Stitching the four images together yields an image with a side length of 2T. The stitched image G is further max-pooled using a 2×2 convolution kernel to generate a T×T image; The set of equal-length signal segments X = {x1, x2, ..., x...} m } is used as data to train the neural network, resulting in the corresponding multi-scale feature matrix set θ={G1,G2,...,G M }, where x m For the m-th signal segment, G M The Mth instance element of the multi-scale feature matrix set; The above operations were performed on abrasive grain fragments and an equal number of random noise fragments to complete the sample set. The abrasive grain fragments were marked as 0 and the noise fragments were marked as 1.

6. The method for identifying oil abrasive particles based on convolutional neural networks and signal matrixing according to claim 5, characterized in that, The formula for calculating the Gram angle field is as follows: First, perform vector normalization: in, Let x be the normalized vector. m The original vector; Perform coordinate system transformation: Where, θ p r is the angle in angular coordinates. p Let P be the radius in angular coordinates, P be the length of the segment sequence, and t be the radius. p The timestamp corresponding to the current element; Perform Gram angle field matrix operations: Where m is the index and I is the unit vector. The normalized vector, for The transpose of .

7. The method for identifying oil abrasive particles based on convolutional neural networks and signal matrixing according to claim 1, 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 involves frequency domain parameter estimation and spectral correction of the entire signal segment, constructing an inverted signal and superimposing it with the original signal to suppress harmonics. The window length for segment division is L. w The step size is L w / 2, where f d The abrasive grain center frequency is denoted as .

8. The method for identifying oil abrasive features based on convolutional neural networks and signal matrixing according to claim 1, characterized in that, The specific steps for abrasive grain positioning in step S8 are as follows: First, the continuous segments of the index are spliced ​​together, and the overlapping parts are multiplied by a coefficient of 0.5 to ensure that the abrasive signal is not distorted; Signal smoothing is achieved through sliding filtering, and based on the abrasive grain center frequency f d Generate a sequence with an amplitude of 1 and a length of . The ideal abrasive grain signal, where f s f is the sampling frequency. d The abrasive grain center frequency; Using this signal as a window, slide it across a one-dimensional signal sample with a step size of 1. Calculate the wear grain position weight using normalized cosine similarity with equal-length segments. The formula for calculating the wear grain position weight using normalized cosine similarity for equal-length vectors η1 and η2 is as follows: Where p(η1,η2) is the weight of the wear grain position for equal-length vectors η1 and η2. Let η1 be a normalized vector. Let η2 be a normalized vector; For any vector Its abrasive grain position weight can be expressed as The index g corresponding to the maximum weight max(Q) is taken as the position of the abrasive grain in the vector. A window function is constructed to obtain the weight vector, and the non-zero interval is extended by 1 / 10 on both sides to complete the abrasive grain positioning. The weight vector w... m The Boolean value of the p-th instance can be calculated using the following formula: Among them, L W L is the window length for dividing the segment. S To identify the length of the window interval, q g For a single weight value, This is the last weight value in the current sample; By vector With weight vector w m The Hadamard product operation is performed to further reduce noise while protecting the abrasive grain characteristics.

9. The method for identifying oil abrasive particles based on convolutional neural networks and signal matrixing according to claim 1, characterized in that, In step S9, the detection signal reconstruction based on the abrasive fragments and their corresponding index positions is performed by padding the signal with zeros to construct a sequence with the same length as the detection signal and accurate abrasive fragment positions. The signal reconstruction is then completed by linear superposition.

Images