A flight control test abnormal timing data prediction method combining uncertainty and deep learning

Abstract

The application belongs to the technical field of data processing and aerospace, and specifically relates to a flight control test abnormal timing data prediction method combining uncertainty and deep learning. The method aims to solve the problems of limited accuracy and insufficient robustness of traditional prediction methods in processing complex flight control timing data. The steps include: constructing a multi-scale block data embedding layer, cutting and high-dimensional mapping of historical multivariate timing; establishing a double-channel expert modeling architecture, modeling the perception machine channel for stationary and non-stationary sequences respectively; extracting statistical features and modeling variable correlation in the stationary channel, and using moving average decomposition and combining cross attention mechanism to capture residual influence to enhance the trend term in the non-stationary channel; using an adaptive gating mechanism to dynamically fuse features according to real-time characteristics and output prediction results. The application can accurately capture the local transient and global trend of flight control data, effectively deal with uncertainty deviation, reduce the computing overhead, improve the prediction accuracy and real-time response ability, and realize early warning of risks such as sensor failure and performance degradation.

Description

Technical Field

[0001] This application belongs to the fields of data processing and aerospace technology, specifically relating to a method for predicting time-series data of flight control test anomalies by combining uncertainty and deep learning. Background Technology

[0002] Flight control testing, a core component of aircraft development and verification, is a crucial means of ensuring aircraft flight quality and safety by comprehensively evaluating flight control laws, sensor networks, and control mechanisms on integrated test benches or in actual flight environments. With advancements in avionics technology, the various signal states collected during flight control testing generate massive amounts of multi-dimensional time-series data as they change over time. This data encompasses stationary sequences during stable cruise phases as well as non-stationary sequences generated by violent maneuvers, revealing the aircraft's operational patterns under different conditions and fault modes. Accurate prediction and analysis of this data are fundamental to achieving real-time monitoring of the test status.

[0003] Among these, the prediction and early warning technology for abnormal time-series data in flight control tests aims to infer future trends by modeling historical multi-dimensional observation data, thereby promptly identifying potential sensor failures, actuator performance degradation, or control logic defects. Existing prediction methods mainly rely on deep neural networks to extract time-series features and attempt to capture the evolution patterns of flight states at different stages by analyzing the nonlinear coupling relationships of multi-dimensional time-series data. An effective prediction mechanism can not only prevent damage to test equipment but also provide important references for the reliability verification of flight control systems.

[0004] However, traditional multivariate time series forecasting methods mostly employ a unified network architecture to model both stationary and non-stationary sequences. This leads to models that tend to learn easily fitted short-term local features during optimization, making it difficult to effectively capture the crucial long-term trends and complex periodic components in non-stationary sequences. Furthermore, existing sequence decomposition models often neglect the dynamic correlation between the small random noise present in the seasonal term and the trend term, failing to address the uncertainties arising from the evolution of data distribution over time. In addition, traditional models are susceptible to overfitting due to noise interference when dealing with stationary sequences, while exhibiting significant underfitting when handling complex non-stationary sequences. This results in limited prediction accuracy and delayed or false alarms in early warning systems when facing sudden spikes, step jumps, or periodic oscillations in flight control experiments.

[0005] Therefore, a method for predicting abnormal time-series data in flight control experiments that combines uncertainty and deep learning is desired. Summary of the Invention

[0006] The purpose of this invention is to provide a method for predicting abnormal time series data of flight control experiments by combining uncertainty and deep learning, which can effectively solve the problems in the background art mentioned above.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0008] A method for predicting anomalous time-series data in flight control experiments, combining uncertainty and deep learning, includes the following specific steps: Step 1: Constructing a multi-scale segmented data embedding layer: The input historical multivariate time series of flight control experiments is processed by multi-scale segmentation to generate segmented data with different time granularities. The segmented data is then mapped to a high-dimensional latent space through the embedding layer to generate segmented feature vectors. Step 2: Establishing a dual-channel expert modeling architecture: For stationary and non-stationary sequences in the flight control experiment data, a multilayer perceptron channel for stationary sequences and a multilayer perceptron channel for non-stationary sequences are constructed respectively. The segmented feature vectors are used as input to achieve targeted modeling of sequences with different distribution characteristics. Step 3: Performing stationary sequence feature extraction: In the stationary sequence multilayer perceptron channel, features are extracted through multiple channels with residual connections. The multilayer perceptron structure calculates the statistical characteristics of each variable and uses the transpose operation to model the correlation between variables to obtain the predicted components of the stationary sequence; Step 4 performs non-stationary sequence decomposition and correlation modeling: In the non-stationary sequence multilayer perceptron channel, the moving average algorithm is used to decompose the sequence into a trend term and a seasonal term, and the influence of random disturbances in the seasonal term on the trend term is captured based on the cross-attention mechanism to generate an enhanced trend term with residual correction; Step 5 performs dynamic feature fusion and anomaly prediction: Using an adaptive gating mechanism, the weighted ratio of the output results of the stationary sequence multilayer perceptron channel and the non-stationary sequence multilayer perceptron channel is calculated according to the real-time dynamic characteristics of the input sequence, and the fused features are mapped to the prediction time dimension to output the prediction results of the flight control test anomaly time series data.

[0009] Preferably, in step 1, the input historical time series sample contains multiple interdependent flight parameter variables, the number of variables is set to a first preset number, the historical observation time step is set to a first preset duration, and the predicted future time step is set to a second preset duration.

[0010] Preferably, the multi-scale segmentation process uses multiple sliding windows of different scales to divide the model into blocks, with the window size of each scale set to the corresponding preset number of time points. Through multi-scale segmentation, the model can simultaneously capture local transient changes and global long-term evolution trends.

[0011] Preferably, the embedding layer maps each data block to a high-dimensional space of a preset dimension, and uses a one-dimensional tiling operation to convert the multi-dimensional block features into a two-dimensional matrix form, providing standard tensor input for subsequent multilayer perceptron operations.

[0012] Preferably, the stationary sequence multilayer perceptron channel in step 2 is configured as a lightweight residual network structure. By limiting the network depth to a preset number of layers, overfitting can be prevented when the model processes engine steady-state vibration signals or cruise phase signals with stable statistical characteristics.

[0013] Preferably, in step 3, the process of extracting time statistical features includes calculating the mean and variance of the input sequence in the block dimension, and injecting the mean and variance as latent features into the multilayer perceptron layer to enhance the robustness of the model to stationary sequence noise.

[0014] Preferably, the correlation between the modeling variables specifically includes: transposing the matrix of learned time features to interchange the original time dimension and variable dimension; using another set of multilayer perceptron weights to extract the nonlinear coupling relationship between different sensor signals such as three-axis angular rate, control surface deflection angle, and flight overload; and finally transposing it again to restore the data structure.

[0015] Preferably, in step 4, the window size of the moving average algorithm is set to a preset size to smooth the sequence and extract the long-term evolution trajectory as the trend term. The residual part after subtracting the trend term from the original sequence is defined as a seasonal term containing periodic features and small random perturbations.

[0016] Preferably, the specific implementation of the cross-attention mechanism includes: using the trend item as the query value, using the seasonal item as the key value and the numerical value respectively, and obtaining the attention weight coefficient by calculating the matrix product of the query value and the key value and performing normalization processing.

[0017] Preferably, after obtaining the attention weight coefficient, it is weighted and summed with the numerical value to obtain the dynamic influence of the small residuals in the seasonal term on the trend term. This influence is then added to the original trend term in an additive form to form an enhanced trend term that can cope with uncertainty shifts.

[0018] Preferably, the non-stationary sequence multilayer perceptron channel performs independent linear mapping processing on the enhanced trend term and the seasonal term, and uses a deep neural network to extract the complex seasonal components of the flight overload sequence generated by violent maneuvers under non-stationary conditions.

[0019] Preferably, in step 5, the adaptive gating mechanism is constructed using a learnable weight parameter matrix, the values ​​of which are within a preset range, and the contribution of stationary and non-stationary channels is dynamically adjusted based on the real-time entropy or variance of the input data.

[0020] Preferably, the calculation process of the weighting ratio follows the following logic: when the input sequence exhibits high stationarity characteristics, the gating mechanism automatically increases the weight of the multilayer perceptron channel for stationary sequences; when a sharp shift or extreme fluctuation is detected in the data distribution, the weight of the multilayer perceptron channel for non-stationary sequences is automatically increased.

[0021] Preferably, the prediction result is output through a linear network containing multiple fully connected layers. This linear network decodes the fused multidimensional latent features into specific physical signal prediction values, and the output dimension is consistent with the set second preset duration.

[0022] Preferably, the flight control test anomaly timing data prediction method is applied to the iron bird integrated test bench or flight simulator environment to perform real-time monitoring and early warning of signal mutations caused by intermittent sensor failures, actuator performance degradation, control law software logic defects, and hardware interface communication anomalies.

[0023] Preferably, the abnormal time series data includes spikes exceeding normal thresholds, step responses, slow offset trends, periodic oscillation instability, and unexpected abrupt changes in characteristic frequencies. The response delay of the prediction system is set to be less than or equal to a preset time threshold to meet the real-time assurance requirements of the flight control system.

[0024] Preferably, the training process of the method uses mean squared error as the loss function, optimizes network parameters using stochastic gradient descent algorithm, and sets the learning rate, batch size, and total number of iterations to corresponding preset values.

[0025] Preferably, the method introduces an early stopping mechanism during training, which terminates training early when the prediction error on the validation set no longer decreases for a preset number of consecutive cycles, in order to ensure the generalization performance of the model under different flight control conditions.

[0026] Preferably, the present invention also includes establishing an anomaly pattern database containing a predetermined scale of historical flight control test records, using a long short-term memory network to perform a secondary evaluation on the predicted future sequence, thereby realizing the prediction of the reliability and remaining service life of the flight control system, with the prediction lead time set to be greater than or equal to a preset lead time.

[0027] Preferably, the prediction results are displayed in real time through a graphical interface, which includes the predicted trajectory of each key flight parameter, the uncertainty confidence interval, and the abnormal trigger probability score. When the prediction probability exceeds the preset alarm threshold, the audible and visual alarm device is automatically triggered and a diagnostic report is generated.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] The technical solution adopted in this invention overcomes the limitations of traditional single architectures: This invention, through a dual-channel expert modeling architecture, completely solves the accuracy limitation problem caused by the use of a unified model for mixed modeling of stationary and non-stationary sequences in traditional multivariate time series prediction methods. For stationary sequences, a lightweight multilayer perceptron is used to avoid overfitting; for non-stationary sequences, enhanced trend modeling is used to capture long-term evolutionary patterns, resulting in a significant reduction in overall prediction error compared to existing technologies.

[0030] The technical solution adopted in this embodiment of the invention has enhanced adaptability to this complex working condition. That is, the flight control test involves a variety of extreme working conditions such as cruise, violent maneuvering, and fault evolution. The multi-scale block data embedding technology of this invention can simultaneously extract local transient features and global trends. It has a very high recognition accuracy for complex abnormal patterns such as spikes, steps and oscillation instability that are common in flight control data, and effectively solves the underfitting defect of the existing technology when facing distribution offset.

[0031] The technical solution adopted in this invention significantly improves robustness against uncertainty. This invention explicitly models the impact of small random noise in the seasonal term on the trend term using a cross-attention mechanism, based on a trend model of seasonal influences. This design effectively combats the problem of predicted trajectory divergence caused by chaotic characteristics of the system or small perturbations in initial conditions, further improving prediction accuracy in long-term prediction tasks compared to traditional linear models and recurrent neural network models. The adaptive gating mechanism allows the model to dynamically adjust its prediction strategy based on the real-time characteristics of the input sequence, ensuring that the model remains in optimal operating mode during different flight state transitions, significantly reducing the false alarm and false negative rates of the early warning system. Attached Figure Description

[0032] Figure 1 This is a schematic diagram of the overall technical solution architecture of a flight control test anomaly time series data prediction method that combines uncertainty and deep learning, as proposed in this invention. Detailed Implementation

[0033] Example 1

[0034] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0035] In the extremely complex core stage of aircraft development—flight control testing—anomaly prediction of multivariate time-series data generated by environments such as iron bird integrated test benches and flight simulators is a key technical means to ensure flight safety and system reliability. The specific execution flow and technical architecture of the flight control test anomaly time-series data prediction method combining uncertainty and deep learning are as follows: The method first executes step 1, constructing a multi-scale block data embedding layer. In this process, the system receives input historical multivariate time-series samples of flight control tests, which are defined as X... input ={X1,X2,…,XM}∈R M×L M represents the first preset quantity, which is the total number of interdependent flight parameter variables. These variables cover key physical parameters such as three-axis angular rate, control surface deflection angle, flight overload, atmospheric data sensor barometric altitude, climb rate, sideslip angle, and angle of attack; L represents the first preset duration, which is the value of the historical observation time step.

[0036] In the specific implementation of step 1, the input flight control test history multivariate time series is subjected to multi-scale segmentation processing. This multi-scale segmentation process uses three sliding windows of different scales for block operation. The first scale window corresponds to a smaller time step, aiming to capture local transient changes in sensor signals, such as small high-frequency oscillations on the control surfaces caused by intermittent actuator jitter. The second scale window corresponds to a medium time step, used to extract transitional features during flight state switching. The third scale window corresponds to a larger time step, i.e., a global window, whose size is set to be close to or equal to a first preset duration, used to identify the long-term evolution trend of the flight trajectory. The window size for each scale is set to the corresponding preset number of time points. Through this multi-scale segmentation, the generated block data with different time granularities can be transformed into a three-dimensional matrix form. , where D is the number of time points within a block, and N is the total number of blocks.

[0037] Next, the segmented data is mapped to a high-dimensional latent space through an embedding layer. The embedding layer uses a linear mapping layer to encode the features of each segment, transforming it into a high-dimensional space vector of a predetermined dimension. The mapping process follows the formula... Where dim is the defined embedding dimension. To meet the input tensor format requirements of the subsequent multilayer perceptron, a one-dimensional tiling operation is performed, and the multidimensional block features are converted into a two-dimensional matrix form using data reshaping instructions. Where B is the product of N and dim. The block feature vectors produced in this step can effectively preserve the structural information of flight control data at different time resolutions.

[0038] After data preparation, step 2 is executed to establish a dual-channel expert modeling architecture. This architecture addresses the two distinct distribution characteristics exhibited in the flight control test data—stationary and non-stationary sequences—by constructing separate Stationary MLP channels for stationary sequences and Non-Stationary MLP channels for non-stationary sequences. The segmented feature vectors... As a common input, the data is split into two independent expert processing modules through data splitting logic. The stationary sequence multilayer perceptron channel is configured with a lightweight residual network structure, with the network depth limited to a preset range of 3 to 5 layers. Its purpose is to efficiently model statistically stable data segments (such as steady-state engine vibration signals generated during stable cruise or long-term state signals maintaining constant altitude and speed flight), avoiding overfitting due to redundant model parameters. The non-stationary sequence multilayer perceptron channel employs a deeper nonlinear transformation structure to handle non-stationary characteristics during severe maneuvers, sudden load changes, or fault evolution.

[0039] Subsequently, step 3 is executed to extract features from the stationary sequence. Within the multilayer perceptron channel for the stationary sequence, statistics of the input sequence in the block dimension are calculated. The process of extracting temporal statistical features specifically includes calculating the mean σmu and variance σsigma of the input sequence tensor along the time step axis. These statistical features represent the core distributional properties of the stationary signal. Injecting them as latent feature vectors into the multilayer perceptron layer can significantly enhance the model's robustness to background electromagnetic interference noise.

[0040] The correlation between the modeling variables is one of the core operations in step 3. First, the matrix of time features learned from the first set of multilayer perceptron outputs is transposed to obtain... At this point, the flight parameter variables, originally located in different rows, are mapped to the processing dimension, enabling another set of independent linear mapping layers to extract the nonlinear coupling relationships between different sensor signals such as triaxial angular rate, control surface deflection, and flight overload. For example, in stable flight, due to the closed-loop effect of the control law, there is a specific and stable proportional coordination relationship between elevator deflection and longitudinal overload. Through this transpose modeling, the model can accurately capture this co-evolutionary pattern between variables. After completing the correlation modeling, the transpose operation is performed again to restore the data structure to its original expression pattern, and the residual connection is used to add it element-wise to the original input, thereby outputting the predicted components of the stationary sequence. .

[0041] Next, step 4 is performed to conduct non-stationary sequence decomposition and correlation modeling. In the multilayer perceptron channel of the non-stationary sequence, considering the complex chaotic characteristics in the flight control experiment, sequence decomposition is first performed using a moving average algorithm. The window size of the moving average algorithm is set to a preset size based on the dynamic response period of the flight control system.

[0042] The decomposition process follows the following formula: The term represents the long-term evolution trajectory extracted through average pooling operations, i.e., the trend term, which reflects the evolution of flight attitude or velocity on a macroscopic scale. The seasonal term is defined as the residual after subtracting the trend term from the original sequence. In the context of flight control testing, this seasonal term includes the periodic signal sampled by the sensor, as well as small random disturbances caused by aerodynamic nonlinearity or actuator wear.

[0043] To address the issue of predicted trajectory deviation caused by minor perturbations in initial conditions, this embodiment introduces association modeling based on a cross-attention mechanism in step 4. Specifically, this includes: [the process of] assigning the trend item... Through the weight matrix Mapped to query value Q; the seasonal item Through the weight matrix respectively and The mapping is between a key value K and a numerical value V. This is achieved by calculating the dot product of the query value Q and the transpose of the key value K, and then dividing by a scaling factor. Normalization (Softmax) is performed to obtain the attention weight coefficients. These coefficients reflect the contribution of each time component in the seasonal term to the evolution of the trend term. A weighted sum of these weight coefficients and the numerical value V is then performed to obtain the dynamic impact of the small residuals in the seasonal term on the trend term. Finally, this influence is additively added to the original trend term to generate an enhanced trend term with residual correction. The non-stationary sequence multilayer perceptron channel then performs independent linear mapping and deep feature extraction on the enhanced trend term and seasonal term, respectively, and finally fuses and outputs the non-stationary prediction component. .

[0044] Finally, step 5, dynamic feature fusion and anomaly prediction, is executed. This embodiment utilizes an adaptive gating mechanism to calculate the weighted ratio of stationary and non-stationary channel outputs based on the real-time dynamic characteristics of the input sequence. The adaptive gating mechanism is constructed from a learnable weight parameter matrix G, whose values ​​are limited to a closed interval between 0 and 1. The system dynamically calculates weights based on the real-time entropy or variance of the input data. When the input sequence exhibits high stationarity characteristics (such as a cruise segment), the gating coefficient G tends to increase, thereby automatically increasing the weight of the stationary sequence multilayer sensor channel; conversely, when a drastic shift in data distribution is detected (such as a sudden pull-up maneuver or a sudden change in sensor fault signal), the gating mechanism automatically increases the contribution of the non-stationary sequence multilayer sensor channel.

[0045] The formula for feature fusion is as follows: in, This represents element-wise multiplication. The fused feature vector. Incorporating a comprehensive understanding of different distribution patterns, the data is then fed into a linear network consisting of multiple fully connected layers. This network decodes the high-dimensional latent features and maps them to the prediction time dimension, outputting the prediction results for the corresponding flight control test anomaly time series data. , where T is the second preset duration.

[0046] In specific flight control test engineering applications, the prediction results are displayed in real time through a graphical interface. The interface includes not only the predicted trajectories of each key flight parameter but also the uncertainty confidence interval calculated based on the prediction deviation. When the predicted value of the prediction system exceeds the set normal operating threshold or exhibits specific fault mode characteristics (such as spikes exceeding a preset threshold, abnormal periodic oscillations, etc.), the system automatically triggers an audible and visual alarm and generates a diagnostic report containing the anomaly type, probability score, and potential root cause analysis. Through long-term operation on the Iron Bird integrated test bench, this embodiment can provide early warnings for intermittent sensor failures, actuator performance degradation, and boundary logic defects in the control law software. The prediction lead is greater than or equal to 500 milliseconds, and the response delay is controlled within 20 milliseconds, greatly improving the safety margin of flight control testing.

[0047] Example 2

[0048] Based on Example 1, this example further refines the training optimization strategy and anomaly pattern recognition mechanism of the flight control test anomaly time series data prediction method under complex non-stationary environments.

[0049] During step 1, considering the flight control characteristics of different aircraft models, the first preset duration L is set to 128 time steps, while the predicted future duration T is set to 96 time steps. To further enhance the model's perception accuracy of multi-scale features, the sliding window step size in the multi-scale segmentation process is set to a sliding method with an overlap rate of 50%. When constructing the block data embedding layer, positional encoding technology is used inside the embedding layer to inject the chronological order information of the time series into the block feature vector, enabling the subsequent expert modeling architecture to perceive the absolute position of data points on the global time axis.

[0050] To further address the feature extraction of stationary sequences in step 3, this embodiment introduces a statistical comparison mechanism. During multilayer perceptron processing, the model not only calculates the mean and variance of the current block but also compares it with historical long-term benchmark statistics, using the resulting delta statistics as additional feature input. This design is particularly suitable for identifying slow drift anomalies. For example, when a barometer in the flight control system exhibits an extremely small upward trend in its measurement due to icing, although the instantaneous value at a single point may be within the normal range, the accumulated delta will be amplified during the transpose modeling of the stationary channel, thereby achieving accurate prediction of slow drift faults.

[0051] In the non-stationary sequence processing step 4, to address the strong nonlinear coupling in aerodynamics, this embodiment extends the cross-attention mechanism with a multi-head cross-attention approach. Specifically, the query value, key value, and numerical value are divided into eight independent subspace heads, each learning the correlation between seasonal and trend terms at different frequency bands. For example, the first head might focus on low-frequency overload evolution, while the eighth head focuses on high-frequency oscillation residuals caused by structural aeroelastic effects. Through multi-head cross-attention, the model generates enhanced trend terms. It can contain richer, multi-dimensional physical constraint information, thereby effectively suppressing the divergence problem of predicted trajectories in long-term prediction caused by the chaotic characteristics of the system.

[0052] Regarding the anomaly prediction in step 5, this embodiment incorporates an anomaly pattern database containing a predetermined size of historical flight control test records. This database pre-stores temporal template features of over 20 typical flight control anomalies, including sensor step failures, actuator jamming, and control surface oscillation instability. After outputting the initial predicted trajectory, the system uses an additional Long Short-Term Memory (LSTM) network to perform a secondary evaluation of the predicted sequence. By calculating the cosine similarity between the predicted trajectory and known anomaly patterns in the database, an anomaly trigger probability score is calculated. When the prediction probability exceeds a preset alarm threshold of 0.8, the prediction system automatically locks onto the anomaly source and highlights the fault-related physical channels (such as the left aileron actuator feedback branch) in the graphical interface.

[0053] During the model training phase, this embodiment employs Weighted Mean Squared Error (MSE) as the loss function. Higher weights are assigned to training samples in abnormally volatile and rapid maneuvering segments to force the model to prioritize learning evolutionary characteristics under extreme conditions. The training process utilizes Stochastic Gradient Descent (SGD) for parameter updates, with an initial learning rate set to 1e-4 and a learning rate decay strategy. To prevent overfitting and ensure generalization performance under different flight control conditions, an early stopping mechanism is introduced. This means that when the prediction error on the validation set no longer decreases within 20 consecutive training epochs, iteration automatically stops and the optimal model parameters are saved. This training strategy ensures that the Chaoslinear framework maintains robust predictive capabilities even when faced with unseen experimental scenarios.

[0054] Example 3

[0055] Based on Embodiments 1 and 2, this embodiment describes in detail the system-level implementation and deployment scheme of the present invention in a specific flight control law verification task, with particular attention to the detection of hardware interface communication anomalies and control law software logic defects.

[0056] In this embodiment, the method is deployed in the core computing unit of the flight control test real-time monitoring system. For the data input in step 1, the system collects high-frequency communication messages between the flight control computer and the actuators in real time via ARINC429 or 1553B avionics bus. The raw messages are decoded into physical signal sequences and input to the multi-scale block data embedding layer. To handle signal glitches caused by electromagnetic interference during hardware communication, the embedding layer performs median filtering preprocessing before segmentation. The filtering window size is set to 3 sampling points to ensure that the data input to the dual-channel expert modeling architecture has a high signal-to-noise ratio.

[0057] In the dual-channel expert modeling architecture constructed in step 2, the stationary sequence channel specifically monitors the fixed-frequency heartbeat signal of bus communication. Since hardware interface communication anomalies often manifest as minor fluctuations in the heartbeat interval or cyclic redundancy check (CRC) errors in specific fields, the stationary sequence channel models the logical dependencies between data fields through transpose operations, enabling it to a priori capture the sub-health state of the communication link.

[0058] To address the non-stationary sequence decomposition in step 4, this embodiment introduces an adaptive window moving average algorithm. The window size is no longer fixed but dynamically adjusted based on the aircraft's current Mach number and angle of attack. In high-speed, high-dynamic-pressure environments, the flight control system's dynamic response is extremely fast, and the moving average window automatically shrinks to capture high-frequency trend evolution; during low-speed takeoff and landing, the window expands accordingly to extract a smoother attitude control trajectory. This dynamic adjustment mechanism allows the enhancement trend term construction process to better match the aircraft's real-time flight envelope.

[0059] In the prediction output stage of step 5, this embodiment not only outputs the time series of future signals but also generates a real-time entropy index to characterize the degree of determinism of the predicted trajectory. When the adaptive gating mechanism detects a surge in non-stationary components in the input data, and the prediction difference between stationary and non-stationary channels reaches a preset threshold, the real-time entropy value will increase significantly. The system uses this entropy value as the basis for the width of the uncertainty confidence interval, enveloping the predicted trajectory in the form of a shaded area on the graphical interface. By observing the changes in the shaded area, test personnel can intuitively determine whether the current flight control system has entered the edge region of the design envelope or whether there is an unexpected risk of control law divergence.

[0060] Furthermore, the method in this embodiment demonstrated extremely high recognition accuracy in closed-loop verification tasks on the Iron Bird Integrated Test Rig. In a specific test targeting control law software logic defects (such as control surface reversal caused by logic flipping), the method successfully triggered a warning 400 milliseconds before the actual fault occurred by analyzing the evolution of the nonlinear correlation between the pilot's input commands and the control surface response sequence. This warning not only prevented potential mechanical damage to the test rig but also accurately pointed to a specific branch in the logic decision tree through the secondary evaluation module. This closed-loop application from time-series prediction to fault location fully demonstrates the significant advantages of this invention in combating uncertainty by combining seasonal residual correction with dual-channel deep learning.

[0061] Example 4

[0062] This embodiment further illustrates the parallel processing and data alignment strategy of the flight control test anomaly time series data prediction method combining uncertainty and deep learning when dealing with large-scale, multi-source heterogeneous sensor data, aiming to improve the breadth and depth of flight control system monitoring.

[0063] During step 1, the input historical time-series samples include not only electrical signals but also wing structure strain data acquired by fiber Bragg grating sensors. For these heterogeneous variables, the multi-scale block data embedding layer employs heterogeneous feature alignment technology. In the block feature vector generation stage, different types of physical signals are first mapped to their respective independent latent spaces, and then projected onto the same dimension (dim) through a common fusion matrix. This approach enables the subsequent expert modeling architecture to understand the complex coupling between the rigid body motion parameters and the flexible structure response of the aircraft within a unified semantic space.

[0064] For stationary sequence extraction in step 3, this embodiment adds a group normalization layer to the multilayer perceptron structure. By independently normalizing interrelated flight parameters (such as three-axis angular velocities as one group and four groups of control surface deflections as another group), the influence of differences in sensor ranges on statistical feature extraction is eliminated. This allows the stationary channel to more sensitively detect small, cooperative drifts within a particular group of variables when calculating the mean and variance.

[0065] In step 4, this embodiment introduces a spectrum sensing mechanism for the seasonal term derived from the non-stationary sequence. Before capturing the impact of random perturbations on the trend term based on cross-attention, the seasonal term component is input into a Fast Fourier Transform (FFT) layer to extract its characteristic frequency distribution. The calculation of the attention weight coefficients considers not only the similarity in the time domain but also the energy concentration in the frequency domain. For example, when a periodic oscillation of a specific frequency appears in the seasonal term and the energy continues to increase, the attention mechanism will significantly increase the weight of the feature of that frequency band in the enhanced trend term, thereby enabling the final generated prediction sequence to accurately predict structural instability caused by aero-servoelasticity (ASE).

[0066] Regarding the dynamic feature fusion in step 5, this embodiment expands the input dimension of the adaptive gating mechanism. The update of the gating parameter matrix not only depends on the current feature vector Χ_{Flatten}, but also introduces contextual identifiers for flight mission phases (such as takeoff, climb, cruise, approach, and landing). In high-risk, high-dynamic phases such as takeoff and landing, the gating mechanism pre-adjusts the initial weight base of non-stationary channels to improve the model's sensitivity to step responses and abrupt anomalies; while in the high-altitude level flight phase, it places more emphasis on stationary channels to achieve fine monitoring of subtle signal distortions.

[0067] Finally, the method of this embodiment was validated in extensive Monte Carlo simulation experiments on a flight simulator. The experimental results show that, under complex conditions simulating intermittent sensor failures (such as discontinuous packet loss or signal jumps), the mean square error (MSE) of the Chaoslinear architecture is reduced by approximately 12% compared to the traditional Transformer model. Particularly when dealing with severe maneuver predictions exhibiting chaotic characteristics, the effective coverage time of the predicted trajectory is extended by approximately 35% due to the introduction of an enhanced trend term containing small residual corrections. This performance improvement provides valuable decision-making time for the emergency takeover mechanism of the flight control system in extreme failure modes, fully demonstrating the industrial application value and technological advancement of this invention in complex aerospace engineering environments.

[0068] Example 5

[0069] This embodiment details the online update and lifecycle prediction strategy of the flight control test anomaly time series data prediction method combining uncertainty and deep learning in processing long-endurance flight control test data, in order to cope with the long-term data distribution drift caused by sensor aging and slow performance degradation of actuators.

[0070] In the dual-channel expert modeling architecture established in step 2, this embodiment introduces a dynamic weight decay and parameter incremental update mechanism. For non-stationary sequence channels, the model can periodically perform online fine-tuning using recently acquired flight control test normal segment data. This means that the modeling ability of non-stationary channels for seasonal and trend terms can automatically adapt as the total flight time increases. For example, with the long-term wear of the actuator hydraulic system, its dynamic response characteristics will undergo subtle evolution. Through online updates, non-stationary channels can learn this slow, regular change caused by physical aging, thereby avoiding false alarms as sudden failures.

[0071] For the stationary sequence feature extraction in step 3, this embodiment introduces a steady-state boundary identification algorithm. When calculating statistical features, the system automatically marks data points deviating from the historical stationary benchmark by more than 3 standard deviations as outlier candidates. The correlation of the transpose operation modeling is no longer limited to the variable dimension but also extended to the dimension of historical operating conditions. By comparing the current stationary sequence prediction components with the "performance degradation trend template" recorded in the outlier pattern database, the model can output the predicted remaining useful life (RUL) value of the sensor or actuator.

[0072] In the non-stationary sequence decomposition process of step 4, to further improve the ability to handle chaotic perturbations, this embodiment employs second-order moving average decomposition. That is, the seasonal term obtained from the initial decomposition is smoothed again to separate pure white noise and structural noise with weak periodicity. The cross-attention mechanism only applies to the structural noise component with weak periodicity, thereby enhancing the generated trend term. It has higher prediction stability and eliminates prediction trajectory jitter caused by purely random measurement noise.

[0073] When the anomaly prediction results from step 5 are output through the graphical interface, a 3D visualization digital twin module is also integrated. The predicted future trajectory is mapped onto the 3D aircraft model in real time, with color depth representing the risk level of different flight envelope positions. When the prediction system anticipates a severe attitude divergence that may occur within the next 2 seconds, the interface will not only issue an alarm but also automatically recommend response strategies through adaptive gating mechanism channel weight allocation analysis, such as switching to backup sensor channels or activating degraded control mode.

[0074] The method described in this embodiment demonstrated outstanding performance in actual long-endurance reliability tests of large unmanned aerial vehicles (UAVs). By monitoring the entire lifecycle of the control surface actuators from normal operation to gradual performance degradation and eventual failure, the system provided accurate performance degradation warnings 15 minutes before the final failure occurred. The predicted abnormal trends were highly consistent with the physical wear observed during subsequent disassembly. This comprehensive monitoring architecture, which combines accurate short-term anomaly prediction with long-term performance degradation assessment, greatly enriches the content of flight control test data prediction methods and provides core technical support for aircraft condition-based maintenance (CBM) and health management (PHM).

[0075] Example 6

[0076] This embodiment describes the anti-interference performance and adaptive noise suppression of the flight control test anomaly time series data prediction method that combines uncertainty and deep learning under complex meteorological environments and strong external interference.

[0077] In the construction of the multi-scale segmented data embedding layer in step 1, this embodiment introduces a wavelet denoising preprocessing module to address the strong airflow disturbance data generated during flight control experiments in wind tunnel environments or complex high-altitude atmospheric environments. Before scale segmentation, the input multivariate time series is first decomposed into approximate and detail components using Discrete Wavelet Transform (DWT). High-frequency detail components are selectively removed or compressed to suppress high-order random noise generated by external gust disturbances. Subsequently, the reconstructed smoothed sequence enters the embedding layer to generate segmented feature vectors.

[0078] In step 3, when extracting stationary sequence features, this embodiment particularly enhances the sensitivity identification of control gain changes. A learnable scaling factor is added to the residual connection structure of the stationary channel. This factor can capture signal jumps caused by adaptive gain switching within the flight control law. Through transpose modeling, the model can distinguish between such normal control logic switching and anomalous sensor signal mutations.

[0079] In step 4, for the correlation modeling of non-stationary sequences, this embodiment introduces a physical prior weight constraint in the cross-attention mechanism. Specifically, the attention weight coefficient matrix is ​​masked based on the aerodynamic derivative matrix of the flight control system. For example, the interaction weights between longitudinal parameters (such as angle of attack and pitch rate) and lateral parameters (such as sideslip angle and roll rate) are set with low prior probabilities under normal flight conditions. When random disturbances in the seasonal term cause this cross-axis coupling to be abnormally enhanced, cross-attention can quickly capture unexpected correlations caused by the system entering a nonlinear unstable region or structural disintegration, thereby improving the targeting of the early warning.

[0080] The output of step 5 is further processed by a Kalman filter smoother to ensure that the predicted future physical signal satisfies the physical constraints of continuity and differentiability of the first derivative. This post-processing mechanism greatly reduces false alarms caused by instantaneous fluctuations in the prediction results of the deep learning model.

[0081] The method presented in this embodiment demonstrates exceptional robustness in simulations of flight control systems encountering lightning strikes and extreme turbulent environments. Even with extremely low signal-to-noise ratios, the method maintains an anomaly prediction accuracy exceeding 92%, with no significant delay in response time. By adjusting the uncertainty confidence interval in real time, the system can effectively identify sensor failure symptoms hidden within turbulent data against a background of strong noise. This superior disturbance rejection capability enables the proposed Chaoslinear framework to meet the monitoring requirements of flight control tests under all-weather, multi-condition conditions, providing robust data assurance for aviation testing missions in complex environments.

[0082] Example 7

[0083] This embodiment focuses on illustrating the distributed implementation of the flight control test anomaly timing data prediction method in large-scale cluster test data processing and the cloud-edge collaborative deployment scheme.

[0084] In step 1, when faced with the simultaneous acquisition of thousands of flight parameters, the task of constructing a multi-scale segmented data embedding layer is distributed across edge computing nodes. Each edge node is responsible for processing time-series data from a specific subsystem (such as the propulsion system subgroup or the navigation system subgroup), utilizing a distributed parallel computing framework to achieve high-speed generation of feature vectors. The generated feature vectors are then aggregated to the central server via a lightweight protocol for expert modeling and fusion prediction in steps 2 through 5.

[0085] For the non-stationary sequence decomposition in step 4, this embodiment employs a streaming decomposition technique, which can incrementally extract trend terms from the TB-level flight control historical data that comes in in real time. The moving average algorithm utilizes sliding window sum optimization during the calculation process, reducing the computational complexity from linear to constant level, thereby meeting the real-time requirements for processing massive flight control messages.

[0086] In step 5, the predicted flight control test anomaly time series data are shared through a cloud database. Prediction results from multiple test sites are aggregated into an anomaly pattern database, and the learning rate and weight parameter matrix of the adaptive gating mechanism are continuously updated using a swarm intelligence optimization algorithm. This allows the model to learn from various failure cases under different laboratory and climatic conditions, significantly improving its ability to predict rare anomaly patterns (corner cases).

[0087] Furthermore, this embodiment also introduces a natural language processing (NLP)-based interpreter into the prediction result display interface, which can automatically convert complex tensor prediction results into engineering terms that are easy for test personnel to understand. When the risk of abnormal flutter on the control surface is predicted, the system will automatically generate text descriptions such as "It is expected that the main servo feedback loop will experience a limiting loop oscillation with a frequency greater than 5 Hz in 3 seconds", which further reduces the technical entry threshold for flight control test monitoring.

[0088] Through its application in a multi-site joint scientific research test of a large transport aircraft, the distributed deployment scheme in this embodiment successfully achieved real-time predictive coverage of more than 5,000 monitoring points across the entire aircraft. The system's high concurrency processing capability and cloud-edge collaborative self-evolutionary characteristics mark a significant leap forward in flight control test data prediction technology from single-task, local processing to a systematic and intelligent approach.

[0089] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

[0090] 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 them; those skilled in the art can modify the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting anomalies in flight control experiment time series data by combining uncertainty and deep learning, characterized in that, The method includes the following steps: Step 1: Perform multi-scale segmentation processing on the input flight control test historical multivariate time series using a preset multi-scale segmented data embedding layer, divide the historical multivariate time series into segmented data with different time granularities, and use the embedding layer to map the segmented data to a high-dimensional latent space to generate segmented feature vectors that characterize the structural features of the time series. Step 2: Construct a dual-channel expert modeling architecture consisting of a stationary sequence multilayer perceptron channel and a non-stationary sequence multilayer perceptron channel, and input the block feature vector into the dual-channel expert modeling architecture to perform classification modeling on the stationary sequence distribution characteristics and non-stationary sequence distribution characteristics in the flight control test data; Step 3: In the stationary sequence multilayer perceptron channel, the time statistical characteristics of each flight parameter variable are calculated through a multilayer perceptron structure with residual connections, and a transpose operation is performed to model the nonlinear coupling correlation between each flight parameter variable to obtain the prediction component of the stationary sequence. Step 4: In the non-stationary sequence multilayer perceptron channel, the input sequence is decomposed into a trend term representing the long-term evolution trajectory and a seasonal term representing periodic perturbations by a moving average algorithm. The influence of random perturbation components in the seasonal term on the evolution of the trend term is captured based on the cross-attention mechanism to generate an enhanced trend term containing residual correction information. Step 5: Using an adaptive gating mechanism, based on the real-time dynamic distribution characteristics of the input sequence, dynamically calculate the weighted ratio between the output results of the stationary sequence multilayer perceptron channel and the output results of the non-stationary sequence multilayer perceptron channel, map the fused features to a preset prediction time dimension, and output the predicted trajectory of the abnormal time series data of the flight control test.

2. The method for predicting anomalies in flight control test time series data by combining uncertainty and deep learning as described in claim 1, characterized in that, The flight control test historical multivariate time series input in step 1 contains multiple interdependent flight parameter variables, which include three-axis angular rate, control surface deflection angle, flight overload, atmospheric data sensor pressure altitude, climb rate, sideslip angle, and angle of attack. The historical observation time step of the historical multivariate time series is set as a first preset duration, and the predicted future time step is set as a second preset duration.

3. The method for predicting anomalies in flight control test time series data by combining uncertainty and deep learning according to claim 1, characterized in that, The multi-scale segmentation process specifically involves: using multiple sliding windows of different scales to perform parallel segmentation of the input flight control test history multivariate time series; the sliding windows include short-time step windows for capturing local transient changes in sensor signals, medium-time step windows for extracting transition features during flight state switching, and long-time step windows for identifying long-term evolution trends of flight trajectories; the size of the sliding window at each scale is set to the corresponding preset number of time points; through multi-scale segmentation, the generated segmented data of different time granularities are transformed into a three-dimensional matrix containing variable dimensions, number of segments, and number of time points within each segment.

4. The method for predicting anomalies in flight control test time series data by combining uncertainty and deep learning as described in claim 1, characterized in that, The process of generating block feature vectors in step 1 includes: using the linear mapping layer in the embedding layer to encode the features of each block of data, transforming it into a high-dimensional space vector of a preset dimension; injecting positional encoding representing the chronological order information of the time series into the high-dimensional space vector; and using data reshaping instructions to perform a one-dimensional tiling operation on the multi-dimensional block features, converting them into a two-dimensional matrix form that meets the operational requirements of a multilayer perceptron.

5. The method for predicting anomalies in flight control test time series data by combining uncertainty and deep learning according to claim 1, characterized in that, Step 2 specifically includes: using the segmented feature vectors as common inputs, and guiding them to the stationary sequence multilayer perceptron channel and the non-stationary sequence multilayer perceptron channel respectively through data splitting logic; the stationary sequence multilayer perceptron channel is configured as a lightweight residual network structure, with its network depth limited to a preset number of layers, and is used to perform feature modeling on the statistically stable engine steady-state vibration signal or cruise phase signal; the non-stationary sequence multilayer perceptron channel is configured as a deep nonlinear transformation structure, and is used to perform feature modeling on the non-stationary features in the flight overload sequence or fault evolution process generated by maneuvering.

6. The method for predicting anomalies in flight control test time series data by combining uncertainty and deep learning according to claim 1, characterized in that, The process of calculating the time statistical characteristics of each flight parameter variable in step 3 includes: calculating the mean and variance of the input sequence tensor on the time step axis; injecting the mean and variance as potential features into the multilayer perceptron layer; comparing the mean and variance of the current block data with the benchmark statistics of the historical period to generate deviation features for identifying slow offset trend anomalies.

7. The method for predicting anomalies in flight control test time series data by combining uncertainty and deep learning according to claim 1, characterized in that, The process of modeling the nonlinear coupling correlation between various flight parameter variables in step 3 includes: transposing the matrix with learned time features to achieve the interchange of time dimension and variable dimension; using mutually independent linear mapping layers to extract the co-evolution mode between different sensor signals, including three-axis angular rate, control surface deflection angle and flight overload; performing the transpose operation again after completing the correlation modeling to restore the original data structure, and adding it element by element to the original input through residual connection.

8. The method for predicting anomalies in flight control test time series data by combining uncertainty and deep learning according to claim 1, characterized in that, The process of performing sequence decomposition using the moving average algorithm in step 4 includes: setting the window size of the moving average algorithm according to the dynamic response period of the flight control system; extracting the long-term evolution trajectory from the block feature vector through average pooling and defining it as the trend term; defining the residual part after subtracting the trend term from the block feature vector as the seasonal term; the seasonal term includes the periodic signal sampled by the sensor and the random disturbance generated by the wear of the actuator.

9. The method for predicting anomalies in flight control test time series data by combining uncertainty and deep learning according to claim 1, characterized in that, The process of capturing influence based on the cross-attention mechanism in step 4 includes: mapping the trend item to a query value through a weight matrix, and mapping the seasonal item to a key value and a value through the corresponding weight matrix; calculating the dot product of the query value and the key value transpose matrix, and performing normalization processing to obtain the attention weight coefficient; the attention weight coefficient characterizes the degree of contribution of each time component in the seasonal item to the evolution of the trend item.

10. The method for predicting time-series flight control experiment anomalies by combining uncertainty and deep learning according to claim 1, characterized in that, The process of generating the enhanced trend term containing residual correction information in step 4 includes: using the attention weight coefficient and the numerical value to perform a weighted summation to obtain the dynamic influence of the small residual in the seasonal term on the trend term; adding the dynamic influence to the original trend term in an additive form to form an enhanced trend term that can cope with uncertainty shifts; and performing independent linear mapping processing on the enhanced trend term and the seasonal term to extract the dramatic evolution characteristics under non-stationary conditions.

Images