Folding wing health state prediction method based on migratory trend grouping self-correlation contrast learning

Abstract

The present application relates to the technical field of civil aircraft maintenance and repair, in particular to a flap health state prediction method based on migratory trend grouping self-correlation contrast learning, first, typical degradation trends are mined from the full data using a clustering algorithm, and individual samples with differentiation are divided into several high-consistency trend data groups; second, a self-correlation contrast mechanism is introduced within each group to capture the common time sequence dependence between samples across long periods, and multiple pre-training models that can represent specific degradation patterns are constructed; then, the most suitable trend category and pre-training model are selected by dynamically matching the characteristics of the flap to be tested; finally, based on the transfer fine-tuning strategy, the pre-training model is fine-tuned with a small amount of data, thereby efficiently obtaining the exclusive prediction model of the flap. Compared with the prior art, it is not necessary to build a model from scratch for each aircraft, and migration can be performed based on a limited model library to effectively manage the health state of a civil aircraft.

Description

Technical Field

[0001] This invention relates to the field of civil aircraft maintenance and repair technology, specifically a flap health status prediction method based on transferable trend grouping autocorrelation contrastive learning that eliminates the need to build a model from scratch for each aircraft and enables transfer based on a limited model library to achieve effective management of the health status of civil aircraft. Background Technology

[0002] The flap system is a critical actuation system for civil aircraft during takeoff and landing, directly impacting flight safety. Takeoff and landing are widely recognized as high-risk phases for accidents; statistics show that takeoff and landing-related flight phases consistently account for approximately 60% of all commercial aviation accidents. Therefore, conducting research on the health status prediction of flap systems is of great significance for improving operational reliability.

[0003] One of the key challenges in predicting the health status of civil aircraft flap systems is that traditional models, limited by local observation windows, struggle to accurately capture long-distance time dependence and actual degradation patterns. On one hand, the signal representing flap physical degradation exhibits extremely slow and weak evolution characteristics; on the other hand, this signal is often superimposed with high-amplitude dynamic noise driven by factors such as seasonal changes and operational condition variations. For traditional models relying on sliding windows, the intensity of environmental noise within such local time segments far exceeds that of the degradation signal, making it difficult to capture global patterns spanning thousands of flights. This results in the model's inability to distinguish between short-term environmental fluctuations and long-term performance degradation.

[0004] Figure 1 This visually illustrates the modeling dilemma caused by window limitations. Certain segments of an aircraft flap deployment time series exhibit significant annual periodicity, with time windows approximately one year apart. and Although geographically distant on the timeline, their physical semantics and signal morphology are highly similar due to their similar seasonal operating ranges; conversely, windows that are closer in distance... and However, their signal characteristics differ significantly due to system state switching. This severe misalignment between "temporal distance" and "physical semantic similarity" constitutes a perceptual bottleneck for traditional models: due to the lack of a correlation mechanism spanning long-distance windows, the model often incorrectly maps temporally proximate but heterogeneous segments (such as...) when mapping high-dimensional representations. and They are categorized as related, but cannot identify evolutionary patterns that are geographically distant but homogeneous (such as...). and ).

[0005] This perceptual limitation is prevalent in current mainstream deep learning models. Transformer-based architectures, such as Autoformer, still confine their autocorrelation mechanisms to a limited lookback window, failing to perceive distant events outside the window and thus unable to understand the consistency of operational patterns across years. Informer improves computational efficiency through its ProbSparse attention mechanism, but when dealing with slowly varying degradation signals like flap systems, its sparse sampling process easily loses weak but globally crucial degradation features, leading to inconsistent representations. While TimesNet attempts to capture multi-scale structures through time-frequency domain transformation, its pre-defined fixed-period basis is ill-suited to the non-strictly periodic, ultra-long-range environmental dependencies of landing gear operation, failing to establish long-range semantic relationships between windows. Even DLinear, known for its simplicity, is limited to fixed-length inputs in its linear decomposition operation, extracting only approximations of local segments and completely stripping away macroscopic evolutionary information about the entire system's lifecycle. In summary, existing models cannot fundamentally overcome the physical limitations of sliding windows, making it difficult to fully explore and utilize long-distance similarity patterns in time series to drive high-precision trend prediction. Summary of the Invention

[0006] This invention addresses the shortcomings and deficiencies of existing technologies by introducing an autocorrelation-based contrastive network (AutoConNet) architecture with a global perspective at the basic feature extraction level. Utilizing an autocorrelation-based contrastive learning strategy, it provides the model with a global supervision signal that transcends the sliding window, enabling it to capture long-term evolutionary patterns across thousands of flight sorties. Furthermore, it integrates a conditional variational patch embedding (CVPE) module, mapping future flight plans to a dynamic environmental context. A joint modeling mechanism directly establishes a deep semantic association between historical degradation patterns and similar future operating conditions, thereby alleviating the prediction challenge of the slow degradation and strong periodic coupling of the flap system.

[0007] This invention achieves its purpose through the following measures:

[0008] A method for predicting the health status of flaps based on transferable trend grouping autocorrelation contrastive learning is characterized by the following steps: First, a clustering algorithm is used to mine typical degradation trends from the full dataset, dividing differentiated individual samples into several highly consistent trend data groups. Second, an autocorrelation contrastive mechanism is introduced within each group to capture common temporal dependencies between samples across long periods, constructing multiple pre-trained models that can characterize specific degradation patterns. Then, dynamic matching is performed based on the characteristics of the flap to be tested to select the most suitable trend category and pre-trained model. Finally, based on a transfer fine-tuning strategy, the pre-trained model is fine-tuned using a small amount of data from the flap, thereby efficiently obtaining a specific prediction model for the flap.

[0009] This invention includes the following steps:

[0010] Step 1: Structured clustering of the full dataset based on differentiated degradation features: Based on the degradation features of the full historical data, a clustering algorithm is used to identify and separate several representative typical degradation modes, including category 1 to category 2. To construct highly consistent trend data grouping: First, the operational degradation characteristics of each aircraft need to be quantitatively represented in low dimension. Principal component analysis is used to project the high-dimensional time series data into a low-dimensional feature space. The extracted principal components capture the direction of variation with the largest variance in the overall data, which represents the macroscopic differences in operational load and environmental experience among different aircraft. Subsequently, the K-Means algorithm is used to process the projected feature vector set. Cluster analysis was performed, and the source domain data was structured into... Categories of different operating modes Each category The flap system deployment and retraction time series of the aircraft exhibited highly similar macroscopic dynamic characteristics, indicating that they experienced similar operating environments and may follow similar performance evolution trajectories.

[0011] Step 2: Grouped Dedicated Model Training for Various Trends: Within each trend group, data preprocessing and sliding window sampling are performed separately. A combined prediction network based on Autocorrelation Contrast Learning (AutoConNet) and Conditional Variational Position Encoding (CVPE) is trained independently to capture common evolutionary patterns across various trends. This results in the construction of a pre-trained model library containing multiple expert models, including Model 1 to Model 2. ;

[0012] Step 3: Differentiated directional fine-tuning of the trend-matching bi-branch decomposition architecture: For the flap to be tested, firstly, match the most similar trend category based on its historical data characteristics, and denote it as the category. Then, the test data is used to train the corresponding pre-trained model, i.e., the model. Differentiated and targeted fine-tuning of the AutoConNet dual-branch decomposition architecture is performed to efficiently obtain a dedicated prediction network for the individual's flaps;

[0013] Step 4: Predict individual differences in flap trends based on a dedicated prediction network. Use the generated dedicated network to extrapolate the future health status of the flap under test and output the final performance degradation trend prediction results.

[0014] The AutoConNet described in step 2 of this invention consists of data input and preprocessing, a dual-branch decomposition architecture, and an autocorrelation contrastive learning mechanism. The data input and preprocessing includes:

[0015] (1) Global sequence definition and sliding window sampling:

[0016] The network input originates from a global perspective of flap health indicator time series. The complete historical operational data of an aircraft flap system is formally defined as a multivariate time series. ,in The time step representing the total number of flights within the entire observation period, and the observed value at each moment. It is a vector containing 𝑐 dimensions, which is composed of the core health indicator, namely flap actuation time, and several key influencing factors, namely flight altitude, pitch angle, and ambient temperature.

[0017] Using the sliding window method to analyze the time series of flap health indicators Sampling is performed using a fixed-length window in the sequence. Swipe up, and each window is divided into a historical observation sequence, i.e., input. The output is the future sequence to be predicted. , No. Sample Represented as an input-output pair ,in: , Represents historical information used for prediction; , This represents the future health indicators that need to be predicted, through the complete sequence. Slide the slider with a step size of 1, and finally generate a sequence containing... Training set of samples ;

[0018] (2) Window normalization and feature splitting:

[0019] For any input sample The data first enters the window normalization module, which calculates the mean within the current window. and standard deviation Transform the original sequence into normalized features. The characteristics of future flight plans are defined as follows It will be with The data are then fused together in the CVPE module. Finally, the total predicted value of the model should undergo window inverse normalization and be superimposed with the prediction results of the two parallel branches. (2).

[0020] The AutoConNet described in this invention employs a dual-branch parallel decomposition architecture to decouple the original time-series data into two orthogonal components: The short-time branch focuses on capturing high-frequency micro-dynamics in the data, i.e., removing residual random disturbances after removing trends and periods, and consists of a lightweight linear layer.

[0021] (3);

[0022] The long-term branch employs a deep encoder-decoder architecture, with global supervision by the AutoCon loss mechanism. In the encoder part, the received data is processed by the CVPE module and fused with future environmental information. Features Temporal Convolutional Network (TCN) is used to map the latent space representations to a series of different scales. :

[0023] (4), of which, It is a global context representation containing temporal features, used to assist the encoder in capturing temporal correlations;

[0024] The decoder employs a multi-scale moving average module, utilizing a set of pooling kernels of different sizes to perform multi-granularity aggregation of temporal features:

[0025] (5),

[0026] Represents pooling kernels of different scales. For the high-dimensional representation of the encoder output, Representing the scale quantity, by... By weighting and fusing the moving average results at various scales, the model can adaptively reconstruct long-term trends that include multiple periodic components.

[0027] The AutoConNet invention introduces an AutoCon module at the bottom, which uses autocorrelation calculation to provide a global calibration signal for the encoder, including:

[0028] (1) Prior calculation driven by autocorrelation: The model first calculates the global sequence Perform a Fast Fourier Transform (FFT) to compute its global autocorrelation function. ,in, Indicates time delay. Represents the mathematical expectation. For time series in The value at time is the time series after a delay. Subsequent values, The mean of the time series. The variance of the time series.

[0029] (6) This function reveals the signal at different lag times The inherent similarity;

[0030] (2) Similar window retrieval and comparative learning: The AutoCon module utilizes sets As an index, historical windows with high autocorrelation to the current window are dynamically retrieved as positive samples during training. To measure the similarity between two windows, two windows are defined. and Theoretical similarity between The absolute value of the global autocorrelation coefficient corresponding to the time interval:

[0031] (7),

[0032] in This represents the global autocorrelation function, while Represents the entire time series In time interval The global autocorrelation coefficient under the given conditions, therefore This represents the theoretical similarity between two time windows derived from the global autocorrelation function; this quantity reflects the lag. Does the sequence as a whole exhibit a fixed periodic pattern? (Based on autocorrelation-based contrastive loss.) As shown in formula (8):

[0033] (8), among which, This indicates the number of windows in a single input model. , and These are all indices for the time window, representing the current window, the candidate positive sample window, and the negative sample window, respectively. Indicates the first The first window and the first Global autocorrelation coefficient between windows Indicates the first The first window and the first Window feature representation vector , Similarity between them Temperature is a parameter used to control the smoothness of the similarity distribution. As an indicator function, when the window With window Distance between Less than or equal to window With window Distance between The value is 1 if it is true, and 0 otherwise.

[0034] The objective function is reconstructed as follows:

[0035] (9), among which, For example, the mean squared error (MSE) is used to measure the error in predicting future release and take-off times. Balance the weights of the two.

[0036] This invention introduces a Conditional Variational Positional Encoding Fusion (CVPE) module before the long-term branch of AutoConNet. The CVPE module receives two inputs: historical data after window normalization. and condition information representing future operating conditions. ,in These represent the sample size for the input window and future flight information, respectively, i.e., the time step; specifically, using past data... Predicting the future from individual sample information One target value, here. Includes the past The number of samples per flight Includes the future Information on each flight; and These represent the number of feature parameters contained in historical data and the number of feature parameters contained in future flight information, respectively, and specifically include the following:

[0037] (1) Multi-source patch embedding and semantic alignment:

[0038] First, open the input window The future operating condition window is divided into two parallel segments. and There are non-overlapping time patches, among which , The lengths of each patch, which is used to divide historical information and future flight information, are respectively. and These represent the number of patches into which each component is divided, and the number of patches into which each component is divided. Each historical data patch is denoted as , No. Each future flight data patch is denoted as Then, each corresponding patch is flattened into a vector using the flatten function. and Due to historical information ( (Wei) and future flight information ( Since each patch has a different feature dimension, independent linear projection layers are used to map the flattened patch to the same feature dimension. A dimensional embedding space is used to achieve semantic alignment, specifically:

[0039] (10)

[0040] (11), among which, , , , The vectors obtained from the two projections are concatenated along the patch dimension in chronological order to form a patch sequence that conforms to a time series relationship.

[0041] (12)

[0042] in, If batch size is taken into account The shape of the overall tensor is At the same time, concatenation does not change the internal dimensions of each vector. It simply extends the sequence length, adding learnable positional codes to the concatenated patch sequence to obtain input with positional information:

[0043] (13), of which , ;

[0044] (2) Router-based bidirectional information interaction: CVPE adopts a bidirectional attention mechanism based on learnable routers. First, through multi-head attention (MHA), the router is used as the query vector, and the patch is embedded as key-value pairs to extract global interaction features from all variables.

[0045] (14)

[0046] in, The aggregated router representation shows that residual connections ensure training stability. Then, CVPE distributes the fused information from the router back to each patch, using the patch as the query vector and the router as the key for further attention calculation.

[0047] (15)

[0048] Obtain enhanced embeddings that include cross-variable information such as geographic and temporal information. The process involves an enhanced patch embedding sequence containing deeply fused cross-temporal and cross-variable information, followed by layer normalization and a feedforward network (MLP) processing.

[0049] (16)

[0050] Output Its sequence length is It is no longer a raw, independent sequence of observations, but an enhanced feature representation that has been dynamically conditionalized for future operating conditions and deeply integrated with internal spatiotemporal correlations. This enhanced... It will be used as the final input to the AutoConNet long-term branch encoder for subsequent decomposition and prediction.

[0051] In step 2 of this invention, for each operating mode category First, global sequence definition and sliding window sampling are performed to transform the raw data into standardized input samples. Then, using the full set of aircraft flap data samples contained therein, a specialized flap health prediction model is independently trained. The model is a deep ensemble network: on the one hand, it uses a bi-branch decomposition architecture to separate long-term and short-term features, and uses the AutoCon mechanism for global supervision to achieve perception of long-distance similarity trends; on the other hand, it embeds a conditional variational location encoding fusion module on the long-term branch, and uses multi-source patch embedding, semantic alignment, and a router-based bidirectional information interaction mechanism to deeply fuse future flight information and historical health information, resulting in a network composed of... A pre-trained model library consisting of several "expert models" .

[0052] In step 3 of this invention, for a civil aircraft flap to be tested, the following specific measures are adopted for knowledge transfer and model adaptation:

[0053] Step 3-1: Trend Matching: First, using the PCA transformation determined in Step 1, the extension and retraction time series of the flap under test is matched. Mapped to feature vectors Then, calculate the vector with all Calculate the distance between each cluster center or all samples within a cluster, and assign each sample to the nearest cluster. :

[0054] (1),

[0055] Determine which type of trend prototype most closely resembles the operating conditions and degradation trend of the aircraft's flaps, and then select the pre-trained expert model. As the best “mentor model” for providing prior knowledge;

[0056] Step 3-2: Differentiated Targeted Fine-Tuning. Based on AutoConNet's unique dual-branch decomposition architecture, a differentiated parameter fine-tuning strategy is implemented:

[0057] For short-term branches, a full parameter freeze and direct transfer strategy is implemented; for long-term branches, the model is trained using a small amount of data from the test subjects to update the parameters.

[0058] Step 4 of this invention enters the final inference and verification stage. In the inference stage, the historical monitoring sequence of the individual to be tested and the future scheduled flight plan are used as joint inputs to drive the dedicated model to perform forward inference and generate a future health status evolution sequence with environmental awareness. In the verification stage, the root mean square error (RMSE) and mean absolute error (MAE) are calculated to quantitatively evaluate the model's fitting accuracy to the degradation trajectory of a specific individual.

[0059] In the task of predicting the health status of flap systems, this invention faces two major challenges in accurately capturing the degradation patterns: First, health indicators are affected by environmental factors such as seasons and exhibit significant long-term changing trends, which limits traditional models to local observation windows and makes it difficult to accurately capture long-distance time dependence and the true degradation patterns; Second, different flaps have different degradation patterns, and general models that rely solely on full data lack specificity and are difficult to accurately characterize the specific degradation trajectory of individuals. To address these challenges, this invention proposes a flap health status prediction method based on transferable trend grouping autocorrelation contrastive learning. First, a clustering algorithm is used to mine typical degradation trends from the full data, dividing differentiated individual samples into several highly consistent trend data groups. Secondly, an autocorrelation comparison mechanism is introduced within each group to capture the common temporal dependencies between samples across long periods, constructing multiple pre-trained models that can characterize specific degradation patterns. Then, dynamic matching is performed based on the features of the flap to be tested to select the most suitable trend category and pre-trained model. Finally, based on a transfer fine-tuning strategy, the pre-trained model is fine-tuned using a small amount of data for that flap, thereby efficiently obtaining a specific prediction model for that flap. This mechanism enables the transfer and reuse of similar degradation data between different flaps, and with the accumulation of operational data, the expansion of samples in each trend group will continuously enhance model performance. Experimental results show that this method effectively utilizes the commonalities in data between different flaps, helping to improve the prediction accuracy of individual samples and the whole. Attached Figure Description

[0060] Appendix Figure 1 This is a curve illustrating the limitations of traditional window mechanisms in capturing similar characteristics of flaps across long periods.

[0061] Appendix Figure 2 This is a flowchart of the present invention.

[0062] Appendix Figure 3 This is a schematic diagram of the network structure of the present invention.

[0063] Appendix Figure 4 This is a schematic diagram of the CVPE module in this invention, which considers the fusion of historical information and future flight information.

[0064] Appendix Figure 5 This is a schematic diagram comparing the prediction accuracy of various methods on various flaps in the embodiments of the present invention, wherein... Figure 5 In the table, (a) is MAE, (b) is MRE, and (c) is RMSE. Detailed Implementation

[0065] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0066] The following will provide a systematic discussion of the TransTG-AutoCon proposed in this invention.

[0067] This example will delve into the design principles and implementation process of TransTG-AutoCon. Overall, addressing the significant differences in individual degradation patterns of civil aircraft flap systems and the insufficient adaptability of a single universal model, this method constructs a structured "grouping-transfer" prediction paradigm. For example... Figure 2 As shown, the complete workflow of this method mainly includes four key steps: clustering, training, fine-tuning, and prediction.

[0068] The specific implementation details of each step are described below:

[0069] Step 1: Clustering – Structured Clustering of Full Data Based on Differentiated Degradation Features: Based on the degradation features of the full historical data, clustering algorithms are used to identify and separate several representative typical degradation modes (Category 1 to Category 2). ), to construct highly consistent trend data groupings.

[0070] To effectively identify differential degradation patterns in the full flap data set, the operational degradation characteristics of each aircraft must first be quantitatively characterized in low dimension. Given the high-dimensional nature of the original flap actuation time series, this example uses principal component analysis to project the high-dimensional time series data into a low-dimensional feature space. The extracted principal components effectively capture the direction of variation with the largest variance in the overall data, representing the macroscopic differences between different aircraft in terms of operational load, environmental experience, etc. Subsequently, the K-Means algorithm is used to analyze the projected feature vector set. Perform cluster analysis. Through this step, the source domain data is structurally divided into... Categories of different operating modes Each category The flap system deployment and retraction time series of the aircraft exhibited highly similar macroscopic dynamic characteristics, indicating that they experienced similar operating environments and may follow similar performance evolution trajectories.

[0071] Step 2: Training – Group-based Dedicated Model Training for Various Trends: Data preprocessing and sliding window sampling are performed separately within each trend group. A combined prediction network based on Autocorrelation Contrast Learning (AutoConNet) and Conditional Variational Position Encoding Module (CVPE) is trained independently to capture common evolutionary patterns across various modes, thereby constructing a pre-trained model library containing multiple expert models (Model 1 to Model 2). ).

[0072] For each operating mode category First, global sequence definition and sliding window sampling are performed to transform the raw data into standardized input samples. Then, using the full set of aircraft flap data samples contained within, a specialized flap health prediction model is independently trained. It is worth noting that this model is a deep ensemble network: on the one hand, it uses a dual-branch decomposition architecture to separate long-term and short-term features, and utilizes the AutoCon mechanism for global supervision to achieve the perception of long-distance similar trends; on the other hand, it embeds a conditional variational location encoding fusion module into the long-term branch, and employs multi-source patch embedding, semantic alignment, and a router-based bidirectional information interaction mechanism to deeply fuse future flight information with historical health information. Thus, a model is obtained... A pre-trained model library consisting of several "expert models" Through this group training mechanism, each expert model Both can, to a certain extent, learn and encode the unique performance evolution laws of their corresponding actuation systems, thus avoiding the averaging effect of a single general model when processing multimodal data.

[0073] Step 3: Fine-tuning – Differentiated directional fine-tuning based on a trend-matching bi-branch decomposition architecture: For the flap to be tested, firstly, match the most similar trend category (category) based on its historical data characteristics. Then, the test data was used to train the corresponding pre-trained model (model). The AutoConNet dual-branch decomposition architecture is differentiated and fine-tuned to efficiently obtain a dedicated prediction network for the individual's flaps.

[0074] For the flaps of a civil aircraft to be tested, the following specific measures are adopted for knowledge transfer and model adaptation:

[0075] (1) Trend matching: First, using the PCA transformation determined in step 1, the extension and retraction time series of the flap under test is matched. Mapped to feature vectors Then, calculate the vector with all Calculate the distance between each cluster center (or all samples within a category) and assign it to the nearest cluster. :

[0076] (1),

[0077] This step allows us to determine which type of trend prototype most closely resembles the operating conditions and degradation trend of the aircraft's flaps, thus enabling the selection of a pre-trained expert model. As the best “mentor model” for providing prior knowledge.

[0078] (2) Differentiated Targeted Fine-Tuning: Next, based on AutoConNet's unique dual-branch decomposition architecture, a differentiated parameter fine-tuning strategy is implemented:

[0079] For the short-time branch, a full parameter freeze and direct transfer strategy is implemented. This decision is based on a deep understanding of the physical mechanisms: the short-time branch is mainly responsible for capturing high-frequency random fluctuations caused by sensor noise, instantaneous disturbances in the hydraulic system, etc. These high-frequency characteristics are influenced by underlying physical laws and exhibit high universality and commonality among different flap individuals, hardly migrating with changes in the individual's decay state. Therefore, directly reusing the short-time branch parameters of the source domain model is equivalent to loading the model of the individual under test with a robust physical noise filter validated by large-scale data. This approach fully utilizes the statistical advantages of training with full data and effectively avoids overfitting the model to random noise when there is only a small amount of individual data, thus ensuring the stability of the prediction.

[0080] For the long-term branch, the model is trained using a small amount of data from the tested individual to update parameters. This is because the long-term branch carries the core of the model, responsible for handling macroscopic evolutionary features with significant individual differences. Its fine-tuning process is essentially a calibration of individual specificity. Specifically, this calibration includes two levels: First, calibration of the relationship between environmental conditions and flap degradation trends. Since this branch integrates a future information fusion module, it is responsible for learning the specific impact of environmental conditions such as temperature and load on the system. Different individuals often have different sensitivities to the same conditions due to differences in component health status. Therefore, fine-tuning this part can correct the model's perception of the environmental response characteristics of a specific individual. Second, calibration of degradation trend parameters. This branch is supervised by the AutoCon contrastive learning mechanism and is responsible for constructing a latent space representing the degradation trend and seasonal cycle. Although different individuals have similar physical causes, their specific degradation rates, seasonal amplitudes, and other macroscopic parameters vary. Through fine-tuning, the model can relatively accurately lock the true evolutionary trajectory and current state of the individual based on the actual data in the pre-trained high-quality representation space, thereby achieving personalized prediction.

[0081] Step 4: Prediction – Individual Differentiation Trend Prediction of Flaps Based on Dedicated Prediction Network: The generated dedicated network is used to extrapolate the future health status of the flap under test, outputting the final performance degradation trend prediction result. After completing the above-mentioned directional fine-tuning, a dedicated prediction network adapted to the flap under test is generated, and then the final inference and verification stage begins. In the inference stage, the historical monitoring sequence of the individual under test and the future scheduled flight plan are used as joint inputs to drive the dedicated model to perform forward inference, generating a future health status evolution sequence with environmental awareness capabilities. In the verification stage, by calculating indicators such as root mean square error (RMSE) and mean absolute error (MAE), not only can the fitting accuracy of the model to the degradation trajectory of a specific individual be quantitatively evaluated, but the generalization effectiveness of the proposed transfer strategy in solving the problem of cross-individual data heterogeneity can also be comprehensively demonstrated.

[0082] AutoConNet is a deep prediction network based on a decomposition structure, designed to capture trends and patterns across long time periods using an autocorrelation-contrast learning mechanism. Figure 3 As shown, the network is not a single black box model, but a structured system consisting of three core components: data input and preprocessing, a bi-branch decomposition architecture, and an autocorrelation contrastive learning mechanism. It aims to explicitly separate the physical causes of the original signal, which is mixed with slow decay, periodic fluctuations, and high-frequency noise.

[0083] Data input and preprocessing include the following:

[0084] (1) Global sequence definition and sliding window sampling: The network input comes from a global perspective of flap health index time series. The complete historical operational data of an aircraft flap system is formally defined as a multivariate time series. ,in This represents the total number of time steps (i.e., flight counts) within the entire observation period. The observation value at each moment. It is an K-dimensional vector composed of core health indicators (flap actuation time) and several key influencing factors (such as flight altitude, pitch angle, ambient temperature, etc.).

[0085] To construct a training dataset suitable for supervised learning, a sliding window method was used to analyze the time series of flap health indicators. Sampling is performed. This method uses a fixed-length window in the sequence. Swipe up; each window is divided into historical observation sequences (input). And the future sequence to be predicted (output) Specifically, the first Sample It can be represented as an input-output pair ,in: , Represents historical information used for prediction; , This represents a future health indicator that needs to be predicted. (This is achieved through the complete sequence.) Slide the slider with a step size of 1, and finally generate a sequence containing... Training set of samples .

[0086] (2) Window normalization and feature splitting: For any input sample To eliminate the differences in measurement units between different variables and to address the non-stationarity issue, the data first enters the window normalization module, which calculates the mean within the current window. and standard deviation Transform the original sequence into normalized features. Furthermore, the characteristics of future flight schedules are defined as follows: It will be with Both predictions are then fed into the CVPE module for information fusion. Finally, the model's total predictions should undergo window inverse normalization and be superimposed with the predictions from the two parallel branches. (2) This design structurally decomposes the prediction task, breaking down the complex prediction objective into independent modeling of the systematic slow-changing components and the high-frequency fast-changing components.

[0087] AutoConNet employs, for example, Figure 3 This illustrates a dual-branch parallel decomposition architecture, which can effectively resolve complex temporal patterns in flap health data. The design aims to decouple the raw time-series data into two orthogonal components:

[0088] (1) Short-term branch: such as Figure 3 As shown by the purple dashed line above, the short-term branch primarily focuses on capturing high-frequency micro-dynamics in the data, i.e., removing residual random disturbances (such as sensor noise) after removing trends and periods. Because these high-frequency fluctuations lack deep temporal dependencies, this branch structure is extremely simple, consisting of a lightweight linear layer:

[0089] (3),

[0090] (2) Long-term branch: The long-term branch is the core of long-term dependency modeling. Its design aims to separate and predict dynamics that combine systemic decay and periodic changes from complex historical information. This branch adopts a deep encoder-decoder architecture and is globally supervised by the AutoCon loss mechanism.

[0091] In the encoder section, the receiver receives information about the future environment after it has been processed by the CVPE module. Features A Temporal Convolutional Network (TCN) is used to map it into a series of high-dimensional latent space representations at different scales. :

[0092] (4), of which, It is a global context representation containing temporal features, used to assist the encoder in capturing temporal correlations.

[0093] In the decoder section, to accurately capture the periodic patterns at different frequencies, the decoder abandons the traditional point-to-point mapping and instead employs a multi-scale moving average module. This module utilizes a set of pooling kernels of different sizes to aggregate temporal features at multiple granularities.

[0094] (5),

[0095] In the above formula, Represents pooling kernels of different scales. For the high-dimensional representation of the encoder output, Indicates the number of scales. This is achieved through... By weighting and fusing the moving average results at various scales, the model can adaptively reconstruct long-term trends that include multiple periodic components.

[0096] The design of long-term and short-term branches ensures to a certain extent that the model will not overfit high-frequency noise, while retaining the necessary instantaneous fluctuation information, thereby improving the reconstruction of the overall signal.

[0097] Autocorrelation-based comparative learning mechanism: AutoCon:

[0098] Traditional models are limited by their local window of view and struggle to perceive seasonal patterns across years. To address this issue, AutoConNet introduces the AutoCon module at the bottom, which uses autocorrelation calculation (FFT & Autocorrelation calculation) to provide a global calibration signal for the encoder.

[0099] (1) Prior calculation driven by autocorrelation

[0100] The model first processes the global sequence. Perform a Fast Fourier Transform (FFT) to compute its global autocorrelation function. .in, Indicates time delay. Represents the mathematical expectation. For time series in The value at time, For time series in delay Subsequent values, The mean of the time series. The variance of the time series;

[0101] (6),

[0102] This function reveals the signal at different lag times. The inherent similarity. For example, if The extremely high autocorrelation coefficient at the celestial sphere indicates a significant annual cycle in the system. Therefore, the system with the highest autocorrelation coefficient is selected. Construct a Top-k lag set with each lag time. .

[0103] (2) Similar window retrieval and comparative learning

[0104] The AutoCon module utilizes a collection As an index, historical windows with high autocorrelation to the current window are dynamically retrieved as positive samples during training. To measure the similarity between two windows, two windows are defined. and Theoretical similarity between The absolute value of the global autocorrelation coefficient corresponding to the time interval:

[0105] (7),

[0106] in This represents the global autocorrelation function, while Represents the entire time series In time interval The global autocorrelation coefficient under the given conditions, therefore This represents the theoretical similarity between two time windows derived from the global autocorrelation function. This quantity reflects the lag... At that time, does the sequence as a whole exhibit a fixed periodic pattern? For the deployment and take-off time of the actuation system, if a certain lag corresponds to a high value, it indicates that the deployment and take-off behavior of the actuation system of different flights during that time interval was macroscopically similar in history.

[0107] Based on the above definitions, the contrastive loss is defined below. The loss function forces the model to learn a globally consistent representation: if two windows are physically similar (based on global autocorrelation), then they must also be similar in the latent space learned by the model. Contrastive loss based on autocorrelation. As shown in formula (8):

[0108] (8), among which, This indicates the number of windows in a single input model. , and These are all indices for the time window, representing the current window, the candidate positive sample window, and the negative sample window, respectively. Indicates the first The first window and the first Global autocorrelation coefficient between windows Indicates the first The first window and the first Window feature representation vector , Similarity between them Temperature is a parameter used to control the smoothness of the similarity distribution. As an indicator function, when the window With window Distance between Less than or equal to window With window Distance between The value is 1 if it is true, and 0 otherwise.

[0109] (3) Joint training loss objective function: Based on the above AutoCon loss design, the objective function was reconstructed as follows:

[0110] (9), among which, Prediction error, such as mean squared error (MSE), is used to measure the error in predicting future release and take-off times. Balance the weights of the two. This is achieved through... To optimize the model, it is forced to cluster window representations that operate under similar conditions (such as similar seasons and city temperatures) in the representation space.

[0111] Conditional variational location coding fusion module that integrates future flight information:

[0112] The future health status of a civil aircraft flap system is influenced not only by historical operational data (such as historical actuation time, flight altitude, and seasonal temperature), but also by known future operational conditions (such as the departure and arrival airports and seasons in future flight plans). To achieve a paradigm shift from "historical extrapolation" to "conditional prediction," this example innovatively introduces a Conditional Variational Position Encoding Fusion (CVPE) module before the long-time branch of AutoConNet. This module aims to semantically align heterogeneous historical influencing factors and, based on this, inject future operational conditions as a conditional context, thereby generating an enhanced feature representation with operational condition understanding.

[0113] like Figure 4 As shown, the CVPE module receives two inputs: historical data after window normalization. and condition information representing future operating conditions. ,in These represent the sample size for the input window and future flight information, respectively, i.e., the time step; specifically, using past data... Predicting the future from individual sample information One target value, here. Includes the past The number of samples per flight Includes the future Information on each flight; and These represent the number of feature parameters contained in historical data and the number of feature parameters contained in future flight information, respectively.

[0114] Multi-source patch embedding and semantic alignment:

[0115] First, open the input window The future operating condition window is divided into two parallel segments. and There are non-overlapping time patches, among which , The lengths of each patch, which is used to divide historical information and future flight information, are respectively. and These represent the number of patches into which each component is divided. Each historical data patch is denoted as , No. Each future flight data patch is denoted as Then, each corresponding patch is flattened into a vector using the flatten function. and Due to historical information ( (Wei) and future flight information ( Since each patch has a different feature dimension, independent linear projection layers are used to map the flattened patch to the same feature dimension. This involves a dimensional embedding space to achieve semantic alignment. Specifically:

[0116] (10)

[0117] (11),

[0118] in, , , , 。

[0119] The vectors obtained from the two projections above are concatenated along the patch dimension in chronological order to form a patch sequence that conforms to a time series relationship:

[0120] (12)

[0121] in, It is worth noting that the above only considers single-sample processing. If batch size is taken into account... The shape of the overall tensor is At the same time, concatenation does not change the internal dimensions of each vector. It simply extends the sequence length.

[0122] To provide explicit temporal and positional information to the concatenated unified patch sequence, enabling the model to distinguish patches from different time periods and sources, thereby improving the selectivity and fusion capability of the Router module, the concatenated patch sequence is further augmented with learnable positional encoding to obtain input with positional information:

[0123] (13)

[0124] in , .

[0125] Router-based two-way information exchange:

[0126] To efficiently capture and fuse the complex interactions between all patches within a window, as well as between different variables (channels), CVPE employs a bidirectional attention mechanism based on learnable routers. This task initializes... A learnable router vector These can be regarded as semantic anchors of the critical states of the system.

[0127] First, a multi-head attention (MHA) mechanism is used, with the router as the query vector and the patch embedding as key-value pairs, to extract global interaction features from all variables.

[0128] (14)

[0129] in, The aggregated router representation shows that residual connections ensure training stability. Then, CVPE distributes the fused information from the router back to each patch, using the patch as the query vector and the router as the key for further attention calculation.

[0130] (15)

[0131] Obtain enhanced embeddings that include cross-variable information such as geographic and temporal information. The enhanced patch embedding sequence contains deeply fused cross-temporal and cross-variable information. Finally, it undergoes layer normalization and a feedforward network (MLP) processing.

[0132] (16)

[0133] Output Its sequence length is It is no longer a raw, independent sequence of observations, but an enhanced feature representation that has been dynamically conditionalized for future operating conditions and deeply integrated with internal spatiotemporal correlations. This enhanced... This will serve as the final input to the AutoConNet long-term branch encoder for subsequent decomposition and prediction. Compared to traditional channel-independent modeling, the CVPE module is computationally efficient and training-stable, while also improving the model's predictive ability in complex scenarios with multiple sources and heterogeneous influencing factors, making it particularly suitable for typical health status prediction problems such as aerospace actuation systems.

[0134] This example systematically verifies the effectiveness and superiority of TransTG-AutoCon using real-world flight path data. First, it clarifies the data sources and the meaning of key physical parameters, establishing the data foundation for model training. Second, it defines the experimental configuration in detail, including the selection of the benchmark model, the setting of evaluation criteria, and the grid search strategy for hyperparameters. Finally, it analyzes the experimental results from multiple dimensions: on the one hand, it verifies the model's prediction accuracy through horizontal comparison with advanced algorithms; on the other hand, it demonstrates the generalization performance of the method through vertical analysis of the transfer strategy, thus providing comprehensive experimental conclusions.

[0135] This example uses the flap system of a certain type of civil airliner as the research object, constructing a high-quality monitoring dataset containing data from 30 aircraft of a certain airline during continuous flight cycles. The original data comes from flight parameters and corresponding flight operation information recorded by the onboard Fast Access Recorder (QAR), covering the complete takeoff, cruise, and landing phases. The variable definitions and feature selection for the prediction task are based on an in-depth analysis of the mechanical actuation principle and aerodynamic characteristics of the flaps. The experiment selects "flap actuation time" as the core indicator characterizing the system's health status. Although the flaps undergo stages such as flap deployment before takeoff, flap retraction before cruise, flap deployment before landing, and flap retraction after landing in a single flight, considering that the aerodynamic load during the landing phase is the most complex and critical, this example specifically selects "flap deployment time before landing" as the prediction target value to capture the performance degradation characteristics of this critical phase.

[0136] To eliminate the interference of environmental and load factors on actuation time, flight altitude, ambient temperature, and maximum pitch angle were introduced as key covariates. These parameters have significant physical coupling with the aerodynamic load of the flap system. For example, temperature affects hydraulic oil viscosity and system damping, while altitude and pitch angle directly determine the aerodynamic pressure distribution on the flap surface. Introducing these parameters helps the model distinguish between fluctuations caused by operating conditions and substantial performance degradation. In addition to the above physical parameters, this example also collected three types of auxiliary information to support the operation of the transfer strategy. Among them, aircraft ID is used to distinguish different individual objects, supporting individual-based heterogeneity analysis and model grouping; take-off and landing cities are used to encode route characteristics, capture the potential impact of different routes on flap wear, and serve as contextual input for future flight information; flight timestamps are used to determine the strict temporal sequence of the same flap actuation event, ensuring the correct causal logic of time series modeling.

[0137] To ensure fairness in the comparative experiments, this example uses uniform hyperparameter settings for shared components involving the AutoConNet structure in all models to eliminate interference from unstructured factors. Specifically, once the AutoConNet hyperparameters are determined, subsequent models involving the AutoConNet structure will use the same settings. The training batch size is uniformly set to 128, and the number of training epochs is 128. The dataset is divided as follows: the first 70% of the dataset for a single aircraft is used for training, and the last 30% is used for testing. The time step (patch length) within a single window is 120, and the observable length (input length) within the window is... Set to 60, the prediction length of the window. Also set it to 60.

[0138] Three indicators were used to evaluate the experiment. Compared with the true value The errors between them are: Mean Absolute Error (MAE), Mean Relative Error (MRE), and Root Mean Square Error (RMSE), calculated using the following formulas:

[0139] (17) (18),

[0140] (19), among which, Represents the index of an element in a vector, for example, Refers to vector The One element; This represents the total number of elements in the predicted value vector and the actual value vector.

[0141] The aim of this study is to determine the optimal hyperparameter configuration for AutoConNet and, based on this, to verify the performance gains of the CVPE module compared to traditional fusion strategies.

[0142] To explore the performance potential of AutoConNet on the flap dataset, this example employs a grid search strategy to systematically optimize key hyperparameters. The search space covers four dimensions: hidden layer depth (Layers: {1, 2}), dropout ratio ({0.1, 0.2}), learning rate (Learning Rate: {0.01, 0.001}), and four window normalization mechanisms: Reversible Instance Normalization (ReVIN), Mean Normalization, Decomp Normalization, and LastVal Normalization.

[0143] To eliminate the influence of differences in the dimensions of different error metrics and to comprehensively evaluate model performance, this example constructs a standardized composite score. This score is composed of the mean of RMSE, MAE, and MRE after Min-Max normalization; a lower score indicates better overall predictive performance. Table 1 shows the experimental results under different parameter combinations.

[0144] Table 1 Performance evaluation of AutoConNet under different hyperparameter combinations

[0145]

[0146] Analysis shows that window normalization has the most significant impact on model performance. Among them, trend decomposition normalization performs best, thanks to its ability to effectively separate non-stationary degradation trends from periodic fluctuations, which is highly compatible with the decomposition architecture of AutoConNet. The final optimal parameter combination is: number of layers = 2, Dropout = 0.1, learning rate = 0.01, and normalization method = Decomp. Under this configuration, the model achieves optimal average performance on 30 aircraft, with an RMSE of 0.8932, a MAE of 0.7275, and a MRE of 0.0661.

[0147] To verify the unique advantages of the CVPE module in multi-source information fusion tasks, an ablation comparison experiment was designed in this example. Under the above optimal parameter configuration, two model variants were constructed respectively: (1) the baseline model (AutoConNet with Linear) simply splices historical health indicators and future flight information using only a fully connected layer (Linear Layer); (2) the enhanced model (AutoConNet with CVPE) introduces the CVPE module to encode flight information as dynamic conditional context embeddings. Table 2 lists the prediction performance comparison of 30 aircraft under the two configurations in detail. The experimental results show that the model with the CVPE module embedded achieves significant accuracy improvement on the vast majority of individuals. From the overall average index, RMSE decreased from 0.9732 to 0.9420, MAE decreased from 0.7832 to 0.7652, and MRE decreased from 0.0742 to 0.0724. This result strongly demonstrates that the CVPE module is not a simple feature concatenation, but rather, through deep semantic fusion, it effectively transforms future operating condition information into conditional constraints on the evolution of the health state, thereby significantly enhancing the model's ability to perceive and adapt to the dynamic operating environment of the aircraft flap features.

[0148] Table 2 Prediction performance between AutoConNet and AutoConNet-CVPE

[0149]

[0150] Comparative experiment:

[0151] To fully validate the performance advantages of TransTG-AutoCon in predicting the health status of civil aircraft flaps, this example selects four representative state-of-the-art models in the field of time series forecasting as benchmarks for comparison. These models include Autoformer, TimesNet, Informer, and DLinear, all of which have demonstrated excellent long-range prediction capabilities on multiple public datasets. The experiments were conducted with identical training and test set partitions, and uniform training and inference were performed on a full dataset of 30 aircraft.

[0152] Table 3 details the average prediction accuracy of each model across all test samples. A cross-sectional comparison shows that TransTG-AutoCon achieves significantly better results in all three key error metrics: MAE, MRE, and RMSE (MAE = 0.7652, MRE = 0.0724, RMSE = 0.9420). Compared to Autoformer and Informer, which are based on the Transformer architecture, this method effectively overcomes the problem of local attention distraction when processing data with strong periodic noise, such as flaps. Compared to DLinear, which is based on linear decomposition, and TimesNet, which is based on multi-periodic convolution, this method more accurately captures long-range dependencies through an autocorrelation mechanism.

[0153] Table 3. Comparison of average accuracy of five methods on flap data of 30 aircraft.

[0154]

[0155] Figure 5 The error distribution of the five methods on each aircraft is further visualized using bar charts. As shown in the figure, in the vast majority of aircraft test cases, the error bar height of the comparative models is significantly higher than that of TransTG-AutoCon, indicating that existing mainstream algorithms are ill-suited to the complex individual heterogeneity in flap data. Although the prediction errors of each model are relatively small or even slightly lower than other models on a few aircraft flaps (such as Flap 9, 15, 19, 26), TransTG-AutoCon demonstrates overwhelming stability on difficult samples with complex and volatile decay trends. This strongly demonstrates that the trend-based grouping transfer strategy not only significantly reduces the average prediction error but also endows the model with extremely strong generalization ability and robustness.

[0156] Building upon the successful validation of the AutoConNet architecture and CVPE module, this example further conducts cross-individual transfer prediction experiments. The experiments aim to verify how TransTG-AutoCon leverages the knowledge system built from the source domain to efficiently adapt to prediction tasks in the target domain, and to delve into the underlying mechanisms of its core fine-tuning strategies.

[0157] First, a reference system for degradation trends is constructed based on large-scale source domain data. To accurately quantify the operational characteristics of individuals, this example extracts nine key statistical features: landing flap deployment time, flight altitude, flight speed, ambient temperature, pitch angle, takeoff city, landing city, takeoff time, and landing time. After standardization and PCA dimensionality reduction, collinearity between features is eliminated and clustering stability is improved.

[0158] Subsequently, the K-means algorithm was used to cluster the dimensionality-reduced feature vectors. To avoid subjectively setting the number of categories... The deviation, the experiment in The search is performed within a specified range, and the silhouette coefficient is used as the evaluation metric. As shown in Table 4, the silhouette coefficient... The peak value is reached when the density is 4, indicating that the intra-cluster compactness and inter-cluster separation are optimally balanced. Therefore, the source domain is ultimately divided into four typical operating mode categories.

[0159] Table 4 Number of clusters Relationship with contour coefficient

[0160]

[0161] Table 5 shows the distribution of the clustering results: class 0 contains 11 aircraft, class 1 contains 13 aircraft, and classes 2 and 3 each contain 3 aircraft. Although the class distribution is unbalanced, it truly reflects the distribution pattern in the actual fleet where "most aircraft follow the conventional pattern, while a few exhibit special patterns." Based on this classification, this example adopts an "intra-class hybrid training" strategy to construct pre-trained expert model libraries corresponding to the four trend prototypes, providing a knowledge foundation for subsequent transfer learning.

[0162] Table 5. Clustering results of flaps from 30 aircraft

[0163]

[0164] To verify the generalization advantage of the proposed method, this example selects 10 new aircraft outside the source domain as test targets. The comparison schemes are set as follows: Traditional single-aircraft flap training and testing—training the model from scratch using only the first 70% of the data from the aircraft under test, and using the last 30% for testing; TransTG-AutoCon—first mapping the features of the aircraft under test to a clustering space, matching the nearest trend prototype (expert model), then using the first 70% of the data to fine-tune the pre-trained model in a targeted manner, and using the last 30% for testing. It is worth noting that the training / test data partitioning used by the two schemes is exactly the same.

[0165] Table 6. Prediction accuracy performance of traditional prediction methods and TransTG-AutoCon on 10 flaps.

[0166]

[0167] The experimental results in Table 6 show that TransTG-AutoCon achieved significantly better prediction accuracy than the benchmark on the vast majority of aircraft. However, on aircraft 5 and 6, the performance difference between the two methods was negligible. Analysis suggests that aircraft 5 possesses the richest historical data sample in the test set. This ample data volume allowed the benchmark model to fully converge and uncover performance evolution patterns, reaching its performance ceiling and resulting in a data saturation effect, thus reducing the small-sample gain brought by transfer learning. For aircraft 6, its degradation pattern may belong to the boundary samples or atypical patterns in the existing clustering space, making it difficult to find a perfectly matching expert model during the matching stage, which to some extent limits the effective reuse of prior knowledge. Nevertheless, TransTG-AutoCon still maintained high accuracy on these two aircraft, comparable to training on the full dataset, and demonstrated robustness under different data feature conditions on this example dataset.

[0168] To verify the rationality of the "freeze short-term branches and fine-tune long-term branches" strategy, this example designed three sets of comparative experiments: full parameter fine-tuning, fine-tuning only short-term branches, and fine-tuning only long-term branches. The results in Table 7 show that the strategy of fine-tuning only long-term branches achieved the best accuracy on the test set. Based on the decomposition architecture of AutoConNet and the CVPE fusion mechanism, this phenomenon can be attributed to three aspects: (1) Long-term branches carry individual differences: Long-term branches are designed to learn global autocorrelation and low-frequency trends outside the window. Their encoder-decoder structure is good at representing the periodic and systematic deviations across the window, which are the main manifestations of the systematic differences between different aircraft flaps. Therefore, the branch should be adapted first in the fine-tuning stage; (2) Short-term branches have universality: Short-term branches usually use low-complexity modules to fit the high-frequency and transient components within the window. Since the original large amount of data has fully trained these universal short-term physical patterns, directly fine-tuning the short-term branches is more likely to fit noise or occasional fluctuations with limited samples to be predicted, resulting in a decrease in generalization ability; (3) Recalibration of CVPE context: The cross-variable context information introduced by the CVPE module is mainly reflected in the global representation of the long-term branches. Since different aircraft models have different responsiveness to changes in operating conditions such as altitude and temperature, the joint representation of these cross-variables must be recalibrated by fine-tuning the long-term branches during migration. In summary, freezing short-term branches and focusing on fine-tuning long-term branches not only aligns with the model's representational division of labor but also provides a more robust transfer path when sample size is limited.

[0169] Table 7. Comparison of prediction accuracy among the three fine-tuning strategies

[0170]

[0171] As a key indicator reflecting mechanical wear and hydraulic performance, the accurate prediction of flap actuation time is crucial for achieving condition-based flap maintenance in civil aircraft. This example addresses the limitations of existing methods in capturing long-distance time dependencies and the lack of specificity in general models when dealing with highly heterogeneous individual degradation patterns. It innovatively integrates the global representation capabilities of autocorrelation contrastive learning networks with a "clustering-pre-training-adaptation" group transfer strategy, alleviating the physical limitations of local sliding windows and capturing slowly varying degradation patterns. Furthermore, by mining common features of group data, it alleviates to some extent the contradiction between the scarcity of individual data and high modeling costs. Thanks to the effective identification of degradation trends and targeted fine-tuning of long-term branches, TransTG-AutoCon achieves superior prediction accuracy and robustness compared to current mainstream time series prediction models such as Autoformer, TimesNet, Informer, and DLinear on the dataset constructed in this example. Compared to the traditional single-aircraft independent modeling paradigm, this method does not require building a model from scratch for each aircraft; instead, it achieves efficient knowledge transfer based on a limited expert model library. This performance improvement does not stem from simply increasing model complexity, but from a deep understanding of physical degradation mechanisms and data distribution characteristics, and from structured modeling.

Claims

1. A method for predicting flap health status based on transferable trend grouping autocorrelation contrastive learning, characterized in that, First, clustering algorithms are used to mine typical degradation trends from the full dataset, dividing differentiated individual samples into several highly consistent trend data groups. Secondly, an autocorrelation comparison mechanism is introduced within each group to capture common temporal dependencies across long periods among samples, constructing multiple pre-trained models that can characterize specific decay patterns. Then, dynamic matching is performed based on the features of the flap under test to select the most suitable trend category and pre-trained model. Finally, based on a transfer fine-tuning strategy, the pre-trained model is fine-tuned using a small amount of data from the flap, thereby efficiently obtaining a specific prediction model for the flap. This includes the following steps: Step 1: Structured clustering of the full dataset based on differentiated degradation features: Based on the degradation features of the full historical data, a clustering algorithm is used to identify and separate several representative typical degradation modes, including category 1 to category 2. To construct highly consistent trend data grouping: First, the operational degradation characteristics of each aircraft need to be quantitatively represented in low dimension. Principal component analysis is used to project the high-dimensional time series data into a low-dimensional feature space. The extracted principal components capture the direction of variation with the largest variance in the overall data, representing the macroscopic differences in operational load and environmental experience among different aircraft. Subsequently, the K-Means algorithm is used to process the projected feature vector set. Cluster analysis was performed, and the source domain data was structured into... Categories of different operating modes Each category The flap system deployment and retraction time series of the aircraft exhibited highly similar macroscopic dynamic characteristics, indicating that they experienced similar operating environments and may follow similar performance evolution trajectories. Step 2: Grouped Dedicated Model Training for Various Trends: Within each trend group, data preprocessing and sliding window sampling are performed separately. A combined prediction network based on Autocorrelation Contrast Learning (AutoConNet) and Conditional Variational Position Encoding (CVPE) is trained independently to capture common evolutionary patterns across various trends. This results in the construction of a pre-trained model library containing multiple expert models, including Model 1 to Model 2. ; Step 3: Differentiated directional fine-tuning of the trend-matching bi-branch decomposition architecture: For the flap to be tested, firstly, match the most similar trend category based on its historical data characteristics, and denote it as the category. Then, the test data is used to train the corresponding pre-trained model, i.e., the model. Differentiated and targeted fine-tuning of the AutoConNet dual-branch decomposition architecture is performed to efficiently obtain a dedicated prediction network for each individual flap; Step 4: Predict individual differences in flap trends based on a dedicated prediction network. Use the generated dedicated prediction network to extrapolate the future health status of the flap under test and output the final performance degradation trend prediction results.

2. The flap health status prediction method based on transferable trend grouping autocorrelation contrastive learning according to claim 1, characterized in that, In step 2, for each operating mode category First, global sequence definition and sliding window sampling are performed to transform the raw data into standardized input samples. Then, using the full set of aircraft flap data samples contained therein, a specialized flap health prediction model is independently trained. The model is a deep ensemble network: on the one hand, it uses a bi-branch decomposition architecture to separate long-term and short-term features, and uses the AutoCon mechanism for global supervision to achieve perception of long-distance similarity trends; on the other hand, it embeds a conditional variational location encoding fusion module on the long-term branch, and uses multi-source patch embedding, semantic alignment, and a router-based bidirectional information interaction mechanism to deeply fuse future flight information and historical health information, resulting in a network composed of... A pre-trained model library consisting of several "expert models" .

3. The flap health status prediction method based on transferable trend grouping autocorrelation contrastive learning according to claim 2, characterized in that, In step 3, for a civil aircraft flap to be tested, the following specific measures are adopted for knowledge transfer and model adaptation: Step 3-1: Trend Matching: First, using the PCA transformation determined in Step 1, the extension and retraction time series of the flap under test is matched. Mapped to feature vectors Then, calculate the vector with all Calculate the distance between each cluster center or all samples within a cluster, and assign each sample to the nearest cluster. : （1）， Determine which type of trend prototype most closely resembles the operating conditions and degradation trend of the aircraft's flaps, and then select the pre-trained expert model. As the best "mentor model" for providing prior knowledge; Step 3-2: Differentiated Targeted Fine-Tuning. Based on AutoConNet's unique dual-branch decomposition architecture, a differentiated parameter fine-tuning strategy is implemented: For short-term branches, implement a full parameter freeze and direct migration strategy; For long-term branches, the model is trained using a small amount of data from the individuals being tested, thus updating the parameters.

4. The flap health status prediction method based on transferable trend grouping autocorrelation contrastive learning according to claim 3, characterized in that, The AutoConNet described in step 2 consists of data input and preprocessing, a two-branch decomposition architecture, and an autocorrelation contrastive learning mechanism. The data input and preprocessing includes: (1) Global sequence definition and sliding window sampling: The network input originates from a global perspective of flap health indicator time series. The complete historical operational data of an aircraft flap system is formally defined as a multivariate time series. ,in The time step representing the total number of flights within the entire observation period, and the observed value at each moment. It is a vector containing 𝑐 dimensions; Using the sliding window method to analyze the time series of flap health indicators Sampling is performed using a fixed-length window in the sequence. Swipe up, and each window is divided into a historical observation sequence, i.e., input. The output is the future sequence to be predicted. , No. Sample Represented as an input-output pair ,in: , Represents historical information used for prediction; , This represents the future health indicators that need to be predicted, through the complete sequence. Slide the slider with a step size of 1, and finally generate a sequence containing... Training set of samples ; (2) Window normalization and feature splitting: For any input sample The data first enters the window normalization module, which calculates the mean within the current window. and standard deviation Transform the original sequence into normalized features. The characteristics of future flight plans are defined as follows It will be with The data are then fused together in the CVPE module. Finally, the total predicted value of the model should undergo window inverse normalization and be superimposed with the prediction results of the two parallel branches. (2).

5. The flap health status prediction method based on transferable trend grouping autocorrelation contrastive learning according to claim 4, characterized in that, The AutoConNet employs a dual-branch parallel decomposition architecture to decouple the original time-series data into two orthogonal components: the short-time branch focuses on capturing high-frequency micro-dynamics in the data, i.e., removing the random perturbations remaining after removing trends and periods, and consists of a lightweight linear layer. （3）； The long-term branch employs a deep encoder-decoder architecture, with global supervision by the AutoCon loss mechanism. In the encoder part, the received data is processed by the CVPE module and fused with future environmental information. Features Temporal Convolutional Network (TCN) is used to map the latent space representations to a series of different scales. : (4), of which, It is a global context representation containing temporal features, used to assist the encoder in capturing temporal correlations; The decoder employs a multi-scale moving average module, utilizing a set of pooling kernels of different sizes to perform multi-granularity aggregation of temporal features: （5）， Represents pooling kernels of different scales. For the high-dimensional representation of the encoder output, The number of scales is represented by the weighted fusion of the moving average results of each scale, which enables the model to adaptively reconstruct long-term trends containing multiple periodic components.

6. The flap health status prediction method based on transferable trend grouping autocorrelation contrastive learning according to claim 5, characterized in that, AutoConNet introduces the AutoCon module at the bottom, which uses autocorrelation calculations to provide a global calibration signal for the encoder, including: (1) Prior calculation driven by autocorrelation: The model first calculates the global sequence Perform a Fast Fourier Transform (FFT) to calculate its global autocorrelation function. ,in, Indicates time delay. Represents the mathematical expectation. For time series in The value at time, For time series in delay Subsequent values, The mean of the time series. The variance of the time series. （6）； (2) Similar window retrieval and comparative learning: The AutoCon module utilizes a collection As an index, historical windows with high autocorrelation to the current window are dynamically retrieved as positive samples during training. To measure the similarity between two windows, two windows are defined. and Theoretical similarity between The absolute value of the global autocorrelation coefficient corresponding to the time interval: （7）， in This represents the global autocorrelation function, while Represents the entire time series In time interval The global autocorrelation coefficient under the given conditions, therefore This represents the theoretical similarity between two time windows derived from the global autocorrelation function; this quantity reflects the lag. Does the sequence as a whole exhibit a fixed periodic pattern? Contrast loss based on autocorrelation As shown in formula (8): (8), among which, This indicates the number of windows in a single input model. , and These are all indices for the time window, representing the current window, the candidate positive sample window, and the negative sample window, respectively. Indicates the first The first window and the first Global autocorrelation coefficient between windows Indicates the first The first window and the first Window feature representation vector , Similarity between them Temperature is a parameter used to control the smoothness of the similarity distribution. As an indicator function, when the window With window Distance between Less than or equal to window With window Distance between The value is 1 if the condition is met, and 0 otherwise; the objective function is as follows: (9), among which, Prediction error, such as mean squared error (MSE), is used to measure the error in predicting future release and take-off times. Balance the weights of the two.

7. The flap health status prediction method based on transferable trend grouping autocorrelation contrastive learning according to claim 6, characterized in that, Before the long-term branch of AutoConNet, a Conditional Variational Location Coding Fusion (CVPE) module is introduced. The CVPE module receives two inputs: window-normalized historical data. and condition information representing future operating conditions. ,in These represent the sample size for the input window and future flight information, respectively, i.e., the time step; specifically, using past data... Predicting the future from individual sample information One target value, here. Includes the past The number of samples per flight Includes the future Information on each flight; and These represent the number of feature parameters contained in historical data and the number of feature parameters contained in future flight information, respectively, and specifically include the following: (1) Multi-source patch embedding and semantic alignment: First, open the input window The future operating condition window is divided into two parallel segments. and Patches of non-overlapping time segments, among which , The lengths of each patch, which is used to divide historical information and future flight information, are respectively. and These represent the number of patches into which each component is divided, and the number of patches into which each component is divided. Each historical data patch is denoted as , No. Each future flight data patch is denoted as Then, each corresponding patch is flattened into a vector using the flatten function. and Due to historical information Maintenance and future flight information Since each patch has a different feature dimension, independent linear projection layers are used to map the flattened patch to the same feature dimension. A dimensional embedding space is used to achieve semantic alignment, specifically: （10）， (11), among which, , , , The vectors obtained from the two projections are concatenated along the patch dimension in chronological order to form a patch sequence that conforms to a time series relationship. （12）， in, If batch size is taken into account The shape of the overall tensor is At the same time, concatenation does not change the internal dimensions of each vector. It simply extends the sequence length, adding learnable positional codes to the concatenated patch sequence to obtain input with positional information: (13), of which , ; (2) Router-based bidirectional information interaction: CVPE adopts a bidirectional attention mechanism based on a learnable router. First, through a multi-head attention mechanism, the router is used as the query vector and the patch is embedded as a key-value pair to extract global interaction features from all variables. （14）， in, The aggregated router representation shows that residual connections ensure training stability. Then, CVPE distributes the fused information from the router back to each patch, using the patch as the query vector and the router as the key for further attention calculation. （15）， Obtain enhanced embeddings that include cross-variable information such as geographic and temporal information. It contains a sequence of enhanced patch embeddings that incorporates deeply fused cross-time and cross-variable information. Finally, it undergoes layer normalization and feedforward network processing. （16）， Output Its sequence length is Enhanced It will be used as the final input to the AutoConNet long-term branch encoder for subsequent decomposition and prediction.

Images