A method and system for predicting the remaining useful life of an aero turbofan engine

Abstract

The application provides a kind of aero turbofan engine remaining useful life prediction method and system, method includes: the original sensor data is preprocessed, obtains the training set and test set after pre-processing;The pre-processed data is decomposed into several intrinsic mode functions using CSI-EMD method, and the non-stationary sensor data is effectively decomposed into different IMF components;The IMF component obtained by decomposing the training set data is input into the double-channel multi-layer feature fusion network for training, and the result obtained by the IMF component through the network is added as the final output, to obtain the trained network;The IMF component obtained by decomposing the test set data is input into the trained network to obtain the remaining useful life of the test set data, and the superiority of the evaluation index evaluation method is evaluated.The application solves the technical problem of the influence of traditional EMD method on end effect, poor feature learning effect, which leads to the accuracy of turbofan engine life prediction.

Description

Technical Field

[0001] This invention relates to the field of aero-engine performance prediction, and specifically to a method and system for predicting the remaining service life of an aero-engine turbofan. Background Technology

[0002] Predictive Health Management (PHM) is a promising technology that can ensure the safety and reliability of high-end equipment. Remaining useful life (RUL) prediction is a crucial component of PHM. Accurately predicting a machine's RUL allows for advance planning of appropriate maintenance decisions, eliminating unnecessary maintenance activities, maximizing equipment utilization and reliability, and reducing operating and maintenance costs. Aero-engine turbofan engines are considered the core high-tech machinery of aircraft. Timely and accurate estimation of aero-engine energy consumption levels helps determine scientific maintenance plans and implement predictive maintenance. In recent years, with the development of the Industrial Internet of Things (IIoT), the amount of detection data collected by various sensors has exploded. How to use this massive amount of sensor data to perform degradation modeling and analysis of aero-engine turbofan engines and construct accurate RUL prediction models has become an urgent problem to be solved.

[0003] Current research on life prediction methods for aero-turbofan engines has yielded several results, which can be categorized into three types: physical model-based methods, data-driven methods, and hybrid methods. Hybrid methods mainly fall into two categories: one involves weighted summation of prediction results from different methods to obtain a combined prediction result, which is susceptible to the influence of weight distribution; the other involves preprocessing the original sequence, decomposing it into multiple subsequences, modeling each subsequence separately, and finally superimposing the prediction results of all subsequences to obtain a combined prediction result that meets accuracy requirements. Empirical Mode Decomposition (EMD), as an excellent time-frequency analysis method, has unique advantages in processing non-stationary and nonlinear signals, but the endpoint effect problem inherent in this method can significantly affect the decomposition accuracy. The existing invention patent application document CN113743016A, entitled "A Method for Predicting the Remaining Service Life of a Turbofan Engine Based on an Improved Stacked Sparse Autoencoder and Attention Echo State Network", firstly uses the 3sigma criterion to remove the original noise to obtain high-quality original data and realize data reconstruction. It uses an improved SSAE to extract the features of each cycle of the engine and performs feature dimensionality reduction. In the encoder, BN layer and Dropout layer are used to solve the gradient vanishing and overfitting problems. Then, the extracted engine features are used to construct HI values ​​to obtain HI curves that characterize the engine degradation trend. Finally, an attention mechanism is introduced into the ESN network to adaptively process different types of features and optimize network parameters, and finally obtain the RUL value to realize the prediction of the remaining service life of the turbofan engine. The existing invention patent document CN112580263A, titled "A Method for Predicting the Remaining Service Life of a Turbofan Engine Based on Spatiotemporal Feature Fusion," first acquires raw, time-varying sensor data. After data selection and normalization, feature data is obtained. This feature data is then input into a one-dimensional fully convolutional neural network and an LSTM neural network, respectively. The one-dimensional fully convolutional neural network extracts spatial feature information from the dataset, while the LSTM neural network extracts time-series feature information. After obtaining these two types of features, a spatiotemporal feature fusion algorithm is used to fuse them. The fused features are then input into a max-pooling one-dimensional convolutional neural network for secondary feature extraction, thus achieving the prediction of the remaining service life of the turbofan engine. While the aforementioned deep learning methods have proven to be effective for RUL prediction, traditional deep learning methods typically only extract single-scale features, neglecting important details at other scales and different layers. Furthermore, these methods usually use the same type of model, failing to realize the advantages of using different types of models.

[0004] In summary, existing technologies suffer from poor feature learning performance due to the end-point effect, leading to technical problems in the accuracy of turbofan engine life prediction. Summary of the Invention

[0005] The technical problem to be solved by this invention is how to solve the problem that the existing technology is affected by the endpoint effect, resulting in poor feature learning effect and thus poor accuracy of turbofan engine life prediction.

[0006] This invention solves the above-mentioned technical problems by employing the following technical solution: A method for predicting the remaining service life of an aero-turbofan engine includes:

[0007] S1. Collect sensor data of the aircraft turbofan engine to obtain sliding window data samples to construct an original training set and an original test set. Preprocess the sensor data in the original training set and the original test set to obtain the training set and the test set respectively.

[0008] S2. Using the CSI-EMD method, decompose the data samples to be decomposed in the training set and the test set to obtain no less than two intrinsic mode functions, which can be used as the difference scale features of the sensor data.

[0009] S3. Decompose the training set to obtain no less than two IMF components. Input each IMF component into a dual-channel multi-layer feature fusion network for training to obtain the training results of the IMF components. Add all the training results of the IMF components together as the final output to obtain the applicable network.

[0010] S4. Input the test set into the applicable network to predict the remaining service life of the aircraft turbofan engine, and select an applicable prediction strategy from no less than two preset strategies by evaluating the logic through preset evaluation indicators.

[0011] This invention uses CSI-EMD to alleviate the end-point effect problem of traditional EMD methods. By using a dual-channel multi-layer feature fusion network, features can be learned more effectively, thereby enabling more accurate prediction of the remaining service life of aero-turbofan engines.

[0012] This invention uses the CSI-EMD decomposition method instead of the traditional empirical mode decomposition method, thereby mitigating errors caused by endpoint effects and decomposing more accurate intrinsic mode functions. Simultaneously, a dual-channel multi-layer feature fusion network is used to model each intrinsic mode function separately. This network is an ensemble model composed of convolutional neural networks and recurrent neural networks, which has a greater advantage in extracting the internal features of the signal. Compared with other lifetime prediction methods, the remaining service life prediction results of aero-turbofan engines obtained by this invention based on CSI-EMD and the dual-channel multi-layer feature fusion network have high prediction accuracy and can effectively solve the engine RUL prediction problem.

[0013] In a more specific technical solution, step S1 includes:

[0014] S11. Acquire sensor signal sets using a pre-set sensor group;

[0015] S12. Select an applicable sensor signal from the set of sensor signals;

[0016] S13. Using preset specification logic, perform max-min specification on each of the applicable sensor signals;

[0017] S14. Select an appropriate time window to obtain the sliding window data sample.

[0018] In a more specific technical solution, in step S13, the following logic is used to perform max-min specification on each of the applicable sensor signals:

[0019]

[0020] Where, x i The original data, For the normalized data, x max x min These represent the maximum and minimum values ​​from the same sensor data, respectively.

[0021] In a more specific technical solution, step S2 includes:

[0022] S21. Expand the data sample x(t) to be decomposed using a signal expansion module;

[0023] S22. Use the extreme point determination module to determine all extreme points of the sensor signal y(t), and then fit the upper and lower envelopes accordingly.

[0024] S23. Based on the upper and lower envelopes, calculate the envelope mean line m using the envelope mean line module. i (t);

[0025] S24. Obtain the intrinsic modulus function (IMF) using the intrinsic modulus function acquisition module. i (t);

[0026] S25. End the CSI-EMD method decomposition operation to obtain n intrinsic mode functions and residual components.

[0027] This invention uses the CSI-EMD method to decompose the signal. The preprocessed data is decomposed into several intrinsic mode function (IMF) components using the CSI-EMD method, which alleviates the endpoint effect of the traditional EMD method and reduces the decomposition error.

[0028] In a more specific technical solution, in step S23, the mean value line m of the envelope is calculated based on the upper and lower envelope lines using the following logic. i (t):

[0029]

[0030] Among them, e max (t) is the upper envelope, e min (t) represents the lower envelope.

[0031] In a more specific technical solution, in step S25, the n intrinsic modulus functions and the residual components are obtained using the following logic:

[0032]

[0033] Where r(t) is...

[0034] In a more specific technical solution, step S3 includes:

[0035] S31. The feature extraction layer in the dual-channel multi-layer feature fusion network includes: at least two multi-scale feature extraction blocks. Each multi-scale feature extraction block uses a convolution kernel of different sizes to perform convolution operations on the IMF components to obtain a different time-scale feature map, which is then used to perform an Add operation to obtain a multi-scale feature extraction feature map.

[0036] S32. The second channel of the feature extraction layer in the dual-channel multi-layer feature fusion network includes an echo state network (ESN), which flattens the sample into one-dimensional data and inputs it into the ESN to extract global features.

[0037] This invention constructs a dual-channel, multi-layer feature fusion network to fully extract features at multiple scales from the signal, utilizing two different networks—convolutional neural networks and echo-state networks—to extract local and global features. This improves the accuracy of turbofan engine life prediction.

[0038] In a more specific technical solution, step S4 includes:

[0039] S41. Using the preset evaluation index assessment logic, calculate the strategy index score SCORE:

[0040]

[0041]

[0042]

[0043] In the formula, RUL i It is the actual remaining service life value of the i-th engine under test. is the predicted remaining service life of the i-th engine under test, and n is the number of engines under test. p1 and p2 are the penalty parameters for lag prediction and advance prediction, respectively;

[0044] S42. Obtain the root mean square error using preset logic processing;

[0045] S43. Select the applicable prediction strategy based on the strategy index score SCORE and the root mean square error. In a more specific technical solution, in step S42, the root mean square error RMSE is calculated using the following logic:

[0046]

[0047] In a more specific technical solution, an aircraft turbofan engine remaining service life prediction system includes:

[0048] The training set and test set acquisition module is used to collect sensor data of the aircraft turbofan engine, thereby obtaining sliding window data samples to construct the original training set and the original test set. The sensor data in the original training set and the original test set are preprocessed respectively to obtain the training set and the test set.

[0049] The sample decomposition module is used to decompose the data samples to be decomposed in the training set and the test set using the CSI-EMD method to obtain no less than two intrinsic mode functions as differential scale features of the sensor data. The sample decomposition module is connected to the training set and test set acquisition module.

[0050] A dual-channel multi-layer feature fusion network training module is used to decompose the training set to obtain no less than two IMF components. Each IMF component is input into the dual-channel multi-layer feature fusion network for training to obtain the training result of the IMF component. The training results of all IMF components are added together as the final output to obtain the applicable network. The dual-channel multi-layer feature fusion network training module is connected to the sample decomposition module.

[0051] The lifespan prediction and strategy selection module is used to input the test set into the applicable network to predict the remaining service life of the aero-turbofan engine, and select an applicable prediction strategy from no less than two preset strategies through preset evaluation index evaluation logic. The lifespan prediction and strategy selection module is connected to the training set and test set acquisition module.

[0052] Compared with the prior art, the present invention has the following advantages: The present invention uses CSI-EMD to alleviate the end-point effect problem of the traditional EMD method, and the use of a dual-channel multi-layer feature fusion network can learn features more effectively, thereby enabling more accurate prediction of the remaining service life of aero-turbofan engines.

[0053] This invention uses the CSI-EMD decomposition method instead of the traditional empirical mode decomposition method, thereby mitigating errors caused by endpoint effects and decomposing more accurate intrinsic mode functions. Simultaneously, a dual-channel multi-layer feature fusion network is used to model each intrinsic mode function separately. This network is an ensemble model composed of convolutional neural networks and recurrent neural networks, which has a greater advantage in extracting the internal features of the signal. Compared with other lifetime prediction methods, the remaining service life prediction results of aero-turbofan engines obtained by this invention based on CSI-EMD and the dual-channel multi-layer feature fusion network have high prediction accuracy and can effectively solve the engine RUL prediction problem.

[0054] This invention uses the CSI-EMD method to decompose the signal. The preprocessed data is decomposed into several intrinsic mode function (IMF) components using the CSI-EMD method, which alleviates the endpoint effect of the traditional EMD method and reduces the decomposition error.

[0055] This invention constructs a dual-channel, multi-layer feature fusion network to fully extract features at multiple scales from the signal, utilizing two different networks—convolutional neural networks and echo-state networks—to extract local and global features. This improves the accuracy of turbofan engine life prediction.

[0056] This invention solves the technical problem that the existing technology suffers from poor feature learning performance due to the end-point effect, which leads to poor accuracy in predicting the lifespan of turbofan engines. Attached Figure Description

[0057] Figure 1 This is a schematic diagram of the basic steps of a method for predicting the remaining service life of an aero-turbofan engine according to Embodiment 1 of the present invention.

[0058] Figure 2 This is a schematic diagram illustrating the specific steps of raw data preprocessing in Embodiment 1 of the present invention;

[0059] Figure 3 This is a schematic diagram of sample selection using a time window sliding window in Embodiment 1 of the present invention;

[0060] Figure 4 This is a schematic diagram showing the specific steps of the CSI-EMD method in Embodiment 1 of the present invention.

[0061] Figure 5 This is a schematic diagram of the intrinsic mode functions of the sample signal after CSI-EMD decomposition in Embodiment 1 of the present invention;

[0062] Figure 6 This is a schematic diagram of the hierarchical structure and data processing of the dual-channel multi-layer feature fusion network in Embodiment 1 of the present invention;

[0063] Figure 7This is a prediction chart of the lifespan of all engines in Embodiment 2 of the present invention;

[0064] Figure 8 This is a life prediction trend chart for engine No. 24 in Embodiment 2 of the present invention. Detailed Implementation

[0065] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0066] Example 1

[0067] like Figure 1 As shown, the present invention provides a method for predicting the remaining service life of an aero-turbofan engine, comprising the following steps:

[0068] Step S1: Preprocess the raw sensor data in the training set and test set folders respectively to obtain the preprocessed training set and test set;

[0069] Step S2: Decompose the preprocessed data into several intrinsic mode functions using the CSI-EMD method, which serve as features of the sensor data at different scales.

[0070] Step S3: Input each IMF component obtained by decomposing the training set data into a dual-channel multi-layer feature fusion network for training. Add the results obtained by the network for all components as the final output to obtain the trained network.

[0071] Step S4: Input the test set into the trained network to obtain the remaining lifetime of the data to be tested, and evaluate the superiority of the method through evaluation metrics.

[0072] like Figure 2 As shown, step 1 involves preprocessing the raw data, specifically including the following steps:

[0073] S11. Collect sensor data from turbofan aero engines;

[0074] S12. Select sensor signals suitable for lifetime prediction. In this embodiment, monotonicity, predictiveness and trend calculations are performed on 21-dimensional sensor data to select appropriate sensor data, thereby reducing unnecessary calculations while ensuring accuracy. Finally, the retained sensor numbers are 2, 3, 4, 7, 11, 12, 15, 17, 20 and 21.

[0075] S13. Perform max-min normalization on each selected sensor signal; In this embodiment, the selected sensor signal data is normalized using the max-min normalization method, thereby stretching and compressing all data ranges to [0, 1], eliminating the influence of different scales in the data on the estimation of equipment health status. The specific calculation formula is as follows:

[0076]

[0077] Where x i The original data, For the normalized data, x max x min These represent the maximum and minimum values ​​from the same sensor data, respectively.

[0078] like Figure 3 As shown, in this embodiment, S14, a suitable time window is selected, and data samples are selected using a sliding window. In this embodiment, a time window with a time step of 30 is selected, and the normalized data is selected by sliding the window, as shown. Figure 3 As shown, the length of the time window is the time step. Each time the time window slides forward by one time unit, a new sample is obtained. The remaining lifespan of the last time step of the time window is taken as the label of the training sample to create a sample set.

[0079] like Figure 4 As shown, in this embodiment, step S2, which uses the CSI-EMD method to decompose the preprocessed data, includes a signal augmentation module, an extreme point determination module, an envelope mean line module, an intrinsic modulus function acquisition module, and a decomposition termination module. The specific steps of decomposing the data include:

[0080] S21. Let the original data sample to be decomposed be x(t). As can be seen from the above steps, the data points of x(t) are 30. x(t) is expanded by the signal expansion module. In this embodiment, specifically, three data points are interpolated between every two data points to obtain the expanded data sample to be decomposed y(t), which has 117 data points.

[0081] S22. Use the extreme point determination module to determine all extreme points of the signal y(t), which makes it easier to fit the upper and lower envelopes;

[0082] S23. Calculate the envelope mean line m using the envelope mean line module. i (t):

[0083]

[0084] Where e max (t) is the upper envelope, e min (t) represents the lower envelope;

[0085] like Figure 5 As shown, in this embodiment, S24, the intrinsic modulus function acquisition module obtains several intrinsic modulus functions (IMFs). i (t);

[0086] S25. Use the decomposition termination module to end the entire CSI-EMD method process, ultimately obtaining n intrinsic mode functions and one residual component, i.e.:

[0087]

[0088] In this embodiment, the dual-channel multi-layer feature fusion network in step S3 specifically consists of a feature extraction layer, a feature fusion layer, and a regression layer;

[0089] like Figure 6 As shown, in this embodiment, the feature extraction layer specifically includes:

[0090] S31. The first channel of the feature extraction layer consists of three multi-scale feature extraction blocks. Each layer uses three different sizes of convolution kernels, namely 8x1, 10x1 and 12x1, to perform convolution operations on the input sample data. The number of convolution kernels is 10, resulting in feature maps at three different time scales. The feature maps at the three different time scales are added together using the Add operation to obtain the feature map extracted by the multi-scale feature extraction block.

[0091] S32. The second channel of the feature extraction layer consists of an echo state network (ESN). The sample is flattened into one-dimensional data and input into the ESN to extract global features.

[0092] The feature fusion layer specifically includes:

[0093] S33. Convert the three feature maps obtained by the convolution operation into one-dimensional feature data by passing them through a convolutional layer and a flattening layer respectively.

[0094] S34. Connect the one-dimensional feature data obtained in the previous step and the feature data obtained from ESN through the Concatenate operation to obtain the feature data after feature fusion.

[0095] The regression layer specifically includes:

[0096] S35. The feature fusion data obtained from the feature fusion layer is output through two fully connected operations. In this embodiment, the first fully connected layer has 100 neurons and the neuron dropout rate is set to 20%. The second fully connected layer has one neuron.

[0097] In this embodiment, all activation functions used in the dual-channel multi-layer feature fusion network are "Mish" activation functions, the loss function is set to mean squared error (MSE), the optimizer is the adaptive moment estimation (Adam) optimizer, the learning rate for the first 100 iterations is set to 0.001, and the learning rate for the last 50 iterations is set to 0.0001.

[0098] like Figure 4 As shown, in step 4, when calculating the remaining lifespan of the data to be tested, random errors are mitigated by averaging multiple experiments. The mean and variance (STD), including the root mean square error (RMSE) and the scoring function (SCORE), are used as evaluation indicators to compare the superiority of this method with other methods.

[0099] Example 2

[0100] like Figure 7 and Figure 8 As shown, in this embodiment, the C-MAPSS dataset of turbofan engine monitoring data provided by NASA is used for the experiment. This dataset consists of simulated degradation data from multiple turbofan engines and contains four sub-datasets FD001-FD004, with different operating conditions and failure modes. Each sub-dataset contains a training dataset and a test dataset. FD001 contains life data of 100 engines used for training, and pre-failure data of 100 engines obtained under the same operating conditions for testing. This sub-dataset contains only one failure mode and one condition. Each row in the dataset contains 21 sensor variables with the same operating cycle, and each column contains a single variable with different operating cycles. This embodiment uses the training and test data from FD001 for the experiment.

[0101] Following the steps outlined above, specific experiments were conducted. To evaluate the prediction performance of the proposed method for predicting the remaining service life of aero-turbofan engines based on CSI-EMD and a dual-channel multi-layer feature fusion network, the predicted remaining service life of the test set of the engines under test was analyzed. The actual remaining useful life (RUL) is calculated using the following two evaluation metrics: SCORE and RMSE.

[0102] (1) Scoring function (SCORE):

[0103]

[0104]

[0105]

[0106] (2) Root Mean Square Error (RMSE):

[0107]

[0108] In the above formula, RUL i It is the actual remaining service life value of the i-th engine under test. is the predicted remaining service life of the i-th engine under test, and n is the number of engines under test. p1 and p2 are the penalty parameters for delayed prediction and early prediction, respectively. Since in real industrial scenarios, later maintenance may lead to more serious consequences than earlier maintenance, late prediction is penalized more than early prediction. The penalty parameters are set as follows: p1 = 13, p2 = 10.

[0109] Analysis of experimental results of examples and comparative examples

[0110] To verify the superiority of the remaining service life prediction method for aero-turbofan engines based on CSI-EMD and dual-channel multi-layer feature fusion network of the present invention, the following comparative examples are analyzed.

[0111] Comparative Example 1: The raw data is preprocessed without CSI-EMD decomposition, and then input into a traditional convolutional neural network to predict the remaining service life of the engine.

[0112] Comparative Example 2: The raw data is preprocessed without CSI-EMD decomposition, and then input into the echo state network for prediction of engine remaining service life.

[0113] Comparative Example 3: The raw data is preprocessed without CSI-EMD decomposition, and then input into a dual-channel multi-layer feature fusion network to predict the remaining service life of the engine.

[0114] Comparative Example 4: After preprocessing, the original data is decomposed using the traditional EMD method. Then, each intrinsic mode function is input into a dual-channel multi-layer feature fusion network to predict the remaining service life of the engine.

[0115] Comparative Example 5: After preprocessing, the raw data is decomposed using the CSI-EMD method. Then, each intrinsic mode function is input into a traditional convolutional neural network to predict the remaining service life of the engine.

[0116] A single experiment is subject to a certain degree of randomness and error. Calculating the average of multiple experiments can reduce random errors and ensure the rigor and accuracy of the experiment. The embodiments and comparative examples of this invention were performed five times.

[0117] Table 1. Experimental results of remaining useful life prediction for embodiments and comparative examples of the present invention.

[0118]

[0119] As shown in Table 1, the remaining service life prediction method for aero-turbofan engines based on CSI-EMD and a dual-channel multi-layer feature fusion network of the present invention has the lowest average score (SCORE) and the smallest average root mean square error (RMSE) compared with the aforementioned comparative examples. Since a lower SCORE and a smaller RMSE indicate higher prediction accuracy, the experiments in Table 1 demonstrate that the remaining service life prediction method for aero-turbofan engines based on CSI-EMD and a dual-channel multi-layer feature fusion network of the present invention can significantly improve the prediction performance.

[0120] In summary, this invention uses CSI-EMD to alleviate the end-point effect problem of traditional EMD methods. The use of a dual-channel multi-layer feature fusion network can learn features more effectively, thereby enabling more accurate prediction of the remaining service life of aero-turbofan engines.

[0121] This invention uses the CSI-EMD decomposition method instead of the traditional empirical mode decomposition method, thereby mitigating errors caused by endpoint effects and decomposing more accurate intrinsic mode functions. Simultaneously, a dual-channel multi-layer feature fusion network is used to model each intrinsic mode function separately. This network is an ensemble model composed of convolutional neural networks and recurrent neural networks, which has a greater advantage in extracting the internal features of the signal. Compared with other lifetime prediction methods, the remaining service life prediction results of aero-turbofan engines obtained by this invention based on CSI-EMD and the dual-channel multi-layer feature fusion network have high prediction accuracy and can effectively solve the engine RUL prediction problem.

[0122] This invention uses the CSI-EMD method to decompose the signal. The preprocessed data is decomposed into several intrinsic mode function (IMF) components using the CSI-EMD method, which alleviates the endpoint effect of the traditional EMD method and reduces the decomposition error.

[0123] This invention constructs a dual-channel, multi-layer feature fusion network to fully extract features at multiple scales from the signal, utilizing two different networks—convolutional neural networks and echo-state networks—to extract local and global features. This improves the accuracy of turbofan engine life prediction.

[0124] This invention solves the technical problem that the existing technology suffers from poor feature learning performance due to the end-point effect, which leads to poor accuracy in predicting the lifespan of turbofan engines.

[0125] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting the remaining service life of an aero-turbofan engine, characterized in that, The method includes: S1. Collect sensor data of the aircraft turbofan engine to obtain sliding window data samples to construct an original training set and an original test set. Preprocess the sensor data in the original training set and the original test set to obtain the training set and the test set respectively. S2. Using the CSI-EMD method, decompose the data samples to be decomposed in the training set and the test set to obtain no less than two intrinsic mode functions (IMFs) as differential scale features of the sensor data. S3. Decompose the training set to obtain at least two IMF components. Input each IMF component into a dual-channel multi-layer feature fusion network for training to obtain the training results of the IMF components. Add all the training results of the IMF components together as the final output to obtain the applicable network. S3 includes: S31. The feature extraction layer in the dual-channel multi-layer feature fusion network includes: at least two multi-scale feature extraction blocks. Each multi-scale feature extraction block uses a convolution kernel of different sizes to perform convolution operations on the IMF components to obtain a different time-scale feature map, which is then used to perform an Add operation to obtain a multi-scale feature extraction feature map. S32. The second channel of the feature extraction layer in the dual-channel multi-layer feature fusion network includes: Echo State Network (ESN), which flattens the sample into one-dimensional data and inputs it into the ESN to extract global features; S4. Input the test set into the applicable network to predict the remaining service life of the aircraft turbofan engine, and select an applicable prediction strategy from no less than two preset strategies by evaluating the logic through preset evaluation indicators.

2. The method for predicting the remaining service life of an aero-turbofan engine according to claim 1, characterized in that, Step S1 includes: S11. Acquire sensor signal sets using a pre-set sensor group; S12. Select an applicable sensor signal from the set of sensor signals; S13. Using preset specification logic, perform max-min specification on each of the applicable sensor signals; S14. Select an appropriate time window to obtain the sliding window data sample.

3. The method for predicting the remaining service life of an aero-turbofan engine according to claim 2, characterized in that, In step S13, the following logic is used to perform max-min specification on each applicable sensor signal: in, This is the original data. For the normalized data, These represent the maximum and minimum values ​​from the same sensor data, respectively.

4. The method for predicting the remaining service life of an aero-turbofan engine according to claim 1, characterized in that, Step S2 includes: S21. Expand the signal enhancement module for the data sample to be decomposed. ; S22. Use the extreme point determination module to determine the sensor signal. All extreme points are used to fit the upper and lower envelope lines; S23. Based on the upper and lower envelopes, calculate the envelope mean line using the envelope mean line module. ; S24. Obtain the intrinsic modulo functions using the intrinsic modulo function acquisition module. ; S25. End the CSI-EMD method decomposition operation and obtain... Each intrinsic modulus function and residual component.

5. The method for predicting the remaining service life of an aero-turbofan engine according to claim 4, characterized in that, In step S23, the mean value line of the envelope is calculated based on the upper and lower envelope lines using the following logic. : in, The upper envelope, This is the lower envelope.

6. The method for predicting the remaining service life of an aero-turbofan engine according to claim 4, characterized in that, In step S25, the following logic is used to obtain... The intrinsic modulo functions and the residual components: 。 7. The method for predicting the remaining service life of an aero-turbofan engine according to claim 1, characterized in that, Step S4 includes: S41. Using the preset evaluation index assessment logic, calculate the strategy index score SCORE: In the formula, It is the first The actual remaining service life of the engine under test. It is the first Predicted remaining service life of each engine under test The number of engines to be tested. These are the penalty parameters for lagging forecasts and leading forecasts, respectively; S42. Obtain the root mean square error using preset logic processing; S43. Select the applicable prediction strategy based on the strategy index score SCORE and the root mean square error.

8. The method for predicting the remaining service life of an aero-turbofan engine according to claim 7, characterized in that, In step S42, the root mean square error (RMSE) is calculated using the following logic: 。 9. A system for predicting the remaining service life of an aircraft turbofan engine, used to execute the method for predicting the remaining service life of an aircraft turbofan engine as described in any one of claims 1 to 8, characterized in that, The system includes: The training set and test set acquisition module is used to collect sensor data of the aircraft turbofan engine, thereby obtaining sliding window data samples to construct the original training set and the original test set. The sensor data in the original training set and the original test set are preprocessed respectively to obtain the training set and the test set. The sample decomposition module is used to decompose the data samples to be decomposed in the training set and the test set using the CSI-EMD method, so as to obtain no less than two intrinsic mode functions (IMFs) as differential scale features of the sensor data. The sample decomposition module is connected to the training set and test set acquisition module. A dual-channel multi-layer feature fusion network training module is used to decompose the training set to obtain no less than two IMF components. Each IMF component is input into the dual-channel multi-layer feature fusion network for training to obtain the training result of the IMF component. The training results of all IMF components are added together as the final output. The parameters are trained through backpropagation to obtain the applicable network. The dual-channel multi-layer feature fusion network training module is connected to the sample decomposition module. The lifespan prediction and strategy selection module is used to input the test set into the applicable network to predict the remaining service life of the aero-turbofan engine, and select an applicable prediction strategy from no less than two preset strategies through preset evaluation index evaluation logic. The lifespan prediction and strategy selection module is connected to the training set and test set acquisition module.

Images