Aero-engine remaining useful life prediction method

Abstract

An aero-engine remaining useful life prediction method belongs to the field of aero-engine remaining useful life prediction, and aims to solve the problems of the existing aero-engine RUL prediction model. The present application predicts the parallel space-time Transformer model, and the parallel space-time Transformer model comprises a time attention module TAM, a space attention module SAM, a space-time feature fusion device and a support vector machine MLP; TAM and SAM are arranged in parallel, and input samples are input into TAM and SAM respectively; time series enhanced features are obtained through TAM; space enhanced features are obtained through SAM; the space-time feature fusion device is used for fusing the time series enhanced features and the space enhanced features; and the support vector machine MLP performs aero-engine remaining useful life prediction according to the fusion features output by the space-time feature fusion device.

Description

Technical Field

[0001] The invention relates to a method for predicting the remaining service life of an aero-engine, and belongs to the field of prediction of the remaining service life of an aero-engine. Background Art

[0002] As the primary power source for aircraft, the operating status and health of aircraft engines are crucial to their normal operation. Remaining Useful Life (RUL) prediction aims to provide information on engine health based on collected monitoring data, helping decision-makers formulate appropriate usage and maintenance plans. This can effectively prevent engine performance deterioration and even catastrophic failure. Therefore, accurately estimating aircraft engine RUL helps improve engine safety and reliability.

[0003] The current mainstream RUL prediction methods can be roughly divided into three categories: model-based methods, data-driven methods, and hybrid model-based methods. First, model-based methods attempt to reveal the degradation or failure mechanism based on physical knowledge, which is quite difficult for aircraft engines. One major reason is that aircraft engines are complex electromechanical and hydraulic systems composed of tens of thousands of parts, and their state changes involve the coupling of many different components. To date, the degradation or failure mechanisms of many aircraft engines have not been clearly revealed. Similarly, although hybrid methods, model-based methods, and data-driven methods have high accuracy and reliability for RUL prediction, the complexity of modeling makes them somewhat unsuitable for practical engineering applications.

[0004] In contrast, data-driven methods rely on historically collected monitoring data, enabling RUL prediction with little domain knowledge and high accuracy. Furthermore, data-driven methods can be regularly updated with collected monitoring data, enhancing their adaptability to non-stationary operating conditions. More importantly, aircraft engines have numerous sensors and collect a vast amount of data, providing sufficient raw data for RUL estimation, making data-driven methods stand out among the three methods.

[0005] Aircraft engines undergo thousands of cycles during their service life. Due to variations in operational levels and operating conditions, the collected monitoring data is often highly nonlinear. Compared to traditional data-driven methods, deep learning offers powerful nonlinear feature extraction capabilities. It can automatically extract useful and informative features from raw nonlinear monitoring data and then use these features to predict the RUL of aircraft engines. This entire process eliminates the need for complex modeling or specialized statistical feature selection by engineers, making the RUL prediction task much simpler.

[0006] To date, many deep learning models, such as convolutional neural networks (CNNs), recurrent neural networks (RNNs), long short-term memory (LSTMs), temporal convolutional networks (TCNs), and deep residual networks, have been widely applied in the field of aircraft engine RUL prediction and have been proven to effectively improve the performance of life prediction. For prediction tasks involving short-sequence data, CNNs and RNNs have demonstrated their superior capabilities in life prediction. However, faced with the challenges of long-sequence data, traditional CNNs and RNNs are unable to capture long-term dependencies. Fortunately, the emergence of LSTMs and TCNs has alleviated this problem to a certain extent. As an improved version of the RNN, the LSTM, by introducing input, forget, and output gates, can effectively store and transmit important long-term information. This structure enables the LSTM to capture critical temporal dependencies when processing long-sequence data, thus playing a vital role in aircraft engine RUL prediction. The TCN, by employing scalable dilated convolution techniques, significantly enhances the model's ability to capture long-term dependencies. This technology enables TCN to handle long-distance temporal dependencies with fewer parameters and higher efficiency, thus performing well in prediction tasks of long sequence data.

[0007] While LSTMs and TCNs have improved their ability to capture long-term dependencies, they still face challenges when processing very long sequences. For example, when handling tasks requiring the capture of long-range dependencies, TCNs typically enhance their capabilities by increasing the number of network layers or expanding the size of the convolutional kernels. While these approaches improve model performance to a certain extent, they also inevitably increase model complexity. More importantly, increasing model complexity also increases the risk of overfitting, which can affect the model's generalization ability in practical applications. While the LSTM's unique gating mechanism enables it to effectively capture long-term dependencies in time series, it still faces the problem of vanishing or exploding gradients when processing long sequences. These problems often affect the model's training process and ultimate prediction performance. Furthermore, the LSTM's step-by-step training mechanism limits its computational efficiency, making it inadequate for processing large-scale data. The Transformer model addresses these issues effectively. Its unique self-attention mechanism can establish dependencies between any two data points in long sequences, regardless of sequence length. Furthermore, the Transformer's parallel computing capabilities significantly reduce the time complexity of training and testing, alleviating issues such as exploding gradients. Although the above methods have achieved good prediction results in the field of aircraft engine RUL prediction, their prognostic performance is still limited by the following three inherent problems:

[0008] (1) There are many types of sensors used to monitor the health of aircraft engines, such as temperature sensors and pressure sensors. However, different types of monitoring parameters have different effects on RUL prediction, and only some parameters have a significant impact on RUL prediction. On the one hand, redundant parameters lead to inefficient data analysis; on the other hand, they may cause deep learning models to over-focus on irrelevant data, ultimately leading to increased RUL prediction errors and increased time consumption.

[0009] (2) During aircraft engine operation, different flight states will cause the signals collected by sensors to contain different degradation information. For aircraft, some violent movements such as lift and dive will make the engine more susceptible to fatigue damage, while stable flight is relatively safe. If the feature extraction process does not focus on violent movements, some useless information may be captured and noise may be amplified, which will have a certain impact on RUL prediction.

[0010] RUL (Remaining Life) is the result of aircraft engine performance degradation over time, reflected in the cumulative effects of various sensor parameters over time. This means that the engine's historical operating conditions significantly influence the current remaining life prediction. Therefore, when predicting RUL, it is necessary to understand the cumulative contribution of sensor parameters to performance degradation at different times, and to emphasize the contributions of those historical moments that have a more significant impact on RUL.

[0011] (3) In the prediction of the remaining life of aircraft engines, if the sensor parameter dimension is regarded as the spatial dimension and the change of sensor parameters over time is regarded as the temporal dimension, the above three inherent problems require that the established prediction model needs to be able to extract both temporal and spatial dependencies. In recent years, the attention mechanism has attracted widespread attention due to its potential in identifying key information in spatial and temporal dimensions. Many researchers have integrated the attention mechanism into deep learning models to extract more representative features, thereby improving the accuracy of remaining life prediction. In this way, the model can more accurately identify and utilize degradation patterns in sensor data, thereby optimizing the prediction results.

[0012] However, most current attention-based methods focus solely on highlighting important information from either the temporal or spatial dimensions. For aircraft engine RUL prediction, both spatial and temporal information have an impact on RUL prediction. Therefore, to further improve RUL prediction performance, some studies have incorporated both temporal and spatial attention mechanisms into deep learning models, simultaneously focusing on important information in both spatial and temporal dimensions. Compared to using either temporal or spatial attention alone, the combined use of temporal and spatial attention mechanisms can further improve the model's predictive performance because it can capture both temporal and spatial dependencies.

[0013] It is worth noting that in the current RUL prediction methods based on the hybrid attention mechanism, temporal attention and spatial attention are used in series, which will lead to a new challenge, that is, temporal attention and spatial attention will interfere with each other. At present, relevant studies have shown that temporal attention and spatial attention mechanisms will achieve different effects when they are in different positions in the series structure, see Figure 1 As shown, the features first pass through the temporal attention and then the spatial attention mechanism, or first pass through the spatial attention and then the spatial attention mechanism. That is to say, different arrangements in the serial structure will cause the features strengthened by the temporal attention and the features strengthened by the spatial attention to influence each other, which may cause some important degenerate information to be lost. For example, if the temporal attention mechanism is applied first, it will strengthen the multidimensional time series from the time dimension, but the subsequent spatial attention mechanism focuses on the spatial features strengthened by the temporal attention, which may cause some of the original spatial relationships of the multidimensional time series to be lost. Similarly, applying the spatial attention mechanism first and then the temporal attention mechanism will also produce similar problems. On the contrary, if the temporal attention and spatial attention mechanisms are connected in parallel, see Figure 2 As shown in the figure, the features are independently processed by the temporal attention and spatial attention mechanisms. These two attention mechanisms can then independently extract temporally enhanced and spatially enhanced features from the input multidimensional time series without interfering with each other, maximizing the temporal and spatial dependencies of the original series. Furthermore, as previously mentioned, RUL (Running Life) is the result of aircraft engine degradation as operating time increases during service. The cumulative effect of parameters over time is generally more important for RUL prediction. In a parallel connection structure, features enhanced by temporal attention can be given greater weight to further emphasize the role of temporal attention. The current challenge is how to effectively integrate the features enhanced by temporal attention with those enhanced by spatial attention to fully benefit RUL prediction. Summary of the Invention

[0014] In view of the problems existing in the existing aircraft engine RUL prediction model, the present invention provides an aircraft engine remaining useful life prediction method.

[0015] The present invention provides a method for predicting the remaining useful life of an aircraft engine. The remaining useful life of the aircraft engine is predicted using a parallel spatiotemporal Transformer model. The parallel spatiotemporal Transformer model includes a temporal attention module (TAM), a spatial attention module (SAM), a spatiotemporal feature fuser, and a support vector machine (MLP).

[0016] The temporal attention module TAM and the spatial attention module SAM are set in parallel, and the input samples are input into the temporal attention module TAM and the spatial attention module SAM respectively; the temporal enhancement features are obtained by the temporal attention module TAM; the spatial enhancement features are obtained by the spatial attention module SAM;

[0017] The temporal enhancement features and spatial enhancement features are fused using a spatiotemporal feature fusion. The spatiotemporal feature fusion takes the temporal enhancement features as the main body and the spatial enhancement features as the auxiliary. The multi-head attention mechanism is used to convert the spatial enhancement features into an attention matrix, and the attention matrix is ​​multiplied with the temporal enhancement features to highlight the dominant position of the temporal enhancement features.

[0018] The support vector machine (MLP) is used to predict the remaining useful life of aircraft engines based on the fusion features output by the spatiotemporal feature fuser.

[0019] Preferably, the process of obtaining the input sample is as follows:

[0020] Collect different types of sensor parameters of various aircraft engine components under different operating conditions, such as temperature, pressure, and speed. These sensor data are used to construct original multidimensional time series.

[0021] A series of preprocessing operations are performed on the original multidimensional time series, including normalization, exponential smoothing, and sliding time window division;

[0022] After the above preprocessing operation, the input sample is obtained, which is a time series of T×m dimensions. Where T represents the length of the time series and m represents the number of sensors.

[0023] Preferably, the process of obtaining temporal enhancement features through the temporal attention module TAM includes:

[0024] Step 11: Perform average pooling along the spatial dimension to convert it into a time series of T×1 dimensions

[0025] Where GAP1D(·) represents one-dimensional global average pooling, express The i-th column in the table represents the sensor number, i=1,2,...,m.

[0026] Step 12: Use multi-scale convolution to pool features Processing is done to obtain the temporal dependencies between different time steps:

[0027]

[0028] In the formula Represents the output feature of the p-th convolution branch in the multi-scale convolution layer, (k p ,b p ) represents the convolution kernel and the corresponding bias vector of the p-th convolution branch, p = 1, 2, ..., n, n represents the number of convolution branches, * represents a one-dimensional convolution operation;

[0029] Step 13: Use the Sigmoid operator to process the output features of each convolution branch in the multi-scale convolution layer to obtain the importance of each time step at different time scales:

[0030]

[0031] In the formula represents the weight information of each time step at the pth time scale, and f(·) represents the Sigmoid activation function;

[0032] Step 14: Obtain the timing enhancement feature F as follows TAM :

[0033]

[0034] Where, represents matrix multiplication, and n represents the number of branches in the multi-scale convolutional layer.

[0035] Preferably, the process of obtaining spatial enhancement features through the spatial attention module SAM includes:

[0036] Step 21: Input sample Get a multidimensional time series X∈R with m×T dimensions m×T ;

[0037] Step 22: Obtain the correlation matrix S:

[0038]

[0039] Where X=[α1,α2,...,α m ] T , α i ∈R 1×T represents the row vector of X, is the inner product between the i-th and j-th sensor parameter vectors;

[0040] Step 23: Obtain the spatial enhancement feature F as follows SAM :

[0041]

[0042] Where, Represents matrix addition.

[0043] Preferably, the process of fusing the temporal enhancement features and the spatial enhancement features using the spatiotemporal feature fuser includes:

[0044] Step 31: Enhance the time series features F TAM and spatial enhancement feature F SAM Perform position encoding to obtain fusion features:

[0045]

[0046] in, are the outputs of temporal enhancement features and spatial enhancement features after position encoding;

[0047] POS t,a represents the position encoding with a time step of t, a is an even-odd option, a = 2s or a = 2s + 1, and s represents the dimension;

[0048] When a=2s,

[0049] When a=2s+1,

[0050] Where D model represents the embedding dimension;

[0051] Step 32: As K and Q matrices respectively, and As the V matrix, it is input into the multi-head attention mechanism to fuse the temporal features and spatial features to obtain the spatial attention matrix W:

[0052]

[0053] Where d represents K T The dimension size, K T stands for K transpose.

[0054] Step 33: Use the spatial attention matrix W to adjust the spatial correlation of V and obtain the output feature F of the multi-head attention mechanism out

[0055]

[0056] Where, F fusion is the result of using the spatial attention matrix to strengthen the temporal features. and γ j Represents the matrix F fusion The row vector of represents the time series feature vector of the i-th sensor variable after spatial dependency weighting, F outRepresents the output features of the multi-head attention mechanism, W proj is the learnable weight matrix, Represents the output feature of the qth attention head in the multi-head attention, q = 1, 2, ..., H, H represents the number of attention heads;

[0057] Step 34: Use normalization and forward propagation to process the output feature F of the multi-head attention out Process and obtain the fusion features output by the spatiotemporal feature fusion device

[0058]

[0059] Where, represents the output of the first residual and normalization layer in the spatiotemporal feature fusion, b represents the bias vector, It is the output of the first residual and normalization layer in the spatiotemporal feature fusion, and serves as the fusion feature output by the spatiotemporal feature fusion.

[0060] Preferably, the process of predicting the remaining useful life of an aircraft engine using the support vector machine MLP according to the fusion features output by the spatiotemporal feature fuser is as follows:

[0061] The support vector machine MLP has one hidden layer and two fully connected layers. The outputs of the two fully connected layers are obtained according to the following equations:

[0062]

[0063] in, and Represent the weight matrices corresponding to the two fully connected layers, and b2∈R 1 Represent the bias matrices corresponding to W1 and W2, X out2 Represents the predicted RUL value.

[0064] Beneficial effects of the present invention: The present invention proposes a parallel spatiotemporal Transformer for aircraft engine RUL prediction. A key point is the design of a novel parallel spatiotemporal attention mechanism, which connects the temporal attention mechanism and the spatial attention mechanism in parallel, and describes the importance of data in the time and space dimensions at the same time. TAM (temporal attention module) generates the importance weights of each time step at different time scales through multi-scale local interactions, and then redistributes the importance of each moment of the multi-dimensional time series data through adaptive weighting to obtain temporal enhancement features. SAM (spatial attention module) uses a multi-head attention mechanism to capture the global importance of each sensor parameter, and then gives each sensor parameter of the multi-dimensional time series data the importance in the spatial dimension through adaptive weighting to obtain spatial enhancement features. Specifically, there are the following advantages:

[0065] 1. Parallel extraction of spatial and temporal dependencies of multidimensional time series. The temporal and spatial attention mechanisms are connected in parallel to simultaneously extract the temporal and spatial features of multidimensional time series. This effectively avoids the mutual interference of temporal and spatial information, extracts the true temporal and spatial dependencies of the original multidimensional time series to the greatest extent possible, and improves the modeling capabilities of complex spatiotemporal relationships in multidimensional time series.

[0066] 2. A novel task-oriented spatiotemporal feature fuser is designed to achieve adaptive fusion of temporal enhancement features and spatial enhancement features guided by the RUL prediction task, so that spatial information and temporal information can complement and fully utilize each other, thereby improving the RUL prediction performance.

[0067] 3. Comprehensive experiments were conducted on the C-MAPSS public dataset. The proposed method achieved state-of-the-art results in the aircraft engine RUL prediction task, demonstrating the effectiveness of the proposed method. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 It is a schematic diagram of the serial structure of temporal attention and spatial attention mechanisms;

[0069] Figure 2 This is a schematic diagram of the parallel structure of temporal attention and spatial attention mechanisms;

[0070] Figure 3 It is a schematic diagram of the parallel spatiotemporal Transformer model of the present invention;

[0071] Figure 4 It is a schematic diagram of the data preprocessing process to generate input samples;

[0072] Figure 5 This is a schematic diagram of the parallel spatiotemporal attention mechanism;

[0073] Figure 6 This is a schematic diagram of the principle of spatiotemporal feature fusion and RUL prediction;

[0074] Figure 7 This is the training flow chart of the parallel spatiotemporal Transformer model;

[0075] Figure 8 It is a diagram of the sliding time window construction process;

[0076] Figure 9 is the RMSE and score distribution of different sliding window lengths on the FD001 test dataset, where Figure 9 (a) is the distribution of scores of different sliding window lengths on FD001. Figure 9 (b) is the RMSE distribution diagram of different sliding window lengths on FD001;

[0077] Figure 10 is the RUL prediction result of all engines in the four test datasets, where Figure 10 (a) is the RUL prediction result of all engines in the FD001 test dataset. Figure 10 (b) is the RUL prediction result of all engines in the FD002 test dataset. Figure 10 (c) is the RUL prediction result of all engines in the FD003 test dataset. Figure 10 (d) is the RUL prediction results of all engines in the FD004 test dataset;

[0078] Figure 11 This is a comparison chart of the RUL prediction results of the three methods in four test data sets, where Figure 11 (a) is the RUL prediction results of all engines in the FD001 test dataset using three methods. Figure 11 (b) is the RUL prediction results of all engines in the FD002 test dataset using three methods. Figure 11 (c) is the RUL prediction results of all engines in the FD003 test dataset using three methods. Figure 11 (d) is the RUL prediction results of all engines in the FD004 test dataset using three methods. DETAILED DESCRIPTION

[0079] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0080] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0081] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but they are not intended to limit the present invention.

[0082] Specific implementation method 1: Figures 1 to 11 This embodiment describes a method for predicting the remaining useful life of an aircraft engine. The remaining useful life of the aircraft engine is predicted using a parallel spatiotemporal Transformer model. The parallel spatiotemporal Transformer model includes a temporal attention module (TAM), a spatial attention module (SAM), a spatiotemporal feature fuser, and a support vector machine (MLP).

[0083] The temporal attention module TAM and the spatial attention module SAM are set in parallel, and the input samples are input into the temporal attention module TAM and the spatial attention module SAM respectively; the temporal enhancement features are obtained by the temporal attention module TAM; the spatial enhancement features are obtained by the spatial attention module SAM;

[0084] The temporal enhancement features and spatial enhancement features are fused using a spatiotemporal feature fusion. The spatiotemporal feature fusion takes the temporal enhancement features as the main body and the spatial enhancement features as the auxiliary. The multi-head attention mechanism is used to convert the spatial enhancement features into an attention matrix, and the attention matrix is ​​multiplied with the temporal enhancement features to highlight the dominant position of the temporal enhancement features.

[0085] The support vector machine (MLP) is used to predict the remaining useful life of aircraft engines based on the fusion features output by the spatiotemporal feature fuser.

[0086] This implementation designs a novel parallel spatiotemporal attention mechanism that connects the temporal attention mechanism and the spatial attention mechanism in parallel, while describing the importance of data in both temporal and spatial dimensions. TAM generates the importance weights of each time step at different time scales through multi-scale local interactions, and then redistributes the importance of each moment in the multidimensional time series data through adaptive weighting to obtain temporal enhancement features. SAM uses a self-attention mechanism to capture the global importance of each sensor parameter, and then assigns the importance of each sensor parameter in the spatial dimension of the multidimensional time series data through adaptive weighting to obtain spatial enhancement features.

[0087] The parallel spatiotemporal Transformer for aircraft engine remaining life prediction consists of three main parts. The first part is data preprocessing, see Figure 4 As shown in the figure, it mainly includes data preprocessing and sliding window construction samples; the second part is to extract temporal enhancement features and spatial enhancement features in parallel through the parallel spatiotemporal attention mechanism, see Figure 5 As shown; the third part is the adaptive fusion of spatiotemporal features through a task-oriented feature fusion. Finally, the fused features are input into a simple regression network to achieve RUL prediction, see Figure 6 As shown in the figure, the parallel spatiotemporal attention mechanism utilizes temporal attention to enhance the degradation effects of key timestamps within a multidimensional time series, while simultaneously utilizing spatial attention to enhance the importance of key sensor parameters within the multidimensional time series, helping the model focus more on important time steps and sensor parameters. Furthermore, to improve prediction results, the feature fusion unit adaptively fuses the two proposed feature types based on RUL prediction requirements, generating high-level features that comprehensively enhance key spatiotemporal information and more effectively characterize engine degradation processes.

[0088] More importantly, in different flight mission scenarios, the data collected by sensors varies greatly, so the degradation information carried by each sensor parameter is unknown and changing, which brings great difficulties to traditional RUL prediction. Fortunately, the parallel spatiotemporal Transformer proposed in this patent combines all the advantages of the parallel spatiotemporal attention mechanism and the task-oriented feature fuser. When the operating conditions or human factors change, the parallel spatiotemporal Transformer can recalibrate the importance of different sensor parameters and the importance of different time steps based on the changed monitoring data, and always extract key information based on RUL requirements, so that it still has good generalization for changing multi-dimensional time series sensor data, and can perform accurate RUL predictions under different flight conditions.

[0089] See also Figure 4 , the process of obtaining input samples:

[0090] Collect different types of sensor parameters of various aircraft engine components under different operating conditions, such as temperature, pressure, and speed. These sensor data are used to construct original multidimensional time series.

[0091] A series of preprocessing operations are performed on the original multidimensional time series, including normalization, exponential smoothing, and sliding time window division;

[0092] After the above preprocessing operation, the input sample is obtained, which is a time series of T×m dimensions. Where T represents the length of the time series and m represents the number of sensors.

[0093] Due to variations in operating conditions and flight missions, the importance of different sensor parameters (e.g., temperature, pressure, and fuel flow) often changes. Furthermore, the degradation information contained in the same sensor parameter varies at different moments during engine operation. Therefore, to simultaneously focus on important time steps and sensor parameters, a parallel spatiotemporal attention mechanism is designed to extract spatiotemporal features from the input multidimensional time series data by employing two different attention mechanisms: a temporal attention module that weights different time steps, and a spatial attention module that weights different sensor parameters.

[0094] About the temporal attention module TAM. During operation, when the aircraft performs some intense actions such as lifting and diving, the engine will be more prone to fatigue damage, while the cruising process consumes very little engine performance. That is to say, in the collected multi-dimensional time series monitoring data, it is necessary to focus on those moments that consume a lot of engine energy, because these moments are the key to affecting the remaining life of the engine. To solve this problem, the present invention uses a temporal attention module as a basis to obtain the time dependency in multi-dimensional time series data. The temporal attention module consists of a global average pooling layer, a multi-scale convolution layer, and a nonlinear activation function, of which multi-scale convolution is the core. For a given multi-dimensional time series The detailed process of temporal dependency extraction using the temporal attention module is described as follows:

[0095] See also Figure 5 ,The process of obtaining temporal enhanced features through the temporal attention module TAM includes:

[0096] Step 11: Perform average pooling along the spatial dimension to convert it into a time series of T×1 dimensions Its purpose is to obtain a global average information representation of all sensor parameters at each time step.

[0097]

[0098] Where GAP1D(·) represents one-dimensional global average pooling, express The i-th column in the table represents the sensor number, i=1 , 2,...,m, The value of each time step in represents the global average information of all sensor parameters at this time step.

[0099] Step 12: During aircraft operation, different important actions have different durations, which may cause the engine performance to show different degradation patterns over time. Some features are more significant in a short time scale, while other features may take a longer time scale to appear. Therefore, multi-scale convolution is used to pool features. Processing is performed to realize the interaction of information at different time steps from different time scales, thereby obtaining the temporal dependency between different time steps.

[0100]

[0101] In the formula Represents the output feature of the p-th convolution branch in the multi-scale convolution layer, (k p ,b p ) represents the convolution kernel and the corresponding bias vector of the p-th convolution branch, p = 1, 2, ..., n, n represents the number of convolution branches, * represents a one-dimensional convolution operation;

[0102] Step 13: Use the Sigmoid operator to process the output features of each convolution branch in the multi-scale convolution layer to obtain the importance of each time step at different time scales:

[0103]

[0104] In the formula represents the weight information of each time step at the pth time scale, and f(·) represents the Sigmoid activation function;

[0105] Step 14: Obtain the timing enhancement feature F as follows TAM :

[0106]

[0107] Where, represents matrix multiplication, and n represents the number of branches in the multi-scale convolutional layer.

[0108] These weights Perform matrix multiplication with the original time series, give the original time series time attention, obtain the time series strengthened based on time dependence, and aggregate the strengthened time series at different time scales together to obtain the final time series strengthened by the time attention mechanism.

[0109] Regarding the spatial attention module SAM, in actual processes, different sensors provide various types of parameters, such as temperature, pressure, speed, etc. However, the effect of the engine degradation information contained in these sensor data on RUL prediction is still unknown. If equal attention is paid to these multi-sensor information, the model will not be able to focus on important features, resulting in a waste of attention. Therefore, analyzing the importance of different sensors has become an indispensable part of RUL prediction. To this end, this patent designs a spatial attention module, which enables the model to automatically establish the importance relationship between sensors under different working conditions, thereby enhancing the sensor signals containing more degradation information and suppressing the sensor signals containing less degradation information. For a given multidimensional time series X∈R m×T The specific process of using the spatial attention module to assign spatial weights to input samples is as follows:

[0110] See also Figure 5 ,The process of obtaining spatial enhanced features through the spatial attention module SAM includes:

[0111] Step 21: Input sample Get a multidimensional time series X∈R with m×T dimensions m×T ;

[0112] Step 22: Obtain the correlation matrix S:

[0113]

[0114] Where X=[α1,α2,...,α m ] T , α i ∈R 1×T represents the row vector of X, is the inner product between the i-th and j-th sensor parameter vectors, which is used to measure their correlation. It can be observed that S is a correlation map between all different sensor parameter vectors, and the i-th row vector in S can represent the correlation between the i-th sensor parameter vector and all other sensor parameter vectors, that is, the global spatial dependency of the i-th sensor parameter vector.

[0115] Step 23. Finally, the correlation matrix S is used to adjust the parameter vectors in the original multidimensional time series X to enhance the important sensor parameters while suppressing the unimportant sensor parameters. Specifically, each row vector in S is processed using Softmax to obtain the spatial attention matrix, which is then multiplied by X to obtain the time series F after spatial dependency enhancement. SAM In addition, in order to ensure the stability of the neural network and avoid gradient disappearance and loss of original information, residual connections are introduced. The spatial enhancement feature F is obtained as follows: SAM :

[0116]

[0117] Where, Represents matrix addition.

[0118] For more information about spatiotemporal feature fusion, see Figure 6 As shown in FIG, the process of fusing the temporal enhancement features and the spatial enhancement features using the spatiotemporal feature fuser includes:

[0119] Step 31: Enhance the time series features F TAM and spatial enhancement feature F SAM Perform position encoding to obtain fusion features:

[0120]

[0121] in, are the outputs of temporal enhancement features and spatial enhancement features after position encoding;

[0122] POS t,a represents the position encoding with a time step of t, a is an even-odd option, a = 2s or a = 2s + 1, and s represents the dimension;

[0123] When a=2s,

[0124] When a=2s+1,

[0125] Where D model represents the embedding dimension;

[0126] Step 32: As K and Q matrices respectively, and As the V matrix, it is input into the multi-head attention mechanism to fuse the temporal features and spatial features to obtain the spatial attention matrix W:

[0127]

[0128] Where d represents K TThe dimension size, K T stands for K transpose.

[0129] The K, Q, and V matrices are three important parameters of the multi-head attention mechanism. Q is the query quantity of the word, K is the "queried" vector, and V is the content vector.

[0130] Step 33: Use the spatial attention matrix W to adjust the spatial correlation of V and obtain the output feature F of the multi-head attention mechanism out

[0131]

[0132]

[0133] Where, F fusion is the result of using the spatial attention matrix to strengthen the temporal features. and γ j Represents the matrix F fusion The row vector of Represents the time series feature vector of the i-th sensor variable after spatial dependency weighting, it can be observed that The weighted sum of all original sensor parameter time series feature vectors, the influence of the remaining m-1 parameters on the mth parameter is considered, that is, the spatial dependence between different parameters is established, which is more conducive to extracting degradation features that can characterize the engine state. out Represents the output features of the multi-head attention mechanism, W proj is the learnable weight matrix, Represents the output feature of the qth attention head in the multi-head attention, q = 1, 2, ..., H, H represents the number of attention heads;

[0134] Step 34: Use normalization and forward propagation to process the output feature F of the multi-head attention out Processing is performed to make the degraded features more obvious, and in order to avoid the problem of gradient disappearance or gradient explosion, residual connections are introduced into the calculation process. In order to highlight important degraded features and suppress unimportant features without changing the time dependency, the fusion features output by the spatiotemporal feature fusion are obtained as follows:

[0135]

[0136] Where, represents the output of the first residual and normalization layer in the spatiotemporal feature fusion, b represents the bias vector, It is the output of the first residual and normalization layer in the spatiotemporal feature fusion, and serves as the fusion feature output by the spatiotemporal feature fusion.

[0137] The process of predicting the remaining useful life of an aircraft engine using the support vector machine MLP based on the fusion features output by the spatiotemporal feature fuser is as follows:

[0138] The support vector machine MLP has one hidden layer and two fully connected layers. The outputs of the two fully connected layers are obtained according to the following equations:

[0139]

[0140] in, and Represent the weight matrices corresponding to the two fully connected layers, and b2∈R 1 Represent the bias matrices corresponding to W1 and W2, X out2 Represents the predicted RUL value.

[0141] See also Figure 7 The flowchart of the proposed parallel spatiotemporal Transformer model for aircraft engine RUL prediction is presented. First, the original multidimensional time series is preprocessed by normalization, exponential smoothing, and sliding time window partitioning. Second, the designed parallel spatiotemporal attention mechanism is used to simultaneously extract temporal enhancement features and spatial enhancement features. The designed task-oriented feature fuser is used to fuse the two extracted features to obtain representative engine degradation features. Third, the RUL prediction task is implemented using a multi-layer perceptron, and the Adam optimizer is used to optimize the weights of the parallel spatiotemporal Transformer. Finally, the test sample is input into the trained parallel spatiotemporal Transformer, and the predicted RUL value is output.

[0142] Verification Example. To verify the effectiveness of the parallel spatiotemporal Transformer in aircraft engine RUL prediction, the parallel spatiotemporal Transformer model was tested on the CMAPSS dataset and the results were compared with other state-of-the-art methods. All experiments in this chapter were conducted on the same workstation, which had an Intel Core i5-9300H CPU, an NVIDIA GeForce GTX 1650 GPU, Windows 10 operating system, and Python 3.8 compiled with PyTorch 1.7.1 + CU110.

[0143] Experimental dataset description:

[0144] NASA developed the Commercial Aerospace Propulsion System Simulator (C-MAPSS), which provides simulated data from 21 sensors during aircraft engine operation. The C-MAPSS dataset consists of four subsets, each divided into a training set and a test set, each consisting of different operating conditions and fault types. The training set covers the entire lifecycle of each engine, from its inception to its final failure, while the test set contains data from the start of each engine's operation to a point in time before failure. The goal is to predict the remaining operating cycle of each engine in the test set before its end of service life. Details of each dataset are shown in Table 1.

[0145] Table 1 Introduction to the C-MAPSS dataset

[0146]

[0147] Dataset preprocessing:

[0148] (1) Sensor signal selection

[0149] Although the C-MAPSS dataset provides monitoring parameters for 21 sensors, the data collected by some sensors remains largely stable throughout the engine's lifecycle, making this data inaccurate and unable to reflect engine degradation trends. Therefore, to reduce the time complexity of data processing and model learning, this data should be discarded. After screening, 14 sensor signals with clear increasing or decreasing trends were selected as the model's raw inputs: sensors S2, S3, S4, S7, S8, S9, S11, S12, S13, S14, S15, S17, S20, and S21.

[0150] (2) Marking URL label

[0151] The model takes multidimensional time series data as input and outputs RULs. To achieve this, RULs must be used as supervisory labels during training. However, the raw data in the C-MAPSS training set does not contain these labels. Therefore, the data must be annotated with corresponding RUL values ​​to facilitate accurate model training.

[0152] The C-MAPSS training dataset details the entire operational life of an engine, from initial deployment to failure. In this dataset, the total service life of each engine is defined by the total number of cycles accumulated up to its final operational point. The RUL at each time point is calculated by subtracting the accumulated cycles recorded at that specific point in time from the total service life.

[0153]

[0154] In the formula represents the RUL of the nth engine at time t, r n represents the total service life of the nth engine, and t represents the number of cycles that occur.

[0155] For aircraft engines, the relationship between RUL and the number of operating cycles is not linear. During the initial engine operation, RUL changes little. However, once the number of cycles exceeds a critical threshold, RUL rapidly decreases as the number of cycles increases. Therefore, the present invention uses a piecewise linear programming method to represent the engine's RUL. The calculated RUL upper limit is set to 125; beyond this value, RUL remains fixed.

[0156]

[0157] Where r max Indicates the RUL threshold during the initial operation of the engine, which is set to 125.

[0158] (3) Data normalization

[0159] After marking the RUL, the next step is to normalize the raw data. Since the raw data comes from different sensors, their dimensions and ranges may be different. This difference often leads to a large number of dimensional data dominating the gradient calculation, thereby affecting the experimental results and hindering the convergence of the model. By normalizing the data, the accuracy and convergence speed of the model can be improved to a certain extent. The present invention first classifies the raw data according to the engine operating conditions, and then uses the Z-Score formula to normalize the data under each operating condition. The Z-Score formula is as follows:

[0160]

[0161] In the formula and Respectively represent the normalized data and original data of the jth sensor under the working mode k, and represent the mean and standard deviation of the jth sensor data under operating mode k respectively.

[0162] (4) Exponential smoothing

[0163] To address the difficulties in extracting degradation trend features and the inability to remove noise, the present invention uses the exponential moving average (EMA) method to process the data. EMA reduces the impact of extreme values ​​in the process of smoothing the data, making it easier to capture long-term degradation trends. The EMA method calculation formula is as follows:

[0164]

[0165] Where α represents the attenuation value (exponential smoothing factor), 0<α≤1, and They represent the data of the jth sensor of the nth engine after exponential smoothing processing and the data before processing, respectively.

[0166] Exponential smoothing combines current and historical data, fully leveraging extended time series information. This approach enables the model to effectively identify deep degradation patterns in time series data while minimizing the disruptive effects of outliers. After debugging, the present invention sets the decay value of the FD001 training set to 1 and the decay value of the remaining training sets to 0.1.

[0167] (5) Sliding window setting

[0168] Since aircraft engine degradation is a continuous time process, inputting multiple consecutive time nodes into the model simultaneously helps capture more useful time series information. Therefore, after the above processing steps, the present invention uses a sliding time window method to batch input data into the model.

[0169] The schematic diagram of using the sliding window method to segment data is as follows Figure 8 As shown, T represents the window size and S represents the sliding step size.

[0170] Evaluation metrics: Two common metrics are used to evaluate the prediction performance of the model on the C-MAPSS dataset, namely score and root mean square error (RMSE). The calculation formulas of the two metrics are as follows:

[0171]

[0172] In the formula and Y i They represent the RUL predicted value and true value of the i-th sample, and N represents the number of samples.

[0173] The RMSE curve is symmetrical, meaning it evaluates early and late predictions equally. The Score curve, however, is asymmetrical. For aircraft engines, overestimating the RUL can lead to serious accidents. Therefore, the Score function imposes a greater penalty on late predictions, making the evaluation more reasonable.

[0174] Hyperparameter settings:

[0175] To determine the optimal hyperparameters for the proposed method, a grid search method was used to determine the hyperparameters for the parallel spatiotemporal Transformer. The relevant training and model hyperparameters are detailed in Tables 2 and 3, respectively. Based on previous research, before conducting experiments on sliding window length selection, the initial sliding window sizes for the four datasets FD001 to FD004 were tentatively set to 30, 20, 35, and 15, respectively. To ensure that the model captures comprehensive degradation information, the sliding step size was uniformly set to 1. This study employed an early stopping strategy, stopping training when the validation set loss begins to increase, thereby reducing the risk of overfitting. Therefore, there was no fixed number of epochs per training session; the maximum number of training iterations and the waiting number of early stopping epochs were set to 200 and 3, respectively. The Adam optimizer was used to update network parameters with a learning rate of 0.001. A dropout rate of 0.1 was used to further reduce model complexity and the possibility of overfitting. Furthermore, training and validation data were randomly selected from the training set in a 4:1 ratio of the number of engines.

[0176] Table 2 Training hyperparameters

[0177]

[0178] Table 3 Model hyperparameters

[0179]

[0180]

[0181] Sliding window length selection:

[0182] In time series forecasting tasks, identifying the relationships between adjacent time nodes is crucial for predicting RUL. To study the impact of different sliding window lengths on the performance of the parallel spatiotemporal Transformer on each sub-dataset and to reveal the optimal configuration of the model on different sub-datasets, we conducted sliding window length selection experiments.

[0183] According to previous research, sliding windows that are too short may not effectively capture degradation information that is critical to model performance, while windows that are too long may introduce irrelevant interference. Therefore, experiments with different sliding window lengths should be conducted for each subset to determine the optimal sliding window length for the parallel spatiotemporal Transformer. The four subsets in the C-MAPSS dataset have different maximum and minimum life cycles. The minimum life cycle number for each subset should be selected as the upper limit of the sliding window length. Table 4 shows the maximum and minimum life cycles of the engines in the four subsets.

[0184] Table 4 Maximum and minimum number of cycles for each subset

[0185]

[0186] Table 4 shows that the minimum number of life cycles in the test set is consistently lower than that in the training set. Therefore, the upper limit of the sliding window length for each subset should be determined by the minimum life cycle of the test set. The minimum number of life cycles in the test set for each subset was 31, 21, 38, and 19, respectively. Therefore, the upper limit of the sliding window length was set to 30, 20, 35, and 15, respectively, and experiments were conducted with increments of 5 sliding window lengths. The hyperparameters for these experiments are shown in Tables 2 and 3. To ensure the reliability of the results, each experiment was repeated ten times, and the average of the results was taken as the final result. The experimental results are shown in Table 5. Figure 9 The box plots show the experimental results for the FD001 dataset. Clearly, setting the sliding window length to 30 not only minimizes both the score and RMSE, but also minimizes the variance. This optimal setting indicates strong stability in the prediction process. Therefore, the optimal window length for FD001 is set to 30. As the sliding window length decreases, the score and RMSE gradually increase. This is because smaller sliding window lengths make it more difficult for features to be fully captured by the model. Similarly, using the same method for the FD002, FD003, and FD004 subsets, the optimal sliding window lengths were determined to be 20, 15, and 15, respectively.

[0187] Table 5 Experimental performance of the method using different window sizes on four subsets.

[0188]

[0189] Comparison between the method of the present invention and other existing methods:

[0190] The performance of the proposed parallel spatiotemporal Transformer was verified using four sub-datasets of C-MAPSS and compared with other advanced RUL prediction methods to highlight the advantages of the parallel spatiotemporal Transformer. The model was trained with the hyperparameters in Table 3 and the sliding window length in Table 4, and then the test data was input into the trained model for RUL prediction. The prediction results are shown in Table 6. In order to reduce the impact of randomness, all experiments were performed ten times, and the average value of these ten experiments was used as the final result. The experimental results show that the model proposed in this invention performs well on the C-MAPSS dataset.

[0191] Table 6. Parallel spatiotemporal Transformer performance on the C-MAPSS dataset.

[0192]

[0193] Figure 10The prediction results of the parallel spatiotemporal Transformer model on four test sets are presented. The true RUL values ​​of each engine in each dataset are sorted for easy visualization. As shown in the figure, the parallel spatiotemporal Transformer performs well on the C-MAPSS dataset, with predictions very close to the true RUL values. Among all datasets, FD001 and FD003 achieve the most accurate predictions, followed by FD002, while FD004 performs the worst. This difference can be attributed to the complexity and diversity of the operating conditions and failure modes of FD002 and FD004. Specifically, FD001 includes one operating condition and one failure mode, while FD003 includes one operating condition and two failure modes. In contrast, FD002 includes six operating conditions and one failure mode, while FD004 includes six operating conditions and two failure modes. Therefore, the FD002 and FD004 datasets are more complex, posing a greater challenge for in-depth analysis of engine degradation trends. Furthermore, engines with smaller true RUL values ​​achieve higher prediction accuracy than engines with larger true RUL values. The reason for this phenomenon is that engines nearing the end of their service life have accumulated a large amount of degradation information, and the model can capture more degradation information, so the predicted value is closer to the true value.

[0194] To further demonstrate the superior performance of the parallel spatiotemporal Transformer model, we conducted comparative experiments with several state-of-the-art methods. These include mainstream models such as DCNN, LSTM, CNN-LSTM, AGCNN, BiGRU-AS, DSAN, and STRUL, as well as the Transformer and its variants, such as PAO Transformer, TATFA-Transformer, and TF-SCN. All of these comparative methods were applied to aircraft engine RUL prediction, consistent with the objectives of the method described in this paper. The experimental results are shown in Table 7.

[0195] The model proposed in this invention has a significant improvement over traditional models such as DCNN, LSTM, and CNN-LSTM, and its performance is also better than some of the latest Transformer improved models, such as PAOTransformer and TF-SCN. It is worth noting that the parallel spatiotemporal Transformer model performs best on the FD002 and FD004 datasets. Compared with the second-best model TF-SCN, the RMSE is reduced by 11.6% and 11.8%, and the Score is reduced by 35.4% and 6.1%, respectively. On the FD001 dataset, TF-SCN achieved the best results, and the parallel spatiotemporal Transformer followed closely behind, ranking second in both RMSE and Score indicators. On the FD003 dataset, TATFA-Transformer achieved the best results, and the parallel spatiotemporal Transformer ranked second and seventh in RMSE and Score indicators, respectively.

[0196] Overall, the Parallel Spatiotemporal Transformer consistently achieved commendable results on the C-MAPSS dataset, particularly on the FD002 and FD004 subsets, where it significantly outperformed other models. The Parallel Spatiotemporal Transformer's advantage stems from its innovative architecture, integrating a spatial attention module for assessing spatial importance from a holistic perspective and a temporal attention module for assessing temporal importance at different scales. These two modules are connected in parallel and cleverly fuse spatiotemporal features through a multi-head attention mechanism. Notably, in the RUL prediction task of our proposed method, the temporal attention features, considered the most critical, are fed into the multi-head attention module as V, tailored to the specific requirements of the task. This approach ensures that critical temporal features are significantly preserved during the feature fusion process. Furthermore, the model employs exponential smoothing to embed historical degradation data into time nodes, thereby mitigating the impact of noise and improving RUL prediction accuracy. Essentially, the Parallel Spatiotemporal Transformer's sophisticated design and strategic data processing contribute to its exceptional performance. Experimental results demonstrate that our proposed method outperforms most existing methods.

[0197] Table 7. Comparison of RMSE and scores of different methods.

[0198]

[0199] Ablation experiment:

[0200] The ablation experiment of the present invention is used to demonstrate the effects of the proposed parallel spatiotemporal attention mechanism and feature fuser, and two different situations are studied. The first situation is a network without adding a feature fuser, and only integrating the parallel spatiotemporal attention mechanism. The output characteristics of the parallel spatiotemporal attention mechanism are merged through a summation process. This specific model is named ST-N. The second scheme is characterized in that the network only contains a feature fuser, in which the Q, K and V of the multi-head attention mechanism are obtained by linear transformation of the same input information. This model is called TN. The above two models and the parallel spatiotemporal Transformer model proposed in the present invention were tested on the test set, and RMSE and Score were used as evaluation indicators. The experiment was repeated 10 times and the average value was taken as the final result. The prediction results of the three models are compared in Table 8. It can be seen from the table that TN is better than ST-N, because except for the Score value of the experimental results on FD003, TN has lower experimental values ​​than ST-N. It is worth noting that TN performs well on datasets with complex operating conditions and diverse fault types, such as FD002, FD003, and FD004, because its RMSE value is consistently low. In addition, the parallel spatiotemporal Transformer has lower RMSE and Score values. Specifically, the performance of ST-N lags far behind that of the parallel spatiotemporal Transformer. On the four subsets, the RMSE value of the parallel spatiotemporal Transformer is reduced by 20.4%, 17.4%, 18.1%, and 24.1% respectively compared with ST-N; the Score is reduced by 45.6%, 18.5%, 16.3%, and 39.0% respectively. This difference highlights the advantages of the feature fusion method introduced in this invention, which can select and integrate the most significant spatial features and fuse them with temporal features, while discarding irrelevant data. The performance of TN also lags behind that of the parallel spatiotemporal Transformer. On the four subsets, the RMSE values ​​of the parallel spatiotemporal Transformer are reduced by 15.8%, 10.2%, 8.0%, and 4.4% respectively compared with TN; the score is reduced by 41.0%, 18.3%, 20.0%, and 17.4% respectively. This further highlights the applicability of the parallel spatiotemporal Transformer in the RUL prediction task. The multi-head attention mechanism in the Transformer can only evaluate the importance of features from a single dimension. In contrast, after integrating the spatiotemporal feature extraction module with the Transformer decoder, the importance of spatiotemporal features can be considered at the same time, so that the model can capture more comprehensive degradation information. The experiment verifies the effectiveness of the spatiotemporal feature extraction method proposed in this invention.

[0201] Table 8 RMSE and scores of the three RUL prediction models on the test set in the ablation study.

[0202]

[0203] To intuitively display the results of the ablation experiment, Figure 8 The prediction results of the three methods on four datasets are shown. As can be seen from the figure, the prediction results of the Parallel Spatiotemporal Transformer on the four datasets are closer to the true RUL values ​​than the other two methods, indicating that the Parallel Spatiotemporal Transformer has higher prediction accuracy.

[0204] In summary, in order to solve the problems of complex working conditions, noise interference of collected signals, difficulty in capturing important degradation information in time series information, and unknown importance of different sensors to RUL prediction in the task of aircraft engine RUL prediction, the present invention proposes a new architecture of parallel spatiotemporal Transformer. First, the temporal features and spatial features are captured independently by the parallel connected temporal feature extractor and spatial feature extractor. Secondly, the spatiotemporal features are deeply fused through the improved Transformer decoder. Finally, the high-dimensional features obtained by feature extraction and fusion are mapped to the final RUL value through the fully connected layer.

[0205] Multiple experiments were conducted on the model to determine its optimal hyperparameters. Ablation experiments were then conducted to confirm that the model had no redundant structure. Finally, the proposed method was compared with current mainstream methods. The experimental results show that the proposed method significantly outperforms mainstream methods in terms of RUL prediction accuracy and also performs well on the more difficult FD002 and FD004 predictions.

[0206] Predicting the RUL of aircraft engines is an important and evolving research topic. In the future, we will focus on studying the impact of factors such as engine materials, structure, and maintenance on RUL, hoping to achieve more accurate RUL predictions by considering more comprehensive factors.

[0207] Specific implementation method 2. This implementation method further optimizes the implementation method 1, proposes a second key point, and designs a novel task-oriented spatiotemporal feature fuser, which realizes the adaptive fusion of temporal enhancement features and spatial enhancement features guided by the RUL prediction task, so that spatial information and temporal information can complement and fully utilize each other, thereby improving the RUL prediction performance.

[0208] Temporal features derived from learning temporal dependencies in multivariate time series focus on describing trends in time series and correlations between different time steps, helping models better understand and predict dynamic behavior and evolution in time series. Spatial features derived from learning spatial dependencies in multivariate time series reflect the correlations and interactions between different variables, as well as how they jointly influence system behavior. This means that these two types of features play different roles for different downstream tasks. For example, in RUL prediction, temporal features play a primary role, while spatial features play a secondary role. Conversely, in certain classification tasks, spatial features play a primary role, while temporal features play a secondary role. However, many current deep learning-based RUL prediction methods embed different attention mechanisms to extract spatial and temporal features. However, on the one hand, most of them aggregate the two types of features through simple concatenation or matrix addition, significantly reducing the final effectiveness. On the other hand, they are designed for specific tasks. When applied to other domains or tasks, these models often need to be redesigned and trained, wasting considerable time and resources. Therefore, how to flexibly use spatial features and temporal features, and adaptively integrate spatial features and temporal features in downstream tasks to achieve the best effect.

[0209] Using a task-oriented spatiotemporal feature fusion can make a model multi-purpose.

[0210] Most current methods mainly use simple matrix addition or splicing to aggregate spatiotemporal features, which may lead to problems such as loss or redundancy of key information. In order to avoid this situation, the present invention proposes a novel task-oriented spatiotemporal feature fuser, which flexibly defines Q, K, and V in the multi-head attention mechanism according to the downstream task, and uses the multi-head attention mechanism to achieve adaptive fusion of spatiotemporal features, so that temporal feature information and spatial feature information are fully integrated and complement each other under the guidance of downstream tasks. Specifically, unlike the traditional transformer, which uses linear transformation to obtain Q, K, and V matrices for any downstream task, the proposed spatiotemporal feature fuser directly assigns different types of features output by the parallel spatiotemporal attention mechanism to Q, K, and V for different downstream tasks. For example, for the RUL prediction task (using an existing multi-sensor time series to predict sensor data at a future time point, implementation method one is this type of task), the more important temporal features are used as the V matrix, and the spatial features are used as Q and K respectively to calculate the spatial correlation attention weight matrix. On the contrary, for classification tasks (such as fault diagnosis, using algorithms to determine the type of fault in a mechanical part), spatial features may be more important, so the spatial features are used as the V matrix, and the temporal features are used as the Q and K matrices to calculate the temporal correlation attention weight matrix. According to the above analysis, by combining the parallel spatiotemporal attention mechanism with the spatiotemporal feature fuser, the deficiency of the traditional Transformer model that cannot simultaneously consider temporal correlation and spatial correlation is improved, thereby achieving multi-purpose use of one model. Figure 6 As shown, the spatiotemporal feature fusion replaces the Transformer encoder with a PSTAM, retaining the decoder structure. More importantly, for the RUL prediction task that the present invention aims to solve, the spatiotemporal feature fusion replaces K and Q in the decoder with spatially enhanced features extracted by the spatial attention module, while V is replaced by temporally enhanced features extracted by the temporal feature module.

[0211] Although the present invention is described herein with reference to specific embodiments, it should be understood that these embodiments are merely illustrative of the principles and applications of the present invention. It should be understood that many modifications may be made to the illustrative embodiments, and that other arrangements may be devised, without departing from the spirit and scope of the present invention as defined by the appended claims. It should be understood that features described in various dependent claims and in this invention may be combined in ways other than those described in the original claims. It should also be understood that features described in conjunction with individual embodiments may be used in conjunction with other described embodiments.

Claims

1. A method for predicting the remaining useful life of an aircraft engine, characterized in that: The remaining useful life of aircraft engines is predicted using a parallel spatiotemporal Transformer model, which includes a temporal attention module (TAM), a spatial attention module (SAM), a spatiotemporal feature fuser, and a support vector machine (MLP). The temporal attention module TAM and the spatial attention module SAM are set in parallel, and the input samples are input into the temporal attention module TAM and the spatial attention module SAM respectively; the temporal enhancement features are obtained by the temporal attention module TAM; the spatial enhancement features are obtained by the spatial attention module SAM; Use the spatiotemporal feature fusion to fuse the temporal enhancement features and spatial enhancement features; The spatiotemporal feature fusion takes temporal enhancement features as the main body and spatial enhancement features as the auxiliary. It uses the multi-head attention mechanism to transform the spatial enhancement features into an attention matrix and multiply the attention matrix with the temporal enhancement features to highlight the dominant position of the temporal enhancement features. The support vector machine (MLP) predicts the remaining useful life of aircraft engines based on the fusion features output by the spatiotemporal feature fuser. The process of fusing temporal enhancement features and spatial enhancement features using the spatiotemporal feature fuser includes: Step 31: Enhance the time series features F TAM and spatial enhancement feature F SAM Perform position encoding to obtain fusion features: in, are the outputs of temporal enhancement features and spatial enhancement features after position encoding; POS t,a represents the position encoding with a time step of t, a is an even-odd option, a = 2s or a = 2s + 1, and s represents the dimension; When a=2s, When a=2s+1, Where D model represents the embedding dimension; Step 32: As K and Q matrices respectively, and As the V matrix, it is input into the multi-head attention mechanism to fuse the temporal features and spatial features to obtain the spatial attention matrix W: Where d represents K T The dimension size, K T stands for K transpose; Step 33: Use the spatial attention matrix W to adjust the spatial correlation of V and obtain the output feature F of the multi-head attention mechanism out ; Where, F fusion is the result of using the spatial attention matrix to strengthen the temporal features. and γ j Represents the matrix F fusion The row vector of represents the time series feature vector of the i-th sensor variable after spatial dependency weighting, F out Represents the output features of the multi-head attention mechanism, W proj is the learnable weight matrix, represents the output feature of the qth attention head in the multi-head attention, q = 1, 2, ..., H, H represents the number of attention heads, and m represents the number of sensors; Step 34: Use normalization and forward propagation to process the output feature F of the multi-head attention out Process and obtain the fusion features output by the spatiotemporal feature fusion device Where, represents the output of the first residual and normalization layer in the spatiotemporal feature fusion, b represents the bias vector, It is the output of the first residual and normalization layer in the spatiotemporal feature fusion, and serves as the fusion feature output by the spatiotemporal feature fusion.

2. The method for predicting the remaining useful life of an aircraft engine according to claim 1, characterized in that: The process of obtaining input samples: Collect different types of sensor parameters of various aircraft engine components under different operating conditions, such as temperature, pressure, and speed. These sensor data are used to construct original multidimensional time series. A series of preprocessing operations are performed on the original multidimensional time series, including normalization, exponential smoothing, and sliding time window division; After the above preprocessing operation, the input sample is obtained, which is a time series of T×m dimensions. Where T represents the length of the time series and m represents the number of sensors.

3. The method for predicting the remaining useful life of an aircraft engine according to claim 2, characterized in that: The process of obtaining temporal enhanced features through the temporal attention module TAM includes: Step 11: Perform average pooling along the spatial dimension to convert it into a time series of T×1 dimensions Where GAP1D(·) represents one-dimensional global average pooling, express The i-th column in the table represents the sensor number, i=1,2,...,m. Step 12: Use multi-scale convolution to pool features Processing is done to obtain the temporal dependencies between different time steps: In the formula represents the output feature of the p-th convolution branch in the multi-scale convolution layer, k p represents the convolution kernel of the p-th convolution branch, b p Represents the bias vector corresponding to the convolution kernel of the p-th convolution branch, p = 1, 2, ..., n, n represents the number of convolution branches, * represents a one-dimensional convolution operation; Step 13: Use the Sigmoid operator to process the output features of each convolution branch in the multi-scale convolution layer to obtain the importance of each time step at different time scales: In the formula represents the weight information of each time step at the pth time scale, and f(·) represents the Sigmoid activation function; Step 14: Obtain the timing enhancement feature F as follows TAM : Where, represents matrix multiplication, and n represents the number of branches in the multi-scale convolutional layer.

4. The method for predicting the remaining useful life of an aircraft engine according to claim 3, characterized in that: The process of obtaining spatial enhanced features through the spatial attention module SAM includes: Step 21: Input sample Get a multidimensional time series X∈R with m×T dimensions m×T ; Step 22: Obtain the correlation matrix S: Where X=[α1,α2,...,α m ] T , α i ∈R 1×T represents the row vector of X, is the inner product between the i-th and j-th sensor parameter vectors; Step 23: Obtain the spatial enhancement feature F as follows SAM : Where, Represents matrix addition.

5. The method for predicting the remaining useful life of an aircraft engine according to claim 1, characterized in that: The process of predicting the remaining useful life of an aircraft engine using the support vector machine MLP based on the fusion features output by the spatiotemporal feature fuser is as follows: The support vector machine MLP has one hidden layer and two fully connected layers. The outputs of the two fully connected layers are obtained according to the following equations: in, and Represent the weight matrices corresponding to the two fully connected layers, and b2∈R 1 Represent the bias matrices corresponding to W1 and W2, X out2 Represents the predicted RUL value.

Images