A method for predicting engine remaining life based on clustering and LSTM

Abstract

A method for predicting the remaining life of an engine based on clustering and LSTM. The invention relates to the field of aviation engine health management, and specifically to a method for predicting the remaining life of an engine based on clustering and LSTM. A method for predicting the remaining life of an engine based on clustering and LSTM is proposed, which uses clustering and LSTM to monitor the health status of an aircraft engine, predict the remaining life of the engine, and thus ensure the normal operation of the flight. The prediction is performed in the following steps: 1) standardization; 2) cluster analysis; 3) mutation analysis; 4) LSTM model prediction; 5) calculation of weight values; 6) multi-stage prediction. Experimental analysis shows that the ILSTMC method we proposed is effective in predicting the remaining life of an aviation engine, and enables operators to better manage the health of the engine based on the prediction results of this case.

Description

Technical Field

[0001] The present invention relates to the field of aviation engine health management, and in particular to a method for predicting the remaining life of an engine based on clustering and LSTM. Background Art

[0002] As the core component of an aircraft, aircraft engines are extremely important for engine health management due to the complexity and variability of their internal components and the long-term operation in a very high-intensity working environment. Predicting the remaining life of aircraft engines is of great significance for improving the safe and reliable operation of aircraft engines and ensuring aircraft flight safety.

[0003] Currently, the remaining life prediction of engines is mostly based on a single-stage remaining life prediction method, which is not accurate for the remaining life prediction of the engine and cannot be well applied to the health management of the engine. Summary of the invention

[0004] In view of the above problems, the present invention proposes an engine remaining life prediction method based on clustering and LSTM, which uses clustering and LSTM to monitor the health status of aircraft engines and predict the remaining life of the engine, thereby ensuring the normal operation of flights.

[0005] The technical solution of the present invention is: prediction is performed according to the following steps:

[0006] 1) Standardization: Input engine data and perform sample normalization and averaging processing;

[0007] 2) Cluster analysis: Based on step 1), the engine performance degradation data of multiple stages are accurately obtained through cluster analysis;

[0008] 3) Mutation analysis: Based on step 2), ILSTMC is used to predict the degradation data of multiple stages. In the prediction process, memory blocks are used instead of ordinary hidden layer nodes to ensure the storage of information and avoid the problems of long-term dependence and gradient disappearance.

[0009] 4) LSTM model prediction: use mutation points for prediction;

[0010] 5) Calculate weight value: Calculate the weight value of each gate according to the LSTM algorithm formula;

[0011] 6) Multi-stage prediction: The data is sequentially correlated. First, clustering is performed to divide it into multiple stages of LSTM, and then prediction is made.

[0012] Step 1) is specifically as follows: first, the input multi-dimensional aircraft engine performance monitoring data is normalized and averaged to [0, 1] to avoid the possibility of positive and negative offset, and then the normalized result is averaged. The specific calculation formula for normalization is as follows:

[0013]

[0014] Among them, x norma l is the normalized data, x t is the original data, x max is the maximum value in the original data sample, x min The normalized and averaged data is defined as the overall health score (OHS) of the engine.

[0015] Step 2) is as follows: use the sum of squared error (SSE) within the cluster to find the optimal number of clusters k, and determine the number of clusters by observing its inflection point graph. The formula is as follows:

[0016]

[0017]

[0018] ω (i,j) =[X1+X2+···+X k ] -1 (4)

[0019] Among them, X k represents the probability of each cluster, ω represents the probability number of the cluster to which the sample belongs, m is the fuzzy coefficient, which is generally taken as 2, and μ is the center point of each cluster.

[0020] Step 3) is as follows: cluster analysis is used to obtain clusters and mutation points, and the engine data is divided into multiple stages. The emergence of mutation points may cause the remaining life of the engine to change, so the processing of these mutation points will become the basis for clustering. The detection of mutation points mainly calculates the distance between each object and these central objects. The specific calculation is shown in formula (2) to formula (4).

[0021] The error at time t-1 can be calculated by formula (11). At the mutation point, the value of the activation function sigmoid of the forget gate, that is, the value of σ, is approximately 0 and the input gate is approximately 1. The cells at the previous moment cannot be transmitted through time and saved to the current moment, and a mutation point will appear.

[0022] Step 4) is specifically as follows: using the mutation point for prediction, the flow of information in the state of the hidden layer cells can be controlled by the input gate, forget gate, and output gate whose value range of the error term σ is [0, 1]:

[0023]

[0024] Among them, C t is the long-term state of the cell at the current moment, F t is the value of the forget gate at the current moment, C t-1 is the cell state at time t-1, I t is the input at the current moment, is the current state, and the forget gate determines the state of the previous moment C t-1 How much of the cell state is retained to the current moment C t .

[0025] The LSTM model removes or adds information to the cell state by carefully designing the structure of three gates, namely the input gate, the forget gate, and the output gate, which are used to protect and update the cell state. The input gate allows information to enter the storage cell, the forget gate allows the cell information to be forgotten or removed from the input storage cell, and the output gate outputs information from the input storage cell.

[0026] The three gates work closely together to update the cell information. First, the input gate uses tanh to extract valid information, and then uses the activation function sigmoid to filter the information and determine the degree of cell information update. t It is a scalar that measures cell updates. The output of the output layer is h t The hidden layer will serve as the output of LSTM.

[0027] Input Gate:

[0028]

[0029]

[0030] Forget Gate:

[0031]

[0032] Output Gate:

[0033]

[0034]

[0035] Among them, σ is the value of the activation function sigmoid, W i is the weight value of the input gate, h t-1is the hidden state value of the cell at time t-1, and b represents the bias term. f is the weight of the forget gate, W o is the weight of the output gate. As the number of data points in the prediction process increases, each weight gradient at time t in the prediction process can be calculated through formula (12) to formula (15). Then, the final weight gradient after summation is obtained by substituting formula (16) to formula (19) to update the weight.

[0036] The specific calculation steps of step 5) are as follows:

[0037] 5.1) Reversely calculate the value of the error term σ corresponding to each moment. The back propagation of the error term is divided into two directions. The first is the back propagation along time, that is, starting from the current moment t, calculating the error term σ-1 at moment t-1; and the second is the propagation of the error term to the upper layer.

[0038]

[0039] 5.2) Calculation of weight gradient: According to the corresponding error term, calculate the gradient of each weight, that is, W fh , W ch , W oh , W ih The weights of , their gradients are the sum of the gradients at each moment. Therefore, we need to first find their gradients at time t, and then find their final gradients.

[0040] Find the error term σ at time t o,t , σ f,t , σ i,t , σ c,t , and further find the weight value W at time t fh , W ch , W oh , W ih Right now:

[0041]

[0042]

[0043]

[0044]

[0045] By adding together the gradients at each moment, we can get the final gradient:

[0046]

[0047]

[0048]

[0049]

[0050] In the LSTM prediction process, the data of the first category obtained by the clustering process is directly predicted as the data of the first stage of prediction; the data of the first and second categories are predicted as the data of the second stage, and so on; the actual value and predicted value of the overall health score OHS of the engine are obtained.

[0051] Step 6) is specifically as follows: for multi-stage remaining life prediction, the data are sequentially associated with each other. First, clustering is performed to divide the data into multiple stages of LSTM, and then prediction is performed. In order to obtain a more accurate prediction effect, the data of the first category obtained by clustering is directly predicted as the data of the first stage of prediction; then the data is iterated for further prediction, and the data of the first and second categories obtained by clustering are predicted as the data of the second stage; the data of the first, second, and third categories obtained by clustering are predicted as the data of the third stage, and then the model comparison analysis is performed.

[0052] This paper studies the prediction of the remaining life of the engine based on clustering and LSTM. Since the deep learning method can learn complex data features, the multi-stage degradation characteristics of the engine can be analyzed by combining clustering to predict the multi-stage remaining life of the engine, which can be applied to the health management of the engine to a certain extent and has certain practical significance.

[0053] The present invention realizes the fusion of multi-dimensional data, establishes a prediction model based on the results of multi-stage information fusion, introduces real-time monitoring data using the parameters of the multi-stage prediction model, and adopts a cluster analysis and LSTM fusion algorithm to realize the update and prediction of the multi-stage model parameters, and finally predicts the performance degradation trend of the aircraft engine and obtains an accurate prediction of the remaining service life of the aircraft engine.

[0054] The present invention has the following beneficial effects: the present invention establishes a new prediction model based on the performance parameters of aircraft engines. The model combines the advantages of LSTM neural network and K-means clustering. K-means clustering has the advantages of short time, fast speed, simple and easy to explain in terms of calculation. LSTM can use memory blocks instead of ordinary hidden layer nodes in prediction to ensure that the information is stored across any delay and the error signal is returned to a time point long ago, so that the network learns to "forget" and stay away from saturation, avoiding long-term dependence, gradient disappearance and explosion. Through experimental analysis and verification of the algorithm, it is found that the error calculated by ILSTMC when predicting the remaining life is better than other traditional models. The experimental analysis shows that when predicting the remaining life of aircraft engines, the ILSTMC method we proposed is effective, which enables operators to better manage the health of the engine in combination with the prediction results of this case, which is of great significance to ensuring the flight safety of aircraft. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is a schematic diagram of the structural framework of the present invention;

[0056] Figure 2 This is the OHS homogenization diagram of the 83rd engine;

[0057] Figure 3 This is the cluster analysis result diagram of the 83rd engine;

[0058] Figure 4(a) shows the ILSTMC stage 1 prediction diagram for the 83rd engine;

[0059] Figure 4(b) shows the ILSTMC stage 2 prediction diagram for the 83rd engine;

[0060] Figure 4(c) shows the ILSTMC stage 3 prediction diagram for the 83rd engine;

[0061] Figure 5 This is the RMSE comparison chart of the prediction results of ILSTMC and other algorithms for the 83rd engine;

[0062] Figure 6 Comparison of multi-stage life cycle prediction errors for the 83rd engine. DETAILED DESCRIPTION

[0063] In order to clearly illustrate the technical features of this patent, this patent is elaborated in detail below through a specific implementation method and in combination with its accompanying drawings.

[0064] The present invention Figure 1-6 As shown in FIG. 1 , the aircraft engine remaining life prediction method based on clustering and LSTM of the present invention mainly includes two modules: clustering and prediction.

[0065] The clustering module includes: the input engine data is normalized to [0, 1] and then the average value is calculated based on the advantages of K-means clustering, such as short calculation time, fast speed, simple and easy to explain. Secondly, the sum of squared error (SSE) within the cluster is used to find the optimal number of clusters, and the number of clusters is determined by observing its inflection point diagram. Finally, cluster analysis is used to obtain clusters and mutation points, and the engine data is divided into multiple stages. The emergence of mutation points may cause changes in the remaining life of the engine, so the processing of these mutation points will become the basis for clustering. The detection of mutation points mainly calculates the distance between each object and these central objects.

[0066] The prediction module includes: multi-stage prediction based on mutation points obtained by clustering. In multi-stage prediction, the data are sequentially associated. First, clustering is performed to divide the data into multiple stages of LSTM, and then prediction is performed. In order to obtain a more accurate prediction effect, the data of the first category obtained by clustering is directly predicted as the data of the first stage of prediction; then the data is iterated for further prediction, and the data of the first and second categories obtained by clustering are predicted as the data of the second stage; the data of the first, second, and third categories obtained by clustering are predicted as the data of the third stage, and then the model comparison analysis is performed. Finally, the prediction error value of each engine is calculated. Considering the prediction of the remaining life of the engine, the life cycle of the engine is divided, and the prediction errors of the five stages of 60%, 70%, 80%, 90%, and 95% are taken out respectively, and compared with the errors of each method. The prediction results are brought into the cost function to obtain RMSE, and compared with the prediction results of the traditional neural network LSTM, RNN, and linear regression method LP method.

[0067] like Figures 1 to 5 As shown, the specific steps of the engine remaining life prediction method based on clustering and LSTM of the present invention are as follows:

[0068] 1) Standardization: Input engine data and perform sample normalization and averaging processing; First, input multi-dimensional aircraft engine performance monitoring data and perform normalization and averaging processing on these data in order to improve the convergence speed and calculation accuracy of the model.

[0069] First, the input multi-dimensional aircraft engine performance monitoring data is normalized and averaged to [0, 1] to avoid the possibility of positive and negative offset. Then, the normalized results are averaged. The specific calculation formula for normalization is as follows:

[0070]

[0071] Among them, x norma l is the normalized data, x t is the original data, x max is the maximum value in the original data sample, x min The normalized and averaged data is defined as the overall health score (OHS) of the engine.

[0072] 2) Cluster analysis: Based on step 1), the engine performance degradation data of multiple stages are accurately obtained through cluster analysis;

[0073] Cluster analysis uses the sum of squared error (SSE) within the cluster to find the optimal number of clusters k, and the number of clusters is determined by observing its inflection point graph. The formula is as follows:

[0074]

[0075]

[0076] ω (i,j) =[X1+X2+···+X k ] -1 (4)

[0077] Among them, X k represents the probability of each cluster, ω represents the probability number of the cluster to which the sample belongs, m is the fuzzy coefficient, which is generally taken as 2, and μ is the center point of each cluster.

[0078] 3) Mutation analysis: Based on step 2), ILSTMC is used to predict the degradation data of multiple stages. In the prediction process, memory blocks are used instead of ordinary hidden layer nodes to ensure the storage of information and avoid the problems of long-term dependence and gradient disappearance.

[0079] Cluster analysis is used to obtain clusters and mutation points, and the engine data is divided into multiple stages. The emergence of mutation points may cause the remaining life of the engine to change, so the processing of these mutation points will become the basis for clustering. The detection of mutation points mainly calculates the distance between each object and these central objects. The specific calculation is shown in formula (2) to formula (4).

[0080] The error at time t-1 can be calculated by formula (11). At the mutation point, the value of the activation function sigmoid of the forget gate, that is, the value of σ, is approximately 0 and the input gate is approximately 1. The cells at the previous moment cannot be transmitted through time and saved to the current moment, and a mutation point will appear.

[0081] 4) LSTM model prediction: Using mutation points for prediction, the flow of information in the state of hidden layer cells can be controlled through the input gate, forget gate, and output gate whose error term σ has a value range of [0, 1]:

[0082]

[0083] Among them, C t is the long-term state of the cell at the current moment, F t is the value of the forget gate at the current moment, C t-1 is the cell state at time t-1, I t is the input at the current moment, is the current state, and the forget gate determines the state of the previous moment C t-1 How much of the cell state is retained to the current moment C t .

[0084] The LSTM model removes or adds information to the cell state by carefully designing the structure of three gates, namely the input gate, the forget gate, and the output gate, which are used to protect and update the cell state. The input gate allows information to enter the storage cell, the forget gate allows the cell information to be forgotten or removed from the input storage cell, and the output gate outputs information from the input storage cell.

[0085] The three gates work closely together to update the cell information. First, the input gate uses tanh to extract valid information, and then uses the activation function sigmoid to filter the information and determine the degree of cell information update. t It is a scalar that measures cell updates. The output of the output layer is h t The hidden layer will serve as the output of LSTM.

[0086] Input Gate:

[0087]

[0088]

[0089] Forget Gate:

[0090]

[0091] Output Gate:

[0092]

[0093]

[0094] Among them, σ is the value of the activation function sigmoid, W i is the weight value of the input gate, h t-1is the hidden state value of the cell at time t-1, and b represents the bias term. f is the weight of the forget gate, W o is the weight of the output gate. As the number of data points in the prediction process increases, each weight gradient at time t in the prediction process can be calculated through formula (12) to formula (15). Then, the final weight gradient after summation is obtained by substituting formula (16) to formula (19) to update the weight.

[0095] 5) Calculate weight value: Calculate the weight value of each gate according to the LSTM algorithm formula. The specific calculation steps are as follows:

[0096] 5.1) Reversely calculate the value of the error term σ corresponding to each moment. The back propagation of the error term is divided into two directions. The first is the back propagation along time, that is, starting from the current moment t, calculating the error term σ-1 at moment t-1; and the second is the propagation of the error term to the upper layer.

[0097]

[0098] 5.2) Calculation of weight gradient: According to the corresponding error term, calculate the gradient of each weight, that is, W fh , W ch , W oh , W ih The weights of , their gradients are the sum of the gradients at each moment. Therefore, we need to first find their gradients at time t, and then find their final gradients.

[0099] Find the error term σ at time t o,t , σ f,t , σ i,t , σ c,t , and further find the weight value W at time t fh , W ch , W oh , W ih Right now:

[0100]

[0101]

[0102]

[0103]

[0104] By adding together the gradients at each moment, we can get the final gradient:

[0105]

[0106]

[0107]

[0108]

[0109] In the LSTM prediction process, the data of the first category obtained by the clustering process is directly predicted as the data of the first stage of prediction; the data of the first and second categories are predicted as the data of the second stage, and so on; the actual value and predicted value of the overall health score OHS of the engine are obtained.

[0110] 6) Multi-stage prediction: For multi-stage remaining life prediction, the data are sequentially associated. First, clustering is performed to divide the data into multiple stages of LSTM, and then prediction is performed. In order to obtain a more accurate prediction effect, the data of the first category obtained by clustering is directly predicted as the data of the first stage of prediction; then the data is iterated for further prediction, and the data of the first and second categories obtained by clustering are predicted as the data of the second stage; the data of the first, second, and third categories obtained by clustering are predicted as the data of the third stage, and then a model comparison analysis is performed. The prediction results of ILSTMC are substituted into the cost function to obtain RSME, and compared with the prediction results of traditional neural network LSTM, RNN, and linear regression LP methods.

[0111] The prediction error value of each engine is calculated. Taking into account the prediction of the remaining life of the engine, the engine life cycle is divided into five stages: 60%, 70%, 80%, 90%, and 95%, and the prediction errors are taken out for comparison with the errors of each method.

[0112] The prediction results are substituted into the cost function to obtain the RMSE, and then compared with the prediction errors of the traditional neural network LSTM, RNN, and linear regression LP methods.

[0113] experiment

[0114] The data used in the experiment is the C-MAPSS data set on the NASA website for verification. The C-MAPSS data set has 4 subsets, each with 27 dimensions, of which the first 2 dimensions are the engine number and the number of times the engine has run, 3 dimensions correspond to the operating environment settings, and 22 dimensions correspond to the gas path performance data of the sensor. Each subset has a different operating environment and number of faults.

[0115] Gas path parameters reflect the health status of each gas path component or unit, so gas path performance parameters are key data that affect the overall performance of the engine and play a leading role in the prediction process. Using the seven characterization parameters in the gas path performance parameters as the determining factors of the overall health of the engine, multi-dimensional parameter data fusion is performed to depict the performance degradation trajectory of the engine.

[0116] Assume that the sequence length of the sample is sequence len=5, and the number of neurons input at each moment is 1. The training set and test set of the data are brought into the neural network for training and testing in turn. In the Adam algorithm of LSTM, one sample value is taken out for batch processing each time, and the learning rate η is set to 0.01. The hidden layer size is 2 layers, and the number of nodes in each layer is 16 and 8 respectively. The training set and test set take 5 epoch cycles respectively to predict the remaining life of the engine.

[0117] After calculation, the cluster number k of each engine is determined. The k value also represents the specific number of stages. Then, model prediction is performed on each engine separately. The prediction results corresponding to each k value are compared, and the cluster number with the smallest error is selected as the specific cluster number of each engine.

[0118] Taking the 83rd engine as an example, when the ILSTMC model based on clustering and LSTM is used for prediction, the following comparison can illustrate the superiority of the prediction method based on clustering and LSTM:

[0119] As shown in Figure 4, compared with RNN, LSTM, and linear regression methods, it is found that the ILSTMC model will make the engine's overall health score OHS value closer to the true value as the time point increases during the prediction process;

[0120] In the case of multi-stage prediction, the RMSE of the last stage is reduced by 0.67% compared with LSTM, and the average RMSE of each stage is reduced by 2.39% compared with LSTM;

[0121] like Figure 5 ,For the whole life cycle prediction, the error value is reduced by 0.58% compared with LSTM in the 90%-95% cases of the life cycle, and the average error of each stage of life is reduced by 2.64% compared with LSTM.

[0122] There are many specific implementation ways of the present invention. The above is only the preferred implementation mode of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements can be made without departing from the principle of the present invention. These improvements should also be regarded as the protection scope of the present invention.

Claims

1. A method for predicting the remaining life of an engine based on clustering and LSTM, characterized in that: The steps include: 1) Input engine data and perform sample normalization and averaging processing; 2) Based on step 1), the engine performance degradation data of multiple stages is accurately obtained through cluster analysis, including: The sum of squared errors within the cluster is used to find the optimal number of clusters k, and the number of clusters is determined by observing its inflection point diagram. The formula is as follows: oh (i,j) =[X1+X2+···+X k ] -1 (4) Among them, X k represents the probability of each cluster, ω represents the probability number of the cluster to which the sample belongs, m is the fuzzy coefficient, which is 2, and μ is the center point of each cluster; 3) Based on step 2), clustering and LSTM are used to predict the degradation data of multiple stages. In the prediction process, memory blocks are used instead of ordinary hidden layer nodes to ensure the storage of information and avoid the problems of long-term dependence and gradient disappearance. Specifically: Cluster analysis is used to obtain clusters and mutation points, and the engine data is divided into multiple stages. The detection of mutation points is to calculate the distance between each object and these central objects. The specific calculation is shown in formula (2) to formula (4). Calculate the error at time t-1. At the mutation point, the value of the activation function sigmoid of the forget gate, that is, the value of σ, is approximately 0 and the input gate is approximately 1. The cells at the previous moment cannot be transmitted through time and saved to the current moment, and a mutation point will appear. 4) LSTM model prediction: Use mutation points for prediction, and use the data of the first category obtained by the clustering process as the data of the first stage of prediction for direct prediction; the data of the first and second categories are used as the data of the second stage for prediction, and so on; the actual value and predicted value of the overall health score OHS of the engine are obtained; 5) Calculate weight value: Calculate the weight value of each gate according to the LSTM algorithm formula; 6) Multi-stage prediction: The data are sequentially related. First, clustering is performed to divide the data into multiple stages of LSTM, and then prediction is performed. The specific prediction process includes: In order to obtain more accurate prediction results, the data of the first category obtained by clustering is directly predicted as the data of the first stage of prediction; secondly, the data is iterated for further prediction, and the data of the first and second categories obtained by clustering are used as the data of the second stage for prediction. Finally, the data of the first, second, and third categories obtained by clustering are used as the data of the third stage for prediction. Then, a comparative analysis of the models is performed, and the prediction results of clustering and LSTM are brought into the cost function to calculate RSME.

2. The engine remaining life prediction method based on clustering and LSTM as claimed in claim 1, characterized in that: Step 1) is specifically as follows: first, the input multi-dimensional aircraft engine performance monitoring data is normalized and averaged to [0, 1] to avoid the possibility of positive and negative offset, and then the normalized result is averaged. The specific calculation formula for normalization is as follows: Among them, x normal is the normalized data, x t is the original data, x max is the maximum value in the original data sample, x min The normalized and averaged data is defined as the overall health score OHS of the engine.

3. The engine remaining life prediction method based on clustering and LSTM as claimed in claim 2, characterized in that: Step 4) is specifically as follows: using the mutation point for prediction, the flow of information in the state of the hidden layer cells is controlled by the input gate, forget gate and output gate whose value range of the error term σ is [0, 1]: Among them, C t is the long-term state of the cell at the current moment, F t is the value of the forget gate at the current moment, C t-1 is the cell state at time t-1, I t is the input at the current moment, is the current state, and the forget gate determines the state of the previous moment C t-1 How much of the cell state is retained to the current moment C t ; The LSTM model removes or adds information to the cell state by carefully designing the structure of three gates, namely the input gate, the forget gate, and the output gate, which are used to protect and update the cell state; the input gate allows information to enter the storage cell, the forget gate allows the cell information to be forgotten or removed from the input storage cell, and the output gate outputs information from the input storage cell; The three gates work closely together to update the cell information. First, the input gate uses tanh to extract valid information, and then uses the activation function sigmoid to filter the information and determine the degree of cell information update. t It is a scalar that measures cell updates. The output of the output layer is h t The hidden layer will be the output of LSTM; Input Gate Forget Gate Output Gate Among them, σ is the value of the activation function sigmoid, W i is the weight value of the input gate, h t-1 is the hidden state value of the cell at time t-1, b represents the bias term, W f is the weight of the forget gate, W o It is the weight of the output gate. As the data points in the prediction process increase, the weight value at time t is calculated, and each weight gradient at time t in the prediction process is calculated. The gradients at each moment are added together to obtain the final weight gradient after summation, and the weight is updated.

4. The method for predicting the remaining life of an engine based on clustering and LSTM according to claim 3, characterized in that: The specific calculation steps of step 5) are as follows: 5.1) Reversely calculate the value of the error term σ corresponding to each moment. The back propagation of the error term is divided into two directions: back propagation along time, that is, starting from the current moment t, calculating the error term σ-1 at moment t-1; propagating the error term to the upper layer; 5.2) Calculation of weight gradient: According to the corresponding error term, calculate the gradient of each weight, that is, W fh , W ch , W oh , W ih The weights of , their gradients are the sum of the gradients at each moment, so we need to first find their gradients at time t, and then find their final gradients; Find the error term σ at time t o,t , σ f,t , σ i,t , σ c,t , and further find the weight value W at time t fh , W ch , W oh ,Right now: By adding together the gradients at each moment, we can get the final gradient:

5. The method for predicting the remaining life of an engine based on clustering and LSTM according to claim 4, characterized in that: It also includes a comparison of the prediction results with traditional neural networks LSTM, RNN, and linear regression LP methods.

Images