Aero-engine residual life prediction method based on deep learning and ensemble learning

Abstract

The invention belongs to the field of aero-engine residual life prediction, and particularly relates to an aero-engine residual life prediction method based on deep learning and integrated learning. The method comprises the following steps: acquiring aero-engine data, and preprocessing the aero-engine data; constructing an aero-engine data deep learning model by adopting a recurrent neural network added with an attention mechanism and ensemble learning; and inputting the aero-engine data into the aero-engine data deep learning model, and outputting a prediction result of the residual life of the aero-engine. The attention mechanism and the integrated learning are introduced on the basis of the classical deep learning model recurrent neural network, the attention mechanism enables the model to pay attention to the most important time point for life prediction in the time sequence data, so that the prediction precision of the model is improved, and the integrated learning reduces the variance of a single model, so that the prediction precision of the model is improved. And the stability and generalization ability of the whole model are further improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of aero-engine residual life prediction, and particularly relates to an aero-engine residual life prediction method based on deep learning and ensemble learning. BACKGROUND

[0002] An aero-engine is a core component of an airplane, accounting for about 20% to 30% of the value of the whole airplane, and the design and development level and manufacturing process of the aero-engine directly affect the performance and reliability of the airplane, and failure of the aero-engine is not allowed to occur during flight. Since the whole engine system structure is extremely complex and is subjected to a harsh working environment, main working parts bear high centrifugal load, aerodynamic load and vibration alternating load, and are also subjected to impact of foreign objects, which may cause accidents endangering the safety of the engine and the airplane, and therefore, it is particularly important to perform health management (PHM) on the aero-engine.

[0003] Residual useful life (RUL) prediction is an important work in the PHM technology. The engine RUL is the time from the current running time t of the engine to the time T when the engine runs to a failure state, i.e., a fault occurs. Prediction of the engine RUL can help maintenance personnel to better make maintenance plans, so as to timely perform necessary maintenance and replacement of parts, thereby increasing the service life of the engine and improving the reliability and safety of the equipment.

[0004] There are three existing aero-engine residual useful life prediction methods: an experience-based prediction method, a physical model-based prediction method and a data-driven prediction method.

[0005] 1. The experience-based prediction method, also known as the fault prediction method based on an expert system, is a program system with a large amount of accumulated knowledge and experience. The method uses computer technology and artificial intelligence technology, and according to professional knowledge and development experience provided by experts in a certain field, finally realizes simulation of decision-making of human experts, and solves various complex problems that need to be handled by human experts through the above method. In short, an expert system is a computer program system simulating human experts to solve problems in a field.

[0006] 2. The physical model-based prediction method. The physical model has been verified by a large amount of data, and has high accuracy. The method takes various electronic product failure models as basic data sources, calculates the cumulative damage value through the environmental stress and actual working stress of each prediction object, and further infers the health state or RUL of each prediction object.

[0007] 3. Data-driven method. This method uses machine learning or probability statistics to reveal the potential correlation and causality between signal data and RUL labels. Since no degradation mechanism model is needed, it is easy to apply to actual industrial activities and has become the most popular RUL prediction method. With the rise of deep learning, more and more scholars have applied deep learning methods to the prediction of the remaining life of aircraft engines. The data-driven prediction method does not require a large amount of prior knowledge and has good generalization ability, and has been applied in the industry.

[0008] However, in the processing of multi-feature data such as aircraft engines, the existing method has many defects. The empirical method and the physical model-based method are time-consuming and laborious, and the deep learning method does not have an effective mechanism to weight the input features, and not all features are rich in aircraft engine degradation information. The weight proportion of some features that do not contribute much to the prediction of the remaining useful life will reduce the prediction accuracy of the model. Therefore, the application introduces an attention mechanism and ensemble learning into the deep learning method in order to achieve more accurate prediction of the remaining useful life of the aircraft engine.

[0009] Therefore, it is desirable to have a technical solution to overcome or at least alleviate at least one of the aforementioned deficiencies of the prior art. SUMMARY

[0010] The purpose of the present application is to provide a deep learning and ensemble learning-based aircraft engine remaining life prediction method to solve at least one problem existing in the prior art.

[0011] The technical solution of the present application is:

[0012] A deep learning and ensemble learning-based aircraft engine remaining life prediction method, comprising:

[0013] Step 1, obtaining aircraft engine data and preprocessing the aircraft engine data;

[0014] Step 2, constructing an aircraft engine data deep learning model, wherein the aircraft engine data deep learning model adopts a recurrent neural network with an attention mechanism and ensemble learning;

[0015] Step 3, inputting the aircraft engine data into the aircraft engine data deep learning model to output a prediction result of the remaining life of the aircraft engine.

[0016] Preferably, in step 1, the preprocessing of the aircraft engine data comprises:

[0017] The aircraft engine data is cleaned, converted and arranged.

[0018] Preferably, in step two, the aero-engine data deep learning model comprises:

[0019] a bagging module for resampling data and randomly selecting features of the data;

[0020] an RNN module for extracting time features in the data;

[0021] an Attention module for assigning different weights to each feature in the data;

[0022] an MLP module for extracting the relevance of the features in the data, extracting hidden features from the hand-crafted features, and outputting a prediction value according to the fusion features of the relevance of the features and the hidden features.

[0023] Preferably, the processing process of the Attention module comprises:

[0024] calculating the similarity of the query vector Q and the input vector K to obtain the weight coefficient corresponding to each input vector K:

[0025] f(Q,K i )=Q T ·K i

[0026] the weight coefficients are normalized by using a softmax function:

[0027]

[0028] the different weights are weighted and summed:

[0029]

[0030] where K i is the input vector, Q is the query vector for finding key information, V i is the weight value, and a i is the weight coefficient of the input vector.

[0031] Preferably, in step two, the aero-engine data deep learning model is trained, comprising:

[0032] obtaining aero-engine data and pre-processing the aero-engine data to obtain data samples;

[0033] resampling the data samples by the bagging module, using the un-sampled data as a test set, randomly selecting features in the data set, and generating a training data set according to the resampled data and the selected features;

[0034] extracting time features from the training data set by the RNN module;

[0035] assigning different weights to each feature in the training data set by the Attention module;

[0036] extracting the correlation of features from the training data set by the MLP module, and extracting hidden features from the hand-crafted features, and outputting the predicted value according to the fusion features of the correlation of features and the hidden features;

[0037] calculating the average loss function, and updating the model parameters by the stochastic gradient descent algorithm.

[0038] Preferably, in step two, further comprising evaluating the performance of the aero-engine data deep learning model by a score function Score and a root mean square error RMSE, wherein,

[0039] The score function Score is:

[0040]

[0041] The root mean square error RMSE is:

[0042]

[0043] wherein d is the error between the detected value and the actual value.

[0044] Preferably, further comprising step four, comparing and analyzing the prediction results of the aero-engine data deep learning model and the remaining life prediction model, wherein the remaining life prediction model comprises: a first remaining life prediction model using a convolutional neural network, a second remaining life prediction model using a convolutional neural network with attention mechanism, a third remaining life prediction model using a recurrent neural network, a fourth remaining life prediction model using a recurrent neural network with attention mechanism, a fifth remaining life prediction model using a long short-term memory network, a sixth remaining life prediction model using a long short-term memory network with attention mechanism, and a seventh remaining life prediction model using a long short-term memory network with attention mechanism and ensemble learning.

[0045] The present application has at least the following beneficial technical effects:

[0046] The remaining life prediction method for aero-engines based on deep learning and ensemble learning in this application introduces attention mechanism and ensemble learning on the basis of the classic deep learning model recurrent neural network (RNN). The attention mechanism enables the model to focus on the time points that are most important for life prediction in time series data, thereby improving the prediction accuracy of the model. Ensemble learning further improves the stability and generalization ability of the overall model by reducing the variance of individual models. Attached Figure Description

[0047] Figure 1 This is a flowchart of an aero-engine remaining life prediction method based on deep learning and ensemble learning, according to one embodiment of this application.

[0048] Figure 2 This is a schematic diagram of a recurrent neural network structure according to one embodiment of this application;

[0049] Figure 3 This is a schematic diagram of the attention mechanism calculation process in one embodiment of this application;

[0050] Figure 4 This is a flowchart of an embodiment of the integrated learning Bagging process in this application;

[0051] Figure 5 This is a flowchart of the training process for a deep learning model for aero-engine data, according to one embodiment of this application. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions in the embodiments of this application will be described in more detail below with reference to the accompanying drawings. In the drawings, the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The described embodiments are some, but not all, embodiments of this application. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application. The embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0053] The following is in conjunction with the appendix Figures 1 to 5 This application will be described in further detail.

[0054] This application provides a method for predicting the remaining life of aero-engines based on deep learning and ensemble learning, such as... Figure 1 As shown, it includes the following steps:

[0055] Step one, obtaining aero-engine data and preprocessing the aero-engine data;

[0056] Step two, constructing an aero-engine data deep learning model, the aero-engine data deep learning model adopts a recurrent neural network with attention mechanism and ensemble learning;

[0057] Step three, inputting the aero-engine data into the aero-engine data deep learning model, and outputting the prediction result of the aero-engine residual life.

[0058] The aero-engine residual life prediction method based on deep learning and ensemble learning of the present application, first, in step one, data preprocessing is the process of data cleaning, data conversion and data arrangement on the original aero-engine data before machine learning task. Due to the problems of noise, missing value, abnormal value and inconsistency in aero-engine data, it will have negative impact on the prediction of aero-engine residual life. Data preprocessing mainly aims to solve these problems, so that the data is more suitable for subsequent analysis and modeling.

[0059] The aero-engine residual life prediction method based on deep learning and ensemble learning of the present application, secondly, in step two, the recurrent neural network with attention mechanism and ensemble learning is used to construct the aero-engine data deep learning model, and the construction of the aero-engine data deep learning model includes:

[0060] The bagging module is used for sampling with replacement and randomly selecting features of the data;

[0061] The RNN module is used for extracting time features in the data;

[0062] The Attention module is used for giving different weights to each feature in the data;

[0063] The MLP module is used for extracting the correlation of the features in the data, and extracting hidden features from the hand-crafted features, and outputting the prediction value according to the fusion features of the correlation of the features and the hidden features.

[0064] Specifically, RNN (recurrent neural network) is a variant of artificial neural network specially designed for processing sequence data or data with time dependence. Unlike traditional feedforward neural networks, RNN has a recurrent connection, allowing information to pass and maintain memory in the network. The main feature of RNN is that it can accept input sequences of arbitrary length and can remember and use previous information in the sequence, which makes RNN very effective in processing time-series aero-engine data. The structure of RNN is as shown in Figure 2 .

[0065] When the input signal is a time axis information related to time t, at time t, the input signal ht is input to the model A, and the output of A is divided into two parts, one part is given to ht, and the other part can be hidden as a signal input back to A. At the next (t+1) time, the input signal is Xt+1, and the previously hidden signal is input back to A as an input. At this time, the input of the model A receives the information at time t and time t+1, and the output signal at this time is transmitted to the next A. On the time axis of the RNN, each information copies itself many times, and each model for the input at a certain time and the output at this time is also input to the next model.

[0066] In the neural network, as the model becomes more complex, the number of parameters increases and the expression ability becomes stronger, and at the same time, the model information becomes redundant, which brings the problem of information overload. In order to solve this problem, the attention mechanism is introduced into the neural network, which can effectively allocate a series of weights, so that the model focuses on key information, thereby improving the algorithm efficiency. The calculation process of the attention mechanism is shown in Figure 3 , which can be divided into three stages.

[0067] In the first stage, the similarity between the query vector Q and the input vector K is calculated to obtain the weight coefficient corresponding to each input vector K. In the second stage, in order to prevent the result from being too large, the weight coefficient is normalized by using the softmax function to convert it into a probability form. In the third stage, the weighted sum of different weights is obtained to obtain the final attention expression:

[0068] f(Q,K i )=Q T ·K i

[0069]

[0070]

[0071] Where K i is the input vector, Q is the query vector used to find the key information, V i is the weight value, and a i is the weight coefficient of the input vector.

[0072] Ensemble learning is a machine learning technique that aims to improve overall prediction performance by combining multiple basic learning models. By combining the prediction results of multiple models, the bias, variance and generalization error of a single model are reduced, thereby improving the overall accuracy and robustness. The process of using ensemble learning Bagging to process the aero-engine data is shown in Figure 4 .

[0073] First, a training set is extracted from the aero-engine data sample set. In each round, n training samples are extracted from the original samples using the Bootstraping method. In the training set, some samples may be extracted multiple times, while some samples may not be extracted at all. A total of k rounds of extraction are performed to obtain k training sets, and the k training sets are independent of each other. Then, one training set is used to obtain one model each time, and k models are obtained from k training sets. Finally, for the regression problem of aero-engine residual life prediction, the mean value of the above models is calculated as the final result.

[0074] The aero-engine residual life prediction method based on deep learning and ensemble learning of the application, in step two, the aero-engine data deep learning model constructed needs to be trained, and the training process includes:

[0075] Obtain aero-engine data and pre-process the aero-engine data to obtain data samples;

[0076] Resample the data samples with replacement through the bagging module, the data not sampled as the test set, randomly select the features in the data set, and generate the training data set according to the data obtained by resampling and the selected features;

[0077] Extract time features from the training data set through the RNN module;

[0078] Different weights are given to each feature in the training data set through the Attention module;

[0079] The correlation of the features is extracted from the training data set through the MLP module, and the hidden features are extracted from the hand-crafted features. The prediction value is output according to the fusion features of the correlation of the features and the hidden features;

[0080] Calculate the average loss function and update the model parameters through the stochastic gradient descent algorithm.

[0081] In the preferred embodiment of the application, as shown in Figure 5 The model training process is as follows:

[0082] Step 1, read the NASA data set, since there are some data columns in the data that do not change, which have no effect on the prediction value, so these columns are excluded.

[0083] Step 2, summarize the flight records each time to obtain the total duration of each aircraft flight.

[0084] Step 3, for each timestamp data, calculate the residual life value of the aircraft at the timestamp, which is the total duration minus the current timestamp.

[0085] Step 4, since the aircraft engine is modeled as a linear degradation model, its remaining life is constant before failure, so the remaining life value of the data with remaining life greater than 150 is set to 150.

[0086] Step 5, maximum and minimum value normalization operation is performed on the data to facilitate fast convergence of the model. In order to make the prediction value more accurate, the remaining life is also subjected to this operation, i.e. divided by 150, and its value is placed in the interval of 0 to 1.

[0087] Step 6, according to the size of the set time window, the input of the data set is generated. Then the remaining life of the first timestamp in the time window is taken as the label of the input.

[0088] Step 7, in order to increase the accuracy of the prediction, handcrafted_feature is generated, which is specifically to calculate the average value of the previous time for each data of the time window and the regression parameter obtained by taking the previous data as the regression value.

[0089] Step 8, set the initial value EnsebleNum = 0.

[0090] Step 9, set the initial value epochNum = 0.

[0091] Step 10, according to the idea of bagging of ensemble learning, the data set is sampled with replacement for times equal to the length of the data set, and those not sampled are taken as the test set.

[0092] Step 11, according to the idea of bagging of ensemble learning, indexNum features in the data set are randomly selected as the features learned by the model this time, since the features selected when predicting are needed, they are saved in the model.

[0093] Step 12, generate the data set for this training according to the selected features and the data obtained by sampling with replacement.

[0094] Step 13, for each data in the data set, combine multiple data into an input according to batch_size, and execute steps 14-21.

[0095] Step 14, since the NASA data set is a time series, the data is input into the RNN module to capture its time features.

[0096] Step 15, in order to make the model give different weight proportions to each feature, the output data of the RNN module is imported into the Attention module.

[0097] Step 16, since the output of the Attention module has multiple feature dimensions, in order to facilitate the subsequent extraction of multiple dimensional features, the data dimensions except batch_size are flattened.

[0098] Step 17, the flattened data is input into the MLP to extract features and further obtain the correlation of high-dimensional features.

[0099] Step 18, the handcrafted_feature is input into the MLP to extract its hidden related features.

[0100] Step 19, in order to fuse two features, the data of step 17 and step 18 are simply spliced.

[0101] Step 20, the fused features are put into the MLP module with output of 1 dimension to output the model prediction value.

[0102] Step 21, since the remaining life value is normalized in the data preprocessing, the label and the output are multiplied by 150 respectively to obtain the true value, the average loss function is calculated, and then the model parameters are updated through the stochastic gradient descent algorithm.

[0103] Step 22, the data obtained in this iteration is input into the test set to train, and the current test set loss and score are obtained.

[0104] Step 23, if the current test set loss is smaller than the last one, save the best_rmse model, otherwise go to the next step.

[0105] Step 24, if the current test set score is smaller than the last one, keep the best_score model, otherwise go to the next step.

[0106] Step 25, if the current training epochNum<maxEpoch, jump back to step 13 and epchNum+1, otherwise go to the next step.

[0107] Step 26, if the current training EnsebleNum<maxEnseble, jump back to step 10, otherwise the training is finished and the model is output.

[0108] In this embodiment, the C-MAPSS data set of NASA is used as data support, which records the normal operation of the engine at the beginning of a certain time sequence, and the failure occurs at a certain time point of the sequence, which is used to predict the remaining service life of the aircraft. C-MAPSS has four data subsets, and the FD001 data subset is used, including training set and test set, a total of 100 flight records, flight state is over sea flight, and there is only one failure state. The training set uses a time window of size 30, and the remaining service time of the first time is used as the label value, and finally 17731 data are generated. In the training set, the size of the failure gradually increases until the system fails. In the test set, the time sequence ends at a certain time before the system failure. The data has a total of 26 features, including flight number, time, operating parameter *3, sensor data *21.

[0109] In this embodiment, the performance of the aviation engine data deep learning model is evaluated by the score function Score and the root mean square error RMSE.

[0110] The score function Score is:

[0111]

[0112] Compared with overestimation, the score function punishes underestimation more, because in many fields such as aerospace industry, underestimation usually leads to more serious consequences.

[0113] The root mean square error RMSE is:

[0114]

[0115] Where d is the error between the detected value and the actual value.

[0116] For engine failure detection problems, a linear degradation model is used in this embodiment, that is, it is assumed that the life of the engine is constant before the engine fails and ages, and the engine is damaged only after the engine fails or ages, and in the application sense, life detection only needs to be detected when it is about to age and cannot be used. In this embodiment, the GPU is GTX 1660Ti, the memory is 16g, and the programming environment is python3.8+Pytorch1.9.1+CUDA11.1.

[0117] The aviation engine remaining life prediction method based on deep learning and ensemble learning of the application, in step three, the aviation engine data deep learning model obtained by the foregoing is used to predict the remaining life of the aviation engine.

[0118] In the preferred embodiment of the present application, step four, comparing and analyzing the prediction results of the aero-engine data deep learning model and the remaining useful life prediction model, is further included. The remaining useful life prediction model includes: a first remaining useful life prediction model using a convolutional neural network (CNN), a second remaining useful life prediction model using a convolutional neural network with attention mechanism (CNN-Attention), a third remaining useful life prediction model using a recurrent neural network (RNN), a fourth remaining useful life prediction model using a recurrent neural network with attention mechanism (RNN-Attention), a fifth remaining useful life prediction model using a long short-term memory network (LSTM), a sixth remaining useful life prediction model using a long short-term memory network with attention mechanism (LSTM-Attention), and a seventh remaining useful life prediction model using a long short-term memory network with attention mechanism and ensemble learning (LSTM-Attention-Bagging).

[0119] 1) CNN

[0120] The selected convolutional neural network is different from the common graph convolution, using 1D convolution, trying to mine the features of the time series data of the same sensor as much as possible, and not considering the relationship features between multiple sensors. Specifically, first, 3 layers of 10*1 convolution and 1 layer of 3*1 convolution are used to extract features, and finally the features are input into a fully connected layer to obtain the final RUL value. Before full connection, a Dropout of 0.5 is cited to avoid overfitting. In this embodiment, an Adam optimizer is used, the squared error is used as the loss function, the learning rate is 0.001, and the batch size is 512.

[0121] 2) CNN-Attention

[0122] On the basis of CNN, an attention mechanism module is added before each layer of convolution layer, so that the model pays attention to the importance of each feature, gives each feature a weight, and generates a more specific description. In this embodiment, an Adam optimizer is used, the squared error is used as the loss function, the learning rate is 0.001, and the batch size is 512. The attention mechanism outputs a feature of dimension 30 from a time window of length 30 through a simple two-layer MLP and Softmax.

[0123] 3) RNN

[0124] The recurrent neural network contains feedback connections from the hidden layer or output layer to the previous layer, so it has the ability to process dynamic information. RNN is a more effective model involving time series data. In this embodiment, a layer of RNN is used to extract features and then input to the fully connected layer, and the mean and trend coefficients of linear regression are added as manual features on the conventional RNN to reflect the characteristics of the original data. In this embodiment, the Adam optimizer is used, the squared error is used as the loss function, the learning rate is 0.001, and the batch size is 512. The input dimension of RNN is 17, the hidden layer dimension is 50, the regression layer is two-layer MLP with dimensions of 1500-50-10, the activation function is Tanh, and the Dropout value is 0.2.

[0125] 4) RNN-Attention

[0126] On the basis of RNN, an attention mechanism module is added before the fully connected layer, so that the model pays attention to the importance of each feature, gives each feature a weight, and generates a more specific description. In this embodiment, the Adam optimizer is used, the squared error is used as the loss function, the learning rate is 0.001, and the batch size is 512. The input dimension of RNN is 17, the hidden layer dimension is 50, the regression layer is two-layer MLP with dimensions of 1500-50-10, the activation function is Tanh, and the Dropout value is 0.2. The attention mechanism outputs a feature with a length of 30 time windows to a feature dimension of 30 features through a simple two-layer MLP and Softmax.

[0127] 5) LSTM

[0128] Because of the gradient vanishing problem in the backpropagation process of model training, the traditional RNN may not be able to capture long-term dependencies. Therefore, the long short-term memory method LSTM, as a variant of RNN, is preferred by many researchers to prevent the disappearance or explosion of backpropagation errors. Manual features in RNN are also cited. The input dimension of LSTM is 17, the hidden layer dimension is 50, the regression layer is two-layer MLP with dimensions of 1500-50-10, the activation function is Tanh, and the Dropout value is 0.2.

[0129] 6) LSTM-Attention

[0130] In this embodiment, an attention mechanism module is added before the fully connected layer based on LSTM, so that the model pays attention to the importance of each feature, gives each feature a weight, and generates a more specific description. The input dimension of LSTM is 17, the hidden layer dimension is 50, the regression layer is two-layer MLP, the dimension is 1500-50-10, the activation function is Tanh, and the Dropout value is 0.2. The attention mechanism outputs a feature with a length of 30 through a simple two-layer MLP and Softmax.

[0131] The comparison results of the C-MAPSS data set FD001 engine life prediction results are shown in Table 1.

[0132] Table 1

[0133]

[0134] On this data set, six popular deep learning methods were each performed 10 times. Among them, Score represents the error between the detected value and the actual value, RMSE represents the root mean square error, MEAN is the average of the 10 results, STD represents the standard deviation, and MIN represents the minimum value. From the above table, under the condition of not using the attention mechanism, RNN performs best among the three models, with an average error of 125 lower than CNN and 21 lower than LSTM; the root mean square error is 1.139 lower than CNN and 1.253 lower than LSTM. The performance of RNN and LSTM is obviously better than that of CNN, the main reason being that they are more suitable for capturing long-term dependencies and time series information in sequences. Their cyclic structure allows them to retain and effectively use information from previous time steps, handling variable-length sequence data, while the gating mechanism (such as LSTM) provides better memory and selective forgetting, which helps to handle long-term temporal relationships. In contrast, CNN is usually used for static data, requires fixed-size input, and parameters are not shared, so additional processing may be required when processing sequence data. Overall, RNN and LSTM are more suitable for tasks involving time and sequence relationships.

[0135] After introducing the attention mechanism, the three models have improved to some extent. The average error of the CNN with attention is 15.6% lower than before; the average error of the RNN with attention is 2.3 lower than before; the average error of the LSTM with attention is 8.2% lower than before. The main reason is that it can better capture key information, handle long-term dependencies, adapt to changes in sequences, and reduce information uncertainty. The attention mechanism allows the model to selectively focus on different parts of the sequence and more flexibly adjust the degree of attention to important information, thereby improving the model's sensitivity and expressiveness for time series data and improving the model's performance and generalization ability.

[0136] The ensemble learning uses the idea of random forest bagging, the number of ensembles is 3, and each model is trained separately. The original data and the subset of features are input into the model for training, as follows. First, 17731 is sampled with replacement for 17731 times, and then 12-dimensional data is randomly extracted from the cleaned 17-dimensional data, which is input into the RNN-Attention and LSTM-Attention models with better performance in the previous section for training. Finally, the obtained data is averaged to obtain the final RUL prediction value.

[0137] LSTM-Attention-Bagging and RNN-Attention-Bagging both use the Adam optimizer, the squared error is the loss function, the learning rate is 0.001, and the batch size is 512. The input dimension of RNN and LSTM is 12, the hidden layer dimension is 50, the regression layer is two-layer MLP, the dimension is 1500-50-10, the activation function is Tanh, and the Dropout value is 0.2. The attention mechanism outputs a feature with a dimension of 30 from a time window with a length of 30 through a simple two-layer MLP and Softmax. The attention mechanism outputs a feature with a dimension of 30 from a time window with a length of 30 through a simple two-layer MLP and Softmax.

[0138] The comparison of the results of the C-MAPSS data set FD001 aircraft engine life prediction is shown in Table 2.

[0139] Table 2

[0140]

[0141] From the above table, the model integrating three RNN-Attention achieves the best aero-engine remaining life prediction effect among all the above models. After using ensemble learning, the prediction accuracy of LSTM-Attention and RNN-Attention increases by 25.1% and 28.6% respectively compared with not using ensemble learning, and the prediction error decreases significantly. There are several key reasons behind this improvement. First, Bagging introduces diversity, which makes each model train on different data subsets by bootstrap sampling, which helps to capture different aspects of data patterns. Second, this ensemble method reduces the risk of overfitting, because each model is less disturbed by noise or outliers on a specific subset. In addition, ensemble learning improves the generalization ability of the overall model by combining the prediction results of multiple models, so that it can also perform well on unseen data. In addition, when individual models make mistakes, ensemble learning can correct these errors through the cooperation of other models, thereby improving the overall robustness.

[0142] The deep learning and ensemble learning based aero-engine remaining life prediction method of the present application compares the effects of three deep learning models CNN, RNN and LSTM in aero-engine remaining life prediction, and conducts comparative analysis on the C-MAPSS time series data set. On the basis of traditional deep learning, in order to further improve the prediction ability, attention mechanism and ensemble learning method are introduced in succession. The results prove that the attention mechanism enables the model to selectively focus on different parts of the sequence and more flexibly adjust the attention degree to important information; ensemble learning improves the generalization ability of the overall model by combining the prediction results of multiple models. This deep learning and ensemble learning based aero-engine remaining life prediction method has higher prediction accuracy and stronger robustness in practical application.

[0143] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto. Any skilled person in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for predicting the remaining life of an aero-engine based on deep learning and ensemble learning, characterized in that, include: Step 1: Acquire aero-engine data and preprocess the aero-engine data; Step 2: Construct a deep learning model for aero-engine data. The deep learning model for aero-engine data adopts a recurrent neural network that incorporates attention mechanism and ensemble learning. Step 3: Input the aero-engine data into the aero-engine data deep learning model and output the prediction result of the remaining life of the aero-engine.

2. The method for predicting the remaining life of aero-engines based on deep learning and ensemble learning according to claim 1, characterized in that, Step one involves preprocessing the aero-engine data, including: The aero-engine data is cleaned, transformed, and organized.

3. The method for predicting the remaining life of aero-engines based on deep learning and ensemble learning according to claim 1, characterized in that, In step two, the deep learning model for aero-engine data includes: The bagging module is used to sample data with replacement and randomly select features from the data. The RNN module is used to extract temporal features from the data; The Attention module is used to assign different weights to each feature in the data. The MLP module is used to extract the correlation of features in the data, extract hidden features from the handcrafted features, and output predicted values ​​based on the fusion features of the feature correlation and the hidden features.

4. The method for predicting the remaining life of aero-engines based on deep learning and ensemble learning according to claim 3, characterized in that, The Attention module processes the data as follows: Calculate the similarity between the query vector Q and the input vector K to obtain the weight coefficient corresponding to each input vector K: f(Q,K i )=Q T ·K i The weight coefficients are normalized using the softmax function: Weighted summation is performed for different weights: Among them, K i Let V be the input vector, Q be the query vector used to find key information, and V be the input vector. i Let a be the weight value. i These are the weight coefficients of the input vector.

5. The method for predicting the remaining life of an aero-engine based on deep learning and ensemble learning according to claim 4, characterized in that, Step two also includes training the deep learning model for the aero-engine data, including: Acquire aero-engine data and preprocess the aero-engine data to obtain data samples; The data samples are sampled with replacement using the bagging module. The unsampled data is used as the test set. Features are randomly selected from the dataset, and a training dataset is generated based on the sampled data with replacement and the selected features. The RNN module extracts temporal features from the training dataset. The Attention module assigns different weights to each feature in the training dataset. The MLP module extracts the correlation of features from the training dataset and extracts hidden features from the handcrafted features. The predicted value is output based on the fusion of the correlation of features and the hidden features. Calculate the average loss function and update the model parameters using the stochastic gradient descent algorithm.

6. The method for predicting the remaining life of an aero-engine based on deep learning and ensemble learning according to claim 5, characterized in that, Step two also includes evaluating the performance of the deep learning model for the aero-engine data using a scoring function (Score) and the root mean square error (RMSE). The scoring function Score is: The root mean square error (RMSE) is: Where d is the error between the detected value and the actual value.

7. The method for predicting the remaining life of an aero-engine based on deep learning and ensemble learning according to claim 6, characterized in that, The method also includes step four, which involves comparing and analyzing the prediction results of the deep learning model for aero-engine data with the remaining lifetime prediction model. The remaining lifetime prediction model includes: a first remaining lifetime prediction model using a convolutional neural network; a second remaining lifetime prediction model using a convolutional neural network with an attention mechanism; a third remaining lifetime prediction model using a recurrent neural network; a fourth remaining lifetime prediction model using a recurrent neural network with an attention mechanism; a fifth remaining lifetime prediction model using a long short-term memory network; a sixth remaining lifetime prediction model using a long short-term memory network with an attention mechanism; and a seventh remaining lifetime prediction model using a long short-term memory network with both an attention mechanism and ensemble learning.

Images