CN-121997452-A - Interpretable aircraft state sensing and predicting system and method

CN121997452ACN 121997452 ACN121997452 ACN 121997452ACN-121997452-A

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 steps of constructing and training a lightweight anomaly diagnosis network with an expansion sensing layer, embedding dynamics, kinematics and failure mechanism formulas of an aircraft into a future state estimation network as soft constraints, sequencing importance of input features by using a SHAP interpretation framework, optimizing a network structure according to the importance, and finally realizing accurate diagnosis of anomaly types and positions and rapid prediction of the state of the aircraft at a plurality of moments in the future. According to the invention, three innovative points of an expansion sensing layer, physical information embedding and feature importance sequencing are introduced, so that the diagnosis accuracy and the model interpretability are remarkably improved, and the high-performance calculation requirement under a limited calculation force scene is met.

Inventors

ZHOU DENGJI
WU YADONG
HUANG DAWEN
WANG CHEN
XU LINFENG

Assignees

上海交通大学

Dates

Publication Date: 20260508
Application Date: 20251226

Claims (9)

1. An interpretable aircraft state awareness and prediction system for use with a high speed aircraft, comprising: The execution mechanism module takes rudder deflection data at the previous moment and the current abnormal state diagnosis result as inputs, and outputs the rudder deflection data at the next moment after feature extraction and learning, wherein the rudder deflection data comprises rudder deflection values of a left rudder, a right rudder and a rudder; The aircraft parameter module is used for extracting characteristics of aircraft parameters to obtain aircraft state parameter characteristics, wherein the aircraft parameters comprise position, attack angle, sideslip angle, roll angle, pitch angle, angular speed, mass, aerodynamic moment, aerodynamic lift, aerodynamic resistance, speed dip angle and track yaw angle; the abnormal state diagnosis module predicts the abnormal state of the aircraft by taking the characteristic of the state parameter of the aircraft and the parameter of the executing mechanism as inputs, wherein the prediction result comprises the type, the position and the severity of the abnormal state; The future state estimation module predicts the aircraft parameters at the next moment by taking the aircraft state parameter characteristics, the prediction output of the execution mechanism module and the abnormal state diagnosis result as inputs; and the model optimization module is used for calculating importance sequences of all the inputs based on the SHAP algorithm and performing redundancy optimization on the system inputs according to the sequence results.
2. The system for sensing and predicting the state of an aircraft according to claim 1, wherein the execution mechanism module adopts a characteristic binning method to discretize the input rudder deflection data, and the processed rudder deflection data and the current abnormal state diagnosis result pass through a characteristic extraction layer and a learning layer to output rudder deflection data at the next moment of learning; the feature extraction layer includes a convolution layer, a normalization layer, a pooling layer, and a regularization layer.
3. The system of claim 1, wherein the anomaly diagnostic module employs an anomaly diagnostic network comprising a plurality of rapid diagnostic layers stacked together and has an expansion awareness layer at the front end; The expansion perception layer comprises a1 multiplied by 1 convolution layer, a3 multiplied by 3 max pooling layer and a three-way output layer, wherein the 1 multiplied by 1 convolution layer is responsible for information fusion and dimension reduction among channels, the 3 multiplied by 3 convolution layer is responsible for local space-time characteristic extraction, the 3 multiplied by 3 max pooling layer is responsible for downsampling and inhibiting high-frequency noise, and the three output layers are spliced in the dimension of the channels; The spliced feature map is subjected to two-layer 1×1 convolution to finish depth separable extraction, and then global average pooling is accessed to compress the high-dimensional features into one-dimensional vectors.
4. An interpretable aircraft state awareness and prediction system according to claim 3, wherein the anomaly diagnostic network depth is controlled at a bit level.
5. An interpretable aircraft state awareness and prediction system according to claim 1, wherein the future state estimation module data processing is as follows: The input aircraft parameters obtain aircraft state parameter characteristics through a characteristic extraction layer, the input execution mechanism parameters and abnormal state diagnosis results obtain execution mechanism prediction output through a characteristic extraction layer and a learning layer, and the input execution mechanism parameters and the abnormal state diagnosis results predict the aircraft parameters at the next moment through a long-period memory recurrent neural network, an attention mechanism and an output layer.
6. An interpretable aircraft state awareness and prediction system according to claim 5 wherein the future state estimation module constructs a loss function for physical information embedding as follows: , Wherein, the And The actual result and the predicted result are respectively, , , Is a super parameter, is used to ensure that the values of the various penalty terms are of the same order of magnitude, In order to achieve the loss of trajectory constraint, In order to account for the loss of attitude constraints, Is an abnormal rudder deflection constraint loss.
7. An interpretable aircraft state awareness and prediction system according to claim 6 wherein the trajectory constraint loss is Expressed as: , Wherein, the 、、 Representation of 、、 The constraint terms in the three directions are, In order to achieve a speed of flight, In order to be the inclination angle, In order to be the speed deflection angle, the speed deflection angle is, 、、 In order to relax the coefficient of the light, Is a harmonic coefficient; The attitude constraint loss Expressed as: , Wherein, the , , Is that , , The constraint terms on the three attitude angles, , , For the angular velocity x-axis, y-axis and z-axis components, , , Is the relaxation coefficient; The abnormal rudder deflection constraint loss Expressed as: , Wherein, the As the current rudder deflection value, For the next rudder deflection value prediction result, Representing the right rudder deflection, the elevator deflection and the left rudder deflection, respectively.
8. An interpretable aircraft state awareness and prediction system according to claim 1 wherein the model optimization module quantifies the marginal contribution of each input in each thrust based on the SHAP algorithm to arrive at a ranking of importance of the inputs; And sending the low-importance features below the threshold value into a fully connected network with a single hidden layer, compressing the fully connected network into a new fusion feature, and splicing the new fusion feature with other input features above the threshold value to jointly serve as the input features.
9. A method of interpretable aircraft state awareness and prediction, implemented based on the interpretable aircraft state awareness and prediction system of claim 1, the method comprising: pre-training an abnormal state diagnosis network and a future state estimation network; Extracting characteristics of aircraft parameters to obtain aircraft state parameter characteristics, wherein the aircraft parameters comprise position, attack angle, sideslip angle, roll angle, pitch angle, angular velocity, speed, mass, aerodynamic moment, aerodynamic lift, aerodynamic resistance, speed dip angle and track yaw angle; Taking the aircraft state parameter characteristics and the actuator parameters as inputs, and predicting the aircraft abnormal state based on a pre-trained abnormal state diagnosis network, wherein the prediction result comprises the type, the position and the severity of the abnormal state; Taking the rudder deflection data at the previous moment and the diagnosis result of the current abnormal state diagnosis network as inputs, and outputting the rudder deflection data at the next moment after feature extraction and learning, wherein the rudder deflection data comprises rudder deflection values of a left rudder, a right rudder and a rudder; The aircraft state parameter characteristics, rudder deflection data prediction output at the next moment and an abnormal state diagnosis result are taken as inputs, and the aircraft parameters at a plurality of moments in the future are predicted based on a pre-trained future state estimation network.

Description

Interpretable aircraft state sensing and predicting system and method Technical Field The invention relates to the technical field of high-speed aircraft health management, in particular to an interpretable aircraft state sensing and predicting system and method. Background The high-speed aircraft operates in a long-time, high-temperature, high-pressure and high-rotation-speed environment, and nonlinear faults such as dead zones, saturation, blocking, loosening and floating, defects and the like are very easy to occur when the steering engine is used as a core executing mechanism. These faults can significantly change the control surface output torque, thereby causing the attitude of the aircraft to be unstable and even causing catastrophic accidents. The traditional method is based on the judgment of thermodynamic parameter threshold values, is large in fluctuation and high in false alarm, cannot accurately position the type and degree of faults, and meanwhile, the existing data driving model is high in onboard calculation force requirement and is difficult to meet the real-time missile-borne requirement. Thus, there is a need for a lightweight, interpretable, on-line, abnormal state awareness technique. Disclosure of Invention The invention aims to provide an interpretable aircraft state sensing and predicting system and method, which solve the problems that in the prior art, the diagnosis precision is insufficient, the calculation burden is heavy, the aircraft limited calculation force condition is difficult to adapt, and the like. In order to achieve the above purpose, the present invention provides the following technical solutions: The invention provides an interpretable aircraft state sensing and predicting system, which is applied to a high-speed aircraft and comprises: The execution mechanism module takes rudder deflection data at the previous moment and the current abnormal state diagnosis result as inputs, and outputs the rudder deflection data at the next moment after feature extraction and learning, wherein the rudder deflection data comprises rudder deflection values of a left rudder, a right rudder and a rudder; The aircraft parameter module is used for extracting characteristics of aircraft parameters to obtain aircraft state parameter characteristics, wherein the aircraft parameters comprise position, attack angle, sideslip angle, roll angle, pitch angle, angular speed, mass, aerodynamic moment, aerodynamic lift, aerodynamic resistance, speed dip angle and track yaw angle; the abnormal state diagnosis module predicts the abnormal state of the aircraft by taking the characteristic of the state parameter of the aircraft and the parameter of the executing mechanism as inputs, wherein the prediction result comprises the type, the position and the severity of the abnormal state; The future state estimation module predicts the aircraft parameters at the next moment by taking the aircraft state parameter characteristics, the prediction output of the execution mechanism module and the abnormal state diagnosis result as inputs; and the model optimization module is used for calculating importance sequences of all the inputs based on the SHAP algorithm and performing redundancy optimization on the system inputs according to the sequence results. Preferably, the executing mechanism module adopts a characteristic box division method to discretize the input rudder deflection data, and the processed rudder deflection data and the current abnormal state diagnosis result pass through a characteristic extraction layer and a learning layer to output rudder deflection data at the next moment of learning; the feature extraction layer includes a convolution layer, a normalization layer, a pooling layer, and a regularization layer. Preferably, the abnormal state diagnosis module adopts an abnormal state diagnosis network with a plurality of rapid diagnosis layers stacked, and an expansion sensing layer is arranged at the front end; The expansion perception layer comprises a1 multiplied by 1 convolution layer, a3 multiplied by 3 max pooling layer and a three-way output layer, wherein the 1 multiplied by 1 convolution layer is responsible for information fusion and dimension reduction among channels, the 3 multiplied by 3 convolution layer is responsible for local space-time characteristic extraction, the 3 multiplied by 3 max pooling layer is responsible for downsampling and inhibiting high-frequency noise, and the three output layers are spliced in the dimension of the channels; The spliced feature map is subjected to two-layer 1×1 convolution to finish depth separable extraction, and then global average pooling is accessed to compress the high-dimensional features into one-dimensional vectors. Preferably, the abnormal state diagnosis network depth is controlled at a bit level. Preferably, the future state estimation module data processing procedure is as follows: The input aircraft parameters obtain aircraft sta