Tmsa-tcn-based residual useful life prediction method for turbofan engine

Abstract

The application provides a TMSA-TCN-based turbofan engine residual service life prediction method and relates to the technical field of aero-engines, and comprises the following steps: collecting historical operation monitoring data of a turbofan engine and performing pretreatment, dividing the pretreated historical operation monitoring data into a training set and a test set; embedding a time multi-scale attention module TMSA into a residual block of a time convolutional neural network TCN to construct a TMSA-TCN prediction model, inputting the training set into the TMSA-TCN prediction model to perform training, and obtaining a trained TMSA-TCN prediction model; inputting the test set into the trained TMSA-TCN prediction model to output a turbofan engine residual service life prediction result. The application introduces a multi-scale attention module TMSA and improves a time convolutional neural network TCN to improve the reliability and accuracy of the residual life prediction of an aero turbofan engine.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of aero-engines, in particular to a turbofan engine residual service life prediction method based on TMSA-TCN. BACKGROUND

[0002] The turbofan engine, also known as a turbofan engine, is one of the main types of aero-engines. As the core power component of an aircraft, the residual service life prediction of the turbofan engine has an extremely important influence on personnel safety, flight safety, and economic benefits. The residual life prediction of the turbofan engine and the accuracy of the prediction are crucial. In the past, the life prediction of the aero-engine mainly relied on experience and regular maintenance, but it was often difficult to ensure the timeliness and accuracy of the prediction. At present, the residual life prediction of the aero-engine mainly includes two categories based on data driving and physical modeling. However, due to the complexity of the structure system of the turbofan engine, it is difficult to ensure the accuracy of the modeling, thereby affecting the accuracy of the engine life prediction.

[0003] The traditional data-driven method has problems such as insufficient feature value extraction and insufficient prediction accuracy. The existing technology proposes a prediction method combining random forest RF, time convolution TCN and Transformer encoder to improve the accuracy of the residual life prediction of the turbofan engine and verifies it on the C-MAPSS data set. The traditional method also has the problem of difficulty in capturing local features simultaneously. The existing technology also proposes a new hybrid model based on parallel TCN and BiLSTM. This model extracts short-term local features through TCN and uses BiLSTM to capture the bidirectional time sequence dependence in the data. This method has high accuracy. To solve the problem of insufficient feature value extraction in prediction, the existing technology uses Spearman rank correlation coefficient (SRCC), Bayesian (Bayesian) and k-nearest neighbor (KNN) algorithm to propose a new method for predicting the residual service life of the aero-engine. This method effectively improves the accuracy of the engine residual life prediction. However, in the field of residual life prediction of the turbofan engine, the prediction accuracy is not high, the timeliness is not strong, and due to the uncertainty of the engine life and fault distribution, the complexity of the engine system and other factors, it is difficult to accurately predict the residual life of the engine. Therefore, it is of great significance to find an effective and high-precision prediction technology method for the residual life prediction of the turbofan engine. SUMMARY

[0004] Therefore, the present application provides a turbofan engine residual service life prediction method based on TMSA-TCN, which introduces a multi-scale attention module TMSA and improves the time convolution neural network TCN to improve the reliability and accuracy of the residual life prediction of the aero turbofan engine.

[0005] The embodiment of the application provides the following technical scheme: a turbofan engine residual service life prediction method based on a TMSA-TCN, comprising:

[0006] Historical operation monitoring data of the turbofan engine are collected and preprocessed, and the preprocessed historical operation monitoring data are divided into a training set and a test set;

[0007] A time multi-scale attention module TMSA is embedded into a residual block of a time convolutional neural network TCN to construct a TMSA-TCN prediction model, the training set is input into the TMSA-TCN prediction model for training, and a trained TMSA-TCN prediction model is obtained;

[0008] The test set is input into the trained TMSA-TCN prediction model, and a turbofan engine residual service life prediction result is output.

[0009] According to an embodiment of the application, the historical operation monitoring data of the turbofan engine are collected and preprocessed, comprising:

[0010] A hybrid resampling method based on density-based oversampling and undersampling is used to preprocess the historical operation monitoring data, so that the imbalance ratio IR of the number of majority class samples and the number of minority class samples is 1, a balanced training set is obtained, the balanced training set is input into the TMSA-TCN prediction model for training, and a trained TMSA-TCN prediction model is obtained.

[0011] According to an embodiment of the application, the hybrid resampling method comprises: in the oversampling stage, a K-means algorithm is used to cluster the minority class samples, new samples are synthesized by using a probability selection method according to sample density, and the imbalance ratio IR of the number of majority class samples and the number of minority class samples is 2; in the undersampling stage, a K-means algorithm is used to cluster the majority class samples, low-density samples are deleted by using a probability selection method according to sample density, and the imbalance ratio IR of the number of majority class samples and the number of minority class samples is 1.

[0012] According to an embodiment of the application, the time multi-scale attention module TMSA is composed of a multi-scale convolution layer and a multi-head attention layer; the multi-scale convolution layer extracts features on different time scales in parallel by connecting multiple one-dimensional convolution layers with different convolution kernel sizes; and the multi-head attention layer performs weighted fusion on multi-scale features by using local attention and global attention mechanisms to enhance the representation of key features.

[0013] According to an embodiment of the application, the time convolutional neural network TCN adopts a network structure comprising a causal convolution, an expansion convolution and a residual connection.

[0014] According to an embodiment of the present application, the time multi-scale attention module TMSA is embedded after a ReLU activation layer of the time convolutional neural network TCN.

[0015] According to an embodiment of the present application, after the training set is input into the TMSA-TCN prediction model for training, the method further comprises: training the TMSA-TCN prediction model based on a Pytorch deep learning framework, and taking a cross-entropy loss function as an optimization objective.

[0016] According to an embodiment of the present application, the method further comprises: analyzing and evaluating the accuracy, F-measure value and G-mean value of the remaining useful life prediction result of the turbofan engine.

[0017] Compared with the prior art, the above at least one technical solution adopted by the embodiments of the present application can achieve at least the following beneficial effects: the embodiments of the present application propose a turbofan engine remaining life prediction method based on a TMSA-TCN neural network, the method improves the TCN neural network, strengthens the expression of key features by introducing a TMSA time multi-scale attention module, weakens the representation of low-quality features, improves the recognition ability of the model for fault features, improves the anti-information interference ability of the TCN network, and enhances the reliability and credibility of the turbofan engine life prediction. The prediction method has the advantages of high prediction accuracy and accurate and reliable prediction. BRIEF DESCRIPTION OF DRAWINGS

[0018] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0019] Figure 1 is a flowchart of the turbofan engine remaining useful life prediction method based on TMSA-TCN of the present application;

[0020] Figure 2 is a first flowchart of the turbofan engine remaining life prediction method based on TMSA-TCN of the embodiments of the present application;

[0021] Figure 3 is a flowchart of engine monitoring data preprocessing of the embodiments of the present application;

[0022] Figure 4 is a schematic diagram of the dilated convolutional network of the embodiments of the present application;

[0023] Figure 5 is a schematic diagram of the TMSA-TCN network structure of the embodiments of the present application;

[0024] Figure 6 is a second flowchart of a TMSA-TCN-based turbofan engine residual life prediction embodiment of the present application. DETAILED DESCRIPTION

[0025] The embodiments of the present application will be described in detail below with reference to the drawings.

[0026] The advantages and effects of the present application can be easily understood by those skilled in the art from the content disclosed in the specification. Obviously, the described embodiments are only a part of the embodiments of the present application, not all. The present application can also be implemented or applied by other different specific embodiments, and the details in the specification can be modified or changed based on different views and applications without departing from the spirit of the present application. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0027] As shown in Figure 1 , the present application provides a TMSA-TCN-based turbofan engine residual life prediction method, comprising:

[0028] S101. Collecting historical operation monitoring data of a turbofan engine and preprocessing, dividing the preprocessed historical operation monitoring data into a training set and a test set;

[0029] S102. Embedding a time multi-scale attention module TMSA into a residual block of a time convolutional neural network TCN to construct a TMSA-TCN prediction model, inputting the training set into the TMSA-TCN prediction model for training to obtain a trained TMSA-TCN prediction model;

[0030] S103. Inputting the test set into the trained TMSA-TCN prediction model to output a turbofan engine residual life prediction result.

[0031] In an embodiment of the present application, collecting historical operation monitoring data of a turbofan engine and preprocessing comprises: preprocessing the historical operation monitoring data by using a hybrid resampling method based on density-based oversampling and undersampling to make the imbalance ratio IR of the number of majority class samples and the number of minority class samples to be 1, obtaining a balanced training set, inputting the balanced training set into the TMSA-TCN prediction model for training to obtain a trained TMSA-TCN prediction model.

[0032] In specific implementation, the hybrid resampling method includes: in the oversampling stage, using the K-means algorithm to cluster the minority class samples, and using a probabilistic selection method to synthesize new samples according to the sample density, until the imbalance ratio (IR) between the number of majority class samples and the number of minority class samples is 2; in the undersampling stage, using the K-means algorithm to cluster the majority class samples, and using a probabilistic selection method to delete low-density samples according to the sample density, until the imbalance ratio (IR) between the number of majority class samples and the number of minority class samples is 1.

[0033] In one embodiment of the present invention, after the training set is input into the TMSA-TCN prediction model for training, the method further includes training the TMSA-TCN prediction model based on the PyTorch deep learning framework, using the cross-entropy loss function as the optimization objective.

[0034] In one embodiment of the present invention, the method further includes: analyzing and evaluating the accuracy, F-measure value, and G-mean value of the predicted remaining service life of the turbofan engine.

[0035] To address the issues of low accuracy, long cycle times, and generally low precision in predicting the remaining service life of turbofan engines due to complex engine system modeling, this invention proposes a TMSA-TCN method for predicting the remaining service life of turbofan engines. First, data acquisition and preprocessing are performed on the turbofan engine. Historical engine operation monitoring data is collected and preprocessed using resampling techniques. Imbalanced datasets collected from the historical engine operation monitoring data are preprocessed to improve the accuracy of life prediction. The preprocessing method used in this invention is resampling, which is divided into two categories: density-based resampling and imbalance-based resampling. Then, the Temporal Multi-Scale Attention Module (TMSA) and Temporal Convolutional Neural Network (TCN) are combined and introduced into the life prediction framework to achieve aero-engine life prediction. The collected and preprocessed data is input into the TMSA-TCN network in a matrix format for training. The TMSA-TCN model is then trained using the PyTorch deep learning framework to output the prediction results.

[0036] This invention proposes a network that combines a multi-scale attention module (TMAS) with a convolutional neural network (TCN). Due to the misleading nature of various interference information, the accuracy of the TCN prediction model can be affected. Therefore, this invention embeds a multi-scale attention module (TMAS) into the TCN, utilizing the attention mechanism to strengthen the feature representation of important information and weaken the effect of low-quality features, thereby constructing a new prediction model, TMSA-TCN. This model, by embedding the multi-scale attention module (TMAS) and improving the temporal convolutional neural network, enhances the accuracy of remaining lifetime prediction for aero-engine turbofan engines.

[0037] In specific implementation, the Temporal Multi-Scale Attention Module (TMSA) consists of multi-scale convolutional layers and multi-head attention layers. The multi-scale convolutional layers extract features at different time scales in parallel by connecting multiple one-dimensional convolutional layers with different kernel sizes. The multi-head attention layers perform weighted fusion of multi-scale features through local attention and global attention mechanisms to enhance the representation of key features.

[0038] In specific implementation, the Temporal Convolutional Neural Network (TCN) adopts a network structure that includes causal convolution, dilated convolution, and residual connections.

[0039] In a preferred embodiment, the temporal multi-scale attention module (TMSA) is embedded after the ReLU activation layer of the temporal convolutional neural network (TCN).

[0040] In one embodiment of the present invention, a method for predicting the remaining life of a turbofan engine based on TMSA-TCN is proposed. This method improves the accuracy of remaining life prediction for aero-turbofan engines by introducing a multi-scale attention module (TMSA) and an improved temporal convolutional neural network (TCN). The method mainly includes three stages: data acquisition and preprocessing, model training, and result evaluation. The method is as follows: Figure 2 As shown.

[0041] (1) Data acquisition and preprocessing: Due to the limitations of engine sensors and the influence of environmental interference, engine monitoring data contains some redundant features and noise, which will increase the burden during the training of the prediction model and affect the accuracy of the model's life prediction. In addition, the collected engine monitoring data contains extremely unbalanced data distribution, which will also reduce the prediction accuracy of the life prediction model. Therefore, the following methods are used: Figure 3 The process shown preprocesses the data.

[0042] This method is a hybrid resampling approach, which includes undersampling and oversampling. It uses the K-means algorithm to cluster the samples and calculates the density of different sample regions. To optimize the dataset, redundant majority class samples are selectively removed. Since samples in dense regions contain more information than samples in low-density regions, they are more likely to be selected.

[0043] In the oversampling stage, suitable minority samples are found through search, and new synthetic samples are synthesized using the found minority samples. Then, the K-means algorithm and the distance between the sample and the mean are used to determine suitable initial cluster centers, thereby improving the quality of the algorithm results. The specific calculation is as follows.

[0044] (5-1)

[0045] In the formula With a small sample size, the clustering results are greatly influenced by the initial cluster centers. Therefore, determining the initial centers requires an intelligent method with independent parameters. Calculate the mean of the few samples and the Euclidean distance between the samples and the mean, sorting them in descending order. The index of the selected cluster centers can be achieved using the following formula.

[0046] (5-2)

[0047] After clustering the minority class samples, the initial density of each sample in the same cluster is set to 0. Then, the nearest neighbor sample is found based on Euclidean distance. For the t-th sample, if its neighbor is the k-th sample, the density of the k-th sample is increased by 1. Finally, the following random tool is used to select samples to synthesize new samples.

[0048]

[0049] In the formula Let represent the fitness of the t-th solution, and N be the number of solutions. Based on the above equations, the sum of the selection probabilities assigned to all solutions according to the following equation should be 1.

[0050]

[0051] Therefore, the higher fitness solution, which benefits from a higher selection probability, is more likely to be selected. The low fitness solution still has a certain chance of being selected, so the selection probability of the low fitness solution is not 0. Then the probability of the t-th sample being selected is as follows.

[0052]

[0053] In the formula Let represent the density of the t-th minority class sample. This represents the density of the k-th minority class sample. Based on the newly synthesized samples, an oversampling method is used to select the actual minority samples. During the oversampling phase, the imbalance ratio IR should be 2, meaning the number of minority class samples should be half the number of majority class samples. The undersampling phase begins when the imbalance ratio IR = 2. The imbalance ratio IR is calculated as follows.

[0054]

[0055] In the formula Indicates minority class samples, This represents the majority class sample.

[0056] The undersampling stage uses the same clustering method as the oversampling stage to calculate the local density of the samples. However, the difference is that the undersampling stage selects samples located in denser regions and removes other samples until the minority sample count equals the majority sample count. Similar to the oversampling stage, the optimal initial center K-means algorithm for clustering is first applied to the majority class samples.

[0057]

[0058] In the formula This represents the number of majority samples. Since the majority sample density is calculated using the nearest neighbor rule during the oversampling phase, the majority samples are selected sequentially based on probability during the undersampling phase, with non-dense regions having a lower probability of selection. This process continues until the number of samples in both classes is equal. The selected samples are considered representative, and unselected samples are deleted. The probabilities of a sample being selected are as follows.

[0059]

[0060] When the IR reaches 1, the training set is balanced, and the output dataset is generated.

[0061] (2) Turbofan Engine Remaining Life (RUL) Based on TMSA-TCN. This invention introduces a temporal multi-scale attention mechanism to improve the temporal convolutional neural network (TCN) and combines it with the TCN to achieve turbofan engine remaining life prediction. TMSA can strengthen the representation of important information features but weaken the representation of low-quality features with weak correlation, while the TCN prediction model is easily misled by interference signals, reducing the accuracy of remaining life prediction. This invention constructs a new life prediction model, MTSA-TCN, by embedding the temporal multi-scale attention mechanism (TMSA) module into the ReLU layer of the temporal convolutional neural network (TCN).

[0062] The essence of Engine Remaining Life (RUL) prediction is to integrate large amounts of data and apply algorithms or AI-trained models to predict the remaining operating time of an engine from its current state until it experiences failure or performance degradation to its normal operating threshold. This training model quantifies and estimates the engine's remaining lifespan. TCN (Tracking Convolutional Network) is a popular model in the cutting-edge field of neural networks, demonstrating powerful data processing capabilities when dealing with high-dimensional nonlinear data devices such as turbofan engines, particularly in detection and classification. TCN incorporates advanced techniques such as causal convolution, dilated convolution, and residual links, which can extract the historical features of time-series data. Causal convolution ensures no information leakage over time, dilated convolution maintains the exponentially increased receptive field of TCN, and residual links ensure deeper stability.

[0063] Causal convolution is a unidirectional structure where previous time steps determine subsequent time steps, and it is not affected by future time steps. The value of the next layer at time t and its preceding values ​​determines the value of the previous layer at time t. Compared to traditional neural networks, causal convolution cannot observe future data and is a model strictly constrained by time. Therefore, when the time series is... and filter are Under these circumstances, The positional causal convolution can be represented by the following formula.

[0064]

[0065] Dilated convolution, due to the strict limitations imposed by kernel size and network depth on the time series modeling of traditional convolutional neural networks, would result in a greater number of linearly stacked layers if traditional neural convolutional networks were used to handle longer and more complex time dependencies. The receptive field (RF) is calculated as follows.

[0066]

[0067] In the formula, n represents the nth layer of the convolutional network, and k represents the kernel size. Introducing dilated convolutions changes the relationship between the receptive field and the hidden layers from linear to exponential. The model achieves a larger receptive field while maintaining a shallower network depth, making it easier to optimize and converge. The receptive field is calculated as follows.

[0068]

[0069] In the formula, n is the number of hidden layers. Let represent the expansion factor of the nth hidden layer, and Dilated convolutional networks, such as... Figure 4 As shown.

[0070] Residual links transmit information in the network model by adding cross-layer connections, allowing input signals to be directly passed to the output of subsequent layers. The output equation of the residual block is as follows.

[0071]

[0072] In the formula, y represents the residual block output, and x represents the residual block input. For a series of transformation operations, This is a one-dimensional convolution operation. The activation function is ReLU. By introducing residual connectivity, TCN can better convey information and utilize historical information, thereby improving modeling capabilities.

[0073] In engine remaining life prediction, when the input data has a high dimensionality, the TCN network may be misled by interference information, thus reducing prediction accuracy. The TMAS attention module can be used flexibly as needed, existing as a standalone module or added to the TCN network to enhance fault feature extraction. The TMAS module utilizes a multi-head attention framework, encompassing multi-scale convolutional layers and multi-head attention layers, with the multi-scale convolutions applying different kernel sizes. and filling The convolutional layers are connected to form the network.

[0074]

[0075] Each convolution operation and final output The calculation is shown in the following formula.

[0076]

[0077]

[0078] The multi-head attention layer consists of a scaling dot product attention mechanism and a linear projection. The process includes local and global attention, and the local attention weights are calculated as follows.

[0079]

[0080] In the formula The key vector dimension is represented by the weights, which are calculated using the softmax function and regularized using dropout, as shown in the following formula.

[0081]

[0082] Local attention is obtained by multiplying the attention weights by the value matrix V.

[0083]

[0084] Global attention, on the other hand, obtains the overall output by summing the results.

[0085]

[0086] In the formula This is the output of the global attention function. The TMSA-TCN network structure is as follows: Figure 5 As shown.

[0087] In the TMSA-TCN network, TMSA enhances the representation of key feature information while mitigating the influence of low-quality features, thus improving the TCN network's robustness against interference. The flowchart for engine life prediction based on TMSA-TCN is shown below. Figure 6 As shown.

[0088] The dataset is input as a matrix into the TMSA-TCN network for training, and the output is the prediction result. The TMSA-TCN network is trained using the PyTorch deep learning framework, and the parameters are adjusted based on the training loss and prediction results until a certain level of accuracy is achieved. Then, the test set is input into the trained model to obtain the remaining lifetime prediction result. The cross-entropy loss function used in classification tasks is employed in both model training and lifetime prediction.

[0089]

[0090] In the formula, L and R represent the number of training samples and the number of fault categories, respectively. , This indicates the predicted fault type and the actual fault type.

[0091] (3) Evaluation of prediction results. In predicting the remaining life of turbofan engines, high accuracy is often the target. However, due to imbalanced data composition, This represents data where the majority class sample is classified as a minority class sample by the classifier. Similarly, This represents data where minority class samples are classified as majority class samples by the classifier. It can reflect the actual situation of misdiagnosis. It can reflect the number of samples that were missed. These represent the number of minority and majority classes correctly classified by the classifier, respectively. Therefore, the following metrics are used to evaluate the accuracy of turbofan engine life prediction.

[0092] Accuracy refers to the average value obtained by a prediction method across multiple predictions, and is calculated as follows:

[0093]

[0094] The f-measure metric is the harmonic mean of precision and recall, defined as follows.

[0095]

[0096] In the formula That is, recall rate. For accuracy, they are calculated using the following formulas respectively.

[0097]

[0098]

[0099] The G-mean metric is the geometric mean of precision and recall, and is often used to evaluate the classification performance of classification models, especially for imbalanced data. It is defined as follows.

[0100]

[0101] Accuracy is the proportion of correctly classified samples out of all samples; precision is the proportion of positive samples correctly identified as true positive samples; and recall is the proportion of correctly identified positive samples out of the total number of positive samples. Scholars have analyzed these metrics and found that f-measure and G-mean, as indicators of classifier performance, can provide relatively objective evaluation metrics when dealing with imbalanced data, and can still achieve relatively more objective evaluation metrics even when the imbalance index (IR) is large.

[0102] This embodiment utilizes the CMPASS dataset for commercial modular aerospace propulsion system simulation provided by NASA's Ames Failure Research Center for simulation verification. This dataset comprises four subsets, FD001-FD004, each containing both training and test sets. The FD001-FD004 datasets represent simulated data under various operating and failure conditions. The training set for each subset includes all cycle data of the engine from the start of aircraft service until operational failure, while the test set includes monitoring data of the engine from its initial state until the failure. This embodiment selects 20 key features as shown in Table 1.

[0103] Table 1. Extracted key features in the time and frequency domains

[0104]

[0105] 20% of the CMPASS dataset was allocated as a validation set for experimental verification. To demonstrate the accuracy of TMSA-TCN in predicting turbofan engine lifespan, four CMPASS datasets were compared with three classic models: the ensemble learning random forest algorithm (RF), its variant Bidirectional Long Short-Time Neural Network (BiLSTM), and the Long Short-Time Neural Network (LSTM). The model in this embodiment achieved good prediction results on all four datasets. The prediction accuracy on the single-condition FD001 and FD003 datasets was similar to that of the three classic models. On the multi-condition FD002 and FD004 datasets, it achieved more ideal results, demonstrating higher accuracy compared to other models, indicating improved model performance.

[0106] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for predicting the remaining service life of a turbofan engine based on TMSA-TCN, characterized in that, include: Historical operational monitoring data of turbofan engines are collected and preprocessed, and the preprocessed historical operational monitoring data is divided into training set and test set. The Temporal Multi-Scale Attention Module (TMSA) is embedded into the residual block of the Temporal Convolutional Neural Network (TCN) to construct the TMSA-TCN prediction model. The training set is then input into the TMSA-TCN prediction model for training to obtain the trained TMSA-TCN prediction model. The test set is input into the trained TMSA-TCN prediction model, and the remaining service life prediction result of the turbofan engine is output. The Temporal Multi-Scale Attention Module (TMSA) consists of multi-scale convolutional layers and multi-head attention layers. The multi-scale convolutional layers extract features at different time scales in parallel by connecting multiple one-dimensional convolutional layers with different kernel sizes. The multi-head attention layers perform weighted fusion of multi-scale features through local and global attention mechanisms to enhance the representation of key features. The temporal multi-scale attention module (TMSA) is embedded after the ReLU activation layer of the temporal convolutional neural network (TCN).

2. The method for predicting the remaining service life of a turbofan engine based on TMSA-TCN according to claim 1, characterized in that, Collect and preprocess historical operational monitoring data of the turbofan engine, including: The historical operation monitoring data is preprocessed using a density-based hybrid oversampling and undersampling resampling method to make the imbalance ratio (IR) of the majority class samples to the minority class samples equal to 1, thus obtaining a balanced training set. The balanced training set is then input into the TMSA-TCN prediction model for training to obtain the trained TMSA-TCN prediction model.

3. The method for predicting the remaining service life of a turbofan engine based on TMSA-TCN according to claim 2, characterized in that, The hybrid resampling method includes: in the oversampling stage, using the K-means algorithm to cluster the minority class samples, and using a probabilistic selection method to synthesize new samples according to the sample density, until the imbalance ratio (IR) between the number of majority class samples and the number of minority class samples is 2; in the undersampling stage, using the K-means algorithm to cluster the majority class samples, and using a probabilistic selection method to delete low-density samples according to the sample density, until the imbalance ratio (IR) between the number of majority class samples and the number of minority class samples is 1.

4. The method for predicting the remaining service life of a turbofan engine based on TMSA-TCN according to claim 1, characterized in that, The Temporal Convolutional Neural Network (TCN) employs a network structure that includes causal convolution, dilated convolution, and residual connections.

5. The method for predicting the remaining service life of a turbofan engine based on TMSA-TCN according to claim 1, characterized in that, After inputting the training set into the TMSA-TCN prediction model for training, the method further includes training the TMSA-TCN prediction model based on the PyTorch deep learning framework, using the cross-entropy loss function as the optimization objective.

6. The method for predicting the remaining service life of a turbofan engine based on TMSA-TCN according to claim 1, characterized in that, The method further includes: analyzing and evaluating the accuracy, F-measure value, and G-mean value of the predicted remaining service life of the turbofan engine.

Images