A Method for Predicting the Remaining Useful Life of Aero-Engines Based on FEFCNet

Abstract

The present invention relates to a method for predicting the remaining useful life of an aero-engine based on FEFCNet, belonging to the field of predicting the remaining useful life of mechanical equipment. The method includes: obtaining aero-engine sensor data and performing clustering, normalization, sliding window processing, and remaining useful life label setting to obtain a training set and a validation set; subsequently, using the fast Fourier transform to extract the significant frequency components of the sample data and calculate the corresponding cycle length, reconstructing the data dimension to obtain upsampled sample data; subsequently, fusing local details and global features through an embedding layer, a multi-scale dilated convolutional layer, an Efficient Multi-Scale Attention (EMA) module, and a multi-head self-attention mechanism to obtain multi-scale fusion features; further using Wavelet Kolmogorov-Arnold Networks (WavKAN) for feature compression; and finally mapping the compressed features to the remaining useful life through a fully connected layer. The present invention can effectively improve the accuracy of predicting the remaining useful life of an aero-engine, provide a reliable basis for timely maintenance and repair, and reduce the operation and maintenance risks.

Description

Technical Field

[0001] This method relates to the technical field of predicting the remaining useful life of mechanical equipment, and particularly to a method for predicting the remaining useful life of an aero-engine based on FEFCNet. Background Art

[0002] Predictive and Health Management (PHM) has been playing an important role in improving operation efficiency, system reliability and performance. As an important part of PHM, the prediction of the remaining useful life (RUL) provides a quantitative basis for maintenance decision-making. Premature maintenance will waste a large amount of resources, while untimely maintenance will cause equipment damage. Therefore, more accurate RUL prediction enables managers to maintain the system in a timely manner, which can not only increase the safety and reliability of the system, but also reduce waste of resources.

[0003] Generally speaking, aero-engine RUL prediction methods can be divided into three categories: model-based prediction methods, data-driven prediction methods and hybrid-driven prediction methods. Both model-based prediction methods and hybrid-driven prediction methods need to establish a physical degradation model of the machine by studying the degradation principle of the machine for prediction, which is relatively accurate and has good interpretability. However, it is limited by challenges such as the high difficulty of modeling complex systems, high cost, and poor generalization ability. In recent years, due to the rapid development of deep learning, data-driven prediction methods have gradually attracted the attention of researchers due to their advantages of high prediction accuracy and low requirements for prior knowledge.

[0004] Traditional data-driven methods highly rely on the quality of features, and the weak correlation between features and degradation information will significantly reduce the prediction accuracy. Therefore, a method with stronger feature extraction ability is needed to improve the prediction accuracy. Summary of the Invention

[0005] Aiming at the above defects or improvement requirements of the prior art, the present invention provides a method for predicting the remaining useful life of an aero-engine based on FEFCNet, which can effectively improve the accuracy of predicting the remaining useful life of an aero-engine, so as to timely overhaul and maintain the aero-engine and extend its service life. To achieve the above object, the technical solution adopted by the present invention is:

[0006] A method for predicting the remaining useful life of an aero-engine based on FEFCNet includes the following steps:

[0007] S1: Obtain the sensor data of the aero-engine and perform preprocessing to determine the training set and the validation set;

[0008] S2: Process the preprocessed data using the Fast Fourier Transform, extract several frequency components with the largest corresponding amplitudes and calculate the corresponding period lengths, and then reconstruct the data dimensions to obtain the sample data after dimensionality increase.

[0009] S3: Extract the degradation features of the data after dimensionality increase through the feature extraction module; the feature extraction module includes an embedding layer, a multi-scale dilated convolutional layer, an Efficient Multi-Scale Attention (EMA) module, and a multi-head self-attention mechanism module. Among them, the embedding layer is used to perform dimensional mapping on the data, mapping the feature dimension of the input data to a fixed dimension d model ; the multi-scale dilated convolutional layer is used to extract local degradation features at different scales; the Efficient Multi-Scale Attention (EMA) module is used to fuse spatio-temporal features, enhance the weights of key time steps and feature channels, and suppress noise interference; the multi-head self-attention mechanism module is used to process the data obtained after the preprocessing in step S1 to capture its global time series dependency relationship.

[0010] S4: Perform dimensionality reduction and compression on the degradation features through the feature compression module to obtain compressed features.

[0011] S5: Map the compressed features to the remaining useful life values through the RUL mapping module and output the prediction results.

[0012] Furthermore, the data preprocessing steps in S1 include:

[0013] Use the K-Means algorithm to cluster the data according to the working conditions to obtain the operating data of the aero-engine under different working conditions, and perform Min-Max normalization on the operating data of the aero-engine under each working condition. The formula is as follows:

[0014]

[0015] where x k,i,j is the original sensor data under the k-th working condition, is the data after standardization under the k-th working condition, k represents the k-th working condition, i represents the i-th sensor, j represents the j-th data point, are the maximum and minimum values of the i-th sensor data under the k-th working condition respectively.

[0016] Adopt the sliding window method to generate two-dimensional time window data, and obtain multiple two-dimensional time window data according to the set sliding window size and sliding step length where represents the set of real numbers, represents d w rows df A two-dimensional matrix of columns, where d w Indicates the sliding window size, d f Represents the number of features.

[0017] A threshold of a RUL label is set, and a RUL label is set for the two-dimensional time window data; specifically, the RUL corresponding to the last row of data in the two-dimensional time window data is used as its RUL.

[0018] The obtained labeled two-dimensional time window data is randomly divided into training set and validation set according to the ratio of 8:2.

[0019] Furthermore, the method in S2 includes:

[0020] The fast Fourier transform is used to process the two-dimensional time window data. The fast Fourier transform is performed on each two-dimensional time window data x along the time dimension. The formula is as follows:

[0021]

[0022] Among them, X(f) represents the spectrum corresponding to f, f represents the frequency, d w represents the sliding window size of the two-dimensional time window data, and x(t) represents the data in the t+1th row of the two-dimensional time window data.

[0023] Calculate the amplitude corresponding to each frequency and average it in the feature dimension to get the average amplitude A, which is as follows:

[0024]

[0025] Among them, d f Represents the number of features of two-dimensional time window data, X n (f) represents the frequency spectrum of the nth feature.

[0026] Select the first k significant frequencies {f1,f2,…,f k}, for each significant frequency f i , calculate the corresponding period length For each candidate period p i , split the time dimension of the original time series into two dimensions, and obtain k three-dimensional data {x1, x2, ..., x k},in

[0027] Furthermore, in S3, the embedding layer is a two-dimensional convolutional layer that performs dimension mapping on the data, mapping the feature dimension of the input data to a fixed dimension d model ;

[0028] The multi-scale dilated convolutional layer includes 2 multi-scale dilated convolutional sub-layers. Each sub-layer includes 7 parallel feature extraction branches. The convolutional layers of each branch adopt different convolutional kernel sizes and dilation rates to extract features of different scales. By setting the padding parameters, the feature maps obtained by each branch are of the same size as the original data. The first sub-layer performs non-linear transformation through the GELU activation function, and the parallel convolutional branches extract local degraded features of different scales.

[0029] The multi-head self-attention mechanism module is used to process the two-dimensional time window data obtained in step S1 to capture its global time series dependencies and make up for the locality limitation of the convolutional operation

[0030] Further, the extraction of the degraded features of the upsampled data by the degraded feature extraction module in S3 includes:

[0031] Passing k upsampled sample data through the two-dimensional convolutional embedding layer respectively, and mapping the last dimension d f to a fixed dimension d model to obtain k new three-dimensional data where

[0032] Pass the k new three-dimensional data through the multi-scale dilated convolutional layer to extract local degraded features of different scales. The multi-scale dilated convolutional layer includes 2 multi-scale dilated convolutional sub-layers. Each sub-layer includes 7 parallel feature extraction branches. The convolutional layers of each branch adopt different convolutional kernel sizes and dilation rates to extract features of different scales. By setting the padding parameters, the feature maps obtained by each branch are of the same size as the original data; the first sub-convolutional layer performs non-linear transformation through the GELU activation function.

[0033] For each three-dimensional data, after passing through each multi-scale dilated convolutional sub-layer, 7 feature maps {x k,1 , x k,2 , …, x k,7} are obtained, where x k,i represents the feature map obtained after the k-th three-dimensional data passes through the i-th branch; the 7 obtained feature maps are concatenated in the last feature dimension and averaged to obtain a fused feature map; since there are k data, finally k fused feature maps are obtained where

[0034] Enhance the weights of the key time steps and feature channels through the EMA module, suppress noise interference, and obtain k enhanced fused feature maps where Next, the first two dimensions of each enhanced fusion feature map are merged, all groups of features are concatenated along the feature dimension, the average of the concatenated result is calculated along the feature dimension, and then the second dimension size d is mapped to d through a linear projection model by f , and finally the overall enhanced fusion feature map is obtained

[0035] The two-dimensional time window data obtained in step S1 is processed by the multi-head self-attention mechanism to capture its global time series dependence; and it is added to the overall enhanced fusion feature map x AE_fused to obtain the final output x of the degradation feature extraction module final , making up for the locality limitation of the convolution operation and realizing the deep fusion of local and global degradation features. The formula is as follows:

[0036] x final = Multiheadself Attention(x) + x AE_fused

[0037] where, Multiheadself Attention represents the multi-head self-attention mechanism module, and x final represents the final output of the degradation feature extraction module.

[0038] Furthermore, the degradation features are dimensionally reduced and compressed by the feature compression module described in S4 to obtain compressed features, including:

[0039] Using 2 parallel WavKANs to compress the fusion feature map obtained in S3 in the time dimension; each WavKAN uses a different type of wavelet basis to specifically extract different types of features; a parameter α is used to perform feature fusion on the features extracted by different WavKANs.

[0040] The different types of wavelet bases include Mexican Hat wavelet and Derivative of Gaussian wavelet; local smooth features and edge mutation features are extracted by the Mexican Hat wavelet and Derivative of Gaussian wavelet respectively, covering the diversity of degradation modes.

[0041] Furthermore, the WavKAN calculation process includes:

[0042] Performing scale and translation transformation on the input data, and the formula is as follows:

[0043]

[0044] where x' represents the data after scale and translation transformation, and are learnable parameters, representing the scale parameter and the translation parameter respectively.

[0045] Perform a transformation on x' using the wavelet basis function to obtain Expand it to through the broadcasting mechanism and then perform an element-wise multiplication operation with the learnable weight matrix and sum the resulting values in dimension d w to obtain the final output of WavKAN

[0046] Perform feature fusion on the features extracted by different WavKANs, automatically optimize the weights of different wavelet branches using α, and enhance the robustness of feature expression. The formula is as follows:

[0047] x' final = αWavKAN1(x final )+(1 - α)WavKAN2(x final )

[0048] where WavKAN1 and WavKAN2 are WavKANs using different wavelet bases respectively; α is a learnable parameter that is dynamically adjusted during training through backpropagation; is the final output of the feature compression layer.

[0049] Furthermore, the RUL mapping module described in S5 includes:

[0050] Two fully connected layers. The first fully connected layer performs a non-linear mapping through the GELU activation function, and the second fully connected layer performs a non-linear mapping through the ReLU activation function to fit the complex non-linear relationship between the degradation features and the remaining useful life, and ensure that the output life value is non-negative.

[0051] Furthermore, mapping the obtained compressed feature data to RUL, as described in S5, includes:

[0052] Map the compressed features obtained in S4 to obtain the RUL prediction value. The formula is as follows:

[0053] RUL pred = ReLU(FC2(GELU(FC1(x' final ))))

[0054] where FC1 and FC2 are the first and second fully connected layers respectively, and GELU and ReLU are activation functions. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 is the flowchart of the present invention

[0056] Figure 2 It is the structure diagram of FEFCNet

[0057] Figure 3 It is the structure diagram of multi-scale dilated convolution

[0058] Figure 4 It is the RUL prediction result diagram of the FD003 test set

[0059] Figure 5 It is the RUL prediction result diagram of the FD004 test set Specific implementation manners

[0060] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0061] The present invention provides a method for predicting the remaining useful life of an aero-engine based on FEFCNet. The method flow is as Figure 1 shown, and the FEFCNet structure is as Figure 2 shown. Specifically, it includes the following steps:

[0062] S1: Obtain a specific subset (FD001 / FD002 / FD003 / FD004) in the NASA-CMAPSS dataset. First, screen out the channels with constant measurement values among the 21 sensors (finally, 14 effective sensors are retained). For the FD002 and FD004 subsets (including 6 working condition conditions), use the K-Means algorithm to cluster the data into 6 categories according to the operating condition characteristics; while the FD001 and FD003 subsets (only containing 1 working condition) do not need to be clustered. Finally, perform Min-Max normalization processing on all data (including the subclass data generated by clustering) respectively. The formula is as follows:

[0063]

[0064] where x k,i,j is the original sensor data under the kth working condition, is the data after standardization under the kth working condition, k represents the kth working condition, i represents the ith sensor, j represents the jth data point, are the maximum and minimum values of the ith sensor data under the kth working condition respectively.

[0065] Concatenate the 14-column sensor measurement values after Min-Max normalization with the three columns of data representing the working conditions to obtain 17-column aero-engine operating state data.

[0066] The sliding window technique is used to process 17 columns of aero-engine operating status data. The size of the sliding window is set as the hyperparameter W, and the sliding step is set to 1, obtaining multiple two-dimensional time window data. Where d w represents the size of the sliding window, and d f represents the number of features, with a size of 17.

[0067] Referring to the practices of most studies, the RUL threshold is set to 125, and the RUL corresponding to the last row data of each two-dimensional time window data x is used as the RUL of x.

[0068] The training set is randomly divided into a training set and a validation set in a ratio of 8:2.

[0069] S2: Use the fast Fourier transform to process the two-dimensional time window data. Perform the fast Fourier transform on each two-dimensional time window data x along the time dimension. The formula is as follows:

[0070]

[0071] Where X(f) represents the spectrum corresponding to f, f represents the frequency, and d w represents the size of the time window of the two-dimensional time window data, and x(t) represents the data of the two-dimensional time window data in the (t + 1)-th row.

[0072] Calculate the amplitude corresponding to each frequency and obtain the average amplitude A by taking the average in the feature dimension. The formula is as follows:

[0073]

[0074] Where d f represents the number of features of the two-dimensional time window data, and X n (f) represents the spectrum of the n-th feature.

[0075] Select the top k significant frequencies {f1, f2,..., f k} with the largest average amplitude. For each significant frequency f i , calculate the corresponding period length For each candidate period p i , split the time dimension into two dimensions to obtain k three-dimensional data {x1, x2,..., x k} after dimensionality increase, where

[0076] S3: Pass the k three-dimensional data after dimensionality increase through a two-dimensional convolutional layer respectively, and map the last dimension d f to a fixed dimension size d model, obtain k new three-dimensional data wherein

[0077] Preliminarily extract the degradation features of aero-engine operation data from these k new three-dimensional data through a multi-scale dilated convolutional layer. The multi-scale dilated convolutional layer is as Figure 3 shown, including 2 multi-scale dilated convolutional sub-layers. Each sub-layer includes 7 parallel feature extraction branches. The convolutional layers of each branch adopt different convolutional kernel sizes and dilation rates to extract features of different scales. Through padding, the feature maps obtained by each branch are made to have the same size as the original data; the first sub-convolutional layer performs a non-linear transformation through the GELU activation function.

[0078] For each three-dimensional data, after passing through the multi-scale dilated convolutional layer, 7 feature maps {x k,1 , x k,2 , …, x k,7} are obtained, where x k,i represents the feature map obtained after the k-th three-dimensional data passes through the i-th branch. The 7 obtained feature maps are concatenated along the last dimension and averaged along the last dimension to obtain a fused feature map; since there are k three-dimensional data, finally k fused feature maps are obtained wherein

[0079] Enhance the weights of key time steps and feature channels through the EMA module, suppress noise interference, and obtain k enhanced fused feature maps wherein Then, merge the first two dimensions of each enhanced fused feature map, concatenate all groups of features along the feature dimension, and average the concatenated result along the feature dimension; then project the size d model of the second dimension to d f through linear projection, and finally obtain the overall enhanced fused feature map

[0080] Process the two-dimensional time window data obtained by S1 through the multi-head self-attention mechanism and add it to the overall enhanced fused feature map x AE_fused to obtain the final output x final of the degradation feature extraction module. Its formula is as follows:

[0081] x final = Multiheadself Attention(x) + x AE_fused

[0082] wherein, Multiheadself Attention represents the multi-head self-attention mechanism module, and x finalRepresents the final output of the degradation feature extraction module.

[0083] S4: Use two parallel WavKANs to compress the enhanced fused feature map in the time dimension. The two WavKANs use the Derivative of Gaussian wavelet and the Mexican Hat wavelet as the wavelet bases respectively. The formulas for the corresponding two wavelets are as follows:

[0084]

[0085] The specific operation of each WavKAN is as follows:

[0086] Perform scale and translation transformation on the input data, and its formula is as follows:

[0087]

[0088] where x' represents the data after scale and translation transformation, and are learnable parameters, representing the scale parameter and the translation parameter respectively.

[0089] Multiply by the broadcast mechanism to expand it to and then perform an element-wise multiplication operation with the learnable weight matrix and sum the resulting values in the d w dimension to obtain the final output of WavKAN

[0090] Fuse the features extracted by different WavKANs to obtain the compressed features Its formula is as follows:

[0091] x' final = αWavKAN DoG (x final )+(1 - α)WavKAN MH (x final )

[0092] where, WavKAN DoG and WavKAN MH are the WavKANs using the Derivative of Gaussian wavelet and the Mexican Hat wavelet as the wavelet bases respectively, α is a learnable parameter (dynamically adjusted by backpropagation during training), is the final output of the feature compression layer.

[0093] S5: Construct two fully-connected layers. The first fully-connected layer has 192 neurons, and the first fully-connected layer performs non-linear mapping through the GELU activation function. The second fully-connected layer has 1 neuron and performs non-linear mapping through the ReLU activation function; x' final After passing through the two fully-connected layers, the RUL prediction value RUL is obtained pred , and its formula is as follows:

[0094] RUL pred = ReLU(FC2(GELU(FC1(x' final ))))

[0095] where FC1 and FC2 represent the two fully-connected layers respectively, and GELU and ReLU are activation functions.

[0096] Use the obtained training set to train the model, and use the validation set to guide the hyperparameter tuning of the model. During the training process, use the MSE loss function and use grid search to find the optimal hyperparameters.

[0097] Based on the above method framework, perform remaining useful life prediction modeling on subsets FD001 - FD004, where hyperparameter optimization is completed using the grid search algorithm: within the preset hyperparameter space (learning rate lr ∈ {0.001, 0.005, 0.01}, batch size b ∈ {32, 64, 128}, sliding window size d w ∈ {25, 30, 35, 40, 45, 50}, number of multi-head self-attention heads n heads ∈ {2, 4}, embedding dimension d model ∈ {16, 32, 64}, number of the first k significant frequency components k ∈ {4, 6, 8, 10}) to optimize the parameters with the minimum MSE of the validation set as the optimization goal. The model performance evaluation uses the root mean square error (RMSE) and score (Score) as evaluation indicators. The calculation formulas of RMSE and Score are as follows:

[0098]

[0099] where n is the number of test sample data, is the remaining life prediction value of the i-th sample, is the true remaining life value of the i-th sample.

[0100] Figure 4 and Figure 5RUL prediction result graphs for the FD003 and FD004 datasets respectively. It can be seen from the graphs that for engines with a relatively small true remaining useful life, since the degradation characteristics are obvious in the later stage of aero-engine operation, the model can easily extract these characteristics, so the prediction error of the model is very small. The remaining useful life prediction errors of these engines in the FD003 dataset are concentrated in the range of 0 - 10, and those in the FD004 dataset are concentrated in the range of 0 - 20. For engines with a relatively large true remaining useful life, since the degradation characteristics are not obvious in the early stage of aero-engine operation, it is more difficult for the model to extract these characteristics, so the prediction error is relatively large. The prediction errors of these engines in the FD003 dataset are concentrated in the range of 0 - 20, and only a small number of prediction errors are greater than 20. The prediction errors of these engines in the FD004 dataset are concentrated in the range of 20 - 30, and only a small number of prediction errors are greater than 30. Therefore, it can be considered that this method has excellent degradation feature extraction ability.

[0101] From the comparison of the prediction performances of different methods on the four sub-datasets of C-MAPSS in Table 1, it can be seen that the FEFCNet method of the present invention shows significant advantages in all four sub-datasets (FD001 - FD004).

[0102] Specifically manifested as:

[0103] RMSE: The average RMSE of FEFCNet is 13.81, which is 6.6% lower than that of GA-Transformer (14.78). Especially in the complex working condition dataset FD004, the RMSE is 16.64, which is 17.4% lower than that of GA-Transformer (20.15), verifying its strong adaptability to multi-condition scenarios;

[0104] Score: The average Score of FEFCNet is 742.58, which is at least 28.7% lower than that of other methods (1047 - 1265). Among them, the Score (1406.54) of the FD004 dataset is significantly optimized by 50.7% compared with DSSN (2852.81), indicating that its prediction deviation is smaller and the timeliness of maintenance decision-making is stronger.

[0105] Table 1 Comparison of prediction performances of different methods on the four sub-datasets of C-MAPSS

[0106]

[0107] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some of the technical features. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A method for predicting the remaining useful life of an aero-engine based on FEFCNet, characterized in that, It includes the following steps: S1: Obtain the sensor data of the aero-engine and perform preprocessing to determine the training set and the validation set; S2: Process the preprocessed data using the fast Fourier transform, extract several frequency components with the largest corresponding amplitudes and calculate the corresponding period lengths, and then reconstruct the data dimension to obtain the sample data with increased dimension; S3: Extract the degradation features of the data after dimensionality increase through the feature extraction module; the feature extraction module includes an embedding layer, a multi-scale dilated convolutional layer, an Efficient Multi-Scale Attention (EMA) module, and a multi-head self-attention mechanism module. Among them, the embedding layer is used to perform dimensional mapping on the data, mapping the feature dimension of the input data to a fixed dimension d model ; the multi-scale dilated convolutional layer is used to extract local degradation features at different scales; the Efficient Multi-Scale Attention (EMA) module is used to fuse spatio-temporal features, enhance the weights of key time steps and feature channels, and suppress noise interference; the multi-head self-attention mechanism module is used to process the data obtained after the preprocessing in step S1 and capture its global time series dependence; S4: Perform dimensionality reduction and compression on the degraded features through the feature compression module to obtain the compressed features; S5: Map the compressed features to the remaining useful life values through the RUL mapping module and output the prediction results.

2. The method for predicting the remaining useful life of an aero-engine based on FEFCNet according to claim 1, wherein, The preprocessing in S1 includes: Use the K-Means algorithm to cluster the data according to the operating conditions, and perform Min-Max normalization on each type of operating condition data respectively; Generate two-dimensional time window data of aero-engine operating states using the sliding window method, and each data sample where represents the set of real numbers, represents d w d rows f d columns of a two-dimensional matrix, where d w represents the sliding window size, and d f represents the number of features; Set the threshold of RUL and set the RUL label for the two-dimensional time window data. Specifically, use the RUL corresponding to the last row of data in the two-dimensional time window data as its RUL; Randomly divide the labeled two-dimensional time window data obtained according to the ratio of 8:2 to obtain the training set and the validation set.

3. The method for predicting the remaining useful life of an aero-engine based on FEFCNet according to claim 1, wherein S2 includes: Process the two-dimensional time window data using the fast Fourier transform, perform the fast Fourier transform on each two-dimensional time window data x along the time dimension, and the formula is as follows: Among them, X(f) represents the spectrum corresponding to f, f represents the frequency, and d w represents the sliding window size of the two-dimensional time window data, and x(t) represents the data of the two-dimensional time window data in the (t + 1)-th row; Calculate the amplitude corresponding to each frequency, and obtain the average amplitude A by taking the average in the feature dimension. The formula is as follows: Among them, d f represents the number of features of two-dimensional time window data, and X n (f) represents the spectrum of the nth feature; Select the first k frequency components {f1, f2, …, f k} with the largest amplitudes, and calculate the corresponding cycle lengths respectively Split the time dimension by frequency and period length to generate k three-dimensional data {x1, x2, …, x k}, where 4. The method for predicting the remaining useful life of an aero-engine based on FEFCNet according to claim 1, wherein: The embedding layer is a two-dimensional convolutional layer that performs dimensional mapping on the data, mapping the feature dimension of the input data to a fixed dimension d model ; The multi-scale dilated convolutional layer includes 2 multi-scale dilated convolutional sub-layers. Each sub-layer includes 7 parallel feature extraction branches. The convolutional layers of each branch use different convolutional kernel sizes and dilation rates to achieve the extraction of features of different scales. By setting the padding parameter, the feature maps obtained by each branch are the same size as the original data. The first sub-layer performs a non-linear transformation through the GELU activation function, and the parallel convolutional branches extract local degraded features of different scales; The multi-head self-attention mechanism module is used to process the two-dimensional time window data obtained in step S1 Capture its global time series dependencies and make up for the locality limitation of the convolution operation.

5. The method for predicting the remaining useful life of an aero-engine based on FEFCNet according to claim 4, wherein The degraded feature extraction module realizes feature extraction through the following steps: Perform dimensional mapping on the three-dimensional data {x1, x2, …, x k} obtained in step S2 through the embedding layer to obtain k new three-dimensional data where Extract local degradation features of different scales from the new k three-dimensional data through the multi-scale dilated convolutional layer; for each three-dimensional data, after passing through each multi-scale dilated convolutional sub-layer, 7 feature maps {x k,1 , x k,2 , …, x k,7} are obtained, where x k,i represents the feature map obtained after the k-th three-dimensional data passes through the i-th branch; the 7 obtained feature maps are concatenated in the last feature dimension and averaged to obtain a fused feature map; since there are k data, finally k fused feature maps are obtained where Enhance the weights of key time steps and feature channels through the EMA module, suppress noise interference, and obtain k enhanced fusion feature maps Among them Then, merge the first two dimensions of each enhanced fusion feature map, splice all groups of features along the feature dimension, average the splicing results along the feature dimension, and then map the size d of the second dimension to d through linear projection model to d f , and finally obtain the overall enhanced fusion feature map The two-dimensional time window data obtained by processing S1 through the multi-head self-attention mechanism Capture its global time series dependencies and fuse them with the overall enhanced feature map x AE_fused Add them together to obtain the final output x of the degradation feature extraction module final , and the formula is as follows: x final = MultiheadselfAttention(x) + x AE_fused Among them, MultiheadselfAttention represents the multi-head self-attention mechanism module, and x final represents the final output of the degradation feature extraction module.

6. The method for predicting the remaining useful life of an aeroengine based on FEFCNet according to claim 1, wherein The feature compression module in S4 includes: Two parallel Wavelet Kolmogorov-Arnold Networks (WavKAN) to specifically extract different types of features, and use a parameter α to perform feature fusion on the features extracted by different WavKANs. The formula is: x' final = αWavKAN1(x final ) + (1 - α)WavKAN2(x final ) where, is the final output of the feature compression module; α is a learnable parameter that is dynamically adjusted through backpropagation during training; WavKAN1 and WavKAN2 use different wavelet bases respectively.

7. The method for predicting the remaining useful life of an aero-engine based on FEFCNet according to claim 6, wherein, The wavelet basis includes Mexican Hat wavelet and Derivative of Gaussian wavelet.

8. The method for predicting the remaining useful life of an aero-engine based on FEFCNet according to claim 6, wherein The calculation process of the WavKAN includes: Perform scale and translation transformation on the input data. The formula is as follows: where x' represents the data after scale and translation transformation, and are learnable parameters, representing the scale parameter and the translation parameter respectively; Transform x' using the wavelet function to obtain Expand it to through the broadcasting mechanism, and then perform an element-wise multiplication operation with the learnable weight matrix ; sum the resulting values along the d w dimension to obtain the final output of WavKAN 9. The method for predicting the remaining useful life of an aero-engine based on FEFCNet according to claim 1, wherein The RUL mapping module in S5 includes: Two fully connected layers. The first fully connected layer performs a non-linear mapping through the GELU activation function, and the second fully connected layer performs a non-linear mapping through the ReLU activation function to fit the complex non-linear relationship between the degraded features and the remaining life, and ensure that the output life value is non-negative. The specific formula is: RUL pred = ReLU(FC2(GELU(FC1(x′ final )))) Wherein, FC1 and FC2 are the first and second fully connected layers respectively, and GELU and ReLU are activation functions.

Images