Two-stage residual life prediction method of aircraft hydraulic system based on data dimension reduction

Abstract

The invention relates to the field of reliability analysis of aircraft hydraulic systems, in particular to a two-stage residual life prediction method of an aircraft hydraulic system based on data dimension reduction, which comprises the following steps: acquiring state data of the aircraft hydraulic system and preprocessing the state data; sequentially carrying out principal component analysis and kernel principal component analysis on the preprocessed data to obtain dimension reduction data; establishing a dual-target prediction model; inputting the trend term and the periodic term into a double-target prediction model to obtain a preliminary life prediction value; inputting the dimension reduction data into a residual prediction network to obtain a life residual prediction value; the residual prediction network is composed of a one-dimensional convolutional neural network and a bidirectional long-short term memory network which are connected in sequence; adding the initial life prediction value and the life residual prediction value to obtain a final residual life prediction value; according to the method, the residual life prediction efficiency and robustness of the aircraft hydraulic system can be improved.

Description

Technical Field

[0001] This invention relates to the field of aircraft hydraulic system reliability analysis, and specifically to a two-stage remaining life prediction method for aircraft hydraulic systems based on data dimensionality reduction. Background Technology

[0002] As a critical power source, the health of aircraft hydraulic systems is essential for flight safety. Remaining Useful Life (RUL) prediction, a crucial aspect of aviation equipment health management, can effectively improve equipment reliability and reduce maintenance costs. Existing RUL prediction methods mainly include physical model-based methods, statistical methods, and data-driven methods. Physical model methods rely on theories such as stress-strain analysis and fatigue damage modeling. While possessing strong interpretability, they require extensive specialized knowledge and are difficult to adapt to complex and dynamically changing aviation systems. Statistical methods (such as Markov processes and Bayesian updates) build degradation models from historical data, providing some predictive capability, but they have limitations when dealing with nonlinear and multivariable systems. In recent years, data-driven methods, leveraging machine learning and deep learning techniques (such as random forests, support vector machines, and long short-term memory networks), have been widely applied in RUL prediction. They do not require complex physical modeling and can more efficiently extract equipment operating characteristics for prediction.

[0003] While data-driven methods have achieved good results in Restricted Length Usage (RUL) prediction, they still face several challenges. First, the sensor data from aircraft hydraulic systems is high-dimensional, containing multiple variables such as temperature, pressure, flow rate, and vibration. Directly using this data for modeling increases computational complexity and may lead to model overfitting. Second, RUL prediction relies on long-term trend analysis, but traditional methods struggle to effectively model time-series characteristics, resulting in poor prediction stability. Furthermore, single prediction models often cannot simultaneously account for short-term fluctuations and long-term trends, making them susceptible to noise and affecting prediction accuracy. Therefore, there is an urgent need to propose a more efficient and robust RUL prediction method. This method should improve computational efficiency through enhanced data dimensionality reduction techniques and combine multi-stage prediction strategies to improve model accuracy, meeting the high-precision prediction requirements of aviation hydraulic systems and providing reliable assurance for equipment maintenance and flight safety. Summary of the Invention

[0004] In view of the above problems, the present invention provides a two-stage remaining life prediction method for aircraft hydraulic systems based on data dimensionality reduction, which solves the technical problems of low remaining life prediction efficiency and poor robustness in the prior art.

[0005] This invention provides a two-stage remaining life prediction method for aircraft hydraulic systems based on data dimensionality reduction, comprising the following steps:

[0006] Step S1: Use sensors to collect status data of the aircraft hydraulic system, and preprocess the status data, including missing data filling, noise reduction and normalization.

[0007] Step S2: Perform principal component analysis and kernel principal component analysis on the preprocessed data in sequence to obtain dimensionality-reduced data;

[0008] Step S3: Use period-trend decomposition to obtain the trend and period items of the dimensionality-reduced data;

[0009] A bi-objective prediction model is established by optimizing the prediction results of random forest and support vector machine to minimize the correlation between the prediction results and the estimation error of the prediction results.

[0010] Input the trend term and periodic term into the bi-objective prediction model to obtain preliminary lifetime prediction values;

[0011] Step S4: Input the dimensionality-reduced data into the residual prediction network to obtain the lifetime residual prediction value; the residual prediction network consists of a one-dimensional convolutional neural network and a bidirectional long short-term memory network connected in sequence.

[0012] Step S5: Add the preliminary lifetime prediction value and the lifetime residual prediction value to obtain the final remaining lifetime prediction value.

[0013] Preferably, step S1 specifically includes:

[0014] Step S1-1: Collect state data of the aircraft hydraulic system in six dimensions using temperature, pressure, flow, vibration, speed and displacement sensors;

[0015] Step S1-2: Calculate the mean of the non-missing state data of the same sensor at different time points to fill in the missing state data;

[0016] Steps S1-3: Using the empirical mode decomposition method, extract multiple intrinsic mode functions from the state data in order of high frequency to low frequency, remove the first preset number of intrinsic mode functions, and reconstruct the denoised data based on the remaining intrinsic mode functions;

[0017] Steps S1-4: Use the Z-score standardization method to normalize the noise reduction data to obtain preprocessed data.

[0018] Preferably, step S2 specifically includes:

[0019] Step S2-1: Calculate the covariance matrix, perform eigenvalue decomposition, and select principal components on the preprocessed data to obtain preliminary dimensionality-reduced data;

[0020] Step S2-2: Perform Gaussian kernel determination, Gaussian kernel centering, eigenvalue decomposition, and projection calculation on the preliminary dimensionality reduction data to obtain the final dimensionality reduction data.

[0021] Preferably, step S2-2 specifically includes:

[0022] Step S2-2-1: The preliminary dimensionality reduction data contains multiple samples. A Gaussian kernel function is selected, and the kernel matrix K is calculated based on the Gaussian kernel function and all the samples.

[0023] Step S2-2-2: Center the kernel matrix, perform eigenvalue decomposition on the centered kernel matrix, and extract the top q eigenvectors with the largest eigenvalues;

[0024] Step S2-2-3: Project the preliminary dimensionality reduction data based on the eigenvalues ​​of the first q feature vectors and the Gaussian kernel function to obtain dimensionality reduction data of dimension q.

[0025] Preferably, step S3 specifically includes:

[0026] Step S3-1: Use the period-trend decomposition algorithm to decompose the dimensionality-reduced data into trend items, periodic items, and random items;

[0027] Step S3-2: Using random forest and support vector machine as base learners, the base learners process historical data in the trend and periodic terms to obtain lifetime prediction values. The optimization objective is to minimize the correlation between the lifetime prediction values ​​obtained by random forest and support vector machine and the estimation error of the lifetime prediction values, thus obtaining a bi-objective prediction model.

[0028] Step S3-3: Input the current data in the trend term and period term into the bi-objective prediction model to obtain the preliminary lifetime prediction value.

[0029] Preferably, step S4 specifically includes:

[0030] Step S4-1: Establish the residual prediction network, which specifically includes:

[0031] Step S4-1-1: Determine the input of the convolutional layer of the residual prediction network, including: using multiple remaining lifetime residuals as inputs to the convolutional layer, wherein the remaining lifetime residuals represent the difference between the predicted and actual remaining lifetime values.

[0032] Step S4-1-2: Determine the network structure. The network structure sequentially passes the input of the convolutional layer through a one-dimensional convolutional layer and a pooling layer, input smoothing, a fully connected layer, and a one-dimensional bidirectional long short-term memory network to finally obtain the lifetime residual prediction value.

[0033] Step S4-2: Determine the residuals between multiple preliminary lifetime predictions and actual values ​​from the dimensionality reduction data, input them into the residual prediction network, and obtain the lifetime residual prediction values.

[0034] Preferably, in step S4-1-4, the input to the convolutional layer is defined as Where n k Indicates the total number of observations. This is the θth observation of the aircraft hydraulic power system. The remaining lifetime residual at time , This indicates that the aircraft hydraulic system obtains the first predicted value for the θth time. At that moment, This represents the max-th predicted value obtained by the aircraft hydraulic system for the θth time.

[0035] Preferably, in step S3-2, the expression for the optimization objective is:

[0036]

[0037] Among them, F1 θ minF1 represents the correlation of the prediction results at the θ-th observation. θ Minimize F1 θ n θ M represents the total number of data points at the θ-th observation. θ Let θ be the number of base learners at the θ-th observation. Let q' represent the weight of the m-th basic learner at the θ-th observation, and q' represent another basic learner different from m. This represents the factorization term of the m-th basic learner at time θ-th observation. The predicted value, * = Trend,Seasonal, indicates that the decomposition term is a trend term or a periodic term. This represents the average value of the decomposition terms of the m-th basic learner at time t during the θ-th observation. This represents the factorization term of the q'-th basic learner at time θ-th observation. The predicted value, This represents the time of the j'-th sample in the initial dimensionality reduction data obtained in the θ-th iteration. This represents the average value of the decomposition terms of the q'-th basic learner at time t during the θ-th observation. This represents the estimation error of the prediction result at the θ-th observation. Minimize This indicates that the decomposition term at time θ represents the value of the term at the θth observation. The true value.

[0038] Compared with the prior art, the present invention has at least the following beneficial effects:

[0039] (1) This invention employs an improved kernel principal component analysis method to reduce the dimensionality of aircraft hydraulic system monitoring data, which overcomes the limitation of traditional principal component analysis, which can only process linear data. By performing principal component analysis and kernel principal component analysis sequentially on the preprocessed state data, the nonlinear characteristics in the hydraulic system monitoring data can be fully preserved and extracted, and the efficiency of data preprocessing can be significantly accelerated.

[0040] (2) This invention provides a two-stage remaining lifetime prediction method. In the first stage, the dimensionality-reduced data is decomposed into a periodic-trend model. By minimizing the correlation and estimation error between the prediction results of random forest and support vector machine, a bi-objective prediction model is established to obtain preliminary lifetime prediction values. In the second stage, the dimensionality-reduced data is input into a residual prediction network composed of a one-dimensional convolutional neural network and a bidirectional long short-term memory network to learn and correct the residuals of the preliminary prediction results. This two-stage division of labor and cooperation takes into account the diversity of models and utilizes the modeling ability of deep learning for temporal nonlinear residuals, effectively improving the accuracy and robustness of the entire remaining lifetime prediction.

[0041] (3) The method proposed in this invention has been verified by actual aircraft hydraulic system data. The mean absolute error (MAE) and root mean square error (RMSE) of the prediction are significantly lower than those of common prediction models such as random forest, decision tree regression, linear regression and CNN-BiLSTM, showing stronger prediction accuracy and anti-interference ability. Compared with existing traditional prediction models, the phased modeling and multi-method fusion of this invention further improves the ability to characterize the life evolution of complex aviation hydraulic systems. Attached Figure Description

[0042] The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of the invention.

[0043] Figure 1 The flowchart shows the two-stage remaining life prediction method for aircraft hydraulic systems based on data dimensionality reduction provided by this invention.

[0044] Figure 2 This is a schematic diagram of the feature variables before and after EMD and Z-score, as provided by the present invention.

[0045] Figure 3 This is a schematic diagram of KPCA graph data before and after dimensionality reduction provided by the present invention.

[0046] Figure 4 A comparison chart of predicted and actual values ​​from the two-stage remaining lifetime prediction model provided by this invention. Detailed Implementation

[0047] To better understand the above-described objectives, features, and advantages of the present invention, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments of the present invention and the features thereof can be combined with each other. Furthermore, the present invention can be implemented in other ways different from those described herein; therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below.

[0048] This invention employs improved kernel principal component analysis (KPCA) for data dimensionality reduction, reducing redundant information and improving data processing efficiency. It also utilizes a two-stage prediction framework combining sequence decomposition, ensemble learning, and deep learning to fully leverage the advantages of multiple prediction techniques, thereby enhancing the accuracy and robustness of RUL prediction. This method can more accurately assess the health status of aircraft hydraulic systems, providing strong data support for equipment maintenance decisions.

[0049] To illustrate the effectiveness of the method proposed in this invention, the following detailed description of the above technical solution is provided through a specific embodiment, such as... Figure 1 As shown, a two-stage remaining life prediction method for aircraft hydraulic systems based on data dimensionality reduction is disclosed. The specific implementation steps are as follows:

[0050] Step S1: Use sensors to collect status data of the aircraft hydraulic system, and preprocess the status data, including missing data filling, noise reduction and normalization.

[0051] The data acquisition sources of this invention include multiple sensors such as pressure, temperature, and flow rate. However, the raw data often suffers from problems such as inconsistent dimensions, missing data, and duplication. Furthermore, sensor signals are susceptible to noise interference in complex operating environments. To ensure data quality and improve model prediction accuracy, this invention preprocesses the acquired data, including missing data imputation, denoising, and normalization, to reduce the impact of data incompleteness and abnormal interference on the prediction results. This lays a solid foundation for subsequent feature extraction and analysis, as described in detail below.

[0052] (1) Real-time data acquisition

[0053] This invention acquires high-precision operational status data by real-time monitoring of key components in an aircraft's hydraulic system. The sensor data collected by this invention can include multiple physical parameters such as temperature, pressure, flow rate, vibration, rotational speed, and displacement, which effectively reflect the health status of the hydraulic system. The sensor types and data units used in this invention are shown in Table 1.

[0054] Table 1

[0055] serial number Sensor type unit 1 Temperature sensor ℃ 2 pressure sensor kPa 3 Flow sensor L / min 4 Vibration sensor μm 5 Speed ​​sensor r / min 6 Displacement sensor mm

[0056] However, due to the complex operating environment of aviation equipment, sensor data may be subject to external interference during acquisition, leading to problems such as data loss, noise pollution, and dimensional differences, which can affect the stability and accuracy of the prediction model. Therefore, preprocessing of the acquired data is necessary to improve data quality and provide reliable data input for subsequent RUL prediction.

[0057] (2) Filling missing data

[0058] To address the issue of missing data, this invention employs mean imputation to handle missing values. Specifically, it calculates the mean of valid data from the same sensor at different time points to fill in missing data, thereby reducing the impact of incomplete data on model training. Mean imputation estimates missing values ​​by calculating the mean of non-missing monitoring data within the variable. First, the mean of the non-missing values ​​in the variable is calculated. Then, the calculated mean is used as the imputation value to replace the original missing data. This process is repeated for each variable with missing values ​​until all missing values ​​are filled. The expression is:

[0059]

[0060] in, Fill in the missing data for the θ-th observation of the i-th sensor parameter. This represents the j-th valid observation value of the θ-th observation of the i-th sensor parameter. Let m be the total number of valid observations for the θ-th observation of the i-th sensor parameter, where θ is the observation index, m is the total number of observations, and i is the sensor parameter index. θ Let θ be the total number of sensor parameters observed in the θth observation.

[0061] (3) Noise interference removal

[0062] To address noise interference in the data, this invention employs Empirical Mode Decomposition (EMD) to denoise the original sensor signal. By decomposing the original signal and removing high-frequency noise components, the stability of the data is effectively improved. The goal of the EMD algorithm is to decompose a complex signal layer by layer into multiple Intrinsic Mode Function (IMF) components. The IMFs are arranged from high frequency to low frequency, with the first IMF containing the highest frequency component of the signal, and subsequent IMFs containing progressively decreasing frequency components. The remaining signal is called the residual component. n (), usually represents the long-term trend component of the signal.

[0063] The expression for the empirical mode decomposition method is:

[0064]

[0065] Among them, X k Let i1 represent the reconstructed signal after noise reduction, and n represent the index of the intrinsic mode function (IMF). Let res represent the i1th intrinsic mode function. n This represents the residual component. Each extracted residual component is equal to the previous extracted residual component minus the extracted intrinsic mode function.

[0066] In some embodiments, for multiple intrinsic mode functions arranged from high frequency to low frequency, the present invention removes a preset number of intrinsic mode functions that are ranked first, and reconstructs the remaining intrinsic mode functions to form noise-reduced data.

[0067] (4) Data normalization

[0068] Different sensors produce output signals with varying dimensions and ranges. The differences in the numerical distribution characteristics of the six sensors provided in this invention can affect the training and prediction performance of the model. Therefore, it is necessary to normalize the data from each sensor. Normalization not only eliminates the influence of dimensionality but also accelerates the learning algorithm and improves the efficiency of model training. This invention employs the Z-score normalization method to convert raw sensor data of different magnitudes into data of the same magnitude, enhancing data comparability.

[0069] After the above data preprocessing, this invention can obtain more stable and standardized input data, ensuring the accuracy of subsequent feature extraction, dimensionality reduction analysis, and RUL prediction. Figure 2 As shown, the data after EMD noise reduction and Z-score normalization is smoother than the original data, and the comparability of data from different sensors is significantly enhanced, providing reliable data support for subsequent modeling.

[0070] Step S2: Perform principal component analysis and kernel principal component analysis on the preprocessed data to obtain dimensionality-reduced data;

[0071] Aircraft hydraulic system monitoring data contains a large amount of redundant information, and directly using it for model training would result in high computational complexity and low efficiency. Therefore, this invention provides a nonlinear data dimensionality reduction method based on improved kernel principal component analysis (Kernel-PCA). First, the covariance matrix of the preprocessed data is calculated, and principal eigenvectors are extracted for initial dimensionality reduction. Subsequently, a Gaussian kernel function is used for nonlinear feature mapping to further extract nonlinear features, achieving efficient dimensionality reduction for aircraft hydraulic system data, as described in detail below.

[0072] Step S2-1: Calculate the covariance matrix, perform eigenvalue decomposition, and select principal components on the preprocessed data to obtain preliminary dimensionality-reduced data.

[0073] First, the data X obtained from the above data preprocessing... * Calculate the covariance matrix Where X *T For X * The covariance matrix is ​​the transpose of the standardized data set, where n is the total number of data points. It measures the linear correlation between variables. The elements of the covariance matrix represent the covariance between different variables, reflecting the degree of correlation between them.

[0074] Perform eigenvalue decomposition on the covariance matrix to obtain eigenvalues ​​and corresponding eigenvectors C = VΛV T , where V,V T Let represent the eigenvector matrix and its transpose, respectively, and Λ represent the eigenvalue diagonal matrix. The magnitude of the eigenvalues ​​reflects the variance of the data along the corresponding eigenvector direction, i.e., the importance of the principal components.

[0075] Based on the magnitude of the eigenvalues, the eigenvectors corresponding to the first p eigenvalues ​​in the eigenvalue diagonal matrix are selected as principal components P1. This ensures that the cumulative variance contribution rate of the selected principal components reaches a preset threshold (e.g., 95%) to preserve the main information of the data and obtain preliminary dimensionality-reduced data X1 = X * P1.

[0076] Step S2-2: Perform Gaussian kernel determination, Gaussian kernel centering, eigenvalue decomposition, and projection calculation on the preliminary dimensionality reduction data to obtain the final dimensionality reduction data.

[0077] To address the issue that PCA cannot effectively handle nonlinear data, this invention further employs an improved kernel principal component analysis (KPCA) method for efficient dimensionality reduction. KPCA transforms the original nonlinear data into a high-dimensional feature space through kernel function mapping, making it easier to extract key features and perform dimensionality reduction in this space. Specifically, the dimensionality reduction process of KPCA includes the following steps:

[0078] Calculate the kernel matrix using a Gaussian kernel: Select the Gaussian kernel function. The preliminary dimensionality reduction data contains multiple samples, x 1i' Let x represent the i'-th sample of the initial dimensionality reduction data. 1j' Let K1(x) represent the j'-th sample. 1i' ,x 1j' ) represents sample x 1i' and x 1j' The kernel function values ​​between the points are given, where ||·|| represents the Euclidean distance and σ represents the width parameter of the Gaussian kernel function. The kernel matrix between all data sample points is calculated to capture the nonlinear characteristics of the data.

[0079] Centering: The kernel matrix is ​​centered to eliminate the influence of data distribution on dimensionality reduction and improve the robustness of feature extraction. The centering process can be achieved by adjusting the mean of the kernel matrix so that the center of the data distribution mapped to the feature space is located at the origin.

[0080] The formula for centering the kernel matrix is ​​as follows:

[0081] K centered =K-1 n ×KK×1 n +1 n ×K×1 n

[0082] Where K is the original kernel matrix, 1 n It is an n×n matrix and each element has a value of n is the number of samples, K centered The kernel matrix after centralization;

[0083] This formula eliminates the mean shift of the data distribution in the feature space, ensuring that the data distribution mapped to the feature space is centered at the origin, thereby improving the robustness of feature extraction and the dimensionality reduction effect.

[0084] Perform eigenvalue decomposition: Perform eigenvalue decomposition on the centered kernel matrix, and extract the top q eigenvectors with the largest eigenvalues ​​to obtain the most representative principal component information.

[0085] Calculate the projection of the new sample: Project the new input data onto the reduced feature space to obtain an optimized low-dimensional data representation. The expression for the projection calculation is as follows.

[0086]

[0087] Where, α 1i” Let λ represent the projected coordinates of the new input data sample onto the i”-th feature vector after dimensionality reduction, where i” = 1, 2, ..., q. 1i” Let u represent the i”th eigenvalue of the centered kernel matrix. 1i”j' K1(x) represents the association weight between the i”-th eigenvector of the centered kernel matrix and the j’-th sample, n’ is the total number of samples, x1 represents the new sample, and K1(x) represents the association weight between the i”-th eigenvector and the j’-th sample. 1i' ,x 1j' ) represents samples x1 and x 1j' The kernel function values ​​between.

[0088] Obtaining the dimensionality reduction result: The final dimensionality reduction data is α1 = (α 11 ,…,α 1q ) TDimensionality reduction can retain the main features of the original data while effectively removing redundant information, thus achieving efficient dimensionality reduction of aircraft hydraulic system data.

[0089] Step S3: Use period-trend decomposition to obtain the trend and period items of the dimensionality-reduced data;

[0090] A bi-objective prediction model is established by optimizing the prediction results of random forest and support vector machine to minimize the correlation between the prediction results and the estimation error of the prediction results.

[0091] Input the trend term and periodic term into the bi-objective prediction model to obtain preliminary lifetime prediction values;

[0092] To improve the accuracy and real-time performance of remaining service (RUL) prediction for aircraft hydraulic systems, this invention provides a two-stage RUL prediction method. This method fully considers the system performance degradation trend and nonlinear variation characteristics, as described below.

[0093] STL Decomposition (Period-Trend Decomposition): Utilizing the STL time series decomposition algorithm, dimensionality-reduced data is decomposed into trend, periodic, and random terms to reduce the complexity of the data sequence and extract system performance degradation patterns. The trend term reflects the long-term trend of the data, the periodic term captures periodic fluctuations, and the random term includes random fluctuations and noise. The expression is as follows:

[0094]

[0095] In the above formula, This represents the j'-th sample obtained from the θ-th time in the aircraft hydraulic system, and the corresponding sequence is... Let θ represent the time of the j'th sample obtained in the θth iteration. f represents the trend term, periodic term, and stochastic term of the aircraft hydraulic system operation sequence, respectively. θ' (·) is the decomposition function of the aircraft hydraulic power system operation sequence.

[0096] Multi-objective optimization: For trend and periodic terms, this invention provides a multi-objective optimization weight ensemble learning model. Using random forest and support vector machine as base learners, a bi-objective optimization algorithm that minimizes prediction error and inter-learner correlation determines the weights of each base learner, obtaining a high-precision preliminary RUL prediction value. This method significantly reduces the error caused by the inadequacy of a single prediction model.

[0097] Considering the limitations of a single prediction model, this invention employs a combined learning and prediction method to predict the trend and periodic terms of aircraft hydraulic system data. Random forest is a non-parametric supervised learning method that constructs multiple decision trees during training and aggregates their prediction results. It handles high-dimensional data well and has strong resistance to overfitting. On the other hand, support vector machine is a powerful classification and regression algorithm. It aims to find the optimal hyperplane in high-dimensional space to separate data of different categories and can effectively handle nonlinear problems. Therefore, this invention selects the widely used random forest and support vector machine as the basic learners and adopts a selective ensemble learning method with the decision objective of minimizing correlation and estimation error. The expression for minimizing correlation and estimation error as the decision objective is:

[0098]

[0099] Among them, F1 θ minF1 represents the correlation of the prediction results at the θ-th observation. θ Minimize F1 θ n θ M represents the total number of data points at the θ-th observation. θ Let θ be the number of base learners at the θ-th observation. Let q' represent the weight of the m-th basic learner at the θ-th observation, and q' represent another basic learner different from m. This represents the factorization term of the m-th basic learner at time θ-th observation. The predicted value, wherein the decomposition term can be a trend term or a period term. This represents the average value of the decomposition terms of the m-th basic learner at time t during the θ-th observation. This represents the factorization term of the q'-th basic learner at time θ-th observation. The predicted value, This represents the average value of the decomposition terms of the q'-th basic learner at time t during the θ-th observation. This represents the estimation error of the prediction result at the θ-th observation. Minimize This indicates that the decomposition term at time θ represents the value of the term at the θth observation. The true value.

[0100] The dual-objective prediction model is obtained through optimization training by minimizing correlation and estimation error as described above. Finally, the current data from the trend and periodic terms are input into the dual-objective prediction model to obtain preliminary lifetime predictions.

[0101] Step S4: Input the dimensionality-reduced data into the residual prediction network to obtain the lifetime residual prediction value; the residual prediction network consists of a one-dimensional convolutional neural network and a bidirectional long short-term memory network connected in sequence.

[0102] This invention further provides a one-dimensional convolutional neural network-bidirectional long short-term memory network (1DCNN-BiLSTM) model to finely model and correct the difference (residual) between the initial lifetime prediction and the true value, thereby improving the accuracy and robustness of aircraft hydraulic system remaining life (RUL) prediction. This method utilizes CNN to extract local features from the data, and BiLSTM to capture long-term dependencies in time-series information. Through deep learning, it effectively reduces the error in the initial prediction and improves the accuracy of the final prediction result, as described in detail below.

[0103] Convolutional layer input determination: The convolutional layer input for predicting the RUL residuals of the aircraft hydraulic power system is defined as follows: Where n k Indicates the total number of observations. This is the θth observation of the aircraft hydraulic power system. The remaining lifetime residual at time , This indicates that the aircraft hydraulic system obtains the first predicted value for the θth time. At that moment, The value represents the max-th predicted value obtained by the aircraft hydraulic system for the θth time, and the remaining life residual represents the difference between the predicted value and the actual value of the remaining life.

[0104] CNN layer feature extraction: When modeling the initial prediction error (residual), the convolutional and pooling layers in a one-dimensional convolutional neural network (1D CNN) are first used to extract local features of the data.

[0105] Input smoothing: After passing through the convolutional and pooling layers of the CNN, and before entering the fully connected layer, the above input needs to be smoothed. In other words, the original multidimensional variables are sequentially transformed into one-dimensional variables as input to the fully connected layer.

[0106] BiLSTM Temporal Modeling: After extracting local features using CNN, a bidirectional long short-term memory network (BiLSTM) is used for further modeling to capture long-term dependencies and improve the temporal stability of residual prediction.

[0107] Residual prediction: Based on a one-dimensional CNN-BiLSTM model, the life residual prediction values ​​of the aircraft hydraulic system were further obtained.

[0108] Step S5: Add the preliminary lifetime prediction value and the lifetime residual prediction value to obtain the final remaining lifetime prediction value.

[0109] Finally, the preliminary lifetime prediction value and lifetime residual prediction value By adding them together, we obtain the final predicted value of the RUL sequence of the aircraft hydraulic system.

[0110] To illustrate the effectiveness of the method proposed in this invention, the following detailed description of the above technical solution of this invention is provided through a specific embodiment.

[0111] Example 1

[0112] Using actual operating data from the hydraulic system of a large domestically produced aircraft as an example, simulation experiments were conducted to verify the effectiveness of the method of this invention. This aircraft hydraulic system comprises two parts: a main hydraulic power system and an auxiliary hydraulic power system, covering several key components such as hydraulic pumps, accumulators, flow-stabilizing solenoid valves, and unloading solenoid valves. The main hydraulic power source system provides hydraulic energy to systems requiring hydraulic power. It consists of left, center, and right hydraulic power source systems, as well as a scavenging system. The auxiliary hydraulic power source system provides emergency hydraulic energy to the flight control system when all four engines of the aircraft stop.

[0113] Step 1: Data Acquisition and Preprocessing

[0114] First, sensor data from the aircraft's hydraulic system are collected, including key indicators such as temperature, pressure, flow rate, vibration, rotational speed, and displacement. Because sensor data may contain missing values, noise interference, and dimensional differences, preprocessing is necessary. Mean imputation is used to complete missing data, followed by EMD (Enhanced Motion Modeling) to process the imputed data. Then, Z-score normalization is used to normalize the data, transforming variables of different magnitudes into the same magnitude. The feature variables before and after EMD reconstruction and Z-score normalization are shown below. Figure 2 As shown.

[0115] Step 2: Data Dimensionality Reduction

[0116] Raw data obtained from aircraft hydraulic systems typically contains multiple high-dimensional variables, and directly using it for model training can lead to high computational complexity and overfitting. Therefore, an improved kernel principal component analysis (KPCA) method is used to reduce the dimensionality of the data. Figure 3 As shown, the data features are more obvious after dimensionality reduction using KPCA, and the data dimensionality is significantly reduced, laying a good data foundation for subsequent prediction work.

[0117] Step 3: Two-stage RUL prediction

[0118] Based on a one-dimensional CNN-BiLSTM model, the prediction residuals for the RUL of the aircraft hydraulic system were further derived. The initial predicted value was added to the prediction residuals to obtain the final predicted value of the RUL of the aircraft hydraulic system. A comparison of the predicted values ​​and actual values ​​from the two-stage RUL prediction model is shown below. Figure 4 As shown.

[0119] Step 31: Preliminary prediction based on STL decomposition

[0120] Step 311: Sequence Decomposition: Using the STL decomposition method, the sensor data sequence is decomposed into trend, seasonal and irregular components, and each component is modeled separately.

[0121] Step 312: Multi-objective optimization ensemble learning: Random forest (RF) and support vector machine (SVM) are used as base learners, and genetic algorithm is used for multi-objective optimization to build an ensemble learning model.

[0122] Step 313: Preliminary RUL prediction: Train the preprocessed data using an ensemble learning model to obtain preliminary RUL prediction results.

[0123] Step 32: Residual Fine Prediction Based on One-Dimensional CNN-BiLSTM

[0124] Step 321: Residual Calculation: Calculate the residual between the preliminary prediction result and the actual RUL, and construct the residual dataset.

[0125] Step 322: One-dimensional CNN feature extraction: A one-dimensional convolutional neural network (CNN) is used to extract features from the residual data to capture local features.

[0126] Step 323: BiLSTM Temporal Modeling: Use a bidirectional long short-term memory network (BiLSTM) to perform temporal modeling on the extracted features to capture long-term dependencies.

[0127] Step 324: Residual Correction: The preliminary prediction results are corrected using the output of BiLSTM to obtain the final RUL prediction value.

[0128] Step 4: Performance Evaluation

[0129] To evaluate the performance of the proposed method, root mean square error (RMSE) and mean absolute error (MAE) were used as evaluation metrics.

[0130] To further verify the predictive performance of the proposed model, this paper uses MAE and RMSE as evaluation metrics to compare the predictive performance of different prediction models. The smaller the RMSE and MAE values, the better the predictive ability of the prediction model. The comparison of evaluation metrics for different prediction models is shown in Table 2.

[0131] Table 2

[0132] Predictive Model RMSE MAE Random Forest (RF) 50.3836 46.4558 Decision Tree Regression 53.3374 49.3374 Linear Regression 49.4302 45.4378 CNN-BiLSTM 47.6002 41.4487 Two-stage Remaining Life (RUL) Prediction Method 20.9965 15.3395

[0133] The prediction results show that, compared with single models (such as RF and CNN-BiLSTM), the two-stage prediction method proposed in this invention achieves better performance in both RMSE and MAE metrics, verifying the effectiveness of the method.

[0134] While the specific embodiments of the present invention depict actions or steps in a particular order, this should be understood as requiring such actions or steps to be performed in the specific order shown or in sequential order, or requiring all illustrated actions or steps to be performed to achieve the desired result. In certain environments, multitasking and parallel processing may be advantageous. Similarly, although several specific implementation details are included in the above discussion, these should not be construed as limiting the scope of this disclosure. Certain features described in the context of individual embodiments may also be implemented in combination in a single implementation. Conversely, various features described in the context of a single implementation may also be implemented individually or in any suitable sub-combination in multiple implementations.

[0135] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A two-stage remaining life prediction method for aircraft hydraulic systems based on data dimensionality reduction, characterized in that, Includes the following steps: Step S1: Use sensors to collect status data of the aircraft hydraulic system, and preprocess the status data, including missing data filling, noise reduction and normalization. Step S2: Perform principal component analysis and kernel principal component analysis on the preprocessed data in sequence to obtain dimensionality-reduced data; Step S3: Use period-trend decomposition to obtain the trend and period items of the dimensionality-reduced data; A bi-objective prediction model is established by optimizing the prediction results of random forest and support vector machine to minimize the correlation between the prediction results and the estimation error of the prediction results. Input the trend term and periodic term into the bi-objective prediction model to obtain preliminary lifetime prediction values; Step S4: Input the dimensionality-reduced data into the residual prediction network to obtain the lifetime residual prediction value; the residual prediction network consists of a one-dimensional convolutional neural network and a bidirectional long short-term memory network connected in sequence. Step S5: Add the preliminary lifetime prediction value and the lifetime residual prediction value to obtain the final remaining lifetime prediction value.

2. The two-stage remaining life prediction method for aircraft hydraulic systems based on data dimensionality reduction according to claim 1, characterized in that, Step S1 specifically includes: Step S1-1: Collect state data of the aircraft hydraulic system in six dimensions using temperature, pressure, flow, vibration, speed and displacement sensors; Step S1-2: Calculate the mean of the non-missing state data of the same sensor at different time points to fill in the missing state data; Steps S1-3: Using the empirical mode decomposition method, extract multiple intrinsic mode functions from the state data in order of high frequency to low frequency, remove the first preset number of intrinsic mode functions, and reconstruct the denoised data based on the remaining intrinsic mode functions; Steps S1-4: Use the Z-score standardization method to normalize the noise reduction data to obtain preprocessed data.

3. The two-stage remaining life prediction method for aircraft hydraulic systems based on data dimensionality reduction according to claim 2, characterized in that, Step S2 specifically includes: Step S2-1: Calculate the covariance matrix, perform eigenvalue decomposition, and select principal components on the preprocessed data to obtain preliminary dimensionality-reduced data; Step S2-2: Perform Gaussian kernel determination, Gaussian kernel centering, eigenvalue decomposition, and projection calculation on the preliminary dimensionality reduction data to obtain the final dimensionality reduction data.

4. The two-stage remaining life prediction method for aircraft hydraulic systems based on data dimensionality reduction according to claim 3, characterized in that, Step S2-2 specifically includes: Step S2-2-1: The preliminary dimensionality reduction data contains multiple samples. A Gaussian kernel function is selected, and the kernel matrix K is calculated based on the Gaussian kernel function and all the samples. Step S2-2-2: Center the kernel matrix, perform eigenvalue decomposition on the centered kernel matrix, and extract the top q eigenvectors with the largest eigenvalues; Step S2-2-3: Project the preliminary dimensionality reduction data based on the eigenvalues ​​of the first q feature vectors and the Gaussian kernel function to obtain dimensionality reduction data of dimension q.

5. The two-stage remaining life prediction method for aircraft hydraulic systems based on data dimensionality reduction according to claim 4, characterized in that, Step S3 specifically includes: Step S3-1: Use the period-trend decomposition algorithm to decompose the dimensionality-reduced data into trend items, periodic items, and random items; Step S3-2: Using random forest and support vector machine as base learners, the base learners process historical data in the trend and periodic terms to obtain lifetime prediction values. The optimization objective is to minimize the correlation between the lifetime prediction values ​​obtained by random forest and support vector machine and the estimation error of the lifetime prediction values, thus obtaining a bi-objective prediction model. Step S3-3: Input the current data in the trend term and period term into the bi-objective prediction model to obtain the preliminary lifetime prediction value.

6. The two-stage remaining life prediction method for aircraft hydraulic systems based on data dimensionality reduction according to claim 5, characterized in that, Step S4 specifically includes: Step S4-1: Establish the residual prediction network, which specifically includes: Step S4-1-1: Determine the input of the convolutional layer of the residual prediction network, including: using multiple remaining lifetime residuals as inputs to the convolutional layer, wherein the remaining lifetime residuals represent the difference between the predicted and actual remaining lifetime values. Step S4-1-2: Determine the network structure. The network structure sequentially passes the input of the convolutional layer through a one-dimensional convolutional layer and a pooling layer, input smoothing, a fully connected layer, and a one-dimensional bidirectional long short-term memory network to finally obtain the lifetime residual prediction value. Step S4-2: Determine the residuals between multiple preliminary lifetime predictions and actual values ​​from the dimensionality reduction data, input them into the residual prediction network, and obtain the lifetime residual prediction values.

7. The two-stage remaining life prediction method for aircraft hydraulic systems based on data dimensionality reduction according to claim 6, characterized in that, In step S4-1-4, the input to the convolutional layer is defined as Where n k Indicates the total number of observations. This is the θth observation of the aircraft hydraulic power system. The remaining lifetime residual at time , This indicates that the aircraft hydraulic system obtains the first predicted value for the θth time. At that moment, This represents the max-th predicted value obtained by the aircraft hydraulic system for the θth time.

8. The two-stage remaining life prediction method for aircraft hydraulic systems based on data dimensionality reduction according to claim 7, characterized in that, In step S3-2, the expression for the optimization objective is: Among them, F1 θ minF1 represents the correlation of the prediction results at the θ-th observation. θ Minimize F1 θ n θ M represents the total number of data points at the θ-th observation. θ Let θ be the number of base learners at the θ-th observation. Let q' represent the weight of the m-th basic learner at the θ-th observation, and q' represent another basic learner different from m. This represents the factorization term of the m-th basic learner at time θ-th observation. The predicted value, * = Trend,Seasonal, indicates that the decomposition item is a trend item or a periodic item. This represents the average value of the decomposition terms of the m-th basic learner at time t during the θ-th observation. This represents the factorization term of the q'-th basic learner at time θ-th observation. The predicted value, This represents the time of the j'-th sample in the initial dimensionality reduction data obtained in the θ-th iteration. This represents the average value of the decomposition terms of the q'-th basic learner at time t during the θ-th observation. This represents the estimation error of the prediction result at the θ-th observation. Minimize This indicates that the decomposition term at time θ represents the value of the term at the θth observation. The true value.

Images