Aircraft engine bearing fault diagnosis method based on STFT-IncepNext

Abstract

The present invention discloses an aero-engine bearing fault diagnosis method based on STFT-IncepNext, comprising the following steps: step 1, collecting acceleration signals of bearings under various working conditions and fault states as sample data; step 2, first determining the number of sampling points of each sensor, and arranging and combining the data collected by sensors at different positions in the same time series to perform data splicing; then using short-time Fourier transform to convert the fused vibration signal into a time-frequency diagram; step 3, data preprocessing: dividing the training set, test set and validation set according to the fault category, and normalizing the converted time-frequency diagram using the mean and standard deviation; step 4, constructing an IncepNext model, and training and validating it; step 5, using the trained IncepNext model to diagnose the aero-engine bearing fault and obtain the diagnosis result. The present invention can effectively identify aero-engine bearing faults, can ensure the recognition rate of faults, and has good accuracy.

Description

Technical Field

[0001] The invention relates to the technical field of aircraft engine bearing fault diagnosis, and in particular to an aircraft engine bearing fault diagnosis method based on STFT-IncepNext. Background Art

[0002] As one of the core components of aircraft, the complex mechanical structure of aircraft engines faces the test of multiple extreme working environments during operation. Among them, aircraft engine bearings, as the supporting components of the rotor system, are subject to huge challenges in harsh environments such as high temperature, heavy load, high speed, and oil shortage. Its main function is to support the rotation of aircraft engines and ensure that the engine can maintain rotation accuracy and reliability of power transmission during operation. Once a bearing fails, it will cause an aircraft to stop in the air, endangering the lives of passengers.

[0003] Traditional rotating machinery fault diagnosis methods mainly rely on the analysis of signals measured by a single sensor. However, due to the complexity of the mechanical structure of the aircraft engine itself and the variability of the actual operating environment, a single signal is difficult to guarantee the accuracy and completeness of the information obtained. In addition, although electronic components such as sensors used in the aerospace field perform well in life tests, potential performance degradation processes may still lead to product failure or malfunction. Therefore, driven by the limitations of the sensor itself and the internal and external factors of the harsh working environment, a single sensor information is often difficult to meet and guarantee the extraction of aircraft engine bearing fault information.

[0004] With the substantial increase in computing power and the rapid development of computer vision, artificial intelligence technologies such as deep learning have made breakthrough progress. Deep learning algorithms that use one-dimensional and two-dimensional vibration signals as network inputs have been widely used in the field of fault diagnosis, realizing end-to-end intelligent diagnosis of rotating machinery. Given that a single sensor signal is difficult to characterize bearing fault information under complex working conditions, researchers have turned to studying multi-source sensor data fusion technology. This technology can be divided into data layer fusion, feature layer fusion, and decision layer fusion according to the fusion level. In bearing fault diagnosis, traditional methods that use one-dimensional vibration signals as input mostly choose to fuse multi-source data at the feature layer. This is because fusion at the data level will increase the computational complexity of the algorithm model, while fusion at the decision layer will cause a large amount of original vibration information to be lost.

[0005] Therefore, how to effectively use multiple sensors at different positions to monitor the vibration state of engine bearings while retaining the original vibration information and not increasing the amount of model calculation, thereby improving the accuracy of fault diagnosis, has become an urgent problem that needs to be solved. Summary of the invention

[0006] The purpose of the present invention is to provide an aircraft engine bearing fault diagnosis method based on STFT-IncepNext. The present invention can effectively identify aircraft engine bearing faults, can ensure the fault recognition rate, and has good accuracy.

[0007] The technical solution of the present invention: an aircraft engine bearing fault diagnosis method based on STFT-IncepNext, comprising the following steps:

[0008] Step 1: Data collection: Acceleration sensors installed at different positions of the aircraft engine bearings are used to collect acceleration signals of the bearings under various working conditions and fault conditions as sample data;

[0009] Step 2, signal processing: first determine the number of sampling points of each sensor, and arrange and combine the data collected by sensors at different positions in the same time series to perform data splicing; then use short-time Fourier transform to convert the fused vibration signal into a time-frequency diagram;

[0010] Step 3: Data preprocessing: divide the training set, test set, and validation set according to the fault category, and normalize the converted time-frequency graph using the mean and standard deviation;

[0011] Step 4: Build the IncepNext model, input the time-frequency diagram in the training set into the IncepNext model for training, and determine the IncepNext model parameters by the validation set and the test set;

[0012] Step 5: Use the trained IncepNext model to diagnose aircraft engine bearing faults and obtain diagnostic results.

[0013] In the above-mentioned aircraft engine bearing fault diagnosis method based on STFT-IncepNext, in step 1, the faulty bearing is installed at the entrance of the driving shaft into the gearbox, and the acceleration sensor is placed on the gearbox housing; during the three time periods of equipment startup, stable operation and about to end, one set of data is collected every minute for 1 minute.

[0014] The aforementioned aircraft engine bearing fault diagnosis method based on STFT-IncepNext, the bearing faults include rolling element faults, inner ring faults, outer ring faults and combined faults, and a single-point fault hole with a diameter of 0.1 mm is arranged at the corresponding position by using an electrospark machining method.

[0015] In the aforementioned aero-engine bearing fault diagnosis method based on STFT-IncepNext, in step 2, the number of sampling points of each sensor is set to 1000, and the data collected by the acceleration sensor in the same time series are sequentially spliced ​​according to the arrangement and combination, and then a time-frequency diagram with a pixel size of 224×224 is generated by short-time Fourier transform; the expression of the short-time Fourier transform is:

[0016]

[0017] S(t,ω)=STFT(t,ω)STFT * (t,ω)=|STFT(t,ω)| 2 ;

[0018] In the formula, x(τ) represents the original signal, ω(τ-t) is the window function used to truncate the signal in time, and e -jωτ d τ is a complex exponential function used to convert the signal from the time domain to the frequency domain; S(t,ω) represents the gray level of the signal at time t and frequency ω.

[0019] In the aforementioned aircraft engine bearing fault diagnosis method based on STFT-IncepNext, in step 3, the expression for normalizing the mean and standard deviation is:

[0020]

[0021] In the formula, x′ x Represents the pixel value of the i-th channel of the original image, mean i and std i Represent the mean and standard deviation of the ith channel, x i Represents the input feature map.

[0022] The aforementioned aircraft engine bearing fault diagnosis method based on STFT-IncepNext, the IncepNext model includes an input layer, a multi-scale branch convolution module, an EdgeNeXt network module, a fully connected layer and an output layer;

[0023] The input form of the input layer is a multidimensional array of B×H×W×C, where B is the input batch size, and H, W, and C represent the length, width, and number of channels of the input image, respectively;

[0024] The multi-scale branch convolution module includes three branches with different convolution kernel sizes and channel numbers, wherein the introduction of the pooling branch retains the global information of the input features;

[0025] The EdgeNeXt network module includes a convolutional encoder with an N×N adaptive kernel size and a separable deep transposed attention encoder;

[0026] The convolution encoder consists of the following parts: dynamic deep convolution with adaptive kernel size, using convolution kernels of size k = 3, 5, 7, and 9 to correspond to different scales in the four stages; then two Point-wiseCNN layers are used to enrich the local representation, and standard layer normalization and Gaussian error linear unit activation are used for nonlinear feature mapping. Finally, jump connections are used to make the previous and next information flow in the network hierarchy, which can be expressed as:

[0027] x i+1 =x i +Linear G (Linear(LN(D ω (x i ))));

[0028] Among them, x i Represents an input feature map of shape H×W×3, Linear G represents the joint layer of Point-wise CNN and GELU, LN represents the normalization layer, and D ω represents dynamic depth convolution;

[0029] The deep transposed attention encoder module consists of two parts. The first part encodes the data at different spatial levels to enhance the network's ability to learn adaptive multi-scale features, expressed as:

[0030]

[0031] Among them, y i-1 Indicates d i-1 The output of ; s represents the number of subsets, which is adaptive based on the number of stages t, t∈(2, 3, 4);

[0032] The second part uses the transposed Q and K attention feature maps. In this way, the dot multiplication operation originally performed in the spatial dimension is transferred to the channel dimension, and the calculation of the cross covariance between channels is realized to capture the global feature representation and implicit knowledge; the attention feature map is expressed as:

[0033]

[0034] stAttention(Q,K,V)Softmax(Q T K);

[0035] Among them, X represents the input feature, Q, K, V represent the attention weights respectively, Attention() represents the attention operation, and softmax is the activation function.

[0036] The fully connected layer is used to integrate the processed information and obtain single-channel data;

[0037] The output layer uses Softmax to back-propagate the error on the data to optimize the IncepNext model until it converges.

[0038] In the aforementioned aero-engine bearing fault diagnosis method based on STFT-IncepNext, in step 4, the training strategy is to select AdamW as the network optimizer, the weight decay Weight Decay value is 0.05, and the hyperparameters eps=1e-8 and betas=(0.9,0.999) are used to further adjust the performance of the optimizer; the training batch is set to 32, the number of iterations is 30, and the cosine annealing learning rate scheduling strategy is adopted to gradually reduce the learning rate according to the training batch, so that the learning rate gradually decreases to 1% of the initial learning rate during the training process; a linear warm-up strategy is adopted, the initial learning rate is 5e-4, and the learning rate is gradually increased within the initial batches, and the cross entropy loss is used as the loss function; the probability values ​​of various fault categories in the test are obtained through the trained IncepNext model, and the fault label corresponding to the value with the largest probability is selected as the final fault type recognition result.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] (1) The present invention converts the one-dimensional vibration signal into a two-dimensional time-frequency diagram as the input of the model through short-time Fourier transform, which solves the problem of additional computational complexity of the model in data layer fusion. While simplifying the fusion steps, the original vibration information of the bearing in different time and space is completely retained.

[0041] (2) The present invention supplements the vibration information of the bearing by fusing the sensor data from different spaces. The fused data is converted into a time-frequency graph, which contains richer vibration features and has a higher degree of recognition of bearing fault features.

[0042] (3) The STFT-IncepNext model proposed in the present invention has a strong time-frequency graph feature extraction capability while maintaining lightweight, which plays an important role in improving the accuracy of aircraft engine bearing fault detection.

[0043] (4) In addition, in actual production practice, technicians can place sensors and determine the number of sensors according to the position of the bearings they need to monitor. By using multi-sensor fusion technology, the performance of data collected in the actual diagnostic system can be optimized, thereby effectively reducing equipment maintenance costs and improving fault identification accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 is a flow chart of the present invention;

[0045] Figure 2 The time-frequency diagrams of the five fault types obtained under working condition 1;

[0046] Figure 3 This is the model framework diagram of IncepNext;

[0047] Figure 4 Dimensionality reduction visualization of the test set fault diagnosis results after multi-source fusion of sensors.

[0048] Figure 5 Comparison chart of the impact of different sensor sets on model accuracy.

[0049] The present invention is further described below in conjunction with the accompanying drawings and embodiments, but they are not intended to limit the present invention.

[0050] Example: An aircraft engine bearing fault diagnosis method based on STFT-IncepNext, such as Figure 1 As shown, the following steps are included:

[0051] Step 1, data collection: Acceleration sensors installed at different positions of aircraft engine bearings are used to collect acceleration signals of bearings under various working conditions and fault conditions as sample data; in this step, the collected data are related to deep groove ball bearings tested on a helicopter transmission system test bench and a main reducer test bench. The faulty bearing is installed at the entrance of the driving shaft into the gearbox, and the acceleration sensor is placed on the gearbox housing. The speed sensor is used to collect the constant speed output by the motor, with a sampling frequency of 10000Hz. During the three time periods of equipment startup, stable operation and about to end, one group of data was collected every minute for 1 minute. Bearing faults include rolling element faults, inner ring faults, outer ring faults and combined faults. Single-point fault holes with a diameter of 0.1mm are arranged at the corresponding parts using electrospark machining technology.

[0052] In order to simulate the real operating environment of the engine and the state when a fault occurs as much as possible, in this embodiment, 5 types of bearing faults are set, and each type of fault is tested under 4 different working conditions. The sample configuration of each working condition is shown in Table 1, and the sample configuration of the fault label is shown in Table 2.

[0053]

[0054] Table 1

[0055] Label Fault type Working mode Sample size 0 normal 4 800 1 Rolling element 4 800 2 joint 4 800 3 Inner Circle 4 800 4 Outer ring 4 800

[0056] Table 2

[0057] Step 2, signal processing: first determine the number of sampling points of each sensor, and arrange and combine the data collected by sensors at different positions in the same time series to perform data splicing; then use short-time Fourier transform (STFT) to convert the fused vibration signal into a time-frequency diagram;

[0058] In this step, the wrapping method is used to determine the optimal fusion parameters, considering the differences in sensor installation positions and signal feature extraction after data fusion of each sensor. The wrapping method is a method of gradually increasing the number of sensors, which is used to compare the impact of each sensor on the accuracy of the model. The specific signal processing measures are: set the number of sampling points of each sensor to 1000, and sequentially splice the data collected by the five sensors in the same time series according to the permutation and combination, and then use the STFT algorithm to generate 200 samples for each of the four working conditions. Each fault category contains a total of 800 time-frequency diagrams with a pixel size of 224×224. The expression of short-time Fourier transform is:

[0059]

[0060] S(t,ω)=STFT(t,ω)STFT * (t,ω)=|STFT(t,ω)| 2 ;

[0061] In the formula, x(τ) represents the original signal, ω(τ-t) is the window function used to truncate the signal in time, and e -jωτ d τ is a complex exponential function used to convert signals from the time domain to the frequency domain; S(t,ω) represents the gray level of the signal at time t and frequency ω. The two-dimensional image converted in this way is usually called a time-frequency diagram. For the specific form of the time-frequency diagram, see Figure 3 .

[0062] Figure 2 The time-frequency diagrams of the five fault types obtained under working condition 1 are shown in Figure 1. The different frequency components in the sample signal are mapped to the two-dimensional diagram to obtain images with different color depths. It can be seen from the figure that the time-frequency diagrams obtained from different sensor data are quite different. After multi-source sensor data fusion, the frequency information contained in the time-frequency diagram is richer and more complex, which better reflects the original vibration state of the bearing.

[0063] Step 3: Data preprocessing: divide the training set, test set, and validation set according to the fault category, and normalize the converted time-frequency graph using the mean and standard deviation;

[0064] In this step, the converted time-frequency graph is divided into data sets according to the fault category, and the training set, test set, and validation set are constructed in a ratio of 0.7:0.15:0.15. The training set is used for iterative training of the model, the test set is used to evaluate the changes in model accuracy during the training process, and the validation set is used to evaluate the generalization effect of the model. In order to reduce the amount of model training, the time-frequency graph data is normalized using the mean and standard deviation method. The expression for normalization of the mean and standard deviation is:

[0065]

[0066] In the formula, x′ x Represents the pixel value of the i-th channel of the original image, mean i and std i Represent the mean and standard deviation of the ith channel, x i Represents the input feature map.

[0067] Step 4: Build the IncepNext model, input the time-frequency diagram in the training set into the IncepNext model for training, and determine the IncepNext model parameters by the validation set and the test set;

[0068] In this step, the IncepNext model includes an input layer, a multi-scale branch convolution module, an EdgeNeXt network module, a fully connected layer and an output layer;

[0069] The input form of the input layer is a multidimensional array of B×H×W×C, where B is the input batch size, and H, W, and C represent the length, width, and number of channels of the input image, respectively;

[0070] The multi-scale branch convolution module includes three branches with different convolution kernel sizes and channel numbers, wherein the introduction of the pooling branch retains the global information of the input features;

[0071] The EdgeNeXt network module includes a convolutional encoder with an N×N adaptive kernel size and a separable deep transposed attention encoder;

[0072] The convolutional encoder consists of the following parts: dynamic deep convolution with adaptive kernel size, using convolution kernels of size k = 3, 5, 7, and 9 to correspond to different scales in the four stages; then two Point-wiseCNN layers are used to enrich the local representation, and standard layer normalization (LN) and Gaussian error linear unit (GELU) activation are used for nonlinear feature mapping. Finally, jump connections are used to make the previous and next information flow in the network hierarchy, which can be expressed as:

[0073] x i+1 =x i +Li nearG (Linear(LN(D ω (x i ))));

[0074] Among them, x i Represents an input feature map of shape H×W×3, Linear G represents the joint layer of Point-wise CNN and GELU, LN represents the normalization layer, and D ω represents dynamic depth convolution;

[0075] The deep transposed attention encoder module consists of two parts. The first part encodes the data at different spatial levels to enhance the network's ability to learn adaptive multi-scale features, expressed as:

[0076]

[0077] Among them, y i-1 Indicates d i-1 The output of ; s represents the number of subsets, which is adaptive based on the number of stages t, t∈(2, 3, 4);

[0078] The second part uses the transposed Q and K attention feature maps. In this way, the dot multiplication operation originally performed in the spatial dimension is transferred to the channel dimension, and the calculation of the cross covariance between channels is realized to capture the global feature representation and implicit knowledge; the attention feature map is expressed as:

[0079]

[0080] stAttention(Q,K,V)Softmax(Q T K);

[0081] Among them, X represents the input feature, Q, K, V represent the attention weights respectively, Attention() represents the attention operation, and softmax is the activation function.

[0082] The fully connected layer is used to integrate the processed information and obtain single-channel data;

[0083] The output layer uses Softmax to perform error back propagation on the data to optimize the IncepNext model until it converges. For the specific IncepNext model framework, see Figure 3 .

[0084] The training set is input into the IncepNext model for training. The training strategy is: select AdamW as the network optimizer, the weight decay value is 0.05, and use the hyperparameters eps=1e-8 and betas=(0.9,0.999) to further adjust the performance of the optimizer. The training batch is set to 32, the number of iterations is 30, and the cosine annealing learning rate scheduling strategy (CosineAnnealing) is adopted. The learning rate is gradually reduced according to the training batch, so that the learning rate gradually decreases to 1% of the initial learning rate during the training process. A linear warm-up strategy is adopted, the initial learning rate is 5e-4, and the learning rate is gradually increased within the first few batches. The cross-entropy loss (CrossEntropyLoss) is used as the loss function. The probability values ​​of various fault categories in the test are obtained through the trained model, and the fault label corresponding to the value with the largest probability is selected as the final fault type recognition result, see Figure 4 .

[0085] Use accuracy to evaluate the installation position of aircraft engine bearings and the effectiveness of multi-source fusion. For specific evaluation, see Figure 5 .according to Figure 5 It can be seen that the installation position of the sensor has a great influence on the accuracy of the model. Sensors NI_2, NI_3 and NI_5 are installed in the axial, horizontal radial position of the middle bearing of the parallel gearbox and the position of the right side of the parallel gearbox near the upper bearing, which can fully retain the vibration information of the bearing. Therefore, the effect of these three sensors as single-source data input is good. When performing multi-source data fusion, the influence of the installation position of the sensor on the accuracy still exists. The accuracy of the training set and test set of the sensor set NI_35, NI_235, and NI_356 reached 100%, and the generalization performance was also the best. This shows that after the fusion of sensor data in different spaces, the vibration signal of the bearing has been well supplemented, and the generated time-frequency diagram has a good recognition of the vibration characteristics of the bearing in different states, and the model can accurately identify the signal type of the bearing.

[0086] Step 5: Use the trained IncepNext model to diagnose aircraft engine bearing faults and obtain diagnostic results.

[0087] Furthermore, the method proposed in the present invention was tested in comparison with several other methods, and the results are shown in Table 3.

[0088]

[0089] Table 3

[0090] As shown in Table 3, after multi-source data fusion, the diagnostic accuracy of the model has been improved. Compared with the other two conversion methods, the short-time Fourier transform can better extract the frequency characteristics of aircraft engine bearings under complex working conditions. Further comparison of various methods shows that the IncepNext model has achieved the highest recognition accuracy when the model parameter is only 1.46M, and it still maintains the best diagnostic effect even on data sets with difficult-to-distinguish frequency characteristics. This shows that the network structure proposed in the present invention has a strong time-frequency graph feature extraction capability.

[0091] In summary, the present invention can effectively identify aircraft engine bearing faults, can ensure the fault recognition rate, and has good accuracy.

Claims

1. The aircraft engine bearing fault diagnosis method based on STFT-IncepNext is characterized by: The steps include: Step 1: Data collection: Acceleration sensors installed at different positions of the aircraft engine bearings are used to collect acceleration signals of the bearings under various working conditions and fault conditions as sample data; Step 2, signal processing: first determine the number of sampling points of each sensor, and arrange and combine the data collected by sensors at different positions in the same time series to perform data splicing; then use short-time Fourier transform to convert the fused vibration signal into a time-frequency diagram; Step 3: Data preprocessing: divide the data into training set, test set and validation set according to the fault category, and normalize the converted time-frequency graph using the mean and standard deviation; Step 4: Build the IncepNext model, input the time-frequency diagram in the training set into the IncepNext model for training, and determine the IncepNext model parameters by the validation set and the test set; Step 5: Use the trained IncepNext model to diagnose the aircraft engine bearing fault and obtain the diagnosis result; The IncepNext model includes an input layer, a multi-scale branch convolution module, an EdgeNeXt network module, a fully connected layer and an output layer; The input form of the input layer is a multidimensional array of B×H×W×C, where B is the input batch size, and H, W, and C represent the length, width, and number of channels of the input image, respectively; The multi-scale branch convolution module includes three branches with different convolution kernel sizes and channel numbers, wherein the introduction of the pooling branch retains the global information of the input features; The EdgeNeXt network module includes a convolutional encoder with an N×N adaptive kernel size and a separable deep transposed attention encoder; The convolution encoder consists of the following parts: dynamic deep convolution with adaptive kernel size, using convolution kernels of size k=3, 5, 7, and 9 to correspond to different scales in the four stages; then two point-wise CNN layers are used to enrich the local representation, and standard layer normalization and Gaussian error linear unit activation are used for nonlinear feature mapping. Finally, jump connections are used to make the previous and next information flow in the network hierarchy, which can be expressed as: x i+1 =x i +Linear G (Linear(LN(D ω (x i )))); Among them, x i Represents an input feature map of shape H×W×3, Linear G represents the joint layer of Point-wise CNN and GELU, LN represents the normalization layer, and D ω represents dynamic depth convolution; The deep transposed attention encoder module consists of two parts. The first part encodes the data at different spatial levels to enhance the network's ability to learn adaptive multi-scale features, expressed as: Among them, y i-1 Indicates d i-1 The output of ; s represents the number of subsets, which is adaptive based on the number of stages t, t∈(2,3,4); The second part uses the transposed Q and K attention feature maps. In this way, the dot multiplication operation originally performed in the spatial dimension is transferred to the channel dimension, and the calculation of the cross covariance between channels is realized to capture the global feature representation and implicit knowledge; the attention feature map is expressed as: s.t.Attention(Q,K,V)Softmax(Q T ·K); Among them, X represents the input feature, Q, K, V represent the attention weights, Attention() represents the attention operation, and softmax is the activation function; The fully connected layer is used to integrate the processed information and obtain single-channel data; The output layer uses Softmax to back-propagate the error on the data to optimize the IncepNext model until it converges.

2. The aircraft engine bearing fault diagnosis method based on STFT-IncepNext according to claim 1 is characterized in that: In step 1, the faulty bearing was installed at the entrance of the driving shaft into the gearbox, and the acceleration sensor was placed on the gearbox housing; during the three time periods of equipment startup, stable operation and about to end, one set of data was collected every minute for 1 minute.

3. The aircraft engine bearing fault diagnosis method based on STFT-IncepNext according to claim 1 is characterized in that: The bearing faults include rolling element fault, inner ring fault, outer ring fault and combined fault, and single-point fault holes with a diameter of 0.1 mm are arranged at corresponding positions by using an electric spark machining method.

4. The aircraft engine bearing fault diagnosis method based on STFT-IncepNext according to claim 1 is characterized in that: In step 2, the number of sampling points of each sensor is set to 1000, and the data collected by the acceleration sensor in the same time series are sequentially spliced ​​according to the permutation and combination, and then a time-frequency diagram with a pixel size of 224×224 is generated by short-time Fourier transform; the expression of the short-time Fourier transform is: S(t,ω)=STFT(t,ω)STFT * (t,ω)=|STFT(t,ω)| 2 ; In the formula, x(τ) represents the original signal, ω(τ-t) is the window function used to truncate the signal in time, and e -jωτ d τ is a complex exponential function used to convert the signal from the time domain to the frequency domain; S(t,ω) represents the gray level of the signal at time t and frequency ω.

5. The aircraft engine bearing fault diagnosis method based on STFT-IncepNext according to claim 1 is characterized in that: In step 3, the expression for normalizing the mean and standard deviation is: In the formula, x′ x Represents the pixel value of the i-th channel of the original image, mean i and std i Represent the mean and standard deviation of the ith channel, x i Represents the input feature map.

6. The aircraft engine bearing fault diagnosis method based on STFT-IncepNext according to claim 1, characterized in that: In step 4, the training strategy is to select AdamW as the network optimizer, the weight decay value is 0.05, and the hyperparameters eps = 1e-8 and betas = (0.9, 0.999) are used to further adjust the performance of the optimizer; the training batch is set to 32, the number of iterations is 30, and the cosine annealing learning rate scheduling strategy is adopted to gradually reduce the learning rate according to the training batch, so that the learning rate gradually decreases to 1% of the initial learning rate during the training process; a linear warm-up strategy is adopted, the initial learning rate is 5e-4, and the learning rate is gradually increased within the first few batches, and the cross entropy loss is used as the loss function; the probability values ​​of various fault categories in the test are obtained through the trained IncepNext model, and the fault label corresponding to the value with the largest probability is selected as the final fault type recognition result.

Images