Method, device, electronic device and medium for predicting remaining service life of engine

Abstract

The present application discloses a method, device, electronic device and medium for predicting the remaining service life of an engine. The method includes: establishing an initial prediction model for the remaining service life of the engine, wherein the initial prediction model for the remaining service life of the engine follows an encoder-decoder structure, and the encoder includes an hourglass-shaped feature extractor and a masked self-attention mechanism; obtaining degradation data of the engine to be tested, and determining the remaining service life of the engine to be tested based on the trained prediction model for the remaining service life of the engine. The degradation data is scaled to multiple scales through the hourglass-shaped feature extractor for feature fusion, and masked processing is performed by the masked self-attention mechanism, so as to obtain time correlation features at different scales, further improving the reliability of the data features, and then improving the accuracy of the prediction model for the remaining service life of the engine in predicting long time series.

Description

Technical Field

[0001] The present invention relates to the technical field of aero-engine fault prediction and health management, and particularly to a method, device, electronic device and medium for predicting the remaining service life of an engine. Background Art

[0002] With the maturity of the manufacturing industry and industrial systems, various mechanical equipment and devices play a very important role in various industries. As a key component of an aircraft, if an aero-engine suddenly fails during operation, it will cause a huge disaster. The Prognostics and Health Management technology (PHM) of equipment is a very good management solution for equipment. The prediction of the remaining service life of an aero-engine is a very important one in the PHM technology.

[0003] In order to obtain the remaining service life of equipment, the commonly used remaining service life prediction methods are roughly divided into three types: methods based on traditional physical models; data-driven methods; and methods that mix the two. Since it is difficult to establish an accurate dynamic model for methods based on physical models and it is difficult to implement mixed methods, data-driven methods are mostly used, and deep learning methods are the mainstream in data-driven methods. However, in existing life prediction methods based on deep learning, most use convolutional neural networks (CNNs) or recurrent neural networks (RNNs) to extract time-dependent information in time series, but these networks have poor extraction effects on the long-term time dependence of feature data, and the recursive structure of RNNs also has problems such as gradient disappearance.

[0004] Therefore, in the prior art, during the process of predicting the remaining service life of an engine, there is a problem of low prediction accuracy for long time series. Summary of the Invention

[0005] In view of this, it is necessary to provide a method, device, electronic device and medium for predicting the remaining service life of an engine to solve the problem of low prediction accuracy existing in the prior art during the process of predicting the remaining service life of an engine.

[0006] To solve the above problems, the present invention provides a method for predicting the remaining service life of an engine, including:

[0007] Obtaining a degradation data sample of a sample engine;

[0008] Establishing an initial prediction model for the remaining service life of an engine, wherein the initial prediction model for the remaining service life of an engine follows an encoder-decoder structure, and the encoder includes an hourglass-shaped feature extractor and a masked self-attention mechanism;

[0009] Input the degraded data samples into the initial engine remaining useful life prediction model, extract data features from the degraded data samples based on the hourglass feature extractor, and use the remaining useful life samples of the sample engine as the output to train the initial engine remaining useful life prediction model to obtain a well-trained engine remaining useful life prediction model;

[0010] Obtain the degraded data of the engine to be tested, and determine the remaining useful life of the engine to be tested based on the well-trained engine remaining useful life prediction model.

[0011] Furthermore, obtaining the degraded data samples of the sample engine includes:

[0012] Obtain the initial degraded data of the sample engine;

[0013] Delete the constant degraded data in the initial degraded data to obtain the transitional degraded data;

[0014] Perform normalization processing and correction processing on the transitional degraded data to obtain the degraded data samples.

[0015] Furthermore, before performing normalization processing and correction processing on the transitional degraded data, it also includes:

[0016] Set up a linear regression model to train the transitional degraded data to obtain the regression coefficient estimation of the transitional degraded data;

[0017] Obtain the sequence average value of the transitional degraded data;

[0018] Concatenate the transitional degraded data with the regression coefficient estimation and the sequence average value respectively.

[0019] Furthermore, the initial engine remaining useful life prediction model includes an input module, an encoding module, a decoding module, and an output module;

[0020] The input module includes a linear layer and a position information encoding unit;

[0021] The encoding module includes an hourglass feature extractor, a three-layer masked self-attention mechanism, and a feed-forward neural network;

[0022] The decoding module includes a single-layer multi-head attention mechanism;

[0023] The output module includes a flattening layer unit, two linear layer units, and an activation function.

[0024] Furthermore, inputting the degraded data samples into the initial engine remaining useful life prediction model, extracting data features from the degraded data samples based on the hourglass feature extractor, and using the remaining useful life samples of the sample engine as the output includes:

[0025] Input the degraded data samples into the input module, process them through a linear layer and a positional information encoding unit to obtain linearly degraded data samples;

[0026] Scale the linearly degraded data samples through an hourglass feature extractor, and fuse the features of different scales to obtain a multi-scale fused feature map;

[0027] Input the multi-scale fused feature map into a three-layer masked self-attention mechanism to obtain time correlation features of different scales, and connect them through a feed-forward neural network to obtain an encoded feature map;

[0028] Input the encoded feature map into the decoder, perform query attention based on a single-layer multi-head attention mechanism to obtain an attention result vector;

[0029] Input the attention result vector into the output module, and pass it through a flattening layer unit, two linear layer units and an activation function in sequence to obtain a remaining useful life prediction value sample.

[0030] Furthermore, the hourglass feature extractor includes two downsampling convolutional layers, two upsampling convolutional layers and a 1×1 one-dimensional convolutional layer; the linearly degraded data samples are subjected to data feature extraction through the hourglass feature extractor, masked through a three-layer masked self-attention mechanism, and data connection is performed through a feed-forward neural network to obtain a remaining useful life encoded sample, including:

[0031] Pass the linearly degraded data samples through two downsampling convolutional layers and two upsampling convolutional layers in sequence to obtain upsampled feature data and downsampled feature data respectively;

[0032] According to the one-dimensional convolutional layer, add and fuse the data with equal time steps in the upsampled feature data and the downsampled feature data, and perform normalization processing to obtain fused feature data;

[0033] Input the fused feature data into the three-layer masked self-attention mechanism for masking, and perform data connection through a feed-forward neural network to obtain a remaining useful life encoded sample.

[0034] Furthermore, the three-layer masked self-attention mechanism includes three-layer pyramid self-attention layers and a masked self-attention layer; input the multi-scale fused feature map into the three-layer masked self-attention mechanism to obtain time correlation features of different scales, and connect them through a feed-forward neural network to obtain an encoded feature map, including:

[0035] Input the multi-scale fused feature map into the three-layer pyramid self-attention layers. The masked self-attention layer first performs masking processing on the multi-scale fused feature map, selects to obtain child nodes, A nodes and parent nodes, and sets the remaining nodes to negative infinity;

[0036] Among them, the principle formula for the masked self-attention layer to perform masking processing on the multi-scale fusion feature map is as follows:

[0037]

[0038] Among them, respectively represent all A nodes, child nodes, and parent nodes corresponding to the l-th data point in the s-th layer, represents the j-th node in the s-th layer, A, C, and P represent the numbers of A node child nodes and parent nodes, and s represents the number of layers.

[0039] Furthermore, to obtain a well-trained prediction model for the remaining useful life of the engine, it further includes:

[0040] Establish evaluation metrics to evaluate the test results of the prediction model for the remaining useful life of the engine;

[0041] According to the evaluation results, adjust the relevant parameters of the prediction model for the remaining useful life of the engine to obtain a well-trained prediction model for the remaining useful life of the engine;

[0042] Among them, the evaluation metrics include:

[0043]

[0044] Among them, n represents the total number of aero-engines, RUL pred，i represents the predicted value of the remaining useful life of the i-th aero-engine, and RUL new，i represents the remaining useful life of the l-th aero-engine after life correction.

[0045] To solve the above problems, the present invention also provides a device for predicting the remaining useful life of an engine, including:

[0046] A sample acquisition module for acquiring degradation data samples of sample engines;

[0047] A model establishment module for establishing an initial prediction model for the remaining useful life of the engine. Among them, the initial prediction model for the remaining useful life of the engine follows an encoder-decoder structure, and the encoder includes an hourglass-shaped feature extractor and a masked self-attention mechanism;

[0048] A model training module for inputting the degradation data samples into the initial prediction model for the remaining useful life of the engine, extracting data features from the degradation data samples based on the hourglass-shaped feature extractor, and using the remaining useful life samples of the sample engines as the output to train the initial prediction model for the remaining useful life of the engine to obtain a well-trained prediction model for the remaining useful life of the engine;

[0049] The remaining useful life determination module is used to obtain the degradation data of the engine to be tested and determine the remaining useful life of the engine to be tested based on the well-trained prediction model of the remaining useful life of the engine.

[0050] To solve the above problems, the present invention also provides an electronic device, including a processor and a memory. A computer program is stored on the memory. When the computer program is executed by the processor, the method for predicting the remaining useful life of the engine as described above is implemented.

[0051] The beneficial effects of adopting the above technical solutions are as follows: The present invention provides a method, device, electronic device and medium for predicting the remaining useful life of an engine. The method includes: establishing an initial prediction model for the remaining useful life of the engine, where the initial prediction model for the remaining useful life of the engine follows an encoder-decoder structure, and the encoder includes an hourglass-shaped feature extractor and a masked self-attention mechanism; obtaining the degradation data of the engine to be tested, and determining the remaining useful life of the engine to be tested based on the well-trained prediction model of the remaining useful life of the engine. The degradation data is scaled to multiple scales for feature fusion through the hourglass-shaped feature extractor, and masked processing is performed by the masked self-attention mechanism, so as to obtain time correlation features at different scales, further improving the reliability of data features, and then improving the accuracy of the prediction model of the remaining useful life of the engine for long time series prediction. Description of the Drawings

[0052] Figure 1 It is a schematic flowchart of an embodiment of the method for predicting the remaining useful life of the engine provided by the present invention;

[0053] Figure 2 It is a schematic flowchart of an embodiment of obtaining a degradation data sample of a sample engine provided by the present invention;

[0054] Figure 3 It is a schematic flowchart of an embodiment of training an initial prediction model for the remaining useful life of the engine provided by the present invention;

[0055] Figure 4 It is a schematic flowchart of an embodiment of obtaining a remaining useful life coding sample provided by the present invention;

[0056] Figure 5 It is a schematic structural diagram of the device for predicting the remaining useful life of the engine provided by the present invention;

[0057] Figure 6 It is a schematic block diagram of an embodiment of the electronic device provided by the present invention. Detailed Embodiments

[0058] The preferred embodiments of the present invention will be specifically described below in conjunction with the accompanying drawings. The accompanying drawings form a part of this application and are used together with the embodiments of the present invention to explain the principles of the present invention, rather than to limit the scope of the present invention.

[0059] Before presenting the embodiments, the convolutional neural network and the recurrent neural network will be elaborated:

[0060] A convolutional neural network (CNN) is a deep neural network with a convolutional structure. The convolutional structure can reduce the memory occupied by the deep network. Its three key operations are local receptive field, weight sharing, and the pooling layer, which can effectively reduce the number of parameters in the neural network and alleviate the overfitting problem of the model.

[0061] A recurrent neural network (RNN) is a type of recursive neural network that takes sequence data as input, recurses in the evolution direction of the sequence, and all nodes (recurrent units) are connected in a chain.

[0062] In the gradient calculation method of the recurrent neural network, problems such as gradient disappearance and error accumulation are likely to occur when the time step is long. Limited by this, it is difficult for the recurrent neural network to capture the temporal correlation features between long-term sequence elements.

[0063] Prognostics Health Management (PHM) is an upgraded development based on Condition Based Maintenance (CBM) proposed to meet the requirements of autonomous guarantee and autonomous diagnosis of mechanical equipment. It was first applied in the field of aeroengines. With the development of industrial systems and PHM technology, it has been gradually applied to medical, military equipment and other fields. The remaining life prediction of aeroengines belongs to a kind of PHM technology. This technology is used to predict problems such as equipment aging faults that may occur in aeroengines, so as to achieve more efficient maintenance of the engines, which is of great significance for improving the reliability, safety and maintenance efficiency of aeroengines.

[0064] Generally speaking, the methods for predicting the remaining life of aero-engines can be roughly divided into three types: traditional model-based methods, data-driven methods, and hybrid methods of the two. Among them, traditional model-based methods require accurate dynamic modeling of mechanical equipment to describe the degradation trend of components, while it is difficult to achieve accurate modeling for modern equipment with complex structures. Therefore, data-driven methods have developed faster. At the same time, as the amount of industrial data is increasing, deep learning methods in data-driven methods have developed rapidly with the ability to automatically extract features and higher accuracy.

[0065] In most of the life prediction methods based on deep learning, convolutional neural networks (CNNs) or recurrent neural networks (RNNs) are used to extract time-dependent information in time series. However, these networks have poor performance in extracting long-term time dependencies of feature data, and the recursive structure of RNNs also has problems such as gradient disappearance. Therefore, in the process of predicting the remaining service life of engines in the prior art, there is a problem of low prediction accuracy.

[0066] To solve the problem of low prediction accuracy in the process of predicting the remaining service life of engines in the prior art, the present invention provides a method, device, electronic device, and storage medium for predicting the remaining service life of an engine, which will be described in detail below.

[0067] As Figure 1 shown, Figure 1 is a schematic flowchart of an embodiment of the method for predicting the remaining service life of an engine provided by the present invention, including:

[0068] Step S101: Obtain the degradation data samples of the sample engine;

[0069] Step S102: Establish an initial prediction model for the remaining service life of the engine, where the initial prediction model for the remaining service life of the engine follows an encoder-decoder structure, and the encoder includes an hourglass feature extractor and a masked self-attention mechanism;

[0070] Step S103: Input the degradation data samples into the initial prediction model for the remaining service life of the engine, extract data features from the degradation data samples based on the hourglass feature extractor, and use the remaining service life samples of the sample engine as the output to train the initial prediction model for the remaining service life of the engine to obtain a trained and complete prediction model for the remaining service life of the engine;

[0071] Step S104: Obtain the degradation data of the engine to be tested, and determine the remaining service life of the engine to be tested based on the trained and complete prediction model for the remaining service life of the engine.

[0072] In this embodiment, first, degradation data samples of a sample engine are obtained; second, an initial prediction model for the remaining useful life of the engine is established. Among them, the initial prediction model for the remaining useful life of the engine follows an encoder-decoder structure, and the encoder includes an hourglass feature extractor and a masked self-attention mechanism; then, the degradation data samples are input into the initial prediction model for the remaining useful life of the engine. Based on the hourglass feature extractor, data features of the degradation data samples are extracted, and the remaining useful life samples of the sample engine are used as outputs to train the initial prediction model for the remaining useful life of the engine, and a well-trained prediction model for the remaining useful life of the engine is obtained; finally, the degradation data of the engine to be tested is obtained, and based on the well-trained prediction model for the remaining useful life of the engine, the remaining useful life of the engine to be tested is determined.

[0073] In this embodiment, by setting an hourglass feature extractor in the encoding module of the prediction model for the remaining useful life of the engine, the degradation data is scaled to multiple scales for feature fusion based on the hourglass feature extractor, and masked processing is performed by the masked self-attention mechanism, so as to obtain time correlation features at different scales, further improve the reliability of data features, and then improve the accuracy of the prediction model for the remaining useful life of the engine in predicting long time series.

[0074] As a preferred embodiment, in step S101, in order to obtain degradation data samples of a sample engine, as Figure 2 shown, Figure 2 is a schematic flow chart of an embodiment for obtaining degradation data samples of a sample engine provided by the present invention, including:

[0075] Step S111: Obtain the initial degradation data of the sample engine;

[0076] Step S112: Delete the constant degradation data in the initial degradation data to obtain transitional degradation data;

[0077] Step S113: Perform normalization processing and correction processing on the transitional degradation data to obtain degradation data samples.

[0078] In this embodiment, first, the initial degradation data of the sample engine is obtained; then, the constant degradation data in the initial degradation data is deleted to obtain transitional degradation data; finally, the transitional degradation data is subjected to normalization processing and correction processing to obtain degradation data samples.

[0079] In this embodiment, by removing the constant degradation data that remains unchanged during the operation cycle, the influence of invalid data on the prediction performance is avoided; by performing normalization processing and correction processing on the data, not only the adverse effects caused by singular sample data can be eliminated, but also the calculation difficulty can be reduced, the data consistency can be improved, and it is helpful to improve the accuracy of the finally obtained prediction result.

[0080] In a specific embodiment, in step S112, all the initial degradation data of the sample engine are listed, and the sensor data that have not changed are excluded to obtain the transitional degradation data; that is, the sensor data that have not changed in the initial degradation data are excluded, and the transitional degradation data are obtained.

[0081] In a specific embodiment, in step S113, the transitional degradation data also need to be normalized and corrected to obtain the degradation data samples. In this embodiment, the Min-Max normalization method is used to scale the transitional degradation data to the range of [0, 1]. The specific formula is as follows:

[0082]

[0083] where x norm represents the normalized data, min(x) represents the minimum value in the time series x, and max(x) represents the maximum value in the time series x.

[0084] Furthermore, the remaining useful life of the engine also needs to be corrected. First, set the maximum life RUL max , and then determine the degradation data samples. The specific formula is as follows:

[0085]

[0086] where RUL new represents the degradation data sample, and RUL origiin represents the remaining useful life value obtained from the original test of the aero-engine.

[0087] In a specific embodiment, the maximum life is set to 125.

[0088] Furthermore, for the convenience of data processing, the degradation data samples need to be processed at intervals. In a specific embodiment, the degradation data samples are first divided to determine the training set and the test set; then, the training set and the test set are intercepted using a sliding time window. In this application, the sliding window size is selected to be 50, and the sliding interval is one time step. Finally, the training set and the test set with a new time step length of 50 are obtained.

[0089] In other embodiments, the sliding window size can also be adaptively adjusted according to actual needs.

[0090] As a preferred embodiment, before normalizing and correcting the transitional degradation data, it is also necessary to splice the transitional degradation data. First, a linear regression model is set to train the transitional degradation data to obtain the regression coefficient estimation of the transitional degradation data; then, the sequence average value of the transitional degradation data is obtained; finally, the transitional degradation data is spliced with the regression coefficient estimation and the sequence average value respectively.

[0091] By splicing the transitional degradation data, it is convenient to extract the data features of the transitional degradation data.

[0092] As a preferred embodiment, in step S102, the initial engine remaining service life prediction model includes an input module, an encoding module, a decoding module, and an output module; wherein, the input module includes a linear layer and a position information encoding unit; the encoding module includes an hourglass feature extractor, a three-layer masked self-attention mechanism, and a feed-forward neural network; the decoding module includes a single-layer multi-head attention mechanism; the output module includes a flattening layer unit, two linear layer units, and an activation function.

[0093] It should be noted that this application is an improvement based on the existing Transformer model. By constructing a Transformer model with an hourglass feature extractor, an initial engine remaining service life prediction model is constructed.

[0094] As a preferred embodiment, in step S103, in order to train the initial engine remaining service life prediction model, as Figure 3 shown, Figure 3 is a schematic flowchart of a method for training an initial engine remaining service life prediction model provided by the present invention, including:

[0095] Step S131: Input the degradation data sample into the input module, and process it through the linear layer and the position information encoding unit to obtain a linear degradation data sample;

[0096] Step S132: Scale the linear degradation data sample through the hourglass feature extractor, and fuse the features of different scales to obtain a multi-scale fusion feature map;

[0097] Step S133: Input the multi-scale fusion feature map into the three-layer masked self-attention mechanism to obtain the time correlation features of different scales, and connect them by the feed-forward neural network to obtain an encoded feature map;

[0098] Step S134: Input the encoded feature map into the decoder, and perform query attention based on the single-layer multi-head attention mechanism to obtain an attention result vector;

[0099] Step S135: Input the attention result vector into the output module, and successively pass through a flattening layer unit, two linear layer units, and an activation function to obtain the remaining useful life prediction value samples.

[0100] In this embodiment, first, input the degradation data samples into the input module, and through processing by the linear layer and the position information encoding unit, obtain the linear degradation data samples; secondly, scale the linear degradation data samples through the hourglass feature extractor, and fuse the features of different scales to obtain the multi-scale fusion feature map; and input the multi-scale fusion feature map into the three-layer masked self-attention mechanism to obtain the time correlation features of different scales, which are connected by the feed-forward neural network to obtain the encoded feature map; then, input the encoded feature map into the decoder, and perform query attention based on the single-layer multi-head attention mechanism to obtain the attention result vector; finally, input the attention result vector into the output module, and successively pass through a flattening layer unit, two linear layer units, and an activation function to obtain the remaining useful life prediction value samples.

[0101] In this embodiment, by extracting the data features in the linear degradation data samples through the hourglass feature extractor and performing feature fusion, the accuracy and reliability of the data features can be improved; masking with the three-layer masked self-attention mechanism can achieve targeted monitoring of the training effect of the initial engine remaining useful life prediction model and adaptively improve the corresponding parameters of the initial engine remaining useful life prediction model, thereby improving the reliability of the initial engine remaining useful life prediction model.

[0102] As a preferred embodiment, in step S132, the hourglass feature extractor includes two downsampling convolutional layers, two upsampling convolutional layers, and a 1×1 one-dimensional convolutional layer. In order to obtain the remaining useful life encoded samples, as Figure 4 shown, Figure 4 is a schematic flowchart of an embodiment for obtaining the remaining useful life encoded samples provided by the present invention, including:

[0103] Step S1321: Pass the linear degradation data samples successively through two downsampling convolutional layers and two upsampling convolutional layers to respectively obtain the upsampled feature data and the downsampled feature data;

[0104] Step S1322: According to the one-dimensional convolutional layer, add and fuse the data with equal time steps in the upsampled feature data and the downsampled feature data, and perform normalization processing to obtain the fused feature data;

[0105] Step S1323: Input the fused feature data into the three-layer masked self-attention mechanism for masking, and perform data connection by the feed-forward neural network to obtain the remaining useful life encoded samples.

[0106] In a specific embodiment, in step S1323, the three-layer masked self-attention mechanism includes three-layer pyramid self-attention layers and a masked self-attention layer; the multi-scale fusion feature map is input into the three-layer pyramid self-attention layers, and the masked self-attention layer first performs a masking process on the multi-scale fusion feature map to select and obtain child nodes, A nodes, and parent nodes, and sets the remaining nodes to negative infinity;

[0107] Among them, the principle formula for the masked self-attention layer to first perform a masking process on the multi-scale fusion feature map is:

[0108]

[0109] Among them, respectively represent all A nodes, child nodes, and parent nodes corresponding to the l-th data point in the s-th layer, represents the j-th node in the s-th layer, A, C, and P represent the numbers of A nodes, child nodes, and parent nodes, and s represents the number of layers.

[0110] In this embodiment, by selecting adjacent nodes at different scales and setting other nodes to negative infinity, and then setting them to zero after passing through the softmax function, the computational complexity can be reduced, and at the same time, the attention between adjacent nodes at different scales can be improved.

[0111] In other embodiments, the above parameters can also be adaptively adjusted according to actual needs.

[0112] Furthermore, in order to evaluate the training results of the engine remaining useful life prediction model, evaluation indicators are also established to evaluate the prediction results. Then, according to the evaluation results, the relevant parameters of the engine remaining useful life prediction model are adjusted to obtain a well-trained engine remaining useful life prediction model.

[0113] Among them, the evaluation indicators include:

[0114]

[0115] Among them, n represents the total number of aero-engines, RUL pred，i represents the predicted value of the remaining useful life of the i-th aero-engine, and RUL new，i represents the remaining useful life of the i-th aero-engine after life correction.

[0116] Finally, in order to clearly and specifically illustrate the prediction effect of the engine remaining useful life prediction model in this application, this application is illustrated by the following specific embodiments.

[0117] In this embodiment, the C-MAPSS dataset of aero turbofan engines provided by NASA is adopted. This dataset contains four sub-datasets (FD001 to FD004). Each sub-dataset has engine historical degradation data under different operating conditions and fault modes. To demonstrate the effect, the FD004 dataset with the highest prediction difficulty is used in the present invention. The FD004 dataset includes a training set of 249 engines with complete degradation until failure data and a test set obtained by randomly intercepting data from the start of degradation to a random time before failure of 248 engines.

[0118] First, list the data changes in the above-mentioned dataset, which include engine numbers, 3 operating condition settings, and 21 sensor data. Among them, the data of sensors No. 1, 5, 10, 16, 18, and 19 hardly change, and the corresponding column data are removed; then the 3 operating condition settings of each engine are integrated and classified into 6 operating conditions, and the remaining 15 sensor data of 249 engines are normalized by Min-max according to the 6 setting conditions. Then the data life is corrected to 125. Subsequently, the sliding window method is used to intercept the data of each engine in the training set to obtain the training data with the last two dimensions of 50×15. For the test set, the first 50 data of each engine are intercepted as the test set, and for the engine data with a step size smaller than the sliding window, the linear interpolation method is used to interpolate the data step size to 50 to obtain the test set of 248×50×15.

[0119] Secondly, the sensor data of the training set and the test set obtained above are put into a linear regression model for preliminary feature extraction, and the corresponding extracted linear regression coefficients and the mean values of the sensor data corresponding to each engine in the dataset are used as new feature data to be spliced with the original data so that the time step dimension changes from 50 to 52.

[0120] Then, an initial prediction model for the remaining useful life of the engine is constructed, and the training data and the test data are input into the model. First, the training prediction results are obtained from the training data, and the prediction results are compared with the true remaining useful life to calculate the training loss. Then, the training loss is used to optimize and update the parameters of the model through the ADAM optimizer. Then, the test results are obtained from the test set, and the test loss and the test score are calculated from the test results and the true values to evaluate the prediction performance of the current model.

[0121] Finally, set the initial optimal loss to a value much larger than the test loss. Then, in each iteration, compare the test loss with the optimal loss. If the test loss is less than the optimal loss, update the optimal loss to the current test loss and save the current model. If the test loss is greater than the optimal loss, enter the learning rate decay strategy, that is, determine whether the current loss is within the range of the optimal loss plus the threshold. If it is, it is a normal fluctuation and enter the next iteration. If it exceeds the range, update the learning rate of the optimizer to a smaller value, and at the same time start the early stopping strategy, that is, increment the stop count by one. When the stop count reaches the set threshold of 5, export the currently saved model as the optimal model.

[0122] Furthermore, the effectiveness of this method can also be evaluated by comparing the prediction accuracy of this method with other methods. The RMSE and Score evaluation results of this method and other methods are shown in the following table:

[0123]

[0124] In summary, by setting a hourglass-shaped feature extractor in the encoding module of the remaining useful life prediction model of the engine, processing the degradation data of the engine based on the hourglass-shaped feature extractor, and fusing the data features, the data features in the degradation data can be effectively extracted, further improving the reliability of the data features, and thus improving the accuracy of the remaining useful life prediction model of the engine for long time series prediction.

[0125] To solve the above problems, the present invention also provides a device for predicting the remaining useful life of an engine, as Figure 5 shown, Figure 5 is a structural schematic diagram of the device for predicting the remaining useful life of an engine provided by the present invention. The device 500 for predicting the remaining useful life of an engine includes:

[0126] A sample acquisition module 501, configured to acquire degradation data samples of a sample engine;

[0127] A model establishment module 502, configured to establish an initial remaining useful life prediction model of the engine, wherein the initial remaining useful life prediction model of the engine follows an encoder-decoder structure, and the encoder includes a hourglass-shaped feature extractor and a masked self-attention mechanism;

[0128] A model training module 503, configured to input the degradation data samples into the initial remaining useful life prediction model of the engine, extract data features from the degradation data samples based on the hourglass-shaped feature extractor, and use the remaining useful life samples of the sample engine as the output to train the initial remaining useful life prediction model of the engine to obtain a trained complete remaining useful life prediction model of the engine;

[0129] A remaining service life determination module 504, configured to obtain degradation data of an engine to be tested, and determine the remaining service life of the engine to be tested based on the trained and complete engine remaining service life prediction model.

[0130] The present invention also correspondingly provides an electronic device, such as Figure 6 shown Figure 6 is a structural block diagram of an embodiment of the electronic device provided by the present invention. The electronic device 600 may be a computing device such as a mobile terminal, a desktop computer, a notebook, a palm computer, and a server. The electronic device 600 includes a processor 601 and a memory 602, wherein an engine remaining service life prediction program 603 is stored on the memory 602.

[0131] In some embodiments, the memory 602 may be an internal storage unit of a computer device, such as a hard disk or a memory of a computer device. In other embodiments, the memory 602 may also be an external storage device of a computer device, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. equipped on the computer device. Further, the memory 602 may also include both an internal storage unit and an external storage device of a computer device. The memory 602 is used to store application software installed on the computer device and various types of data, such as program codes installed on the computer device. The memory 602 may also be used to temporarily store data that has been output or will be output. In one embodiment, the engine remaining service life prediction program 603 can be executed by the processor 601, so as to implement the engine remaining service life prediction method of each embodiment of the present invention.

[0132] In some embodiments, the processor 601 may be a central processing unit (CPU), a microprocessor, or other data processing chips, configured to run program codes stored in the memory 602 or process data, such as executing the engine remaining service life prediction program.

[0133] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Any reference to a memory, storage, database, or other medium used in the embodiments provided in the present application can include non-volatile and / or volatile memories. Non-volatile memories can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memories can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and Rambus dynamic RAM (RDRAM), etc.

[0134] As described above, the above are only the preferred specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims

1. A method for predicting the remaining service life of an engine, characterized in that Including: Obtaining a degradation data sample of a sample engine; Establishing an initial prediction model for the remaining useful life of the engine, wherein the initial prediction model for the remaining useful life of the engine follows an encoder-decoder structure, and the encoder includes an hourglass feature extractor and a masked self-attention mechanism; the initial prediction model for the remaining useful life of the engine includes an input module, an encoding module, a decoding module, and an output module; the input module includes a linear layer and a position information encoding unit; the encoding module includes an hourglass feature extractor, a three-layer masked self-attention mechanism, and a feed-forward neural network; the decoding module includes a single-layer multi-head attention mechanism; the output module includes a flattening layer unit, two linear layer units, and an activation function; Inputting the degradation data sample into the input module, processing it through the linear layer and the position information encoding unit to obtain a linear degradation data sample; scaling the linear degradation data sample through the hourglass feature extractor, and fusing features of different scales to obtain a multi-scale fusion feature map; inputting the multi-scale fusion feature map into the three-layer masked self-attention mechanism to obtain time correlation features of different scales, and connecting them by the feed-forward neural network to obtain an encoded feature map; Inputting the encoded feature map into the decoder, performing query attention based on the single-layer multi-head attention mechanism to obtain an attention result vector; inputting the attention result vector into the output module, and successively passing through the flattening layer unit, the two linear layer units, and the activation function to obtain a predicted value sample of the remaining useful life; training the initial prediction model for the remaining useful life of the engine to obtain a trained and complete prediction model for the remaining useful life of the engine; Obtaining the degradation data of the engine to be tested, and determining the remaining useful life of the engine to be tested based on the trained and complete prediction model for the remaining useful life of the engine.

2. The method for predicting the remaining service life of an engine according to claim 1, wherein Obtaining a degradation data sample of a sample engine, including: Obtaining the initial degradation data of the sample engine; Deleting the constant degradation data in the initial degradation data to obtain transitional degradation data; Performing normalization processing and correction processing on the transitional degradation data to obtain a degradation data sample.

3. The method for predicting the remaining service life of an engine according to claim 2, wherein Before performing normalization processing and correction processing on the transitional degradation data, it further includes: Setting a linear regression model to train the transitional degradation data to obtain an estimated regression coefficient of the transitional degradation data; Obtaining the sequence average value of the transitional degradation data; Concatenating the transitional degradation data with the estimated regression coefficient and the sequence average value respectively.

4. The method for predicting the remaining service life of an engine according to claim 1, characterized in that, The hourglass feature extractor includes two downsampling convolutional layers, two upsampling convolutional layers, and a 1×1 one-dimensional convolutional layer; extracting data features of the linear degradation data sample through the hourglass feature extractor, masking through the three-layer masked self-attention mechanism, and connecting data by the feed-forward neural network to obtain a remaining useful life encoded sample, including: The linear degradation data samples are sequentially passed through the two downsampling convolutional layers and the two upsampling convolutional layers to obtain upsampled feature data and downsampled feature data respectively; According to the one-dimensional convolutional layer, the data with equal time steps in the upsampled feature data and the downsampled feature data are added and fused correspondingly, and normalized to obtain fused feature data; The fused feature data is input into the mask of the three-layer masked self-attention mechanism, and data connection is performed by the feed-forward neural network to obtain the remaining useful life coding samples.

5. The method for predicting the remaining service life of an engine according to claim 4, wherein, The three-layer masked self-attention mechanism includes three-layer pyramid self-attention layers and a masked self-attention layer; the multi-scale fused feature map is input into the three-layer masked self-attention mechanism to obtain time correlation features at different scales, and connection is performed by the feed-forward neural network to obtain an encoded feature map, including: The multi-scale fused feature map is input into the three-layer pyramid self-attention layer, and the masked self-attention layer first performs masking processing on the multi-scale fused feature map, selects to obtain child nodes, A nodes, and parent nodes, and sets the remaining nodes to negative infinity; Among them, the principle formula for the masked self-attention layer to first perform masking processing on the multi-scale fused feature map is: in, Respectively represent s Tier l All A nodes, child nodes and parent nodes corresponding to the data point, Indicates the s Tier j nodes, A, C, and P represent the number of child nodes and parent nodes of node A. s Indicates the number of layers.

6. The method for predicting the remaining service life of an engine according to claim 1, wherein Obtaining a trained and complete remaining useful life prediction model for the engine further includes: Establishing evaluation indicators to evaluate the test results of the remaining useful life prediction model for the engine; According to the evaluation results, adjust the relevant parameters of the remaining useful life prediction model for the engine to obtain a trained and complete remaining useful life prediction model for the engine; Among them, the evaluation indicators include: Among them, n represents the total number of aero-engines, represents the i predicted remaining service life value of the th aero-engine, i and i represents the remaining service life of the i th aero-engine after life correction.

7. An engine remaining service life prediction device, characterized in that Including: A sample acquisition module for acquiring degradation data samples of a sample engine; A model establishment module for establishing an initial remaining useful life prediction model for the engine, where the initial remaining useful life prediction model for the engine follows an encoder-decoder structure, and the encoder includes an hourglass feature extractor and a masked self-attention mechanism; the initial remaining useful life prediction model for the engine includes an input module, an encoding module, a decoding module, and an output module; the input module includes a linear layer and a position information encoding unit; the encoding module includes an hourglass feature extractor, a three-layer masked self-attention mechanism, and a feed-forward neural network; the decoding module includes a single-layer multi-head attention mechanism; the output module includes a flattening layer unit, two linear layer units, and an activation function; A model training module, configured to input the degraded data samples into the input module, process them through the linear layer and the position information encoding unit to obtain linearly degraded data samples; scale the linearly degraded data samples through the hourglass feature extractor, and fuse features of different scales to obtain a multi-scale fusion feature map; input the multi-scale fusion feature map into the three-layer masked self-attention mechanism to obtain time correlation features of different scales, and connect them through the feed-forward neural network to obtain an encoded feature map; input the encoded feature map into the decoder, perform query attention based on the single-layer multi-head attention mechanism to obtain an attention result vector; input the attention result vector into the output module, and successively pass through the flattening layer unit, the two linear layer units and the activation function to obtain the remaining useful life prediction value samples; train the initial remaining useful life prediction model of the engine to obtain a trained remaining useful life prediction model of the engine. A remaining useful life determination module, configured to obtain the degraded data of the engine to be tested, and determine the remaining useful life of the engine to be tested based on the trained remaining useful life prediction model of the engine.

8. An electronic device, characterized in that, It includes a processor and a memory, and a computer program is stored on the memory. When the computer program is executed by the processor, the method for predicting the remaining useful life of the engine according to any one of claims 1-6 is implemented.

Images