An aviation gear residual service life prediction method based on self-driven independent degradation trajectory

Abstract

This application provides a method for predicting the remaining service life of aircraft gears based on self-driven independent degradation trajectories. The specific steps are as follows: First, multi-source vibration signals throughout the entire lifecycle of the aircraft gear are collected and multi-domain features are extracted and optimized to obtain a high-dimensional feature tensor. Second, a multi-modal spatiotemporal fusion coding clustering network is used to complete deep fusion and feature reconstruction, constructing a unified degradation index characterizing gear performance. Third, based on this index, kernel density estimation of distribution differences and state functions is used to adaptively identify the degradation initiation prediction point and failure threshold, completing adaptive segmentation of the degradation trajectory. Finally, a Temporal Attention Long Short-Term Memory (TALSTM) network with a hysteresis-constrained loss function is trained using the segmented data, and the trained TALSTM network is used to predict the remaining service life of the aircraft gear. This method eliminates the reliance on manual experience in degradation stage segmentation, effectively reduces the risk of hysteresis prediction, and achieves high prediction accuracy, providing technical support for intelligent predictive maintenance of aircraft gears.

Description

Technical Field

[0001] This invention relates to the field of gear remaining service life prediction, and designs a method for predicting the remaining service life of aircraft gears based on self-driven independent degradation trajectories. Background Technology

[0002] Aviation gears are core components of aviation transmission systems. Accurate assessment of their degradation status and prediction of remaining useful life (RUL) is fundamental to ensuring the safe and reliable operation of aviation equipment and formulating scientific preventative maintenance strategies. It directly determines the operational safety and maintenance economy of aviation equipment throughout its entire lifecycle. Aviation gears operate under extreme conditions of high altitude, high speed, and variable loads for extended periods, exhibiting strong individual variability in their degradation process. Failure modes and degradation trajectories display significant spatiotemporal evolution characteristics, placing stringent demands on the adaptability, accuracy, and reliability of RUL prediction technology.

[0003] Construction of degradation index (DI) is a core prerequisite for gear degradation characterization and life prediction. Among existing DI construction methods, traditional time-domain and frequency-domain statistical features have shortcomings such as insufficient state representation ability and poor noise resistance; multi-feature fusion methods rely heavily on human experience and lack adaptive optimization capabilities; while traditional deep learning reconstruction methods only aim to minimize reconstruction error, and the hidden layer features lack physical degradation prior constraints, resulting in insufficient correlation with the actual degradation process of gears, which easily leads to fluctuations and distortions in the degradation trajectory and cannot effectively characterize the performance evolution law of aerospace gears throughout their entire life cycle.

[0004] Degradation stage segmentation and failure threshold setting are key benchmarks for RUL prediction. In existing technologies, the detection of degradation initiation prediction points mostly relies on manual observation and fixed statistical methods, which are affected by individual equipment differences and operating condition fluctuations, resulting in high uncertainty in the identification results and insufficient sensitivity to degradation evolution. Failure thresholds are mostly based on empirical fixed values ​​based on experimental data, which are easily affected by environmental interference and changes in failure modes, lacking the ability to adapt to individual degradation differences, and are prone to false alarms and false negatives in life prediction.

[0005] Existing RUL prediction models are mostly developed for gears in conventional operating conditions in industries such as railways and wind power, making it difficult to directly adapt to the spatiotemporal degradation characteristics of aerospace gears under extreme operating conditions. On the one hand, existing models lack the ability to capture long-term dependencies in long-term degradation data, and their generalization performance is limited in samples with significant individual differences. On the other hand, existing models generally ignore the safety risks to aerospace equipment caused by prediction lags, failing to meet the stringent requirements of high-reliability operation and maintenance of aerospace gears, and making it difficult to achieve high-precision and high-reliability prediction of aerospace gear RUL. Summary of the Invention

[0006] The purpose of this invention is to provide a method for predicting the remaining service life of aircraft gears based on self-driven independent degradation trajectories, which can be used to predict the remaining service life of gears.

[0007] To achieve the above objectives, the specific steps are as follows:

[0008] 1) Data Acquisition and Multi-Domain Feature Extraction and Screening: Taking T as the sampling time for a single sample... To obtain a dataset with N samples, multi-channel vibration signals of aircraft gears throughout their entire lifecycle were collected at sampling intervals. Variational mode decomposition was performed on the collected vibration signals to obtain several intrinsic modal functions (IMFs). Multidimensional time-domain and frequency-domain features were extracted for each IMF component. The monotonicity (Mon), robustness (Rob), and trendiness (Tre) of all features were calculated. Based on the entropy-weighted index (EWI), the top 16 features were selected to construct a multimodal feature tensor X.

[0009] 2) Self-driven DI construction: The multimodal feature tensor X is input into the multimodal spatiotemporal fusion based on evidence clustering autoencoder (MSFECA) network to generate a unified DI sequence based on self-driven independent degradation trajectories. The MSFECA network includes a cascaded multimodal spatiotemporal fusion (MSTF) module and a variational autoencoding and deep evidence clustering (AEEC) module.

[0010] 3) Adaptive segmentation of degradation trajectory: Based on the distribution difference of Gaussian kernel density estimation and the adaptive threshold method of K-fold optimization, the degradation start prediction point (first prediction time, FPT) is determined; based on the state function method of normal state confidence, the failure threshold (FT) is adaptively determined; with FPT as the start point and FT as the end point, the independent degradation trajectory (DT) of a single aerospace gear is adaptively segmented.

[0011] 4) Prediction Model Construction: Construct a prediction model for the remaining service life of aircraft gears using a Temporal Attention Long Short-Term Memory (TALSTM) network with hysteresis constraint loss. The TALSTM network includes a cascaded Temporal Convolutional Network (TCN) module, a bidirectional linear attention long short-term memory (LALSTM) module, and a fully connected layer.

[0012] 5) Model training: Use the degenerate stage DI sequence after segmentation in step 3) to construct a training sample set, and update the parameters of the TALSTM network model with the goal of minimizing the hysteresis constraint loss function;

[0013] 6) Remaining service life prediction: The first k DI values ​​at time t are used as the input of the updated TALSTM network model to recursively predict the DI values ​​at future times. If the predicted DI value is greater than or equal to the failure threshold FT determined in step 3), the remaining service life RUL of the aircraft gear at the current time is output.

[0014] The working steps of the MSFECA network described in step 2) are as follows:

[0015] 2-1) The MSTF module takes the multimodal feature tensor X as input, decomposes X into independent feature vectors at the IMF scale, stacks multiple spatiotemporal blocks at each scale channel, and mines the spatial correlation between multiple sensors and the dynamic dependence of the feature sequence in the temporal dimension. After 1×1 convolution dimensionality reduction, average pooling and channel attention weighting, it outputs a unified spatiotemporal fusion feature S.

[0016] 2-2) The AEEC module takes the spatiotemporal fusion feature S as input, and obtains the mean μ and log-variance logσ² of the latent variable through the encoder mapping. The latent variable z is generated through the reparameterization technique. The decoder reconstructs the input feature based on the latent variable z. The joint loss of reconstruction loss, KL divergence loss, evidence loss and consistency loss is used as the optimization objective. Finally, the latent variable z is output as the unified degradation index DI.

[0017] The joint loss function of the AEEC module is:

[0018]

[0019] In the formula, To reconstruct the loss, Let KL divergence be the KL divergence. To reduce the evidence loss in order to generate clear categorical evidence for latent variables, To enhance the consistency loss of the clustering structure, , , These are hyperparameters used to balance various losses.

[0020] The steps for determining the degradation initiation prediction point (FPT) mentioned in step 3) are as follows:

[0021] 3-1) Normalize the DI sequence and take the top 20% of the data as the health baseline data to construct the baseline health distribution;

[0022] 3-2) Five-fold cross-validation is used to optimize the Gaussian kernel bandwidth. The kernel density of the DI sequence after sliding window segmentation is estimated with the optimal bandwidth, and the first-order Wasserstein distance between adjacent window distributions is calculated.

[0023] 3-3) Based on The distribution difference threshold is calculated in principle. When the Wasserstein distance between adjacent windows exceeds the threshold, the corresponding time is the degradation start prediction point FPT.

[0024] The steps for determining the failure threshold FT mentioned in step 3) are as follows:

[0025] Define the confidence function E(t) for the normal state:

[0026]

[0027] In the formula, The DI value represents the initial health status. The DI value at time t; set the lower limit of confidence for the normal state. Solve The failure threshold FT in the DI dimension is obtained as follows:

[0028]

[0029] The hysteresis constraint loss function mentioned in step 5) is:

[0030]

[0031] In the formula, N is the sample size. , This is the actual lifespan value. Predict lifetime values ​​for the model.

[0032] Furthermore, the calculation steps for the information entropy weighted index EWI mentioned in step 1) are as follows:

[0033] 1-1) Normalize Mon, Rob, and Tre for all features;

[0034] 1-2) Calculate the information entropy of each evaluation indicator. Calculate the information entropy of each evaluation indicator. , , Normalize(∙) represents the normalization operation, n is the number of features, and the weights of each indicator are calculated based on information entropy. :

[0035]

[0036] In the formula, j=1 corresponds to monotonicity Mon, j=2 corresponds to robustness Rob, and j=3 corresponds to trend Tre;

[0037] 1-3) Calculate the EWI value for each feature:

[0038]

[0039] The top 16 features with the highest EWI scores are selected from highest to lowest, and a multimodal feature tensor X is constructed.

[0040] Furthermore, the spatiotemporal block mentioned in step 2-1) includes a spatial block and a temporal block. The spatial block fuses spatial information from multiple sensors and multiple IMF components through three layers of spatial convolution, and is then input into the temporal block after 1×1 convolution processing. The temporal block is based on a temporal Transformer structure, and models long-range temporal dependencies through position encoding and multi-head self-attention mechanism. Finally, it outputs spatiotemporal fusion features with the same input dimension after 1×1 convolution and residual connection.

[0041] Furthermore, the reconstruction loss described in step 2-2) For reconstructing signals With the original signal The mean square error is calculated using the following formula:

[0042]

[0043] The KL divergence of a single sample The calculation formula is:

[0044]

[0045] The loss of evidence To minimize the category uncertainty of the latent variable z, the calculation formula is as follows:

[0046]

[0047] In the formula, S(z) represents the total evidence corresponding to the latent variable z. Uncertainty regarding unknown categories;

[0048] The consistency loss To constrain the consistency between the latent variable distribution and the cluster centers, the calculation formula is as follows:

[0049]

[0050] In the formula, Let z be the pseudo-label category with the minimum distance. The evidence value corresponding to category k. Marginal parameters to ensure category differentiation.

[0051] Furthermore, the specific steps for model training in step 5) are as follows:

[0052] 5-1) For the DI sequence between FPT and FT, segment it using a sliding window of length k and step size 1, and construct the input matrix and label Y:

[0053]

[0054]

[0055] In the formula, Let be the DI value at time i, and m be the total length of the DI sequence in the degradation phase;

[0056] 5-2) The first k rows of matrix W are used as the input of the TALSTM network model, and matrix Y is used as the output label of the model. The weight matrix and bias matrix of the TALSTM network model are updated with the goal of minimizing the hysteresis constraint loss function.

[0057] Furthermore, the specific steps for predicting the remaining useful life in step 6) are as follows:

[0058] 6-1) Use the first k DI values ​​at time t as input to the trained TALSTM network model to predict the DI value at time t+1.

[0059] 6-2) Repeat step 6-1) to recursively predict the DI value at time t+2, t+3... Stop recursive prediction when the predicted DI value is greater than or equal to the FT.

[0060] 6-3) The number of DI vectors obtained from statistical recursive prediction is used to determine the remaining service life (RUL) of the aircraft gear, which is equal to the number of predicted vectors multiplied by the sampling interval. .

[0061] Furthermore, the residual block of the TCN module described in step 4-1) includes a dilated causal convolutional layer, a weight normalization layer, a nonlinear activation layer, and a Dropout layer. The dilated causal convolution expands the receptive field while ensuring temporal causality, and the residual connection ensures effective gradient propagation.

[0062] Furthermore, the LALSTM module described in step 4-2) adopts a bidirectional stacked structure, integrates forward and backward sequence dependencies, and reduces the attention computation complexity from O(N²) to O(N) through a linear attention mechanism, where N is the length of the input sequence.

[0063] Because of the adoption of the above technical solution, the present invention has the following advantages:

[0064] 1. This invention constructs a unified degradation index (DI) for aerospace gears by using a multimodal spatiotemporal fusion evidence clustering autoencoder (MSFECA). This achieves deep fusion of multi-scale spatiotemporal features of multi-source sensor signals and unsupervised degradation state clustering. The constructed degradation index is significantly superior to traditional methods in terms of monotonicity, noise robustness, and trend characterization ability. It can accurately characterize the performance evolution process of aerospace gears throughout their entire life cycle and effectively adapt to the individual degradation differences and complex failure modes of aerospace gears under extreme working conditions.

[0065] 2. This invention proposes an adaptive identification method for the degradation initiation prediction point FPT based on Gaussian kernel distribution differences and K-fold optimization, as well as an adaptive setting method for the failure threshold FT based on state functions. This completely eliminates the dependence on manual experience and fixed thresholds for degradation stage division, and realizes reference-free, adaptive segmentation of the independent degradation trajectory of a single aerospace gear. It provides an objective and reliable time benchmark for high-precision life prediction, and greatly improves the scientific nature and engineering practicality of degradation assessment.

[0066] 3. This invention constructs a temporal attention long short-term memory (TALSTM) network with hysteresis constraint loss. It extracts multi-scale local features through a temporal convolutional network and models long-term temporal dependencies through a bidirectional linear attention long short-term memory network, effectively improving the learning ability and prediction accuracy of the evolution law of degenerate trajectories. At the same time, through the hysteresis constraint loss function with asymmetric penalty, it significantly suppresses the safety risks of aviation equipment caused by prediction hysteresis and meets the stringent requirements of high reliability operation and maintenance of aviation gears.

[0067] 4. This invention forms an integrated framework for self-driven construction of degradation indicators, adaptive segmentation of degradation trajectories, and high-precision prediction of remaining service life. The entire process requires no prior manual intervention and can construct independent degradation trajectories that match the service performance evolution of aviation gears under different working conditions and with different individual differences, providing effective technical support for intelligent predictive maintenance of aviation equipment.

[0068] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained from the following description and claims. Attached Figure Description

[0069] The accompanying drawings of this invention are as follows.

[0070] Figure 1 This is a flowchart of the method of the present invention.

[0071] Figure 2 A diagram illustrating the multimodal spatiotemporal fusion DI construction method based on evidence clustering autoencoder networks.

[0072] Figure 3 This is a structural framework diagram of TALSTM.

[0073] Figure 4 This is a flowchart for RUL prediction.

[0074] Figure 5 The diagram shows the structural results of the experimental gear DI.

[0075] Figure 6 The figure shows the experimental results for RUL prediction. Detailed Implementation

[0076] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0077] like Figure 1 The method for predicting the remaining service life of aircraft gears based on self-driven independent degradation trajectories, as shown below, includes the following specific steps:

[0078] 1) Vibration Signal Data Acquisition: A self-built aviation gear fatigue life test bench was used to conduct full life cycle fatigue tests on aviation gears to collect vibration signal data throughout the entire service life of the aviation gears. The test bench consists of a gear fatigue test platform, a lubrication and temperature control system, a drive control system, and a vibration signal acquisition system. The vibration signal acquisition system includes four PCB single-axis accelerometers, an NI-9234 acquisition card, a cDAQ-9188 chassis, and supporting data acquisition software. The four accelerometers are respectively arranged in the horizontal and vertical directions at the drive end and load end of the gearbox to realize multi-channel, full-dimensional vibration signal acquisition of aviation gears.

[0079] The key parameters of the aerospace gear used are: modulus 0.5, number of teeth 60, pressure angle 20°, and material 17Cr, meeting the lightweight and high-strength design requirements of aerospace gears. During data acquisition, continuous time-series sampling was performed at 1-minute intervals (0.1-second sampling per sample, 51.2 kHz sampling frequency). Each sampling point corresponds to a data file containing 5120 data points (51.2 kHz × 0.1 s), for a total of 6 sets of full-lifecycle gear data collected (detailed parameters are shown in Table 1) to fully verify the adaptability of the proposed method under different operating conditions.

[0080] Table 1 Test Condition Settings

[0081] 2) Construction of a unified degradation index (DI) for aircraft gears: Multi-domain feature extraction and screening are performed on the collected vibration signals. A unified degradation index (DI) capable of characterizing the degradation state of aircraft gears is constructed using a multimodal spatiotemporal fusion evidence clustering autoencoder (MSFECA). The specific method is as follows:

[0082] 2-1) Multi-domain feature extraction and screening: Variational mode decomposition is used to decompose the original vibration signal collected in each group into 9 intrinsic mode function (IMF) components; for each IMF component, 15 time-domain features and 13 frequency-domain features are extracted to construct a high-dimensional initial feature matrix.

[0083] For all extracted features, calculate the three evaluation metrics for each feature: monotonicity (Mon), robustness (Rob), and trend (Tre). The calculation formula is as follows:

[0084]

[0085]

[0086]

[0087] In the formula, , These are the feature values ​​and their corresponding time series, respectively, where N is the total length of the sequence. Let i be the eigenvalue at time i. for The exponentially smoothed eigenvalue at time t; , These are the mean values ​​of the feature value sequence and the time series, respectively; all indicators are normalized to the [0,1] interval, and the closer the value is to 1, the better the degradation performance of the feature.

[0088] Based on information entropy weights, an entropy-weighted comprehensive index EWI is constructed, which integrates the three evaluation indicators. The calculation formula is as follows:

[0089]

[0090]

[0091] In the formula, The information entropy weights are the three evaluation indicators. Let be the information entropy of the j-th evaluation index.

[0092] For each IMF component, the top D=16 features with the highest EWI scores are selected to construct the feature vector at that IMF scale. Where M is the number of sensor channels; finally, a multimodal feature tensor sample is constructed. Where I is the total number of IMF components, and in this embodiment I=9.

[0093] 2-2) Multimodal spatiotemporal feature fusion: The multimodal feature tensor X is input into the multimodal spatiotemporal fusion MSTF module. The MSTF module contains multiple stacked spatiotemporal blocks, convolutional layers, average pooling layers, fully connected layers and sigmoid activation functions; each spatiotemporal block is composed of stacked spatial blocks and temporal blocks, and residual connections are set.

[0094] The spatial block, through three layers of spatial convolution, fuses spatial correlation information from multiple sensors and multimodal IMF components, and is then input into the temporal block after 1×1 convolution. The temporal block, based on a temporal Transformer structure, models long-term temporal dependencies through positional encoding and multi-head self-attention mechanisms. After fusion via 1×1 convolution and residual connections, the output spatiotemporal fusion feature has the same dimension as the input. The MSTF module decomposes the input feature tensor X into independent feature vectors at nine IMF scales. Each scale channel stacks multiple spatiotemporal blocks, simultaneously mining the spatial correlation between multiple sensors and the temporal evolution of feature sequences at each scale. After dimensionality reduction via 1×1 convolution and extraction of key information via average pooling, the outputs of each channel are weighted by channel attention through fully connected layers and a sigmoid activation function, and finally fused to obtain a unified spatiotemporal fusion feature tensor S.

[0095] 2-3) DI generation based on evidence clustering: The spatiotemporal fusion feature tensor S is input into the variational autoencoder and deep evidence clustering AEEC module. Through the pipeline of "encoder-latent space-decoder-evidence clustering", the unified degradation index DI of aerospace gears is obtained.

[0096] The encoder extracts a high-dimensional representation of S through multiple convolutional layers, and outputs the mean of the latent variables through a fully connected layer. With log variance Generate latent variables through reparameterization techniques ,in The noise is standard normal; the decoder reconstructs the original spatiotemporal fusion features from the latent variable z through a transposed convolutional layer, using the mean square error between the reconstructed signal and the original input S as the reconstruction loss. The calculation formula is:

[0097]

[0098] In the formula, S represents the spatiotemporal fusion feature reconstructed by the decoder.

[0099] Introducing KL divergence as a regularization term constrains the posterior distribution of latent variables to approximate the prior distribution, resulting in KL divergence loss. The calculation formula is:

[0100]

[0101] In the formula, d represents the dimension of the latent variable.

[0102] The evidence clustering module flattens the latent variable z into a (1,2)-dimensional vector using a 1×1 convolution kernel, corresponding to the "normal" and "abnormal" service states of the aircraft gears, respectively. The evidence parameters are then obtained by mapping using the softplus activation function. Calculate the confidence level of the sample for each cluster based on evidence theory. Uncertainty related to the unknown The calculation formula is:

[0103]

[0104]

[0105] In the formula, This represents the total amount of evidence.

[0106] Constructing evidence loss With consistency loss Evidence loss, by minimizing overall uncertainty, guides latent variables to generate explicit categorical evidence; consistency loss, by minimizing inter-class evidence differences, strengthens the clustering structure. The calculation formula is as follows:

[0107]

[0108]

[0109] In the formula, The pseudo-labels are obtained by minimizing the Euclidean distance between the latent variables and the cluster centers. To ensure the marginal parameters of class discrimination, cluster centers are initialized using the K-Means algorithm and iteratively optimized during training.

[0110] Total loss function of the AEEC module for:

[0111]

[0112] In the formula, , , Hyperparameters are used to balance the weights of each loss term.

[0113] Using the feature tensors of Gear1-1 and Gear2-1 as training sets, the total loss function L was optimized until the model converged. The feature tensors of the six datasets were input into the trained AEEC module, and the output latent variable z was extracted as the unified degradation index (DI) of the corresponding aviation gear, resulting in six full lifecycle DI sequences. The original DI sequences were smoothed using a sliding window averaging method with a window length of 30 and a step size of 1 to eliminate noise interference.

[0114] 3) Adaptive Segmentation of Degradation Trajectory for Aircraft Gears: Based on the constructed DI sequence, the degradation initiation prediction point FPT and the failure threshold FT are adaptively identified to complete the autonomous segmentation of the independent degradation trajectory of a single aircraft gear. The specific method is as follows:

[0115] 3-1) Adaptive identification of degradation initiation prediction point (FPT):

[0116] DI Sequence Preprocessing and Benchmark Health Distribution Construction: For Full Life Cycle DI Sequences (Late variables output for AEEC) are normalized to eliminate differences in units and values ​​using φ. The first 20% of the normalized data are used as baseline health data, and their distribution is as follows: This represents the steady-state characteristics before degradation.

[0117] Kernel bandwidth optimization: The Gaussian kernel bandwidth *l* determines the accuracy of kernel density estimation. Using K=5 cross-validation, the DI data is divided into 5 mutually exclusive subsets; each subset is used sequentially as the validation set, and the remaining subsets are used as the training set. For the bandwidth set... Calculate the log-likelihood loss on the validation set:

[0118]

[0119] in, It is the size of the k-th validation set. This is a kernel density estimate over bandwidth l. The optimal bandwidth minimizes the average log-likelihood loss.

[0120]

[0121] Partial distribution kernel density estimation of DI: A sliding window (size W=30, stride 1) is used to segment the normalized sequence; the t-th window is... Optimal bandwidth Gaussian kernel density estimation fitting The probability density function in:

[0122]

[0123] In the formula For Gaussian kernel function, for Normalized DI in the middle.

[0124] Wasserstein distance metric: First-order Wasserstein distance quantization and Differences in distribution between them:

[0125]

[0126] in, yes The cumulative distribution function captures distribution changes without requiring overlap.

[0127] FPT determination: Extraction of distribution difference sequence of healthy phase Calculate its mean and standard deviation . use The principle of standard deviation and the threshold are ,when t is FPT.

[0128] 3-2) Failure threshold FT adaptive setting

[0129] A confidence function E(t) for the normal state of an aircraft gear is constructed to characterize the confidence that the gear is in normal service condition at time t. The calculation formula is as follows:

[0130]

[0131] In the formula, The DI value represents the initial healthy state of the gear. Furthermore, its value decreases monotonically as the gears degenerate.

[0132] Set the lower confidence limit for the normal state. In this embodiment, the lower confidence limit of condition A is... The lower confidence limit for condition B ;when When the gear is determined to have entered a failure state, the failure threshold FT in the DI dimension is obtained by inverse kinematics:

[0133]

[0134] Degeneracy trajectory construction: Selecting the corresponding FPT values The DI sequence constitutes the independent degradation trajectory DT of the aircraft gear; the degradation stage segmentation results of the 6 test gears are shown in Table 2.

[0135] Table 2. Segmentation results of degradation stages for each test gear

[0136]

[0137] 4) Constructing a RUL prediction model for aerospace gears with lag constraints: A RUL prediction model based on a temporal attention long short-term memory (TALSTM) network is constructed, and a lag constraint loss function is designed to suppress prediction lag risk. Specifically:

[0138] The TALSTM network comprises a TCN module, a bidirectional linear attention long short-term memory (LALSTM) module, and a fully connected layer. The TCN module consists of two residual blocks, each containing dilated causal convolution, weight normalization, a nonlinear activation layer, and a dropout layer. The dilated causal convolution expands the receptive field while maintaining temporal causality, extracting multi-scale local temporal features of the DI sequence. The LALSTM module combines LSTM, linear attention, and a bidirectional stacked structure. While retaining the long-range information storage capacity of LSTM, it reduces the computational complexity of attention from O(N²) to O(N) through linear attention. The bidirectional structure fuses forward and backward sequence dependencies to extract long-range global degradation features of the DI sequence. The fully connected layer fuses the local features of the TCN module and the global features of the LALSTM module, and after linear transformation and nonlinear activation, outputs the prediction result of the DI sequence.

[0139] The design incorporates a lag-constrained loss function, employing asymmetric penalty weights for prediction lag and lead to suppress high-risk prediction lag issues. The calculation formula is as follows:

[0140]

[0141] In the formula, N is the sample size. , This is the actual lifespan value. Predict lifetime values ​​for the model. If the time frame indicates a forecast lag, a higher penalty weight should be applied. If the time indicates an over-prediction, a lower penalty weight is set.

[0142] 5) Prediction model training and DI sequence recursive prediction:

[0143] 5-1) Construction of the prediction dataset: For the degradation trajectory DT of each aircraft gear, a sliding window with a window length k=30 and a step size of 1 is used to segment the data, constructing the model input set W and the supervision label Y:

[0144]

[0145] 5-2) Model Training: The input sets W of Gear1-1 and Gear2-1 are used as training inputs, and the corresponding supervision labels Y are used as training outputs to train the TALSTM network. During training, the learning rate is set to 0.001, the training epochs are 100, and the Adam optimizer is used. The weight parameters and bias parameters of the TALSTM network are iteratively updated with the goal of minimizing the hysteresis constraint loss function.

[0146] 5-3) Recursive DI Sequence Prediction: Using the 30 consecutive DI values ​​before the current time step as input to the trained TALSTM network, predict the DI value at the next time step; add the predicted DI values ​​to the input sequence, and recursively predict the DI sequence at future time steps. The step-by-step prediction steps are as follows:

[0147]

[0148]

[0149] …

[0150]

[0151] In the formula, f is the mapping function of the trained TALSTM network. Let be the predicted DI value at time t+i.

[0152] 6) Prediction of remaining service life of aircraft gears: Continuously and recursively predict the DI value at future moments. When the predicted DI value is greater than or equal to the failure threshold FT of the gear, the gear is determined to have reached a failure state; at the current moment... The corresponding Remaining Useful Life (RUL) calculation formula is as follows:

[0153]

[0154] In the formula, r is the time interval from the current time to the predicted failure time, i.e., the remaining service life.

[0155] To verify the feasibility and superiority of the method of the present invention, the degradation trajectories of 6 sets of gears were selected for RUL prediction verification. The DI sequence from FPT to the current time was used as training data, and the DI sequence from the current time to FT corresponding to the actual RUL was used as test label; finally, the RUL prediction result was generated.

[0156] The prediction results show that the RUL prediction error rate of the method of the present invention for the six sets of gears is less than 30%, with the prediction error rates of Gear1-1 and Gear2-1 as low as 2.94% and -2.78% respectively. The comprehensive prediction score reaches 0.5974, which is significantly better than mainstream time series prediction models such as 1DCNN, TCN, BiLSTM, and BiGRU. The specific comparison results are shown in Table 3.

[0157] Table 3 Comparison of Degradation Trajectory Construction and Prediction Results

[0158] The results show that the method proposed in this invention can construct independent degradation trajectories that match the service performance evolution of aircraft gears under different working conditions and for different individuals, realize referenceless adaptive identification of FPT and FT, and achieve high-precision and high-reliability RUL prediction through the TALSTM network with hysteresis constraints. Compared with the existing mainstream models, it has better prediction performance and engineering application value.

[0159] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for predicting the remaining life of aircraft gears based on self-driven independent degradation trajectories, characterized in that, The specific steps are as follows: 1) Data Acquisition and Multi-Domain Feature Extraction and Screening: Using T as the sampling time for a single sample... To obtain a dataset with N samples, multi-channel vibration signals of aircraft gears throughout their entire lifecycle were collected at sampling intervals. Variational mode decomposition was performed on the collected vibration signals to obtain several intrinsic modal functions (IMFs). Multidimensional time-domain and frequency-domain features were extracted for each IMF component. The monotonicity (Mon), robustness (Rob), and trendiness (Tre) of all features were calculated. Based on the entropy-weighted index (EWI), the top 16 features were selected to construct a multimodal feature tensor X. 2) Construction of self-driven degradation index (DI): The multimodal feature tensor X is input into the multimodal spatiotemporal fusion based onevidence clustering autoencoder (MSFECA) network to generate a unified degradation index (DI) sequence based on self-driven independent degradation trajectories. The MSFECA network includes a cascaded multimodal spatiotemporal fusion (MSTF) module and a variational autoencoding and deep evidence clustering (AEEC) module. 3) Adaptive segmentation of degradation trajectory: Based on the adaptive threshold method of Gaussian kernel density estimation distribution difference and K-fold optimization, the degradation start prediction point (first prediction time, FPT) is determined; based on the state function method of normal state confidence, the failure threshold (FT) is adaptively determined; with FPT as the start point and FT as the end point, the independent degradation trajectory (DT) of a single aerospace gear is adaptively segmented. 4) Prediction Model Construction: Construct a prediction model for the remaining service life of aircraft gears using a temporal attention long short-term memory (TALSTM) network with hysteresis constraint loss. The TALSTM network includes a cascaded time convolutional network (TCN) module, a bidirectional linear attention long short-term memory (LALSTM) module, and a fully connected layer. 5) Model training: Use the degenerate stage DI sequence after segmentation in step 3) to construct a training sample set, and update the parameters of the TALSTM network model with the goal of minimizing the hysteresis constraint loss function; 6) Remaining useful life prediction: The first k DI values ​​at time t are used as the input of the updated TALSTM network model to recursively predict the DI values ​​at future times. If the predicted DI value is greater than or equal to the FT determined in step 3), the remaining useful life (RUL) of the aircraft gear at the current time is output. The working steps of the MSFECA network described in step 2) are as follows: 2-1) The MSTF module takes the multimodal feature tensor X as input, decomposes X into independent feature vectors at the IMF scale, stacks multiple spatiotemporal blocks at each scale channel, and mines the spatial correlation between multiple sensors and the dynamic dependence of the feature sequence in the temporal dimension. After 1×1 convolution dimensionality reduction, average pooling and channel attention weighting, it outputs a unified spatiotemporal fusion feature S. 2-2) The AEEC module takes the spatiotemporal fusion feature S as input, and obtains the mean μ and log-variance logσ² of the latent variable through the encoder mapping. The latent variable z is generated through the reparameterization technique. The decoder reconstructs the input feature based on the latent variable z. The joint loss of reconstruction loss, KL divergence loss, evidence loss and consistency loss is used as the optimization objective. Finally, the latent variable z is output as the unified degradation index DI. The joint loss function of the AEEC module is: In the formula, To reconstruct the loss, Let KL divergence be the KL divergence. To reduce the evidence loss in order to generate clear categorical evidence for latent variables, To enhance the consistency loss of the clustering structure, , , These are hyperparameters used to balance various losses; The steps for determining the FPT mentioned in step 3) are as follows: 3-1) Normalize the DI sequence and take the top 20% of the data as the health baseline data to construct the baseline health distribution; 3-2) Five-fold cross-validation is used to optimize the Gaussian kernel bandwidth. The kernel density of the DI sequence after sliding window segmentation is estimated with the optimal bandwidth, and the first-order Wasserstein distance between adjacent window distributions is calculated. 3-3) Based on The principle is to calculate the distribution difference threshold. When the Wasserstein distance between adjacent windows exceeds the threshold, the corresponding time is the FPT. The steps for determining the FT mentioned in step 3) are as follows: Define the normal state confidence function E(t): In the formula, The DI value represents the initial health status. Let be the DI value at time t; Set the lower confidence limit for the normal state. Solve The failure threshold FT of the DI dimension is obtained as follows: The working steps of the TALSTM network described in step 4) are as follows: 4-1) The TCN module takes the segmented DI sequence as input, extracts multi-scale local temporal features through residual blocks with dilated causal convolution, and outputs intermediate feature tensors; 4-2) The LALSTM module takes the output of the TCN module as input, models bidirectional temporal dependencies through a bidirectional long short-term memory network LSTM, reduces computational complexity through a linear attention mechanism, and extracts global temporal features; 4-3) The fully connected layer fuses the local features of the TCN module and the global features of the LALSTM module, and outputs the DI prediction value after linear transformation and nonlinear activation; The hysteresis constraint loss function mentioned in step 5) is: In the formula, N is the sample size. , This is the actual lifespan value. Predict lifetime values ​​for the model.

2. The method for predicting the remaining service life of aircraft gears based on self-driven independent degradation trajectories as described in claim 1, characterized in that, The calculation steps for EWI described in step 1) are as follows: 1-1) Normalize Mon, Rob, and Tre for all features; 1-2) Calculate the information entropy of each evaluation indicator. , , Normalize(∙) represents the normalization operation, n is the number of features, and the weights of each indicator are calculated based on information entropy. : In the formula, j=1 corresponds to Mon, j=2 corresponds to Rob, and j=3 corresponds to Tre; 1-3) Calculate the EWI value for each feature: The top 16 features with the highest EWI scores are selected from highest to lowest, and a multimodal feature tensor X is constructed.

3. The method for predicting the remaining service life of aircraft gears based on self-driven independent degradation trajectories as described in claim 1, characterized in that, The spatiotemporal block mentioned in step 2-1) includes a spatial block and a temporal block. The spatial block fuses spatial information from multiple sensors and multiple IMF components through three layers of spatial convolution, and is then input into the temporal block after 1×1 convolution. The temporal block is based on a temporal Transformer structure, and models long-range temporal dependencies through position encoding and multi-head self-attention mechanism. Finally, it outputs a spatiotemporal fusion feature S with the same dimension as the input through 1×1 convolution and residual connection.

4. The method for predicting the remaining service life of aircraft gears based on self-driven independent degradation trajectories as described in claim 1, characterized in that, The reconstruction loss described in step 2-2) For reconstructing signals With the original signal The mean square error is calculated using the following formula: The KL divergence of a single sample The calculation formula is: The loss of evidence To minimize the category uncertainty of the latent variable z, the calculation formula is as follows: In the formula, S(z) represents the total evidence corresponding to the latent variable z. Uncertainty regarding unknown categories; The consistency loss To constrain the consistency between the latent variable distribution and the cluster centers, the calculation formula is as follows: In the formula, Let z be the pseudo-label category with the minimum distance. The evidence value corresponding to category k. Marginal parameters to ensure category differentiation.

5. The method for predicting the remaining service life of aircraft gears based on self-driven independent degradation trajectories as described in claim 1, characterized in that, The specific steps for model training in step 5) are as follows: 5-1) For the DI sequence between FPT and FT, segment it using a sliding window of length k and step size 1, and construct the input matrix W and label Y: In the formula, Let be the DI value at time i, and m be the total length of the DI sequence in the degradation phase; 5-2) The first k rows of matrix W are used as the input of the TALSTM network model, and matrix Y is used as the output label of the model. The weight matrix and bias matrix of the TALSTM network model are updated with the goal of minimizing the hysteresis constraint loss function.

6. The method for predicting the remaining service life of aircraft gears based on self-driven independent degradation trajectories as described in claim 1, characterized in that, The specific steps for predicting the remaining useful life in step 6) are as follows: 6-1) Use the first k DI values ​​at time t as input to the trained TALSTM network model to predict the DI value at time t+1. 6-2) Repeat step 6-1) to recursively predict the DI value at time t+2, t+3... Stop recursive prediction when the predicted DI value is greater than or equal to the FT. 6-3) The number of DI vectors obtained from statistical recursive prediction is used to determine the remaining service life (RUL) of the aircraft gear, which is equal to the number of predicted vectors multiplied by the sampling interval. .

7. The method for predicting the remaining service life of aircraft gears based on self-driven independent degradation trajectories as described in claim 1, characterized in that, The residual block of the TCN module described in step 4-1) includes a dilated causal convolutional layer, a weight normalization layer, a nonlinear activation layer, and a Dropout layer. The dilated causal convolution expands the receptive field while ensuring temporal causality, and the residual connection ensures effective gradient propagation.

8. The method for predicting the remaining service life of aircraft gears based on self-driven independent degradation trajectories as described in claim 1, characterized in that, The LALSTM module described in step 4-2) adopts a bidirectional stacked structure, integrates forward and backward sequence dependencies, and reduces the attention computation complexity from O(N²) to O(N) through a linear attention mechanism, where N is the length of the input sequence.

Images