Aero-engine remaining useful life prediction method based on feature-model transfer

Abstract

The application discloses an aero-engine residual useful life prediction method based on feature-model migration. Based on the feature-model migration, in the data preprocessing stage, the application selects state monitoring parameters and operation parameters of the aero-engine as inputs of a convolutional neural network, and carries out normalization and sliding time window processing on the data, and divides source domain data and target domain data. In the convolutional neural network construction stage, the application selects a plurality of convolutional layers to sufficiently extract data features. In the feature-model migration stage, the application firstly performs feature migration by using source domain labeled data and target domain unlabeled data, and then fine tunes the feature-migrated model by using target domain labeled data, and performs model migration, so as to sufficiently utilize the source domain data and the target domain data. Results show that the model based on the feature-model migration learning can achieve good prediction effect on the test set of the target domain.

Description

Technical Field

[0001] This invention relates to the fields of transfer learning and aircraft engine remaining life prediction, and particularly to an aircraft engine remaining life prediction method based on transfer learning. Background Technology

[0002] As a crucial component of aircraft, the aero-engine is a key part of the aircraft's safe operation during its service life. Engine failure can lead to safety hazards in the aircraft system, increase the risk of unplanned grounding, and even cause significant casualties and economic losses. Therefore, predicting the remaining service life of aero-engines has been a research focus and challenge in recent years. With the continuous development of machine learning, aero-engine remaining service life prediction based on transfer learning has attracted increasing research attention. However, classic transfer learning methods have limited applicability and still have significant room for improvement in prediction accuracy. During actual engine operation, data collection methods may change, resulting in situations where both labeled and unlabeled data exist. Classical transfer learning methods struggle to fully utilize all available data. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of existing transfer learning methods for predicting the remaining service life of aero engines, and to provide a method for predicting the remaining service life of aero engines based on feature-model transfer.

[0004] The technical solution provided to solve the above technical problems is as follows:

[0005] A method for predicting the remaining service life of aero-engines based on feature-model transfer includes the following steps:

[0006] S1. Create a training dataset using historical data of aircraft engine condition monitoring parameters and operating parameters, and preprocess the training dataset to form source domain labeled data, target domain unlabeled data, and target domain labeled data.

[0007] S2. Construct a convolutional neural network consisting of multiple convolutional layers, fully connected layers, and an output layer cascaded together. The network input is the aero-engine condition monitoring parameters and operating parameters, and the network output is the remaining service life of the aero-engine.

[0008] S3. Input the source domain labeled data and the target domain unlabeled data into the convolutional neural network at the same time to perform feature transfer, and then use the target domain labeled data to perform model transfer on the feature-transferred model.

[0009] S4. Input the target domain aero-engine condition monitoring parameters and operating parameters data to be predicted into the convolutional neural network after feature transfer and model transfer. The output of the convolutional neural network is the remaining service life of the aero-engine.

[0010] Preferably, step S1 specifically includes the following steps:

[0011] S101. Using sensor-monitored variables from different parts of the aero-engine as condition monitoring parameters, and flight altitude, Mach number, and throttle lever solver angle as operating parameters, historical data sequences of aero-engine condition monitoring parameters and operating parameters are obtained.

[0012] S102. Normalize each parameter in the historical data sequence;

[0013] S103. Training samples are extracted from the normalized historical data sequence using a time window sliding method, and a training dataset is created. Each training sample contains time-series data of all status monitoring parameters and operating parameters within a time window, with or without the remaining service life label of the aircraft engine.

[0014] S104. Based on the domain source of each training sample and whether it carries the label of the remaining service life of the aircraft engine, divide the training dataset into source domain labeled data, target domain unlabeled data, and target domain labeled data.

[0015] Furthermore, the historical data sequence is derived from the C-MAPSS dataset. Different subsets in the dataset exhibit different data distributions based on different combinations of their operating modes and fault types, with each subset serving as a domain.

[0016] Furthermore, in step S102, the parameter normalization adopts the max-min normalization method.

[0017] Furthermore, in step S103, the time window is moved forward by one moment each time, and each time window forms a sample; the subset is divided into a source domain and a target domain, wherein the sample data in the source domain carries the label of the remaining service life of the aircraft engine, and only a small number of sample data in the target domain carry the label of the remaining service life of the aircraft engine, while the remaining sample data do not carry the label of the remaining service life of the aircraft engine.

[0018] Preferably, in step S2, the convolutional neural network has 5 convolutional layers, and its output layer is preferably implemented using a fully connected layer.

[0019] Preferably, step S3 specifically includes the following steps:

[0020] S301. Input the source domain labeled data and the target domain unlabeled data into the convolutional neural network. The stacked convolutional layers extract features from the input data. Then, the fully connected layer and the output layer perform domain adaptation to achieve feature transfer between the source domain and the target domain.

[0021] S302. Freeze the convolutional layers in the convolutional neural network after feature transfer, but keep the fully connected layers and output layers in a trainable state. Input the labeled data of the target domain into the convolutional neural network for retraining and achieve model transfer through fine-tuning.

[0022] Furthermore, in step S301, after the source domain labeled data and the target domain unlabeled data pass through multiple convolutional layers, domain adaptation is performed in the fully connected layer and the output layer, respectively. The domain adaptation process is as follows:

[0023] The Maximum Mean Discrepancy (MMD) method is used to measure the distribution distance between labeled data from the source domain and unlabeled data from the target domain in the fully connected layer and the output layer. The domain adaptive loss is calculated as follows:

[0024] l d =MMD(D′) S ,D′ T )+σ*MMD(D″ S ,D″ T )

[0025] Among them, l d Let D′ be the domain adaptive loss. S ,D′ T D″ represents the input data of the fully connected layer, namely the source domain labeled data and the target domain unlabeled data. S ,D″ T σ represents the input data of the source domain labeled data and the target domain unlabeled data in the output layer, respectively; σ is the weighted ratio of the two-layer domain adaptation loss, which is the hyperparameter of the model; the kernel function used in the maximum mean difference method is preferably the sum of multiple radial basis kernel functions, and more preferably the sum of 5 radial basis kernel functions;

[0026] The total loss for domain-adaptive feature transfer is then calculated as follows:

[0027] loss=l pre +αl d

[0028] In the above formula, l pre α is the mean squared error loss between the predicted remaining lifetime and the actual remaining lifetime label value of the source domain labeled data; α is the weight of the domain adaptation loss in the total loss, and is the hyperparameter of the model.

[0029] Finally, the total loss is backpropagated, and the Adam optimizer is used to optimize the model parameters of the convolutional neural network.

[0030] Furthermore, in step S302, the loss function of the convolutional neural network in the model transfer process is the mean squared error loss between the predicted remaining lifetime and the actual remaining lifetime label value of the labeled data in the target domain.

[0031] Preferably, step S4 specifically includes the following steps:

[0032] S401. For the state monitoring parameters and operating parameters of the aero-engine to be predicted belonging to the target domain at a specified time, construct them as the input data of the network according to the same preprocessing method as the training samples.

[0033] S402. Input the input data obtained in S401 into the convolutional neural network after model transfer. The final output of the convolutional neural network is the predicted value of the remaining service life of the aero-engine at a specified time.

[0034] Compared with the prior art, the present invention has the following advantages:

[0035] This invention is based on feature-model transfer. In the data preprocessing stage, the invention selects the state monitoring parameters and operating parameters of an aero-engine as inputs to a convolutional neural network, and performs data normalization and sliding time window processing to separate source domain data and target domain data. In the convolutional neural network construction stage, this invention uses multiple convolutional layers to fully extract data features. In the feature-model transfer stage, this invention first performs feature transfer using labeled data from the source domain and unlabeled data from the target domain, and then fine-tunes the feature-transferred model using labeled data from the target domain to perform model transfer, thereby making full use of both source and target domain data. The results show that the model based on feature-model transfer learning in this invention can achieve good prediction results on the test set of the target domain. Attached Figure Description

[0036] Figure 1 This is a schematic diagram of a feature-model transfer-based method for predicting the remaining service life of aero-engines.

[0037] Figure 2 The results are the test results of a feature-model transfer-based method for predicting the remaining service life of aero-engines. Detailed Implementation

[0038] To better understand this application, various aspects of this application will be described in more detail with reference to the accompanying drawings. It should be understood that these detailed descriptions are merely descriptions of exemplary embodiments of this application and are not intended to limit the scope of this application in any way.

[0039] The following specific embodiment uses the C-MAPSS simulation dataset released by NASA to illustrate the specific operation steps and verify the effectiveness of the proposed method.

[0040] The C-MAPSS simulation dataset comprises four subsets, each with varying operating conditions and fault types. Subsets 1 and 3 have only one operating mode, while subsets 2 and 4 have six. Subsets 1 and 2 have one fault type, while subsets 3 and 4 have two. These operating modes and fault types are key reasons for the differences in data distribution across the four subsets. Each subset contains 21 condition monitoring parameters and 3 operating variables. The condition monitoring parameters include important monitoring variables acquired by sensors at various engine locations, such as temperature, pressure, and engine speed, as detailed in Table 1. The operating parameters are flight altitude, Mach number, and throttle lever solver angle; the levels of these three operating parameters determine the engine's operating mode.

[0041] Table 1. Variables monitored by 21 sensors at different parts of the engine

[0042]

[0043] This invention discloses a method for predicting the remaining service life of aero-engines based on feature-model transfer, the method comprising the following steps:

[0044] S1. Using historical data of aero-engine condition monitoring parameters and operating parameters, create a training dataset and preprocess the training dataset. In this embodiment, step S1 uses the C-MAPSS simulation dataset released by NASA to create the training dataset, specifically including the following sub-steps:

[0045] S101. Select the input parameters for the network model from the C-MAPSS simulation dataset. The input parameters are divided into two categories: aero-engine condition monitoring parameters and operating parameters. The condition monitoring parameters consist of 21 sensor-monitored variables, including temperature, pressure, and speed variables of various engine parts as shown in Table 1 above. The operating parameters are the three aforementioned operating parameters: flight altitude, Mach number, and throttle lever solver angle. Historical data for each condition monitoring parameter and operating parameter are recorded in the C-MAPSS simulation dataset as time-series data sequences.

[0046] S102. Normalize the historical data sequence of each input parameter selected in S101 using the min-max normalization method. The normalization formula used in the min-max normalization method is as follows:

[0047]

[0048] in, This represents the normalized value of the i-th parameter at time k. This represents the true value of the i-th parameter at time k. This represents the minimum value of the i-th parameter. This represents the maximum value of the i-th parameter.

[0049] S103. Training samples are extracted from the normalized historical data sequence using a time window sliding method, and a training dataset is created. Each training sample needs to meet the input requirements of a convolutional neural network. In this embodiment, each training sample contains time-series data of all state monitoring parameters and operating parameters within a time window, with or without an engine remaining service life (RUL) label. The sliding step size of the time window can be set to one sampling time, meaning that each time the time window moves forward by one time, a sample is formed. The final training set is {W}. t ,R t}, where W t For the input parameters at a certain moment, R t This corresponds to the RUL label. In this embodiment, the time window length is 30. It should be noted that when a sample is used as the target domain, an RUL label may not be set.

[0050] S104. Based on the domain source of each training sample and whether it carries the RUL label of the aircraft engine, the training dataset is divided into source domain data and target domain data. The target domain data is further divided according to whether it carries a label, and finally three types of data are formed: source domain labeled data, target domain unlabeled data, and target domain labeled data.

[0051] In this embodiment, since different subsets of the C-MAPSS dataset exhibit different data distributions based on different combinations of operating modes and fault types, each subset can be considered a domain, thus dividing the C-MAPSS dataset into a source domain and a target domain. The source domain data is labeled with RUL tags, while the target domain contains a large amount of unlabeled data and only a small amount labeled data. In the migration experiments between the four subsets, due to the differences in data distribution, a total of 12 migration experiments can be conducted. For example, migration from subset 1 to subset 2 constitutes one experiment, and migration from subset 1 to subset 3 constitutes another. In this embodiment, the labeled data in the target domain of each migration experiment consists of 30% of the total data volume of the randomly divided target domain training set.

[0052] S2. Construct a convolutional neural network consisting of multiple convolutional layers, fully connected layers, and an output layer cascaded together. The multiple convolutional layers are introduced to improve the feature extraction capability from the data. The network input consists of aircraft engine condition monitoring parameters and operating parameters, and the network output is the remaining service life of the aircraft engine.

[0053] In this embodiment, step S2 specifically includes the following sub-steps:

[0054] S201. Use stacked multi-layer convolutional layers to extract features from the input data, which are the aero-engine condition monitoring parameters and operating parameters after sliding time window processing.

[0055] S202, the convolutional layer is followed by a fully connected layer and an output layer, the output of which is the remaining service life of the corresponding aero-engine at the corresponding moment.

[0056] In this embodiment, the structure of the convolutional neural network is as follows: Figure 1 As shown, there are 5 stacked convolutional layers, and the output layer can also be implemented using a fully connected layer.

[0057] S3. Input the source domain labeled data and the target domain unlabeled data simultaneously into the convolutional neural network to perform feature transfer. Then, use the target domain labeled data to perform model transfer on the feature-transferred model. In this embodiment, step S3 specifically includes the following sub-steps:

[0058] S301. Input the source domain labeled data and the target domain unlabeled data into the convolutional neural network. The stacked convolutional layers extract features from the input data. Then, the fully connected layer and the output layer perform domain adaptation to achieve feature transfer between the source domain and the target domain.

[0059] In this embodiment, the source domain data in each migration experiment is labeled, while the target domain data consists of a large amount of unlabeled data and a small amount of labeled data. First, domain adaptation is performed using the labeled source domain data and the unlabeled target domain data. The domain adaptation process is as follows:

[0060] S3011. The Maximum Mean Discrepancy (MMD) method is used to measure the distribution distance between labeled data from the source domain and unlabeled data from the target domain in the fully connected layer and the output layer. The MMD method is an existing technique. Its basic principle is to map two samples with different distributions onto a Reproducing Kernel Hilbert Space (RKHS), where each distribution corresponds to a point on the RKHS. The difference between the two distributions can be represented by the inner product of their corresponding points. Assuming there are datasets x and y in the sample space, with distributions p and q respectively, their MMD distance is as follows:

[0061]

[0062] Where sup represents finding the upper bound (i.e., finding the maximum value), E x～p Find the expectation, where f represents the mapping function, ||f|| H≤1 This indicates that the norm of the function f in the regenerated Hilbert space cannot exceed 1.

[0063] Assuming datasets x and y contain n and m samples respectively, the MMD distance can be expressed as:

[0064]

[0065] To calculate the spatial distance between two distributions, a kernel function is typically introduced when calculating MMD, and the kernel function is set to... The radial basis function is generally used to calculate the distance of a high-dimensional spatial distribution, as shown in the following formula:

[0066]

[0067] Therefore, the MMD method used to measure the final distribution distance in this embodiment can be further expressed as:

[0068]

[0069] As can be seen from the above formula, different kernel functions lead to different computational results, therefore, the choice of kernel function is crucial. To reduce the impact of a single kernel function on the final result, this embodiment chooses to sum multiple radial basis function kernels as the kernel function in the MMD method, that is:

[0070]

[0071] Experience shows that the best results are achieved when the five kernel functions are added together. Therefore, this embodiment uses the sum of the five radial basis functions as the kernel function in the MMD method.

[0072] S3012. Based on the above MMD method, the formula for calculating the domain adaptive loss is:

[0073] l d =MMD(D′) S ,D′ T )+σ*MMD(D″ S ,D″ T )

[0074] Among them, l d Let D′ be the domain-adaptive loss sum, and MMD(,) represent the MMD method used to ultimately measure the distribution distance. S ,D′ TD″ represents the input data of the fully connected layer, namely the source domain labeled data and the target domain unlabeled data. S ,D″ T σ represents the input data of the source domain labeled data and the target domain unlabeled data in the output layer, respectively; σ is the weighted ratio of the two-layer domain adaptation loss, which is a hyperparameter of the model and is set to 1 in this embodiment.

[0075] The total loss for domain-adaptive feature transfer is then calculated as follows:

[0076] loss = l pre +αl d

[0077] In the above formula, l pre α is the mean squared error loss between the predicted RUL and the true RUL label values ​​of the source domain labeled data; α is the weight of the domain adaptation loss in the total loss and is a hyperparameter of the model, which is set to 0.01 in this embodiment.

[0078] Finally, the total loss is backpropagated, and the Adam optimizer is used to optimize all learnable model parameters of the convolutional neural network, achieving feature transfer learning between the source and target domains.

[0079] S302. Input the target domain labeled data into the convolutional neural network after feature transfer to perform model transfer. In this embodiment, after the convolutional neural network is trained based on domain-adaptive feature transfer, the convolutional layers need to be frozen, but the fully connected layers and output layer remain trainable. Input the target domain labeled data into the network, retrain the convolutional neural network, and achieve model transfer by fine-tuning the fully connected layers and output layer. The loss function in the model transfer process is the mean squared error loss between the predicted RUL and the true RUL label values ​​of the target domain labeled data.

[0080] S4. Input the target domain aero-engine condition monitoring parameters and operating parameters data to be predicted into the convolutional neural network after feature transfer and model transfer. The output of the convolutional neural network is the remaining service life of the aero-engine. The specific steps are as follows:

[0081] S401. For the state monitoring parameters and operating parameters of the aero-engine to be predicted belonging to the target domain at a specified time, construct them as the input data of the network according to the same preprocessing method as the training samples.

[0082] S402. Input the input data obtained in S401 into the convolutional neural network after model transfer. The final output of the convolutional neural network is the predicted value of the remaining service life of the aero-engine at a specified time.

[0083] In this embodiment, following the general practice of predictive network performance testing, a portion of samples is pre-extracted from the target domain data to form the target domain test set data, in order to verify the specific prediction effect of the present invention (of course, in practical applications, the state monitoring parameters and operating parameters of the aero-engine to be predicted at the target time belonging to the target domain can be selected to replace the test set data). Therefore, in the above step S4, the target domain test set data is directly input into the convolutional neural network, and the output of the convolutional neural network is the remaining service life of the aero-engine in the test set, specifically including the following sub-steps:

[0084] 1) Preprocess the test set data of the target domain, including parameter selection, data normalization, and sliding time window processing. The data preprocessing method is the same as the training set data preprocessing method in S2 above, to form test samples.

[0085] 2) Input the state monitoring parameters and operation parameters from the test samples into the convolutional neural network after model transfer. The output of the convolutional neural network is the predicted RUL value of the engine in the target domain test set. The prediction effect can be evaluated based on the predicted RUL value and the actual RUL label in the test samples.

[0086] In the C-MAPSS dataset of this embodiment, according to step S4, in the transfer experiment from subset 4 to subset 3, the feature transfer model is trained using the labeled data of subset 4 and the unlabeled data of subset 3. Then, the model after feature transfer is fine-tuned (i.e., model transfer) using the labeled data of subset 3. The fine-tuned model has a root mean square error of 21.5 on the test set of subset 3. The prediction results are shown in the attached figure. Figure 2 As shown, it can be observed that the model based on feature-model transfer learning performs well in predicting the target domain on the test set.

[0087] In summary, this invention, based on feature-model transfer, utilizes the NASA-released C-MAPSS dataset for modeling experiments. In the data preprocessing stage, this invention selects the condition monitoring parameters and operating parameters of an aero-engine as inputs to the convolutional neural network, and performs data normalization and sliding time window processing to separate source and target domain data. In the convolutional neural network construction stage, this invention employs multiple convolutional layers to fully extract data features. In the feature-model transfer stage, this invention first performs feature transfer using labeled source domain data and unlabeled target domain data, and then fine-tunes the feature-transferred model using labeled target domain data to achieve model transfer, thereby fully utilizing both source and target domain data.

[0088] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the invention. Therefore, all technical solutions obtained through equivalent substitution or transformation fall within the protection scope of the present invention.

Claims

1. A method for predicting the remaining service life of an aero-engine based on feature-model transfer, characterized in that, Includes the following steps: S1. Create a training dataset using historical data of aircraft engine condition monitoring parameters and operating parameters, and preprocess the training dataset to form source domain labeled data, target domain unlabeled data, and target domain labeled data. S2. Construct a convolutional neural network consisting of multiple convolutional layers, fully connected layers, and an output layer cascaded together. The network input is the aero-engine condition monitoring parameters and operating parameters, and the network output is the remaining service life of the aero-engine. S3. Input the source domain labeled data and the target domain unlabeled data into the convolutional neural network at the same time to perform feature transfer, and then use the target domain labeled data to perform model transfer on the feature-transferred model. S4. Input the target domain aero-engine condition monitoring parameters and operation parameters data to be predicted into the convolutional neural network after feature transfer and model transfer. The output of the convolutional neural network is the remaining service life of the aero-engine. Step S3 specifically includes the following steps: S301. Input the source domain labeled data and the target domain unlabeled data into the convolutional neural network. The stacked convolutional layers extract features from the input data. Then, the fully connected layer and the output layer perform domain adaptation to achieve feature transfer between the source domain and the target domain. S302. Freeze the convolutional layers in the convolutional neural network after feature transfer, but keep the fully connected layers and the output layer in a trainable state. Input the target domain labeled data into the convolutional neural network for retraining and achieve model transfer through fine-tuning. In step S301, after the source domain labeled data and the target domain unlabeled data pass through multiple convolutional layers, domain adaptation is performed in the fully connected layer and the output layer, respectively. The domain adaptation process is as follows: The Maximum Mean Discrepancy (MMD) method is used to measure the distribution distance between labeled data from the source domain and unlabeled data from the target domain in the fully connected layer and the output layer. The domain adaptive loss is calculated as follows: l d =MMD(D′ S ,D′ T )+σ*MMD(D″ S ,D″ T ) Among them, l d Let D′ be the domain adaptive loss. S ,D′ T D″ represents the input data of the fully connected layer, namely the source domain labeled data and the target domain unlabeled data. S ,D″ T σ represents the input data of the source domain labeled data and the target domain unlabeled data in the output layer, respectively; σ is the weighted ratio of the two-layer domain adaptation loss, and σ is the hyperparameter of the model; the kernel function used in the maximum mean difference method is the sum of multiple radial basis function kernels; The total loss for domain-adaptive feature transfer is then calculated as follows: loss=l pre +αl d In the above formula, l pre α is the mean squared error loss between the predicted remaining lifetime and the actual remaining lifetime label value of the source domain labeled data; α is the weight of the domain adaptation loss in the total loss, and α is also used as a hyperparameter of the model. Finally, the total loss is backpropagated, and the Adam optimizer is used to optimize the model parameters of the convolutional neural network.

2. The method for predicting the remaining service life of aero-engines based on feature-model transfer as described in claim 1, characterized in that, Step S1 specifically includes the following steps: S101. Using sensor-monitored variables from different parts of the aero-engine as condition monitoring parameters, and flight altitude, Mach number, and throttle lever solver angle as operating parameters, historical data sequences of aero-engine condition monitoring parameters and operating parameters are obtained. S102. Normalize each parameter in the historical data sequence; S103. Training samples are extracted from the normalized historical data sequence using a time window sliding method, and a training dataset is created. Each training sample contains time-series data of all status monitoring parameters and operating parameters within a time window, with or without the remaining service life label of the aircraft engine. S104. Based on the domain source of each training sample and whether it carries the label of the remaining service life of the aircraft engine, divide the training dataset into source domain labeled data, target domain unlabeled data, and target domain labeled data.

3. The method for predicting the remaining service life of aero-engines based on feature-model transfer as described in claim 2, characterized in that, The historical data sequence is derived from the C-MAPSS dataset. Different subsets of the dataset exhibit different data distributions based on different combinations of their operating modes and fault types, with each subset serving as a domain.

4. The method for predicting the remaining service life of aero-engines based on feature-model transfer as described in claim 2, characterized in that, In step S102, the parameter normalization adopts the max-min normalization method.

5. The method for predicting the remaining service life of aero-engines based on feature-model transfer as described in claim 2, characterized in that, In step S103, the time window is moved forward by one moment each time, and each time window forms a sample; the subset is divided into a source domain and a target domain, wherein the sample data in the source domain carries the remaining service life label of the aircraft engine, and only a small number of sample data in the target domain carry the remaining service life label of the aircraft engine, while the remaining sample data do not carry the remaining service life label of the aircraft engine.

6. The method for predicting the remaining service life of aero-engines based on feature-model transfer as described in claim 1, characterized in that, In step S2, the convolutional neural network has 5 convolutional layers, and its output layer is implemented using a fully connected layer.

7. The method for predicting the remaining service life of aero-engines based on feature-model transfer as described in claim 1, characterized in that, The kernel function used in the maximum mean difference method is the sum of five radial basis function kernels.

8. The method for predicting the remaining service life of aero-engines based on feature-model transfer as described in claim 1, characterized in that, In step S302, the loss function of the convolutional neural network in the model transfer process is the mean squared error loss between the predicted remaining lifetime and the actual remaining lifetime label value of the labeled data in the target domain.

9. The method for predicting the remaining service life of aero-engines based on feature-model transfer as described in claim 1, characterized in that, Step S4 specifically includes the following steps: S401. For the state monitoring parameters and operating parameters of the aero-engine to be predicted belonging to the target domain at a specified time, construct them as the input data of the network according to the same preprocessing method as the training samples. S402. Input the input data obtained in S401 into the convolutional neural network after model transfer. The final output of the convolutional neural network is the predicted value of the remaining service life of the aero-engine at a specified time.

Images