Aero-engine remaining service life interval prediction method based on spatial-temporal feature fusion network

Abstract

The invention relates to the technical field of aero-engine residual life prediction, and provides an aero-engine residual service life interval prediction method based on a spatial-temporal feature fusion network. The method comprises the following steps: firstly, extracting time features by adopting an attention mechanism-based long-short-term memory neural network (LSTM), then extracting spatial features associated with the interior of the aero-engine by using three-dimensional convolution, and finally, fusing the two features to realize extraction learning of the spatial-temporal features. Meanwhile, many residual service life prediction algorithms at present only perform point estimation, and neglect the uncertainty of prediction. For the interval prediction of the residual life of the aero-engine, the interval prediction of the residual life is realized by adopting a nonparametric quantile regression method. The method provides a research basis for accurate prediction of the residual life of the aero-engine.

Description

Technical Field

[0001] The present invention relates to the technical field of aircraft engine remaining service life prediction, and in particular to an aircraft engine remaining service life interval prediction method based on a spatiotemporal feature fusion network. Background Art

[0002] Aircraft engines are the core components of aircraft. As the main power source of aircraft, their safety and reliability are crucial to ensuring aviation safety. Therefore, establishing an effective prediction and health management mechanism for aircraft engines is of great significance to ensure their safety and economy. In recent years, thanks to the rapid development of sensor technology and data mining technology, data-driven research on the remaining life prediction of aircraft engines has received widespread attention. In order to effectively monitor the health status of aircraft engines, sensors are usually used to monitor multiple important performance parameters. Data-driven remaining life prediction methods often integrate these monitoring data to achieve accurate prediction of the remaining life. In 2017, Zheng et al. proposed an aircraft engine remaining life prediction algorithm based on time series features, which is superior to traditional machine learning methods such as extreme learning machine (ELM) and random forest (RF) in life prediction. Although prediction algorithms based on time series features have achieved good results, they mainly focus on the information of historical moments and often ignore the associated spatial features between internal structures. In contrast, our study found that the spatiotemporal fusion network can not only pay attention to the time series feature information at the same time, but also effectively measure the relationship between different feature information. Therefore, the spatiotemporal fusion network can extract more comprehensive feature information and provide more accurate predictions based on it.

[0003] In addition, most current data-driven prediction methods only provide point estimates of the target remaining life and lack uncertainty measures for RUL predictions. Therefore, it is very important to provide relevant confidence intervals for RUL predictions. To this end, a RUL prediction framework based on non-parametric methods is constructed to achieve interval prediction of the remaining life of aircraft engines. Summary of the invention

[0004] The present invention aims to solve the problems existing in the prior art and provides a method for predicting the remaining useful life interval of an aircraft engine based on a spatiotemporal feature fusion network.

[0005] The technical solution of the present invention is as follows:

[0006] A method for predicting the remaining useful life interval of an aircraft engine based on a spatiotemporal feature fusion network comprises the following steps:

[0007] Step 1: Process the raw sensor data and construct the training set and test set;

[0008] Step 2: Build an LSTM module based on the attention mechanism to effectively extract the sensor's temporal feature information. At the same time, combine the attention mechanism to assign attention weights to each feature information.

[0009] Step 3: Build a 3D convolutional neural network to extract the internal spatial features of the aircraft engine. In the 3D convolutional neural network, the convolution operation considers the width and height of the image as well as the time dimension to capture the spatial information in the data.

[0010] Step 4: Build a parameter-free quantile regression module to achieve interval prediction of the remaining life of aircraft engines.

[0011] Furthermore, the step 1 is specifically as follows:

[0012] Step 1.1: Effective feature selection;

[0013] First, the resulting disturbance is feature screened according to data changes, and sensor data related to engine degradation is selected, combined with other three operation information as the final feature input for subsequent model training and prediction;

[0014] Furthermore, the sensor data include: flight altitude, Mach number, throttle lever angle, low-pressure turbine outlet temperature, high-pressure compressor outlet temperature, low-pressure turbine outlet temperature, high-pressure compressor outlet total pressure, uncorrected fan speed, uncorrected core engine speed, high-pressure gas outlet static pressure, fuel flow to P30 ratio, fan correction speed, core engine correction speed, bypass ratio, bleed air enthalpy, high-pressure turbine cooling air flow and low-pressure turbine cooling air flow;

[0015] Furthermore, the three operation information include flight altitude, Mach number and throttle lever angle.

[0016] Step 1.2: Normalize and standardize the sensor data selected in step 1.1;

[0017] The sensor data is processed by normalization and standardization, and the formula is as follows:

[0018]

[0019] where x (t,f) represents the fth feature information at time t, Represents the data after standardization. and Respectively represent the maximum and minimum values ​​of the f-th feature;

[0020] Step 1.3: Label the sample data processed in step 1.2;

[0021] In the test set, the RUL label of the remaining service life is found from RUL_FD00* and annotated; in the training set, the RUL label of the remaining service life is calculated from the training set itself and annotated;

[0022] Step 1.4: Process the URL label with a piecewise linear function;

[0023] The sample labels are divided into a stable period and a degenerate period. The sample labels in the stable period remain constant, while the RUL labels in the degenerate period will degenerate to 0 as the running cycle progresses. The piecewise function is defined as follows:

[0024]

[0025] Where t represents the current operating cycle, T represents the maximum flight cycle of the current engine, and R max is the piecewise function value of the setting;

[0026] Step 1.5: Set a sliding window for training samples;

[0027] Set sliding windows for training samples and test samples, divide the samples according to the window size, and finally generate training sets and test sets.

[0028] Furthermore, the step 2 is specifically as follows:

[0029] Step 2.1: Construct an LSTM module to extract time series features;

[0030] The LSTM module mainly realizes the feature extraction of time series information. The internal structure of LSTM is divided into three parts, namely the forget gate, input gate, and output gate. According to the LSTM structure, it is expressed as follows:

[0031] f t =σ(w f [h t-1 ,x t ]+b f )

[0032] i t =σ(w i [h t-1 ,x t ]+b i )

[0033]

[0034]

[0035] o t =σ(w o [h t-1 ,x t]+b o )

[0036] h t =o t *tanh(C t )

[0037] Among them, f t represents the output of the forget gate, x t Represents the input of the current time step, i t Indicates that the input gate retains information, represents the candidate memory cell of the current time step, C t represents the state of the memory cell that combines the previous state and the current input information, o t Represented as the output gate output, h t represents the hidden state at time step t, C t-1 Represents the memory cell state at the previous moment, w f , w i , w c , w o represents the weight, b f , b i , b c , b o represents bias, σ(·) and tanh are sigmoid function and tanh function respectively; after LSTM extracts features, it is necessary to combine the attention mechanism to assign weights to the features according to the contribution of each feature, and finally form a feature vector with attention by combining features and feature weights;

[0038] Step 2.2: Build an attention module based on step 2.1, and assign weights to the features extracted in step 2.1 through the attention mechanism module;

[0039] The features learned by the LSTM network for a sample are represented as X = {X 1 ,X 2 , …, X n} T , T represents the transposition operation, where X i ∈R d , where d is the number of consecutive steps of the feature; based on the self-attention mechanism, for the input X i Feature importance is expressed as:

[0040] s i =φ(W T X i +b)

[0041] Where W and b are the weight matrix and bias vector respectively, φ is the activation function; after obtaining W i The weight coefficient s corresponding to the featurei Afterwards, the softmax function is used for normalization as follows:

[0042]

[0043] Step 2.3: After obtaining the feature weights in step 2.2, the time series information extracted based on the attention mechanism module is formed after weighting with the corresponding features;

[0044] In step 2.2, we obtain the i-th feature weight a i Afterwards, the feature weights are combined with the features to form the final time series information features. The final output of the time series feature extraction module can be expressed as:

[0045] O=X*A

[0046] Where X is represented by {X 1 ,X 2 , …, X n}, A is represented by {a 1 , a 2 , a 3 ..., a n}.

[0047] Furthermore, the step 3 is specifically as follows:

[0048] Step 3.1: Based on the data processed in step 1.4, construct a three-dimensional data sample;

[0049] When using CNN for aircraft engine sensor fault diagnosis, the aircraft engine sensor data should be reconstructed first, so that its data size changes from 1×k to k×k; at the same time, the third dimension of CNN is represented as a time window, and the sensor information features are extracted through three-dimensional CNN, which can effectively extract spatial information features and internal correlation feature information at the same time;

[0050] Step 3.2: Build a three-dimensional convolutional neural network module to extract relevant features between sensor signals;

[0051] After reconstructing the data in step 3.1, the reconstructed data is input into the input layer of the three-dimensional convolution. Then, the convolution operation is performed on the input data using the three-dimensional convolution kernel through the convolution layer to extract spatial and temporal features. Finally, three-dimensional pooling is performed on the output of the convolution layer to reduce the feature dimension.

[0052] Step 3.3: Combine the two features extracted in step 2.3 and step 3.2 and use them as input features for regression prediction;

[0053] After flattening the features extracted in step 3.2 through the fully connected layer output, combined with the time series information features extracted in step 2, the two parts of features are used as high-level features for RUL prediction.

[0054] Furthermore, the step 4 is specifically as follows:

[0055] Given the true remaining life value y i , and the neural network f(x;θ), at the quantile level q∈[0,1], the quantile regression loss is defined as:

[0056]

[0057] In the formula, Represents the predicted remaining life value;

[0058] Step 4.1: Design quantile regression levels to determine the predicted value of RUL, including the upper and lower predicted values ​​of RUL;

[0059] Step 4.2: Design the loss function of the neural network, perform backpropagation according to the loss function and update the parameters for each training;

[0060] Given multiple quantiles Q = {q 1 ,q 2 ,...q M}, the optimal loss function is defined as follows:

[0061]

[0062] According to the optimal loss function value, back propagation is performed to optimize the network parameters, where q m Expressed as given multiple quantiles Q = {q 1 ,q 2 ,...q M}, N represents the number of samples in the data set, and M represents the number of selected quantiles;

[0063] Step 4.3: Use the root mean square error RMSE and the scoring function Score to evaluate the predicted value, and select the quantile loss as the evaluation indicator of interval prediction performance;

[0064] The root mean square error RMSE and the scoring function Score are defined as follows:

[0065]

[0066]

[0067] From the definition of the Score function above, we can see that the penalty for delayed prediction will be increased; at the same time, the smaller the RMSE and Score values ​​are, the higher the prediction accuracy of the model represented is.

[0068] Beneficial effects of the present invention: This method has significant advantages in aircraft engine prediction and health management. First, compared with traditional machine learning methods, it has higher remaining life prediction accuracy. In particular, the spatiotemporal fusion network can simultaneously focus on time series information and the spatial correlation between different features, extract more comprehensive feature information, and improve the stability and adaptability of the prediction. In addition, this method makes full use of sensor monitoring data to achieve real-time monitoring of the engine health status, and combines the interval prediction framework based on non-parametric methods to provide confidence intervals for remaining life prediction and enhance the reliability of the prediction results. Through accurate life prediction, this method can optimize maintenance decisions, promote on-demand maintenance, reduce unplanned downtime, reduce operating costs, and improve engine utilization and service life. More importantly, this method helps to identify potential faults in advance, reduce the risk of sudden accidents, improve aviation safety, and meet the needs of future intelligent aviation development, and has important engineering value and economic benefits. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 This is a flow chart of the remaining useful life interval prediction method for aircraft engines based on spatiotemporal feature fusion network.

[0070] Figure 2 This is a neural network architecture diagram for the remaining useful life interval prediction method for aircraft engines based on a spatiotemporal feature fusion network.

[0071] Figure 3 This is the internal structure diagram of LSTM.

[0072] Figure 4 This is the LSTM structure diagram based on the attention mechanism.

[0073] Figure 5 This is the predicted fitting effect diagram for FD001 to FD004.

[0074] Figure 6 Prediction fitting effect diagram for a single engine.

[0075] Figure 7 A comparison chart of ablation experiment results. DETAILED DESCRIPTION

[0076] A technology for predicting the remaining useful life interval of aircraft engines based on a spatiotemporal feature fusion network. The specific process is as follows: Figure 1 As shown, the specific steps are as follows:

[0077] Step 1: Process the original sensor data and construct the training set and test set.

[0078] Step 1.1: Select effective features;

[0079] First, screen the features of the result perturbation amount according to the data change, and select the sensor data related to the engine degradation. It is found through experiments that not all sensor information will affect the remaining life of the aero-engine. Specifically, among the 21 sensor information in the remaining life data set of the aero-engine, the values of sensors 1, 5, 6, 10, 16, 18, and 19 always remain unchanged, indicating that these 7 sensors have nothing to do with the engine degradation, that is, the value changes of these sensors will not significantly reflect the engine degradation trend. Therefore, finally, only the data of 17 sensors are selected for training. The sensor data includes: flight altitude, Mach number, throttle lever angle, low-pressure turbine outlet temperature, high-pressure compressor outlet temperature, low-pressure turbine outlet temperature, high-pressure compressor outlet total pressure, uncorrected fan speed, uncorrected core speed, high-pressure air outlet static pressure, ratio of fuel flow to P30, fan corrected speed, core corrected speed, bypass ratio, bleed air enthalpy value, high-pressure turbine cooling air flow, and low-pressure turbine cooling air flow. Among these 17 sensors, flight altitude, Mach number, and throttle lever angle are operating conditions.

[0080] Step 1.2: Normalize and standardize the original sensor data;

[0081] The method of normalizing and standardizing the original sensor data is as follows:

[0082]

[0083] Step 1.3: Label the sample data processed in Step 1.2;

[0084] After normalizing and standardizing the data in Step 1.2, it is necessary to label the training samples. When labeling the sample labels, there are differences in the labeling methods between the training set and the test set. This is because in the test set, the remaining service life at the last moment of the aero-engine is unknown and needs to be found from RUL_FD00*, while in the training set, the remaining life of each aero-engine will gradually decay from the healthy state to zero with the cycle, and the remaining service life label can be calculated from the training set itself.

[0085] Step 1.4: Process the RUL label with a piecewise linear function;

[0086] The sample labels are divided into a stable period and a degenerate period. The sample labels in the stable period remain constant, while the RUL labels in the degenerate period will slowly degenerate to 0 as the running cycle progresses. The piecewise function is defined as follows:

[0087]

[0088] In the present invention, R max Set to 125. Set RUL values ​​above 125 to 125, and set RUL values ​​below 125 to their own values.

[0089] Step 1.5: Set a sliding window for training samples;

[0090] At the same time, the samples are divided according to the sample window size. The sliding window can enrich the data samples and make the information carried by each sample richer. In the present invention, for FD001 and FD003, since the failure mode is relatively simple and the data samples are relatively small, the sample window is set to 32. For FD002 and FD004, since the failure mode is relatively complex, the window is set to 64. The division of all training data is achieved through the divided sample windows. After the data sets FD001 to FD004 are divided, the corresponding sample training numbers of FD001 to FD004 are 17700, 48756, 21789, and 57459, and the corresponding sample test numbers are 100, 259, 100, and 248.

[0091] Step 2: Build an LSTM module based on the attention mechanism to effectively extract sensor timing information.

[0092] When constructing the LSTM module based on the attention mechanism, the structure of the neural network is divided into an input layer (17×50), and at the same time through the attention mechanism (s×s, s represents the sample window), and the final attention-based output is (s×50).

[0093] Step 2.1: Construct an LSTM module to extract the temporal features of the extracted information;

[0094] The LSTM module mainly realizes the feature extraction of time series information. The internal structure of LSTM is divided into three parts, namely the forget gate, input gate, and output gate. According to the LSTM structure, it is expressed as follows:

[0095] f t =σ(w f [h t-1 ,x t ]+b f )

[0096] i t =σ(w i [h t-1 ,xt ]+b i )

[0097]

[0098]

[0099] o t =σ(w o [h t-1 ,x t ]+b o )

[0100] h t =o t *tanh(C t )

[0101] When the feature h extracted by LSTM is calculated t After that, the attention mechanism is needed to assign weights to the features, and finally a feature vector with attention is formed by combining the features and feature weights.

[0102] Step 2.2: Build an attention module based on step 2.1, and assign weights to the features extracted in step 2.1 through the attention mechanism module;

[0103] When predicting the remaining life of aircraft engines, it is an effective operation to focus on the area of ​​interest by assigning different weights to features at each time step. At the same time, the self-attention mechanism is used to learn features and time steps. The features learned by the LSTM network for a sample can be expressed as

[0104] X={X 1 ,X 2 , …, X n} T , T represents the transposition operation, where X i ∈R d , where d is the number of consecutive steps of the feature. Based on the self-attention mechanism, for the input X j Feature importance can be expressed as:

[0105] s i =φ(W T X i +b)

[0106] Where W and b are the weight matrix and bias vector respectively, and φ is the score function, which can be designed as an activation function in a neural network, such as sigmoid and linear functions. i The weight coefficient s corresponding to the feature i Afterwards, the softmax function can be used to normalize as follows:

[0107]

[0108] Step 2.3: After obtaining the feature weights in step 2.2, they are weighted with the corresponding features to form the temporal information extracted based on the attention mechanism module.

[0109] In step 2.2, we obtain the i-th feature weight a i Afterwards, the feature weights are combined with the features to form the final time series information features, and the final output can be expressed as:

[0110] O=X*A

[0111] Where X represents {X 1 ,X 2 , …, X n}, A is represented by {a 1 , a 2 , a 3 ..., a n}.

[0112] Step 3: Build a three-dimensional convolutional neural network to extract relevant features between sensor signals.

[0113] In traditional two-dimensional convolutional neural networks, convolution operations are only performed on the width and height of the image, which limits its ability to capture dynamic information. In contrast, three-dimensional convolutional neural networks not only perform convolutions in the spatial dimension, but also consider the temporal dimension at the same time, thereby more comprehensively capturing spatial and temporal information in the data. Applying three-dimensional convolution to the prediction of the remaining life of aircraft engines can effectively extract and capture the associated features in space and improve the accuracy of prediction.

[0114] Step 3.1: Based on the data processed in step 1.4, construct a three-dimensional data sample;

[0115] If only the method in step 2.1 is used for feature extraction, the spatial feature information associated with the internal structure of the aircraft engine at the same time will be ignored. In order to fully extract the spatial feature information, considering the characteristics of the local receptive field of CNN, when using CNN for aircraft engine sensor fault diagnosis, the aircraft engine sensor data should be reconstructed first. When reconstructing the aircraft engine sensor data, the data size is changed from 1×k to k×k. At the same time, the third dimension of CNN is represented as a time window. The sensor information features are extracted through a three-dimensional CNN, which can effectively extract spatial information. By constructing a three-dimensional data sample, the data size becomes s×k×k.

[0116] Step 3.2: Build a three-dimensional convolutional neural network module to extract relevant features between sensor signals;

[0117] The core idea of ​​3D convolution is to use a 3D convolution kernel to slide on the input 3D data to extract spatial features. The three dimensions refer to the width, height, and time (depth). After reconstructing the data in step 3.1, the reconstructed data is input into the input layer of the 3D convolution. Then, a convolution operation is performed on the input data using a 3D convolution kernel (3×3×3) through the convolution layer to extract spatial and temporal features. Finally, 3D pooling, such as maximum pooling or average pooling, is performed on the output of the convolution layer to reduce the feature dimension.

[0118] Step 3.3: Combine the two parts of features extracted in step 2.3 and step 3.2 and use them as input features for regression prediction.

[0119] After flattening the features extracted in step 3.2 through the fully connected layer output, combined with the time series information features extracted in step 2, the two parts of features are used as high-level features for RUL prediction.

[0120] Step 4: Build a parameter-free quantile regression module to achieve interval prediction of the remaining life of aircraft engines.

[0121] When it is difficult to determine the specific distribution, non-parametric methods can be used to achieve interval prediction of the remaining life of aircraft engines. For non-parametric methods, multiple RUL prediction values ​​at different quantile levels can be obtained through quantile regression.

[0122] Step 4.1: Design quantile regression levels to determine the predicted RUL, the upper predicted value of RUL, and the lower predicted value of RUL.

[0123] Given the true remaining life value y t , and the neural network f(x;θ), at the quantile level q∈[0,1], the quantile regression loss is defined as:

[0124]

[0125] Step 4.2: Design the loss function of the neural network, perform backpropagation according to the loss function and update the parameters for each training.

[0126] Given multiple quantiles Q = {q 1 ,q 2 ,...q M}, the optimal loss function is defined as follows:

[0127]

[0128] According to the optimal loss function value, back propagation is performed to optimize the network parameters.

[0129] Step 4.3: Use the root mean square error (RMSE) and the scoring function (Score) to evaluate the predicted value, and select the quantile loss as the evaluation indicator of the interval prediction performance.

[0130] The root mean square error RMSE and the scoring function Score are defined as follows:

[0131]

[0132]

[0133] From the above definition of the Score function, we can see that the penalty for delayed prediction will be increased. At the same time, the smaller the RMSE and Score values ​​are, the higher the prediction accuracy of the model represented is.

[0134] The following examples are used to verify the prediction effect of the present invention:

[0135] The C-MAPSS dataset was selected for testing, including FD001 to FD004. In order to verify the effectiveness of the present invention, the present invention also conducted an ablation experiment. The experimental results are as follows: Figure 7 In addition, the comparison results of the method proposed in the present invention and some current advanced RUL prediction methods are shown in Table 1.

[0136] Table 1

[0137]

[0138]

[0139] Table 1 compares the algorithm of the present invention (TSFNet) with several advanced prediction algorithms, including weighted adaptive deep learning strategy (SSDS), multi-head convolutional long short-term memory network (Multi-head CNN-LSTM), adaptive graph convolutional network (AGCNN), bidirectional gated recurrent unit and temporal self-attention mechanism (BiGRU-TSAM), trend attention fully convolutional network (TaFCN), self-attention architecture (Transformer), incremental multi-feature deep learning and space-time network (IMDSSN). Different from these comparison algorithms, this patent proposes a method for predicting the remaining useful life interval of aircraft engines based on a spatiotemporal feature fusion network. Compared with the comparison algorithm, it can more effectively extract the correlation features inside the engine and has better robustness and accuracy. It can be seen from the comparison results in Table 1 that the method proposed by the present invention is superior to the existing advanced RUL prediction methods in multiple evaluation indicators. The root mean square error (RMSE) and score function (Score) performed well in the four data sets (FD001 to FD004), especially in FD002 (15.18) and FD004 (16.09), where the errors were significantly lower than most of the comparison algorithms. This shows that TSFNet has a strong advantage in prediction accuracy. At the same time, the quantitative indicators of interval prediction ability, QL-0.1 and QL-0.9, also reflect the advantages of TSFNet, showing the stability and robustness of this method at different quantization levels. The ablation experiment verifies that the proposed method can effectively improve the prediction accuracy, especially when dealing with complex flight degradation modes, TSFNet can better capture the degradation trend and further improve the prediction performance. The experimental results show that TSFNet not only improves the training efficiency of the model through precise feature selection and optimized model structure, but also enhances its generalization ability on a variety of flight data sets, fully proving its effectiveness in the prediction of the remaining life of aircraft engines.

[0140] It can be seen that the TSFNet method proposed in this paper has significant advantages in the prediction of remaining useful life. Through comparative analysis, TSFNet is superior to existing advanced methods in multiple evaluation indicators, especially in prediction accuracy and interval prediction stability.

Claims

1. A method for predicting the remaining useful life interval of an aircraft engine based on a spatiotemporal feature fusion network, characterized in that: The steps include: Step 1: Process the raw sensor data and construct the training set and test set; Step 2: Build an LSTM module based on the attention mechanism to effectively extract the sensor's temporal feature information. At the same time, combine the attention mechanism to assign attention weights to each feature information. Step 3: Build a 3D convolutional neural network to extract the internal spatial features of the aircraft engine. In the 3D convolutional neural network, the convolution operation considers the width and height of the image as well as the time dimension to capture the spatial information in the data. Step 4: Build a parameter-free quantile regression module to achieve interval prediction of the remaining life of aircraft engines.

2. The method for predicting the remaining useful life interval of an aircraft engine based on a spatiotemporal feature fusion network according to claim 1 is characterized in that: The step 1 is specifically as follows: Step 1.1: Effective feature selection; First, the resulting disturbance is screened based on data changes, and sensor data related to engine degradation is selected. This is combined with other three operational information as the final feature input for subsequent model training and prediction. Step 1.2: Normalize and standardize the sensor data selected in step 1.1; The sensor data is processed by normalization and standardization, and the formula is as follows: where x (t,f) represents the fth feature information at time t, Represents the data after standardization. and Respectively represent the maximum and minimum values ​​of the f-th feature; Step 1.3: Label the sample data processed in step 1.2; In the test set, the RUL label of the remaining service life is found from RUL_FD00* and annotated; in the training set, the RUL label of the remaining service life is calculated from the training set itself and annotated; Step 1.4: Process the URL label with a piecewise linear function; The sample labels are divided into a stable period and a degenerate period. The sample labels in the stable period remain constant, while the RUL labels in the degenerate period will degenerate to 0 as the running cycle progresses. The piecewise function is defined as follows: Where t represents the current operating cycle, T represents the maximum flight cycle of the current engine, and R max is the piecewise function value of the setting; Step 1.5: Set a sliding window for training samples; Set sliding windows for training samples and test samples, divide the samples according to the window size, and finally generate training sets and test sets.

3. The method for predicting the remaining useful life interval of an aircraft engine based on a spatiotemporal feature fusion network according to claim 1 is characterized in that: The step 2 is specifically as follows: Step 2.1: Construct an LSTM module to extract time series features; The LSTM module mainly realizes the feature extraction of time series information. The internal structure of LSTM is divided into three parts, namely the forget gate, input gate, and output gate. According to the LSTM structure, it is expressed as follows: f t =σ(w f [h t-1 ,x t ]+b f ) i t =σ(w i [h t-1 ,x t ]+b i the t =σ(w o [h t-1 ,x t ]+b o ) h t =o t *tanh(C t ) Among them, f t represents the output of the forget gate, x t represents the input of the current time step, i t Indicates that the input gate retains information, represents the candidate memory cell of the current time step, C t represents the state of the memory cell that combines the previous state and the current input information, o t Represented as the output gate output, h t represents the hidden state at time step t, C t-1 Represents the memory cell state at the previous moment, w f , w i , w c , w o represents the weight, b f , b i , b c , b o represents bias, σ(·) and tanh are sigmoid function and tanh function respectively; after LSTM extracts features, it is necessary to combine the attention mechanism to assign weights to the features according to the contribution of each feature, and finally form a feature vector with attention by combining features and feature weights; Step 2.2: Build an attention module based on step 2.1, and assign weights to the features extracted in step 2.1 through the attention mechanism module; The features learned by the LSTM network for a sample are represented as X = {X1, X2, ..., X n } T , T represents the transposition operation, where X i ∈R d , where d is the number of consecutive steps of the feature; based on the self-attention mechanism, for the input X i Feature importance is expressed as: s i =φ(W T X i +b) Where W and b are weight matrix and bias vector respectively, φ is the activation function; after obtaining X i The weight coefficient s corresponding to the feature i Afterwards, the softmax function is used for normalization as follows: Step 2.3: After obtaining the feature weights in step 2.2, the time series information extracted based on the attention mechanism module is formed after weighting with the corresponding features; In step 2.2, we obtain the i-th feature weight a i Afterwards, the feature weights are combined with the features to form the final time series information features. The final output of the time series feature extraction module can be expressed as: O=X*A Where X is represented by {X1, X2, …, X n }, A is represented by {a1, a2, a3..., a n }.

4. The method for predicting the remaining useful life interval of an aircraft engine based on a spatiotemporal feature fusion network according to claim 1, characterized in that: The step 3 is as follows: Step 3.1: Based on the data processed in step 1.4, construct a three-dimensional data sample; When using CNN for aircraft engine sensor fault diagnosis, the aircraft engine sensor data should be reconstructed first, so that its data size changes from 1×k to k×k; at the same time, the third dimension of CNN is represented as a time window, and the sensor information features are extracted through three-dimensional CNN, which can effectively extract spatial information features and internal correlation feature information at the same time; Step 3.2: Build a three-dimensional convolutional neural network module to extract relevant features between sensor signals; After reconstructing the data in step 3.1, the reconstructed data is input into the input layer of the three-dimensional convolution. Then, the convolution operation is performed on the input data using the three-dimensional convolution kernel through the convolution layer to extract spatial and temporal features. Finally, three-dimensional pooling is performed on the output of the convolution layer to reduce the feature dimension. Step 3.3: Combine the two features extracted in step 2.3 and step 3.2 and use them as input features for regression prediction; After flattening the features extracted in step 3.2 through the fully connected layer output, combined with the time series information features extracted in step 2, the two parts of features are used as high-level features for RUL prediction.

5. The method for predicting the remaining useful life interval of an aircraft engine based on a spatiotemporal feature fusion network according to claim 1, characterized in that: The step 4 is specifically as follows: Given the true remaining life value y i , and the neural network f(x;θ), at the quantile level q∈[0,1], the quantile regression loss is defined as: In the formula, Represents the predicted remaining life value; Step 4.1: Design quantile regression levels to determine the predicted value of RUL, including the upper and lower predicted values ​​of RUL; Step 4.2: Design the loss function of the neural network, perform backpropagation according to the loss function and update the parameters for each training; Given multiple quantiles Q = {q1,q2,...q M }, the optimal loss function is defined as follows: According to the optimal loss function value, back propagation is performed to optimize the network parameters, where q m It is expressed as a given number of quantiles Q = {q1, q2, ...q M }, N represents the number of samples in the data set, and M represents the number of selected quantiles; Step 4.3: Use the root mean square error RMSE and the scoring function Score to evaluate the predicted value, and select the quantile loss as the evaluation indicator of interval prediction performance; The root mean square error RMSE and the scoring function Score are defined as follows: From the definition of the Score function above, we can see that the penalty for delayed prediction will be increased; at the same time, the smaller the RMSE and Score values ​​are, the higher the prediction accuracy of the model represented is.

6. The method for predicting the remaining useful life interval of an aircraft engine based on a spatiotemporal feature fusion network according to claim 2, characterized in that: The sensor data include: flight altitude, Mach number, throttle lever angle, low-pressure turbine outlet temperature, high-pressure compressor outlet temperature, low-pressure turbine outlet temperature, high-pressure compressor outlet total pressure, uncorrected fan speed, uncorrected core engine speed, high-pressure gas outlet static pressure, fuel flow to P30 ratio, fan correction speed, core engine correction speed, bypass ratio, bleed air enthalpy, high-pressure turbine cooling air flow and low-pressure turbine cooling air flow; the three operation information include flight altitude, Mach number and throttle lever angle.

Images