Aero-engine residual life prediction method and system based on multi-task semi-supervised learning

Abstract

The invention belongs to the technical field of equipment residual service life prediction, and discloses an aero-engine residual service life prediction method and system based on multi-task semi-supervised learning, and the method comprises the steps: firstly constructing a three-dimensional evaluation system through employing monotonicity, tendency and Spearman correlation coefficients, and effectively screening out key operation parameters with a strong degradation characterization capability; unsupervised pre-training is carried out by using label-free data of the whole life cycle, and the model is helped to better understand data distribution; a degradation stage is divided by adopting an H-Kmeans method fused with mode voting, three tasks of degradation stage identification, key operation parameter prediction and residual service life prediction are combined to construct a multi-task combined prediction model, and the prediction precision of the residual service life of the aero-engine is improved by fully exerting the advantages of multi-task learning and semi-supervised learning.

Description

Technical Field

[0001] This invention belongs to the field of equipment remaining service life prediction technology, and in particular relates to a method and system for predicting the remaining service life of aero-engines based on multi-task semi-supervised learning. Background Technology

[0002] In the aviation field, the health of an engine is directly related to flight safety and economic efficiency. The operation of an aircraft engine relies on adjusting its rotational speed and the thrust it generates. Its internal structure consists of several key components, including the air intake, fan, bypass duct, high-pressure and low-pressure compressors, combustion chamber, high-pressure and low-pressure turbines, and exhaust pipe. The parameters related to speed, pressure, temperature, and thrust involved in the operation of these components are crucial for predicting the engine's range of performance (RUL).

[0003] The operational status of aero-engines is typically influenced by a variety of factors, which exhibit certain regularities over time. CN202411479629.9 discloses a GRU-CNN-based method for predicting the remaining service life of aero-engines; however, this method does not fully consider the impact of key aero-engine operating parameters on the prediction of remaining service life, resulting in limited model accuracy.

[0004] In research on predicting the remaining service life of aero-engines, multi-task learning can simultaneously consider multiple pieces of information related to the remaining service life of the aero-engine, such as the prediction of key operating parameters and the identification of degradation stages. Meanwhile, semi-supervised learning can utilize a large amount of unlabeled data combined with a small amount of labeled data to improve the model's generalization ability to unknown situations. By combining multi-task learning and semi-supervised learning, the potential information in the data can be more effectively mined, improving the accuracy and robustness of predictions.

[0005] Based on the above analysis, the problems and shortcomings of the existing technology are as follows:

[0006] A GRU-CNN-based method for predicting the remaining service life of aero-engines has been developed, but this method does not fully consider the impact of key operating parameters of aero-engines on the prediction of remaining service life, resulting in limited model accuracy. Summary of the Invention

[0007] To address the problems existing in the prior art, this invention provides a method for predicting the remaining life of aero-engines based on multi-task semi-supervised learning.

[0008] This invention is implemented as follows: A method for predicting the remaining life of an aero-engine based on multi-task semi-supervised learning includes:

[0009] Step 1, Data Acquisition and Preprocessing

[0010] Collect the operational status monitoring signals of the aero-engine throughout its entire life cycle, from its initial health state to the end of its life, and perform data preprocessing to obtain the aero-engine operational status monitoring signals with complete timestamps.

[0011] Step 2, Key Operating Parameter Screening

[0012] By using monotonicity and trend indicators and their Spearman correlation coefficient with the remaining service life, the preprocessed aero-engine operating status monitoring signals were screened to obtain key operating parameters with strong degradation characterization capabilities.

[0013] Step 3, Unsupervised pre-training

[0014] An auto-encoder (AE) was used to perform unsupervised pre-training on the key operating parameter sequences of the entire life cycle of an aero-engine. The network training was completed by minimizing the reconstruction error, which helped the model better understand the intrinsic distribution of the data and learn an unsupervised feature extractor, thus promoting the learning of subsequent tasks.

[0015] Step 4, Degradation Stage Division

[0016] The H-Kmeans (Hierarchical Clustering and K-means) clustering method, which integrates the majority vote, is used to classify the engine degradation stages. This provides another auxiliary task for the prediction of the remaining useful life (RUL) of aero-engines in subsequent multi-task joint training, namely, the accurate labels required for the degradation stage identification task.

[0017] Step 5, Multi-task Joint Training

[0018] A multi-task joint training model is constructed, comprising three tasks: degradation stage identification, prediction of future trends of key operating parameters, and RUL prediction. The encoding layer of the AE after unsupervised pre-training in step 3 is used as the feature sharing layer for multi-task learning. A bidirectional long short-term memory network (Bi-LSTM) is used for RUL prediction and key operating parameter prediction, while a convolutional neural network (CNN) is used for degradation stage identification. Through the internal parameter sharing mechanism in multi-task learning, the accuracy of prediction of the remaining service life of aero-engines is improved.

[0019] Step 6, Model Application Testing

[0020] The aero-engine operating parameters of the test object, after being filtered in step 2, are input into the trained multi-task prediction model to obtain the predicted values ​​of remaining service life, degradation stage, and future trends of key operating parameters. The average value of the multi-step RUL prediction values ​​output by the statistical model at the prediction time is used as the final RUL prediction value. The mean square error is used as the indicator to evaluate the remaining service life prediction performance of the model.

[0021] Furthermore, in step 1, the original operating status monitoring signal data of the aero-engine is first read, and the multi-sensor signals are grouped and extracted according to the engine number. Then, the data is converted into a format suitable for analysis, and the sliding window method is used to divide it into fixed-length sequence samples. At the same time, a corresponding label (such as remaining life) is generated for each sample, missing values ​​are processed, and the data is normalized.

[0022] Furthermore, in step 2, a three-dimensional evaluation system is constructed by using the monotonicity index Mon, the trend index Tred, and selecting the Spearman correlation coefficient between the higher comprehensive score and RUL, which effectively screens out key operating parameters with strong degradation characterization capabilities.

[0023] Furthermore, in step 3, an autoencoder network is used to perform unsupervised pre-training on the full lifecycle sequence of key operating parameters selected in step 2. The intrinsic structure of the data is learned by minimizing the reconstruction error, and the obtained network parameters are used as initial parameters. Unlabeled data is used to provide a good initialization for the network, and labeled training can further optimize the network parameters and improve the performance of the model on specific tasks.

[0024] Furthermore, in step 4, the H-Kmeans clustering method is used to cluster the engine degradation stages, and the cluster labels are optimized by adding mode voting post-processing to achieve accurate division of degradation stages.

[0025] Furthermore, in step 5, the historical key operating parameters of the aero-engine are used as input, and the target output includes three items: the predicted future trend value of the key operating parameters, the predicted value of the remaining service life, and the degradation stage classification result. In time series modeling, in order to match the model input dimension, the original time series needs to be windowed, that is, it is divided into multiple time series segments according to a set window length, and then a model training and testing sample dataset is constructed based on the key operating parameters of the aero-engine. This dataset corresponds to three types of label sets: one is the key operating parameter prediction label, the second is the remaining service life prediction label, and the third is the degradation stage classification label.

[0026] In step 6, the average value of the multi-step RUL predictions output by the statistical model at the prediction time is used as the final RUL prediction value.

[0027] Another objective of this invention is to provide an aero-engine remaining life prediction system based on multi-task semi-supervised learning, comprising:

[0028] The data acquisition and preprocessing unit is used to acquire, clean, and convert raw aero-engine operating status monitoring data;

[0029] The key operating parameter screening unit, based on the monotonicity and trend characteristics of the operating parameters and their Spearman correlation coefficient with RUL, screens out key operating parameters with strong degradation characterization capabilities, providing high-quality data input for subsequent modeling;

[0030] The unsupervised pre-training unit utilizes an autoencoder (AE) to pre-train unlabeled data from the entire lifecycle of an aero-engine.

[0031] The degradation stage segmentation unit utilizes clustering methods to segment the engine degradation stages, providing the necessary accurate labels for the degradation stage identification task;

[0032] A multi-task joint training unit is used to construct a multi-task model for predicting the RUL of aero engines. The encoding layer of the pre-trained AE is used as a feature sharing layer, Bi-LSTM is used as an independent task layer for predicting the trend of key operating parameters and RUL, and CNN is used as an independent task layer for identifying the degradation stage.

[0033] The model application testing unit is used to test the trained model and calculate the performance metrics of all test results.

[0034] Another object of the present invention is to provide a computer device including a memory and a processor, the memory storing a computer program, which, when executed by the processor, causes the processor to perform the steps of the aircraft engine remaining life prediction method based on multi-task semi-supervised learning.

[0035] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method for predicting the remaining life of an aero-engine based on multi-task semi-supervised learning.

[0036] Another objective of this invention is to provide an information data processing terminal for implementing the aero-engine remaining life prediction system based on multi-task semi-supervised learning.

[0037] Based on the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solution to be protected by this invention are as follows:

[0038] This method addresses the shortcomings of existing aero-engine remaining life prediction methods, which rely on single-supervised models and lack sufficient understanding of degradation stages. By introducing a semi-supervised and multi-task collaborative mechanism, it achieves unified modeling of the entire engine life degradation process, significantly improving the stability and generalization ability of life prediction. At the data processing level, it solves the problems of inconsistent time series and severe noise interference from multi-source sensor data. Through standardized preprocessing and time windowing, it improves the accuracy of subsequent models in perceiving changes in operating status. Regarding parameter selection, it addresses the issues of traditional feature selection relying on experience and insufficient degradation representation capabilities. By combining monotonicity, trend, and life correlation constraints, it makes the selected parameters more physically meaningful in relation to degradation, enhancing the effectiveness of model input. In the feature learning stage, it solves the problem of insufficient model training due to insufficient labeled samples. By utilizing unsupervised pre-training to learn the inherent structure of operating data in advance, it effectively reduces the dependence on the scale of labeled data. In terms of degradation stage modeling, it addresses the problem of ambiguous degradation boundaries by obtaining stable stage labels through clustering and voting fusion, enabling the model to explicitly distinguish different degradation evolution states.

[0039] In predictive modeling, the challenge of balancing lifetime accuracy and degradation awareness in a single task is addressed. Multi-task joint optimization enables shared features to simultaneously serve lifetime prediction, trend prediction, and stage identification. In time-series modeling, the correlation between historical and future states is fully depicted, improving the lifetime prediction's responsiveness to early, subtle degradation. At the system level, the difficulty of engineering deployment of the algorithm is resolved through modular unit partitioning, achieving reusability and scalability of the prediction process. At the prediction output level, statistical fusion of multi-time-stamp prediction results reduces the impact of instantaneous prediction fluctuations on lifetime assessment.

[0040] This technology achieves collaborative optimization from data, features, model to system, and has made significant progress in prediction accuracy, stability and engineering applicability. (1) It considers that unlabeled operational data in aero-engine operation data is usually much more than labeled fault data. Semi-supervised learning can effectively utilize these unlabeled data to help the model better understand the inherent distribution of data, enhance the model's prediction ability, reduce the need for a large amount of labeled data, and reduce the cost of data preparation. (2) It considers the positive role of key operational parameter trend prediction and degradation stage identification tasks in predicting the remaining service life of aero-engines. Through the shared representation layer of multi-task learning, the model can learn general features across tasks, while the task-specific layer can capture specific information for remaining service life prediction, improving the accuracy of remaining service life prediction. At the same time, multi-task joint training can reduce the model's overfitting to specific datasets and improve the model's generalization ability on unknown data.

[0041] (1) The expected benefits and commercial value of the technical solution of this invention after transformation are significant, mainly reflected in the in-depth mining of data asset value to achieve cost reduction and efficiency improvement in operation and maintenance throughout the entire life cycle. On the one hand, in response to the high maintenance costs of aero engines, this invention constructs a high-precision condition-based maintenance strategy through a three-dimensional evaluation system and multi-task joint prediction technology, effectively avoiding the risk of unplanned downtime caused by prediction lag and the waste of excessive maintenance caused by prediction too early, helping enterprises reduce potential economic losses and improve overall operational efficiency; on the other hand, this invention innovatively introduces an unsupervised pre-training mechanism, transforming the massive amount of unlabeled flight data, which is regarded as sunk cost in the traditional model, into high-value assets, overcoming the over-reliance on expensive and scarce full life cycle fault label data, significantly reducing the data collection threshold for model development and deployment, and significantly improving the market competitiveness and commercial transformation prospects of this technical solution.

[0042] (2) This invention proposes a method for predicting the remaining service life of aero-engines based on semi-supervised and multi-task learning. Currently, the mainstream technologies in the industry at home and abroad mainly rely on a large amount of labeled data throughout the entire life cycle for fully supervised training. However, in actual engineering, it is extremely difficult to obtain complete labeled data of an engine from operation to failure. This invention innovatively introduces an unsupervised pre-training mechanism, filling the gap in how to improve the generalization ability of the model using massive process data with very few labeled samples. In addition, this invention constructs a three-in-one multi-task model that integrates H-Kmeans stage division by mode voting, prediction of key operating parameters, and prediction of remaining service life, filling the technical gap in the industry of lacking explicit guidance on service life prediction using degradation stage information, and forming a complete, data-driven closed-loop solution for engine health management. Attached Figure Description

[0043] Figure 1 This is a flowchart of the aero-engine remaining life prediction method based on multi-task semi-supervised learning provided in an embodiment of the present invention.

[0044] Figure 2 This is a block diagram of the aero-engine remaining life prediction system based on multi-task semi-supervised learning provided in an embodiment of the present invention.

[0045] Figure 3 This is a flowchart of the aero-engine life prediction technology based on multi-task semi-supervised learning provided in an embodiment of the present invention.

[0046] Figure 4 This is a comparison chart of the monotonicity, trend, and comprehensive score of key operating parameters provided in the embodiments of the present invention.

[0047] Figure 5 These are data visualization diagrams of three sets of key operating parameters (5, 11, and 12) provided in the embodiments of the present invention.

[0048] Figure 6 This is a graph showing the H-Kmeans mode voting results provided in an embodiment of the present invention.

[0049] Figure 7 These are the contour coefficients after the mode vote for different clustering methods provided in the embodiments of the present invention.

[0050] Figure 8 This is a diagram of the multi-task joint training model architecture provided in an embodiment of the present invention.

[0051] Figure 9 This is a graph showing the multi-task joint training prediction results provided in an embodiment of the present invention. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0053] like Figure 1 As shown in the figure, the method for predicting the remaining life of an aero-engine based on multi-task semi-supervised learning provided by this invention includes the following steps:

[0054] S101, Data Acquisition and Preprocessing;

[0055] Collect the operational status monitoring signals of the aero-engine throughout its entire life cycle, from its initial health state to the end of its life, and perform data preprocessing to obtain the aero-engine operational status monitoring signals with complete timestamps.

[0056] S102, Key Operating Parameter Screening;

[0057] By using monotonicity and trend indicators and their Spearman correlation coefficient with the remaining service life, the preprocessed aero-engine operating status monitoring signals were screened to obtain key operating parameters with strong degradation characterization capabilities.

[0058] S103, unsupervised pre-training;

[0059] An auto-encoder (AE) was used to perform unsupervised pre-training on the key operating parameter sequences of the entire life cycle of an aero-engine. The network training was completed by minimizing the reconstruction error, which helped the model better understand the intrinsic distribution of the data and learn an unsupervised feature extractor, thus promoting the learning of subsequent tasks.

[0060] S104, Degradation Stage Classification;

[0061] The H-Kmeans (Hierarchical Clustering and K-means) clustering method, which integrates the majority vote, is used to classify the engine degradation stages. This provides another auxiliary task for the prediction of the remaining useful life (RUL) of aero-engines in subsequent multi-task joint training, namely, the accurate labels required for the degradation stage identification task.

[0062] S105, multi-task joint training;

[0063] A multi-task joint training model was constructed, comprising three tasks: degradation stage identification, prediction of future trends of key operating parameters, and RUL prediction. The encoding layer of the AE after unsupervised pre-training of S103 was used as the feature sharing layer for multi-task learning. A bidirectional long short-term memory network (Bi-LSTM) was used for RUL prediction and key operating parameter prediction, while a convolutional neural network (CNN) was used for degradation stage identification. Through the internal parameter sharing mechanism in multi-task learning, the accuracy of prediction of the remaining service life of aero-engines was improved.

[0064] S106, Model Application Testing;

[0065] The operating parameters of the aero-engines of the test object, after being screened by S102, are input into the trained multi-task prediction model to obtain the predicted values ​​of remaining service life, degradation stage, and future trends of key operating parameters. The average value of the multi-step RUL prediction values ​​output by the statistical model at the prediction time is used as the final RUL prediction value. The mean square error is used as the indicator to evaluate the prediction performance of the model's remaining service life.

[0066] In S101 provided in this embodiment of the invention, the original operating status monitoring signal data of the aero-engine is first read, and the multi-channel sensor signals are grouped and extracted according to the engine number. Then, the data is converted into a format suitable for analysis, and the sliding window method is used to divide it into fixed-length sequence samples. At the same time, a corresponding label (such as remaining life) is generated for each sample, missing values ​​are processed, and the data is normalized.

[0067] In S102 provided by the present invention, a three-dimensional evaluation system is constructed by using the monotonicity index Mon, the trend index Tred, and selecting the Spearman correlation coefficient between the higher comprehensive score and RUL, which effectively screens out key operating parameters with strong degradation characterization capabilities.

[0068] In S103 of this embodiment, an autoencoder network is used to perform unsupervised pre-training on the full lifecycle sequence of key operating parameters selected in S102. The intrinsic structure of the data is learned by minimizing the reconstruction error, and the obtained network parameters are used as initial parameters. Unlabeled data is used to provide a good initialization for the network, and labeled training can further optimize the network parameters and improve the performance of the model on specific tasks.

[0069] In S104 of this embodiment of the invention, the H-Kmeans clustering method is used to cluster the engine degradation stages, and the clustering labels are optimized by adding the mode voting post-processing to achieve accurate division of the degradation stages.

[0070] In S105 of this embodiment, the historical key operating parameters of the aero-engine are used as input, and the target output includes three items: the predicted future trend value of the key operating parameters, the predicted value of the remaining service life, and the degradation stage classification result. In time series modeling, in order to match the input dimension of the model, the original time series needs to be windowed, that is, it is divided into multiple time series segments according to a set window length. Then, a model training and testing sample dataset is constructed based on the key operating parameters of the aero-engine. This dataset corresponds to three types of label sets: one is the key operating parameter prediction label, the second is the remaining service life prediction label, and the third is the degradation stage classification label.

[0071] In S106, the average value of the multi-step RUL predictions output by the statistical model at the prediction time is used as the final RUL prediction value.

[0072] like Figure 2 As shown, an embodiment of the present invention provides an aero-engine remaining life prediction system based on multi-task semi-supervised learning, comprising:

[0073] The data acquisition and preprocessing unit is used to acquire, clean, and convert raw aero-engine operating status monitoring data;

[0074] The key operating parameter screening unit, based on the monotonicity and trend characteristics of the operating parameters and their Spearman correlation coefficient with RUL, screens out key operating parameters with strong degradation characterization capabilities, providing high-quality data input for subsequent modeling;

[0075] The unsupervised pre-training unit utilizes an autoencoder (AE) to pre-train unlabeled data from the entire lifecycle of an aero-engine.

[0076] The degradation stage segmentation unit utilizes clustering methods to segment the engine degradation stages, providing the necessary accurate labels for the degradation stage identification task;

[0077] A multi-task joint training unit is used to construct a multi-task model for predicting the RUL of aero engines. The encoding layer of the pre-trained AE is used as a feature sharing layer, Bi-LSTM is used as an independent task layer for predicting the trend of key operating parameters and RUL, and CNN is used as an independent task layer for identifying the degradation stage.

[0078] The model application testing unit is used to test the trained model and calculate the performance metrics of all test results.

[0079] This invention, based on the concept of multi-task semi-supervised learning, performs unified modeling of the entire life cycle operational data of aero-engines. Through the collaborative work of multiple units, it achieves accurate prediction of the remaining service life of aero-engines. During system operation, the data acquisition and preprocessing unit first collects multi-source operational status monitoring data generated by the aero-engine at different operational stages, and cleans outliers, missing values, and noise in the raw data. Simultaneously, it performs data format conversion and normalization to ensure comparability and consistency of data from different operating conditions and different sensor channels, providing a stable data foundation for subsequent analysis.

[0080] After completing the basic preprocessing, the key operating parameter screening unit performs feature-level constraints and screening on the operating status monitoring data. Starting from the engine degradation mechanism, this unit comprehensively analyzes the monotonicity and trend of each operating parameter throughout its entire life cycle, as well as its correlation with the remaining service life. Redundant parameters with weak characterization of the degradation process are eliminated, and only key operating parameters that can reflect the characteristics of engine health decline are retained. This reduces the dimensionality of the model input, reduces noise interference, and improves modeling efficiency and stability.

[0081] Building upon this foundation, the unsupervised pre-training unit utilizes a large amount of unlabeled key operational parameter data throughout the entire lifespan of the aero-engine, learning the inherent distribution characteristics of the data through an autoencoder structure. This process does not rely on remaining lifetime labels; by minimizing the difference between the input data and the reconstructed data, the model can pre-learn the potential feature representations of operational parameters evolving over time. This provides a reasonable initial feature space for subsequent supervised or semi-supervised tasks, alleviating the problem of insufficient training caused by a lack of labeled data.

[0082] Subsequently, the degradation stage segmentation unit divides the engine operation process into degradation stages based on the temporal variation characteristics of key operating parameters. Through cluster analysis, the engine operating state is divided into several stages with similar degradation characteristics, and the stage results are stabilized to ensure that the obtained degradation stage labels accurately reflect the engine's evolution from a healthy state to a failure state, providing a reliable basis for degradation stage identification.

[0083] In the multi-task joint training unit, the system uses the encoding structure obtained from unsupervised pre-training as a feature-sharing layer to guide different prediction tasks to learn collaboratively within the same feature space. Through time series modeling, it simultaneously predicts the future trends of key operating parameters and the remaining useful life, while also executing the degradation stage identification task in parallel. The multiple tasks constrain each other by sharing feature information, enabling the model to learn the degradation patterns while also considering the structural information of the degradation stages, thereby improving the overall accuracy and robustness of the predictions.

[0084] During the model application testing phase, the model application testing unit inputs the operational data of the aero-engine under test into the trained prediction model to obtain the remaining service life prediction results, degradation stage discrimination results, and future trend prediction results of operating parameters. Statistical analysis is then performed on the prediction output to evaluate the model's predictive performance and engineering applicability. Through this collaborative working mechanism, this invention achieves high-precision and stable prediction of the remaining service life of aero-engines, possessing significant potential for widespread application.

[0085] like Figure 3 As shown, the method for predicting the remaining service life of aero-engines based on multi-task semi-supervised learning according to the present invention includes the following steps:

[0086] Step 1: Acquisition and Preprocessing of Aircraft Engine Operation Status Monitoring Signals

[0087] First, the raw operating status monitoring signal data of the aero-engine is read, and the multi-sensor signals are grouped and extracted according to the engine number. Then, the data is converted into a format suitable for analysis and divided into fixed-length sequence samples using the sliding window method. At the same time, a corresponding label (such as remaining life) is generated for each sample. Missing values ​​are processed, and the data is normalized to lay the foundation for subsequent model training and prediction.

[0088] Step 2: Screening of key operating parameters based on monotonicity, trend, and Spearman correlation coefficient.

[0089] Because aero-engines have numerous operating parameters, including a large amount of redundancy and noise, a preliminary screening of engine operating parameters is conducted based on their monotonicity and trend characteristics. This is achieved by calculating the monotonicity index Mon and the trend index Tred, identifying several operating parameters with high overall scores. Following this initial screening, Spearman correlation coefficients between these operating parameters and their Remaining Service Life (RUL) are calculated to identify the key operating parameters most relevant to RUL, ensuring that subsequent analysis focuses on critical information.

[0090] The formula for calculating the monotonicity index Mon is as follows:

[0091] ,

[0092] in, yes The statistical characteristic value of time, where K is the number of observed times. It is a unit step function.

[0093] The formula for calculating the trend indicator Tred is as follows:

[0094] ,

[0095] in, This represents the weighted sum of the squares of the eigenvalues ​​at all time points. This represents the sum of squares at all points in time.

[0096] The formula for calculating the overall score is as follows:

[0097] ,

[0098] in It is the minimum / maximum value of the monotonicity index. It is the minimum / maximum value of the trend indicator, and the coefficient 1 / 2 corresponds to equal weight (making the comprehensive score range [0,1]).

[0099] The formula for calculating the Spearman coefficient is as follows:

[0100] ,

[0101] in, It is the difference in rank between two variables, where n is the sample size.

[0102] Step 3, unsupervised pre-training.

[0103] An autoencoder network is used to perform unsupervised pre-training on the lifecycle sequences of key operational parameters selected in step 2. The intrinsic structure of the data is learned by minimizing the reconstruction error, and the resulting network parameters are used as initial parameters. Unlabeled data is used to provide a good initialization for the network, while labeled training can further optimize the network parameters and improve the model's performance on specific tasks.

[0104] Step 4: Integrate the degeneracy stage division of H-Kmeans and mode voting.

[0105] The degradation stages are divided using H-Kmeans clustering, which combines hierarchical clustering and K-means. The clustering results are then processed through a mode vote to achieve accurate division of the aero-engine degradation stages. The specific process is as follows:

[0106] Hierarchical clustering is used to initially divide the data and generate an initial cluster structure;

[0107] (1) Extract the cluster centroids from the hierarchical clustering results and use them as the initial centroids of K-Means.

[0108] (2) Use K-Means to optimize these initial centroids to obtain the final clustering results;

[0109] (3) Based on the sliding window, the label at the center of the window is replaced with the label that appears most frequently in the window by voting on the cluster labels in each sliding window.

[0110] The steps of the H-Kmeans algorithm are as follows:

[0111] (1) First, read the data of the aero-engine, extract the feature columns that need to be clustered, and check and process missing values ​​to ensure data integrity.

[0112] (2) The data is initially grouped by hierarchical clustering, and the mean of each cluster is calculated as the initial cluster center of K-means to improve the stability and effectiveness of clustering.

[0113] (3) Using the above center point as the initial value, execute the K-means algorithm to iteratively cluster the data and finally obtain the cluster label of each sample.

[0114] (4) Apply a sliding window to smooth the clustering results to reduce noise and random fluctuations, making the labels more continuous and representative.

[0115] (5) Repeat steps 2 and 3 until the updated cluster centers no longer change.

[0116] After clustering, the initial clustering results undergo a mode vote post-processing to reduce discrete points and improve the continuity of the partitions. The following is a description of the mode vote steps and their formula:

[0117] (1) Let the size of the sliding window be w, and the center point of the window be the i-th data point. For each data point i, the range of the sliding window is .

[0118] ,

[0119] Where N is the total number of data points, This indicates rounding down to the nearest integer.

[0120] (2) For each sliding window Extract cluster labels from the window.

[0121]

[0122] in, Represents the cluster label of the j-th data point. Calculate the mode of the labels within the window (i.e., the label that appears most frequently).

[0123] ,

[0124] in, It is an indicator function, it is when The value is 1 if the condition is met, and 0 otherwise; k is the value of the cluster label.

[0125] (3) Change the label of the center point i of the window to the mode label.

[0126] ,

[0127] in, This indicates the smoothed label.

[0128] (4) Repeat the above steps for each data point i (i = 0, 1, ..., N−1) to obtain the smoothed label sequence.

[0129] ,

[0130] This method can effectively capture local structural features in data, and is especially suitable for processing degraded data with complex distribution characteristics.

[0131] Step 5: Building and training a multi-task semi-supervised RUL prediction model.

[0132] To build a multi-task prediction model, historical key operating parameters of the aero-engine are used as input. The target output includes three items: predicted future trends of key operating parameters, predicted remaining service life, and degradation stage classification results. In time series modeling, to match the model input dimensions, the original time series needs to be windowed, i.e., divided into multiple time series segments according to a set window length. Then, a model training and testing sample dataset is constructed based on the key operating parameters of the aero-engine. This dataset corresponds to three types of label sets: first, key operating parameter prediction labels; second, remaining service life prediction labels; and third, degradation stage classification labels.

[0133] Given a complete sequence of key operating parameters Its corresponding RUL sequence is Where n is the number of moments in the sequence of key operating parameters. This represents the key operating parameter value at time n. The problem of RUL prediction is essentially solving... to sequence The mapping relationship is shown in the following formula.

[0134] ,

[0135] in, This represents the nonlinear mapping relationship to be solved.

[0136] Knowing the sequence of key operating parameters at a certain moment The essence of predicting key operating parameters is trend prediction. The sequence labels for each time point are as follows:

[0137] ,

[0138] in, The window size for the time series. The prediction step size for key operating parameters.

[0139] The RUL prediction task employs a multi-step prediction and averaging method. When the prediction step size is set to 5, for the key operational parameter sequence at time n... The corresponding multi-step predicted RUL label is The corresponding labels for the predicted key operating parameters are: .

[0140] The multi-task semi-supervised RUL prediction model uses the encoding layer of the pre-trained AE as the feature sharing layer for multi-task learning. It performs RUL prediction and key runtime parameter prediction tasks based on a bidirectional LSTM network, and simultaneously performs degradation stage identification tasks based on a CNN. The network architecture and loss function are defined as follows:

[0141] 1) Backbone Network

[0142] First, an autoencoder is pre-trained on unlabeled data, preserving the training parameters. Then, the trained autoencoder is used as the backbone network for multi-task learning. By sharing this encoder layer, not only is the model complexity reduced, but information sharing and interaction between different tasks are also achieved, providing a more effective feature representation for the RUL prediction task.

[0143] 2) Task-Specific Networks

[0144] Three networks were designed for specific tasks: one for RUL prediction, one for critical operating parameter prediction, and one for degradation stage identification. For critical operating parameter prediction and RUL prediction, a Bi-LSTM was employed, enhancing the processing capability of sequence data by introducing a bidirectional information transfer mechanism. Two LSTM layers started from both ends of the sequence and converged towards the middle, ultimately merging their outputs at each time step. This merging could be a simple connection, a weighted average, or other more complex fusion strategies to produce the final output representation.

[0145] ,

[0146] ,

[0147] ,

[0148] in, The input features of the sequence data at each time step are represented. and These are the hidden states of the LSTM in two directions at time step t. This represents the weights used to merge the hidden states in both directions at time step. This represents the weights used for the backward hidden state at time step t. Indicates at time step The bias term.

[0149] For the task of identifying degradation stages, a CNN network structure is adopted to capture the stage feature differences of temporal data through local convolutional operations. The network consists of convolutional layers, pooling layers, and fully connected layers: the convolutional layers extract local features from the windowed temporal input, the pooling layers compress the feature dimensions and retain key patterns, and finally, the fully connected layers output the predicted probability distribution of each degradation stage.

[0150] ,

[0151] ,

[0152] ,

[0153] in, It is a windowed timing input. The convolution kernel matrix, For the bias of the convolutional layer, For activation function, It is the output feature map of the convolutional layer; It is the feature map output by the pooling layer; For the weights of the fully connected layer, For biasing the fully connected layer, The L2 regularization coefficient is... It is a prediction result of the degradation stage.

[0154] 3) Loss Function

[0155] The choice of loss function is crucial when designing a network. Since the network needs to handle three tasks simultaneously, three loss functions are required: the root mean square error (MSE) loss for prediction and the cross-entropy loss for classification. The expression for the root mean square error loss function is as follows:

[0156] ,

[0157] In the formula, n represents the sample size, and y represents the true value. This represents the predicted value.

[0158] The cross-entropy loss is expressed as follows:

[0159] ,

[0160] Where P is the target distribution and Q is the estimated distribution.

[0161] To enable collaborative training of the three tasks within the network, simply summing the three loss functions to obtain the final loss function is not optimal. Therefore, hyperparameters are introduced. and The weights of the auxiliary tasks are determined. The total loss is expressed as follows:

[0162] ,

[0163] in, This is the loss from the RUL prediction task. The loss is the cost of predicting key operating parameters. It is the loss identified during the degradation stage. and It is a weighting factor that controls auxiliary tasks. This is the loss from unsupervised learning using the AE autoencoder.

[0164] Step 6, Model application and RUL prediction.

[0165] The test dataset is preprocessed. A sliding window is applied to construct a dataset using the key operating parameters selected in step 2. This dataset is then input into the trained model, which outputs multi-step remaining useful life (RUL) predictions. The average of these multi-step RUL predictions at each prediction time is calculated and used as the final RUL prediction. Mean squared error (MSE) is used as the metric to evaluate the model's RUL prediction performance.

[0166] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

[0167] This invention provides a method for predicting the remaining life of aero-engines based on multi-task semi-supervised learning. The invention will be further described below with reference to specific embodiments.

[0168] Step 1, aircraft engine signal acquisition.

[0169] Experimental data were analyzed using data generated by the Commercial Modular Aviation Power Simulation System (C-MAPSS). C-MAPSS is a simulation tool designed for large civil aircraft engines, capable of simulating engines with thrust up to 90,000 pounds. It is suitable for simulating engine performance at flight altitudes of 0 to 40,000 feet, Mach numbers of 0 to 0.9, and sea-level temperatures of -60 to 103°F. Researchers introduced noise into the simulation process to simulate interference that might be encountered when data is collected by actual sensors. The resulting simulation data is presented as a 26-dimensional time series, detailing engine serial number, cycle count, flight parameters, and parameters from 21 sensors.

[0170] Step 2: Screening of key operating parameters based on monotonicity, trend, and Spearman correlation coefficient.

[0171] RUL prediction forecasts remaining service life from key parameter sequences using a nonlinear mapping relationship. The label decreases linearly from the maximum service life to 0, with a maximum value set at 125 cycles. Analysis was performed using the train_FD001 subset from C-MAPSS, which contains operating parameters for 100 engines with significantly different service life termination points. Key operating parameter labels are based on trend predictions of the time series; therefore, for the key operating parameter sequence at time n... The corresponding 5-step RUL prediction label is The corresponding labels for the 5-step prediction of key operational parameters for the auxiliary task are: The following is an example table 1 showing the construction of the labels for the FD001 dataset:

[0172] Table 1. Example of label construction for the FD001 dataset.

[0173]

[0174] The average monotonicity and trend of all key operating parameters for 100 engines in the train_FD001 file were calculated, and the monotonicity and trend were normalized using the min-max normalization method to obtain a comprehensive score. After selecting key operating parameters with good monotonicity and trend, the Spearman rank correlation coefficient was calculated for each data set and its RUL (Relative Limiting Rank), and then the parameters most correlated with the RUL prediction were selected. The calculation results are shown in Table 2. Figure 4 As shown:

[0175] Table 2. Monotonicity, Trend, and Overall Score of Key Operating Parameters

[0176]

[0177] To conduct an in-depth study, five sets of key operating parameters (5, 8, 11, 12, and 13) with excellent performance were selected. These parameters not only demonstrated good monotonicity but also good stability and resistance to interference in terms of trend. After selecting the key operating parameters with good monotonicity and trend, Spearman rank correlation coefficients were calculated for each set of data and its RUL (Relative Limiting Parameter), and the parameters most correlated with RUL prediction were then selected. The Spearman correlation coefficients between the key operating parameters and RUL are shown in Table 3.

[0178] Table 3. Correlation coefficients between key operating parameters and RUL's Spearman coefficients.

[0179]

[0180] As shown in the table above, the key operating parameters 5, 11, and 12 have high Spearman correlation coefficients with their corresponding RULs, with parameter 11 having the highest average correlation coefficient. Therefore, parameter 11 was selected as the key operating parameter for trend prediction in subsequent experiments. Partial data visualization of the three key operating parameters 5, 11, and 12 is shown below. Figure 5 As shown.

[0181] Step 3, unsupervised pre-training.

[0182] For the train_FD001 folder in the C-MAPSS dataset, the key operating parameters of all aero-engines were pre-trained unsupervised using an autoencoder AE network. The network structure is shown in Table 4.

[0183] Table 4 Parameter Settings for Unsupervised Pre-training Layers

[0184]

[0185] Step 4: H-Kmeans Degeneration Stage Division Based on Mode Voting

[0186] For the train_FD001 folder in the C-MAPSS dataset, key operating parameters 11 for all aero-engines were processed. Some visualization results are shown below. Figure 6 As shown in the figure, after applying the mode voting method, the distribution of data points exhibits a clearer clustering trend. This trend indicates that mode voting can further optimize the clustering results, making the data points within each cluster more compact and the boundaries between clusters more explicit. To evaluate the clustering effect of H-Kmeans, the results of other clustering methods were also post-processed using the mode voting strategy, and the processed silhouette coefficients were calculated. The specific results are shown in the figure. Figure 7 As shown.

[0187] Given the significant advantages of the H-Kmeans algorithm in clustering performance, applying it to the degradation stage segmentation of aero-engine data can provide accurate labels for the degradation stage identification task required for subsequent multi-task joint training of aero-engine RUL prediction, thus laying a solid foundation for further improving the accuracy and reliability of aero-engine RUL prediction.

[0188] Step 5: Training the RUL prediction model for aero-engines based on multi-task joint training

[0189] Building upon key operating parameter prediction and RUL prediction, a degradation stage identification task is incorporated to construct an aero-engine RUL prediction model based on multi-task joint training (Multi-Task Bidirectional Long Short-term Memory Neural Network, MT-Bi-LSTM). This model simultaneously performs RUL prediction, key operating parameter prediction, and degradation stage identification on aero-engine data based on multi-task learning. By treating key operating parameter prediction and degradation stage identification as auxiliary tasks, the model learns richer feature representations. An unsupervised feature extractor obtained through pre-training is used as a feature sharing layer in the multi-task learning process, sharing internal parameters and thus improving the accuracy of RUL prediction.

[0190] Table 5 Parameter Settings for Supervised Training Layers

[0191]

[0192] Table 6 CNN Network Parameter Settings

[0193]

[0194] The weights of the two auxiliary tasks were set to 0.3 and 0.4 respectively using a hyperparameter grid search. The model architecture is as follows: Figure 8 As shown. Detailed network structure settings are shown in Tables 5 and 6.

[0195] To verify the superiority of the multi-step RUL prediction averaging method, two different RUL prediction methods were tested: single-step prediction and multi-step prediction with overlapping averaging. The prediction results for three key operating parameters were compared. In single-step prediction, the model outputs a unique RUL value for each input sequence. These RUL values ​​are then concatenated to obtain the RUL prediction curve. In multi-step prediction, each input sequence generates multiple RUL prediction values ​​at the same time. To obtain more representative prediction results, the RUL prediction values ​​at the same time are averaged. The performance comparison between single-step and multi-step prediction is shown in Table 7.

[0196] Table 7 Comparison of prediction performance between single-step and multi-step overlapping mean-based methods

[0197]

[0198] From Table 7 above, Figure 9 As can be seen, the proposed model exhibits excellent prediction performance on all three test sets. Among them, the prediction accuracy of multi-step prediction with overlapping and averaging is better than that of single-step prediction, and this advantage is particularly significant in test sets 1 and 3.

[0199] To further demonstrate the superiority of the multi-task semi-supervised model proposed in this invention, a series of experiments were conducted to compare it with the model of this invention. Four comparison methods were used:

[0200] (1) The two-task model 2T-Bi-LSTM removes the degradation stage division task and only performs key parameter prediction and RUL prediction.

[0201] (2) Remove all auxiliary tasks, namely degradation stage division and key parameter prediction, and the Bi-LSTM model does not perform multi-task learning, but is first pre-trained on unlabeled data.

[0202] (3) RUL prediction model based on DCNN. DCNN is known for its powerful feature extraction capabilities and is particularly suitable for automatically learning complex spatial hierarchical structures from static images or time series data.

[0203] (4) RUL prediction model based on CNN-GRU. CNN-GRU is a supervised learning model that combines CNN and Gated Cycle Unit (GRU). It combines the advantages of CNN and GRU, so that the model can not only capture the local spatial features of the data, but also effectively handle temporal dependencies.

[0204] The above methods were used to conduct comparative analysis on the same test set, and the average value of the results was taken. The prediction accuracy evaluation is shown in Table 8.

[0205] As shown in Table 8, compared with the 2T-Bi-LSTM model and the single-task Bi-LSTM model, MT-Bi-LSTM exhibits superior performance in predicting the RUL of aero-engines. This improvement is mainly attributed to the parameter sharing mechanism of multi-task learning and the addition of two auxiliary tasks. When comparing these models, MT-Bi-LSTM demonstrates three core advantages: First, it fully utilizes a large amount of unlabeled data for pre-training and retains the model's initial training parameters, enabling the model to demonstrate better generalization ability in subsequent prediction tasks; second, the multi-step prediction overlap and averaging method significantly improves prediction robustness; third, the two auxiliary tasks of degradation stage segmentation and key operating parameter prediction provide the model with additional information dimensions, enabling the model to understand the engine degradation process from a more comprehensive perspective, which plays a positive role in improving model performance.

[0206] Table 8 Comparison of Model Prediction Accuracy

[0207]

[0208] It should be noted that embodiments of the present invention can be implemented in hardware, software, or a combination of both. The hardware portion can be implemented using dedicated logic; the software portion can be stored in memory and executed by a suitable instruction execution system, such as a microprocessor or dedicated-design hardware. Those skilled in the art will understand that the above-described devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuitry such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field-programmable gate arrays, programmable logic devices, etc., or by software executed by various types of processors, or by a combination of the above-described hardware circuitry and software, such as firmware.

[0209] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for predicting the remaining service life of an aero-engine based on multi-task semi-supervised learning, characterized in that, Includes the following steps: Acquire the operational status monitoring signals of the aero-engine from its initial health state to the end of its lifespan, and preprocess the operational status monitoring signals to obtain operational status monitoring data with a time sequence relationship. Based on the monotonicity and trend characteristics of the operational status monitoring data, as well as the Spearman correlation between the operational status monitoring data and the remaining service life of the aero-engine, the operational status monitoring data is filtered to obtain key operational parameters for characterizing engine degradation behavior. Unsupervised feature learning is performed using unlabeled key operational parameter lifecycle data to obtain shared feature representations for subsequent prediction tasks; Based on the temporal distribution characteristics of key operating parameters, the degradation stage of the aero-engine operation process is divided to obtain stage labels for degradation stage identification tasks. A multi-task joint learning model is constructed based on shared feature representation to simultaneously complete the tasks of identifying the degradation stage of aero-engines, predicting the future trend of key operating parameters, and predicting the remaining service life. The key operating parameters of the aero-engine under test are input into the multi-task joint learning model, and the corresponding remaining service life prediction results are output.

2. The method for predicting the remaining service life of aero-engines based on multi-task semi-supervised learning as described in claim 1, characterized in that, Preprocessing of aircraft engine operating status monitoring signals includes: Group the multi-source sensor signals according to the engine number; Perform missing value processing and normalization on the signal; The operating status monitoring signal is divided into time series samples of fixed length using a sliding window method.

3. The method for predicting the remaining service life of aero-engines based on multi-task semi-supervised learning as described in claim 1, characterized in that, The key operating parameters are selected using the following method: Calculate the monotonicity and trend indicators of each operating parameter throughout the entire life cycle; A comprehensive score is constructed based on the aforementioned monotonicity and trend indicators; Calculate the Spearman correlation coefficients between each operating parameter and the remaining service life; The operating parameters that meet the preset conditions in terms of both the comprehensive score and the Spearman correlation coefficient are selected as the key operating parameters.

4. The method for predicting the remaining service life of aero-engines based on multi-task semi-supervised learning as described in claim 1, characterized in that, The unsupervised feature learning is accomplished through an autoencoder network. By minimizing the reconstruction error between the input key operating parameters and the reconstructed output, a low-dimensional feature representation of the key operating parameters is learned. The learned network parameters are then used as the initial parameters for the multi-task joint learning model.

5. A method for predicting and modeling the remaining service life of aero-engines based on multi-task semi-supervised learning, characterized in that, Construct a multi-task joint learning model that includes a feature-sharing layer and multiple task branches. The feature sharing layer consists of an encoding network obtained through unsupervised pre-training; Based on the feature sharing layer, a classification task branch for identifying the degradation stage of aero-engines, a time series prediction task branch for predicting the future trend of key operating parameters, and a regression task branch for predicting the remaining service life are constructed respectively. During model training, the loss functions of the multiple tasks are jointly optimized so that the feature sharing layer can simultaneously serve the tasks of degradation stage identification, parameter trend prediction, and remaining useful life prediction.

6. The modeling method as described in claim 5, characterized in that, The degradation stage identification task divides key operational parameters into stages by fusing hierarchical clustering and cluster center updates. The clustering results are then subjected to a majority vote to obtain stable degradation stage labels.

7. The modeling method as described in claim 5, characterized in that, Both the task of predicting the future trends of key operating parameters and the task of predicting remaining useful life are implemented based on a bidirectional long short-term memory neural network. The degradation stage identification task is implemented based on a convolutional neural network. Different task branches are modeled independently after the feature sharing layer.

8. A system for predicting the remaining service life of an aero-engine based on multi-task semi-supervised learning, characterized in that, include: The data acquisition and preprocessing unit is used to acquire and preprocess the operating status monitoring signals of the aero-engine. The key operating parameter filtering unit is used to filter key operating parameters based on their degradation characteristics. Unsupervised feature learning unit is used to perform unsupervised feature learning on key operating parameters and generate shared feature representations; Degradation stage division unit, used to divide the degradation stage of the aero-engine operation process; The multi-task joint prediction unit is used to simultaneously output prediction results of the aero-engine degradation stage, prediction results of the future change trend of key operating parameters, and prediction results of the remaining service life based on shared feature representations.

9. The aircraft engine remaining service life prediction system as described in claim 8, characterized in that, The multi-task joint prediction unit includes a feature sharing module and multiple task prediction modules. The feature sharing module consists of an unsupervised pre-trained encoding network. Each task prediction module corresponds to the degradation stage identification task, the key operating parameter prediction task, and the remaining service life prediction task, respectively.

10. The aircraft engine remaining service life prediction system as described in claim 8, characterized in that, The system performs statistical processing on the predicted remaining useful life values ​​output at multiple prediction times during the prediction phase. The average of multiple prediction results is used as the final prediction result for the remaining service life of the aero-engine.

Images