Physical information neural network-based behavior identification method for strong maneuvering target

Abstract

The invention discloses a behavior identification method for a strong maneuvering target based on a physical information neural network, and the method achieves the high-precision maneuvering identification through building the precise mapping relation between six types of typical maneuvering modes and the data of a multi-source time sequence sensor, and combining with the physical constraint of six-degree-of-freedom dynamics. The method is characterized by comprising the following steps: constructing a maneuvering mode-sensor data feature corresponding relation library; designing a multi-scale spatial-temporal feature extraction network, fusing CNN, BiLSTM and an attention mechanism, and capturing local mutation and long-time dependence features of maneuvering behaviors; introducing a six-degree-of-freedom kinetic equation as a physical constraint; and developing a self-adaptive mixed loss function, and dynamically balancing a data driving error and a physical rule residual error. According to the method, a physical model and deep learning are fused, so that the physical reasonability of a prediction result is ensured, the problem of generalization under a small sample condition is solved, and the method has the capability of capturing transient characteristics and long-term modes of maneuvering behaviors at the same time.

Description

Technical Field

[0001] The present invention relates to a method for identifying the maneuvering behavior of a strong maneuvering target, and in particular to a method for identifying the behavior of a strong maneuvering target based on a physical information neural network. Background Art

[0002] With the development of modern military technology, the maneuverability of various high-speed aircraft (such as hypersonic missiles, stealth fighters, and intelligent drones) has significantly increased. Their complex maneuvers pose a severe challenge to traditional tracking and recognition methods. In the field of military defense, accurately identifying the target's maneuvering behavior is a key prerequisite for effective interception and is directly related to national aerospace security. Furthermore, in the field of civil aviation, target maneuvering behavior recognition is also of great value in applications such as drone traffic management and air collision avoidance systems.

[0003] Existing technologies for identifying maneuvering behavior of highly maneuvering targets face the following major bottlenecks: Traditional model-based methods (such as IMM filters) rely on a fixed set of maneuver models, making them difficult to adapt to new complex maneuvers such as the "spiral-snake" maneuver of hypersonic missiles. This results in high response delays for step maneuvers and a high rate of misjudgment of barrel rolls. Purely data-driven methods (such as LSTM) experience a sharp drop in accuracy in small sample sizes and may produce predictions that violate dynamics (such as continuous acceleration without thrust). By integrating physical models with deep learning, it is necessary to ensure that all predictions (such as the periodic forces in snake maneuvers and the conservation of angular momentum in barrel rolls) strictly conform to Newtonian mechanics and eliminate counter-physics predictions. Summary of the Invention

[0004] In order to solve the technical problems mentioned in the above background technology, the present invention proposes a behavior identification method for highly maneuverable targets based on a physical information neural network. By establishing a mapping relationship between typical maneuvering patterns and multi-source time-series sensor data and combining dynamic physical constraints, high-precision maneuvering identification is achieved.

[0005] In order to achieve the above technical objectives, the technical solution of the present invention is:

[0006] A method for identifying a highly maneuverable target's behavior based on a physical information neural network is characterized by comprising the following steps:

[0007] Step S1: establishing a feature mapping relationship between a strong maneuvering target's maneuvering mode and sensor detection data;

[0008] Step S2: Construct a physical information neural network (PINN), which includes: a multi-scale feature extraction module for extracting local and global temporal features from sensor data; a physical constraint embedding module that uses the target's dynamic equation as a regularization term in the loss function to constrain the network output to conform to physical laws; and a multi-task output module that simultaneously outputs the target's predicted state, maneuvering force estimation, and maneuvering mode classification probability.

[0009] Step S3: define a hybrid loss function, including data fitting loss, physical equation residual loss, and regularization term, and dynamically adjust the weight coefficients of each loss according to the residual ratio of the previous round of training;

[0010] Step S4: training the network using the labeled sensor data and physical simulation data to optimize network parameters;

[0011] Step S5: input sensor data in real time, and output the maneuvering mode classification result and motion state prediction through the network.

[0012] Furthermore, in step S1:

[0013] The time series sensor detection data includes the position, speed and angular velocity information of the strong maneuvering target;

[0014] The maneuvering behaviors of the strong maneuvering target include snake maneuver, spiral maneuver, dive maneuver, step maneuver, barrel roll maneuver and conventional flight.

[0015] Furthermore, in step S2, the multi-scale feature extraction module is composed of three core components, which are intended to capture features of different time scales from time series sensor data, including: a one-dimensional convolutional layer, which is used to extract rich local features and capture short-term mutations or high-frequency changes; a bidirectional long short-term memory network BiLSTM, wherein the bidirectional long short-term memory network BiLSTM includes forward and backward LSTM, and the forward and backward LSTM respectively process forward and reverse time series information to identify periodic maneuvers; a self-attention mechanism, which enhances the feature weights of mutations or turning points and captures the relationship between different feature spaces in parallel.

[0016] Furthermore, in step S2, the physical constraint embedding module includes the 12-DOF strong maneuvering high-speed aircraft model equation:

[0017]

[0018] Among them, u, v, and w are the velocity components of the strong maneuvering target, is the acceleration component, p, q, r are the angular velocity components, is the angular acceleration component, F x 、F y 、F z I is the combined external force component of the body axis, including aerodynamic force, thrust, and gravity components; xx , I yy , I zz is the moment of inertia around the body axis, M x 、M y 、M zis the target body axis external torque component, including aerodynamic torque, control torque, and thrust eccentric torque; φ, θ, ψ are the roll, pitch, and yaw Euler angles, is the rate of change of Euler angles.

[0019] Furthermore, in step S2, the multi-task output module includes the following branches: a state prediction branch, which outputs the position, velocity and acceleration of the target; a maneuvering force estimation branch, which outputs the net force and torque acting on the target; and a maneuvering mode classification branch, which outputs the probability distribution of each preset maneuvering type.

[0020] Furthermore, in step S3, the hybrid loss function ensures that the model satisfies both data fitting accuracy and physical law consistency by fusing data-driven error, physical constraint residual, and regularization term. The specific form of the hybrid loss function is:

[0021] L total =λ1L data +λ2L physics +λ3L reg

[0022] Among them, λ1, λ2, λ3 are weight coefficients, L data Data fitting loss, L physics Physical constraint residual loss, L total is the mixed loss function, L reg is a regularization term used to prevent overfitting;

[0023] The data fitting loss is used to minimize the deviation between the model prediction state and the actual sensor data. The specific expression is:

[0024]

[0025] in, is the state vector of the strong maneuvering target predicted by the network, which includes position vector, velocity vector, attitude angle vector and angular velocity vector, X i is the true state of a strong maneuvering target, is the classification probability of maneuver behavior, P i is the true classification label, N is the number of strong maneuvering target state vectors, and its value is 12.

[0026] Furthermore, in step S3, the physical constraint loss is used to force the network-predicted forces, moments, and state changes to conform to the high-speed aircraft dynamics equations. The specific expression is:

[0027]

[0028] in, is the residual of the target translational dynamics equation, is the residual of the target rotational dynamics equation, The network predicts the body axial forces and moments, and has:

[0029]

[0030] The regularization term is used to prevent overfitting and improve the generalization ability of the model. The specific expression is:

[0031]

[0032] Among them, ω i is the network weight, α is the weight coefficient;

[0033] Strategies for dynamically adjusting weight coefficients include:

[0034] In the initial stage, λ1 is set to be approximately equal to λ2, and the regularization weight λ3 is set between 0.001 and 0.01;

[0035] Dynamically adjust λ2 according to the residual ratio of the previous round of training, that is:

[0036]

[0037] Where t represents the training round, is the initial value of λ2.

[0038] Furthermore, in step S4, the training process of the physical information neural network for identifying the maneuvering behavior of a strong maneuvering target includes: simulating a strong maneuvering target combat scenario based on a six-degree-of-freedom dynamic model of the strong maneuvering target, wherein the relationship between the pre-set maneuvering pattern of the strong maneuvering target and the sensor detection data is embedded in the simulation as follows:

[0039] Snake maneuver: The lateral acceleration changes periodically, and the lateral acceleration and yaw rate oscillate at the same frequency in the range of 0.1Hz to 2Hz; the autocorrelation function shows a significant peak, and its expression is: F y =Asin(ωt), A is the amplitude of the fluctuation in the y direction, ω represents the angular frequency;

[0040] Spiral maneuver: angular velocity is stable, the pitch angular velocity changes in the range of 10 to 30 degrees per second, the three-dimensional trajectory curvature radius gradually decreases, and the roll-yaw moment is coupled

[0041] Dive maneuver: vertical acceleration changes suddenly, and the acceleration is greater than 50 meters per second squared; the altitude change rate increases or decreases sharply, and the gravity component F z = -mgcosθcosφ;

[0042] Step maneuver: acceleration step (time rate of change of acceleration>100m / s 3), force mutation ||F(t+Δt)-F(t)||>F th , F th Indicates thrust;

[0043] Barrel roll maneuver: synchronous change of three-axis angular velocity, moment of inertia constraint

[0044] Normal flight: acceleration and angular velocity are close to zero, physical residual is very small, ||R physics ||<ε;

[0045] The flight status data of the strong maneuvering target and the corresponding maneuvering mode during the sampling simulation process are used to obtain the training set and its labels and the test set and its labels; the improved physical information neural network is verified based on the test set and its labels.

[0046] Furthermore, in step S5, sensor detection data is input in real time, and the maneuvering mode classification result and motion state prediction of the strong maneuvering target are output through the established physical information network.

[0047] Compared with the prior art, the present invention adopts the above technical solution and has the following beneficial effects:

[0048] (1) The present invention proposes a behavior recognition method for highly maneuverable targets based on a physical information neural network, which integrates physical models with deep learning and embeds the 6-DOF rigid body dynamics equations as hard constraints into the neural network, which not only ensures the physical rationality of the prediction results but also solves the generalization problem under small sample conditions.

[0049] (2) The multi-scale spatiotemporal feature extraction architecture designed in this paper can simultaneously capture the transient characteristics and long-term patterns of maneuvering behavior.

[0050] (3) The present invention develops an adaptive hybrid loss function to dynamically balance the data-driven error and the physical law residual, ensuring that the model simultaneously meets the data fitting accuracy and physical law consistency. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0052] Figure 1 This is a specific flow chart of a method for identifying maneuvering behavior of a strong maneuvering target based on a physical information network proposed by the present invention;

[0053] Figure 2This is a principle block diagram of a method for identifying the maneuvering behavior of a highly maneuverable target based on a physical information network proposed in the present invention. DETAILED DESCRIPTION

[0054] To make the purpose, technical solutions, and advantages of the embodiments of this application more clear, the technical solutions in the embodiments of this application will be clearly and completely described below in conjunction with the drawings in the embodiments of this application. Obviously, the described embodiments are part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0055] like Figure 1 As shown in the figure, in this embodiment, a method for identifying maneuvering behavior of a strong maneuvering target based on a physical information network is provided:

[0056] Step S1: Establish a feature mapping relationship between the maneuvering pattern of a strong maneuvering target and sensor detection data. The time-series sensor detection data includes the position, velocity, and angular velocity information of the strong maneuvering target. Strong maneuvering target maneuvers include serpentine maneuvers, spiral maneuvers, dive maneuvers, step maneuvers, barrel rolls, and conventional flight.

[0057] The following table details how the six types of maneuver patterns appear in typical sensor data:

[0058] Table 1 Feature mapping relationship between maneuvering mode of strong maneuvering target and sensor detection data

[0059]

[0060] Among them, u, v, and w are the velocity components of the strong maneuvering target in the x, y, and z coordinates in the body coordinate system, and p, q, and r are the roll angular velocity, pitch angular velocity, and yaw angular velocity.

[0061] Step S2: Construct a physical information neural network (PINN). The network includes: a multi-scale feature extraction module for extracting local and global temporal features from sensor data; a physical constraint embedding module that uses the target's dynamic equation as a regularization term in the loss function to constrain the network output to conform to physical laws; and a multi-task output module that simultaneously outputs the target's predicted state, maneuvering force estimation, and maneuvering mode classification probability.

[0062] The multi-scale feature extraction module consists of three core components, designed to capture features at different time scales from time series sensor data. These components include: a one-dimensional convolutional layer for extracting rich local features and capturing short-term mutations or high-frequency changes; a bidirectional long short-term memory network, with forward and backward LSTMs processing forward and reverse time series information, respectively, to identify periodic maneuvers; and a self-attention mechanism that enhances the feature weights of mutations or turning points and captures relationships in different feature spaces in parallel.

[0063] The physical constraint embedding module includes the 12-DOF strong maneuvering high-speed aircraft model equations:

[0064]

[0065] Among them, u, v, and w are the velocity components of the strong maneuvering target, is the acceleration component, p, q, r are the angular velocity components, is the angular acceleration component, F x 、F y 、F z I is the combined external force component of the body axis, including aerodynamic force, thrust, and gravity components; xx , I yy , I zz is the moment of inertia around the body axis, M x 、M y 、M z is the target body axis external torque component, including aerodynamic torque, control torque, and thrust eccentric torque; φ, θ, ψ are the roll, pitch, and yaw Euler angles, is the rate of change of Euler angles.

[0066] The multi-task output module includes the following branches: a state prediction branch that outputs the position, velocity, and acceleration of the target; a maneuvering force estimation branch that outputs the net force and torque acting on the target; and a maneuvering mode classification branch that outputs the probability distribution of each preset maneuvering type.

[0067] Step S3: Define the hybrid loss function, including data fitting loss, physical equation residual loss and regularization term.

[0068] The hybrid loss function ensures that the model satisfies both data fitting accuracy and physical law consistency by integrating data-driven error, physical constraint residual, and regularization term. The specific form of the hybrid loss function is:

[0069] L total =λ1L data +λ2L physics +λ3L reg

[0070] Among them, λ1, λ2, λ3 are weight coefficients, L data Data fitting loss, Lphysics Physical constraint residual loss, L total is the mixed loss function, L reg is a regularization term used to prevent overfitting.

[0071] The data fitting loss is used to minimize the deviation between the model prediction state and the actual sensor data. The specific expression is:

[0072]

[0073] in, is the state vector of the strong maneuvering target predicted by the network, which includes position vector, velocity vector, attitude angle vector and angular velocity vector, X i is the true state of a strong maneuvering target, is the classification probability of maneuver behavior, P i is the true classification label, N is the number of strong maneuvering target state vectors, and its value is 12.

[0074] The physical constraint loss forces the network to predict forces / torques and state changes that conform to the high-speed aircraft dynamics equations, which are expressed as:

[0075]

[0076] in, is the residual of the target translational dynamics equation, is the residual of the target rotational dynamics equation, The network predicts the body axial forces and moments, and has:

[0077]

[0078] The regularization term is used to prevent overfitting and improve the generalization ability of the model. The specific expression is:

[0079]

[0080] Among them, ω i is the network weight, and α is the weight coefficient.

[0081] The strategy for dynamically adjusting the weight coefficients includes: setting λ1≈λ2 in the initial stage, and the regularization weight λ3 to [0.001-0.01]. Dynamically adjust λ2 based on the residual ratio of the previous round of training:

[0082]

[0083] Where t represents the training round, is the initial value of λ2.

[0084] Step S4: using the labeled sensor data and physical simulation data to train the network and optimize network parameters;

[0085] The training process of the physical information neural network for identifying the maneuvering behavior of a strong maneuvering target includes: simulating the combat scenario of a strong maneuvering target based on the six-degree-of-freedom dynamic model of the strong maneuvering target. The relationship between the pre-set maneuvering pattern of the strong maneuvering target and the sensor detection data is embedded in the simulation as follows:

[0086] Snake maneuver: lateral acceleration changes periodically (0.1-2Hz), and the autocorrelation function shows a significant peak, F y =Asin(ωt), A is the amplitude of the fluctuation in the y direction, ω represents the angular frequency;

[0087] Spiral maneuver: stable angular velocity (10-30° / s), gradually decreasing radius of curvature of the three-dimensional trajectory, coupled roll-yaw moment

[0088] Dive maneuver: sudden change in vertical acceleration (>5G), steep increase / decrease in altitude change rate, gravity component F z = -mgcosθcosφ;

[0089] Step maneuver: acceleration step (jerk>100m / s 3 ), force mutation ||F(t+Δt)-F(t)||>F th ;

[0090] Barrel roll maneuver: Synchronous change of three-axis angular velocity (roll + yaw), moment of inertia constraint

[0091] Normal flight: acceleration and angular velocity are close to zero, physical residual is very small, ||R physics ||<ε.

[0092] The flight status data of the strong maneuvering target and the corresponding maneuvering mode during the sampling simulation process are used to obtain the training set and its labels and the test set and its labels; the improved physical information neural network is verified based on the test set and its labels.

[0093] Step S5: Input sensor detection data in real time, and output the maneuvering mode classification results and motion state prediction of the strong maneuvering target through the established physical information network.

[0094] The foregoing is merely a list of specific embodiments of the present application, intended to enable those skilled in the art to understand or implement the present application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application is not limited to the embodiments shown herein, but is intended to conform to the broadest scope consistent with the principles and novel features of the present application.

Claims

1. A behavior recognition method for a strong maneuvering target based on a physical information neural network, characterized in that: The following steps are involved: Step S1: establishing a feature mapping relationship between a strong maneuvering target's maneuvering mode and sensor detection data; Step S2: Construct a physical information neural network (PINN), which includes: a multi-scale feature extraction module for extracting local and global temporal features from sensor data; a physical constraint embedding module that uses the target's dynamic equation as a regularization term in the loss function to constrain the network output to conform to physical laws; and a multi-task output module that simultaneously outputs the target's predicted state, maneuvering force estimation, and maneuvering mode classification probability. Step S3: define a hybrid loss function, including data fitting loss, physical equation residual loss, and regularization term, and dynamically adjust the weight coefficients of each loss according to the residual ratio of the previous round of training; Step S4: training the network using the labeled sensor data and physical simulation data to optimize network parameters; Step S5: input sensor data in real time, and output the maneuvering mode classification result and motion state prediction through the network.

2. The method for identifying a highly maneuverable target based on a physical information neural network according to claim 1, characterized in that: In the step S1: The time series sensor detection data includes the position, speed and angular velocity information of the strong maneuvering target; The maneuvering behaviors of the strong maneuvering target include snake maneuver, spiral maneuver, dive maneuver, step maneuver, barrel roll maneuver and conventional flight.

3. The method for identifying a highly maneuverable target based on a physical information neural network according to claim 1, characterized in that: In step S2, the multi-scale feature extraction module is composed of three core components, which are designed to capture features of different time scales from time series sensor data, including: a one-dimensional convolutional layer for extracting rich local features and capturing short-term mutations or high-frequency changes; a bidirectional long short-term memory network BiLSTM, wherein the bidirectional long short-term memory network BiLSTM includes forward and backward LSTMs, and the forward and backward LSTMs process forward and reverse time series information respectively to identify periodic maneuvers; and a self-attention mechanism to enhance the feature weights of mutations or turning points and capture the relationship between different feature spaces in parallel.

4. The method for identifying a highly maneuverable target based on a physical information neural network according to claim 1, characterized in that: In step S2, the physical constraint embedding module includes the 12-DOF strong maneuvering high-speed aircraft model equation: Among them, u, v, and w are the velocity components of the strong maneuvering target, is the acceleration component, p, q, r are the angular velocity components, is the angular acceleration component, F x 、F y 、F z I is the combined external force component of the body axis, including aerodynamic force, thrust, and gravity components; xx , I yy , I zz is the moment of inertia around the body axis, M x 、M y 、M z is the target body axis external torque component, including aerodynamic torque, control torque, and thrust eccentric torque; φ, θ, ψ are the roll, pitch, and yaw Euler angles, is the rate of change of Euler angles.

5. The method for identifying a highly maneuverable target based on a physical information neural network according to claim 1, characterized in that: In step S2, the multi-task output module includes the following branches: a state prediction branch that outputs the position, velocity, and acceleration of the target; a maneuvering force estimation branch that outputs the net force and torque acting on the target; and a maneuvering mode classification branch that outputs the probability distribution of each preset maneuvering type.

6. The method for identifying a highly maneuverable target based on a physical information neural network according to claim 1, characterized in that: In step S3, the hybrid loss function ensures that the model satisfies both data fitting accuracy and physical law consistency by fusing data-driven error, physical constraint residual, and regularization term. The specific form of the hybrid loss function is: L total =λ1L data +λ2L physics +λ3L reg Among them, λ1, λ2, λ3 are weight coefficients, L data Data fitting loss, L physics Physical constraint residual loss, L total is the mixed loss function, L reg is a regularization term used to prevent overfitting; The data fitting loss is used to minimize the deviation between the model prediction state and the actual sensor data. The specific expression is: in, is the state vector of the strong maneuvering target predicted by the network, which includes position vector, velocity vector, attitude angle vector and angular velocity vector, X i is the true state of a strong maneuvering target, is the classification probability of maneuver behavior, P i is the true classification label, N is the number of strong maneuvering target state vectors, and its value is 12.

7. The method for identifying a highly maneuverable target based on a physical information neural network according to claim 1, characterized in that: In step S3, the physical constraint loss is used to force the network-predicted forces, moments, and state changes to conform to the high-speed aircraft dynamics equations. The specific expression is: in, is the residual of the target translational dynamics equation, is the residual of the target rotational dynamics equation, The network predicts the body axial forces and moments, and has: The regularization term is used to prevent overfitting and improve the generalization ability of the model. The specific expression is: Among them, ω i is the network weight, α is the weight coefficient; Strategies for dynamically adjusting weight coefficients include: In the initial stage, λ1 is set to be approximately equal to λ2, and the regularization weight λ3 is set between 0.001 and 0.01; Dynamically adjust λ2 according to the residual ratio of the previous round of training, that is: Where t represents the training round, is the initial value of λ2.

8. The method for identifying behavior of a highly maneuverable target based on a physical information neural network according to claim 1, characterized in that: In step S4, the training process of the physical information neural network for identifying the maneuvering behavior of a strong maneuvering target includes: simulating a strong maneuvering target combat scenario based on a six-degree-of-freedom dynamic model of the strong maneuvering target, wherein the relationship between a pre-set maneuvering pattern of the strong maneuvering target and sensor detection data is embedded in the simulation as follows: Snake maneuver: The lateral acceleration changes periodically, and the lateral acceleration and yaw rate oscillate at the same frequency in the range of 0.1Hz to 2Hz; the autocorrelation function shows a significant peak, and its expression is: F y =Asin(ωt), A is the amplitude of the fluctuation in the y direction, ω represents the angular frequency; Spiral maneuver: angular velocity is stable, the pitch angular velocity changes in the range of 10 to 30 degrees per second, the three-dimensional trajectory curvature radius gradually decreases, and the roll-yaw moment coupling M x ∝qr; Dive maneuver: vertical acceleration changes suddenly, and the acceleration is greater than 50 meters per second squared; the altitude change rate increases or decreases sharply, and the gravity component F z = -mgcosθcosφ; Step maneuver: acceleration step (time rate of change of acceleration>100m / s 3 ), force mutation ||F(t+Δt)-F(t)||>F th , F th Indicates thrust; Barrel roll maneuver: synchronous change of three-axis angular velocity, moment of inertia constraint Normal flight: acceleration and angular velocity are close to zero, physical residual is very small, ||R physics ||<ε; The flight status data of the strong maneuvering target and the corresponding maneuvering mode during the sampling simulation process are used to obtain the training set and its labels and the test set and its labels; the improved physical information neural network is verified based on the test set and its labels.

9. The method for identifying a highly maneuverable target based on a physical information neural network according to claim 1, characterized in that: In the step S5, sensor detection data is input in real time, and the maneuvering mode classification result and motion state prediction of the strong maneuvering target are output through the established physical information network.

Images