Aero-engine lubricating oil abrasive particle fault identification method based on convolutional neural network

Abstract

The invention belongs to the technical field of aviation machinery state monitoring and fault diagnosis, and particularly relates to an aero-engine lubricating oil abrasive particle fault identification method based on a convolutional neural network, which comprises the following steps: S1, collecting color image data of aero-engine lubricating oil abrasive particles, and converting the color image data into gray level image data through a weighted average method; s2, bilateral filtering is carried out on the grayscale image data of the aero-engine lubricating oil abrasive particles; s3, segmenting the lubricating oil abrasive particle image and the image background of the aero-engine by using an Otsu method; and S4, classifying the processed black and white single-channel aero-engine lubricating oil abrasive particle image data, and performing fault identification and output by using a convolutional neural network model. According to the method, the abrasive particle image acquired by the lubricating oil sensor of the aero-engine is processed, the convolutional neural network model is constructed to train and recognize the image, the fault type of the engine is quickly and accurately recognized, and the efficiency and accuracy of aero-engine fault diagnosis are improved.

Description

Technical Field

[0001] The invention belongs to the field of aviation machinery condition monitoring and fault diagnosis, and in particular relates to an aviation engine lubricating oil wear particle fault identification method based on convolutional neural network. Background Art

[0002] As the key power source of aircraft, the operating status of aircraft engines is related to flight safety. Since aircraft engines need to be in high-temperature and high-load working conditions for a long time, the incidence of wear and failure is relatively high. The abrasive particles in the lubricating oil of aircraft engines carry important information about the internal wear of the engine. By analyzing the number, size, shape and material of these abrasive particles, the wear status and potential failures inside the engine can be effectively judged. Therefore, accurate identification of abrasive particle information in lubricating oil can improve the monitoring accuracy of the engine's operating status, which is of great significance to ensuring the smooth completion of flight missions.

[0003] There are two detection methods for lubricating oil wear particles: offline detection and online detection. Although offline detection can obtain more comprehensive monitoring information, it cannot realize the feedback of the real-time status of the engine and has the disadvantage of information lag; online detection includes capacitance method, resistance method, magnetic induction method, acoustic method, image method, etc. Among them, the image method can obtain the size, quantity, material and morphology information of the wear particles. In particular, the morphological characteristics can only be obtained from the wear particle image, which can provide valuable information for identifying the wear form and finding the cause of wear. Therefore, it is urgent to propose a fault diagnosis method for the wear particle image of the lubricating oil sensor of the aircraft engine, so as to accurately identify the engine fault and thus improve the accuracy of the fault diagnosis of the aircraft engine. Summary of the invention

[0004] In view of the shortcomings of the prior art, the present invention provides an aircraft engine lubricating oil wear particle fault identification method based on convolutional neural network. By processing the wear particle images collected by the aircraft engine lubricating oil sensor, a convolutional neural network model is constructed to train and identify the images, so as to achieve rapid and accurate identification of engine fault types, aiming to improve the efficiency and accuracy of aircraft engine fault diagnosis.

[0005] To achieve the above object, the present invention discloses the following technical solution:

[0006] A method for identifying aero-engine lubricating oil wear particle faults based on a convolutional neural network, comprising:

[0007] S1: Collect color image data of aircraft engine lubricating oil wear particles and convert the color image data into grayscale image data by weighted average method;

[0008] S2: performing bilateral filtering on the grayscale image data of the aircraft engine lubricating oil wear particles in step S1;

[0009] S21: To reduce image noise while maintaining edge information of grayscale image data A(i,j) of aircraft engine lubricating oil particles, calculate the spatial proximity factor w of the bilateral filtering weight function. s (i,j) and gray similarity factor w r (i, j), the bilateral filtering weight coefficient w(i, j) is:

[0010]

[0011] Among them, w(i,j) is the bilateral filtering weight coefficient; w s (i, j) is the spatial proximity factor of the bilateral filter weight function; w r (i, j) is the grayscale similarity factor of the bilateral filtering weight function; c is the bilateral filtering parameter; i is the horizontal coordinate number of the grayscale image data pixel; j is the vertical coordinate number of the grayscale image data pixel;

[0012] S22: bilateral filtering is performed on the grayscale image data pixel A(i,j) of the aircraft engine lubricating oil wear particles, and the filtered image data pixel A'(x,y) is obtained as follows:

[0013]

[0014] Wherein, A'(x, y) is the image data pixel after grayscale image data filtering; A(i, j) is the grayscale image data pixel; M x,y is a (2Q+1)×(2Q+1) spatial neighborhood set centered on pixel (x, y); x is the horizontal coordinate of the pixel in the filtered image data; y is the vertical coordinate of the pixel in the filtered image data; Q is the spatial neighborhood parameter;

[0015] S23: repeating steps S21 and S22 until the grayscale image data of the entire aircraft engine lubricating oil wear particles are processed, and the grayscale image data A(i, j) of the aircraft engine lubricating oil wear particles is subjected to bilateral filtering to obtain filtered image data pixel A'(x, y);

[0016] S3: Use Otsu's method to segment the aircraft engine oil wear particle image and image background;

[0017] S31: Obtain the image data pixel A'(x, y) after the grayscale image data is filtered by bilateral filtering in step S2, extract the number of image pixels as N, the grayscale range as [0, K], and calculate the probability P of each grayscale level appearing r for:

[0018]

[0019] Among them, P r is the probability of each gray level appearing; nr is the number of pixels with gray level r; K is the upper limit of the gray level of the image pixel; l is the gray level number of the image pixel; r is the gray level number;

[0020] S32: Set the gray level threshold of the aircraft engine oil wear particle image data to T, then the gray level range of the aircraft engine oil wear particle image is [0, T], and the gray level range of the background is [T+1, K], and calculate the probability of wear particles appearing in the aircraft engine oil wear particle image ω 磨粒 and grayscale mean μ 磨粒 , background appearance probability ω 背景 and grayscale mean μ 背景 , the grayscale mean μ of the aircraft engine lubricating oil wear particle image data is obtained as:

[0021] μ=ω 磨粒 μ 磨粒 +ω 背景 μ 背景 ;

[0022] Wherein, μ is the grayscale mean of the aircraft engine lubricating oil wear particle image data; ω 磨粒 is the probability of abrasive particles appearing; μ 磨粒 is the mean gray value of the wear particles; ω 背景 is the background occurrence probability; μ 背景 is the background grayscale mean;

[0023] S33: Since the gray level threshold T of the aircraft engine lubricating oil wear particle image data is in the range of [0, K], it increases in sequence with a step size of 1. When the inter-class variance σ 2 The gray level threshold T corresponding to the maximum is the optimal threshold, and the inter-class variance σ 2 for:

[0024] σ 2 =ω 磨粒 (μ 磨粒 -μ) 2 +ω 背景 (μ 背景 -μ) 2 ;

[0025] Among them, σ 2 is the between-class variance at each gray level threshold;

[0026] S34: using the found optimal grayscale threshold of the aircraft engine lubricating oil wear particle image data to perform binarization processing on the image, pixels greater than the grayscale threshold T are set to white, the grayscale value is set to 255, and pixels less than or equal to the grayscale threshold T are set to black, the grayscale value is set to 0; obtaining black and white single-channel aircraft engine lubricating oil wear particle image data;

[0027] S4: Classify the black and white single-channel aircraft engine oil wear particle image data processed in step S3, and divide it into a training set, a validation set, and a test set according to the proportion; establish a convolutional neural network model and train it to obtain the real-time collected aircraft engine oil wear particle image data for fault identification and output.

[0028] Preferably, the color image data is converted into grayscale image data by weighted averaging method in step S1, specifically:

[0029] S11: extracting RGB three-component data from the color image data of the aircraft engine lubricating oil particles;

[0030] S12: The grayscale image data is obtained by weighted average calculation of the three components of RGB according to the following formula:

[0031] Gray(i,j)=0.299×R(i,j)+0.578×G(i,j)+0.114×B(i,j);

[0032] Among them, Gray(i,j) is the grayscale image data; R(i,j) is the first component of the color image data; G(i,j) is the second component of the color image data; B(i,j) is the third component of the color image data;

[0033] S13: traverse each pixel point to calculate the gray value and generate a gray image of the aircraft engine lubricating oil wear particles.

[0034] Preferably, the spatial proximity factor w in step S21 is s (i,j) is:

[0035]

[0036] Among them, w s (i,j) is the spatial proximity factor; σ s is the spatial proximity parameter.

[0037] Preferably, the grayscale similarity factor w in step S21 is r (i,j) is:

[0038]

[0039] Among them, w r (i,j) is the grayscale similarity factor; σ r is the grayscale similarity parameter.

[0040] Preferably, the probability of occurrence of abrasive particles in step S32 is 磨粒 and grayscale mean μ 磨粒 , specifically:

[0041] The probability of wear particles appearing ω 磨粒 for:

[0042]

[0043] Wherein, T is the gray level threshold of the aircraft engine lubricating oil wear particle image data;

[0044] The mean gray value of the abrasive particles μ 磨粒 for:

[0045]

[0046] Preferably, the background occurrence probability ω in step S32 is 背景 and grayscale mean μ 背景 , specifically:

[0047] The background occurrence probability ω 背景 for:

[0048]

[0049] The background grayscale mean μ 背景 for:

[0050]

[0051] Preferably, in step S4, the black and white single-channel aircraft engine oil wear particle image data processed in step S3 is classified as follows:

[0052] According to the historical database of aircraft engine lubricating oil wear particle images, the correspondence between the wear particle morphology and the fault type is analyzed and determined, and then the corresponding labels are set for the output layer of the convolutional neural network according to the correspondence, including: the first category label is normal state, the second category label is cutting wear, the third category label is fatigue spalling and the fourth category label is severe sliding wear.

[0053] Preferably, the convolutional neural network model in step S4 specifically includes a 10-layer network, specifically:

[0054] S41: Use the input layer to receive black and white single-channel aircraft engine oil wear particle image data input, the image size is 256×256;

[0055] S42: Use the first convolutional layer for processing, 32 convolution kernels of size 3×3 for convolution operation, stride 1, padding 1, and activation function ReLU for processing;

[0056] S43: Use the first pooling layer for processing, with a pooling window of 2×2, a step size of 2, no padding, and halving the image size to reduce the amount of calculation;

[0057] S44: Use the second convolutional layer for processing, 64 convolution kernels of size 3×3 for convolution, stride 1, padding 1, and activation function ReLU for processing;

[0058] S45: Use the second pooling layer for processing, with a 2×2 pooling window, a step size of 2, and no padding;

[0059] S46: Use the third convolutional layer for processing, 128 convolution kernels of size 3×3 for convolution, stride 1, padding 1, and activation function ReLU for processing;

[0060] S47: Use the third pooling layer for processing, with a pooling window of 2×2, a step size of 2, and no padding to obtain a feature map;

[0061] S48: Use a fully connected layer to expand the feature map into a one-dimensional vector and input it into the fully connected layer. The number of nodes in the fully connected layer is determined according to the number of categories of the training data.

[0062] S49: Use dropout layer: set to 0.5; Setting the dropout rate to 0.5 means that during each training iteration, each neuron in the network has a 50% probability of being temporarily dropped;

[0063] S410: Use the softmax function in the output layer to perform classification output, and the output dimension is set to the number of categories E; in the output layer, for the input real number vector z=[z1,z2,...,z E ], the Softmax function converts the vector into a probability distribution The probability distribution of each real vector for:

[0064]

[0065] in, is the probability distribution of the lth real vector, l = 1, 2, ..., E; z l is the lth real number vector; E is the number of categories; l is the number of categories.

[0066] Compared with the prior art, the present invention has the following beneficial effects:

[0067] (1) The present invention can effectively improve the safety and reliability of aircraft engine operation, reduce the need for manual operation and improve the speed of fault detection through automated image processing and analysis processes; by utilizing the powerful feature extraction capability of convolutional neural networks, combined with bilateral filtering and Otsu segmentation, the present invention can effectively distinguish different types of wear particles and achieve high-precision fault identification; the method of the present invention is not only applicable to the current data set, but can also update the model through continuous learning to adapt to new fault types and image conditions.

[0068] (2) The present invention provides an automated method for aircraft engine lubricating oil wear particle image fault identification, which reduces the need for manual detection; through precise image segmentation and feature extraction, the accuracy of fault detection is improved, and the possibility of false alarms and missed alarms is reduced; the maintenance and fault prevention capabilities of aircraft engines are enhanced, and the safety and reliability of flight are improved; by using deep learning technology, a new technical means is provided for the field of aircraft engine fault diagnosis; for the model, it can be directly applied to similar lubricating oil sensors and has strong migration capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 A control block diagram of an aircraft engine lubricating oil wear particle fault identification method based on a convolutional neural network according to the present invention;

[0070] Figure 2 It is a hierarchical structure diagram of the convolutional neural network model of the present invention;

[0071] Figure 3 It is a fault diagnosis flow chart of the present invention;

[0072] Figure 4 This is the convolutional neural network model running process of the present invention. DETAILED DESCRIPTION

[0073] The exemplary embodiments, features and aspects of the present invention will be described in detail below with reference to the accompanying drawings. The same reference numerals in the accompanying drawings represent elements with the same or similar functions. Although various aspects of the embodiments are shown in the accompanying drawings, the drawings are not necessarily drawn to scale unless otherwise specified.

[0074] The embodiment of the present invention provides a method for identifying aero-engine lubricating oil wear particle faults based on a convolutional neural network, such as Figure 1 As shown, the color image data of aircraft engine lubricating oil wear particles are collected, and the color image data are converted into grayscale image data by weighted average method; bilateral filtering is performed on the grayscale image data of aircraft engine lubricating oil wear particles; the aircraft engine lubricating oil wear particle image and image background are segmented by Otsu method; the processed black and white single-channel aircraft engine lubricating oil wear particle image data are classified, and the convolutional neural network model is used for fault identification and output; it includes:

[0075] Step S1: Collect color image data of aircraft engine lubricating oil wear particles, and convert the color image data into grayscale image data by weighted average method, specifically:

[0076] Step S11: Extract RGB three-component data from the color image data of the aircraft engine lubricating oil wear particles. Step S12: Perform weighted average calculation on the RGB three-components according to the following formula to obtain grayscale image data:

[0077] Gray(i,j)=0.299×R(i,j)+0.578×G(i,j)+0.114×B(i,j);

[0078] Among them, Gray(i,j) is the grayscale image data; R(i,j) is the first component of the color image data; G(i,j) is the second component of the color image data; B(i,j) is the third component of the color image data.

[0079] Step S13: traverse each pixel point to calculate the gray value and generate a gray image of the aircraft engine lubricating oil wear particles.

[0080] Step S2: bilateral filtering is performed on the grayscale image data of the aircraft engine lubricating oil wear particles in step S1; compared with the classical Gaussian filtering, the bilateral filtering algorithm combines the pixel grayscale similarity, fully considers the image spatial information and grayscale information, and greatly retains the detail information while suppressing the coherent spots in the homogeneous area.

[0081] Step S21: To reduce image noise while maintaining edge information of the grayscale image data A(i, j) of the aircraft engine lubricating oil wear particles, the spatial proximity factor w of the bilateral filtering weight function is calculated respectively. s (i,j) and gray similarity factor w r (i,j), specifically:

[0082] Spatial proximity factor w s (i,j) is:

[0083]

[0084] Among them, w s (i,j) is the spatial proximity factor; σ s is the spatial proximity parameter.

[0085] Gray similarity factor w r (i,j) is:

[0086]

[0087] Among them, w r (i,j) is the grayscale similarity factor; σ r is the grayscale similarity parameter.

[0088] The bilateral filtering weight coefficient w(i,j) is obtained as:

[0089]

[0090] Among them, w(i,j) is the bilateral filtering weight coefficient; w s(i, j) is the spatial proximity factor of the bilateral filter weight function; w r (i, j) is the grayscale similarity factor of the bilateral filtering weight function; c is the bilateral filtering parameter; i is the horizontal coordinate number of the grayscale image data pixel; j is the vertical coordinate number of the grayscale image data pixel.

[0091] Step S22: bilateral filtering is performed on the grayscale image data pixel A(i, j) of the aircraft engine lubricating oil wear particles to obtain the filtered image data pixel A'(x, y):

[0092]

[0093] Wherein, A'(x, y) is the image data pixel after grayscale image data filtering; A(i, j) is the grayscale image data pixel; M x,y is a (2Q+1)×(2Q+1) spatial neighborhood set centered on the pixel (x, y); x is the horizontal coordinate of the pixel of the filtered image data; y is the vertical coordinate of the pixel of the filtered image data; Q is the spatial neighborhood parameter.

[0094] Step S23: repeat steps S21 and S22 until the grayscale image data of the entire aircraft engine lubricating oil wear particles are processed, and the grayscale image data A(i,j) of the aircraft engine lubricating oil wear particles is subjected to bilateral filtering to obtain filtered image data pixels A'(x,y).

[0095] Step S3: Use Otsu's method to segment the aircraft engine oil wear particle image and image background; Use Otsu's method to segment the wear particle and image background. The core idea of ​​this algorithm is to maximize the inter-class variance, that is, to find a threshold in the grayscale histogram of the image to maximize the difference between the segmented foreground and background. The purpose of this is to minimize the classification error, that is, to minimize the possibility of foreground pixels being mistakenly classified as background pixels, or background pixels being mistakenly classified as foreground pixels.

[0096] Step S31: Obtain the image data pixel A'(x, y) after the grayscale image data is obtained by bilateral filtering in step S2, extract the number of image pixels as N, the grayscale range as [0, K], and calculate the probability P of each grayscale level appearing r for:

[0097]

[0098] Among them, P r is the probability of each gray level appearing; n r is the number of pixels with gray level r; K is the upper limit of the gray level of the image pixel; l is the gray level number of the image pixel; r is the gray level number.

[0099] Step S32: Set the grayscale threshold of the aircraft engine oil wear particle image data to T, then the grayscale range of the aircraft engine oil wear particle image is [0, T], and the grayscale range of the background is [T+1, K], and calculate the wear particle appearance probability ω 磨粒 for:

[0100]

[0101] Wherein, T is the gray level threshold of the aircraft engine lubricating oil wear particle image data;

[0102] Calculate the mean gray value μ of the wear particles 磨粒 for:

[0103]

[0104] Calculate the background occurrence probability ω 背景 for:

[0105]

[0106] Calculate the background gray mean μ 背景 for:

[0107]

[0108] The grayscale mean μ of the aircraft engine oil wear particle image data is obtained as:

[0109] μ=ω 磨粒 μ 磨粒 +ω 背景 μ 背景 ;

[0110] Wherein, μ is the grayscale mean of the aircraft engine oil wear particle image data; ω 磨粒 is the probability of abrasive particles appearing; μ 磨粒 is the mean gray value of the wear particles; ω 背景 is the background occurrence probability; μ 背景 is the background grayscale mean.

[0111] Step S33: Since the gray level threshold T of the aircraft engine lubricating oil wear particle image data is in the range of [0, K], it is increased in sequence with a step size of 1. When the inter-class variance σ 2 The gray level threshold T corresponding to the maximum is the optimal threshold, and the inter-class variance σ 2 for:

[0112] σ 2 =ω 磨粒 (μ 磨粒 -μ) 2 +ω 背景 (μ 背景 -μ) 2 ;

[0113] Among them, σ 2 is the between-class variance at each gray level threshold.

[0114] Step S34: Use the found optimal grayscale threshold of the aircraft engine lubricating oil wear particle image data to binarize the image, set pixels greater than the grayscale threshold T to white, and the grayscale value is set to 255, and set pixels less than or equal to the grayscale threshold T to black, and the grayscale value is set to 0; obtain black and white single-channel aircraft engine lubricating oil wear particle image data.

[0115] Step S4: classify the black and white single-channel aircraft engine oil wear particle image data processed in step S3, specifically:

[0116] According to the historical database of aircraft engine lubricating oil wear particle images, the correspondence between the wear particle morphology and the fault type is analyzed and determined, and then the corresponding labels are set for the output layer of the convolutional neural network according to the correspondence, including: the first category label is normal state, the second category label is cutting wear, the third category label is fatigue spalling, and the fourth category label is severe sliding wear. 500 images are selected as training data for each state, and divided into training set, validation set and test set in a ratio of 6:2:2. The specific division is shown in Table 1.

[0117] Table 1 Training set, validation set and test set table of training data

[0118] state Training set (sheets) Validation set (sheets) Test set (sheets) Normal state 300 100 100 Cutting wear state 300 100 100 Fatigue peeling state 300 100 100 Severe active wear status 300 100 100

[0119] Build a convolutional neural network model and train it, such as Figure 2 The figure shows the hierarchical structure diagram of the convolutional neural network model of the present invention. The convolutional neural network model specifically includes 10 layers of networks, specifically:

[0120] Step S41: Use the input layer to receive black and white single-channel aircraft engine oil wear particle image data input, and the image size is 256×256.

[0121] Step S42: Use the first convolutional layer for processing, perform convolution operations with 32 convolution kernels of size 3×3, a step size of 1, a padding of 1, and use the activation function ReLU for processing.

[0122] Step S43: Use the first pooling layer for processing, with a 2×2 pooling window, a step size of 2, and no padding, to halve the image size and reduce the amount of calculation.

[0123] Step S44: Use the second convolutional layer for processing, use 64 convolution kernels of size 3×3 for convolution, with a step size of 1, a padding of 1, and use the activation function ReLU for processing.

[0124] Step S45: Use the second pooling layer for processing, with a 2×2 pooling window, a step size of 2, and no padding.

[0125] Step S46: Use the third convolutional layer for processing, use 128 convolution kernels of size 3×3 for convolution, the step size is 1, the padding is 1, and the activation function ReLU is used for processing.

[0126] Step S47: Use the third pooling layer for processing, with a 2×2 pooling window, a step size of 2, and no padding to obtain a feature map.

[0127] Step S48: Use a fully connected layer to expand the feature map into a one-dimensional vector and input it into the fully connected layer. The number of nodes in the fully connected layer is determined according to the number of categories of the training data.

[0128] Step S49: Use a dropout layer: set to 0.5; Setting the dropout rate to 0.5 means that during each training iteration, each neuron in the network has a 50% probability of being temporarily dropped.

[0129] Step S410: Use the softmax function in the output layer to perform classification output, and the output dimension is set to the number of categories E; in the output layer, for the input real number vector z = [z1, z2, ..., z E ], the Softmax function converts the vector into a probability distribution The probability distribution of each real vector for:

[0130]

[0131] in, is the probability distribution of the lth real vector, l = 1, 2, ..., E; z l is the lth real number vector; E is the number of categories; l is the number of categories.

[0132] Table 2 shows the specific training parameters for the neural network model, which are used to complete the pre-setting of the neural network model and ensure the accuracy of the output of the neural network model.

[0133] Table 2 Neural network model training parameters

[0134]

[0135]

[0136] After completing all the training, the model weight file with the highest accuracy is selected as the parameter basis of the state recognition model to identify the unknown lubricating oil wear particle image data.

[0137] like Figure 3The fault diagnosis flow chart of the present invention is shown as follows; the real-time collected aircraft engine oil wear particle image data is obtained for fault identification and output. After the unknown wear particle image is processed according to steps S2-S3, it is input into the trained model file to realize fault diagnosis. Figure 4 This is a convolutional neural network model operation flow chart of the present invention, which is a convolutional neural network model flow chart in the fault diagnosis process. The training of the neural network model is completed by inputting a data set, and the aircraft engine lubricating oil wear particle fault is completed by the trained neural network model.

[0138] The beneficial effects of the present invention are as follows: the embodiment of the present invention proposes an aircraft engine lubricating oil wear particle fault identification method based on a convolutional neural network, and uses bilateral filtering technology to filter the image to reduce noise interference, retain edge information, and ensure the clarity of the wear particle characteristics; the image is segmented by the Otsu method, the wear particles are separated from the background, and a clear single-channel black-and-white image is generated to more accurately identify the morphological characteristics of the wear particles; a convolutional neural network model is constructed and trained using a divided data set, which can effectively diagnose and identify unknown aircraft engine lubricating oil sensor faults, greatly reducing the need for manual operation, thereby improving the accuracy and efficiency of fault detection, and reducing the possibility of false alarms and missed alarms; the maintenance and fault prevention capabilities of aircraft engines are enhanced, and the safety and reliability of flight are improved.

[0139] The embodiments described above are only descriptions of the preferred implementation modes of the present invention, and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A method for identifying aircraft engine lubricating oil wear particle faults based on convolutional neural networks, characterized in that: It includes: S1: Collect color image data of aircraft engine lubricating oil wear particles and convert the color image data into grayscale image data by weighted average method; S2: bilateral filtering is performed on the grayscale image data of the aircraft engine lubricating oil wear particles in step S1, specifically including the following sub-steps: S21: To reduce image noise while maintaining edge information of grayscale image data A(i,j) of aircraft engine lubricating oil particles, calculate the spatial proximity factor w of the bilateral filtering weight function. s (i,j) and gray similarity factor w r (i, j), the bilateral filtering weight coefficient w(i, j) is: Among them, w(i,j) is the bilateral filtering weight coefficient; w s (i, j) is the spatial proximity factor of the bilateral filter weight function; w r (i, j) is the grayscale similarity factor of the bilateral filtering weight function; c is the bilateral filtering parameter; i is the horizontal coordinate number of the grayscale image data pixel; j is the vertical coordinate number of the grayscale image data pixel; S22: bilateral filtering is performed on the grayscale image data pixel A(i,j) of the aircraft engine lubricating oil wear particles, and the filtered image data pixel A'(x,y) is obtained as follows: Wherein, A'(x, y) is the image data pixel after grayscale image data filtering; A(i, j) is the grayscale image data pixel; M x,y is a (2Q+1)×(2Q+1) spatial neighborhood set centered on pixel (x, y); x is the horizontal coordinate of the pixel in the filtered image data; y is the vertical coordinate of the pixel in the filtered image data; Q is the spatial neighborhood parameter; S23: repeating steps S21 and S22 until the grayscale image data of the entire aircraft engine lubricating oil wear particles are processed, and the grayscale image data A(i,j) of the aircraft engine lubricating oil wear particles is subjected to bilateral filtering to obtain filtered image data pixel A'(x,y); S3: Use the Otsu method to segment the aircraft engine oil wear particle image and image background, which specifically includes the following sub-steps: S31: Obtain the image data pixel A'(x, y) after the grayscale image data is filtered by bilateral filtering in step S2, extract the number of image pixels as N, the grayscale range as [0, K], and calculate the probability P of each grayscale level appearing r for: Among them, P r is the probability of each gray level appearing; n r is the number of pixels with gray level r; K is the upper limit of the gray level of the image pixel; l is the gray level number of the image pixel; r is the gray level number; S32: Set the gray level threshold of the aircraft engine oil wear particle image data to T, then the gray level range of the aircraft engine oil wear particle image is [0, T], and the gray level range of the background is [T+1, K], and calculate the probability of wear particles appearing in the aircraft engine oil wear particle image ω 磨粒 and grayscale mean μ 磨粒 , background appearance probability ω 背景 and grayscale mean μ 背景 , the grayscale mean μ of the aircraft engine lubricating oil wear particle image data is obtained as: m = oh 磨粒 m 磨粒 +oh 背景 m 背景 Wherein, μ is the grayscale mean of the aircraft engine lubricating oil wear particle image data; ω 磨粒 is the probability of abrasive particles appearing; μ 磨粒 is the mean gray value of the wear particles; ω 背景 is the background occurrence probability; μ 背景 is the background grayscale mean; S33: The gray level threshold T of the aircraft engine lubricating oil wear particle image data ranges from [0, K]. Within this range, it increases in steps of 1. When the inter-class variance σ 2 The gray level threshold T corresponding to the maximum is the optimal threshold, and the inter-class variance σ 2 for: s 2 =ω 磨粒 (m 磨粒 -m) 2 +oh 背景 (m 背景 -m) 2 Among them, σ 2 is the between-class variance at each gray level threshold; S34: using the obtained optimal grayscale threshold of the aircraft engine lubricating oil wear particle image data to perform binarization processing on the image, setting pixels greater than the grayscale threshold T to white, with the grayscale value set to 255, and setting pixels less than or equal to the grayscale threshold T to black, with the grayscale value set to 0; obtaining black and white single-channel aircraft engine lubricating oil wear particle image data; S4: Classify the black and white single-channel aircraft engine oil wear particle image data processed in step S3, and divide it into a training set, a validation set, and a test set according to the proportion; establish a convolutional neural network model and train it to obtain the real-time collected aircraft engine oil wear particle image data for fault identification and output.

2. The method for identifying aircraft engine oil wear particle faults based on a convolutional neural network according to claim 1, characterized in that: The color image data is converted into grayscale image data by weighted averaging method in step S1, specifically: S11: extracting RGB three-component data from the color image data of the aircraft engine lubricating oil particles; S12: The grayscale image data is obtained by weighted average calculation of the three components of RGB according to the following formula: Gray(i,j)=0.299×R(i,j)+0.578×G(i,j)+0.114×B(i,j) Among them, Gray(i,j) is the grayscale image data; R(i,j) is the first component of the color image data; G(i,j) is the second component of the color image data; B(i,j) is the third component of the color image data; S13: traverse each pixel point to calculate the gray value and generate a gray image of the aircraft engine lubricating oil wear particles.

3. The method for identifying aircraft engine oil wear particle faults based on a convolutional neural network according to claim 2, characterized in that: The spatial proximity factor w in step S21 s (i,j) is: Among them, w s (i,j) is the spatial proximity factor; σ s is the spatial proximity parameter.

4. The method for identifying aircraft engine oil wear particle faults based on a convolutional neural network according to claim 3, characterized in that: The grayscale similarity factor w in step S21 r (i,j) is: Among them, w r (i,j) is the grayscale similarity factor; σ r is the grayscale similarity parameter.

5. The method for identifying aircraft engine oil wear particle faults based on convolutional neural network according to claim 1, characterized in that: The probability of wear particles appearing in step S32 磨粒 and grayscale mean μ 磨粒 The specific calculation formula is: The probability of wear particles appearing ω 磨粒 for: Wherein, T is the gray level threshold of the aircraft engine lubricating oil wear particle image data; The mean gray value of the abrasive particles μ 磨粒 for:

6. The method for identifying aircraft engine oil wear particle faults based on convolutional neural network according to claim 1, characterized in that: The background occurrence probability ω in step S32 背景 and grayscale mean μ 背景 The specific calculation formula is: The background occurrence probability ω 背景 for: The background grayscale mean μ 背景 for:

7. The method for identifying aircraft engine oil wear particle faults based on convolutional neural network according to claim 3, characterized in that: In step S4, the black and white single-channel aircraft engine oil wear particle image data processed in step S3 is classified, specifically: According to the historical database of aircraft engine lubricating oil wear particle images, the correspondence between the wear particle morphology and the fault type is analyzed and determined, and then the corresponding labels are set for the output layer of the convolutional neural network according to the correspondence, including: the first category label is normal state, the second category label is cutting wear, the third category label is fatigue spalling and the fourth category label is severe sliding wear.

8. The method for identifying aircraft engine oil wear particle faults based on convolutional neural network according to claim 1, characterized in that: The convolutional neural network model in step S4 includes a 10-layer network, specifically: S41: Use the input layer to receive black and white single-channel aircraft engine oil wear particle image data input, the image size is 256×256; S42: Use the first convolutional layer for processing, 32 convolution kernels of size 3×3 for convolution operation, stride 1, padding 1, and activation function ReLU for processing; S43: Use the first pooling layer for processing, a 2×2 pooling window, a step size of 2, and halve the image size; S44: Use the second convolutional layer for processing, 64 convolution kernels of size 3×3 for convolution, stride 1, padding 1, and activation function ReLU for processing; S45: Use the second pooling layer for processing, with a pooling window of 2×2 and a step size of 2; S46: Use the third convolutional layer for processing, 128 convolution kernels of size 3×3 for convolution, stride 1, padding 1, and activation function ReLU for processing; S47: Use the third pooling layer for processing, with a pooling window of 2×2 and a step size of 2 to obtain a feature map; S48: Use a fully connected layer to expand the feature map into a one-dimensional vector, and input it into the fully connected layer. The number of nodes in the fully connected layer is determined according to the number of categories of the training data; S49: Use dropout layer: The dropout rate is set to 0.5; Setting the dropout rate to 0.5 means that during each training iteration, each neuron in the network has a 50% probability of being temporarily dropped; S410: Use the softmax function in the output layer to perform classification output, and the output dimension is set to the number of categories E; in the output layer, for the input real number vector z=[z1,z2,...,z E ], the Softmax function converts the vector into a probability distribution The probability distribution of each real vector for: in, is the probability distribution of the lth real vector, l = 1, 2, ..., E; z l is the lth real vector; E is the number of categories; l is the number of category numbers.

Images