Aero-engine unit bearing fault diagnosis method based on deep learning

Abstract

The invention discloses an aero-engine unit bearing fault diagnosis method based on deep learning. The method comprises the following steps: (1) extracting vibration signal data of an aero-engine unit bearing and preprocessing original data; (2) preprocessing the original data by adopting EMD and SG filters, and constructing an SWAFormer-FMNet diagnosis model by fusing SWAFormer with FMNet so as to diagnose a bearing vibration signal; (3) applying Tent chaotic mapping, a Gaussian random walk strategy and a differential variation strategy to obtain an improved dung beetle optimization algorithm; (4) using the improved dung beetle algorithm to optimize three parameters of a learning rate, a hidden layer number and a batch processing size in the diagnosis model; and (5) establishing an SWAFormer-FMNet diagnosis model based on the improved dung beetle algorithm. Compared with the prior art, the method can be applied to the aero-engine unit bearing fault diagnosis modeling process, has the advantage of high optimization precision, and finally ensures the accuracy of aero-engine unit bearing fault diagnosis.

Description

Technical Field

[0001] The invention relates to a bearing fault diagnosis technology, in particular to an aero-engine group bearing fault diagnosis method. Background Art

[0002] In the modern aviation industry, the aircraft engine unit is an important power supply system for aircraft, and its operating reliability is directly related to the safety and economy of flight. The engine unit supports various key equipment on the aircraft by providing a stable power supply. Therefore, ensuring the stable operation of the engine unit is one of the core tasks of aviation safety. In the engine unit, the bearing is a key rotating component, and its health status directly affects the performance and life of the engine unit. However, under long-term high-speed operation and high load, the bearing is easily affected by various factors such as wear, fatigue, and corrosion, which may cause failures. If these faults are not detected and handled in time, they may cause the failure of the engine unit and even cause flight accidents.

[0003] At present, bearing fault diagnosis mainly relies on traditional methods such as vibration signal analysis, acoustic signal analysis, and temperature monitoring. These methods are usually based on the extraction and analysis of signal features, and judge the health status of bearings by changes in characteristic parameters such as acceleration, velocity, and displacement. However, the effectiveness of traditional methods depends largely on empirical models and manual judgment, and is easily affected by external noise and complex working conditions, with low detection accuracy and reliability. In addition, traditional methods often show shortcomings when facing complex fault modes and nonlinear characteristics, and it is difficult to meet the high-precision and high-real-time requirements of modern aviation equipment for fault diagnosis systems. Therefore, a diagnosis system based on deep learning came into being. Through the SWAFormer-FMNet fusion diagnosis model, SWAFormer can effectively extract time series and global features from the input data, and FMNet introduces a focus modulation mechanism to enhance the local feature capture capability of the network architecture; the data after model training is connected to the softmax neural network layer, which can effectively extract the fault information in the vibration signal for the bearing fault diagnosis of the aero-engine group and improve the accuracy of the bearing fault diagnosis of the aero-engine group. Summary of the invention

[0004] Purpose of the invention: In view of the problems in the background technology, the present invention provides an efficient and accurate method for diagnosing faults of aircraft engine group bearings, which realizes accurate diagnosis of aircraft engine group bearings by combining empirical mode decomposition (EMD), Savitzky-Golay filter (SG), Stochastic Weight Averaging Transformer (SWAFormer), focal modulation network mechanism (FMNet) and improved dung beetle optimization algorithm.

[0005] Technical solution: The present invention discloses a method and system for diagnosing bearing faults of an aero-engine assembly based on deep learning, comprising the following steps:

[0006] Step 1: Collect vibration signals of four different situations of inner ring fault, outer ring fault, rolling element fault and normal situation of the rolling bearing of the aircraft engine group at the same speed.

[0007] Step 2: Use EMD (empirical mode decomposition) and SG filter (Savitzky-Golay filter) to decompose and reduce noise of four different vibration signals;

[0008] Step 3: Apply the generalization ability of SWAFormer and the focal attention mechanism of FMNet to diagnose the original data;

[0009] Step 4: Tent chaos mapping is used to increase the diversity of the initialization of the intelligent particle algorithm; the Gaussian random walk strategy is used to enhance the local search capability of the algorithm; the differential mutation strategy is used to enhance the global search capability of the algorithm, and an improved dung beetle optimization algorithm is obtained; the improved dung beetle optimization algorithm is used to optimize the hyperparameters of the SWAFormer-FMNet diagnostic model, and the hyperparameters include learning rate, number of hidden layers, and batch processing size;

[0010] Step 5: Use the optimized SWAFormer-FMNet fusion model to diagnose the bearing fault of the aircraft engine group, and classify the output of the fusion model using the softmax function;

[0011] Step 6: Use the test set data to test the trained SWAFormer-FMNet fault diagnosis model.

[0012] Further, in step 2, the steps of using EMD and SG filters to decompose and reduce noise of 4 different vibration signals are as follows:

[0013] Step 2.1: EMD (Empirical Mode Decomposition).

[0014] The original bearing vibration signal x(t) is solved into several intrinsic mode functions (IMFs). The formula is expressed as:

[0015]

[0016] Among them, IMF i (t) is the i-th intrinsic mode function, r n (t) is the residual, which represents the final remaining non-oscillatory part.

[0017] Step 2.2: Savitzky-Golay filter (SG filter)

[0018] An SG filter is used on each IMF to smooth the signal. The core formula of the SG filter is:

[0019]

[0020] Among them, y i is the smoothed signal value, x i+j is the value of the original signal within the window. j are the filter coefficients, obtained by local polynomial fitting. m is half the window size.

[0021] Step 2.3: Extract the time-frequency domain features of the above data after preprocessing through Hilbert transform.

[0022] The formula for the Hilbert transform is:

[0023]

[0024] Where PV is the Cauchy principal value integral, and H(x(t)) is the instantaneous amplitude and phase information of the signal.

[0025] The extracted feature matrix:

[0026] X={x1,x2,...,x n}

[0027] where each x i Feature vectors representing different time points or signal segments.

[0028] Step 2.4: Divide the sample set into training set and test set in a ratio of 8:2.

[0029] Further, in step 3, the steps of applying the generalization ability of SWAFormer and the FMNet focus attention mechanism to diagnose the original data are as follows:

[0030] Step 3.1: Extract temporal dependencies and global features from preprocessed data using Tranformer:

[0031]

[0032] MultiHead(Q,K,V)=Concat(head1,...,head h )W O

[0033] Z l =LayerNorm(X+MultiHead(Q,K,V))

[0034] Among them, Q, K, and V are matrix representations of query, key, and value respectively, and d kis the dimension of the query and key. is the weight matrix. L The final feature representation is input into FMNet for local modulation.

[0035] Step 3.2: Use SWA (random weight averaging) to smoothly optimize the weights of the Tranformer model.

[0036] SWA generates a smoother solution by averaging weights at different training steps. Suppose during training, the weight at iteration t is θ t , SWA uses the following formula to update the weights:

[0037]

[0038] Among them, θ t are the model parameters at the tth iteration. T0 is the time point at which the average value is calculated. T is the total number of training steps.

[0039] Step 3.3: Use FMNet to extract the global feature Z from Transformer L The modulation mechanism combines local and global information. The formula of FMNet is

[0040] F modulated (Z L )=F local (Z L )+α·F context (Z L )

[0041] where F local (Z L ) is the local feature module processing Transformer output Z L , F context (Z L ) is the module processing Z for global context information L , α is the modulation coefficient, which controls the degree of fusion of local and global features.

[0042] The data features after FMNet modulation are used as the input of the softmax neural network layer to output the prediction results of bearing faults.

[0043] Further, the improved dung beetle optimization algorithm obtained by improving the dung beetle optimization algorithm in step 4 includes the following steps:

[0044] Step 4.1: Set the population size of the dung beetle optimization algorithm to 35, the number of iterations to 50, the upper and lower limits of the search space include: learning rate [0.001-0.1], the number of neural network layers to [10-50], and the batch size to [16-256];

[0045] Step 4.2: Use Tent chaotic map to initialize the population position of the algorithm.

[0046] Step 4.2.1: The Tent mapping formula is as follows:

[0047]

[0048] Among them, x k is the individual value of the kth generation, r is the parameter of the Tent mapping, and r=0.5 is usually taken.

[0049] Step 4.2.2: Generate chaotic sequences and map the search space.

[0050] Randomly select an initial value x0 in the range [0,1], and iteratively generate a chaotic sequence of length N×D from the initial seed x0, where N is the number of individuals in the population and D is the dimension of the individual. The generated chaotic sequence is:

[0051] {x1,x2,...,x N}

[0052] For each individual, the Tent mapping formula is used for recursion.

[0053] Assuming the search space range is [a, b], the mapping formula is:

[0054] X i =a+(ba)·x i

[0055] Where X i is the position of individual i in the search space, x i is a value in the chaotic sequence.

[0056] Step 4.2.3: Generate the initial population.

[0057] For each individual i and each dimension j, the value of the chaotic sequence is mapped to the corresponding dimension of the search space to generate each position of the population. i The generation method is as follows:

[0058] X i =[X i,1 ,X i,2 ,...,X i,D ]

[0059] Each X i,j All are mapped through the values ​​generated by the Tent mapping.

[0060] After completing population initialization, the initial population X iIt is used as the initial input of the dung beetle optimization algorithm for subsequent iterations and optimization.

[0061] Step 4.3: Use Gaussian random walk strategy to enhance the local search capability of the algorithm.

[0062] Step 4.3.1: Calculate the fitness value and apply the standard update rule of the dung beetle optimization algorithm

[0063] The initial population X is obtained by the Tent mapping in step 4.2 i , calculate the fitness value f(X i ), and determine the current optimal individual X best .

[0064] According to the position update mechanism of the dung beetle optimization algorithm, update the position X of each individual i , position update formula:

[0065] X i (t+1)=X i (t)+V i (t)

[0066] Where V i (t) is the velocity of individual i in generation t.

[0067] Step 4.3.2: Introduce Gaussian random walk, fitness evaluation and selection.

[0068] Introduce Gaussian distributed random perturbations to search the local neighborhood, Gaussian random walk formula:

[0069]

[0070] in is the new position after applying the Gaussian random walk, is the current individual position, σ is the step size factor, is a standard normal distribution with mean 0 and variance 1.

[0071] Step 4.4: Use differential mutation strategy to enhance the global search capability of the algorithm.

[0072] Step 4.4.1: Differential mutation strategy.

[0073] For each individual after Gaussian random walk , apply the differential mutation strategy to generate the mutation vector V i :

[0074] V i =X r1 +F·(X r2 -X r3 )

[0075] Among them, X r1 ,X r2 ,X r3 are different individuals randomly selected from the population, and F is a scaling factor.

[0076] Step 4.4.2: Crossover and selection operations.

[0077] The mutation vector Vi is combined with the individual after Gaussian random walk Perform crossover to generate candidate solution U i :

[0078]

[0079] Where CR is the crossover probability, j rand is a randomly chosen dimension.

[0080] Calculate the candidate solution U i The fitness value f(U i ), and decide whether to accept the candidate solution based on the fitness value:

[0081]

[0082] Repeat the above steps until the maximum number of iterations is reached.

[0083] Step 4.5: The optimized dung beetle algorithm iteratively updates a set of solutions, and each update is based on the previous solution to avoid retreating to a suboptimal solution during the search process; decode the optimal solution dung beetle position, output the training error and the optimal learning rate, number of neural network layers and batch processing size of the SWAFormer-FMNet fusion diagnosis model.

[0084] Further, the step 5 includes: using the optimized SWAFormer-FMNet fusion model to perform aircraft engine group bearing fault diagnosis, and classifying the output of the fusion model using a softmax function.

[0085] Step 5.1: Use the softmax function to calculate the probability P of the feature vector being classified into each category after the fully connected layer dimension reduction of the aircraft engine group bearing fault. The calculation formula of the softmax function is as follows:

[0086]

[0087] Where k represents the number of outputs or categories of the neural network, v is the output vector, and v jis the value of the jth output or category in v, i represents the category that needs to be calculated currently, the calculation result is between 0 and 1, and the sum of the softmax values ​​of all categories is 1.

[0088] Step 5.2: Model training uses the cross entropy loss function Loss to adjust the network parameters. The calculation formula is as follows:

[0089]

[0090] In the formula, y represents the actual label, n represents the total number of samples, and x represents traversing all possible categories.

[0091] Beneficial effects:

[0092] The present invention uses EMD+SG filter to decompose the vibration signal, effectively removes the noise information in the vibration signal, and reduces the noise interference in the signal. A SWAFormer model of the improved dung beetle optimization algorithm is established, wherein the improved dung beetle optimization algorithm improves the defects of the dung beetle optimization algorithm that it is easy to fall into local optimality and has limited applicability, and finds the optimal solution at a relatively faster convergence speed. SWAFormer can effectively extract time series and global features from input data. FMNet introduces a focus modulation mechanism to enhance the local feature capture capability of the network architecture, and the combination of SWAFormer and FMNet further enhances its processing capability for complex time series tasks, making the modeling more accurate and improving the accuracy and efficiency of aircraft engine group bearing fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0093] Figure 1 It is a flow chart of the present invention;

[0094] Figure 2 Flowchart of the improved dung beetle optimization algorithm;

[0095] Figure 3 This is the SWAFormer-FMNet neural network structure diagram;

[0096] Figure 4 Schematic diagram of bearing signal decomposition by EMD+SG filter;

[0097] Figure 5 This is a comparison chart of the confusion matrix between the present invention and other models. DETAILED DESCRIPTION

[0098] The present invention is described in detail below with reference to specific examples.

[0099] The present invention discloses a method for diagnosing bearing faults of an aero-engine assembly based on deep learning. Figures 1 to 5 As shown, the specific steps include:

[0100] Step 1: Collect vibration signals of four different situations of inner ring fault, outer ring fault, rolling element fault and normal situation of the rolling bearing of the aircraft engine group at the same speed.

[0101] Step 2: Use EMD (Empirical Mode Decomposition) and SG filter (Savitzky-Golay filter) to decompose and reduce noise of 4 different vibration signals. Figure 4 This is a schematic diagram of bearing signal decomposition using EMD+SG filter in this implementation manner.

[0102] Step 2 specifically includes the following steps:

[0103] Step 2.1: EMD (Empirical Mode Decomposition).

[0104] The original bearing vibration signal x(t) is solved into several intrinsic mode functions (IMFs). The formula is expressed as:

[0105]

[0106] Among them, IMF i (t) is the i-th intrinsic mode function, r n (t) is the residual, which represents the final remaining non-oscillatory part.

[0107] Step 2.2: Savitzky-Golay filter (SG filter)

[0108] An SG filter is used on each IMF to smooth the signal. The core formula of the SG filter is:

[0109]

[0110] Among them, y i is the smoothed signal value, x i+j is the value of the original signal within the window. j are the filter coefficients, obtained by local polynomial fitting. m is half the window size.

[0111] Step 2.3: Extract the time-frequency domain features of the above data after preprocessing through Hilbert transform.

[0112] The formula for the Hilbert transform is:

[0113]

[0114] Where PV is the Cauchy principal value integral, and H(x(t)) is the instantaneous amplitude and phase information of the signal.

[0115] The extracted feature matrix:

[0116] X={x1,x2,...,x n}

[0117] where each x i Feature vectors representing different time points or signal segments.

[0118] Step 2.4: Divide the sample set into training set and test set in a ratio of 8:2.

[0119] Step 3: Apply the generalization ability of SWAFormer and the focal attention mechanism of FMNet to diagnose the original data.

[0120] Step 3 specifically includes the following steps:

[0121] Step 3.1: Extract temporal dependencies and global features from preprocessed data using Tranformer:

[0122]

[0123] MultiHead(Q,K,V)=Concat(head1,...,head h )W O

[0124] Z l =LayerNorm(X+MultiHead(Q,K,V))

[0125] Among them, Q, K, and V are matrix representations of query, key, and value respectively, and d k is the dimension of the query and key. is the weight matrix. L The final feature representation is input into FMNet for local modulation.

[0126] Step 3.2: Use SWA (random weight averaging) to smoothly optimize the weights of the Tranformer model.

[0127] SWA generates a smoother solution by averaging weights at different training steps. Suppose during training, the weight at iteration t is θ t , SWA uses the following formula to update the weights:

[0128]

[0129] Among them, θ t are the model parameters at the tth iteration. T0 is the time point at which the average value is calculated. T is the total number of training steps.

[0130] Step 3.3: Use FMNet to extract the global feature Z from TransformerL The modulation mechanism combines local and global information. The formula of FMNet is

[0131] F modulated (Z L )=F local (Z L )+α·F context (Z L )

[0132] where F local (Z L ) is the local feature module processing Transformer output Z L , F context (Z L ) is the module processing Z for global context information L , α is the modulation coefficient, which controls the degree of fusion of local and global features.

[0133] The data features after FMNet modulation are used as the input of the softmax neural network layer to output the prediction results of bearing faults.

[0134] Step 4: Improve the dung beetle optimization algorithm to obtain an improved dung beetle optimization algorithm. Use the improved dung beetle optimization algorithm to optimize the hyperparameters of the SWAFormer-FMNet diagnostic model, wherein the hyperparameters include learning rate, number of hidden layers, and batch size;

[0135] Step 4 specifically includes the following steps:

[0136] Step 4.1: Set the population size of the dung beetle optimization algorithm to 35, the number of iterations to 50, the upper and lower limits of the search space include: learning rate [0.001-0.1], the number of neural network layers to [10-50], and the batch size to [16-256];

[0137] Step 4.2: Use Tent chaotic map to initialize the population position of the algorithm.

[0138] Step 4.2.1: The Tent mapping formula is as follows:

[0139]

[0140] Among them, x k is the individual value of the kth generation, r is the parameter of the Tent mapping, and r=0.5 is usually taken.

[0141] Step 4.2.2: Generate chaotic sequences and map the search space.

[0142] Randomly select an initial value x0 in the range [0,1], and iteratively generate a chaotic sequence of length N×D from the initial seed x0, where N is the number of individuals in the population and D is the dimension of the individual. The generated chaotic sequence is:

[0143] {x1,x2,...,x N}

[0144] For each individual, the Tent mapping formula is used for recursion.

[0145] Assuming the search space range is [a, b], the mapping formula is:

[0146] X i =a+(ba)·x i

[0147] Where X i is the position of individual i in the search space, x i is a value in the chaotic sequence.

[0148] Step 4.2.3: Generate the initial population.

[0149] For each individual i and each dimension j, the value of the chaotic sequence is mapped to the corresponding dimension of the search space to generate each position of the population. i The generation method is as follows:

[0150] X i =[X i,1 ,X i,2 ,...,X i,D ]

[0151] Each X i,j All are mapped through the values ​​generated by the Tent mapping.

[0152] After completing population initialization, the initial population Z i It is used as the initial input of the dung beetle optimization algorithm for subsequent iterations and optimization.

[0153] Step 4.3: Use Gaussian random walk strategy to enhance the local search capability of the algorithm.

[0154] Step 4.3.1: Calculate the fitness value and apply the standard update rule of the dung beetle optimization algorithm

[0155] The initial population X is obtained by the Tent mapping in step 4.2 i , calculate the fitness value f(X i ), and determine the current optimal individual X best .

[0156] According to the position update mechanism of the dung beetle optimization algorithm, update the position X of each individual i , position update formula:

[0157] X i (t+1)=X i (t)+V i (t)

[0158] Where V i (t) is the velocity of individual i in generation t.

[0159] Step 4.3.2: Introduce Gaussian random walk, fitness evaluation and selection.

[0160] Introduce Gaussian distributed random perturbations to search the local neighborhood, Gaussian random walk formula:

[0161]

[0162] in is the new position after applying the Gaussian random walk, is the current individual position, σ is the step size factor, is a standard normal distribution with mean 0 and variance 1.

[0163] Step 4.4: Use differential mutation strategy to enhance the global search capability of the algorithm.

[0164] Step 4.4.1: Differential mutation strategy.

[0165] For each individual after Gaussian random walk Apply the differential mutation strategy to generate the mutation vector V i :

[0166] V i =X r1 +F·(X r2 -X r3 )

[0167] Among them, X r1 ,X r2 ,X r3 are different individuals randomly selected from the population, and F is a scaling factor.

[0168] Step 4.4.2: Crossover and selection operations.

[0169] The mutation vector Vi is combined with the individual after Gaussian random walk Perform crossover to generate candidate solution U i :

[0170]

[0171] Where CR is the crossover probability, j rand is a randomly chosen dimension.

[0172] Calculate the candidate solution U i The fitness value f(U i ), and decide whether to accept the candidate solution based on the fitness value:

[0173]

[0174] Repeat the above steps until the maximum number of iterations is reached.

[0175] Step 4.5: The optimized dung beetle algorithm iteratively updates a set of solutions, and each update is based on the previous solution to avoid retreating to a suboptimal solution during the search process; decode the optimal solution dung beetle position, output the training error and the optimal learning rate, number of neural network layers and batch processing size of the SWAFormer-FMNet fusion diagnosis model.

[0176] Step 5: Use the optimized SWAFormer-FMNet fusion model to diagnose aircraft engine bearing faults, and classify the output of the fusion model using the softmax function.

[0177] Step 5.1: Use the softmax function to calculate the probability P of the feature vector being classified into each category after the fully connected layer dimension reduction of the aircraft engine group bearing fault. The calculation formula of the softmax function is as follows:

[0178]

[0179] Where k represents the number of outputs or categories of the neural network, v is the output vector, and v j is the value of the jth output or category in v, i represents the category that needs to be calculated currently, the calculation result is between 0 and 1, and the sum of the softmax values ​​of all categories is 1.

[0180] Step 5.2: Model training uses the cross entropy loss function Loss to adjust the network parameters. The calculation formula is as follows:

[0181]

[0182] In the formula, y represents the actual label, n represents the total number of samples, and x represents traversing all possible categories.

[0183] The following bearing fault experiments were carried out using different diagnostic models for the above scheme. Figure 5 This is a comparison chart of the confusion matrix of the present invention and other models. The fault diagnosis errors of different models are as follows:

[0184] Table 1. Error table of the results of the model of the present invention and other control models

[0185]

[0186]

[0187] As shown in Table 1, by comparing Transformer, EMD+SG-Transformer, SWAFormer, EMD+SG-SWAFormer, and SWAFormer-FMNet, it can be seen that the model proposed in the present invention has the highest Accuracy, Recall, and F1-score, indicating that the proposed model has the best diagnostic effect. Taking Accuracy as an example: the value of Transformer is 0.8714, the value of EMD+SG-Transformer is 0.9283, the value of SWAFormer is 0.9682, the value of EMD+SG-SWAFormer is 0.9704, and the value of SWAFormer-FMNet is 0.9959, which can prove the effectiveness of the proposed model. By comparing Transformer and EMD+SG-Transformer, it can be seen that Transformer has a higher prediction effect after data preprocessed by EMD+SG. By comparing Transformer and SWAFormer, it can be found that introducing a new optimizer SWA in Transformer can further improve the diagnostic accuracy of Transformer. By comparing EMD+SG-SWAFormer and SWAFormer-FMNet, it can be seen that the proposed focal attention mechanism FMNet can supplement the local feature capture ability of the SWAFormer model and further improve the diagnostic accuracy of the model.

[0188] The above embodiments are only for illustrating the technical concept and features of the present invention, and their purpose is to enable people familiar with the technology to understand the content of the present invention and implement it accordingly, and they cannot be used to limit the protection scope of the present invention. Any equivalent transformation or modification made according to the spirit of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for diagnosing bearing faults of an aircraft engine block based on deep learning, characterized in that: The following steps are involved: Step 1: Collect vibration signals of four different situations: inner ring fault, outer ring fault, rolling element fault and normal situation of the rolling bearing of the aircraft engine group at the same speed; Step 2: Use empirical mode decomposition (EMD) and Savitzky-Golay filter to decompose and reduce noise of four different vibration signals; Step 3: Apply the generalization ability of SWAFormer and the focal attention mechanism of FMNet to build the SWAFormer-FMNet diagnosis model; Step 4: Tent chaos mapping is used to increase the diversity of intelligent particle algorithm initialization; Gaussian random walk strategy is used to enhance the local search capability of the algorithm; differential mutation strategy is used to enhance the global search capability of the algorithm, and an improved dung beetle optimization algorithm is obtained; the improved dung beetle optimization algorithm is used to optimize the hyperparameters of the SWAFormer-FMNet diagnostic model, including learning rate, number of hidden layers, and batch processing size; Step 5: Use the optimized SWAFormer-FMNet diagnostic model to diagnose the aircraft engine group bearing fault, and classify the output of the SWAFormer-FMNet diagnostic model using the softmax function; Step 6: Use the test set data to test the trained SWAFormer-FMNet fault diagnosis model.

2. The method for diagnosing bearing faults of an aero-engine assembly based on deep learning according to claim 1, characterized in that: The steps of using EMD and Savitzky-Golay filter to decompose and reduce noise of 4 different vibration signals are as follows: Step 2.1: Empirical mode decomposition (EMD) is used to decompose the original bearing vibration signal x(t) into several intrinsic mode functions (IMFs); Step 2.2: Savitzky-Golay filter, use Savitzky-Golay filter on each IMF to smooth the signal; Step 2.3: extracting the time-frequency domain features of the data preprocessed in step 2.2 by Hilbert transform; Step 2.4: Divide the sample set into training set and test set in a ratio of 8:

2.

3. The method for diagnosing bearing faults of an aero-engine assembly based on deep learning according to claim 1, characterized in that: In step 3, the generalization ability of SWAFormer and the focus attention mechanism of FMNet are used to build the SWAFormer-FMNet diagnosis model. The steps are as follows: Step 3.1: Extract temporal dependencies and global features from preprocessed data using Tranformer: MultiHead(Q,K,V)=Concat(head1,...,head h )W O Z l =LayerNorm(X+MultiHead(Q,K,V)) Among them, Q, K, and V are matrix representations of query, key, and value respectively, and d k are the dimensions of the query and key, is the weight matrix, Z L The final feature representation is input into FMNet for local modulation; Step 3.2: Use random weight average SWA to smoothly optimize the weights of the Tranformer model; Assume that during the training process, the weight of the tth iteration is θ t , SWA uses the following formula to update the weights: Among them, θ t is the model parameter of the tth iteration, T0 is the time point when the average value is calculated, and T is the total number of training steps; Step 3.3: Use FMNet to extract the global feature Z from Transformer L The modulation mechanism combines local and global information. The formula of FMNet is F modulated (WITH L )=F local (WITH L )+α F context (WITH L ) where F local (Z L ) is the local feature module processing Transformer output Z L , F context (Z L ) is the module processing Z for global context information L , α is the modulation coefficient, which controls the degree of fusion of local and global features; The data features after FMNet modulation are used as the input of the softmax neural network layer to output the prediction results of bearing faults.

4. The method for diagnosing bearing faults of an aircraft engine block based on deep learning according to claim 1, characterized in that: The improved dung beetle optimization algorithm obtained by improving the dung beetle optimization algorithm in step 4 includes the following steps: Step 4.1: Set the population size of the dung beetle optimization algorithm to 35, the number of iterations to 50, the upper and lower limits of the search space include: learning rate [0.001-0.1], the number of neural network layers to [10-50], and the batch size to [16-256]; Step 4.2: Use the population position of the Tent chaotic map initialization algorithm to generate a chaotic sequence and map the search space to generate the initial population X i ; Step 4.3: Use Gaussian random walk strategy to enhance the local search capability of the algorithm; Step 4.3.1: Calculate the fitness value and apply the standard update rule of the dung beetle optimization algorithm Through the initial population X i , calculate the fitness value f(X i ), and determine the current optimal individual x best ; According to the position update mechanism of the dung beetle optimization algorithm, the position x of each individual is updated i , position update formula: X i (t+1)=X i (t)+V i (t) Among them, V i (t) is the velocity of individual i in generation t; Step 4.3.2: Introduce Gaussian random walk, fitness evaluation and selection; Introduce Gaussian distributed random perturbations to search the local neighborhood, Gaussian random walk formula: in, is the new position after applying the Gaussian random walk, is the current individual position, σ is the step size factor, It is a standard normal distribution with mean 0 and variance 1; Step 4.4: Use differential mutation strategy to enhance the global search capability of the algorithm; Step 4.4.1: Differential mutation strategy; For each individual after Gaussian random walk Apply the differential mutation strategy to generate the mutation vector V i : V i =X r1 +F·(X r2 -X r3 ) Among them, X r1 ,X r2 ,X r3 are different individuals randomly selected from the population, and F is the scaling factor; Step 4.4.2: Crossover and selection operations; The mutation vector V i and individuals undergoing a Gaussian random walk Perform crossover to generate candidate solution U i : Where CR is the crossover probability, j rand is a randomly chosen dimension; Calculate the candidate solution U i The fitness value f(U i ), and decide whether to accept the candidate solution based on the fitness value: Repeat the above steps until the maximum number of iterations is reached; Step 4.5: The optimized dung beetle algorithm iteratively updates a set of solutions, each update is based on the previous solution, to avoid retreating to a suboptimal solution during the search process; decode the optimal solution dung beetle position, and output the training error and the optimal learning rate, number of neural network layers, and batch size hyperparameters of the SWAFormer-FMNet diagnostic model.

5. The method for diagnosing bearing faults of an aircraft engine block based on deep learning according to claim 1, characterized in that: In step 5, the output of the SWAFormer-FMNet diagnostic model is reduced in dimension through a fully connected layer and then classified using a softmax function, as follows: Step 5.1: Use the softmax function to calculate the probability P of the feature vector being classified into each category after the fully connected layer dimension reduction of the aircraft engine group bearing fault. The calculation formula of the softmax function is as follows: Where k represents the number of outputs or categories of the neural network, v is the output vector, and v j is the value of the jth output or category in v, i represents the category that needs to be calculated currently, the calculation result is between 0 and 1, and the sum of the softmax values ​​of all categories is 1; Step 5.2: Model training uses the cross entropy loss function Loss to adjust the network parameters. The calculation formula is as follows: In the formula, y represents the actual label, n represents the total number of samples, and x represents traversing all possible categories.

Images