FEFCNet-based aero-engine residual life prediction method

Abstract

The invention relates to an FEFCNet-based aero-engine residual life prediction method, and belongs to the field of residual life prediction of mechanical equipment. The method comprises the following steps: acquiring aero-engine sensor data, and performing clustering, normalization, sliding window processing and residual life label setting to obtain a training set and a verification set; then, extracting significant frequency components of the sample data by using fast Fourier transform, calculating a corresponding period length, and reconstructing data dimensions to obtain dimension-raised sample data; then, the local details and the global features are fused through an embedded layer, a multi-scale expansion convolution layer, an Effient Multi-Scale Attention (EMA) module and a multi-head self-attention mechanism, and multi-scale fusion features are obtained; the method comprises the following steps of: further performing feature compression by using Wavelet Kolmogorov-Arnold Networks (WavelKAN), and further performing feature compression by using Wavelet Kolmogorov-Arnold Networks (WavelKAN); and finally, mapping the compression characteristics into residual life through a full connection layer. The method can effectively improve the prediction accuracy of the residual life of the aero-engine, provides a reliable basis for timely overhaul and maintenance, and reduces the operation and maintenance risk.

Description

Technical Field

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

[0002] Predictive and Health Management (PHM) has been playing an important role in improving operating efficiency, system reliability and performance. Remaining Useful Life (RUL) prediction, as an important part of PHM, 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 resource waste.

[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 complex system modeling, 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 because of 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 existing technology, the present invention provides a method for predicting the remaining life of an aero-engine based on FEFCNet, which can effectively improve the accuracy of predicting the remaining life of an aero-engine, so as to timely repair and maintain the aero-engine and extend its service life. To achieve the above object, the technical solution adopted by the present invention is as follows: A method for predicting the remaining life of an aero-engine based on FEFCNet, comprising 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 (FFT), extract several frequency components with the largest corresponding amplitudes, calculate the corresponding period lengths, and then reconstruct the data dimension to obtain the upsampled sample data. S3: Extract the degradation features of the upsampled data through the feature extraction module. S4: Reduce the dimensionality of the degradation 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.

[0006] Further, the data preprocessing steps in S1 include: Use the K-Means algorithm to cluster the data according to the operating conditions to obtain the operating data of the aero-engine under different operating conditions, and perform Min-Max normalization on the operating data of the aero-engine under each operating condition. The formula is as follows: where is the original sensor data under the k-th operating condition, is the data after standardization under the k-th operating condition, k represents the k-th operating condition, i represents the i-th sensor, and j represents the j-th data point. , are the maximum and minimum values of the data of the i-th sensor under the k-th operating condition, respectively.

[0007] Generate two-dimensional time window data using the sliding window method, and obtain multiple two-dimensional time window data according to the set sliding window size and sliding step. , where represents the set of real numbers, represents rows and columns of the two-dimensional matrix, where represents the sliding window size, and

[0008] Set the threshold of the RUL label and set the RUL label for the two-dimensional time window data; specifically, use the RUL corresponding to the last row data of the two-dimensional time window data as its RUL.

[0009] 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.

[0010] Further, the method in S2 includes: Process the two-dimensional time window data using the Fast Fourier Transform, and perform the Fast Fourier Transform on each two-dimensional time window data along the time dimension. The formula is as follows: Among them, represents the corresponding frequency spectrum, represents the frequency, represents the sliding window size of the two-dimensional time window data, represents the data of the two-dimensional time window data in the th row.

[0011] Calculate the amplitude corresponding to each frequency and obtain the average amplitude by taking the average in the feature dimension , and its formula is as follows: Among them, represents the number of features of the two-dimensional time window data, represents the frequency spectrum of the nth feature.

[0012] Select the top k significant frequencies with the largest average amplitude , for each significant frequency , calculate the corresponding period length . For each candidate period , split the time dimension of the original time series into two dimensions to obtain k three-dimensional data after dimension elevation , where .

[0013] Furthermore, the degraded features of the data after dimension elevation extracted by the degraded feature extraction module described in S3 include: Pass the k three-dimensional data after dimension elevation through the two-dimensional convolutional embedding layer respectively, and map the last dimension to a fixed dimension to obtain new k three-dimensional data , where ; Pass the new k 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, and the convolutional layer of each branch uses different convolutional kernel sizes and dilation rates to achieve the extraction of features of different scales, and the feature maps obtained by each branch are made to be the same size as the original data by setting the padding parameters; the first sub-convolutional layer performs non-linear transformation through the GELU activation function.

[0014] For each three-dimensional data, after passing through each multi-scale dilated convolutional sub-layer, 7 feature maps are obtained , where , 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 feature dimension (the last dimension) and averaged along the feature dimension to obtain a fused feature map; since there are k data, k fused feature maps are finally obtained , where .

[0015] Enhance the weights of key time steps and feature channels through the EMA module, suppress noise interference, and obtain k enhanced fused feature maps , where ; then merge the first two dimensions of each enhanced fused feature map, concatenate all groups of features along the feature dimension, average the concatenated result along the feature dimension, and then map the size of the second dimension to , and finally obtain the overall enhanced fused feature map ; Process the two-dimensional time window data obtained in step S1 through the multi-head self-attention mechanism , capture its global time series dependence; and add it to the overall enhanced fused feature map to obtain the final output of the degradation feature extraction module , make up for the locality limitation of the convolution operation, and realize the deep fusion of local and global degradation features. The formula is as follows: where, represents the multi-head self-attention mechanism module, represents the final output of the degradation feature extraction module.

[0016] Furthermore, the feature compression module described in S4 is used to perform dimensionality reduction and compression on the degradation features to obtain compressed features, including: Use 2 parallel WavKANs to compress the fused feature maps obtained in S3 in the time dimension; each WavKAN uses different types of wavelet bases to extract different types of features specifically; through a parameter to perform feature fusion on the features extracted by different WavKANs.

[0017] The different types of wavelet bases include Mexican Hat wavelet and Derivative of Gaussian wavelet; local smooth features and edge mutation features are extracted through the Mexican Hat wavelet and Derivative of Gaussian wavelet respectively to cover the diversity of degradation patterns.

[0018] Furthermore, the WavKAN calculation process includes: Perform scaling and translation transformations on the input data, and the formula is as follows: Where represents the data after scaling and translation transformations, and are learnable parameters, representing the scaling parameter and the translation parameter respectively.

[0019] Use the wavelet basis function to perform a transformation to obtain , and expand it to through the broadcasting mechanism, then perform an element-wise multiplication operation with the learnable weight matrix , and sum the obtained results in the dimension to obtain the final output of WavKAN .

[0020] Perform feature fusion on the features extracted by different WavKANs, and use to automatically optimize the weights of different wavelet branches to enhance the robustness of feature expression. The formula is as follows: Where, and are WavKANs using different wavelet bases respectively, is a learnable parameter (dynamically adjusted through backpropagation during training), is the final output of the feature compression layer.

[0021] Furthermore, the RUL mapping module described 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 degradation features and the remaining useful life, and ensure that the output life value is non-negative.

[0022] Furthermore, the mapping of the obtained compressed feature data to RUL described in S5 includes: Map the compressed features obtained in S4 to obtain the RUL prediction value. The formula is as follows: Where, and are the first and second fully connected layers respectively, and are activation functions. Description of the Drawings

[0023] Figure 1 Is the flowchart of the present invention Figure 2 It is the structure diagram of FEFCNet Figure 3 It is the structure diagram of multi-scale dilated convolution Figure 4 It is the RUL prediction result diagram of the FD003 test set Figure 5 It is the RUL prediction result diagram of the FD004 test set Specific implementation manners 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.

[0024] 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: S1: Obtain a specific subset (FD001 / FD002 / FD003 / FD004) in the NASA-CMAPSS dataset. First, filter out the channels with constant measurement values among the 21 sensors (finally retaining 14 effective sensors). For the FD002 and FD004 subsets (including 6 operating 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 operating 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: Among them, is the original sensor data under the kth operating condition, is the data after standardization under the kth operating condition. k represents the kth operating condition, i represents the ith sensor, and j represents the jth data point. , are the maximum and minimum values of the ith sensor data under the kth operating condition respectively.

[0025] Concatenate the 14-column sensor measurement values after Min-Max normalization with the three columns of data representing the operating conditions to obtain 17-column aero-engine operating state data.

[0026] Use the sliding window technique to process the 17-column aero-engine operating state data. Set the sliding window size as the hyperparameter W and the sliding step as 1 to obtain multiple two-dimensional time window data , where Denotes the sliding window size, Represents the number of features, with a size of 17.

[0027] Set the RUL threshold to 125 according to the practices of most studies, and use the RUL corresponding to the last row of data for each two-dimensional time window data as the RUL.

[0028] Randomly divide the training set into a training set and a validation set according to an 8:2 ratio.

[0029] S2: Process the two-dimensional time window data using the fast Fourier transform. Perform the fast Fourier transform on each two-dimensional time window data along the time dimension. The formula is as follows: where, denotes the corresponding spectrum, denotes the frequency, denotes the size of the time window of the two-dimensional time window data, denotes the data of the two-dimensional time window data in the row.

[0030] Calculate the amplitude corresponding to each frequency and obtain the average amplitude by taking the average in the feature dimension , and the formula is as follows: where, denotes the number of features of the two-dimensional time window data, denotes the spectrum of the nth feature.

[0031] Select the top significant frequencies with the largest average amplitude . For each significant frequency , calculate the corresponding period length . For each candidate period , split the time dimension into two dimensions to obtain k three-dimensional data after dimensionality increase , where .

[0032] S3: Pass the k three-dimensional data after dimensionality increase through a two-dimensional convolutional layer respectively, and map the last dimension to a fixed dimension size to obtain k new three-dimensional data , where .

[0033] The new k three-dimensional data are used to initially extract the degradation features of aero-engine operation data through a multi-scale dilated convolutional layer. The multi-scale dilated convolutional layer is as shown in Figure 3 , and it includes 2 multi-scale dilated convolutional sub-layers. Each sub-layer includes 7 parallel feature extraction branches. The convolutional layer of each branch uses different convolutional kernel sizes and dilation rates to extract features of different scales, and padding is used to make the feature maps obtained by each branch have the same size as the original data; the first sub-convolutional layer performs non-linear transformation through the GELU activation function.

[0034] For each three-dimensional data, after passing through the multi-scale dilated convolutional layer, 7 feature maps are obtained , where , 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 , where .

[0035] The weights of key time steps and feature channels are enhanced through the EMA module to suppress noise interference, and k enhanced fused feature maps are obtained , where ; then the first two dimensions of each enhanced fused feature map are merged, all groups of features are concatenated along the feature dimension, and the concatenated result is averaged along the feature dimension; then the size of the second dimension is mapped to through linear projection, and finally the overall enhanced fused feature map is obtained; The two-dimensional time window data obtained by processing S1 through the multi-head self-attention mechanism is added to the overall enhanced fused feature map to obtain the final output of the degradation feature extraction module, and its formula is as follows: Among them, represents the multi-head self-attention mechanism module, represents the final output of the degradation feature extraction module.

[0036] S4: Use 2 parallel WavKANs to compress the enhanced fused feature maps in the time dimension. The two WavKANs use the Derivative of Gaussian wavelet and the Mexican Hat wavelet as wavelet bases respectively. The formulas of the corresponding two wavelets are as follows: The specific operation of each WavKAN is as follows: Perform scaling and translation transformation on the input data, and its formula is as follows: Where represents the data after scaling and translation transformation, and are learnable parameters, representing the scale parameter and the translation parameter respectively.

[0037] Expand to through the broadcasting mechanism and then perform element-wise multiplication operation with the learnable weight matrix , and sum the obtained results in the dimension to obtain the final output of WavKAN .

[0038] Perform feature fusion on the features extracted by different WavKANs to obtain compressed features , and its formula is as follows: Where, and are WavKANs using the Derivative of Gaussian wavelet and the Mexican Hat wavelet as the wavelet basis respectively, is a learnable parameter (dynamically adjusted through backpropagation during training), is the final output of the feature compression layer.

[0039] 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; After passing through the two fully connected layers, the RUL prediction value is obtained, and its formula is as follows: Where, and represent the two fully connected layers respectively, and are activation functions.

[0040] 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.

[0041] Based on the above method framework, remaining useful life prediction modeling is implemented for the FD001 - FD004 subsets, where hyperparameter optimization is completed using the grid search algorithm: within the preset hyperparameter space (learning rate , batch size , sliding window size , number of multi - head self - attention heads , embedding dimension , the first k significant frequency components ), parameter optimization is carried out with the goal of minimizing the MSE of the validation set. Model performance evaluation uses the root mean square error (RMSE) and score (Score) as evaluation metrics. The calculation formulas for RMSE and Score are as follows: where n is the number of test sample data, is the predicted remaining life value of the i - th sample, is the true remaining life value of the i - th sample.

[0042] Figure 4 and Figure 5 are the RUL prediction result graphs of the FD003 and FD004 datasets respectively. It can be seen from the figure that for engines with a relatively small true remaining 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 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 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.

[0043] By comparing the prediction performance 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). Specifically, RMSE: The average RMSE of FEFCNet is 13.81, a 6.6% decrease compared to GA-Transformer (14.78). Especially in the complex working condition dataset FD004, the RMSE is 16.64, a 17.4% increase compared to GA-Transformer (20.15), verifying its strong adaptability to multi-working condition scenarios. Score: The average Score of FEFCNet is 742.58, at least 28.7% lower than other methods (1047 - 1265). Among them, the Score of the FD004 dataset (1406.54) is significantly optimized by 50.7% compared to DSSN (2852.81), indicating that its prediction deviation is smaller and the timeliness of maintenance decisions is stronger.

[0044] Table 1 Comparison of prediction performance of different methods on four sub-datasets of C-MAPSS 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 for 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 life of an aircraft engine based on FEFCNet, characterized in that: The following steps are involved: S1: Obtain and preprocess the sensor data of the aircraft engine to determine the training set and validation set; S2: Use fast Fourier transform to process the preprocessed data, extract several frequency components with the largest amplitude and calculate the corresponding period length, then reconstruct the data dimension to obtain the sample data after dimensionality increase; S3: extracting the degraded features of the data after dimensionality increase through the feature extraction module; S4: The degraded features are compressed by reducing the dimension through the feature compression module to obtain compressed features; S5: Map the compressed features to the remaining useful life value through the RUL mapping module and output the prediction result.

2. The method for predicting the remaining life of an aircraft engine based on FEFCNet according to claim 1, characterized in that: The preprocessing in S1 includes: The K-Means algorithm is used to cluster the data according to the working conditions, and Min-Max normalization is performed on each type of working condition data; The sliding window method is used to generate the two-dimensional time window data of the aircraft engine operation status. Each data sample ,in represents the set of real numbers, represent OK A two-dimensional matrix of columns, represents the sliding window size, represents the number of features; Set the threshold of RUL and set the RUL label 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; The obtained labeled two-dimensional time window data is randomly divided into training set and validation set according to the ratio of 8:

2.

3. The method for predicting the remaining life of an aircraft engine based on FEFCNet according to claim 1, characterized in that: The S2 includes: The two-dimensional time window data is processed by fast Fourier transform. The fast Fourier transform is performed along the time dimension, and the formula is as follows: in, express The corresponding spectrum, Indicates frequency, Indicates the sliding window size of the two-dimensional time window data, Indicates that the two-dimensional time window data is Row data; Calculate the amplitude corresponding to each frequency and average it in the feature dimension to get the average amplitude , the formula is as follows: in, represents the number of features of the two-dimensional time window data, represents the spectrum of the nth feature; Select the first k frequency components with the largest amplitude , respectively calculate the corresponding period length ; Split the time dimension by frequency and cycle length to generate k three-dimensional data ,in .

4. The method for predicting the remaining life of an aircraft engine based on FEFCNet according to claim 1, characterized in that: The feature extraction module in S3 includes: The embedding layer is a two-dimensional convolutional layer that performs dimension mapping on the data and maps the feature dimension of the input data to a fixed dimension. ; The multi-scale dilated convolution layer includes two multi-scale dilated convolution sub-layers, each of which includes seven parallel feature extraction branches. The convolution layer of each branch uses different convolution kernel sizes and dilation rates to extract features of different scales. The feature map obtained by each branch is made consistent with the original data size by setting the padding parameters. The first sub-layer is nonlinearly transformed by the GELU activation function. The parallel convolution branches extract local degradation features of different scales. Efficient Multi-Scale Attention (EMA) module, which is used to fuse spatiotemporal features, enhance the weights of key time steps and feature channels, and suppress noise interference; Multi-head self-attention mechanism module, used to process the two-dimensional time window data obtained in step S1 , capturing its global time series dependencies and making up for the local limitations of convolution operations.

5. The method for predicting the remaining life of an aircraft engine based on FEFCNet according to claim 4, characterized in that: The degradation feature extraction module implements feature extraction through the following steps: The three-dimensional data obtained in step S2 is embedded in the Perform dimension mapping to obtain new k three-dimensional data ,in ; The new k 3D data are passed through a multi-scale dilated convolutional layer to extract local degradation features of different scales; For each three-dimensional data, after passing through each multi-scale dilated convolution sublayer, 7 feature maps are obtained ,in , Represents the feature map obtained after the k-th three-dimensional data passes through the i-th branch; the 7 feature maps obtained are spliced ​​along the feature dimension (the last dimension) and averaged in the feature dimension to obtain a fused feature map; since there are k data, k fused feature maps are finally obtained ,in ; The EMA module is used to enhance the weights of key time steps and feature channels, suppress noise interference, and obtain k enhanced fusion feature maps. ,in ; Then merge the first two dimensions of each enhanced fusion feature map, splice all group features along the feature dimension, average the splicing results in the feature dimension, and then use linear projection to reduce the size of the second dimension Map to , 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 , capturing its global time series dependencies and fusion feature maps with the overall enhancement Add together to get the final output of the degradation feature extraction module , the formula is as follows: in, Represents the multi-head self-attention mechanism module, Represents the final output of the degradation feature extraction module.

6. The method for predicting the remaining life of an aircraft engine based on FEFCNet according to claim 1, characterized in that: The feature compression module in S4 includes: Two parallel WaveletKolmogorov-ArnoldNetworks (WavKAN) are used to extract different types of features in a targeted manner, through a parameter To fuse the features extracted by different WavKANs, the formula is: in, is the final output of the feature compression module, α is a learnable parameter (dynamically adjusted through back propagation during training), and Use different wavelet bases respectively.

7. The method for predicting the remaining life of an aircraft engine based on FEFCNet according to claim 6, characterized in that: The wavelet bases include Mexican Hat wavelet and Derivative of Gaussian wavelet.

8. The method for predicting the remaining life of an aircraft engine based on FEFCNet according to claim 6, characterized in that: The calculation process of WavKAN includes: The scale and translation transformation of the input data is performed as follows: in Represents the data after scale and translation transformation, and are learnable parameters, representing scale parameters and translation parameters respectively; Using wavelet function Transform and get ; Expanded to After that, with the learnable weight matrix Perform element-by-element multiplication; the result is The final output of WavKAN is obtained by summing the dimensions. .

9. The method for predicting the remaining life of an aircraft engine based on FEFCNet according to claim 1, characterized in that: The RUL mapping module in S5 includes: Two fully connected layers, the first fully connected layer performs nonlinear mapping through the GELU activation function, and the second fully connected layer performs nonlinear mapping through the ReLU activation function to fit the complex nonlinear relationship between degradation characteristics and remaining life, and ensure that the output life value is non-negative; the specific formula is: in, and are the first and second fully connected layers, and is the activation function.

Images