Aero-engine life prediction method based on multi-source time series feature fusion

Abstract

The application discloses an aero-engine life prediction method based on multi-source time sequence feature fusion. The method comprises the following steps: acquiring an initial training data set, wherein each initial training data is the full life cycle data of each aero-engine, and the full life cycle data of each aero-engine is 14 sensor parameters related to the remaining useful life of the aero-engine at multiple time points; and constructing an F2Net network model, wherein the F2Net network model comprises a GRU, a multi-head self-attention mechanism, a feature fusion module, a first full connection layer, a linear projection layer, a ReLU activation function, a first Dropout layer, a second full connection layer, a feature fusion connection layer, a second Dropout layer, a RUL linear regression layer and a prediction result. The application solves the technical problem that, in the prior art, when the remaining life of an aero-engine is predicted, valuable information cannot be extracted from massive data, resulting in low prediction accuracy of the remaining life of the aero-engine.

Description

Technical Field

[0001] This invention relates to the field of intelligent analysis technology for aerospace big data, and more specifically, to a method for predicting the lifespan of aero-engines based on the fusion of multi-source time series features. Background Technology

[0002] The operational status of aero-engines is crucial to the flight safety of aircraft. Remaining service life prediction technology is an important part of aero-engine fault prediction and health management. It can predict potential faults under certain conditions so as to dynamically adjust maintenance plans and extend the service life of engines. How to extract sufficient valuable information from massive amounts of data and make full use of it is an important task to achieve the final prediction of the system's remaining service life.

[0003] Generally speaking, while most existing methods perform well in certain situations, their models have poor generalizability and the modeling process is extremely complex. For example, Li et al. in "Li CJ, Lee H. Gear Fatigue Crack Prognosis using Embedded Model, Gear Dynamic Model and Fracture Mechanics[J]" Mechanical Systems and Signal Processing ; 2005, 19(4): 836-846. ” combines rigid body, dynamics and fracture mechanics models to propose a method for predicting the remaining service life of gears with fatigue cracks; and Ghodrati et al. in “Ghodrati B, Kumar U, Ahmadzadeh F. Remaining Useful Life Estimation of Mining Equipment: A Case Study[C]. New Delhi: International Symposium on Mine Planning and Equipment Selection In 2012, a scale-adaptive population counting network was designed to estimate the remaining service life of mining equipment through reliability analysis and provide maintenance recommendations accordingly. Although these models employ a combination of statistical and data-driven methods, they lack versatility and face difficulties in real-time application in large-scale industrial equipment. Therefore, they are not suitable for predicting the remaining service life of aero-engines. The versatility of a model refers to its ability to maintain good performance in various application environments without significant adjustments to the model structure, parameters, or methods. However, the limitations of the aforementioned methods (over-reliance on the physical characteristics of specific equipment and high-quality historical data, resulting in insufficient versatility and computational complexity) hinder the application of this process in predicting the remaining service life of aero-engines.

[0004] Recent researchers have focused on using deep learning and other methods to identify system performance degradation trends from large amounts of sensor monitoring data, and using data-driven methods to predict remaining service life. For example, Li Hao et al., in "Li Hao, Wang Zhuo-jian, Li Zhe, et al. Prediction of Remaining Useful Life of Aero-Engine Based on StackedAutoencoder and Deep-AR[J]. Propulsion Technology, 2022, 43(11): 210645. (LI Hao, WANG Zhuo-jian, LI Zhe, et al. Prediction of Remaining Useful Life of Aero-Engine Based on StackedAutoencoder and Deep-AR[J]. Journal of Propulsion Technology , 2022, 43(4):210645.)” proposed a method using stacked autoencoders to extract features from the multivariate time series of engines, and used bidirectional LSTM to construct a DeepAR model for predicting remaining service life. Nie Lei et al. in “Nie Lei, Xu Shi-yi, ZHANG Lyu-fan, et al. Remaining Useful LifePrediction of Aeroengine Based on Multi-Head Attention[J]. Propulsion Technology, 2023, 44(8):2204040. (NIE Lei, XU Shi-yi, ZHANG Lyu-fan, et al. Remaining Useful LifePrediction of Aeroengine Based on Multi-Head Attention[J]. Journal of Propulsion Technology , 2023, 44(8): 2204040.)” A one-dimensional convolutional neural network model was established for the multi-dimensional features of engine data, and a multi-head attention mechanism was used for weighted processing. However, although the above methods improved the prediction accuracy of the remaining service life of aero engines, there is still a problem of insufficient data feature extraction. It is impossible to extract enough valuable information from these massive data. The overall performance and prediction accuracy of the model have not seen a breakthrough improvement. Therefore, they are still not suitable for the task of predicting the remaining service life of aero engines. Summary of the Invention

[0005] This invention provides a method for predicting the lifespan of aero-engines based on multi-source time series feature fusion, which at least solves the technical problem that existing technologies cannot extract sufficiently valuable information from massive amounts of data when predicting the remaining lifespan of aero-engines, resulting in low accuracy in predicting the remaining lifespan of aero-engines.

[0006] According to one aspect of the present invention, a method for predicting the lifespan of aero-engines based on multi-source time series feature fusion is provided. The method may include: acquiring an initial training dataset, wherein each initial training dataset represents the full lifespan data of each aero-engine, and the full lifespan data of each aero-engine comprises 14 sensor parameters related to the remaining lifespan of the aero-engine at multiple time points; constructing an F2Net network model, wherein the F2Net network model includes: GRU, a multi-head self-attention mechanism, a feature fusion module, a first fully connected layer, a linear projection layer, a ReLU activation function, a first Dropout layer, a second fully connected layer, a feature fusion connection layer, a second Dropout layer, a RUL linear regression layer, and prediction results; preprocessing the initial training dataset to obtain a target training dataset; inputting the target training dataset into the F2Net network model to obtain a successfully trained F2Net network; acquiring a test dataset, wherein each test dataset represents the half-lifespan data of each aero-engine; inputting the test dataset into the successfully trained F2Net network to obtain a prediction result for each test dataset, wherein the prediction result represents the remaining lifespan of each aero-engine.

[0007] Optionally, the preprocessing of the initial training dataset to obtain the target training dataset includes: normalizing each sensor parameter related to the remaining service life of the aerospace engine at each moment in each initial training dataset to obtain the target training dataset.

[0008] Optionally, the step of inputting the target training dataset into the F2Net network model to obtain a successfully trained F2Net network includes: processing the target training dataset using a sliding window to obtain multiple target training samples, each target training sample being 14 sensor parameters related to the remaining service life of the aerospace engine at a target time, wherein the size of the sliding window is 30; and training the F2Net network model sequentially using each target training sample to obtain a successfully trained F2Net network.

[0009] Optionally, the step of training the F2Net network model sequentially using each target training sample to obtain a successfully trained F2Net network includes: inputting each target training sample into a GRU to obtain the sequence features of each target training sample; inputting the sequence features of each target training sample into a multi-head self-attention mechanism to obtain a target attention score; inputting the sequence features and target attention score of each target training sample into a feature fusion module to obtain a first fused feature; inputting the first fused feature into a first fully connected layer to obtain a first target feature; sequentially inputting each target training sample into a linear projection layer, a ReLU activation function, a first Dropout layer, and a second fully connected layer to obtain a second target feature; inputting the first target feature and the second target feature into a feature fusion connection layer to obtain a second fused feature; sequentially inputting the second fused feature into a second Dropout layer and a RUL linear regression layer to obtain the prediction result of each target training sample, and iterating in a loop to obtain a successfully trained F2Net network.

[0010] Optionally, the expression for the process of inputting each target training sample into the GRU to obtain the sequence features of each target training sample is:

[0011]

[0012]

[0013]

[0014]

[0015] in, For a target training sample, consider 14 sensor parameters at time t that are related to the remaining service life of the space engine. Let be the sequence features of a target training sample at time t. Let be the candidate hidden state of a target training sample at time t. It is the Sigmoid activation function. For the XOR operation, , , The weight matrix, For the first The door is constantly being updated. For the first The door to reset time. For a target training sample at time t-1, For dot product operation, The activation function is used to concatenate the sequence features of each target training sample at each time step to obtain the sequence features of each target training sample.

[0016] Optionally, after inputting the test dataset into the successfully trained F2Net network and obtaining the prediction result for each test data, the method further includes: comparing the prediction result with the true value, and calculating the mean squared error (RMSE) and performance metric (Score) of the F2Net network model.

[0017] Optionally, the optimizer of the F2Net network model is the Adam optimizer, and the learning rate of the optimizer is 0.001.

[0018] The beneficial effects of this invention are:

[0019] This invention proposes a method for predicting the remaining service life of aero-engines based on multi-source time-series feature fusion. It introduces a feature fusion network that can fully extract and utilize effective features from massive amounts of sensor data with extremely high dimensionality and ultra-long durations to achieve accurate prediction of the remaining service life of aero-engines. Due to the novel network architecture and sequence feature learning, it can effectively capture short-term and long-term dependencies in the sequence. By combining features learned by the deep network with the original features obtained through linear mapping, it overcomes the limitations of single feature extraction methods, providing a more comprehensive and integrated feature representation for the model. This invention significantly improves the accuracy of remaining service life prediction for aero-engines, and compared to previous methods, it effectively enhances the performance of remaining service life prediction. Attached Figure Description

[0020] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0021] Figure 1 This is a flowchart of an aero-engine life prediction method based on multi-source time series feature fusion according to an embodiment of the present invention;

[0022] Figure 2 This is a framework diagram of the F2Net network model according to an embodiment of the present invention. Detailed Implementation

[0023] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0024] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and to describe a specific order or sequence. It should be understood that such terms can be used interchangeably where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0025] Example 1

[0026] According to embodiments of the present invention, a method for predicting the lifespan of an aero-engine based on multi-source time series feature fusion is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system containing at least one set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0027] Figure 1 This is a flowchart of an aero-engine life prediction method based on multi-source time series feature fusion according to an embodiment of the present invention, as shown below. Figure 1 As shown, the method may include the following steps:

[0028] Step S101: Obtain the initial training dataset, wherein each initial training dataset is the full life cycle data of each aerospace engine, and the full life cycle data of each aerospace engine consists of 14 sensor parameters related to the remaining service life of the aerospace engine at multiple times.

[0029] In the technical solution provided by step S101 of the present invention, among the 21 sensor monitoring parameters in the FD001 dataset, sensor parameters that are not related to performance degradation parameters are removed, and 14 sensor parameters that are related to the remaining service life are accurately selected to improve the training efficiency of the network and serve as the input of the network.

[0030] Step S102: Construct the F2Net network model, which includes: GRU, multi-head self-attention mechanism, feature fusion module, first fully connected layer, linear projection layer, ReLU activation function, first Dropout layer, second fully connected layer, feature fusion connection layer, second Dropout layer, RUL linear regression layer and prediction results.

[0031] In the technical solution provided by step S102 of the present invention, Figure 2 This is a framework diagram of the F2Net network model according to an embodiment of the present invention. Figure 2 As can be seen, the F2Net network model includes: Gated Recurrent Units (GRUs), multi-head self-attention mechanism, feature fusion module, first fully connected layer, linear projection layer, ReLU activation function, first Dropout layer, second fully connected layer, feature fusion connection layer, second Dropout layer, RUL linear regression layer, and prediction results.

[0032] Step S103: Preprocess the initial training dataset to obtain the target training dataset.

[0033] In the technical solution provided by step S103 of the present invention, each initial training data in the initial training dataset is normalized to obtain the target training dataset.

[0034] Step S104: Input the target training dataset into the F2Net network model to obtain the successfully trained F2Net network.

[0035] In the technical solution provided in step S104 of the present invention, the F2Net network model is trained on the target training dataset to obtain a successfully trained F2Net network.

[0036] Step S105: Obtain the test dataset, wherein each test dataset is half-life data of each aerospace engine.

[0037] In the technical solution provided by step S105 of the present invention, test set data is obtained, and each test set data is the predicted remaining service life of each aerospace engine.

[0038] Step S106: Input the test dataset into the successfully trained F2Net network to obtain the prediction result for each test data, where the prediction result is the remaining service life of each aerospace engine.

[0039] In the technical solution provided in step S106 of the present invention, the successfully trained F2Net network processes the test dataset to obtain the remaining service life of each aerospace engine in the test dataset.

[0040] The method described in this embodiment will be further described below.

[0041] As an optional embodiment, step S103, which involves preprocessing the initial training dataset to obtain the target training dataset, includes: normalizing each sensor parameter related to the remaining service life of the aerospace engine at each moment in each initial training dataset to obtain the target training dataset.

[0042] In this embodiment, the 14 selected sensor parameters are smoothed to reduce the impact of noise on model learning. Simultaneously, a normalization expression is applied:

[0043]

[0044] in, This represents the maximum value of each sensor parameter before normalization. This represents the minimum value of each sensor parameter before normalization. For each sensor parameter before normalization, To normalize each sensor parameter, we standardize each sensor parameter to the range of [0,1] to eliminate the difference in magnitude between different sensor parameters, thereby improving the stability and convergence efficiency of model training.

[0045] As an optional embodiment, step S104, which involves inputting the target training dataset into the F2Net network model to obtain a successfully trained F2Net network, includes: processing the target training dataset using a sliding window to obtain multiple target training samples, each target training sample consisting of 14 sensor parameters related to the remaining service life of the aerospace engine at a target time point, wherein the size of the sliding window is 30; and training the F2Net network model sequentially using each target training sample to obtain a successfully trained F2Net network.

[0046] In this embodiment, a sliding window technique is used to segment the target training dataset (i.e., time series data). Each window contains 30 consecutive time steps to capture the temporal patterns of the input data. The step size of the sliding window can be set according to the specific experiment to generate multiple training samples while ensuring the temporal correlation and consistency between features. The F2Net network model processes the multi-target training samples to obtain a successfully trained F2Net network.

[0047] As an optional embodiment, the step of training the F2Net network model sequentially using each target training sample to obtain a successfully trained F2Net network includes: inputting each target training sample into a GRU to obtain the sequence features of each target training sample; inputting the sequence features of each target training sample into a multi-head self-attention mechanism to obtain a target attention score; inputting the sequence features and target attention score of each target training sample into a feature fusion module to obtain a first fused feature; inputting the first fused feature into a first fully connected layer to obtain a first target feature; sequentially inputting each target training sample into a linear projection layer, a ReLU activation function, a first Dropout layer, and a second fully connected layer to obtain a second target feature; inputting the first target feature and the second target feature into a feature fusion connection layer to obtain a second fused feature; sequentially inputting the second fused feature into a second Dropout layer and a RUL linear regression layer to obtain the prediction result of each target training sample, and iterating in a loop to obtain a successfully trained F2Net network.

[0048] In this embodiment, such as Figure 2 As shown, each target training sample is input into a GRU to obtain the sequence features of each target training sample; the sequence features of each target training sample are input into a multi-head self-attention mechanism to obtain a target attention score; the sequence features and target attention scores of each target training sample are input into a feature fusion module to obtain a first fused feature; the first fused feature is input into a first fully connected layer to obtain a first target feature; each target training sample is sequentially input into a linear projection layer, a ReLU activation function, and a first Dropout layer to obtain the original features of each target training sample; the original features of each training sample are input into a second fully connected layer to obtain a second target feature; the first target feature and the second target feature are input into a feature fusion connection layer to obtain a second fused feature; the second fused feature is sequentially input into a second Dropout layer and a ReLU linear regression layer to obtain the prediction result of each target training sample. This process is iterated and looped to obtain a successfully trained F2Net network.

[0049] The sequence features of each target training sample are input into the multi-head self-attention mechanism (MHSAM). MHSAM identifies key time points that have a significant impact on RUL prediction by calculating the importance weight of each time step in the input sequence. Each head independently calculates attention weights and extracts features. The features of all heads are concatenated and then linearly transformed to obtain the target attention score.

[0050] As an optional embodiment, the expression for the process of inputting each target training sample into the GRU to obtain the sequence features of each target training sample is:

[0051]

[0052]

[0053]

[0054]

[0055] in, For a target training sample, consider 14 sensor parameters at time t that are related to the remaining service life of the space engine. Let be the sequence features of a target training sample at time t. Let be the candidate hidden state of a target training sample at time t. It is the Sigmoid activation function. For the XOR operation, , , The weight matrix, For the first The door is constantly being updated. For the first The door to reset time. For a target training sample at time t-1, For dot product operation, The activation function is used to concatenate the sequence features of each target training sample at each time step to obtain the sequence features of each target training sample.

[0056] In this embodiment, in order to effectively preserve the temporal sequence of data, a gated recurrent unit (GRU) improved from the Long Short-Term Memory (LSTM) network structure is used to capture sequence dependencies with long temporal distances. Each processed target training sample is first input into the GRU network for sequence feature learning, which can effectively capture the dependencies in the sequence.

[0057] The preprocessed sliding window data is input into the GRU. The GRU effectively captures short-term and long-term dependencies in time series data through its gating mechanism, while reducing the number of parameters in traditional RNN and LSTM networks. The process of GRU extracting time series features is as follows: First, the gates are updated... Used to control the hidden state of the previous time step. The degree of preservation;

[0058] Reset door Control the hidden state of the previous time step In the current input The degree to which something is forgotten under the influence of [the system], and then the candidate hidden state. Based on current input The hidden state is calculated based on the reset hidden state, and finally, the candidate hidden state is used. and the update gate The weighted fusion yields the current hidden state. By dynamically adjusting the information flow between time steps using GRU, key patterns in time series data can be learned.

[0059] As an optional embodiment, after inputting the test dataset into the successfully trained F2Net network and obtaining the prediction result for each test data, the method further includes: comparing the prediction result with the true value, and calculating the mean squared error (RMSE) and performance metric (Score) of the F2Net network model.

[0060] In this embodiment, the predicted results are compared with the true values, and the expressions for calculating the mean squared error (RMSE) and performance metric (Score) of the F2Net network model are as follows:

[0061]

[0062]

[0063]

[0064]

[0065] in, It's the sample size. It is the first The real test sample value, Indicates the first One test sample The difference between the predicted value and the actual value This indicates the final score. Indicates the first The score of each test sample.

[0066] As an optional implementation, the optimizer of the F2Net network model is the Adam optimizer, and the learner rate of the optimizer is 0.001.

[0067] Experimental section:

[0068] This invention was run on an NVIDIA GeForce GTX3090 and an Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz.

[0069] The dataset used in the experiment was C-MAPSS, which was developed by Frederick et al. in the paper "DK Frederick, JA DeCastro, and JS Litt, 'User's Guide for the Commercial Modular Aero-Propulsion System Simulation (C-MAPSS)'". NASA Technical Report The C-MAPSS protocol, proposed in NASA / TM-2007-215026, 2007, comprises four subsets: FD001, FD002, FD003, and FD004. The FD001 subset was used in this experiment and contains 200 sample data points. The training set includes full lifecycle data from 100 engines, while the test set includes partial operating cycle data from 100 engines.

[0070] 2. Experiment Content

[0071] First, the FD001 subset of the C-MAPSS dataset was selected as the experimental data, and its training set was used to train the feature fusion network model. The number of hidden nodes in the network was adjusted according to the given training data. Specifically, the input dimension of GRU was 17 and the number of hidden nodes was set to 50; the number of heads of MHSAM was set to 5; the output hidden nodes of the two FC layers were both set to 10; and the learning rate of the Adam optimization algorithm was set to 0.001.

[0072] Then, the trained model was used to infer the remaining service life of each engine on the test set of the FD001 subset, and the performance index of the model was calculated. In the experiment, the performance of each algorithm on RMSE and Score was measured to evaluate the predictive performance of the model. In order to avoid the randomness of the results, five experiments were conducted on the FD001 dataset and the average value was taken.

[0073] To demonstrate the effectiveness of the algorithm, experiments were conducted to compare the performance of several mainstream residual life prediction models, such as DLSTM, LSTM, and DCNN. Among them, DLSTM was mentioned in the paper "MA Qi-you, LIU Ke-wei, DU Jian, et al. Prediction of Residual Life of Engine Blades Based on Deep Short Term Memory Network [J]. Propulsion Technology, 2021, 42(8): 1888-1897. (MA Qi-you, LIU Ke-wei, DU Jian, et al. Prediction of Residual Life of Engine Blades Based on Deep Short Term Memory Network [J]. Propulsion Technology, 2021, 42(8): 1888-1897.) Journal of Propulsion TechnologyThe D-convolutional neural network is described in detail in the paper “Li X, Ding Q, Sun J Q. Remaining Useful Life Estimation in Prognostics Using DeepConvolution Neural Networks [J]” by Li X et al. Reliability Engineering&System Safety The invention was proposed in , 2018, 172: 1-11.”. This invention uses RMSE and Score to measure the predictive performance of the model. Experimental results show that the feature fusion network outperforms other comparative models in both metrics, demonstrating a significant performance advantage. The comparison results are shown in Table 1:

[0074] Table 1 Comparison of results from different algorithms

[0075]

[0076] As shown in Table 1, compared with the best-performing method on the FD001 dataset, the AG convolutional neural network framework significantly improves both RMSE and Score performance metrics, with RMSE reduced by 10.54% and Score reduced by 23.45%. This demonstrates the effectiveness of the proposed method in improving prediction accuracy. Furthermore, although the AG convolutional neural network exhibits strong feature extraction capabilities under specific conditions, the feature fusion network, by combining GRU and MHSAM, achieves more comprehensive feature fusion, significantly improving the modeling ability for complex time series data. This proves the significant advantages of this invention in terms of effectiveness and robustness.

[0077] In this embodiment of the invention, an initial training dataset is obtained, wherein each initial training dataset is the full lifecycle data of each aerospace engine, and the full lifecycle data of each aerospace engine consists of 14 sensor parameters related to the remaining service life of the aerospace engine at multiple time points; an F2Net network model is constructed, wherein the F2Net network model includes: GRU, multi-head self-attention mechanism, feature fusion module, first fully connected layer, linear projection layer, ReLU activation function, first Dropout layer, second fully connected layer, feature fusion connection layer, second Dropout layer, RUL linear regression layer, and prediction results; the initial training dataset is preprocessed to obtain the target training dataset; and then... The target training dataset is input into the F2Net network model to obtain a successfully trained F2Net network. A test dataset is obtained, where each test dataset represents the half-life data of each aerospace engine. The test dataset is input into the successfully trained F2Net network to obtain the prediction result for each test dataset, where the prediction result represents the remaining service life of each aerospace engine. Existing technologies for predicting the remaining service life of aerospace engines cannot extract sufficiently valuable information from massive amounts of data, resulting in low accuracy in predicting the remaining service life of aerospace engines. This paper achieves the technical effect of improving the accuracy of predicting the remaining service life of aerospace engines by using the F2Net feature fusion network.

[0078] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0079] In the above embodiments of the present invention, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0080] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual couplings, direct couplings, or communication connections may be through some interfaces; indirect couplings or communication connections between units or modules may be electrical or other forms.

[0081] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0082] Furthermore, the functional units in the various embodiments of the present invention can be integrated into a first processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0083] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for predicting the lifespan of aero-engines based on multi-source time series feature fusion, characterized in that, include: Obtain the initial training dataset, where each initial training dataset is the full life cycle data of each aero-engine, and the full life cycle data of each aero-engine consists of 14 sensor parameters related to the remaining service life of the aero-engine at multiple times. Construct an F2Net network model, which includes: GRU, multi-head self-attention mechanism, feature fusion module, first fully connected layer, linear projection layer, ReLU activation function, first Dropout layer, second fully connected layer, feature fusion connection layer, second Dropout layer and RUL linear regression layer; The initial training dataset is preprocessed to obtain the target training dataset; The target training dataset is input into the F2Net network model to obtain a successfully trained F2Net network model. Specifically, this involves: inputting each target training sample into a GRU to obtain the sequence features of each target training sample; inputting the sequence features of each target training sample into a multi-head self-attention mechanism to obtain a target attention score; inputting the sequence features and target attention score of each target training sample into a feature fusion module to obtain a first fused feature; inputting the first fused feature into a first fully connected layer to obtain a first target feature; inputting each target training sample sequentially into a linear projection layer, a ReLU activation function, a first Dropout layer, and a second fully connected layer to obtain a second target feature; inputting the first and second target features into a feature fusion connection layer to obtain a second fused feature; and inputting the second fused feature sequentially into a second Dropout layer and a ReLU linear regression layer to obtain the prediction result for each target training sample. This process is iterated and looped to obtain a successfully trained F2Net network model. Obtain the test dataset, where each test dataset represents half-life data for each aircraft engine; The test dataset is input into the successfully trained F2Net network model to obtain the prediction result for each test dataset, where the prediction result is the remaining service life of each aircraft engine.

2. The method according to claim 1, characterized in that, The preprocessing of the initial training dataset to obtain the target training dataset includes: For each initial training data point, each sensor parameter related to the remaining service life of the aero-engine at each time step is normalized to obtain the target training dataset.

3. The method according to claim 2, characterized in that, The step of inputting the target training dataset into the F2Net network model to obtain a successfully trained F2Net network model includes: A sliding window is used to process the target training dataset to obtain multiple target training samples. Each target training sample consists of 14 sensor parameters related to the remaining service life of the aero-engine at a specific time. The size of the sliding window is 30. The F2Net network model is trained sequentially using each target training sample to obtain a successfully trained F2Net network model.

4. The method according to claim 3, characterized in that, The expression for the process of obtaining the sequence features of each target training sample by inputting it into the GRU is as follows: in, This refers to 14 sensor parameters related to the remaining service life of an aero-engine at time t in a target training sample. Let be the sequence features of a target training sample at time t. Let be the candidate hidden state of a target training sample at time t. It is the Sigmoid activation function. For the XOR operation, , , The weight matrix, For the first The door is constantly being updated. For the first The door to reset time. For a target training sample at time t-1, For dot product operation, For activation functions; The sequence features of each target training sample at each time step are concatenated to obtain the sequence features of each target training sample.

5. The method according to claim 1, characterized in that, After inputting the test dataset into the successfully trained F2Net network model and obtaining the prediction result for each test dataset, the method further includes: The predicted results are compared with the true values, and the root mean square error (RMSE) and performance metric (Score) of the F2Net network model are calculated.

6. The method according to claim 5, characterized in that, The optimizer for the F2Net network model is the Adam optimizer, with a learning rate of 0.

001.

7. A computer system, characterized in that... include: One or more processors, a computer-readable storage medium for storing one or more programs, wherein, when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the method of claim 1.

8. A computer-readable storage medium, characterized in that... The device stores computer-executable instructions, which, when executed, are used to implement the method of claim 1.

Images