Method and system for predicting residual service life of engine based on space-time state selection

Abstract

The invention discloses an engine residual service life prediction method and system based on space-time state selection, and belongs to the field of equipment residual service life prediction. According to the method, multiple technologies such as a time sequence-space double-branch coding and hierarchical fusion mechanism, a selective state space model, a cross-layer residual jump connection and parameter sharing mechanism and an efficient self-attention mechanism are integrated, and an STSSFormer network model is constructed to predict the residual service life of the aero-engine; the system mainly comprises a data selection and preprocessing module, an input sample construction module, a model training module and a residual service life prediction module. According to the method, the feature expression and fusion capability is comprehensively enhanced, and the method has remarkable advantages in the aspects of prediction accuracy of the remaining service life, model robustness and engineering practicability.

Description

Technical Field

[0001] The present invention belongs to the field of predicting the remaining useful life of equipment, and particularly relates to a method and system for predicting the remaining useful life of an engine based on spatio-temporal state selection. Background Art

[0002] With the advent of the big data era and the rise of low-cost computing power, modern industries have put forward higher requirements for the reliability and predictive maintenance of complex systems. As an important part of Predictive Health Management (PHM), the Remaining Useful Life (RUL) has been widely applied to various equipment and systems, especially in those fields that require regular maintenance and management. For example, in industries such as aviation, energy, batteries, and mechanical equipment, by analyzing the sensor data monitoring the operation status of equipment, the RUL of the equipment can be estimated, so as to achieve preventive maintenance and avoid economic losses caused by unexpected failures.

[0003] However, due to the uncertainty of the equipment operation 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 aero engines. These methods generally include physical degradation models constructed based on the equipment failure principle, or by designing corresponding machine learning algorithms to learn the degradation trend in sensor data to predict RUL. However, with the development of industry, the complexity of mechanical systems and the changes in the operation environment make it increasingly difficult to accurately construct physical degradation models. Therefore, data-driven methods have received more and more attention and have made remarkable progress in practical applications.

[0004] The design based on the Transformer model usually processes temporal and spatial features independently. In multi-dimensional data, there is a lack of an effective cross-modal fusion mechanism, resulting in information loss or error accumulation. At the same time, when extracting spatial features, a traditional convolutional network is used. Traditional convolution usually requires processing a large number of parameters and calculations. Especially when the number of channels of the input feature map and the size of the convolution kernel are large, the amount of calculation increases sharply; existing solutions often use fixed weighting coefficients or simple splicing methods to combine spatio-temporal features, and cannot dynamically adjust the weights of temporal and spatial features according to different input data, which may lead to some key features not being fully concerned, thus affecting the model performance; the computational complexity of the Transformer increases quadratically with the increase of the sequence length and is not suitable for processing long-time sequence 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 conversion between state spaces. Summary of the Invention

[0005] To solve the above problems, the present invention proposes a method and system for predicting the remaining useful life of an aeroengine based on spatio-temporal state selection, which combines multiple technologies such as time-series and spatial double-branch coding and hierarchical fusion mechanism, selective state space model, cross-layer residual skip connection and parameter sharing mechanism, and efficient self-attention mechanism, constructs a STSSFormer (Spatial-Temporal Selective State Transformer) network model for predicting the remaining useful life of an aeroengine, and effectively realizes the real-time health management of an aeroengine and reduces the maintenance cost.

[0006] The technical solution of the present invention is as follows: A method for predicting the remaining useful life of an engine based on spatio-temporal state selection, comprising the following steps: Step 1, obtain the sensor monitoring data of the aeroengine, and perform data selection and preprocessing; Step 2, use the sliding time window technology to construct multi-period multi-sensor samples; Step 3, construct a 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 predict the remaining useful life of the aeroengine; Step 4, initialize the weights of the STSSFormer network model, use the mean square error to optimize the parameters of the network model, and use the multi-period multi-sensor samples constructed in Step 2 to train the network model; Step 5, after the training is completed, obtain the trained STSSFormer network model, collect the sensor monitoring data of the current aeroengine to be predicted and process it according to the processes of Step 1 and Step 2, input the processed data into the trained STSSFormer network model, and obtain the prediction result of the remaining useful life of the aeroengine in real time.

[0007] Further, the specific process of Step 1 is as follows: Step 1.1, obtain the sensor monitoring values within different measurement ranges of the aeroengine to form a sensor time series data set , is the monitoring data of all sensors at the th time step, is the total number of time steps, is the transpose symbol; , where represents the The data of a sensor at the th time step, where is the total number of sensors; The sensor monitoring data of the aero-engine includes temperature parameters, pressure parameters, rotational speed parameters, some cooling-related parameters, exhaust and emission parameters, vibration and noise parameters, fuel system parameters, and lubricating oil system parameters; Analyze the time-varying characteristics of each sensor's data, eliminate the invalid sensor channels with constant values, and retain the effective sensor data with significant change trends; For the data under different working conditions, adopt a differential processing strategy for data selection according to their complexity differences. The differential processing strategy is as follows: When the equipment operates under fixed working conditions, only the data collected by the key sensors that can best reflect the equipment aging situation need to be selected; When the equipment operates under complex and changeable working conditions, all the complete data collected by all sensors need to be retained; Step 1.2, In the data preprocessing stage, use the min-max scaling method to normalize the sensor monitoring data to the range of [-1, 1].

[0008] Furthermore, in the step 2, when using the sliding time window technique to construct samples, divide the sensor monitoring data into time windows of a fixed length. The data within one time window is a multi-period multi-sensor sample; The time window slides along the time dimension, and the remaining useful life corresponding to the last time step within each window will be used as the remaining useful life label for all the data within the current window.

[0009] Furthermore, in the step 3, the specific working process of the STSSFormer network model is as follows: Step 3.1, Input the multi-period multi-sensor sample into the embedding and position encoding module to obtain a set of sensor time series data containing embedded position information; Step 3.2, Input the set of sensor time series data containing embedded position information into the spatial and temporal feature extractor to obtain the features after spatial and temporal fusion; Step 3.3, Predict the remaining useful life of the aero-engine based on the predictor.

[0010] Furthermore, in the 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: Step 3.1.1, The embedding layer maps the set of sensor time series data to a higher dimension to obtain an embedding vector ; Step 3.1.2, The position encoding is generated using sine and cosine functions, and the formula is: ; ; Among them, is the positional encoding; is the position index; is the dimension; is the dimension of the embedding vector; Step 3.1.3: Add the embedding vector and the positional encoding to obtain a set of sensor time series data containing embedded position information : ; Among them, is the time window size.

[0011] Furthermore, in the said Step 3.2, a temporal-spatial double-branch encoding and hierarchical fusion mechanism is adopted 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: Step 3.2.1: Extract spatial features based on the spatial attention module; the spatial attention module is stacked by two encoder layers, and the core design of each encoder layer is as follows: First, use depth convolution to process along the time dimension, and then use three pointwise convolutions to dynamically generate the query matrix, key matrix, and value matrix of self-attention. The formula is: ; ; ; ; Among them, is the feature representation after depth convolution processing along the time dimension; is the depth convolution; , , are respectively the query matrix, key matrix, and value matrix generated by the spatial attention module; is the pointwise convolution; The generated , , are split into multiple attention heads, and the input is embedded into different subspaces; the self-attention mechanism in the traditional Transformer is improved to an efficient self-attention mechanism, and the calculation process is as follows: ; ; Among them, is the global context vector, , are respectively the dimensions of the key and the value; is the output of the attention head, is the batch size, is the number of attention heads; The outputs of multiple attention heads are concatenated and then passed through a linear mapping to obtain the multi-head feature map , and it is added to using a residual connection, and then the result is normalized using a normalization layer. The formula is: ; ; where, is the concatenation; , , , represent the outputs 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 a feed-forward network and used as the input to the next encoder layer, and it is passed layer by layer until the spatial feature output by the final spatial attention module is obtained: ; where, is the spatial feature; represents the feed-forward network; Step 3.2.2. Extract temporal features based on the temporal attention module ; The principle of the temporal attention module is basically the same as that of the spatial attention module, except that when 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: ; ; ; where, , , are the query matrix, key matrix, and value matrix generated by the temporal attention module, respectively; , , are respectively , , 's weight matrices; 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 method is as follows: ; Among them, is the feature after spatial and temporal fusion; , are the weight matrices of different features respectively.

[0012] Furthermore, in step 3.3, the predictor adopts a cross-layer residual skip connection and parameter sharing mechanism, specifically including two cascaded selective state space models; the specific working process of the predictor is as follows: Step 3.3.1, normalize the feature after spatial and temporal fusion, and obtain the residual branch through linear transformation: ; Among them; is the residual branch; is the linear transformation; is the root mean square normalization; Step 3.3.2, input the residual branch into the first-layer selective state space model. After input projection by the selective state space model, three vectors for state transition are generated: state step , state input weight , state output weight ; among them, the state step is activated after linear projection: ; Among them, is the activated state step; is the activation function; is the weight matrix of the state step; Define the state transition matrix , The initial value of is the all-1 matrix, and its subsequent update method is determined by the learnable parameter : ; The hidden state update process of the selective state space model is: ; Among them, , are the th, th hidden states at the input sequence at the th time step; and are two different parameters modulated by , and their calculation formulas are: ; ; Among them, is the th time step of the th time step; is the state step size after activation at the th time step; 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 passing through the activation function and the Dropout layer, it serves 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 are fused to obtain the predicted value of the remaining useful life : ; Among them, is the learnable gating coefficient.

[0013] Furthermore, in step 4, the mean squared 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 mean squared error The formula is as follows: ; Among them, represents the number of samples; and respectively correspond to the actual observed value and the predicted value of the remaining useful life of the th sample.

[0014] An engine remaining useful life prediction system based on spatio-temporal state selection adopts the engine remaining useful life prediction method based on spatio-temporal state selection as described above. This system includes the following modules: Data selection and preprocessing module, which is used to perform data selection and preprocessing on the acquired aero-engine sensor monitoring data; Input sample construction module, which uses the sliding time window technique to construct multi-period multi-sensor samples as the input of the STSSFormer network model; Model training module, which is used to initialize the weights of the STSSFormer network model and optimize the parameters of the network model using the mean squared error; The remaining useful life prediction module uses the STSSFormer network model including an embedding and position encoding module, a spatio-temporal feature extractor, and a predictor to predict the remaining useful life of an aero-engine.

[0015] The beneficial technical effects brought by the present invention: Through the innovative spatio-temporal decoupled parallel encoding structure, selective state space modeling, and efficient attention mechanism, the present invention comprehensively enhances the feature expression and fusion ability, effectively overcomes the problems of insufficient accuracy, weak temporal dependence modeling, and poor generalization performance in existing RUL predictions, and has significant advantages in prediction accuracy, model robustness, and engineering practicability. Description of the Drawings

[0016] Figure 1 It is a flowchart of the method for predicting the remaining useful life of an aero-engine based on spatio-temporal state selection according to the present invention.

[0017] Figure 2 It is an example diagram of the sliding time window and RUL processing according to the present invention.

[0018] Figure 3 It is a comparison diagram of the predicted value and the true value of the remaining useful life prediction of the FD001 sub-dataset in the embodiment of the present invention.

[0019] Figure 4 It is a comparison diagram of the predicted value and the true value of the remaining useful life prediction of the FD004 sub-dataset in the embodiment of the present invention. Detailed Embodiments

[0020] The present invention will be further described in detail below in conjunction with the drawings and specific embodiments: As Figure 1 shown, a method for predicting the remaining useful life of an aero-engine based on spatio-temporal state selection includes the following steps: Step 1: Obtain the monitoring data of the aero-engine sensors and perform data selection and preprocessing; the specific process is as follows: Step 1.1: Sensors are installed at different parts of the aero-engine to collect the operation status parameters of key engine components throughout their entire life cycle. The sensor monitoring data of the aero-engine includes temperature parameters, pressure parameters, speed parameters, some cooling-related parameters, exhaust and emission parameters, vibration and noise parameters, fuel system parameters, lubricating oil system parameters, etc. 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, differential pressure monitoring, filter differential pressure, air cooling channel differential pressure, 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 speed, etc.; vibration and noise parameters include blade passing frequency, bearing fault 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.

[0021] Among the above data, some key channels have strong dynamic change characteristics throughout the process. By analyzing the time-varying characteristics of the data of each sensor, the invalid sensor channels with constant values are removed, and the effective sensor data with significant change trends is retained. For the data under different working conditions, different processing strategies are adopted for data selection according to their complexity differences. The different processing strategies are as follows: when the equipment operates under fixed working conditions, only the data collected by several key sensors that can best reflect the equipment aging situation need to be selected; while when the equipment operates under complex and changeable working conditions, the complete data collected by all sensors need to be retained so as to comprehensively capture the impact of different operating states on the equipment.

[0022] Step 1.2: In the data preprocessing stage, the time series data collected by each sensor needs to be systematically processed to improve the model performance. In order to improve the data processing speed and accelerate the convergence speed of training, and improve the efficiency of data processing and the generalization ability of the model, the min-max (minimum-maximum) scaling method is used to normalize the sensor monitoring data to the range of [-1, 1].

[0023] Obtain the sensor monitoring values within different measurement ranges to form a sensor time series data set , is the monitoring data of all sensors at the th time step, is the total number of time steps, is the transpose symbol; , where Indicates the data of the th sensor at the th time step, where [[ID=]] is the total number of sensors; the following formula is used to normalize each data to the range of [-1, 1].

[0024] ; where, is the value after normalization; and respectively represent the maximum and minimum values in the data of the th sensor in the original data. There is no obvious change trend in the sensors before normalization, and no effective degradation information can be provided for RUL prediction. In contrast, the data shows an obvious change trend after normalization, which is beneficial to subsequent degradation feature learning.

[0025] Step 2: Use the sliding time window technique to construct multi-period multi-sensor samples as the input of the network model.

[0026] In practical applications, the degradation of aeroengines usually starts after being used for a period of time, that is, the RUL of the engine should remain unchanged in the early stage. Therefore, a piecewise linear degradation function is used to describe the degradation process of the engine, which limits the maximum RUL of the engine and starts to degrade linearly after running for a period of time. When using the sliding time window technique to construct samples, the sensor monitoring data is divided into time windows of a fixed length, and the data within one 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 the data in the current window. As shown in Figure 2 the figure, a complete dashed box is a time window, and different arrows indicate that the RUL of the last time step in 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.

[0027] 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, takes 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 achieve RUL prediction.

[0028] 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 a set of sensor time series data containing embedded position information; the embedding and position encoding module includes an embedding layer and a position encoding layer, and the specific working process is as follows: Step 3.1.1: The embedding layer maps the set of sensor time series data into a higher dimension to obtain an embedding vector . The embedding layer enables the input to have a richer expressive ability, thus providing a more informative representation for subsequent feature extraction.

[0029] 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 retain the order information of the time series data. The position encoding provides unique information for each time step, helping the model understand the relative and absolute positions of the time steps in the input data, so that the STSSFormer model can capture the temporal relationship and long-term dependence in the time series; the position encoding is generated using sine and cosine functions, and the formula is: ; ; where, is the position encoding; is the position index; is the dimension; is the dimension of the embedding vector; Step 3.1.3: Finally, the embedding vector and the position encoding are added together to obtain a set of sensor time series data containing embedded position information : ; where, is the time window size.

[0030] Step 3.2: Input the set of sensor time series data containing embedded position information into the spatial and temporal feature extractor to obtain the features after spatial and temporal fusion; the present invention adopts 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 contains four attention heads. 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 stacked by multiple encoder layers, and the core design of each encoder layer is as follows: First, use depthwise convolution (DC) to process along the time dimension. After that, use three pointwise convolutions (PC) to dynamically generate the query matrix, key matrix, and value matrix of self-attention. The formula is: ; ; ; ; where, is the feature representation after depthwise convolution processing along the time dimension; is the depthwise convolution; , , are the query matrix, key matrix, and value matrix generated by the spatial attention module respectively; is the pointwise convolution; To enhance the feature representation ability of the attention layer, the generated , , are split into multiple attention heads and the input is embedded into different subspaces to improve the performance of the attention layer. To reduce the computational complexity of the model, the self-attention mechanism in the traditional Transformer is improved to an efficient self-attention mechanism. Its calculation process is as follows: ; ; where, is the global context vector, , are the dimensions of the 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 to .

[0031] Concatenate the outputs of multiple attention heads and then obtain the multi-head feature map through linear mapping. Use residual connection to add it to to retain the original information and alleviate the vanishing gradient. Then use the normalization layer to normalize the result to improve the training stability. The formula is: ; ; Among them, is splicing; 、 、 、 represent the outputs of four attention heads; is the weight matrix; is the output after normalization and residual connection; is the normalization layer.

[0032] Finally, the output of the current encoder layer is obtained through the feed-forward network and used as the input of the next encoder layer, and is passed layer by layer until the spatial features output by the final spatial attention module are obtained: ; Among them, is the spatial feature; represents the feed-forward network.

[0033] Step 3.2.2: Extract temporal features based on the temporal attention module; the temporal attention module is also stacked by multiple encoder layers. The structure of each encoder layer is similar to that of the spatial attention module, calculating the query matrix, key matrix, and value matrix, using the efficient attention mechanism to calculate the temporal dependence relationship, stabilizing the training through residual connection and normalization, and using the feed-forward network and normalization layer to obtain the temporal features output by the final temporal attention module . The temporal attention module is different from the spatial attention module when calculating the query matrix, key matrix, and value matrix. The temporal attention module obtains the three matrices through linear projection, and its calculation process is as follows: ; ; ; Among them, 、 、 are the query matrix, key matrix, and value matrix generated by the temporal attention module respectively; 、 、 are respectively 、 、 weight matrices.

[0034] Step 3.2.3: The attention of the temporal attention module focuses on the time dimension and models based on long-term temporal dependencies; while the attention of the spatial attention module focuses on the spatial dimension, enhances local features through convolution, and models based on local spatial correlations. Finally, a feature fusion layer is designed to fuse the features obtained from the spatial attention module and the temporal attention module. The specific implementation method is as follows: ; Among them, is the feature after spatial and temporal fusion; 、 are the weight matrices of different features respectively, which are two learnable weight coefficients, both initialized to 0.5, and dynamically optimized during training.

[0035] Step 3.3: Predict the RUL based on the predictor; The present invention introduces a Selective State Space Model (SelectiveSSM) to construct a predictor, which is used to model the features extracted by the spatial and temporal feature extractors, that is, to decode and predict the fused features, and output the RUL of the engine. This predictor adopts a cross-layer residual skip connection and parameter sharing mechanism, which specifically includes two cascaded SelectiveSSMs, and each layer represents the evolution pattern of the input features through state space dynamic modeling. The specific process is as follows: Step 3.3.1: First, normalize the features after spatial and temporal fusion, and obtain the residual branch through linear transformation: ; Among them; is the residual branch; is the linear transformation; is the root mean square normalization; Step 3.3.2: Then input the residual branch into the first layer of SelectiveSSM. After input projection by SelectiveSSM, three vectors for state transition are generated: the state step , the state input weight , and the state output weight . Among them, the state step is activated after linear projection, that is: ; Among them, is the activated state step; is the activation function to ensure that is a positive value; is the weight matrix of the state step; Define the state transition matrix , The initial value of is a matrix of all 1s, and its subsequent update method is determined by the learnable parameter whose initial value is ln(1): ; The core state update process of SelectiveSSM is as follows: ; Among them, and are the hidden states at the th and th time steps respectively. The th time step represents the current moment, and the th time step represents the previous moment; is the feature of the input sequence at the th time step; and are two different parameters modulated by , and their calculation formulas are: ; ; Among them, is the th of the th time step; is the state step size activated at the th time step; is the element-wise multiplication.

[0036] The final output of the first layer of SelectiveSSM is calculated as follows: ; The output of the first layer, after passing through the activation function and the Dropout layer, is used as the input of the second layer of SelectiveSSM. Finally, after fusing the output of the second layer of SelectiveSSM and the residual branch , the predicted value of the remaining useful life is obtained: ; Among them, is the learnable gating coefficient used to control the strength of the residual path information.

[0037] Step 4: Initialize the weights of the STSSFormer network model, optimize the parameters of the network model using the mean squared error, and train the network model using the multi-period multi-sensor samples constructed in Step 2. The process of training the network model using the samples and the network model configuration are as follows: To avoid gradient explosion or gradient disappearance during training, use the Xavier uniform initializer to initialize the weights of the linear and convolutional layers in the model, ensuring that the gradients remain stable between layers. Select the mean squared error (MSE) as the loss function of the model. The role of the loss function is to make the predicted value infinitely close to the true value. The mean squared error The calculation formula is as follows: ; where, represents the number of samples; and correspond to the actual observed value and the predicted value of the remaining useful life of the th sample respectively. During the training process, optimize the model parameters according to the loss function.

[0038] Step 5: After the training is completed, obtain the trained STSSFormer network model, collect the monitoring data of the aero-engine sensors to be predicted currently, and process it according to the processes of Step 1 and Step 2. Input the processed data into the trained STSSFormer network model to predict the remaining useful life prediction result of the aero-engine in real time.

[0039] The present invention also proposes an aero-engine remaining useful life prediction system based on spatio-temporal state selection, including the following modules: Data selection and preprocessing module, used to perform data selection and preprocessing on the obtained monitoring data of aero-engine sensors; Input sample construction module, which uses the sliding time window technology to construct multi-period multi-sensor samples as the input of the STSSFormer network model; Model training module, used to initialize the weights of the STSSFormer network model and optimize the parameters of the network model using the mean squared error; Remaining useful life prediction module, which uses the STSSFormer network model including the embedding and position encoding module, spatial and temporal feature extractor, and predictor to predict the remaining useful life of the aero-engine; In addition, a model evaluation module is also set up to compare the predicted value of the network model with the true value to evaluate the performance of the network model.

[0040] To prove the feasibility and superiority of the present invention, the following specific embodiments are given.

[0041] This embodiment uses the famous turbofan engine dataset (C-MAPSS) of the National Aeronautics and Space Administration of the United States. This 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 specifically. 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 the true 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, the monitoring data of not all sensors is 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. Therefore, they need to be removed. In the four sub-datasets, the monitoring data of sensors numbered 1, 5, 6, 10, 16, 18, and 19 in the FD001 and FD003 sub-datasets are removed. Since the FD002 and FD004 sub-datasets are data collected under 6 operating conditions and are more complex compared to the other two sub-datasets, all sensor data in these two sub-datasets are retained. After selecting the sensor data, the min-max normalization method is used to normalize the sensor data to the range of [-1, 1].

[0042] Table 1 C-MAPSS dataset 。

[0043] According to the process in Step 2, the sliding time window technique is used to construct multi-period multi-sensor samples as the input of the model. The corresponding sliding window sizes are different for different data distributions and complexities. Among them, the time window size corresponding to the FD001 sub-dataset is 50, the time window sizes corresponding to the FD002 and FD004 sub-datasets are 60, and the time window size corresponding to the FD003 sub-dataset is 70. The piecewise linear degradation function is used to describe the degradation process of the engine, and the maximum RUL of the engine is restricted to make the engine degradation process more in line with the actual scenario. When using the sliding time window technique to construct samples, the data within one time window is one sample. The time window slides along the time dimension, moving forward one time step each time. The RUL corresponding to the last time step within each window will be used as the RUL label for all data within the current window.

[0044] Construct the STSSFormer network model according to the process of step 3. Then initialize the weights of the network model using Xavier Uniform, optimize the parameters of the network model using mean squared error, and train the network model using the constructed samples. During the training process, iterate through the batch data in the data loader, use the Adam optimizer with a learning rate set to 1e-3, and use MSE as the loss function to guide parameter updates (perform backpropagation to update the model parameters). In this embodiment, after training for 30 epochs to obtain the optimal parameters, process the test set using the same processing method as the training set. After inputting the processed test set samples into the trained network model to obtain the RUL prediction values, compare them with the true RUL values to evaluate the model performance. Specifically, use two evaluation metrics, Root Mean Square Error (RMSE) and Score function, to verify the performance of the model. Their formula definitions are as follows: ; ; where, is the Root Mean Square Error; is the value of the Score function; represents the number of samples in the test set; represents the difference between the predicted value and the true value of the RUL of the th engine in the test set. Among them, RMSE directly reflects the absolute error of the predicted RUL, and the smaller the value, the higher the model accuracy; the Score function is an asymmetric evaluation metric, aiming to punish late predictions. As the error increases, the penalty weight will increase exponentially. Late prediction means that the RUL of the engine predicted by the model is greater than the actual RUL, which will cause maintenance actions to be delayed. If the engine fails before the predicted maintenance time point, it will lead to catastrophic consequences due to untimely maintenance.

[0045] In this embodiment, the comparison graphs of the predicted values and the true values of the remaining useful life predictions for the two sub-datasets FD001 and FD004 are shown in Appendix Figure 3 and Appendix Figure 4 . It can be clearly seen from Figure 3 and Figure 4 that the true value and the predicted value of the RUL of the present invention are basically in agreement, and the present invention can achieve high-precision prediction of the RUL.

[0046] 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 those skilled in the art within the scope of the essence of the present invention should also fall within the protection scope of the present invention.

Claims

1. A method for predicting the remaining useful life of an engine based on spatio-temporal state selection, characterized in that It includes the following steps: Step 1: Obtain the monitoring data of the aero-engine sensors, and perform data selection and preprocessing; Step 2: Use the sliding time window technique to construct multi-period multi-sensor samples; Step 3: Construct the 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, takes 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 predict the remaining useful life of the aero-engine; Step 4: Initialize the weights of the STSSFormer network model, use the mean square error to optimize the parameters of the network model, and use the multi-period multi-sensor samples constructed in Step 2 to train the network model; Step 5: After the training is completed, obtain the trained STSSFormer network model, collect the monitoring data of the aero-engine sensors to be predicted currently and process them according to the processes of Step 1 and Step 2, input the processed data into the trained STSSFormer network model, and obtain the prediction result of the remaining useful life of the aero-engine in real time.

2. The method for predicting the remaining service life of an engine based on spatio-temporal state selection according to claim 1, characterized in that, The specific process of Step 1 is as follows: Step 1.1: Obtain the sensor monitoring values within different measurement ranges of the aero-engine to form a sensor time series data set , is the monitoring data of all sensors at the -th time step, is the total number of time steps, is the transpose symbol; , where represents the data of the -th sensor at the -th time step, is the total number of sensors; the sensor monitoring data of the aero-engine includes temperature parameters, pressure parameters, rotational speed parameters, some cooling-related parameters, exhaust and emission parameters, vibration and noise parameters, fuel system parameters, and lubricating oil system parameters; analyze the time-varying characteristics of each sensor data, eliminate the invalid sensor channels with constant values, and retain the effective sensor data with significant change trends; for the data under different working conditions, adopt a differentiated processing strategy for data selection according to their complexity differences. The differentiated processing strategy is: when the equipment operates under fixed working conditions, only select the data collected by the key sensors that can best reflect the aging of the equipment; when the equipment operates under complex and variable working conditions, retain the complete data collected by all sensors; Step 1.2: In the data preprocessing stage, use the min-max scaling method to normalize the sensor monitoring data to the range of [-1, 1].

3. The method for predicting the remaining service life of an engine based on spatio-temporal state selection according to claim 1, wherein In Step 2, when using the sliding time window technique to construct samples, divide the sensor monitoring data into time windows of a fixed length, and the data within one time window is a multi-period multi-sensor sample; the time window slides along the time dimension, and the remaining useful life corresponding to the last time step within each window will be used as the remaining useful life label of all the data within the current window.

4. The method for predicting the remaining service life of an engine based on spatio-temporal state selection according to claim 1, characterized in that, In Step 3, 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 a set of sensor time series data containing embedding position information; Step 3.2: Input the set of sensor time series data containing embedding position information into the spatial and temporal feature extractor to obtain the spatially and temporally fused features; Step 3.3: Predict the remaining useful life of the aero-engine based on the predictor.

5. The method for predicting the remaining service life of an engine based on spatio-temporal state selection according to claim 4, characterized in that, 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: Step 3.1.1: The embedding layer maps the sensor time series data set to a higher dimension to obtain an embedding vector ; Step 3.1.2: The position encoding is generated using sine and cosine functions, and the formula is: ； ； Among them, is the positional encoding; is the position index; is the dimension; is the dimension of the embedding vector; Step 3.1.3: Add the embedding vector and the positional encoding to obtain a set of sensor time series data containing embedded position information : ； Among them, is the time window size.

6. The method for predicting the remaining service life of an engine based on spatio-temporal state selection according to claim 5, wherein, In Step 3.2, adopt a temporal-spatial double-branch encoding and hierarchical fusion mechanism to construct the 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: Step 3.2.1: Extract spatial features based on the spatial attention module; the spatial attention module is composed of two encoder layers stacked together, and the core design of each encoder layer is as follows: First, use depth convolution to process along the time dimension. Then, use three pointwise convolutions to dynamically generate the query matrix, key matrix, and value matrix of self-attention. The formula is as follows: ； ； ； ； Among them, is the feature representation after depth convolution processing in the time dimension; is the depth convolution; , , are the query matrix, key matrix, and value matrix generated by the spatial attention module respectively; is the pointwise convolution; The generated , , are split into multiple attention heads and the input is embedded into different subspaces; the self-attention mechanism in the traditional Transformer is improved to an efficient self-attention mechanism, and the calculation process is as follows: ； ； Among them, is the global context vector, , are the dimensions of the key and value respectively; is the output of the attention head, is the batch size, is the number of attention heads; Concatenate the outputs of multiple attention heads and then obtain the multi-head feature map through a linear mapping , use residual connection to add it to , and then use a normalization layer to normalize the result. The formula is as follows: ； ； Among them, is for splicing; , , , represent the outputs 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 feed-forward network and used as the input of the next encoder layer, and is passed layer by layer until the spatial features output by the final spatial attention module are obtained: ； Among them, is a spatial feature; represents a feedforward network; Step 3.2.2: Extract temporal features based on the temporal attention module ; The principle of the temporal attention module is basically the same as that of the spatial attention module, except that they are different when calculating the query matrix, key matrix, and value matrix. The temporal attention module obtains the three matrices through linear projection, and its calculation process is as follows: ； ； ； Among them, , , are the query matrix, key matrix, and value matrix generated by the temporal attention module, respectively; , , are the , , weight matrices, respectively. 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 method is as follows: ； Among them, is the feature after spatial and temporal fusion; , are the weight matrices of different features respectively.

7. The method for predicting the remaining service life of an engine based on spatio-temporal state selection according to claim 6, wherein, In step 3.3, the predictor adopts a cross-layer residual skip connection and parameter sharing mechanism, which specifically includes two cascaded selective state space models; the specific working process of the predictor is: Step 3.3.

1. Normalize the features after spatio-temporal fusion and obtain the residual branch through linear transformation: ； Wherein; is the residual branch; is a linear transformation; is root mean square normalization; Step 3.3.2: Input the residual branch into the first-layer selective state space model. After input projection by the selective state space model, three vectors for state transition are generated: the state step , the state input weight , and the state output weight ; among them, the state step is activated after linear projection: ； Among them, is the state step size after activation; is the activation function; is the weight matrix of the state step size; Define the state transition matrix , The initial value of which is a matrix of all 1s, and its subsequent update method is determined by the learnable parameter : ； The hidden state update process of the selective state space model is: ； Among them, and are the hidden states at the -th and -th time steps respectively; is the feature of the input sequence at the -th time step; and are two different parameters modulated by , and their calculation formulas are: ； ； Among them, is the th time step of the th time step; is the state step size after activation at the th time step; is element-wise multiplication; Final output of the first-layer selective state space model The calculation process is as follows: ； Output of the first-layer selective state space model After passing through the activation function and the Dropout layer, it serves as the input to the second-layer selective state space model; finally, after fusing the output of the second-layer selective state space model and the residual branch the predicted value of the remaining useful life is obtained : ； Among them, is a learnable gating coefficient.

8. The method for predicting the remaining service life of an engine based on spatio-temporal state selection according to claim 4, characterized in that, In step 4, the mean squared error is selected as the loss function of the model. During the training process, the model parameters are optimized according to the loss function, and the mean squared error The formula is as follows: ； Among them, represents the number of samples; and respectively correspond to the actual observed value and the predicted value of the remaining service life of the th sample.

9. An engine remaining service life prediction system based on spatio-temporal state selection, characterized in that, Using the method for predicting the remaining useful life of an engine based on spatio-temporal state selection according to any one of claims 1-8, the system includes the following modules: A data selection and preprocessing module for data selection and preprocessing of the acquired aero-engine sensor monitoring data; An input sample construction module that uses a sliding time window technique to construct multi-period multi-sensor samples as the input of the STSSFormer network model; A model training module for initializing the weights of the STSSFormer network model and using the mean square error to optimize the parameters of the network model; A remaining useful life prediction module that uses the STSSFormer network model including an embedding and position encoding module, a spatial and temporal feature extractor, and a predictor to predict the remaining useful life of the aero-engine.

Images