Method for predicting residual service life of turbofan engine

Abstract

The invention discloses a method for predicting the remaining service life of a turbofan engine. The method comprises the steps of 1, collecting multi-dimensional sensor monitoring data of the turbofan engine, performing data processing, and dividing the data into a source domain and a target domain; 2, constructing a double-flow adaptive domain fusion convolutional network model; step 3, training a double-flow adaptive domain fusion convolutional network model, performing forward propagation on input data, extracting time and space features, performing fusion, calculating a loss function, optimizing model parameters by using a back propagation algorithm, and adjusting the model to minimize the loss function until training reaches a preset iteration number upper limit; and 4, inputting target domain test data into the trained model, predicting the cross-domain remaining service life, and evaluating the model. According to the method, high-precision residual life prediction under the small sample condition is realized, an efficient solution can be provided for predictive maintenance of high-value equipment such as aero-engines, and the method is suitable for health management of turbofan engines under complex working conditions.

Description

Technical Field

[0001] This invention relates to the field of industrial equipment health management technology, specifically to a method for predicting the remaining service life of a turbofan engine. Background Technology

[0002] In the field of industrial equipment health management, remaining useful life prediction is a key technology for ensuring equipment reliability and safety. However, existing methods face significant challenges when processing data under varying operating conditions. Complex equipment such as turbofan engines experience frequent changes in operating conditions, such as speed and load, causing the distribution of sensor monitoring data to shift over time. Traditional data-driven methods, such as convolutional neural networks and long short-term memory networks, typically assume that training and test data are independently and identically distributed, which is difficult to apply in real-world industrial scenarios. When the distributions of the training data (source domain) and the test data (target domain) differ significantly, directly applying these methods leads to a significant decrease in prediction accuracy and a high error rate.

[0003] Meanwhile, acquiring labeled data for the entire lifecycle of industrial equipment is costly and impractical, especially for high-value equipment such as aero engines. In real-world scenarios, obtaining sensor status monitoring data from the engine's complete lifespan to failure is both impractical and costly. The ratio of historical data from normal operation to failure data is severely imbalanced, making it expensive and difficult to obtain sufficient labeled data for remaining service life prediction tasks. Furthermore, existing technologies mainly rely on single-modal feature extraction methods, which struggle to simultaneously capture the temporal dynamic characteristics and spatial distribution characteristics of turbofan engine sensor data, resulting in insufficient ability to characterize complex degradation patterns.

[0004] To address these challenges, a method is urgently needed that can simultaneously preserve the essential attributes of the target domain's degradation features during domain alignment. Existing technologies have significant shortcomings in feature extraction and domain alignment. Current domain-adaptive remaining lifetime prediction methods primarily utilize domain adversarial or multi-kernel maximum mean difference (MMD) methods to obtain domain-invariant feature information from the data, such as adversarial domain adaptation and MMD alignment. While these methods can theoretically alleviate data distribution differences, they have significant limitations in practical applications. These methods mainly focus on extracting domain-invariant features, but often neglect the target domain's unique degradation information during alignment, particularly failing to effectively fuse temporal dynamic features with spatial local features, making it difficult for the model to adapt to the target domain's unique degradation patterns. Furthermore, existing technologies lack effective mechanisms for preserving target domain-specific information when dealing with complex scenarios involving multiple operating conditions and multiple failure modes, leading to decreased prediction performance under strong domain shifts. The failure to fully utilize the advantages of a dual-stream feature fusion architecture makes it difficult to simultaneously preserve the target domain's unique degradation features during dynamic domain alignment, limiting the model's generalization ability in cross-domain scenarios. Summary of the Invention

[0005] To address the problems of decreased prediction accuracy due to data distribution shifts, insufficient model generalization ability in small sample scenarios, and neglect of target domain-specific degradation features in the prediction of the remaining service life of turbofan engines under various operating conditions, this invention proposes a method for predicting the remaining service life of turbofan engines. The specific technical solution is as follows: A method for predicting the remaining service life of a turbofan engine includes the following steps: Step 1: Collect multi-dimensional sensor monitoring data of the turbofan engine, process the data and divide it into source domain and target domain to obtain a labeled source domain training set, an unlabeled target domain training set and a labeled target domain test set. Step 2: Construct a two-stream adaptive domain fusion convolutional network model; Step 3: Train the two-stream adaptive domain fusion convolutional network model. Input the labeled source domain training set and the unlabeled target domain training set into the two-stream adaptive domain fusion convolutional network model at the same time, perform forward propagation, extract and fuse temporal and spatial features, calculate the loss function and use the backpropagation algorithm to calculate the gradient of the loss function with respect to the model parameters, and then use the gradient descent algorithm to update the model parameters, adjust the model to minimize the loss function, until the training reaches the preset maximum number of iterations. Step 4: Input the labeled target domain test set into the trained model to predict the cross-domain remaining lifetime and evaluate the model.

[0006] Furthermore, step 1 specifically includes: Step 1.1: Collect real-time monitoring data from multiple sensors of the turbofan engine and sort the data according to the ID and cycle number columns. The monitoring data includes vibration, temperature and pressure. Step 1.2: Normalize the data using the min-max normalization method to scale the data range; Step 1.3: Divide the data into fixed-size windows. For training data, generate windowed data and corresponding labels: RUL, i.e., remaining lifetime. For test data, if the data length is less than the window size, use spline interpolation to expand the data to the window size. Step 1.4: Divide the labeled data of the source domain into a training set, that is, obtain the labeled source domain training set. Divide the target domain data into an unlabeled target domain training set and a labeled target domain test set at a ratio of 90% and 10%, respectively. The true RUL labels of the target domain test set are retained for the final evaluation.

[0007] Furthermore, in step 1.3, when the length of the test data is less than the preset window size, cubic spline interpolation is used to smoothly expand the data so that the data dimension matches the window size.

[0008] Furthermore, in step 2, the dual-stream adaptive domain fusion convolutional network model includes a source domain feature extraction module, a target domain feature extraction module, a RUL predictor, a domain discriminator, and a mutual information estimation module; The source domain feature extraction module includes a ConvLSTM layer and a Convolutional Neural Network (CNN) layer. The source domain temporal sensor data is input into the ConvLSTM layer to extract time-dependent features and obtain the features of the temporal flow. The source domain sensor grid data or the spatial representation after dimensionality reduction is input into the CNN layer to extract local spatial features and obtain the features of the spatial flow. By concatenating and integrating the features of the temporal flow and the spatial flow, the source domain comprehensive features are output. The target domain feature extraction module has the same structure as the source domain feature extraction module and shares the architecture parameters of ConvLSTM and CNN, but the parameters of the two are independent. It takes unlabeled target domain data as input and outputs comprehensive target domain features. The RUL predictor is built on a multilayer perceptron. The domain discriminator guides the source domain feature extraction module and the target domain feature extraction module to learn domain-invariant features through an adversarial learning mechanism; The mutual information estimation module maximizes the mutual information between the extracted target features and the target samples while learning the invariant representation of the target domain, so as to preserve the degradation feature information unique to the target domain.

[0009] Furthermore, in the multilayer perceptron of the RUL predictor, the input layer receives input data and passes it to the hidden layer. The hidden layer consists of multiple neurons, each of which performs linear combination and nonlinear activation function operations to obtain the output value of the hidden layer. Finally, the output layer of the last layer of the multilayer perceptron generates the final prediction result.

[0010] Furthermore, the construction of the domain discriminator specifically involves: employing a binary classification neural network, defining an adversarial adaptation loss, and using a mini-maximum game strategy during the training process of the binary classification neural network: when the target domain feature extraction module and the source domain feature extraction module are fixed, the discriminator parameters are optimized to maximize the adversarial adaptation loss, i.e., accurately distinguishing between source and target domain features; when the domain discriminator is fixed, the gradient direction is reversed through a gradient reversal layer (GRL), and the feature extractor parameters are optimized to minimize the adversarial adaptation loss, making it difficult for the domain discriminator to distinguish between source and target domain features; finally, through adversarial training, the target domain feature extraction module and the source domain feature extraction module generate domain-invariant features, achieving alignment of the source and target domain feature distributions.

[0011] Furthermore, the mutual information estimation module estimates the mutual information between the comprehensive features of the target domain and the original input data through a neural network, preserving the target domain task information during the domain alignment process. The original input data includes the time-series sensor data and sensor grid data of the target domain. The neural network has a multilayer perceptron interwoven with two feedforward neural networks and Gaussian error linear unit functions, and adopts the Donsker-Varadhan notation to transform the mutual information estimation into the optimization objective of the neural network.

[0012] Furthermore, in step 3, the loss function includes the RUL prediction loss of the RUL predictor, the adversarial adaptation loss of the domain discriminator, and the mutual information estimation loss of the mutual information estimation module.

[0013] Furthermore, the RUL prediction loss uses mean squared error, and the adversarial adaptation loss is defined using binary cross-entropy loss.

[0014] Furthermore, in step 4, the model evaluation includes prediction performance evaluation and cross-domain generalization ability analysis. The prediction performance evaluation quantitatively assesses the model's prediction accuracy and error distribution in the target domain by calculating the root mean square error and a scoring function, and combining this with a visual comparison of predicted and true values. The root mean square error is used to measure the average deviation between predicted and true values, and the scoring function is used to penalize errors. The cross-domain generalization ability analysis verifies whether the model's adaptability under diverse working conditions and its dependence on key features conform to domain knowledge by grouping and testing data under different working conditions and analyzing feature importance.

[0015] The advantages and beneficial effects of this invention are as follows: The method of this invention significantly improves the accuracy of remaining lifetime prediction and the model generalization ability in cross-domain scenarios through the collaborative design of a dual-stream feature fusion architecture and a mutual information estimation adversarial domain adaptation mechanism. The constructed dual-stream feature extraction architecture, which integrates temporal and spatial streams, extracts dynamic features from sensor temporal data using ConvLSTM while capturing spatial distribution features from sensor grid data using CNN. Finally, a feature fusion module integrates multimodal information to form a more comprehensive representation of engine health status. This addresses the problem that traditional single-stream feature extraction cannot simultaneously capture temporal dynamics and spatial heterogeneity, significantly enhancing the ability to identify complex degradation patterns. Through adversarial learning and mutual information maximization, it effectively solves the performance degradation problem caused by data distribution differences across operating conditions. A target domain adaptation strategy based on few-sample learning solves the dependence of traditional cross-domain models on large-scale labeled data, enabling fine-tuning of model parameters with only a small amount of unlabeled data in the target domain, quickly adapting to new operating conditions and significantly improving engineering deployment efficiency. A multi-stage joint optimization strategy solves the overfitting or underfitting problem caused by a single loss function. Through the dynamic synergy of RUL prediction loss, adversarial adaptation loss, and mutual information estimation loss, it effectively balances RUL prediction accuracy and domain adaptation objectives, further enhancing the model's generalization performance. Attached Figure Description

[0016] Figure 1 This is a flowchart of a method for predicting the remaining service life of a turbofan engine according to an embodiment of the present invention; Figure 2 This is a data processing flowchart in an embodiment of the present invention; Figure 3 This is a structural diagram of the dual-stream adaptive domain fusion convolutional network model in an embodiment of the present invention; Figure 4 This is a flowchart illustrating the principle of a method for predicting the remaining service life of a turbofan engine in an embodiment of the present invention. Figure 5 These are parameters in the embodiments of the present invention. Performance impact diagrams in 12 cross-domain scenarios. Detailed Implementation

[0017] To make the objectives, technical solutions, and technical effects of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0018] like Figure 1 As shown in this embodiment, a method for predicting the remaining service life of a turbofan engine includes the following steps: Step 1: Collect multi-dimensional sensor data from the turbofan engine, process the data, and divide it into source and target domains. For example... Figure 2 As shown, the specific steps include the following: Step 1.1: Sensor Screening: Collect real-time monitoring data from multiple sensors on the turbofan engine, including vibration, temperature, and pressure. Remove unnecessary columns and sort the data by ID and cycle count columns for easier subsequent processing and analysis.

[0019] Step 1.2: Normalization: Normalize the data using the min-max normalization method to scale the data to a specified range.

[0020] Data normalization is performed on the grouped data after standardization is applied to column setting1. The formula for standardization is as follows: ; The formula for calculating min-maximum linear normalization is as follows: ; in, x For sensor data, It is a normalized value. and These are the minimum and maximum values ​​for each feature (column). By default, the normalized data range is [0,1].

[0021] Step 1.3: Time window segmentation: Divide the data into fixed-size windows. For training data, generate windowed data and corresponding labels: RUL, i.e. remaining lifetime. For test data, if the data length is less than the window size, use spline interpolation to expand the data to the window size.

[0022] When the length of the test data is less than the preset window size, cubic spline interpolation is used to smoothly expand the data to ensure that the data dimension matches the window size. Specifically, this method constructs a piecewise cubic polynomial to achieve a smooth transition from the original data points to the target data points while ensuring data continuity. The cubic spline polynomial is used in each interval... The calculation formula is as follows: ; in, In the first i Interpolation polynomial over intervals, , , , The coefficients of the spline polynomial are determined by the interpolation conditions and continuity constraints. The location of the interpolation point. For the first i The starting point of each interval.

[0023] Step 1.4: Divide the labeled data of the source domain into a training set to obtain the labeled source domain training set. Divide the target domain data into 90% and 10% to obtain the unlabeled target domain training set and the labeled target domain test set, respectively. Retain the real RUL labels of the target domain test set for final evaluation.

[0024] Step 2: As Figure 3 As shown, a two-stream adaptive domain fusion convolutional network model is constructed, including a source domain feature extraction module, a target domain feature extraction module, a RUL predictor, a domain discriminator, and a mutual information estimation module. The specific steps for constructing the two-stream adaptive domain fusion convolutional network model are as follows: Step 2.1: Constructing the Source Domain Feature Extraction Module: The source domain feature extraction module includes a Convolutional Long Short Memory (ConvLSTM) layer and a Convolutional Neural Network (CNN) layer. The ConvLSTM layer inputs source domain temporal sensor data to extract time-dependent features, obtaining the temporal flow features. The CNN layer inputs source domain sensor grid data or a dimensionality-reduced spatial representation to extract local spatial features, obtaining the spatial flow features. The feature extraction module, by concatenating and integrating the temporal and spatial flow features, outputs the comprehensive source domain features. .

[0025] Time-series streams are used to process time-series sensor data from the source domain and extract time-dependent features. This is achieved by inputting time-series sensor data from the source domain. ,in t This represents the time step. ConvLSTM is used to extract time-dependent features. The specific formula for ConvLSTM is as follows: ; in, , , These are the forget gate, input gate, and output gate, respectively. Indicates a time step. In cellular state, and For the weight matrix and bias terms, It is the Sigmoid activation function. The hyperbolic tangent activation function is used. This represents the convolution operation. This represents the Hadamard product, which is an element-wise multiplication.

[0026] Spatial flow is used to represent sensor mesh data from the source domain, extracting local spatial features. This is achieved by inputting sensor mesh data from the source domain. CNN is used to process the input data and extract spatial dependency features. The specific formula for CNN is as follows: ; ; ; in, This is the output feature map of the convolution operation; For bias terms; For activation functions; This is a max pooling operation; The feature map after activation; This is the feature map after pooling.

[0027] Feature fusion involves concatenating features extracted from temporal and spatial streams along the channel dimension to form a comprehensive feature from the source domain. The specific formula for feature fusion is as follows: ; in, This represents the comprehensive characteristics of the source domain. It is a time-dependent feature. These are local spatial features.

[0028] Step 2.2: Construct the target domain feature extraction module: Construct a target domain feature extraction module with the same structure as the source domain feature extraction module and sharing the ConvLSTM and CNN architecture parameters, but with independent parameters. The input is unlabeled target domain data, and the output is comprehensive target domain features. .

[0029] Step 2.3: Constructing the RUL predictor: Based on the comprehensive features extracted from the source domain, construct a multilayer perceptron-based RUL predictor to predict the remaining lifespan of the device.

[0030] In a multilayer perceptron, the input layer receives input data: synthesized features from the source domain. Combined features of the target domain , recorded as The results are then passed to the hidden layer. The hidden layer consists of multiple neurons, each performing linear combination and non-linear activation function operations. The intermediate value is first obtained through linear combination. The specific formula is as follows: ; in, The number of input features, As weight, For bias, This refers to the input data.

[0031] The output value of the hidden layer is then obtained through a non-linear activation function. The specific formula is as follows: ; in, This represents the activation value of the j-th neuron, i.e., the output value after processing by the ReLU activation function. For the ReLU activation function itself, This is an intermediate value obtained from a linear combination.

[0032] The output layer is the last layer of the multilayer perceptron and is responsible for generating the final prediction result. The specific formula is as follows: ; in, The output of the hidden layer, As weight, The number of input data.

[0033] Step 2.4: Constructing a domain discriminator: Through an adversarial learning mechanism, guide the feature extraction module to learn domain-invariant features and construct a domain discriminator.

[0034] The core objective of constructing a domain discriminator is to guide the feature extraction module to learn domain-invariant features, thereby reducing the distributional differences between the source and target domains. A binary classification neural network is used as the domain discriminator D, and an adversarial adaptation loss is defined. During training, a mini-maximum game strategy is employed: when the feature extraction module is fixed, the discriminator parameters are optimized to maximize... This means accurately distinguishing features between the source and target domains; when the discriminator is fixed, the gradient inversion layer (GRL) is used to invert the gradient direction, optimizing the parameters of the feature extraction module to minimize [the desired feature size]. This makes it difficult for the discriminator to distinguish between source and target domain features. Ultimately, adversarial training enables the feature extraction module to generate domain-invariant features, aligning the feature distributions of the source and target domains and improving the model's generalization ability in the target domain.

[0035] Step 2.5: Construct a mutual information estimation module: While learning the invariant representation of the target domain, maximize the mutual information between the extracted target features and the target samples to preserve the degenerate feature information unique to the target domain.

[0036] The mutual information estimation module is used to estimate the comprehensive features of the target domain. Compared to the raw input data: time-series sensor data in the target domain Sensor grid data in the target domain The mutual information between the domains preserves task-related information specific to the target domain during domain alignment. This mutual information can be estimated using a neural network, with the specific formula as follows: ; in, Represents entropy, Represents conditional entropy. Represents the marginal probability distribution. This represents the joint probability distribution. Mutual information can be equivalently represented as the joint probability distribution. With marginal probability distribution The KL divergence between them. Therefore, we have the following formula: ; in, Let KL divergence be the KL divergence. As expected.

[0037] For ease of computation, the Donsker-Varadhan representation is used to transform mutual information estimation into a neural network. The optimization objective is given by the following formula: ; in, The neural network, consisting of two feedforward neural networks and a multilayer perceptron interleaved with Gaussian error linear unit functions, is used to estimate the joint probability distribution. For the splicing operation, and and Concatenate as joint input. To and The sample order is shuffled to generate negative sample pairs to estimate the marginal distribution. This is to take the average value of the batch samples.

[0038] Step 3: Train the network using training samples, optimize the network parameters, and obtain the network model.

[0039] During the training phase, labeled (marked) source domain and unlabeled (unmarked) target domain training data are simultaneously input into a two-stream adaptive domain fusion convolutional network model for forward propagation. The overall loss function is obtained by calculating the remaining lifetime prediction loss, adversarial adaptation loss, and mutual information estimation loss. The model parameters are then optimized using the backpropagation algorithm to minimize the loss function. Training continues until a preset maximum number of iterations is reached. Figure 4 As shown, the model domain adaptive training includes the following steps: Step 3.1: Input Data: Input the labeled source domain training set and unlabeled target domain training set Simultaneously, the inputs are fed into a two-stream adaptive domain fusion convolutional network model. , and , Represent the source domain and the target domain respectively. iThe and the first j The temporal and spatial feature vectors of each sample For source domain samples , The corresponding real tags, and These represent the number of samples in the source domain and the target domain, respectively. In this embodiment of the invention, the selection of the source domain and the target domain is based on different subsets of the C-MAPSS dataset: FD001~FD004. Each subset has different fault modes and operating conditions to simulate cross-domain scenarios in actual industry. In each experiment, one dataset is selected as the source domain, and the other three datasets are selected as the target domains. During the model training phase, labeled source domain training data and 90% unlabeled target domain training data are used for training, while 10% of the labeled target domain training data is used as the validation set.

[0040] In this embodiment of the invention, 12 cross-domain RUL prediction tasks were designed to comprehensively evaluate the performance of the two-stream adaptive domain fusion convolutional network model, as shown in Table 1: Table 1: 12 sets of cross-domain RUL prediction tasks of the present invention

[0041] Step 3.2: Forward Propagation: Perform forward propagation calculations on the input source and target domain samples.

[0042] During the forward propagation process, the model extracts feature representations from samples in both the source and target domains. (Source domain samples) , The comprehensive features obtained after forward propagation are: Target domain samples , The obtained comprehensive features are represented as .

[0043] Step 3.3: Calculate the loss function: During training, three main loss functions need to be optimized, namely RUL prediction loss, adversarial adaptation loss, and mutual information estimation loss, to finally obtain the overall loss function.

[0044] Mean squared error is used as the loss function for RUL prediction. For source domain samples, the predicted value is... The formula for predicting loss using RUL is as follows: ; in, The loss function for RUL prediction. The number of samples in the source domain. For source domain samples , The corresponding real tags, The model is based on the comprehensive features of the source domain samples. The predicted RUL value.

[0045] The adversarial adaptation loss is defined using the binary cross-entropy loss, and the specific formula is as follows: ; in, For the loss of antagonistic adaptation, and These represent the number of samples in the source domain and the target domain, respectively. For the discriminator to identify the features of the i-th source domain sample The probability of belonging to the source domain. For the discriminator to identify the features of the j-th target domain sample The probability that it belongs to the source domain.

[0046] Based on the mutual information estimation method, the formula for mutual information estimation loss is as follows: ; in, To estimate the loss using mutual information, To take the average value, For neural networks, For splicing operations, For mixed washing operation, For neural network pairs and , The joint probability distribution of the concatenated variables is estimated. For neural network pairs and , shuffle variables The product of estimated marginal probabilities.

[0047] The overall loss function formula is as follows: ; in, For RUL, predict loss. For the loss of antagonistic adaptation, To estimate the loss using mutual information, These are hyperparameters and need to be adjusted experimentally to balance the impact of different loss terms on model training.

[0048] Step 3.4: Update model parameters: Calculate the overall loss function using the backpropagation algorithm. The gradient of the model parameters is then calculated, and the gradient descent algorithm is used to update the model parameters to obtain the minimum. And update the model's training parameters. .

[0049] The specific formula for updating model parameters using the gradient descent algorithm is as follows: ; in, For each trainable parameter in the model, The learning rate controls the step size for each parameter update.

[0050] Step 3.5: Repeat the above steps until the preset maximum number of iterations is reached. n In each iteration, the model continuously adjusts its parameters to better adapt to the data distribution in the source and target domains, thereby improving the RUL prediction performance in the target domain.

[0051] Step 4: Input the test data from the target domain into the trained model to predict the remaining useful life across domains. The model outputs the prediction results, which can be used to evaluate the model's prediction performance and generalization ability under different operating conditions. Specifically, this includes the following steps: Step 4.1: Input and prediction of target domain test data: Input the preprocessed target domain test data into the trained model, perform forward propagation, output the RUL prediction value and inverse normalize it to obtain the prediction result in the actual time unit; Step 4.2: Prediction performance evaluation: By calculating indicators such as root mean square error and scoring function, and combining the visualization comparison between predicted values ​​and true values, the prediction accuracy and error distribution of the model in the target domain are quantitatively evaluated. Root mean square error (RMSE) measures the average deviation between predicted and actual values. The specific formula is as follows: ; in, N For the number of test samples, For the first i The true remaining useful life of each sample For the first i Predicted remaining useful life for each sample.

[0052] Step 4.3: Cross-domain generalization ability analysis: By testing data under different working conditions in groups and analyzing the importance of features, we can verify whether the model's adaptability under diverse working conditions and the dependence of key features are consistent with domain knowledge.

[0053] In this embodiment of the invention, to verify the prediction performance of the unsupervised domain adaptation method, eight repeated experiments were conducted on 12 cross-domain RUL prediction tasks, based on the RMSE and Score evaluation metrics. The performance of the two-stream adaptive domain fusion convolutional model was compared with that of the Only-Source model and the Only-Target model. The Only-Source model was trained using only source domain data and did not employ a domain adaptation method; its pre-trained model directly performed RUL predictions on test data from other domains. The Only-Target model, on the other hand, was trained using target domain data and then used test data from the target domain for prediction. The experimental mean was used as the final experimental result, as shown in Table 2. Table 2: Comparison of RUL prediction performance of the present invention with Only-Source model and Only-Target model

[0054] Table 3 shows a comparison of the cross-domain RUL prediction performance of this invention based on the RMSE standard and several unsupervised domain adaptive methods:

[0055] The optimal value is marked in bold, and the second-best value is marked with an underline.

[0056] Table 4 shows a comparison of the cross-domain RUL prediction performance of this invention based on the Score metric with several other unsupervised domain adaptive methods: Table 4: Score results of cross-domain RUL prediction of this invention and 5 other unsupervised domain adaptive methods

[0057] The optimal value is marked in bold, and the second-best value is marked with an underline.

[0058] like Figure 5 As shown, the mutual information estimation loss in this embodiment of the invention... The impact of weights may vary in different cross-domain scenarios, and it can be observed that different cross-domain scenarios may require different weights. To improve performance. Mutual information is used to estimate the loss. The optimal values ​​for the weights in the 12 cross-domain scenarios are shown in Table 5: Table 5: Best corresponding results in 12 cross-domain RUL prediction tasks value

[0059] The results show that the remaining life prediction method proposed in this invention can extract more representative cross-domain degradation features, more accurately describe the degradation process of turbofan engines, and achieve more accurate remaining life prediction.

[0060] In summary, this invention presents a method for predicting the remaining service life of turbofan engines based on adaptive dual-stream mutual information estimation, which can solve the problem of remaining service life prediction in cross-domain scenarios. This invention collects and processes multi-dimensional sensor data from turbofan engines, constructs a dual-stream adaptive domain fusion convolutional network model, extracts features from temporal and spatial streams and fuses them to form a comprehensive representation of the engine's health status; it addresses the problem of data distribution differences across operating conditions through adversarial learning mechanisms and mutual information maximization; and it constructs a multi-task learning framework, dynamically adjusting hyperparameters to balance the influence of different loss values, thereby improving the model's prediction accuracy and generalization ability in the target domain. This invention achieves high-precision remaining service life prediction under small sample conditions, providing an efficient solution for predictive maintenance of high-value equipment such as aero-engines, and is applicable to the health management of turbofan engines under complex operating conditions and the predictive maintenance of complex equipment such as aero-engines.

[0061] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting the remaining service life of a turbofan engine, characterized in that, The steps include: Step 1: Collect multi-dimensional sensor monitoring data of the turbofan engine, process the data and divide it into source domain and target domain to obtain a labeled source domain training set, an unlabeled target domain training set and a labeled target domain test set. Step 2: Construct a two-stream adaptive domain fusion convolutional network model; Step 3: Train the two-stream adaptive domain fusion convolutional network model. Input the labeled source domain training set and the unlabeled target domain training set into the two-stream adaptive domain fusion convolutional network model at the same time, perform forward propagation, extract and fuse temporal and spatial features, calculate the loss function and use the backpropagation algorithm to calculate the gradient of the loss function with respect to the model parameters, and then use the gradient descent algorithm to update the model parameters, adjust the model to minimize the loss function, until the training reaches the preset maximum number of iterations. Step 4: Input the labeled target domain test set into the trained model to predict the cross-domain remaining lifetime and evaluate the model.

2. The method for predicting the remaining service life of a turbofan engine according to claim 1, characterized in that, Step 1 specifically includes: Step 1.1: Collect real-time monitoring data from multiple sensors of the turbofan engine and sort the data according to the ID and cycle number columns. The monitoring data includes vibration, temperature and pressure. Step 1.2: Normalize the data using the min-max normalization method to scale the data range; Step 1.3: Divide the data into fixed-size windows. For training data, generate windowed data and corresponding labels: RUL, i.e., remaining lifetime. For test data, if the data length is less than the window size, use spline interpolation to expand the data to the window size. Step 1.4: Divide the labeled data of the source domain into a training set, that is, obtain the labeled source domain training set. Divide the target domain data into an unlabeled target domain training set and a labeled target domain test set at a ratio of 90% and 10%, respectively. The true RUL labels of the target domain test set are retained for the final evaluation.

3. The method for predicting the remaining service life of a turbofan engine according to claim 2, characterized in that, In step 1.3, when the length of the test data is less than the preset window size, cubic spline interpolation is used to smoothly expand the data so that the data dimension matches the window size.

4. The method for predicting the remaining service life of a turbofan engine according to claim 1, characterized in that, In step 2, the dual-stream adaptive domain fusion convolutional network model includes a source domain feature extraction module, a target domain feature extraction module, a RUL predictor, a domain discriminator, and a mutual information estimation module. The source domain feature extraction module includes a ConvLSTM layer and a ConvNN layer. The source domain time-series sensor data is input into the ConvLSTM layer to extract time-dependent features and obtain the features of the time-series flow. Input source domain sensor grid data or dimensionality-reduced spatial representation into the CNN layer, extract local spatial features to obtain spatial flow features, and output source domain comprehensive features by splicing and integrating temporal flow and spatial flow features. The target domain feature extraction module has the same structure as the source domain feature extraction module and shares the architecture parameters of ConvLSTM and CNN, but the parameters of the two are independent. It takes unlabeled target domain data as input and outputs comprehensive target domain features. The RUL predictor is built on a multilayer perceptron. The domain discriminator guides the source domain feature extraction module and the target domain feature extraction module to learn domain-invariant features through an adversarial learning mechanism; The mutual information estimation module maximizes the mutual information between the extracted target features and the target samples while learning the invariant representation of the target domain, so as to preserve the degradation feature information unique to the target domain.

5. The method for predicting the remaining service life of a turbofan engine according to claim 4, characterized in that, In the multilayer perceptron of the RUL predictor, the input layer receives input data and passes it to the hidden layer. The hidden layer consists of multiple neurons, each of which performs linear combination and nonlinear activation function operations to obtain the output value of the hidden layer. Finally, the output layer of the last layer of the multilayer perceptron generates the final prediction result.

6. The method for predicting the remaining service life of a turbofan engine according to claim 4, characterized in that, The construction of the domain discriminator is specifically as follows: a binary classification neural network is adopted, an adversarial adaptation loss is defined, and a minimum-maximum game strategy is adopted during the training process of the binary classification neural network: when the target domain feature extraction module and the source domain feature extraction module are fixed, the discriminator parameters are optimized to maximize the adversarial adaptation loss, that is, to accurately distinguish the source domain and target domain features. When the domain discriminator is fixed, the gradient direction is reversed by the gradient reversal layer GRL, and the feature extractor parameters are optimized to minimize the adversarial adaptation loss, making it difficult for the domain discriminator to distinguish between source domain and target domain features. Finally, adversarial training enables the target domain feature extraction module and the source domain feature extraction module to generate domain-invariant features, thereby aligning the feature distributions of the source domain and the target domain.

7. The method for predicting the remaining service life of a turbofan engine according to claim 4, characterized in that, The mutual information estimation module estimates the mutual information between the comprehensive features of the target domain and the original input data through a neural network, preserving the target domain task information during the domain alignment process. The original input data includes the time-series sensor data and sensor grid data of the target domain. The neural network has a multilayer perceptron interwoven with two feedforward neural networks and Gaussian error linear unit functions, and adopts the Donsker-Varadhan notation to transform the mutual information estimation into the optimization objective of the neural network.

8. The method for predicting the remaining service life of a turbofan engine according to claim 4, characterized in that, In step 3, the loss function includes the RUL prediction loss of the RUL predictor, the adversarial adaptation loss of the domain discriminator, and the mutual information estimation loss of the mutual information estimation module.

9. The method for predicting the remaining service life of a turbofan engine according to claim 8, characterized in that, The RUL prediction loss uses mean squared error, and the adversarial adaptation loss uses binary cross-entropy loss.

10. The method for predicting the remaining service life of a turbofan engine according to claim 1, characterized in that, In step 4, the model evaluation includes prediction performance evaluation and cross-domain generalization ability analysis. The prediction performance evaluation quantitatively assesses the model's prediction accuracy and error distribution in the target domain by calculating the root mean square error and a scoring function, combined with a visual comparison of predicted and true values. The root mean square error is used to measure the average deviation between predicted and true values, and the scoring function is used to penalize errors. The cross-domain generalization ability analysis verifies whether the model's adaptability under diverse working conditions and its dependence on key features conform to domain knowledge by grouping and testing data under different working conditions and analyzing feature importance.

Images