An online health monitoring system for hypersonic aircraft

Abstract

The present invention discloses an online health monitoring system for a hypersonic aircraft, which includes abnormal state diagnosis, future state estimation and model performance autonomous optimization modules, wherein the abnormal state diagnosis module is a neural network model fused with an extended perception layer, which can accurately diagnose the abnormal state of a hypersonic aircraft actuator in a relatively short time; the future state estimation module is a block interpretable neural network model embedded with physical information, which can accurately estimate the future state of the aircraft under abnormal conditions and provide sufficient information for safe flight decisions of the aircraft; the online health monitoring system for a hypersonic aircraft disclosed by the present invention has the advantages of less online computing resource occupation, fast computing speed, low data dependence, high accuracy of abnormal state diagnosis and future state estimation, and high reliability of health monitoring results.

Description

Technical Field

[0001] The present invention relates to the technical field of hypersonic aircraft, and in particular to an online health monitoring system for a hypersonic aircraft. Background Technology

[0002] Hypersonic aircraft is a new type of long-distance carrier that can achieve long-distance unpowered flight with the help of aerodynamics and has the ability to reach long distances quickly. Hypersonic aircraft fly in harsh airspace conditions. Their system functions are complex, and their actuators are frequently actuated, which makes them prone to abnormalities. Once an abnormal state occurs, it can cause the aircraft to become unstable at best, or even crash at worst, and fail to complete the scheduled mission. Therefore, timely and effective detection of abnormal states is crucial to ensure the successful execution of the mission.

[0003] However, the flight process of hypersonic vehicles is characterized by strong coupling, strong nonlinearity and high uncertainty, which brings great challenges to abnormal flight state detection and flight state estimation. Traditional state perception and anomaly detection methods based on mechanism models are difficult to meet application requirements. Data-driven modeling does not need to pay attention to the specific physical information in the movement of the aircraft. It can mine the complex nonlinear relationship between parameters based on a certain amount of data. Compared with the mechanism model, the data-driven model contains more complex mathematical descriptions and can be compatible with a variety of mechanism equations and physical relationships. It can significantly enhance the robustness of the model and improve the accuracy of abnormal state detection and estimation. However, hypersonic vehicle flight data is difficult to obtain, the data transmission cost is high, the data set is low in completeness, and the low interpretability and trust of the traditional data-driven model itself limits its effective application in abnormal state detection and estimation tasks. Based on this, the present invention proposes a block interpretable neural network modeling method based on physical information embedding based on the flight mechanism equation of hypersonic aircraft, and establishes an online health monitoring system with abnormal state diagnosis, future state estimation and autonomous optimization of model performance as the main functions to meet the needs of online autonomous health monitoring of hypersonic aircraft, ensure safe and reliable flight and mission execution success rate, and provide technical guarantee for the improvement of my country's hypersonic aircraft state monitoring capability and long-range defense capability. SUMMARY OF THE INVENTION

[0004] The purpose of the present invention is to provide an online health monitoring system for a hypersonic aircraft to solve the problems in the background technology.

[0005] To achieve the above purpose, the invention provides the following technical solutions:

[0006] A hypersonic vehicle online health monitoring system, including an abnormal state diagnosis module, a future state estimation module and a model performance autonomous optimization module;

[0007] Among them, the abnormal state diagnosis module can achieve rapid detection and diagnosis of the abnormal state of the aircraft actuator under high-speed and high-dynamic flight conditions, improving flight safety;

[0008] The future state estimation module quickly estimates the motion and attitude parameters of the aircraft in the abnormal state based on the abnormal state detection results and flight state parameters, providing sufficient information for the safe flight decision-making of the aircraft and improving the mission success rate;

[0009] The model performance self-optimization module optimizes the input parameters and sorting of the model through the contribution degree of the model input to the output, improving the online calculation performance of the model.

[0010] A hypersonic aircraft online health monitoring method includes the following steps:

[0011] Step 1: Design the overall framework of the hypersonic aircraft online health monitoring system including an abnormal state diagnosis module, a future state estimation module, and a model performance self-optimization module, as Figure 1 , where the abnormal state diagnosis module can achieve rapid detection and diagnosis of the abnormal state of the aircraft actuator under high-speed and high-dynamic flight conditions, improving flight safety; the future state estimation module quickly estimates the motion and attitude parameters of the aircraft in the abnormal state based on the abnormal state detection results and flight state parameters, providing sufficient information for the safe flight decision-making of the aircraft and improving the mission success rate; the model performance self-optimization module optimizes the input parameters and sorting of the model through the contribution degree of the model input to the output, improving the online calculation performance of the model.

[0012] Step 2: Design the abnormal state diagnosis module in the hypersonic aircraft online health monitoring system, which includes an extended perception layer and a rapid diagnosis layer. The extended perception layer extracts features from high-dimensional flight state parameters and actuator parameters with large feature differences, improving the perception field of the diagnosis model, and then stacking with the rapid diagnosis layer to achieve rapid and accurate diagnosis of the abnormal state of the actuator.

[0013] Step 3: Design the future state estimation module in the hypersonic aircraft online health monitoring system, which includes an actuator module and a flight parameter module. The actuator module preprocesses the aircraft rudder deflection value through feature binning and uses it as the model input together with the abnormal state diagnosis result to realize online estimation of the future actuation state of the actuator; the flight parameter module fuses the rudder deflection result of the next moment obtained through the actuator module and realizes online estimation of the future flight state through an interpretable neural network embedded with physical information.

[0014] Step 4: Design an interpretable neural network model for embedding physical information in the flight parameter module. The flight vehicle's motion equations, attitude equations, and abnormal rudder deflection equations are embedded as mechanism constraints in the neural network model, simplifying the model optimization space, improving the transparency and credibility of the model, and thus obtaining more accurate future state estimation results.

[0015] Step 5: Design a model performance self-optimization module in the online health monitoring system of a hypersonic vehicle. By calculating the contribution of the input parameter features of the model at each moment to the model output, analyze the importance of the input parameters, and sort them, reducing or ignoring the non-critical input parameters of the model, reducing the input dimension, improving the model's ability to mine limited data, and reducing the computational burden.

[0016] Design the communication mode of the abnormal state diagnosis module, future state estimation module, and model performance self-optimization module according to the overall framework, and construct an online health monitoring system for hypersonic vehicles.

[0017] Technical effects: There are two optimization schemes for using the importance ranking results. One is to directly ignore non-critical features, reducing the input dimension of the model, which can directly reduce the amount of data for operation and the complexity of the model; the other is to connect a fully connected layer behind these features with relatively small importance, synthesize multiple groups of features into a single feature, and then form a new input data combination with other important features, retaining non-important features to a certain extent and also reducing the input data dimension, reducing part of the computational burden. Description of the Drawings

[0018] Figure 1 This is the overall technical framework of the present invention.

[0019] Figure 2 This is the abnormal state diagnosis model of the hypersonic vehicle of the present invention.

[0020] Figure 3 This is the actuator module of the present invention.

[0021] Figure 4 This is the future state estimation model of the hypersonic vehicle of the present invention. Detailed Embodiments

[0022] In order to make the objectives, technical solutions, and advantages of the invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0023] In one embodiment, as Figures 1 - 4 shown, an online health monitoring system for a hypersonic vehicle includes an abnormal state diagnosis module, a future state estimation module, and a model performance self-optimization module;

[0024] Among them, the abnormal state diagnosis module can realize the rapid detection and diagnosis of the abnormal state of the aircraft actuator under high-speed and high-dynamic flight conditions, improving flight safety;

[0025] The future state estimation module quickly estimates the motion and attitude parameters of the aircraft in the abnormal state according to the abnormal state detection results and flight state parameters, providing sufficient information for the safe flight decision-making of the aircraft and improving the mission success rate;

[0026] The model performance self-optimization module optimizes the input parameters and sorting of the model through the contribution degree of the model input to the output, improving the online calculation performance of the model.

[0027] A hypersonic aircraft online health monitoring method includes the following steps:

[0028] Step 1: The hypersonic aircraft online health monitoring system includes an abnormal state diagnosis module, a future state estimation module, and a model performance self-optimization module. Among them, an extended perception layer is integrated into the abnormal state diagnosis, expanding the feature perception field of view. The state detection model can perceive the existence of abnormal conditions in the shortest possible time; in the future state estimation module, a physical feature constraint network modeling method based on the block idea is adopted, which can estimate the state of the aircraft in a short future time after the abnormality occurs, so as to provide sufficient information for the subsequent processing of abnormal state perception, such as Figure 1 .

[0029] Step 2: The abnormal state diagnosis network is mainly used for the timely judgment of the abnormal state of the aircraft. Since the aircraft state changes rapidly and the sampling feature information varies greatly, how to ensure that each data is fully utilized under limited information to the greatest extent becomes the key to the abnormal state diagnosis of hypersonic aircraft. The designed neural network modeling method integrating the extended perception layer can better fuse multi-dimensional data features without increasing the number of network layers and computational load. Through the multi-scale feature information fusion of the extended perception layer and the classification and diagnosis of the fast diagnosis layer, the final diagnosis result is obtained. The structure of the abnormal state diagnosis model is as Figure 2Considering the computational conditions and computing power resources of hypersonic vehicles, a simplified extended perception layer network design is proposed. By utilizing the feature extraction capabilities of convolutional layers and pooling layers, more comprehensive feature perception results can be obtained through different scales and fields of view. The structure of the extended perception layer mainly includes two feature convolutional layers with different scales and a pooling layer. Different-sized convolutional kernels are used to enable the model to autonomously select more appropriate input data. Since an overly large convolutional kernel will bring a greater computational burden, the convolutional kernel sizes of feature convolutional layer 1 and feature convolutional layer 2 are 1×1 and 3×3 respectively, and the pooling layer uses the 3×3 max-pooling method. In addition, using multiple feature convolutional layers will result in a too large thickness of the feature map and a large number of parameters. By adding a dimensionality reduction convolutional layer with a 1×1 convolutional kernel to integrate information from different channels, the number of channels can be quickly reduced. While reducing the number of parameters, the effect of the non-linear activation function is enhanced, and the expressive ability of the model is improved.

[0030] Step 3: The neural network highly depends on the generalization performance of the dataset. However, for hypersonic vehicles, it is difficult to obtain data. Design an interpretable neural network based on embedded physical information to improve the reliability of the neural network model. Since the abnormal states of hypersonic vehicles are often related to the actuators, various faults and non-linear characteristics occur more significantly in the actuators. The interpretable neural network with embedded physical information is divided into an actuator module and a vehicle parameter module.

[0031] There are three actuators for hypersonic vehicles, namely the left rudder, right rudder, and direction rudder. The influence of abnormal states on future state estimation is also considered. Therefore, the input of the actuator module of the hypersonic vehicle is 3 rudder deflection values and the diagnosis results of the abnormal state diagnosis model. The calculation process is as follows Figure 3 Through the method of feature binning, the rudder deflection data is preprocessed. The binning interval is set to be every 0.1 between [-30, 30], with a total of 600 interval ranges. The rudder deflection data is discretized, which is more conducive to the data-driven model to understand, and at the same time allows the vector to fluctuate within a smaller range, improving the robustness of the model.

[0032] In the actuator module, the input data is the rudder deflection data at the previous time node within the sample and the current state diagnosis result. After passing through the feature extraction layer and the learning layer, the rudder deflection value for the next time node is output for learning. In the vehicle parameter module, by fusing the rudder deflection result for the next moment obtained from the actuator module, through the neural network, attention mechanism, and output layer, the motion and attitude parameters of the vehicle for the next moment are estimated in combination with the vehicle motion parameters.

[0033] Step 4: During the aircraft simulation process, there are many kinematic and dynamic equations. Embedding the corresponding aircraft flight mechanism equations into the data-driven model as model training constraints can simplify the model optimization space, increase the transparency and credibility of the data-driven model, and obtain better model future state estimation results. The mechanism constraint equations used are as follows:

[0034] Motion constraint equation:

[0035]

[0036] The motion constraint equation is derived from the kinematic mechanism equation. Among them, T x , T y , T z correspond to the constraint terms of the model in the x, y, and z directions, and K x , K y , K z are relaxation coefficients. Considering the influence of system uncertainty, it allows the model learning results to deviate slightly from the mechanism formula constraints within a small range. At the same time, since the basic dynamic equation cannot be violated, it is hoped that the relaxation coefficient is as close to 1 as possible. In the motion constraint equation, the part on the right side of the plus sign constrains the relaxation coefficient to punish the model estimation results that deviate greatly from the mechanism process. λ is a harmonic coefficient used to reconcile the calculation results on both sides. Generally, the harmonic coefficient is infinitely close to 0. MSE T is the motion constraint loss, which is obtained by directly summing the three constraint terms of T x , T y , and T z .

[0037] Attitude constraint equation:

[0038]

[0039] The attitude constraint equation is derived from the dynamic mechanism equation. Among them, G ψ , G γ correspond to the constraint terms of the model on the ψ, γ three attitude angles, K ψ , K γ are relaxation coefficients. MSE P is the attitude constraint loss, which is obtained by directly summing the three constraint terms of G ψ , and G γ .

[0040] Abnormal rudder deflection constraint:

[0041] MSE F = ∑(δ i - δi ) 2

[0042] where δ i is the rudder deflection value at the current moment, δ i ' is the estimated result of the rudder deflection value at the next moment containing the abnormal type result. i = a, r, e respectively represent the right rudder deflection, elevator deflection, and left rudder deflection, and MSE F is the abnormal rudder deflection constraint result.

[0043] The motion constraint, attitude constraint, and abnormal rudder deflection constraint modules introduce additional penalty terms to penalize the learning results of the model that do not conform to the decision rules. Under the motion constraint, attitude constraint, and abnormal rudder deflection constraint, for the actual result Y and the model prediction result Y', the mean square error loss function of the model can be written as:

[0044] MSE = (Y - Y') 2 + αMSE T + βMSE P + γMSE F

[0045] The above loss function includes four parts: the model prediction loss, motion constraint loss, attitude constraint loss, and abnormal rudder deflection constraint loss. Among them, α, β, and γ are hyperparameters used to ensure that the values of each loss term have the same order of magnitude. In the experiment, α = 0.003, β = 0.01, and γ = 0.05. The model continuously optimizes its internal parameters during the training process to minimize this loss function.

[0046] Step Five: During the model training process, better model performance indicators are often required. It is necessary to understand the operating principle of the model and which features play a key role, which is important for further optimizing the model.

[0047] By calculating the marginal contribution of each feature, that is, the difference between the contribution when the feature is added and the contribution before the feature is added, which is the marginal contribution of the feature, to describe the influence of the input on the output. By explaining how much each sample and each feature contribute to the corresponding estimated value, the local interpretation of the model can be quantified. The contribution degree calculation formula is as follows:

[0048]

[0049] where φ i is the feature contribution value, that is, the degree of influence on the result, S is the subset of features that need to be retrained in the model, F represents the set of all features, and f S∪{i} represents the model after training with the current feature, and f S represents the model after training with the hidden features, and f S∪{i} (x s∪{i})-f S (x s ) represents the difference after the two models are predicted. x s represents the input feature setting under the S set, being the weight.

[0050]

[0051] Among them, g is the interpretation model, z' ∈ {0, 1} M characterizing whether the corresponding feature is observed, M is the number of input features, φ i is the contribution value of each feature, φ 0 is a constant, referring to the predicted mean of all training samples.

[0052] The ranking results of the importance of input features at each moment are obtained through calculation. Since it is not only for the preference of a single type of feature, but also to ensure the generalization of the training model, the calculation results of this importance are of great reference value for the entire model. There are two optimization schemes for using the ranking results of importance. One is to directly ignore non-critical features and reduce the input dimension of the model, which can directly reduce the amount of data for operation and the complexity of the model. The other is to connect fully connected layers in series behind these features with smaller importance, synthesize multiple groups of features into a single feature, and then form a new input data combination with other important features, which retains non-important features to a certain extent and also reduces the input data dimension and reduces part of the calculation burden.

[0053] As described above, it is only the specific implementation manner of the present disclosure, but the protection scope of the present disclosure is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present disclosure can easily think of changes or substitutions, which should be covered by the protection scope of the present disclosure. Therefore, the protection scope of the present disclosure should be subject to the protection scope of the claims.

Claims

1. A hypersonic vehicle online health monitoring system, characterized in that: It includes abnormal state diagnosis module, future state estimation module and model performance autonomous optimization module; The abnormal state diagnosis module can realize rapid detection and diagnosis of abnormal states of aircraft actuators under high-speed and high-dynamic flight conditions, thus improving flight safety. The future state estimation module quickly estimates the motion and attitude parameters of the aircraft under abnormal conditions based on the abnormal state detection results and flight state parameters, providing sufficient information for the aircraft's safe flight decision and improving the mission success rate; The model performance autonomous optimization module optimizes the model input parameters and sorting by the contribution of the model input to the output, thereby improving the model's online computing performance; The future state estimation module includes an actuator module and a flight parameter module, wherein the actuator module pre-processes the rudder deviation value of the aircraft through feature binning, and uses it together with the abnormal state diagnosis result as a model input to realize online estimation of the future actuation state of the actuator; the flight parameter module realizes online estimation of the flight state at the future moment by fusing the rudder deviation result of the next moment obtained by the actuator module and an interpretable neural network embedded with physical information; The interpretable neural network with physical information embedding simplifies the model optimization space and improves the transparency and credibility of the model by embedding the aircraft motion equation, attitude equation and abnormal rudder deflection equation in the neural network model as mechanism constraints, thereby obtaining more accurate future state estimation results. The embedded motion constraint equation, attitude constraint equation and abnormal rudder deflection constraint equation are as follows: The motion constraint equations are: Among them, T x , T y , T z are the constraints in the x, y, and z directions respectively, and K x , K y , K z is the relaxation coefficient, which is close to 1, λ is the harmonic coefficient which is close to 0, MSE T is the motion constraint loss, which is represented by T x , T y , T z The three constraints are calculated by directly adding them together; Attitude constraint equation: in, , G ψ , G γ They are , ψ, γ constraints, , K ψ , K γ is the relaxation coefficient; MSE P is the posture constraint loss; Abnormal rudder deflection constraint equation: MSE F =∑(δ i -δ i ′) 2 , Among them, δ i is the rudder deviation value at the current moment, δ i ′ is the estimated value of the rudder deviation at the next moment under abnormal state, i=a, r, e represent the right rudder deviation, elevator deviation and left rudder deviation respectively, MSE F It is the abnormal rudder deviation constraint; Under motion constraints, attitude constraints and abnormal rudder constraints, the model mean square error loss function is: MSE=(YY′) 2 +αMSE T +βMSE P +γMSE F Among them, α, β, γ are hyperparameters, Y is the actual result, and Y′ is the model prediction result.

2. A hypersonic vehicle online health monitoring system according to claim 1, characterized in that: The abnormal state diagnosis module includes an extended perception layer and a rapid diagnosis layer. The extended perception layer extracts features of high-dimensional flight state parameters and actuator parameters with large feature differences to improve the perception field of the diagnosis model, and is further stacked with the rapid diagnosis layer to achieve rapid and accurate diagnosis of abnormal states of actuators.

3. The hypersonic vehicle online health monitoring system according to claim 1, characterized in that: The model performance autonomous optimization module calculates the contribution of the model input parameter features to the model output at each moment, analyzes the importance of the input parameters, and sorts the parameters, reducing or ignoring the non-critical input parameters of the model, reducing the input dimension, improving the model's ability to mine limited data, and reducing the computational burden; the feature contribution calculation equation is as follows: Among them, φ i is the feature contribution, S is the feature subset that needs to be retrained in the model, F represents the entire feature set, and f S∪{i} represents the model trained with the current features, f s represents the model trained after hiding the features, f S∪{i} (x s∪{i} )-f S (x s ) represents the difference between the two models’ predictions, x s represents the input feature setting under the S set, is the weight; Where g is the model explanation function, z′∈{0,1} M Indicates whether the corresponding feature is observed, M is the number of input features, φ i is the contribution of each feature, φ0 is a constant, which refers to the predicted mean of all training samples.

Images