Engine remaining service life prediction method and system based on spatiotemporal state selection

Abstract

This invention discloses a method and system for predicting the remaining useful life of an engine based on spatiotemporal state selection. This method, which belongs to the field of equipment remaining useful life prediction, integrates multiple technologies, including temporal-spatial dual-branch encoding and hierarchical fusion, a selective state space model, a cross-layer residual skip connection and parameter sharing mechanism, and an efficient self-attention mechanism, to construct an STSSFormer network model for aircraft engine remaining useful life prediction. The system primarily includes a data selection and preprocessing module, an input sample construction module, a model training module, and a remaining useful life prediction module. This invention comprehensively enhances feature expression and fusion capabilities, offering significant advantages in remaining useful life prediction accuracy, model robustness, and engineering practicality.

Description

Technical Field

[0001] The present invention belongs to the field of equipment remaining service life prediction, and in particular relates to a method and system for predicting the remaining service life of an engine based on spatiotemporal state selection. Background Art

[0002] With the advent of the big data era and the rise of low-cost computing power, modern industry is placing higher demands on the reliability and predictive maintenance of complex systems. Remaining Useful Life (RUL), a key component of predictive health management (PHM), has been widely applied to various types of equipment and systems, particularly in areas requiring regular maintenance and management. For example, in industries such as aviation, energy, batteries, and mechanical equipment, analyzing sensor data that monitors equipment operating status can estimate the RUL of equipment, enabling preventive maintenance and avoiding economic losses caused by unexpected failures.

[0003] However, due to the uncertainty of the equipment operating environment and conditions, RUL prediction remains a challenging problem. Existing research methods usually combine data analysis and RUL prediction models, and integrate expert knowledge to achieve RUL prediction of equipment such as aircraft engines. These methods generally include physical degradation models based on equipment failure principles, or by designing corresponding machine learning algorithms to learn degradation trends in sensor data to predict RUL. However, with the development of industry, the complexity of mechanical systems and changes in the operating environment have made it increasingly difficult to accurately construct physical degradation models. Therefore, data-driven methods have received increasing attention and have made significant progress in practical applications.

[0004] Designs based on Transformer models usually process temporal and spatial features independently. For multi-dimensional data, there is a lack of effective cross-modal fusion mechanisms, resulting in information loss or error accumulation. At the same time, traditional convolutional networks are used to extract spatial features. Traditional convolutions usually require processing a large number of parameters and calculations, especially when the number of channels and convolution kernel size of the input feature map are large, the amount of calculation increases sharply; existing solutions often use fixed weighting coefficients or simple splicing methods to merge temporal and spatial features, and are unable to dynamically adjust the weights of temporal and spatial features according to different input data. This may cause some key features to not be fully paid attention to, thereby affecting model performance; the computational complexity of the Transformer increases quadratically with the increase in sequence length, and is not suitable for processing long time series data; traditional models (such as LSTM and Transformer) mainly rely on explicit temporal modeling or spatial feature extraction, and rarely involve how to effectively model the transition between state spaces. Summary of the Invention

[0005] In order to solve the above problems, the present invention proposes a method and system for predicting the remaining useful life of an aircraft engine based on spatiotemporal state selection. It integrates multiple technologies such as temporal-spatial dual-branch coding and hierarchical fusion mechanism, selective state space model, cross-layer residual skip connection and parameter sharing mechanism, and efficient self-attention mechanism, and constructs the STSSFormer (Spatial-Temporal Selective State Transformer, STSSFormer) network model to predict the remaining useful life of an aircraft engine, effectively realizing real-time health management of the aircraft engine and reducing maintenance costs.

[0006] The technical solutions of the present invention are as follows:

[0007] A method for predicting the remaining useful life of an engine based on spatiotemporal state selection comprises the following steps:

[0008] Step 1: Acquire aircraft engine sensor monitoring data and perform data selection and preprocessing;

[0009] Step 2: Use the sliding time window technique to construct a multi-period multi-sensor sample;

[0010] Step 3: Construct the STSSFormer network model. The STSSFormer network model includes an embedding and positional encoding module, a spatial and temporal feature extractor, and a predictor. The STSSFormer network model uses the embedding and positional encoding module to process the input data, uses the processed data as input to the spatial and temporal feature extractor to extract features, and finally constructs a predictor based on the selective state space model to predict the remaining useful life of aircraft engines.

[0011] Step 4: Initialize the weights of the STSSFormer network model, optimize the parameters of the network model using mean square error, and train the network model using the multi-cycle multi-sensor samples constructed in step 2;

[0012] Step 5. After the training is completed, the trained STSSFormer network model is obtained. The current aircraft engine sensor monitoring data to be predicted is collected and processed according to the process of steps 1 and 2. The processed data is input into the trained STSSFormer network model to obtain the aircraft engine remaining service life prediction results in real time.

[0013] Furthermore, the specific process of step 1 is as follows:

[0014] Step 1.1: Obtain sensor monitoring values ​​within different measurement ranges of the aircraft engine to form a sensor time series data set , For the The monitoring data of all sensors in time steps, is the total number of time steps, is the transpose symbol; ,in Indicates the The sensor in time step data, is the total number of sensors; the sensor monitoring data of aircraft engines includes temperature parameters, pressure parameters, speed parameters, some cooling-related parameters, exhaust and emission parameters, vibration and noise parameters, fuel system parameters, and lubricating oil system parameters; the time-varying characteristics of each sensor data are analyzed, invalid sensor channels with constant values ​​are eliminated, and valid sensor data with significant change trends are retained; for data under different operating conditions, a differentiated processing strategy is adopted for data selection based on their complexity differences. The differentiated processing strategy is: when the equipment is operating under fixed conditions, only the data collected by the key sensors that best reflect the aging of the equipment need to be selected; when the equipment is operating under complex and changeable conditions, the complete data collected by all sensors need to be retained;

[0015] Step 1.2: In the data preprocessing stage, the sensor monitoring data is normalized to the range of [-1, 1] using the minimum and maximum scaling method.

[0016] Furthermore, in step 2, when using the sliding time window technology to construct the sample, the sensor monitoring data is divided into time windows of fixed length, and the data in a time window is a multi-period multi-sensor sample; the time window slides along the time dimension, and the remaining service life corresponding to the last time step in each window will be used as the remaining service life label of all data in the current window.

[0017] Furthermore, in step 3, the specific working process of the STSSFormer network model is as follows:

[0018] Step 3.1: Input the multi-period multi-sensor samples into the embedding and position encoding module to obtain a sensor time series data set containing embedded position information;

[0019] Step 3.2: Input the sensor time series data set containing embedded location information into the spatial and temporal feature extractor to obtain the spatial and temporal fused features;

[0020] Step 3.3: Predict the remaining useful life of the aircraft engine based on the predictor.

[0021] Furthermore, in step 3.1, the embedding and position encoding module includes an embedding layer and a position encoding layer, and the specific working process is as follows:

[0022] Step 3.1.1, embedding layer integrates sensor time series data Mapping to a higher dimension to get an embedding vector ;

[0023] Step 3.1.2: Position code is generated using sine and cosine functions, using the formula:

[0024] ;

[0025] ;

[0026] in, Encode for position; is the position index; It is a dimension; is the dimension of the embedding vector;

[0027] Step 3.1.3: Add the embedding vector and the position code to obtain the sensor time series data set containing the embedded position information. :

[0028] ;

[0029] in, is the time window size.

[0030] Furthermore, in step 3.2, a temporal-spatial dual-branch encoding and hierarchical fusion mechanism is used to construct a spatial and temporal feature extractor, which specifically includes a spatial attention module, a temporal attention module, and a feature fusion layer. The specific working process of the spatial and temporal feature extractor is as follows:

[0031] Step 3.2.1: Extract spatial features based on the spatial attention module. The spatial attention module is composed of two stacked encoder layers. The core design of each encoder layer is as follows:

[0032] First, we use depthwise convolution to After processing, three point-by-point convolutions are used to dynamically generate the query matrix, key matrix, and value matrix of self-attention. The formula is:

[0033] ;

[0034] ;

[0035] ;

[0036] ;

[0037] in, It is the feature representation after deep convolution processing in the time dimension; is the depth convolution; 、 、 These are the query matrix, key matrix, and value matrix generated by the spatial attention module respectively; is point-by-point convolution;

[0038] Will generate 、 、 Split into multiple attention heads and embed the input into different subspaces; improve the self-attention mechanism in the traditional Tranformer into an efficient self-attention mechanism. The calculation process is as follows:

[0039] ;

[0040] ;

[0041] in, is the global context vector, 、 are the dimensions of key and value respectively; is the output of the attention head, is the batch size, is the number of attention heads;

[0042] The outputs of multiple attention heads are concatenated and then linearly mapped to obtain a multi-head feature map. , and use residual connections to combine them with Add them together, and then use the normalization layer to normalize the result. The formula is:

[0043] ;

[0044] ;

[0045] in, For splicing; 、 、 、 Represents the output of four attention heads; is the weight matrix; is the output after normalization and residual connection; is the normalization layer;

[0046] Finally, the output of the current encoder layer is obtained through the feedforward network and used as the input of the next encoder layer. This is passed layer by layer until the spatial features output by the final spatial attention module are obtained:

[0047] ;

[0048] in, is the spatial feature; represents a feedforward network;

[0049] Step 3.2.2: Extracting temporal features based on temporal attention module The principle of the temporal attention module is basically the same as that of the spatial attention module. The only difference is the calculation of the query matrix, key matrix, and value matrix. The temporal attention module obtains the three matrices through linear projection. The calculation process is as follows:

[0050] ;

[0051] ;

[0052] ;

[0053] in, 、 、 These are the query matrix, key matrix, and value matrix generated by the temporal attention module respectively; 、 、 They are 、 、 The weight matrix of

[0054] Step 3.2.3: Design a feature fusion layer to fuse the features obtained by the spatial attention module and the temporal attention module. The specific implementation is as follows:

[0055] ;

[0056] in, It is the feature after spatial and temporal fusion; 、 are the weight matrices of different features respectively.

[0057] Furthermore, in step 3.3, the predictor adopts a cross-layer residual skip connection and parameter sharing mechanism, specifically including two selective state space models connected in series; the specific working process of the predictor is:

[0058] Step 3.3.1: Features after spatial and temporal fusion Normalize and obtain the residual branch through linear transformation:

[0059] ;

[0060] in; is the residual branch; is a linear transformation; is the root mean square normalization;

[0061] Step 3.3.2: Branch the residual Input to the first layer of selective state space model, the selective state space model generates three vectors for state transfer after input projection: state step , state input weight , state output weight ; The state step is activated after linear projection:

[0062] ;

[0063] in, is the state step after activation; is the activation function; is the weight matrix of the state step;

[0064] Define the state transition matrix , The initial value is a full 1 matrix, and its subsequent update method is determined by the learnable parameters Decide:

[0065] ;

[0066] The hidden state update process of the selective state space model is:

[0067] ;

[0068] in, 、 Respectively , The hidden state of the time step; The input sequence is Features at time steps; and For the Sutra The two different parameters after modulation are calculated as follows:

[0069] ;

[0070] ;

[0071] in, For the The time step time steps; For the The state step length after activation of time steps; is element-wise multiplication;

[0072] The final output of the first layer selective state space model The calculation process is as follows:

[0073] ;

[0074] The output of the first layer selective state space model After the activation function and Dropout layer, it is used as the input of the second-layer selective state space model; finally, the output of the second-layer selective state space model and the residual branch After fusion, the predicted value of the remaining service life is obtained :

[0075] ;

[0076] in, is the learnable gating coefficient.

[0077] Furthermore, in step 4, the mean square error is selected as the loss function of the model. During the training process, the model parameters are optimized according to the loss function. The formula is as follows:

[0078] ;

[0079] in, Indicates the number of samples; and Corresponding to the The actual observed and predicted values ​​of the remaining useful life of the samples.

[0080] A system for predicting the remaining useful life of an engine based on spatiotemporal state selection, using the above-mentioned method for predicting the remaining useful life of an engine based on spatiotemporal state selection, includes the following modules:

[0081] A data selection and preprocessing module is used to select and preprocess the acquired aircraft engine sensor monitoring data;

[0082] Input sample construction module, which uses sliding time window technology to construct multi-period multi-sensor samples as input to the STSSFormer network model;

[0083] The model training module is used to initialize the weights of the STSSFormer network model and optimize the parameters of the network model using mean square error;

[0084] The remaining useful life prediction module uses the STSSFormer network model including embedding and position encoding modules, spatial and temporal feature extractors, and predictors to predict the remaining useful life of aircraft engines.

[0085] The beneficial technical effects brought about by the present invention: The present invention comprehensively enhances the feature expression and fusion capabilities through an innovative spatiotemporal decoupling parallel coding structure, selective state space modeling and efficient attention mechanism, effectively overcoming the problems of insufficient accuracy, weak temporal dependency modeling and poor generalization performance in existing RUL prediction, and has significant advantages in prediction accuracy, model robustness and engineering practicality. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] Figure 1 This is a flow chart of the method for predicting the remaining useful life of an aircraft engine based on spatiotemporal state selection according to the present invention.

[0087] Figure 2 This is an example diagram of the sliding time window and RUL processing of the present invention.

[0088] Figure 3 This is a comparison chart of the predicted value and the actual value of the remaining useful life prediction of the FD001 sub-dataset in an embodiment of the present invention.

[0089] Figure 4 This is a comparison chart of the predicted value and the actual value of the remaining useful life prediction of the FD004 sub-dataset in an embodiment of the present invention. DETAILED DESCRIPTION

[0090] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0091] like Figure 1 As shown, a method for predicting the remaining useful life of an aircraft engine based on spatiotemporal state selection includes the following steps:

[0092] Step 1: Obtain aircraft engine sensor monitoring data and perform data selection and preprocessing. The specific process is as follows:

[0093] Step 1.1: Sensors are installed in various locations on the aircraft engine to collect operational parameters of key engine components throughout their lifecycle. Sensor monitoring data for aircraft engines includes temperature, pressure, speed, some cooling-related parameters, exhaust and emissions parameters, vibration and noise parameters, fuel system parameters, and lubricating oil system parameters. Temperature parameters include exhaust temperature, turbine inlet temperature, inter-stage turbine temperature, combustion chamber wall temperature, turbine blade temperature, lubricating oil temperature, fuel temperature, cooling air temperature, etc.; pressure parameters include total inlet pressure, compressor outlet / combustion chamber inlet pressure, turbine outlet pressure, fuel / lubricating oil pressure, fuel supply pressure, lubricating oil supply pressure, pressure difference monitoring, filter pressure difference, air cooling channel pressure difference, etc.; speed parameters include low-pressure rotor speed, high-pressure rotor speed, accessory gearbox speed, etc.; some cooling-related parameters include cooling air flow rate, cooling hole blockage rate, thermal barrier coating status, etc.; exhaust and emission parameters include exhaust pressure, pollutant concentration, exhaust velocity, etc.; vibration and noise parameters include blade passing frequency, bearing failure frequency, acoustic emission, etc.; fuel system parameters include fuel flow rate, fuel injection pressure / atomization particle size, etc.; lubricating oil system parameters include metal chip content, lubricating oil viscosity / acid value, etc.

[0094] In the above data, some key channels exhibit strong dynamic changes throughout the entire process. By analyzing the time-varying characteristics of each sensor data, invalid sensor channels with constant values ​​are eliminated, retaining valid sensor data with significant change trends. For data from different operating conditions, a differentiated processing strategy is adopted to select data based on their complexity. When the equipment operates under fixed conditions, only the data collected by a few key sensors that best reflect the equipment's aging status needs to be selected. However, when the equipment operates under complex and changing conditions, the complete data collected by all sensors must be retained to fully capture the impact of different operating conditions on the equipment.

[0095] Step 1.2: During the data preprocessing phase, the time series data collected by each sensor needs to be systematically processed to improve model performance. To increase data processing speed and training convergence, improving data processing efficiency and model generalization, the sensor monitoring data is normalized to the range [-1, 1] using the min-max scaling method.

[0096] Obtain sensor monitoring values ​​within different measurement ranges to form a sensor time series data set , For the The monitoring data of all sensors in time steps, is the total number of time steps, is the transpose symbol; ,in Indicates the The sensor in time step data, is the total number of sensors; use the following formula to normalize each data to the range of [-1,1].

[0097] ;

[0098] in, is the normalized value; and Represents the original data The maximum and minimum values ​​in the sensor data. Before normalization, the sensor has no obvious change trend and cannot provide effective degradation information for RUL prediction. In contrast, after normalization, the data shows a clear change trend, which is conducive to subsequent degradation feature learning.

[0099] Step 2: Use the sliding time window technology to construct multi-period multi-sensor samples as the input of the network model.

[0100] In practical applications, the degradation of aircraft engines usually begins after a period of use, which means that the RUL of the engine should remain unchanged in the early stages. Therefore, a piecewise linear degradation function is used to describe the degradation process of the engine. This function limits the maximum RUL of the engine and begins to degrade linearly after running for a period of time. When using the sliding time window technology to construct samples, the sensor monitoring data is divided into time windows of fixed length. The data in a time window is a multi-period multi-sensor sample. The time window slides along the time dimension, and the RUL corresponding to the last time step in each window will be used as the RUL label for all data in the current window, as shown in the attached figure. Figure 2 As shown in FIG, a complete dotted box is a time window, and different arrows indicate that the RUL of the last time step of each time window is used as the RUL of the current window. During the sliding process of the time window, it slides down one step each time.

[0101] Step 3. Construct the STSSFormer (Spatial-Temporal Selective State Transformer, STSSFormer) network model; the STSSFormer network model includes an embedding and position encoding module, a spatial and temporal feature extractor, and a predictor; the STSSFormer network model uses the embedding and position encoding module to process the input data, and uses the processed data as the input of the spatial and temporal feature extractor to extract features, and finally constructs a predictor based on the selective state space model to realize RUL prediction.

[0102] The specific working process of the STSSFormer network model is as follows:

[0103] Step 3.1: Input the multi-period multi-sensor samples into the embedding and position encoding module to obtain a sensor time series data set containing embedded position information. The embedding and position encoding module includes an embedding layer and a position encoding layer. The specific working process is as follows:

[0104] Step 3.1.1, embedding layer integrates sensor time series data Mapping to a higher dimension to get an embedding vector , the embedding layer makes the input more expressive, thus providing a more informative representation for subsequent feature extraction.

[0105] Step 3.1.2: Considering that the relative position of the current data plays a very important role in the RUL prediction task, position encoding is added to the embedding vector to preserve the sequential information of the time series data. Position encoding provides unique information for each time step, helping the model understand the relative and absolute position of the time step in the input data, so that the STSSFormer model can capture the temporal relationship and long-term dependency in the time series; position encoding is generated using sine and cosine functions, and the formula is:

[0106] ;

[0107] ;

[0108] in, Encode for position; is the position index; It is a dimension; is the dimension of the embedding vector;

[0109] Step 3.1.3. Finally, the embedding vector and the position code are added to obtain the sensor time series data set containing the embedded position information. :

[0110] ;

[0111] in, is the time window size.

[0112] Step 3.2: Input the sensor time series data set containing embedded location information into the spatial and temporal feature extractor to obtain the spatial and temporal fused features. This paper uses a temporal-spatial dual-branch encoding and hierarchical fusion mechanism to construct the spatial and temporal feature extractor, which specifically includes a spatial attention module (SAM), a temporal attention module (TAM), and a feature fusion layer. The number of encoder layers in each module is 2, and the multi-head attention mechanism includes four attention heads. The specific working process of the spatial and temporal feature extractor is as follows:

[0113] Step 3.2.1: Extract spatial features based on the spatial attention module. The spatial attention module is composed of multiple encoder layers. The core design of each encoder layer is as follows:

[0114] First, use depthwise convolution (DC) to convolution along the time dimension. After processing, three pointwise convolutions (PCs) are used to dynamically generate the query matrix, key matrix, and value matrix of self-attention. The formula is:

[0115] ;

[0116] ;

[0117] ;

[0118] ;

[0119] in, It is the feature representation after deep convolution processing in the time dimension; is the depth convolution; 、 、 These are the query matrix, key matrix, and value matrix generated by the spatial attention module respectively; is point-by-point convolution;

[0120] In order to improve the feature representation ability of the attention layer, the generated 、 、 The attention layer is split into multiple attention heads and the input is embedded into different subspaces to improve the performance of the attention layer. In order to reduce the computational complexity of the model, the self-attention mechanism in the traditional Transformer is improved to an efficient self-attention mechanism. The calculation process is as follows:

[0121] ;

[0122] ;

[0123] in, is the global context vector, 、 are the dimensions of key and value respectively; is the output of the attention head, is the batch size, is the number of attention heads. Through the efficient self-attention mechanism, the computational complexity of the model can be reduced from Reduce to .

[0124] The outputs of multiple attention heads are concatenated and then linearly mapped to obtain a multi-head feature map. , and use residual connections to combine them with Adding them preserves the original information and alleviates the vanishing gradient. Then, a normalization layer is used to normalize the results to improve training stability. The formula is:

[0125] ;

[0126] ;

[0127] in, For splicing; 、 、 、 Represents the output of four attention heads; is the weight matrix; is the output after normalization and residual connection; is the normalization layer.

[0128] Finally, the output of the current encoder layer is obtained through the feedforward network and used as the input of the next encoder layer. This is passed layer by layer until the spatial features output by the final spatial attention module are obtained:

[0129] ;

[0130] in, is the spatial feature; represents a feed-forward network.

[0131] Step 3.2.2: Extract temporal features based on the temporal attention module; the temporal attention module is also composed of multiple encoder layers stacked together. The structure of each encoder layer is similar to the spatial attention module. It calculates the query matrix, key matrix, and value matrix, uses an efficient attention mechanism to calculate temporal dependencies, and uses residual connections and normalization to stabilize training. The final temporal features output by the temporal attention module are obtained using a feedforward network and normalization layer. The temporal attention module and the spatial attention module are different in calculating the query matrix, key matrix, and value matrix. The temporal attention module obtains the three matrices through linear projection. The calculation process is as follows:

[0132] ;

[0133] ;

[0134] ;

[0135] in, 、 、 These are the query matrix, key matrix, and value matrix generated by the temporal attention module respectively; 、 、 They are 、 、 The weight matrix of .

[0136] In step 3.2.3, the temporal attention module focuses on the temporal dimension and models based on long-term temporal dependencies, while the spatial attention module focuses on the spatial dimension, enhancing local features through convolution and modeling based on local spatial correlations. Finally, a feature fusion layer is designed to fuse the features obtained by the spatial and temporal attention modules. The specific implementation is as follows:

[0137] ;

[0138] in, It is the feature after spatial and temporal fusion; 、 They are weight matrices for different features, which are two learnable weight coefficients, both initialized to 0.5 and dynamically optimized during training.

[0139] Step 3.3: predict RUL based on the predictor;

[0140] This paper introduces a Selective State Space Model (SelectiveSSM) to construct a predictor, which models the features extracted by the spatial and temporal feature extractors. This predictor decodes and predicts the fused features, outputting the engine's RUL. This predictor employs a cross-layer residual skip connection and parameter sharing mechanism. Specifically, it consists of two serially connected SelectiveSSMs, each layer representing the evolution of input features through dynamic state space modeling. The specific process is as follows:

[0141] Step 3.3.1: First, the features after spatial and temporal fusion Normalize and obtain the residual branch through linear transformation:

[0142] ;

[0143] in; is the residual branch; is a linear transformation; is the root mean square normalization;

[0144] Step 3.3.2, then branch the residual Input to the first layer SelectiveSSM. After the input projection, SelectiveSSM generates three vectors for state transfer: state step , state input weight , state output weight The state step is activated after linear projection, that is:

[0145] ;

[0146] in, is the state step after activation; is the activation function, ensuring is a positive value; is the weight matrix of the state step;

[0147] Define the state transition matrix , The initial value is a matrix of all 1s, and its subsequent update method is a learnable parameter with an initial value of ln(1) Decide:

[0148] ;

[0149] The SelectiveSSM core status update process is as follows:

[0150] ;

[0151] in, 、 Respectively , The hidden state of the time step, The time step represents the current moment, A time step represents the previous moment; The input sequence is Features at time steps; and For the Sutra The two different parameters after modulation are calculated as follows:

[0152] ;

[0153] ;

[0154] in, For the The time step time steps; For the The state step length after activation of time steps; is element-wise multiplication.

[0155] The final output of the first layer SelectiveSSM The calculation process is as follows:

[0156] ;

[0157] The output of the first layer After the activation function and Dropout layer, it is used as the input of the second layer SelectiveSSM. Finally, the output of the second layer SelectiveSSM and the residual branch After fusion, the predicted value of the remaining service life is obtained :

[0158] ;

[0159] in, is a learnable gating coefficient used to control the strength of the residual path information.

[0160] Step 4: Initialize the weights of the STSSFormer network model, use the mean square error to optimize the parameters of the network model, and train the network model using the multi-cycle multi-sensor samples built in step 2. The process of training the network model using samples and the network model configuration are as follows:

[0161] In order to avoid gradient explosion or gradient vanishing during training, the Xavier uniform initializer is used to initialize the weights of the linear layer and convolution layer in the model to ensure that the gradient remains stable between layers. The mean square error (MSE) is selected as the loss function of the model. The role of the loss function is to make the predicted value as close as possible to the true value. The calculation formula is as follows:

[0162] ;

[0163] in, Indicates the number of samples; and Corresponding to the The actual observed and predicted values ​​of the remaining useful life of samples are obtained. During the training process, the model parameters are optimized according to the loss function.

[0164] Step 5. After the training is completed, the trained STSSFormer network model is obtained. The current aircraft engine sensor monitoring data to be predicted is collected and processed according to the process of steps 1 and 2. The processed data is input into the trained STSSFormer network model to predict the remaining useful life of the aircraft engine in real time.

[0165] The present invention also proposes a system for predicting the remaining useful life of an aircraft engine based on spatiotemporal state selection, which includes the following modules:

[0166] A data selection and preprocessing module is used to select and preprocess the acquired aircraft engine sensor monitoring data;

[0167] Input sample construction module, which uses sliding time window technology to construct multi-period multi-sensor samples as input to the STSSFormer network model;

[0168] The model training module is used to initialize the weights of the STSSFormer network model and optimize the parameters of the network model using mean square error;

[0169] The remaining useful life prediction module uses the STSSFormer network model, which includes an embedding and position encoding module, a spatial and temporal feature extractor, and a predictor, to predict the remaining useful life of aircraft engines.

[0170] In addition, a model evaluation module is set up to compare the network model prediction value with the actual value to evaluate the network model performance.

[0171] In order to demonstrate the feasibility and superiority of the present invention, the following specific examples are given.

[0172] This embodiment uses the famous turbofan engine dataset (C-MAPSS) of NASA. The dataset contains four sub-datasets FD001, FD002, FD003, and FD004. Each sub-dataset corresponds to different operating conditions and fault states, as shown in Table 1. The rows corresponding to the training set and the test set represent the number of engines in the current dataset. Each sub-dataset contains a training set, a test set, and a real RUL. The data in the training set is used to train the model, and then the data in the test set is used to test the performance of the model. However, in the test set, not all sensor monitoring data are valid. The values ​​of some sensors are constant throughout the life cycle of the engine and cannot play a role in the RUL prediction process, so they need to be eliminated. Of the four sub-datasets, the monitoring data for sensors numbered 1, 5, 6, 10, 16, 18, and 19 in the FD001 and FD003 sub-datasets were removed. Since the FD002 and FD004 sub-datasets were collected under six operating conditions and were more complex than the other two sub-datasets, all sensor data in these two sub-datasets was retained. After selecting the sensor data, the min-max normalization method was used to normalize the sensor data to the range [-1, 1].

[0173] Table 1 C-MAPSS dataset

[0174] .

[0175] Following the process in step 2, a sliding time window technique is used to construct multi-period, multi-sensor samples as model input. The corresponding sliding window sizes vary for different data distributions and complexities. The time window size for the FD001 sub-dataset is 50, the time window size for the FD002 and FD004 sub-datasets is 60, and the time window size for the FD003 sub-dataset is 70. A piecewise linear degradation function is used to describe the engine degradation process, limiting the engine's maximum RUL and making the engine degradation process more consistent with actual scenarios. When using the sliding time window technique to construct samples, the data within a time window is a sample. The time window slides along the time dimension, moving forward one time step each time. The RUL corresponding to the last time step in each window will be used as the RUL label for all data in the current window.

[0176] Construct the STSSFormer network model according to the process of step 3. Then use Xavier Uniform to initialize the weights of the network model, use the mean square error to optimize the parameters of the network model, and use the constructed samples to train the network model. During the training process, traverse the data to load the batch data, use the Adam optimizer, set the learning rate to 1e-3, and use MSE as the loss function to guide parameter updates (perform back propagation to update model parameters). In this embodiment, after 30 rounds of training to obtain the optimal parameters, the test set is processed using the same processing method as the training set, and the processed test set samples are input into the trained network model. After obtaining the RUL predicted value, it is compared with the actual RUL value to evaluate the model performance. Specifically, the root mean square error (RMSE) and the score function are used as two evaluation indicators to verify the performance of the model. The formulas are defined as follows:

[0177] ;

[0178] ;

[0179] in, is the root mean square error; is the Score function value; Indicates the number of samples in the test set; Indicates the first The difference between the predicted RUL and the actual RUL for each engine. The RMSE directly reflects the absolute error in the predicted RUL; smaller values ​​indicate higher model accuracy. The Score function is an asymmetric evaluation metric designed to penalize late predictions. The penalty weight increases exponentially as the error increases. A late prediction occurs when the model predicts an engine RUL that is greater than the actual RUL, which can delay maintenance. If an engine fails before the predicted maintenance time, the consequences of not performing timely maintenance can be catastrophic.

[0180] In this embodiment, the comparison chart of the remaining useful life prediction values ​​and the actual values ​​of the two sub-data sets FD001 and FD004 is shown in the attached figure. Figure 3 and attached Figure 4 As shown. Figure 3 and Figure 4 It can be clearly seen that the actual RUL value of the present invention is basically consistent with the predicted value, and the present invention can achieve high-precision prediction of RUL.

[0181] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by technicians in this technical field within the essential scope of the present invention should also fall within the scope of protection of the present invention.

Claims

1. A method for predicting the remaining useful life of an engine based on spatiotemporal state selection, characterized in that: The steps include: Step 1: Acquire aircraft engine sensor monitoring data and perform data selection and preprocessing; Step 2: Use the sliding time window technique to construct a multi-period multi-sensor sample; Step 3: Construct the STSSFormer network model. The STSSFormer network model includes an embedding and positional encoding module, a spatial and temporal feature extractor, and a predictor. The STSSFormer network model uses the embedding and positional encoding module to process the input data, uses the processed data as input to the spatial and temporal feature extractor to extract features, and finally constructs a predictor based on the selective state space model to predict the remaining useful life of aircraft engines. The specific working process of the STSSFormer network model is as follows: Step 3.1: Input the multi-period multi-sensor samples into the embedding and position encoding module to obtain the sensor time series data set containing embedded position information. ; Step 3.2: Input the sensor time series data set containing embedded location information into the spatial and temporal feature extractor to obtain the spatial and temporal fused features; A temporal-spatial dual-branch encoding and hierarchical fusion mechanism is used to construct a spatial and temporal feature extractor, which includes a spatial attention module, a temporal attention module, and a feature fusion layer. The specific working process of the spatial and temporal feature extractor is as follows: Step 3.2.1: Extract spatial features based on the spatial attention module. The spatial attention module is composed of two stacked encoder layers. The core design of each encoder layer is as follows: First, we use depthwise convolution to After processing, three point-by-point convolutions are used to dynamically generate the query matrix, key matrix, and value matrix of self-attention. The formula is: ； ； ； ； in, It is the feature representation after deep convolution processing in the time dimension; is the depth convolution; 、 、 These are the query matrix, key matrix, and value matrix generated by the spatial attention module respectively; is point-by-point convolution; Will generate 、 、 Split into multiple attention heads and embed the input into different subspaces; improve the self-attention mechanism in the traditional Tranformer into an efficient self-attention mechanism. The calculation process is as follows: ； ； in, is the global context vector, 、 are the dimensions of key and value respectively; is the output of the attention head, is the batch size, is the number of attention heads; is the time window size; is the transpose symbol; The outputs of multiple attention heads are concatenated and then linearly mapped to obtain a multi-head feature map. , and use residual connections to combine them with Add them together, and then use the normalization layer to normalize the result. The formula is: ； ； in, For splicing; 、 、 、 Represents the output of four attention heads; is the weight matrix; is the output after normalization and residual connection; is the normalization layer; Finally, the output of the current encoder layer is obtained through the feedforward network and used as the input of the next encoder layer. This is passed layer by layer until the spatial features output by the final spatial attention module are obtained: ； in, is the spatial feature; represents a feedforward network; Step 3.2.2: Extracting temporal features based on temporal attention module The temporal attention module is also composed of multiple encoder layers stacked together. Each encoder layer calculates the query matrix, key matrix, and value matrix, and uses an efficient attention mechanism to calculate the temporal dependency. Through residual connection and normalization stabilization training, the feedforward network and normalization layer are used to obtain the temporal features output by the final temporal attention module. The temporal attention module obtains three matrices through linear projection. The calculation process is as follows: ； ； ； in, 、 、 These are the query matrix, key matrix, and value matrix generated by the temporal attention module respectively; 、 、 They are 、 、 The weight matrix of Step 3.2.3: Design a feature fusion layer to fuse the features obtained by the spatial attention module and the temporal attention module. The specific implementation is as follows: ； in, It is the feature after spatial and temporal fusion; 、 are the weight matrices of different features respectively; Step 3.3: Predict the remaining useful life of the aircraft engine based on the predictor. The predictor uses a cross-layer residual skip connection and parameter sharing mechanism and specifically includes two serially connected selective state space models. The specific working process of the predictor is as follows: Step 3.3.1: Features after spatial and temporal fusion Normalize and obtain the residual branch through linear transformation: ； in; is the residual branch; is a linear transformation; is the root mean square normalization; Step 3.3.2: Branch the residual Input to the first layer of selective state space model, the selective state space model generates three vectors for state transfer after input projection: state step , state input weight , state output weight ; The state step is activated after linear projection: ； in, is the state step after activation; is the activation function; is the weight matrix of the state step; Define the state transition matrix , The initial value is a full 1 matrix, and its subsequent update method is determined by the learnable parameters Decide: ； The hidden state update process of the selective state space model is: ； in, 、 Respectively , The hidden state of the time step; The input sequence is Features at time steps; and For the Sutra The two different parameters after modulation are calculated as follows: ； ； in, For the The time step time steps; For the The state step length after activation of time steps; is element-wise multiplication; The final output of the first layer selective state space model The calculation process is as follows: ； The output of the first layer selective state space model After the activation function and Dropout layer, it is used as the input of the second-layer selective state space model; finally, the output of the second-layer selective state space model and the residual branch After fusion, the predicted value of the remaining service life is obtained : ； in, is the learnable gating coefficient; Step 4: Initialize the weights of the STSSFormer network model, optimize the parameters of the network model using mean square error, and train the network model using the multi-cycle multi-sensor samples constructed in step 2; Step 5. After the training is completed, the trained STSSFormer network model is obtained. The current aircraft engine sensor monitoring data to be predicted is collected and processed according to the process of steps 1 and 2. The processed data is input into the trained STSSFormer network model to obtain the aircraft engine remaining service life prediction results in real time.

2. The engine remaining service life prediction method based on spatiotemporal state selection according to claim 1 is characterized in that: The specific process of step 1 is: Step 1.1: Obtain sensor monitoring values ​​within different measurement ranges of the aircraft engine to form a sensor time series data set , For the The monitoring data of all sensors in time steps, is the total number of time steps; ,in Indicates the The sensor in time step data, is the total number of sensors; the sensor monitoring data of aircraft engines includes temperature parameters, pressure parameters, speed parameters, some cooling-related parameters, exhaust and emission parameters, vibration and noise parameters, fuel system parameters, and lubricating oil system parameters; the time-varying characteristics of each sensor data are analyzed, invalid sensor channels with constant values ​​are eliminated, and valid sensor data with significant change trends are retained; for data under different operating conditions, a differentiated processing strategy is adopted for data selection based on their complexity differences. The differentiated processing strategy is: when the equipment is operating under fixed conditions, only the data collected by the key sensors that best reflect the aging of the equipment need to be selected; when the equipment is operating under complex and changeable conditions, the complete data collected by all sensors need to be retained; Step 1.2: In the data preprocessing stage, the sensor monitoring data is normalized to the range of [-1, 1] using the minimum and maximum scaling method.

3. The engine remaining service life prediction method based on spatiotemporal state selection according to claim 1 is characterized in that: In step 2, when using the sliding time window technology to construct the sample, the sensor monitoring data is divided into time windows of fixed length. The data in a time window is a multi-period multi-sensor sample; the time window slides along the time dimension, and the remaining service life corresponding to the last time step in each window will be used as the remaining service life label of all data in the current window.

4. The method for predicting the remaining service life of an engine based on spatiotemporal state selection according to claim 1, characterized in that: In step 3.1, the embedding and position encoding module includes an embedding layer and a position encoding layer. The specific working process is as follows: Step 3.1.1, embedding layer integrates sensor time series data Mapping to a higher dimension to get an embedding vector ; Step 3.1.2: Position code is generated using sine and cosine functions, using the formula: ； ； in, Encode for position; is the position index; It is a dimension; is the dimension of the embedding vector; Step 3.1.3: Add the embedding vector and the position code to obtain the sensor time series data set containing the embedded position information. : 。 5. The method for predicting the remaining service life of an engine based on spatiotemporal state selection according to claim 1, characterized in that: In step 4, the mean square error is selected as the loss function of the model. During the training process, the model parameters are optimized according to the loss function. The formula is as follows: ； in, Indicates the number of samples; and Corresponding to the The actual observed and predicted values ​​of the remaining useful life of the samples.

6. An engine remaining service life prediction system based on spatiotemporal state selection, characterized in that: The method for predicting the remaining service life of an engine based on spatiotemporal state selection according to any one of claims 1 to 5 is adopted, and the system includes the following modules: A data selection and preprocessing module is used to select and preprocess the acquired aircraft engine sensor monitoring data; Input sample construction module, which uses sliding time window technology to construct multi-period multi-sensor samples as input to the STSSFormer network model; The model training module is used to initialize the weights of the STSSFormer network model and optimize the parameters of the network model using mean square error; The remaining useful life prediction module uses the STSSFormer network model including embedding and position encoding modules, spatial and temporal feature extractors, and predictors to predict the remaining useful life of aircraft engines.

Images