Method and system for predicting remaining life of aircraft engines based on deep learning
By introducing multi-scale convolutional feature extraction and fusion feature attention GRU module of graph attention mechanism in the remaining life prediction of aircraft engines, the problems of insufficient feature spatial correlation and temporal dependency are solved, and more accurate life prediction is achieved.
Patent Information
- Application Number
- CN202411630892.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-15
- Publication Date
- 2025-09-09
- Estimated Expiration
- 2044-11-15
AI Technical Summary
Existing deep learning-based aero-engine remaining life prediction methods fail to adequately capture the spatial correlation and temporal dependency of features, which affects the prediction effect.
A multi-scale convolutional feature extraction module based on the graph attention mechanism and a GRU module fused with the feature attention mechanism are adopted to capture the correlation between feature maps through multi-scale convolution blocks, and dynamically adjust the feature weights at different time steps to build a more comprehensive life degradation model.
The accuracy of aircraft engine remaining life prediction is improved, and a more accurate life prediction model is established by more comprehensively capturing characteristic degradation information.
Smart Images
Figure CN119558184B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of engine remaining life prediction, and in particular to a method and system for predicting the remaining life of an aero-engine based on deep learning. Background Art
[0002] As one of the most critical aircraft components, the engine operates under complex conditions, making failures inevitable over long periods of operation. Real-time assessment of engine health and accurate prediction of its remaining useful life (RUL) can effectively reduce unplanned downtime, unnecessary maintenance, and excessive maintenance, and provide a reference for the development of subsequent maintenance strategies.
[0003] The performance and service life of aircraft engines are affected by numerous factors, including flight conditions, operating environment, and internal structure. Engine remaining life prediction involves processing and analyzing acquired sensor data. Then, using relevant models or machine learning methods, a mapping model is constructed between operating status data and remaining service life. This allows for a prediction of the engine's lifespan from its current operation to end of life. Currently, commonly used methods fall into two main categories: physics-based and data-driven. The former primarily characterizes degradation processes by building physical models based on the equipment's operating mechanisms. However, this method relies heavily on historical monitoring data, and for complex systems, insufficient expertise and the presence of uncertainties can limit the integrity and accuracy of the constructed models. Data-driven methods, on the other hand, are independent of physical principles and utilize statistical methods or machine learning algorithms to make direct predictions based on acquired sensor data. In recent years, with the advancement of sensor technology, aircraft engine life prediction requires processing vast amounts of degradation data. Deep learning methods, owing to their superior data processing and learning capabilities, have gained widespread application. At present, the relevant methods based on deep learning mainly use network structures such as CNN and LSTM to extract feature degradation information from a large amount of sensor data, capture the complex nonlinear relationship between sensor signal characteristics and remaining life, and thus realize the remaining life prediction of aircraft engines.
[0004] However, existing methods based on deep learning models typically employ channel attention or spatial attention mechanisms after convolution and pooling to enhance key feature information in different feature maps, and then fuse the features using simple concatenation or summation operations. However, these attention mechanisms are often limited to a single convolution kernel, thus ignoring the influence of the weight matrix in the convolution operation and the correlation between feature maps formed by convolution kernels of different scales. Furthermore, in terms of capturing temporal features, existing methods use the same feature extraction process at different time steps, ignoring the differences between different features at different time steps. When the spatial correlation and temporal dependency between features are not fully extracted, the model's prediction performance will be affected. Summary of the Invention
[0005] (1) Technical problems solved
[0006] In response to the shortcomings of the existing technology, the present invention provides a method and system for predicting the remaining life of an aircraft engine based on deep learning, which solves the technical problem that related methods based on deep learning models cannot adequately capture the spatial correlation and temporal dependency of features.
[0007] (2) Technical solution
[0008] To achieve the above objectives, the present invention is implemented through the following technical solutions:
[0009] In a first aspect, the present invention provides a method for predicting the remaining life of an aircraft engine based on deep learning, comprising:
[0010] Obtain historical engine operation data and in-service engine monitoring data, and pre-process them to obtain a data set;
[0011] The dataset is labeled based on the operating characteristics of aircraft engines, and the data in the dataset is divided into normal operation stage data and life degradation stage data;
[0012] The sliding window method is used to segment the data of adjacent time periods of each engine unit to construct a training sample data set;
[0013] Training a pre-built neural network using a training sample data set to obtain an aircraft engine remaining life prediction model, wherein the aircraft engine remaining life prediction model is used to predict the life of the aircraft engine;
[0014] Among them, the pre-constructed neural network includes a multi-scale convolutional feature extraction module based on the graph attention mechanism, a GRU module fused with the feature attention mechanism, and a RUL prediction module; the multi-scale convolutional feature extraction module based on the graph attention mechanism and the GRU module fused with the feature attention mechanism are operated in parallel, and the outputs of both are input into the RUL prediction module; the multi-scale convolutional feature extraction module based on the graph attention mechanism captures multi-level degradation features through several parallel convolution blocks of different sizes, and uses the graph attention mechanism to regard different filters as nodes, constructs multiple different graph structures, and assigns weights to each feature map according to the correlation between different feature maps; the GRU module fused with the feature attention mechanism dynamically adjusts the weights of the input features at different time steps to achieve differential updates of sequence data.
[0015] Preferably, the preprocessing includes feature selection and standardization;
[0016] in,
[0017] The feature selection includes: combining a correlation analysis method to perform feature selection on multidimensional variables to obtain variables related to the remaining life of the aircraft engine; the multidimensional variables include relevant sensor data and operating parameters;
[0018] The standardization process includes a normalization process.
[0019] Preferably, labeling the data set according to the operating characteristics of the aircraft engine, and dividing the data in the data set into normal operating stage data and life degradation stage data, includes:
[0020] The operation phases of aircraft engines are divided using the piecewise linear function method. According to their own operating characteristics, the aircraft engine operation cycle is divided into two phases: normal operation phase and life degradation phase. It is assumed that the engine has a constant RUL value in the normal operation phase:
[0021]
[0022] Among them, y label Indicates the label of the constructed sample data, y label =R early Indicates the normal operation stage, y label = RUL represents the life degradation stage, RUL represents the actual remaining service life of the aircraft engine, R early Indicates the set engine life threshold;
[0023] According to the division criteria, the data in the data set are processed and divided into normal operation stage data and life degradation stage data.
[0024] Preferably, the multi-scale convolutional feature extraction module based on the graph attention mechanism includes three parallel convolution blocks of different sizes and a graph attention mechanism;
[0025] Among them, three parallel convolution blocks of different sizes capture multi-level degenerate features, and use the graph attention mechanism to treat the convolution kernels in different convolution blocks as nodes to construct multiple different graph structures. Based on the correlation between different feature maps, each feature map is weighted;
[0026] Among them, the expression of CNN feature extraction is as follows:
[0027]
[0028] in, and is the weight matrix and bias matrix under the i-th convolution kernel, represents the feature map formed after the convolution operation, and σ(·) represents the activation function;
[0029] The weighting process of the graph attention mechanism is as follows:
[0030]
[0031] Among them, i and j represent nodes in the graph structure, and The feature vector representation of nodes i and j, e ij represents the attention coefficient of node i to node j, represents all adjacent nodes of node i, || represents vector splicing;
[0032] After obtaining the attention scores of all nodes, the softmax function is used to normalize them, obtain the weights of different nodes, and assign weights to them.
[0033] Preferably, the GRU module integrating the feature attention mechanism is a double-layer GRU network structure, and the feature attention mechanism is introduced after the first GRU, and the feature attention mechanism is combined with the hidden state h of the previous moment in the GRU. t-1 The historical running data retained in the , weights the different input features at the next moment according to their importance, and the weighted input variables are used as the input of the next layer of GRU; the operation process of GRU is as follows:
[0034] r t =σ(W r ·[h t-1 ,x t ]+b r )
[0035]
[0036] z t =σ(W z ·[h t-1 ,x t ]+b z )
[0037]
[0038] Among them, r t represents the output of the reset gate, z t represents the output of the update gate, σ(·) represents the activation function; W r 、W h and W z represents the weight matrix; b r 、b h and b z represents the bias matrix; h t Represents the information of the hidden state passed to the next moment.
[0039] Preferably, the RUL prediction module includes a splicing layer and a fully connected layer, and the two vector matrices are spliced together by the concat function in the splicing layer.
[0040] In the RUL prediction module, after obtaining the outputs of the SDMCN module and the F-GRU module, the output vectors of the multi-scale convolutional feature extraction module based on the graph attention mechanism and the GRU module with the fusion feature attention mechanism are matrix-concatenated through the concat function, and the concatenated vectors are input into the fully connected layer to obtain the engine remaining life prediction result.
[0041] In a second aspect, the present invention provides an aircraft engine remaining life prediction system based on deep learning, comprising:
[0042] The data acquisition module is used to obtain historical engine operation data and in-service engine monitoring data, and perform preprocessing to obtain a data set;
[0043] The data partitioning module is used to label the data set according to the operating characteristics of the aircraft engine, and divide the data in the data set into normal operation stage data and life degradation stage data;
[0044] A training sample construction module is used to segment the data of adjacent time periods of each engine unit using a sliding window method to construct a training sample data set;
[0045] A model training module is used to train a pre-built neural network using a training sample data set to obtain an aircraft engine remaining life prediction model, wherein the aircraft engine remaining life prediction model is used to predict the life of the aircraft engine;
[0046] Among them, the pre-constructed neural network includes a multi-scale convolutional feature extraction module based on the graph attention mechanism, a GRU module fused with the feature attention mechanism, and a RUL prediction module; the multi-scale convolutional feature extraction module based on the graph attention mechanism and the GRU module fused with the feature attention mechanism are operated in parallel, and the outputs of both are input into the RUL prediction module; the multi-scale convolutional feature extraction module based on the graph attention mechanism captures multi-level degradation features through several parallel convolution blocks of different sizes, and uses the graph attention mechanism to regard different filters as nodes, constructs multiple different graph structures, and assigns weights to each feature map according to the correlation between different feature maps; the GRU module fused with the feature attention mechanism dynamically adjusts the weights of the input features at different time steps to achieve differential updates of sequence data.
[0047] Preferably, labeling the data set according to the operating characteristics of the aircraft engine, and dividing the data in the data set into normal operating stage data and life degradation stage data, includes:
[0048] The operation phases of aircraft engines are divided using the piecewise linear function method. According to their own operating characteristics, the aircraft engine operation cycle is divided into two phases: normal operation phase and life degradation phase. It is assumed that the engine has a constant RUL value in the normal operation phase:
[0049]
[0050] Among them, y label Indicates the label of the constructed sample data, y label =R early Indicates the normal operation stage, y label = RUL represents the life degradation stage, RUL represents the actual remaining service life of the aircraft engine, R early Indicates the set engine life threshold;
[0051] According to the division criteria, the data in the data set are processed and divided into normal operation stage data and life degradation stage data.
[0052] In a third aspect, the present invention provides a computer-readable storage medium storing a computer program for predicting the remaining life of an aircraft engine based on deep learning, wherein the computer program enables a computer to execute the method for predicting the remaining life of an aircraft engine based on deep learning as described above.
[0053] In a fourth aspect, the present invention provides an electronic device, comprising:
[0054] One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, and the programs include a method for executing the deep learning-based aircraft engine remaining life prediction method as described above.
[0055] (3) Beneficial effects
[0056] This invention provides a method and system for predicting the remaining life of an aircraft engine based on deep learning. Compared with existing technologies, it has the following advantages:
[0057] In response to the problem of insufficient extraction of spatial correlation between features in existing methods, the present invention establishes a multi-scale convolution feature extraction module based on the graph attention mechanism and a GRU module that integrates the feature attention mechanism. Through the multi-scale feature extraction module, the spatial correlation of features is fully extracted from two perspectives: the correlation between feature maps formed under the same size convolution kernel (internal similarity) and the correlation between feature maps formed under different size convolution kernels (external similarity). At the same time, the F-GRU module is used to dynamically adjust the weights of input features at different time steps to achieve differential updates of sequence data. By adopting these two parallel branches, the extraction of spatial correlation and temporal dependency between features is achieved, thereby more comprehensively capturing feature degradation information, establishing a more accurate life degradation model, and improving the accuracy of model prediction. BRIEF DESCRIPTION OF THE DRAWINGS
[0058] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.
[0059] Figure 1 This is a block diagram of the deep learning-based aircraft engine remaining life prediction method;
[0060] Figure 2 This is a detailed flowchart of the remaining life prediction method for aircraft engines based on deep learning;
[0061] Figure 3 A schematic diagram of a segmentation process for segmenting data of adjacent time periods of each engine unit using a sliding window method;
[0062] Figure 4 This is a schematic diagram of the structure of the aircraft engine remaining life prediction model;
[0063] Figure 5This is a schematic diagram of the structure of the multi-scale convolutional feature extraction module based on the graph attention mechanism;
[0064] Figure 6 Schematic diagram of the structure of the GRU module that integrates the feature attention mechanism;
[0065] Figure 7a to Figure 7d All of them are comparison charts of prediction effects, among which Figure 7a The predicted and true values for the engine unit #76 on the four data subsets are: Figure 7b The predicted and true values for the engine unit #64 on the four data subsets are: Figure 7c The predicted and true values for the engine unit #99 on the four data subsets are: Figure 7d The predicted and true values for the engine unit #47 on the four data subsets. DETAILED DESCRIPTION
[0066] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention are clearly and completely described. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.
[0067] The embodiments of the present application provide a method and system for predicting the remaining life of an aircraft engine based on deep learning, addressing the technical issues of insufficiently capturing the spatial correlation between different feature maps and the temporal dependencies of input features in related methods based on deep learning models. This method simultaneously addresses the correlation between feature maps formed with convolution kernels of the same size and those formed with convolution kernels of different sizes, as well as the dynamic updating of input feature weights at different time steps. This method more comprehensively captures feature degradation information from both temporal and spatial perspectives, enabling more accurate life prediction.
[0068] The technical solution in the embodiments of the present application is to solve the above technical problems, and the overall idea is as follows:
[0069] Existing deep learning-based methods primarily utilize network structures such as CNN, LSTM, and CNN-BiGRU to extract feature degradation information from large amounts of sensor data, capturing the complex nonlinear relationship between sensor signal characteristics and remaining life, thereby enabling aircraft engine remaining life prediction. However, existing technologies have the following drawbacks:
[0070] (1) When using CNN networks for feature extraction, a single-scale convolution kernel is often unable to fully obtain feature information. In addition, existing model methods usually use channel attention mechanisms or spatial attention mechanisms to enhance the key feature information in different feature maps after convolution and pooling operations, and then use simple splicing or summation operations to perform feature fusion. However, these attention mechanisms are limited to a single convolution kernel, thus ignoring the influence of the weight matrix in the convolution operation and the correlation between features of different scales. When the spatial correlation between features is not fully extracted, the prediction effect of the model will be affected to a certain extent.
[0071] (2) Traditional methods, when using the Gated Recurrent Unit (GRU) model to learn time series data, focus on the temporal dependency of sensor data, that is, assigning different weights to features at different time steps to highlight the importance of data at different time steps in the RUL prediction process. The feature update process of GRU is indifferent across the entire data sequence, that is, all features at each time step are processed in the same way. However, in actual situations, different feature vectors in each time step play different roles in the RUL prediction process, especially in the critical stage of engine life degradation. Therefore, if all features at different time periods are processed in the same way, it will affect the capture of key feature degradation information, thereby affecting the accuracy of model prediction.
[0072] In order to better understand the above technical solution, the above technical solution will be described in detail below with reference to the accompanying drawings and specific implementation methods.
[0073] The embodiment of the present invention provides a method for predicting the remaining life of an aircraft engine based on deep learning, such as Figure 1 As shown, the method includes:
[0074] S1. Obtain historical engine operation data and in-service engine monitoring data, and pre-process them to obtain a data set;
[0075] S2. Label the dataset based on the operating characteristics of the aircraft engine, dividing the data in the dataset into normal operation phase data and life degradation phase data;
[0076] S3. Segment the data of adjacent time periods of each engine unit using a sliding window method to construct a training sample data set;
[0077] S4. Training a pre-built neural network using a training sample data set to obtain an aircraft engine remaining life prediction model, and predicting the aircraft engine using the aircraft engine remaining life prediction model to obtain a prediction result;
[0078] The pre-built neural network includes a multi-scale convolutional feature extraction module based on the graph attention mechanism, a GRU module fused with the feature attention mechanism, and a RUL prediction module. The multi-scale convolutional feature extraction module based on the graph attention mechanism and the GRU module fused with the feature attention mechanism are operated in parallel, and the outputs of both are input into the RUL prediction module. The multi-scale convolutional feature extraction module based on the graph attention mechanism captures multi-level degraded features through several parallel convolution blocks of different sizes, and uses the graph attention mechanism to treat different filters as nodes, constructing multiple different graph structures, and assigning weights to each feature graph based on the correlation between different feature graphs. The GRU module fused with the feature attention mechanism dynamically adjusts the weights of the input features at different time steps to achieve differential updates of sequence data.
[0079] Firstly, in order to address the problem of insufficient extraction of spatial correlation between features in existing methods, the embodiment of the present invention establishes a multi-scale convolution feature extraction module based on the graph attention mechanism, which fully extracts the spatial correlation of features from two perspectives: the correlation between feature maps formed under the same-sized convolution kernels (internal similarity) and the correlation between feature maps formed under different-sized convolution kernels (external similarity). Secondly, in order to address the shortcomings of existing methods in capturing temporal dependencies of features, the feature attention mechanism is integrated with GRU to achieve differentiated updates of input features at different time steps. Through two parallel branches, feature degradation information is captured more comprehensively, thereby improving the accuracy of the prediction model for the remaining life of aircraft engines.
[0080] The following is combined with Figure 2 The flowchart shown describes the individual steps in detail:
[0081] In step S1, historical engine operation data and in-service engine monitoring data are obtained and pre-processed to obtain a data set. The specific implementation process is as follows:
[0082] Historical engine operation data and in-service engine monitoring data both include relevant sensor data and operating parameters. Preprocessing mainly includes feature selection and standardization. Specifically, it includes:
[0083] By combining correlation analysis with multidimensional variables such as relevant sensor data and operating parameters, we selected features to identify variables related to the remaining life of aircraft engines. These variables include engine operating parameters, high- and low-pressure compressor temperature parameters, and fan speed. The selected variables were standardized using the "min-max" normalization method, using the following formula:
[0084]
[0085] Among them, xi For each sensor value, X max and X min are the maximum and minimum values of each column of sensor variables, respectively, x i ' is the corresponding normalized variable value.
[0086] The standardized variables related to aircraft engine remaining life form a dataset, which includes a training set and a test set. The training set is generated by preprocessing historical engine operation data, while the test set is generated by preprocessing in-service engine monitoring data. The training set is used for model training and parameter adjustment, while the test set is used to verify the effectiveness of the trained model.
[0087] In step S2, the dataset is labeled according to the operating characteristics of the aircraft engine, and the data in the dataset is divided into normal operation stage data and life degradation stage data. The specific implementation process is as follows:
[0088] The operation phases of aircraft engines are divided using the piecewise linear function method. According to their own operating characteristics, the aircraft engine operation cycle is divided into two phases: normal operation phase and life degradation phase. It is assumed that the engine has a constant RUL value in the normal operation phase:
[0089]
[0090] Among them, y label Indicates the label of the constructed sample data, RUL represents the actual remaining service life of the aircraft engine, R early Indicates the set engine life threshold.
[0091] According to this division standard, the data in the data set are processed and divided into normal operation stage data and life degradation stage data.
[0092] The purpose of dividing the data in the dataset into normal operation stage data and life degradation stage data is to better perform prediction processing. If the data is not divided, it is equivalent to accurately predicting the data at each moment. After the division, the predicted label is equivalent to a piecewise linear degradation.
[0093] In step S3, the data of adjacent time periods of each engine unit are segmented using the sliding window method to construct a training sample data set. The specific implementation process is as follows:
[0094] Specific segmentation methods such as Figure 3As shown in the figure, the time series features between individual independent data are not obvious, so here the data of each engine is processed by sliding window, and the data within a certain interval is packaged as one data. The sliding step is 1, W represents the sliding window size, and the segmented data is used as the training sample data set.
[0095] In step S4, the pre-built neural network is trained using the training sample data set to obtain an aircraft engine remaining life prediction model, which is used to predict the life of the aircraft engine. The specific implementation process is as follows:
[0096] The structure of the pre-built neural network in the embodiment of the present invention (the structure of the trained aircraft engine remaining life prediction model is consistent with it, at least the parameters of the model are different) is as follows: Figure 4 As shown in Figure 1, it includes two parallel multi-scale convolutional feature extraction modules based on graph attention mechanism (SDMCN modules), a GRU module fused with feature attention mechanism (F-GRU module) and a RUL prediction module. Figure 5 As shown, the GRU module that integrates the feature attention mechanism is as follows Figure 6 As shown in Figure 2, the SDMCN module and the F-GRU module are run in parallel, and the outputs of both are input into the RUL prediction module.
[0097] like Figure 5 As shown, the multi-scale convolution feature extraction module based on the graph attention mechanism captures multi-level degradation features through several (three in the embodiment of the present invention) parallel convolution blocks of different sizes, and uses the graph attention mechanism to treat different filters as nodes (i.e., feature maps formed after feature convolution operations with different weights under different convolution kernels and the same convolution kernel), constructs multiple different graph structures, and assigns weights to each feature map based on the correlation between different feature maps. Among them, the expression of CNN feature extraction is as follows:
[0098]
[0099] in, and is the weight matrix and bias matrix under the i-th convolution kernel, represents the feature map formed after the convolution operation, and σ(·) represents the activation function.
[0100] The weighting process of GAT (graph attention mechanism) is as follows
[0101]
[0102] Among them, i and j represent nodes in the graph structure (each convolution kernel is regarded as a node in the graph structure), and The feature vector representation of nodes i and j, e ij represents the attention coefficient of node i to node j, Represents all adjacent nodes of node i. || represents vector concatenation. After obtaining the attention scores of all nodes, the softmax function is used to normalize them, obtain the weights of different nodes, and assign weights to them.
[0103] like Figure 6 As shown in the figure, the GRU module integrating the feature attention mechanism is a two-layer GRU network structure. The feature attention mechanism is introduced after the first GRU. With the help of the feature attention mechanism, the hidden state h of the previous moment in the GRU is combined with the t-1 The retained historical running data is used to assign weights to different input features at the next moment according to their importance, and then the weighted input variables are used as the input of the next layer of GRU. The operation process of GRU is as follows:
[0104] r t =σ(W r ·[h t-1 ,x t ]+b r )
[0105]
[0106] z t =σ(W z ·[h t-1 ,x t ]+b z )
[0107]
[0108] Among them, r t represents the output of the reset gate, z t represents the output of the update gate, σ(·) represents the activation function; W r 、W h and W z represents the weight matrix; b r 、b h and b z represents the bias matrix; h t Represents the information of the hidden state passed to the next moment.
[0109] In the RUL prediction module, after obtaining the outputs of the SDMCN module and the F-GRU module, the two vector matrices are concatenated using the concat function. They are then input into the fully connected layer to obtain the engine remaining life prediction result RUL. pre .
[0110] out con =[out conv ,out gru ]
[0111] RUL pre =LeakyReLU(W out con +b)
[0112] The model is trained using the training set, and the grid search method is used to optimize the parameters to determine the best parameter set. The effectiveness of the trained model is verified through the test set, and the trained aircraft engine remaining life prediction model is obtained.
[0113] The effectiveness of the embodiment of the present invention is verified by simulation experiments below:
[0114] The dataset C-MAPSS selected in this simulation experiment was developed by NASA using commercial modular simulation software in the MATLAB environment. It simulates the entire degradation process data of a turbofan engine from operation to failure. It consists of four data subsets FD001-FD004, each of which has 26 columns of data (engine ID, current number of operating cycles, 3 operating settings, and 21 columns of sensor signal monitoring data).
[0115] Step 1: Plot the curves of each variable in the training set as it changes over the life of the aircraft engine. Combined with correlation analysis and other methods, 17 columns of variable data were selected as model inputs (operating condition 1, operating condition 2, sensor 2, 3, 4, 6, 7, 8, 9, 11, 12, 13, 14, 15, 17, 20, and 21). Each column of variables was then normalized using the minimum-maximum normalization method.
[0116] Step 2: Segment the lifespan prediction labels in the training set and the test set. In this invention, R early =120.
[0117] Step 3: Analyze the training sets in the four data subsets in C-MAPSS respectively, select appropriate time window sizes W1, W2, W3, and W4 for the training sets according to the minimum number of cycles in the corresponding test sets, set the step size to 1, and construct N×W i The sample input of (i=1,2,3,4) has the corresponding predicted label y t+W-1 The test set is processed in the same way, but unlike the training set, the test sample only uses the last sliding window data of each engine unit.
[0118] X t =[x1,x2,......,x N ]
[0119]
[0120] Among them, X t and X t' They represent the original input data and the preprocessed input data sequence at the t-th time step, N represents the feature dimension, and T represents the length of the time window.
[0121] Step 4: Select RMSE and Score as the evaluation indicators of the model prediction effect. RMSE is used to measure the gap between the model prediction value and the true value, and Score reflects the penalty value of the model for different prediction situations (advance prediction or delayed prediction). The principle is as follows:
[0122]
[0123] Where N is the total number of samples, is the predicted value of engine life, is the actual value of engine life, d i Indicates the difference between the predicted value and the true value.
[0124] Step 5: Use the training data processed in Step 3 as the model input data. Select three different convolution kernel sizes (3, 5, and 7) to construct the remaining life prediction model and perform parameter optimization. This validation process determines the optimal parameter set based on four different data subsets and fine-tunes them. Some hyperparameter values are shown in Table 1:
[0125] Table 1 Some hyperparameter values
[0126]
[0127] Step 6: The four test subsets are respectively put into the model trained in step 5 for verification, and the model's predicted value of the remaining service life of each engine after random stop is obtained, and it is compared with the true label. After multiple independent repeated experiments, the RMSE and Score mean on the four test sets are obtained and compared with the prediction effect of the SOTA method respectively to analyze the effectiveness of the method proposed in this invention.
[0128] Step 7: To more intuitively demonstrate the prediction effect of the model, plot the predicted values and true values of the four data subsets for engine units #76, #64, #99, and #47 for visual analysis, as shown in the following figure: Figure 7a to Figure 7d shown.
[0129] An embodiment of the present invention further provides an aircraft engine remaining life prediction system based on deep learning, comprising:
[0130] The data acquisition module is used to obtain historical engine operation data and in-service engine monitoring data, and perform preprocessing to obtain a data set;
[0131] The data partitioning module is used to label the data set according to the operating characteristics of the aircraft engine, and divide the data in the data set into normal operation stage data and life degradation stage data;
[0132] A training sample construction module is used to segment the data of adjacent time periods of each engine unit using a sliding window method to construct a training sample data set;
[0133] A model training module is used to train a pre-built neural network using a training sample data set to obtain an aircraft engine remaining life prediction model, wherein the aircraft engine remaining life prediction model is used to predict the life of the aircraft engine;
[0134] The pre-built neural network includes a multi-scale convolutional feature extraction module based on the graph attention mechanism, a GRU module fused with the feature attention mechanism, and a RUL prediction module. The multi-scale convolutional feature extraction module based on the graph attention mechanism and the GRU module fused with the feature attention mechanism are operated in parallel, and the outputs of both are input into the RUL prediction module. The multi-scale convolutional feature extraction module based on the graph attention mechanism captures multi-level degraded features through several parallel convolution blocks of different sizes, and uses the graph attention mechanism to treat different filters as nodes, constructing multiple different graph structures, and assigning weights to each feature graph based on the correlation between different feature graphs. The GRU module fused with the feature attention mechanism dynamically adjusts the weights of the input features at different time steps to achieve differential updates of sequence data.
[0135] It can be understood that the data center energy system security risk assessment system provided by the embodiment of the present invention corresponds to the above-mentioned data center energy system security risk assessment method. The explanation, examples, beneficial effects, etc. of its relevant contents can refer to the corresponding contents in the data center energy system security risk assessment method, and will not be repeated here.
[0136] An embodiment of the present invention further provides a computer-readable storage medium storing a computer program for data center energy system security risk assessment, wherein the computer program enables a computer to execute the data center energy system security risk assessment method as described above.
[0137] An embodiment of the present invention also provides an electronic device, comprising: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, and the programs include a method for executing the data center energy system security risk assessment method as described above.
[0138] In summary, compared with the existing technology, the present invention has the following beneficial effects:
[0139] In order to address the problem of insufficient extraction of spatial correlation between features in existing methods, the embodiments of the present invention establish a multi-scale convolution feature extraction module based on a graph attention mechanism and a GRU module that integrates a feature attention mechanism. Through the multi-scale feature extraction module, the spatial correlation of features is fully extracted from two perspectives: the correlation between feature maps formed under convolution kernels of the same size (internal similarity) and the correlation between feature maps formed under convolution kernels of different sizes (external similarity). At the same time, the GRU module that integrates a feature attention mechanism is used to dynamically adjust the weights of input features at different time steps to achieve differential updates of sequence data. These two parallel branches are used to realize the extraction of spatial correlation and temporal dependency between features, thereby more comprehensively capturing feature degradation information, establishing a more accurate life degradation model, and improving the accuracy of model prediction.
[0140] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply the existence of any such actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or device comprising the element.
[0141] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.
Claims
1. A method for predicting the remaining life of an aircraft engine based on deep learning, characterized in that: include: Obtain historical engine operation data and in-service engine monitoring data, and pre-process them to obtain a data set; The dataset is labeled based on the operating characteristics of aircraft engines, and the data in the dataset is divided into normal operation phase data and life degradation phase data; The sliding window method is used to segment the data of adjacent time periods of each engine unit to construct a training sample data set; Training a pre-built neural network using a training sample data set to obtain an aircraft engine remaining life prediction model, wherein the aircraft engine remaining life prediction model is used to predict the life of the aircraft engine; Among them, the pre-constructed neural network includes a multi-scale convolutional feature extraction module based on the graph attention mechanism, a GRU module fused with the feature attention mechanism, and a RUL prediction module. The multi-scale convolutional feature extraction module based on the graph attention mechanism and the GRU module fused with the feature attention mechanism are parallel, and the outputs of both are input into the RUL prediction module; the multi-scale convolutional feature extraction module based on the graph attention mechanism captures multi-level degradation features through several parallel convolution blocks of different sizes, and uses the graph attention mechanism to regard different filters as nodes, constructs multiple different graph structures, and assigns weights to each feature graph according to the correlation between different feature graphs; the GRU module fused with the feature attention mechanism dynamically adjusts the weights of the input features at different time steps to achieve differential updates of sequence data.
2. The method for predicting the remaining life of an aircraft engine based on deep learning according to claim 1, characterized in that: The preprocessing includes feature selection and standardization; in, The feature selection includes: combining a correlation analysis method to perform feature selection on multidimensional variables to obtain variables related to the remaining life of the aircraft engine; the multidimensional variables include relevant sensor data and operating parameters; The standardization process includes a normalization process.
3. The method for predicting the remaining life of an aircraft engine based on deep learning according to claim 1, wherein: The data set is labeled according to the operating characteristics of the aircraft engine, and the data in the data set is divided into normal operation stage data and life degradation stage data, including: The operation phases of aircraft engines are divided using the piecewise linear function method. According to their own operating characteristics, the aircraft engine operation cycle is divided into two phases: normal operation phase and life degradation phase. It is assumed that the engine has a constant RUL value in the normal operation phase: Among them, y label Indicates the label of the constructed sample data, y label =R early Indicates the normal operation stage, y label = RUL represents the life degradation stage, RUL represents the actual remaining service life of the aircraft engine, R early Indicates the set engine life threshold; According to the division criteria, the data in the data set are processed and divided into normal operation stage data and life degradation stage data.
4. The method for predicting the remaining life of an aircraft engine based on deep learning according to any one of claims 1 to 3, characterized in that: The multi-scale convolutional feature extraction module based on the graph attention mechanism includes three parallel convolution blocks of different sizes and a graph attention mechanism; Among them, three parallel convolution blocks of different sizes capture multi-level degenerate features, and use the graph attention mechanism to treat the convolution kernels in different convolution blocks as nodes to construct multiple different graph structures. Based on the correlation between different feature maps, each feature map is weighted; Among them, the expression of CNN feature extraction is as follows: in, and is the weight matrix and bias matrix under the i-th convolution kernel, represents the feature map formed after the convolution operation, and σ(·) represents the activation function; The weighting process of the graph attention mechanism is as follows: Among them, i and j represent nodes in the graph structure, and The feature vector representation of nodes i and j, e ij represents the attention coefficient of node i to node j, represents all adjacent nodes of node i, || represents vector splicing; After obtaining the attention scores of all nodes, the softmax function is used to normalize them, obtain the weights of different nodes, and assign weights to them.
5. The method for predicting the remaining life of an aircraft engine based on deep learning according to any one of claims 1 to 3, characterized in that: The GRU module integrating the feature attention mechanism is a two-layer GRU network structure. The feature attention mechanism is introduced after the first GRU. The feature attention mechanism is combined with the hidden state h of the previous moment in the GRU. t-1 The historical running data retained in the , weights the different input features at the next moment according to their importance, and the weighted input variables are used as the input of the next layer of GRU; the operation process of GRU is as follows: r t =σ(W r ·[h t-1 ,x t ]+b r ) z t =σ(W z ·[h t-1 ,x t ]+b z ) Among them, r t represents the output of the reset gate, z t represents the output of the update gate, σ(·) represents the activation function; W r 、W h and W z represents the weight matrix; b r 、b h and b z represents the bias matrix; h t Represents the information of the hidden state passed to the next moment.
6. The method for predicting the remaining life of an aircraft engine based on deep learning according to any one of claims 1 to 3, characterized in that: The RUL prediction module includes a splicing layer and a fully connected layer. The concat function in the splicing layer is used to splice the two vector matrices. In the RUL prediction module, after obtaining the outputs of the SDMCN module and the F-GRU module, the output vectors of the multi-scale convolutional feature extraction module based on the graph attention mechanism and the GRU module with the fusion feature attention mechanism are matrix-concatenated through the concat function, and the concatenated vectors are input into the fully connected layer to obtain the engine remaining life prediction result.
7. A deep learning-based aircraft engine remaining life prediction system, characterized by: include: The data acquisition module is used to obtain historical engine operation data and in-service engine monitoring data, and perform preprocessing to obtain a data set; The data partitioning module is used to label the data set according to the operating characteristics of the aircraft engine, and divide the data in the data set into normal operation stage data and life degradation stage data; A training sample construction module is used to segment the data of adjacent time periods of each engine unit using a sliding window method to construct a training sample data set; A model training module is used to train a pre-built neural network using a training sample data set to obtain an aircraft engine remaining life prediction model, wherein the aircraft engine remaining life prediction model is used to predict the life of the aircraft engine; Among them, the pre-constructed neural network includes a multi-scale convolutional feature extraction module based on the graph attention mechanism, a GRU module fused with the feature attention mechanism, and a RUL prediction module; the multi-scale convolutional feature extraction module based on the graph attention mechanism and the GRU module fused with the feature attention mechanism are operated in parallel, and the outputs of both are input into the RUL prediction module; the multi-scale convolutional feature extraction module based on the graph attention mechanism captures multi-level degradation features through several parallel convolution blocks of different sizes, and uses the graph attention mechanism to regard different filters as nodes, constructs multiple different graph structures, and assigns weights to each feature map according to the correlation between different feature maps; the GRU module fused with the feature attention mechanism dynamically adjusts the weights of the input features at different time steps to achieve differential updates of sequence data.
8. The deep learning-based aircraft engine remaining life prediction system according to claim 7, characterized in that: The data set is labeled according to the operating characteristics of the aircraft engine, and the data in the data set is divided into normal operation stage data and life degradation stage data, including: The operation phases of aircraft engines are divided using the piecewise linear function method. According to their own operating characteristics, the aircraft engine operation cycle is divided into two phases: normal operation phase and life degradation phase. It is assumed that the engine has a constant RUL value in the normal operation phase: Among them, y label Indicates the label of the constructed sample data, y label =R early Indicates the normal operation stage, y label = RUL represents the life degradation stage, RUL represents the actual remaining service life of the aircraft engine, R early Indicates the set engine life threshold; According to the division criteria, the data in the data set are processed and divided into normal operation stage data and life degradation stage data.
9. A computer-readable storage medium, characterized in that It stores a computer program for predicting the remaining life of an aircraft engine based on deep learning, wherein the computer program enables a computer to execute the method for predicting the remaining life of an aircraft engine based on deep learning as described in any one of claims 1 to 6.
10. An electronic device, characterized in that: include: One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, and the programs include a method for executing the deep learning-based aircraft engine remaining life prediction method according to any one of claims 1 to 6.