An aircraft engine life prediction method based on improved LSTM

Abstract

The present invention belongs to the technical field of aero-engines, and specifically relates to an aero-engine life prediction method based on an improved LSTM. It comprises: step 1: processing the original sensor data acquired by the sensor, constructing a training sample, and the training sample includes a training set and a test set; step 2: constructing an LSTM structural model based on the training sample constructed in step 1 as an engine remaining life prediction model; step 3: inputting the test set of step 1 into the LSTM structural model constructed in step 2 to obtain a predicted RUL value, and using RMSE and Score evaluation indicators to evaluate the obtained predicted RUL value. The present invention proposes an aero-engine life prediction method of SDAE and LSTM, which utilizes the advantages of unsupervised feature extraction of deep encoders to effectively extract features from engine sensor signals, avoids the low efficiency of manually extracting features and the resulting prediction uncertainty, and utilizes the advantages of LSTM model processing time series data to predict the remaining life of the engine.

Description

Technical Field

[0001] The invention belongs to the technical field of aero-engines, and in particular relates to an aero-engine life prediction method based on an improved LSTM. Background Art

[0002] Aircraft engines are important components for the normal flight of aircraft. Due to the complex and changeable operating conditions of engines and the relatively harsh operating environment, once a failure occurs, it will pose a huge threat to flight safety and the safety of passengers' lives. The remaining life prediction of aircraft engines is based on condition monitoring data, such as the temperature, pressure, speed and other historical data of the fan and compressor inlet and outlet, and features are extracted to build a life prediction model, providing technical support for preventive maintenance, which has a wide range of application value.

[0003] In recent years, data-driven methods have gradually become the mainstream technology in the field of remaining life prediction. There are two main types of commonly used data-driven methods: one is a data-driven algorithm based on statistical analysis, and the other is an intelligent learning algorithm based on deep learning.

[0004] The first type of method is to model the degradation of nonlinear random coefficients and propose a method for predicting the remaining life that integrates prior information and field information. This method can effectively use historical data for modeling and relatively accurately predict the actual remaining life of the equipment. However, since this method uses strict mathematical modeling, it often requires relatively ideal assumptions, which limits its development and application.

[0005] The second method is based on deep learning intelligent learning algorithms, which do not need to build a specific mathematical model. Instead, they analyze sensor monitoring data, extract features that describe engine performance degradation, and use deep learning technology to predict the remaining life of the engine. First, the high-dimensional features are reduced in dimension, and then the life prediction model is built. However, the characteristics of the time series degradation of engine sensor data are ignored. When building the model, the correlation between the past moment and the current moment cannot be paid attention to.

[0006] There are also fault prediction methods based on bidirectional LSTM (long short-term memory, LSTM) neural networks, which can process sequence data features in both forward and reverse directions. However, since it only uses a single-layer network structure, the feature extraction capability and prediction accuracy need to be improved. Deep learning is an effective method to solve the problem of engine life prediction. Compared with traditional statistical methods, it can process larger sample sizes of data, so the prediction reliability is higher.

[0007] Feature extraction is an important prerequisite for using deep learning technology to predict remaining life. Before life prediction, sensor data needs to be preprocessed, such as feature extraction, feature fusion, etc. There are many research results on feature extraction. Kernel principal component analysis (KPCA) can be used to remove information redundancy between features and construct a simplified feature matrix. The complexity of the algorithm is closely related to the dimension of the data. Since the monitoring data collected by the engine sensor has the characteristics of high dimension, multi-parameter and large scale, the feature extraction model is highly complex. Summary of the invention

[0008] In view of the problems existing in the prior art, the present invention proposes an aircraft engine life prediction method that integrates a deep sparse autoencoder (Sparse Deep AutoEncoder, SDAE) and a long short-term memory network, introduces the SDAE model, utilizes the advantages of unsupervised feature extraction of the deep encoder, effectively extracts features from engine sensor signals, avoids the inefficiency of manual feature extraction and the resulting prediction uncertainty, and utilizes the advantages of the LSTM model in processing time series data to predict the remaining life of the engine.

[0009] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0010] An aircraft engine life prediction method based on improved LSTM includes the following steps:

[0011] Step 1: Process the original sensor data obtained by the sensor and construct training samples, which include training sets and test sets;

[0012] Step 2: Based on the training samples constructed in step 1, an LSTM structure model is constructed as the engine remaining life prediction model;

[0013] Step 3: Input the test set of step 1 into the LSTM structure model constructed in step 2 to obtain the predicted RUL value. Use RMSE and Score evaluation indicators to evaluate the obtained predicted RUL value and evaluate the life prediction effect.

[0014] Preferably, the specific steps of step 1 include:

[0015] Step 1.1: Normalize and standardize the raw sensor data acquired by the sensor;

[0016] Step 1.2: Construct a RUL tag for the raw sensor data processed in step 1.1;

[0017] Step 1.3: Use a piecewise linear function to process the label data RUL in step 1.2, set RUL in the early cycle to a constant value, and start linear degradation after the cycle until it reaches 0;

[0018] Step 1.4: Perform feature selection on the data processed in step 1.3;

[0019] Step 1.5: Construct training samples for the data after feature selection in step 1.4. The training samples include training set and test set.

[0020] Preferably, the specific method of the normalization and standardization processing in step 1.1 is to use the Min-Max model for normalization, as shown in formula (1), to convert the normalized data into a distribution with a mean of 0 and a standard deviation of 1;

[0021]

[0022] Data standardization is shown in formula (2):

[0023]

[0024] In formula (1) and (2), x′ i,j (t) represents the dimensionless sample, x i,j (t) represents the original sample, max(x :,j ) represents the maximum value of samples of the same dimension, min(x :,j ) represents the minimum value of the same dimension; represents the sample mean; s represents the sample standard deviation.

[0025] Preferably, the specific steps of step 2 include:

[0026] Step 2.1: Build an LSTM structure model;

[0027] Step 2.2: Input the test set in step 1 into SDAE, extract the deep features of the training set data in step 1.5 through SDAE encoding, perform unsupervised pre-training on the training set data, and construct the engine health factor HI curve through training to characterize the engine degradation trend.

[0028] Step 2.3: Use the time series data encoded by SDAE in step 2.2 as the input of the LSTM structure model, use the reverse number of the engine's full life cycle as the label of the LSTM structure model, set the learning rate and the number of hidden layer neurons, and then train the LSTM structure model;

[0029] Step 2.4: Calculate the loss of LSTM structure model training and update the LSTM structure model parameters.

[0030] Preferably, in step 2.1, a double-layer LSTM structure model is constructed as the engine remaining life prediction model. The forget gate, input gate and output gate of the double-layer LSTM structure model select the sigmoid activation function, denoted as σ, and the output range is [0,1], which represents the importance of the current input information. The closer to 1, the more important it is. When generating candidate memories, the tanh activation function is selected to adjust the value flowing through the network, and the output value is always limited to [-1,1]. Specifically,

[0031] 4) Forget Gate

[0032] f t =σ(W f [h t-1 ,x t ]+b f ) (3)

[0033] 5) Input Gate

[0034] i t =σ(W i [h t-1 ,x t ]+b i ) (4)

[0035]

[0036] 6) Output Gate

[0037] o t =σ(W o [h t-1 ,x t ]+b o ) (6)

[0038]

[0039] h t =o t ⊙tanh(C t ) (8)

[0040] Among them, W and b are model weights and biases, respectively, and x t is the input sample, w f is the forget gate weight, h t-1 is the hidden layer state at time t-1, b f is the forget gate bias, w i is the input gate weight, b i is the input gate bias, h is the hidden layer state at time t, b c is the cell state bias, w c is the cell state weight, w o is the output gate weight, ot Output gate status.

[0041] Preferably, the loss of the LSTM model in step 2.4 is calculated using mean square error:

[0042]

[0043] in, Represents the predicted value of remaining life, y i Represents the true value of remaining life, and n is the sample size.

[0044] Preferably, the LSTM structure model parameter update formula of step 2.4 is as follows, and the moment weight and momentum factor are updated:

[0045] m t =β1·m t-1 +(1-β1)·g t

[0046] w t+1 =w t -α·m t (10)

[0047] Among them, w t is the weight at time t, g t is the gradient at time t, α is the initial learning rate, m t is the first-order momentum factor, and the empirical value of β1 is 0.9.

[0048] Preferably, the specific operations of step 3 include:

[0049] Step 3.1: Input the test set data in step 1.5 into the input gate of the LSTM structure model updated in step 2.4 to obtain the predicted RUL value;

[0050] Step 3.2: Use the root mean square error RMSE and Score function to evaluate the predicted RUL value in step 3.1;

[0051] Step 3.3: Use the predicted life deviation δ to evaluate the life prediction effect.

[0052] Preferably, the root mean square error RMSE evaluation method of step 3.2 is:

[0053]

[0054] In the formula, Represents the predicted value of remaining life, y i Indicates the true value of remaining life. The smaller the RMSE value, the better the prediction effect.

[0055] If the RMSE index is similar, the Score function is used for further evaluation, as shown in formula (12). The smaller the Score, the better the prediction effect.

[0056]

[0057] Preferably, the method of evaluating the life prediction effect by using the predicted life deviation δ in step 3.3 is:

[0058]

[0059] in, Represents the predicted value of remaining life, y i Indicates the true value of remaining life.

[0060] Compared with the prior art, the present invention has the following beneficial effects:

[0061] 1. An aircraft engine life prediction method that integrates deep sparse autoencoder SDAE and long short-term memory network introduces the SDAE model and uses the advantages of unsupervised feature extraction of deep encoders to effectively extract features from engine sensor signals, avoiding the inefficiency of manual feature extraction and the resulting prediction uncertainty;

[0062] 2. After data encoding and dimensionality reduction, the complexity of the prediction model is greatly reduced, and the life prediction effect is significantly improved. The proposed method has strong generalization ability and feasibility, and can effectively predict life. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0064] In the attached picture:

[0065] Figure 1 is a flow chart of the method of the present invention;

[0066] Figure 2 This is the preprocessing result of some sensor data;

[0067] Figure 3 (a) HI curve before filtering, (b) HI curve after filtering;

[0068] Figure 4 (a) shows the life prediction effect of three random engines in the FD001 subset, (b) shows the life prediction effect of three random engines in the FD002 subset, (c) shows the life prediction effect of three random engines in the FD003 subset, and (d) shows the life prediction effect of three random engines in the FD004 subset.

[0069] Figure 5(a) is the error curve of FD001, Figure 5(b) is the error curve of FD003, Figure 5(c) is the error curve of FD002, and Figure 5(d) is the error curve of FD004;

[0070] Figure 6 This is a model structure diagram of the present invention. DETAILED DESCRIPTION

[0071] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0072] Example:

[0073] Refer to the attached Figure 1-6 As shown, an aircraft engine life prediction method based on improved LSTM includes the following steps:

[0074] Step 1: Process the raw sensor data obtained by the sensor and construct training samples. The training samples include training sets and test sets. The specific steps include:

[0075] Step 1.1: The raw sensor data acquired by the sensor is normalized and standardized. The specific method is to use the Min-Max model for normalization, as shown in formula (1), and convert the normalized data into a distribution with a mean of 0 and a standard deviation of 1;

[0076]

[0077] Data standardization is shown in formula (2):

[0078]

[0079] In formula (1) and (2), x i ' ,j (t) represents the dimensionless sample, x i,j (t) represents the original sample, max(x :,j ) represents the maximum value of samples of the same dimension, min(x :,j ) represents the minimum value of the same dimension; represents the sample mean; s represents the sample standard deviation.

[0080] Step 1.2: Construct RUL labels for the raw sensor data processed in step 1.1. The second column in the training set is the number of engine cycles, and the reverse number of the cycle is used as the label of the training set;

[0081] Step 1.3: The label data RUL of step 1.2 is processed by a piecewise linear function. The RUL in the early cycle is set to a constant value, and linear degradation begins after the cycle. Specifically, the value of the part of the RUL label data greater than 125 is set to 125, and the part less than 125 is set to a linear function with a slope of -1 until it reaches 0;

[0082] Step 1.4: Perform feature selection on the data processed in step 1.3. The data of the 1st, 5th, 10th, 16th, 18th, and 19th sensors remain unchanged or change slightly during the entire life cycle of the engine. In order to improve the accuracy of life prediction, such data are eliminated and the remaining data with changes are selected for experiments;

[0083] Step 1.5: Construct training samples for the data after feature selection in step 1.4. The training samples include a training set and a test set. After feature selection, both the training set and the test set are 17 dimensions. The training set includes data from the entire life cycle of the engine, and the test set is the monitoring data of the engine from the first cycle to the cycle before failure.

[0084] Step 2: Based on the training samples constructed in step 1, an LSTM structure model is constructed as the engine remaining life prediction model. The specific steps include:

[0085] Step 2.1: Construct a double-layer LSTM structure model as the engine remaining life prediction model. The forget gate, input gate and output gate of the double-layer LSTM structure model select the sigmoid activation function, denoted as σ, and the output range is [0,1], which represents the importance of the current input information. The closer to 1, the more important it is. When generating candidate memories, select the tanh activation function to adjust the value flowing through the network and always limit the output value to [-1,1]. The details are as follows:

[0086] 1) Forget Gate

[0087] f t =σ(W f [h t-1 ,x t ]+b f ) (3)

[0088] 2) Input Gate

[0089] i t =σ(W i [h t-1 ,x t ]+b i ) (4)

[0090]

[0091] 3) Output gate

[0092] o t =σ(W o [h t-1 ,x t ]+b o ) (6)

[0093]

[0094] h t =o t ⊙tanh(C t ) (8)

[0095] Among them, W and b are model weights and biases, respectively, and x t is the input sample, w f is the forget gate weight, h t-1 is the hidden layer state at time t-1, b f is the forget gate bias, w i is the input gate weight, b i is the input gate bias, h is the hidden layer state at time t, b c is the cell state bias, w c is the cell state weight, w o is the output gate weight, o t Output gate status.

[0096] Step 2.2: Use SDAE as the feature extraction model, input the test set of step 1 into SDAE, extract the deep features of the training set data of step 1.5 through SDAE encoding, select the SDAE encoding network structure as five layers, and select the number of neurons in each layer as 17-16-15-8-1. Perform unsupervised pre-training on the training set data, and construct the engine health factor HI curve through training to characterize the engine degradation trend. The HI curve has noise, poor curve smoothness, and insufficient discrimination, so it is smoothed and filtered. The HI curve shows obvious monotonous increasing or decreasing characteristics throughout the life cycle of the engine, which can better characterize the degradation process of the engine.

[0097] SDAE is a feature expression network that attempts to learn a function through an unsupervised learning algorithm so that the output value is approximately equal to the input value. It consists of an input layer, one or more hidden layers, and an output layer, where the hidden layer h can generate a code to represent the input. The encoding network and the decoding network are generally constructed using neural networks. The encoding process can be expressed as h = f(x), and the decoding process can be expressed as r = g(h). Finally, the input x is approximately equal to the output g(f(x)), and the output of the hidden layer is the encoding result.

[0098] Step 2.3: Use the time series data encoded by SDAE in step 2.2 as the input of the LSTM structure model, use the reverse number of the engine's full life cycle as the label of the LSTM structure model, set the learning rate and the number of hidden layer neurons, and then train the LSTM structure model;

[0099] Step 2.4: Calculate the loss of LSTM structure model training using mean square error:

[0100]

[0101] in, Represents the predicted value of remaining life, y i Represents the true value of remaining life, and n is the sample size.

[0102] Update the LSTM structure model parameters. The LSTM structure model parameter update formula is as follows, and the moment weight and momentum factor are updated:

[0103] m t =β1·m t-1 +(1-β1)·g t

[0104] w t+1 =w t -α·m t (10)

[0105] Among them, w t is the weight at time t, g t is the gradient at time t, α is the initial learning rate, m t is the first-order momentum factor, and the empirical value of β1 is 0.9.

[0106] Step 3: Input the test set in step 1 into the LSTM structure model constructed in step 2 to obtain the predicted RUL value, and use RMSE and Score evaluation indicators to evaluate the predicted RUL value and the life prediction effect. The specific operations include:

[0107] Step 3.1: Input the test set data in step 1.5 into the input gate of the LSTM structure model updated in step 2.4 to obtain the predicted RUL value;

[0108] Step 3.2: Use the root mean square error RMSE and Score function to evaluate the predicted RUL value in step 3.1.

[0109] The root mean square error RMSE evaluation method is:

[0110]

[0111] In the formula, Represents the predicted value of remaining life, y i Represents the true value of the remaining life, when y i and When the deviation is small, the RMSE value is small, and the smaller the value, the better the prediction effect;

[0112] If the RMSE index is similar, the Score function is used for further evaluation, as shown in formula (12). The smaller the Score, the better the prediction effect.

[0113]

[0114] Step 3.3 The method of evaluating the life prediction effect using the predicted life deviation δ is:

[0115]

[0116] in, Represents the predicted value of remaining life, y i It represents the true value of the remaining life. The smaller the life deviation is, the more accurate the prediction is.

[0117] Simulation experiment results and analysis:

[0118] The present invention selects the C-MAPSS data set published by NASA Fault Prediction Research Center for experimental verification. The simulation environment uses NVIDIA GeForce GTX 1650Ti, Intel Core i5-10200HCPU, 16G RAM, Windows 10, and MATLAB R2020b.

[0119] The dataset consists of multiple multivariate time series and is divided into four subsets, FD001 to FD004. Each subset is divided into a training set and a test set at a ratio of 1:1. The sample size of each subset is shown in Table 1. FD001 and FD003 contain 100 engine degradation data, FD002 and FD004 contain 260 and 249 engine degradation data, respectively. The training samples of each subset are the sensor monitoring data of the turbofan engine throughout its life cycle, and the test samples are the monitoring data of the engine from the first cycle to a cycle before failure.

[0120] Table 1 Original sample set files

[0121]

[0122] Simulation experiment:

[0123] The simulation experiment specifically includes three aspects: data preprocessing, model parameter training, and engine remaining life prediction.

[0124] First, the data of FD001 to FD004 are preprocessed. Taking the degradation data of the first engine in FD001 as an example, the original data is normalized and standardized. Figure 2 These are the data preprocessing results of the fan inlet temperature, high-pressure compressor outlet temperature, fan inlet pressure, high-pressure compressor outlet pressure, and actual core engine speed in the engine. As can be seen from the figure, the preprocessed data is limited to the range of [-3,3], eliminating the impact of inconsistent dimensions on the prediction effect.

[0125] Secondly, the preprocessed data is input into SDAE to perform unsupervised pre-training on the training set data. Since the data comes from 21 sensors, the SDAE encoding network structure is selected as five layers, and the number of neurons in each layer is selected as 17-16-15-8-1. The last layer is the final extracted feature value. Finally, the engine health factor (HI) curve is obtained, and the horizontal axis of the HI curve is the time step. Figure 3 (a) takes FD001 as an example to obtain the HI curve of 100 engines in the training set. However, there is noise in the HI curve, the curve smoothness is poor, and the discrimination is insufficient. Therefore, it is smoothed and filtered. The filtered image is shown in Figure 3 (b).

[0126] It can be seen that due to the differences in the weights of each sensor in the degradation process of different engines, the deep features extracted by SDAE show obvious monotonous increasing or decreasing characteristics throughout the life cycle of the engine, which can better characterize the degradation process of the engine. Experiments show that the other three subsets FD002~FD004 also have obvious monotonicity.

[0127] Finally, the cycle inversion number of each engine's full life cycle is selected as the training set label to train the LSTM network model parameters. According to the four operating conditions, four failure modes and sensor acquisition parameters of the data set, the weights of the sensor data under different operating conditions and failure modes on the remaining life are trained in groups, and the trained weights are introduced into the LSTM model to predict the remaining life of the engine.

[0128] The number of neurons in the hidden layer is set to 400, and the learning rate is set in stages. The initial learning rate is 0.005, and it is reduced by 1 / 5 every 30 rounds of iteration. In order to prevent overfitting caused by too many neurons and a large number of iterations, which leads to a decrease in the generalization performance of the model, a dropout layer is added with a value of 0.25.

[0129] During the training process, in order to improve the life prediction effect, the optimal solution of the loss function is found in the back-propagation process to update the weight and bias parameters, and a suitable gradient descent optimization algorithm should be selected.

[0130] The prediction errors of three gradient descent algorithms Adam, RMSProp and SGDM are compared in the four test subsets FD001 to FD004. The prediction error of SGDM algorithm is the smallest, so this paper uses SGDM for gradient calculation. The predicted RMSE and Score results are shown in Table 2.

[0131] Table 2 Comparison of different optimization methods

[0132]

[0133] Experimental results:

[0134] Three engines are randomly selected from each test subset to draw the life prediction result curve. Figure 4(a) shows the life prediction effect of three random engines in the FD001 subset, Figure 4(b) shows the life prediction effect of three random engines in the FD002 subset, Figure 4(c) shows the life prediction effect of three random engines in the FD003 subset, and Figure 4(d) shows the life prediction effect of three random engines in the FD004 subset.

[0135] The horizontal axis is the time step. The predicted value curve has a high degree of fit with the true value curve. When the engine with an early failure time is close to the failure inflection point, the curve has a small deviation, but the overall prediction accuracy remains at a high level.

[0136] In order to fully verify the prediction effect of the model on different data sets, the ratio of the prediction error of the remaining life of the engine in each data set to the actual life is defined as the predicted life deviation. The predicted life and the actual life of all engines in the four test sets are compared, and the error curves are calculated and drawn (Figure 5 (a) is the error curve of FD001, Figure 5 (b) is the error curve of FD003, Figure 5 (c) is the error curve of FD002, and Figure 5 (d) is the error curve of FD004). The horizontal axis is the engine number. As can be seen from Table 3, FD001 has the smallest deviation, which is 16.35%, and its maximum error is 55.22; FD002 has the largest deviation, which is 22.14%, and the maximum error is 67.16.

[0137] Table 3 Maximum error and deviation

[0138] FD001 FD002 FD003 FD004 Maximum error 55.22 67.16 64.79 103.98 Deviation 16.35％ 22.14％ 20.06％ 18.80％

[0139] Under the same experimental conditions, three methods, DNN, BiLSTM and single-layer LSTM, are selected to compare with the method in this paper, as shown in Table 4.

[0140] Table 4 Comparison of prediction effects between this model and other models

[0141]

[0142] The SDAE-LSTM model proposed in this paper has an average RMSE reduction of 6.6%, 25%, and 19.8% compared with DNN, BiLSTM, and single-layer LSTM respectively; and an average Score reduction of 39.1%, 92.7%, and 66.7% respectively. Except for the FD003 dataset, where the RMSE and Score are slightly higher than those of DNN, the models in this paper are the best in the other subsets.

[0143] in conclusion:

[0144] The simulation results show that the SDAE+LSTM model used in the present invention can take advantage of the feature extraction of the autoencoder to achieve deep feature extraction of data, obtain the HI curve, and establish a characteristic relationship with the RUL curve. After data encoding and dimensionality reduction, the complexity of the prediction model is greatly reduced, and the life prediction effect is significantly improved. The method proposed in the present invention has strong generalization ability and feasibility, and can effectively predict life.

[0145] The above shows and describes the basic principles, main features and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. An aircraft engine life prediction method based on improved LSTM, characterized by: Step 1: Process the original sensor data obtained by the sensor and construct training samples, which include training sets and test sets; Step 2: Based on the training samples constructed in step 1, an LSTM structure model is constructed as an engine remaining life prediction model; the specific steps of step 2 include: Step 2.1: Build an LSTM structure model; In step 2.1, a double-layer LSTM structure model is constructed as the engine remaining life prediction model. The forget gate, input gate and output gate of the double-layer LSTM structure model select the sigmoid activation function, denoted as σ, and the output range is [0,1], which represents the importance of the current input information. The closer to 1, the more important it is. When generating candidate memories, the tanh activation function is selected to adjust the value flowing through the network, and the output value is always limited to [-1,1]. The details are as follows: 1) Forget Gate f t =σ(W f [h t-1 ,x t ]+b f ) (3) 2) Input Gate i t =σ(W i [h t-1 ,x t ]+b i ) (4) 3) Output gate the t =σ(W o [h t-1 ,x t ]+b o ) (6) h t =o t ⊙tanh(C t ) (8) Among them, W and b are model weights and biases, respectively, and x t is the input sample, w f is the forget gate weight, h t-1 is the hidden layer state at time t-1, b f is the forget gate bias, w i is the input gate weight, b i is the input gate bias, h is the hidden layer state at time t, b c is the cell state bias, w c is the cell state weight, w o is the output gate weight, o t is the output gate state; Step 2.2: Input the test set in step 1 into SDAE, extract the deep features of the training set data in step 1.5 through SDAE encoding, perform unsupervised pre-training on the training set data, and construct the engine health factor HI curve through training to characterize the engine degradation trend; Step 2.3: Use the time series data encoded by SDAE in step 2.2 as the input of the LSTM structure model, use the reverse number of the engine's full life cycle as the label of the LSTM structure model, set the learning rate and the number of hidden layer neurons, and then train the LSTM structure model; Step 2.4: Calculate the loss of LSTM structural model training and update the LSTM structural model parameters; the loss of the LSTM model in step 2.4 is calculated using mean square error: in, Represents the remaining life prediction value, y i represents the true value of remaining life, and n is the sample size; The LSTM structure model parameter update formula of step 2.4 is as follows, which updates the moment weight and momentum factor: m t =β1·m t-1 +(1-β1)·g t w t+1 =w t -α·m t (10) Among them, w t is the weight at time t, g t is the gradient at time t, α is the initial learning rate, m t is the first-order momentum factor, and the empirical value of β1 is 0.9; Step 3: Input the test set in step 1 into the LSTM structure model constructed in step 2 to obtain the predicted RUL value. Use RMSE and Score evaluation indicators to evaluate the obtained predicted RUL value and evaluate the life prediction effect. Step 3.1: Input the test set data in step 1.5 into the input gate of the LSTM structure model updated in step 2.4 to obtain the predicted RUL value; Step 3.2: Use the root mean square error RMSE and Score function to evaluate the predicted RUL value of step 3.1; the root mean square error RMSE evaluation method of step 3.2 is, In the formula, Represents the remaining life prediction value, y i Indicates the true value of remaining life. The smaller the RMSE value, the better the prediction effect. If the RMSE index is similar, the Score function is used for further evaluation, as shown in formula (12). The smaller the Score, the better the prediction effect. Step 3.3: Use the predicted life deviation δ to evaluate the life prediction effect.

2. The aircraft engine life prediction method based on improved LSTM according to claim 1 is characterized in that: The specific steps of step 1 include: Step 1.1: Normalize and standardize the raw sensor data acquired by the sensor; Step 1.2: Construct a RUL tag for the raw sensor data processed in step 1.1; Step 1.3: Use a piecewise linear function to process the label data RUL in step 1.2, set RUL in the early cycle to a constant value, and start linear degradation after the cycle until it reaches 0; Step 1.4: Perform feature selection on the data processed in step 1.3; Step 1.5: Construct training samples for the data after feature selection in step 1.

4. The training samples include training set and test set.

3. The aircraft engine life prediction method based on improved LSTM according to claim 2 is characterized in that: The specific method of normalization and standardization processing in step 1.1 is: Normalization uses the Min-Max model, as shown in formula (1), which converts the normalized data into a distribution with a mean of 0 and a standard deviation of 1; Data standardization is shown in formula (2): In formula (1) and (2), x i ' ,j (t) represents the dimensionless sample, x i,j (t) represents the original sample, max(x :,j ) represents the maximum value of samples of the same dimension, min(x :,j ) represents the minimum value of the same dimension; represents the sample mean; s represents the sample standard deviation.

4. The aircraft engine life prediction method based on improved LSTM according to claim 3 is characterized in that: The method of evaluating the life prediction effect by using the predicted life deviation δ in step 3.3 is: in, Represents the remaining life prediction value, y i Indicates the true value of remaining life.

Images