Inplausible aircraft state sensing and predicting system and method

Abstract

The invention relates to the technical field of high-speed aircraft health management, and particularly discloses an interpretable aircraft state sensing and predicting system and method. The method comprises the following steps: constructing and training a lightweight anomaly diagnosis network with an expansion perception layer; aircraft dynamics, kinematics and fault mechanism formulas are used as soft constraints to be embedded into the future state estimation network; an SHAP interpretation framework is used for sorting the importance of the input features, and a network structure is optimized accordingly; and finally, accurate diagnosis of abnormal types and positions and rapid prediction of aircraft states at a plurality of moments in the future are realized. According to the method, by introducing three innovation points of an expansion perception layer, physical information embedding and feature importance sorting, the diagnosis accuracy and the model interpretability are remarkably improved, and the high-performance calculation requirement in a scene with limited calculation power is met.

Description

Technical Field

[0001] This invention relates to the field of high-speed aircraft health management technology, and in particular to an interpretable aircraft state perception and prediction system and method. Background Technology

[0002] High-speed aircraft operate under prolonged, high-temperature, high-pressure, and high-speed environments, making their servo motors, as core actuators, highly susceptible to nonlinear faults such as dead zones, saturation, jamming, loosening, and damage. These faults significantly alter the output torque of the control surfaces, leading to aircraft attitude instability and even catastrophic accidents. Traditional methods often rely on thermodynamic parameter thresholds for judgment, resulting in large fluctuations, high false alarm rates, and an inability to accurately pinpoint the type and extent of faults. Furthermore, existing data-driven models require significant airborne computing power, making them unsuitable for real-time onboard applications. Therefore, a lightweight, interpretable, and online-operating anomaly detection technology is urgently needed. Summary of the Invention

[0003] The present invention aims to provide an interpretable aircraft state perception and prediction system and method to solve the problems of insufficient diagnostic accuracy and heavy computational burden in the prior art, which are difficult to adapt to the limited computing power of aircraft.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] This invention provides an interpretable aircraft state perception and prediction system, applied to high-speed aircraft, comprising:

[0006] The actuator module takes the rudder deflection data from the previous moment and the current abnormal state diagnosis result as input, and after feature extraction and learning, outputs the rudder deflection data for the next moment; the rudder deflection data includes the rudder deflection values ​​of the left rudder, right rudder and rudder.

[0007] The aircraft parameter module extracts features from the aircraft parameters to obtain the aircraft state parameter features; the aircraft parameters include: position, angle of attack, sideslip angle, roll angle, pitch angle, angular velocity, velocity, mass, aerodynamic torque, aerodynamic lift, aerodynamic drag, velocity tilt angle, and track yaw angle;

[0008] The abnormal state diagnosis module takes the aircraft's state parameter characteristics and actuator parameters as input to predict the abnormal state of the aircraft; the prediction results include: abnormal state type, location and severity.

[0009] The future state estimation module takes the characteristics of the aircraft's state parameters, the predicted output of the actuator module, and the abnormal state diagnosis results as inputs to predict the aircraft parameters at the next moment.

[0010] The model optimization module calculates the importance ranking of all inputs based on the SHAP algorithm and performs redundancy optimization on the system inputs according to the ranking results.

[0011] Preferably, the actuator module uses a feature binning method to discretize the input rudder deflection data. The processed rudder deflection data and the current abnormal state diagnosis result are then processed through a feature extraction layer and a learning layer to output the rudder deflection data for the next time step.

[0012] The feature extraction layer includes a convolutional layer, a normalization layer, a pooling layer, and a regularization layer.

[0013] Preferably, the abnormal state diagnosis module adopts an abnormal state diagnosis network with multiple fast diagnosis layers stacked, and an extended perception layer is set at the front end;

[0014] The extended perception layer includes: a 1×1 convolutional layer responsible for information fusion and dimensionality reduction between channels, a 3×3 convolutional layer responsible for local spatiotemporal feature extraction, a 3×3 max pooling layer responsible for downsampling and suppressing high-frequency noise, and the three outputs are spliced ​​together in the channel dimension.

[0015] The concatenated feature maps are subjected to two layers of 1×1 convolution to achieve depthwise separable extraction, and then global average pooling is applied to compress the high-dimensional features into a one-dimensional vector.

[0016] Preferably, the depth of the abnormal state diagnosis network is controlled at the single-level.

[0017] Preferably, the data processing procedure of the future state estimation module is as follows:

[0018] The input aircraft parameters are processed through a feature extraction layer to obtain aircraft state parameter features. The input actuator parameters and abnormal state diagnosis results are processed through a feature extraction layer and a learning layer to obtain actuator prediction outputs. Both are then used to predict the aircraft parameters at the next moment through a long short-term memory recurrent neural network, an attention mechanism, and an output layer.

[0019] Preferably, the future state estimation module constructs a loss function for embedding physical information, as follows: , in, and These are the actual results and the predicted results, respectively. , , These are hyperparameters used to ensure that the values ​​of all loss terms are of the same order of magnitude. For trajectory constraint loss, For attitude constraint loss, This is the loss due to abnormal rudder deflection constraint.

[0020] Preferably, the trajectory constraint loss Represented as:

[0021] ,

[0022] in, , , express , , Constraints in three directions, For flight speed, It is the angle of inclination. For velocity deflection angle, , , The relaxation coefficient is... The harmonic coefficient;

[0023] The attitude constraint loss Represented as:

[0024] ,

[0025] in, , , for , , Constraints on the three attitude angles, , , The components of angular velocity are the x-axis, y-axis, and z-axis components. , , The relaxation coefficient;

[0026] The abnormal rudder deflection constraint loss Represented as:

[0027] ,

[0028] in, This represents the rudder deflection value at the current moment. The result is the prediction of the rudder offset for the next moment. These represent right rudder deflection, elevator deflection, and left rudder deflection, respectively.

[0029] Preferably, the model optimization module is based on the SHAP algorithm to quantify the marginal contribution of each input in each inference and obtain the importance ranking of the inputs;

[0030] Low-importance features below a threshold are fed into a fully connected network with a single hidden layer, compressed into a new fused feature, and then concatenated with other input features above the threshold to serve as the input feature.

[0031] This invention also provides an interpretable aircraft state perception and prediction method, implemented based on the aforementioned interpretable aircraft state perception and prediction system, the method comprising:

[0032] Pre-train the abnormal state diagnosis network and the future state estimation network;

[0033] Feature extraction is performed on the aircraft parameters to obtain the aircraft state parameter features; the aircraft parameters include: position, angle of attack, sideslip angle, roll angle, pitch angle, angular velocity, velocity, mass, aerodynamic torque, aerodynamic lift, aerodynamic drag, velocity tilt angle, and track yaw angle;

[0034] Using aircraft state parameter characteristics and actuator parameters as input, an abnormal state of the aircraft is predicted based on a pre-trained abnormal state diagnosis network; the prediction results include: abnormal state type, location and severity.

[0035] The system takes the rudder deflection data from the previous moment and the diagnostic results from the current abnormal state diagnostic network as input, and after feature extraction and learning, outputs the rudder deflection data for the next moment; the rudder deflection data includes the rudder deflection values ​​of the left rudder, right rudder and rudder.

[0036] Using the aircraft's state parameter characteristics, the predicted output of the rudder deflection data at the next moment, and the abnormal state diagnosis results as inputs, the aircraft parameters at several future moments are predicted based on a pre-trained future state estimation network.

[0037] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0038] This invention provides an interpretable aircraft state perception and prediction system, introducing three major innovations: an extended perception layer, physical information embedding, and SHAP interpretable optimization, to more accurately capture the characteristics of minor servo motor faults. To reduce false alarms and delays caused by traditional threshold monitoring, this invention monitors changes in multi-source time-series parameters in real time, infers the offset of servo motor geometric / dynamic characteristics using a physical constraint network, and identifies the type, location, and severity of anomalies online. Simultaneously, it predicts the aircraft state several moments in the future, providing millisecond-level, low-computing-power, and highly reliable technical assurance for the safe flight of high-speed aircraft. This invention uses the SHAP interpretation tool to score each input feature, pruning or compressing unimportant dimensions, resulting in a smaller and faster model; achieving millisecond-level online diagnosis and predicting flight state several steps in advance; and immediately triggering fault tolerance upon detecting a problem, ensuring the safety and controllability of high-speed aircraft under extreme conditions throughout the entire process. Attached Figure Description

[0039] Figure 1 A schematic diagram of the interpretable aircraft abnormal state perception and prediction system architecture provided by the present invention;

[0040] Figure 2 A schematic diagram of the abnormal state diagnosis module provided by the present invention;

[0041] Figure 3 A schematic diagram of the extended sensing layer structure provided by the present invention;

[0042] Figure 4 A schematic diagram of the future state estimation module provided by this invention;

[0043] Figure 5 This is a schematic diagram of the experimental results of control group 2 in this embodiment of the invention;

[0044] Figure 6 This is a schematic diagram of the thermogram experimental results of control group 3 in an embodiment of the present invention. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the embodiments and accompanying drawings. Here, the illustrative embodiments and descriptions of this invention are used to explain the invention, but are not intended to limit the invention.

[0046] It should also be noted that, in order to avoid obscuring the invention with unnecessary details, only the structures and / or processing steps closely related to the solution according to the invention are shown in the accompanying drawings, while other details that are not closely related to the invention are omitted.

[0047] It should be emphasized that the term "including / comprises" as used herein refers to the presence of a feature, element, step, or component, but does not exclude the presence or addition of one or more other features, elements, steps, or components.

[0048] It should also be noted that, unless otherwise specified, the term "connection" in this article can refer not only to a direct connection, but also to an indirect connection involving an intermediary.

[0049] In the following description, embodiments of the invention will be illustrated with reference to the accompanying drawings. In the drawings, the same reference numerals represent the same or similar parts, or the same or similar steps.

[0050] It should be emphasized here that the step markers mentioned below are not a limitation on the order of the steps, but should be understood as meaning that the steps can be executed in the order mentioned in the embodiments, or in a different order than in the embodiments, or several steps can be executed simultaneously.

[0051] To address the issues of insufficient accuracy in servo status diagnosis and difficulties in airborne deployment in existing technologies for high-speed aircraft, this invention proposes to first characterize typical anomalies that may occur in high-speed aircraft servos, such as "dead zones, saturation, jamming, looseness, and defects," into simulateable data using mathematical models. Then, this data is used to train a "lightweight yet intelligent" neural network—a front-end layer of "expanded perception" multi-scale convolution cleverly expands the receptive field without increasing its workload; the middle stage incorporates the aircraft's dynamics and kinematic equations into the loss function, allowing the network to learn from the data while adhering to physics; the back-end uses the SHAP (SHapley Additive exPlanations) interpretation tool to score each input feature, pruning or compressing unimportant dimensions, resulting in a smaller and faster model. This achieves millisecond-level online diagnosis and can predict flight status several steps in advance. Once a problem is detected, fault tolerance is immediately triggered, ensuring the safety and controllability of the high-speed aircraft under extreme conditions throughout the entire process.

[0052] Based on the above inventive concept, the present invention provides an interpretable aircraft state perception and prediction system, see [link to relevant documentation]. Figure 1 Mainly includes:

[0053] The actuator module takes the previous moment's rudder deflection data and the current abnormal state diagnosis result as input, passes through the feature extraction layer and the learning layer, and outputs the learned rudder deflection value for the next moment.

[0054] The aircraft parameter module extracts features from the aircraft parameters and fuses them with the rudder deflection values ​​obtained from the actuator module for the next moment. The results are then input into a long short-term memory recurrent neural network. The attention mechanism module and the output layer predict the final aircraft parameters for the next moment through the aircraft motion parameters. The aircraft parameters include: position, angle of attack, sideslip angle, roll angle, pitch angle, angular velocity, velocity, aerodynamic torque, lift, drag, left rudder deflection, right rudder deflection, elevator deflection, and abnormal status of the actuators.

[0055] The abnormal state diagnosis module adopts a multi-layer fast diagnosis structure, aiming to improve network performance while reducing parameters. Naturally, sparsification constraints are considered to be added to the fully connected layer. However, in reality, the reduction in computation time after sparsification is not significant. Therefore, a simplified extended perception layer network design is proposed, which utilizes the feature extraction capabilities of convolutional and pooling layers to obtain more comprehensive abnormal state perception results through different scales and perspectives.

[0056] The future state estimation module introduces constraints such as physical laws when estimating the future state of the system, adds a block layer and attention mechanism, and analyzes the state parameters of the aircraft and the state parameters of the actuators to obtain the change of the aircraft state at the next moment.

[0057] In the model optimization module, higher performance metrics are often required during model training, especially in the field of weaponry and equipment where high reliability is required. The SHAP method is a game theory approach that can explain the output of a machine learning model through the calculation of Shapley values. By using the classic Shapley value method in game theory, it describes the impact of input on output, thereby quantifying the local interpretation of the model.

[0058] The high-speed aircraft has three actuators: left rudder, right rudder, and steering rudder. The impact of abnormal states on future state estimation is also considered. Therefore, the actuator module of this invention takes as input the rudder offset values ​​of the three rudders and the diagnostic results from the abnormal state diagnosis network. The actuator module includes a feature extraction layer composed of convolutional layers, normalization layers, pooling layers, and regularization layers. This layer effectively highlights the features of the parameter data and performs data feature reconstruction.

[0059] Furthermore, the actuator module preprocesses the rudder deflection data using a feature binning method, setting the binning interval to [-30, 30] with each interval being 0.1, resulting in a total of 600 intervals. This method discretizes the rudder deflection data, which is more conducive to the understanding of the data-driven model. At the same time, it allows the vector to fluctuate within a small range, which can improve robustness and the current state diagnosis results. After passing through the feature extraction layer and the learning layer, the system outputs the rudder deflection value for the next time step.

[0060] See Figure 2 The abnormal state diagnosis module of this invention employs an abnormal state diagnosis network consisting of multiple stacked rapid diagnosis layers, aiming to improve network performance while reducing parameters. An "extended perception layer" is set at the network front end. For example... Figure 3 As shown, this layer consists of three parallel paths: a 1×1 convolution for information fusion and dimensionality reduction between channels, a 3×3 convolution for detailed characterization of local spatiotemporal features, and a 3×3 max pooling for downsampling and suppressing high-frequency noise. The three outputs are concatenated along the channel dimension, preserving the original details while gaining a broader contextual view, effectively amplifying even weak signals generated by tiny dead zones, saturation, or cracks in the control surfaces. The concatenated feature map then undergoes two more 1×1 convolutions for depth-separable extraction, followed by global average pooling to compress high-dimensional features into a one-dimensional vector. The entire network depth is controlled at single-level depth, and the convolution kernel size and pooling window are kept lightweight, ensuring stable operation on airborne embedded devices without additional heat dissipation.

[0061] See Figure 4The future state estimation module of this invention obtains aircraft state parameter features by inputting aircraft parameters through a feature extraction layer, and obtains actuator prediction outputs by inputting actuator parameters and anomaly diagnosis results through a feature extraction layer and a learning layer. Both are used to predict the flight state changes at the next moment through a long short-term memory recurrent neural network, an attention mechanism, and an output layer, thereby obtaining aircraft parameters such as position, angle of attack, sideslip angle, roll angle, pitch angle, angular velocity, velocity, aerodynamic torque, lift, drag, left rudder deflection, right rudder deflection, elevator deflection, and actuator abnormal states at the final next moment.

[0062] Furthermore, in the future state estimation module, aircraft dynamics, kinematics, and fault mechanism formulas are embedded as soft constraints into the future state estimation network, forming a loss function that embeds physical information. Based on the fault features extracted above, this embodiment no longer allows the network to be "purely data-driven," but instead incorporates the inherent kinematic and dynamic laws of the aircraft into the loss function, forming a "data-mechanism" fusion learning framework. Specifically, trajectory, attitude, and rudder deflection fault constraints are established, as follows:

[0063] Trajectory constraint equations:

[0064] ,

[0065] The trajectory constraint equations are derived from the kinematic mechanism equations. Among them, , , The corresponding model is , , Constraints in three directions, For flight speed, The angle of inclination, For velocity deflection angle, , , The relaxation coefficient, taking into account the effects of system uncertainties, allows the model learning results to deviate from the constraints of the mechanism formula within a small range. Meanwhile, since the fundamental dynamic equations cannot be violated, it is desirable for the relaxation coefficient to be as close to 1 as possible. In the trajectory constraint equation, the part to the right of the plus sign constrains the relaxation coefficient to penalize model predictions that deviate significantly from the mechanism process. This is the harmonic coefficient, used to harmonize the calculation results on both sides. Generally, it is desirable for the harmonic coefficient to be infinitely close to 0. For trajectory constraint loss, by , , The result is obtained by directly adding the three constraint terms.

[0066] Attitude constraint equations:

[0067] ,

[0068] The attitude constraint equations are derived from the dynamic mechanism equations. Among them, , , The corresponding model is , , Constraints on the three attitude angles, , , The components of angular velocity are the x-axis, y-axis, and z-axis components. , , is the relaxation coefficient. The attitude constraint loss is borne by , , The three constraint terms are simply added together to obtain the result.

[0069] Rudder deflection fault constraints:

[0070] ,

[0071] in, This represents the rudder deflection value at the current moment. This is the next time step rudder offset estimate, which includes the abnormal diagnosis results. These represent right rudder deflection, elevator deflection, and left rudder deflection, respectively. This represents the result of abnormal rudder deflection constraints.

[0072] The trajectory constraint, attitude constraint, and abnormal rudder deflection constraint modules introduce additional penalty terms to penalize model learning results that do not conform to the decision rules. Under trajectory constraint, attitude constraint, and abnormal rudder deflection constraint, the actual results... Compared with model prediction results The mean squared error loss function of the model can be written as:

[0073] ,

[0074] The loss function described above comprises four parts: model prediction loss, trajectory constraint loss, attitude constraint loss, and abnormal rudder deviation constraint loss. , , This is a hyperparameter used to ensure that the values ​​of all loss terms have the same order of magnitude. In the experiment, we take... , , The model continuously optimizes its internal parameters during training to minimize the loss function.

[0075] The model optimization module of this invention uses the SHAP method to calculate the importance ranking of input features and performs pruning or feature fusion optimization on the input of the entire system based on the ranking results. The inputs of this invention include aircraft state parameters and actuator parameters, specifically including: position x-axis component, position y-axis component, position z-axis component, angle of attack, sideslip angle, roll angle, pitch angle, roll angle, angular velocity x-axis component, angular velocity y-axis component, angular velocity z-axis component, velocity, mass, aerodynamic torque x-axis component, aerodynamic torque y-axis component, aerodynamic torque z-axis component, aerodynamic drag, aerodynamic lift, aerodynamic side force, velocity tilt angle, and track yaw angle.

[0076] After training with the aforementioned physical information constraints, although the system achieves high accuracy, it still suffers from the potential for redundant input dimensions and an excessive number of parameters. To address this, this embodiment introduces a SHAP-based interpretation framework to quantify the marginal contribution of each input feature in every inference step. Through global aggregation and local visualization, the system generates an intuitive heatmap: darker colors indicate a greater impact of the feature on anomaly diagnosis or future state prediction, while lighter colors mean the feature participates almost entirely in the decision-making process. After obtaining the importance ranking, two parallel lightweight strategies are employed: the first strategy directly discards low-importance features below a threshold, simplifying the input dimension to a "key subset," retraining on the validation set, and monitoring for accuracy decline. This strategy directly reduces the amount of computational data and model complexity. The second strategy does not completely discard low-importance features but instead feeds them into a single-hidden-layer fully connected network, compressing them into a new fusion feature, which is then concatenated with other high-importance features. This strategy retains information while significantly reducing the number of parameters.

[0077] In one embodiment, we constructed a high-speed aircraft abnormal state perception model based on an interpretable neural network and trained and validated the model using a dataset generated by a six-degree-of-freedom simulation platform. This dataset covers various fault types, including dead zone, saturation, jamming, loose float, and defects, as well as a fault-free complete control prototype, ensuring the comprehensiveness and robustness of the model training. Furthermore, we detailed 25-dimensional aircraft parameters, including x-axis, y-axis, and z-axis components of position, angle of attack, sideslip angle, roll angle, pitch angle, roll angle, x-axis, y-axis, and z-axis components of angular velocity, velocity, mass, x-axis, y-axis, and z-axis components of aerodynamic torque, aerodynamic drag, aerodynamic lift, aerodynamic side force, velocity tilt angle, yaw angle, left rudder deflection, right rudder deflection, elevator deflection, and abnormal state information, providing comprehensive input information for the model.

[0078] Control group 1: Expansion of the sensory layer (with or without contrast)

[0079] Control group 1 primarily compares the anomaly diagnosis results of the network with and without an extended sensing layer on the same test dataset. As shown in Table 1, model 1-1 incorporates an extended sensing layer. Table 2 presents the evaluation metrics for the anomaly diagnosis results of model 1-1 on the test dataset, and Table 3 presents the evaluation metrics for the anomaly diagnosis results of model 1-2 on the test dataset. Tables 2 and 3 show that the model with the extended sensing layer significantly outperforms the model without it in both accuracy and generalization ability. This indicates that the extended sensing layer can effectively extract multi-scale features and enhance the model's ability to identify different fault types.

[0080] Table 1. Control Group 1 Model

[0081]

[0082] Table 2 Evaluation metrics for anomaly diagnosis results of Model 1-1 on the test dataset

[0083]

[0084] Table 3 Evaluation metrics for anomaly diagnosis results of Models 1-2 on the test dataset.

[0085]

[0086] Control group 2: Physical information embedding comparison,

[0087] Control group 2 mainly compares the accuracy and running efficiency of models 2-1 to 2-4. As shown in Table 4, model 2-1 is a commonly used long short-term memory neural network method for processing time series data; model 2-2 is a commonly used convolutional neural network method for feature extraction; model 2-3 first uses a feature extraction network to extract data features, and then uses a long short-term memory neural network method for time series prediction; model 2-4 incorporates physical information embedding from aircraft dynamics, kinematics, and fault constraint modules into the neural network.

[0088] like Figure 5 As shown in the performance evaluation convergence curve and absolute error curve, the physical information embedding network outperforms other models in both accuracy and computational efficiency. This indicates that embedding aircraft dynamics, kinematics, and fault mechanism formulas as soft constraints into the network can significantly improve the model's prediction accuracy and computational efficiency.

[0089] Table 4. Control Group 2 Model

[0090]

[0091] Control Group 3: Comparison of importance ranking optimization

[0092] In control group 3, two different feature optimization methods were designed for validation, as shown in Table 5. Model 3-2 used feature optimization method 1, which directly discarded the five least important inputs, reducing the network input dimension by five dimensions, and then performed new training. Model 3-3 used feature optimization method 2, which connected the five least important inputs to a fully connected layer, integrating them into a one-dimensional input, and then used this data along with other input data for subsequent training.

[0093] Table 5. Control Group 3 Model

[0094]

[0095] like Figure 6 The results show that feature optimization method 2 significantly reduces model complexity and inference latency while maintaining accuracy. This further demonstrates that optimization driven by the SHAP method can effectively improve the computational efficiency and interpretability of the model.

[0096] Based on these results, we can conclude that optimization driven by the SHAP method can effectively improve the computational efficiency and interpretability of the model, providing strong technical support for the safe flight of high-speed aircraft. It not only improves the performance of the model but also enhances its interpretability, making the model's decision-making process more transparent and credible.

[0097] Based on the same inventive concept, the present invention also provides an interpretable aircraft state perception and prediction method, implemented based on the interpretable aircraft state perception and prediction system of the above embodiments, the method comprising:

[0098] Pre-train the abnormal state diagnosis network and the future state estimation network;

[0099] Feature extraction is performed on the aircraft parameters to obtain the aircraft state parameter features; the aircraft parameters include: position, angle of attack, sideslip angle, roll angle, pitch angle, angular velocity, velocity, mass, aerodynamic torque, aerodynamic lift, aerodynamic drag, velocity tilt angle, and track yaw angle;

[0100] Using aircraft state parameter characteristics and actuator parameters as input, an abnormal state of the aircraft is predicted based on a pre-trained abnormal state diagnosis network; the prediction results include: abnormal state type, location and severity.

[0101] The system takes the rudder deflection data from the previous moment and the diagnostic results from the current abnormal state diagnostic network as input, and after feature extraction and learning, outputs the rudder deflection data for the next moment; the rudder deflection data includes the rudder deflection values ​​of the left rudder, right rudder and rudder.

[0102] Using the aircraft's state parameter characteristics, the predicted output of the rudder deflection data at the next moment, and the abnormal state diagnosis results as inputs, the aircraft parameters at several future moments are predicted based on a pre-trained future state estimation network.

[0103] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0104] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0105] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0106] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0107] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. An interpretable aircraft state perception and prediction system, applied to high-speed aircraft, characterized in that, include: The actuator module takes the rudder deflection data from the previous moment and the current abnormal state diagnosis result as input, and after feature extraction and learning, outputs the rudder deflection data for the next moment; the rudder deflection data includes the rudder deflection values ​​of the left rudder, right rudder and rudder. The aircraft parameter module extracts features from the aircraft parameters to obtain the aircraft state parameter features; The aircraft parameters include: position, angle of attack, sideslip angle, roll angle, pitch angle, angular velocity, velocity, mass, aerodynamic torque, aerodynamic lift, aerodynamic drag, velocity tilt angle, and track yaw angle; The abnormal state diagnosis module takes the aircraft's state parameter characteristics and actuator parameters as input to predict the abnormal state of the aircraft; the prediction results include: abnormal state type, location and severity. The future state estimation module takes the characteristics of the aircraft's state parameters, the predicted output of the actuator module, and the abnormal state diagnosis results as inputs to predict the aircraft parameters at the next moment. The model optimization module calculates the importance ranking of all inputs based on the SHAP algorithm and performs redundancy optimization on the system inputs according to the ranking results.

2. The interpretable aircraft state perception and prediction system according to claim 1, characterized in that, The actuator module uses a feature binning method to discretize the input rudder deflection data. The processed rudder deflection data and the current abnormal state diagnosis result are then processed through a feature extraction layer and a learning layer to output the rudder deflection data for the next moment. The feature extraction layer includes a convolutional layer, a normalization layer, a pooling layer, and a regularization layer.

3. The interpretable aircraft state perception and prediction system according to claim 1, characterized in that, The abnormal state diagnosis module adopts an abnormal state diagnosis network with multiple fast diagnosis layers stacked together, and sets an extended perception layer at the front end. The extended perception layer includes: a 1×1 convolutional layer responsible for information fusion and dimensionality reduction between channels, a 3×3 convolutional layer responsible for local spatiotemporal feature extraction, a 3×3 max pooling layer responsible for downsampling and suppressing high-frequency noise, and the three outputs are spliced ​​together in the channel dimension. The concatenated feature maps are subjected to two layers of 1×1 convolution to achieve depthwise separable extraction, and then global average pooling is applied to compress the high-dimensional features into a one-dimensional vector.

4. The interpretable aircraft state perception and prediction system according to claim 3, characterized in that, The depth of the abnormal state diagnosis network is controlled at the single-level.

5. The interpretable aircraft state perception and prediction system according to claim 1, characterized in that, The data processing procedure of the future state estimation module is as follows: The input aircraft parameters are processed through a feature extraction layer to obtain aircraft state parameter features. The input actuator parameters and abnormal state diagnosis results are processed through a feature extraction layer and a learning layer to obtain actuator prediction outputs. Both are then used to predict the aircraft parameters at the next moment through a long short-term memory recurrent neural network, an attention mechanism, and an output layer.

6. The interpretable aircraft state perception and prediction system according to claim 5, characterized in that, The future state estimation module constructs a loss function for embedding physical information, as follows: ， in, and These are the actual results and the predicted results, respectively. , , These are hyperparameters used to ensure that the values ​​of all loss terms are of the same order of magnitude. For trajectory constraint loss, For attitude constraint loss, This is the loss due to abnormal rudder deflection constraint.

7. The interpretable aircraft state perception and prediction system according to claim 6, characterized in that, The trajectory constraint loss Represented as: ， in, , , express , , Constraints in three directions, For flight speed, It is the angle of inclination. For velocity deflection angle, , , The relaxation coefficient is... The harmonic coefficient; The attitude constraint loss Represented as: ， in, , , for , , Constraints on the three attitude angles, , , The components of angular velocity are the x-axis, y-axis, and z-axis components. , , The relaxation coefficient; The abnormal rudder deflection constraint loss Represented as: ， in, This represents the rudder deflection value at the current moment. The result is the prediction of the rudder offset for the next moment. These represent right rudder deflection, elevator deflection, and left rudder deflection, respectively.

8. The interpretable aircraft state perception and prediction system according to claim 1, characterized in that, The model optimization module is based on the SHAP algorithm, which quantifies the marginal contribution of each input in each inference and obtains the importance ranking of the inputs. Low-importance features below a threshold are fed into a fully connected network with a single hidden layer, compressed into a new fused feature, and then concatenated with other input features above the threshold to serve as the input feature.

9. An interpretable method for aircraft state perception and prediction, characterized in that, Based on the interpretable aircraft state perception and prediction system of claim 1, the method includes: Pre-train the abnormal state diagnosis network and the future state estimation network; Feature extraction is performed on the aircraft parameters to obtain the aircraft state parameter features; the aircraft parameters include: position, angle of attack, sideslip angle, roll angle, pitch angle, angular velocity, velocity, mass, aerodynamic torque, aerodynamic lift, aerodynamic drag, velocity tilt angle, and track yaw angle; Using aircraft state parameter characteristics and actuator parameters as input, an abnormal state of the aircraft is predicted based on a pre-trained abnormal state diagnosis network; the prediction results include: abnormal state type, location and severity. The system takes the rudder deflection data from the previous moment and the diagnostic results from the current abnormal state diagnostic network as input, and after feature extraction and learning, outputs the rudder deflection data for the next moment; the rudder deflection data includes the rudder deflection values ​​of the left rudder, right rudder and rudder. Using the aircraft's state parameter characteristics, the predicted output of the rudder deflection data at the next moment, and the abnormal state diagnosis results as inputs, the aircraft parameters at several future moments are predicted based on a pre-trained future state estimation network.

Images