Behavior intention prediction method based on multi-target threat situation and GRU network

Abstract

The invention discloses a behavior intention prediction method based on a multi-target threat situation and a GRU network. The behavior intention prediction method comprises the following steps: establishing a static threat field model and calculating a static threat field value of each target to our party; establishing a dynamic threat field model and calculating a dynamic threat field value of each target to our party; updating a static threat field value and a dynamic threat field value based on the historical trajectory data and the target state prediction information; based on a GRU model, time sequence modeling is carried out on the historical trajectory data, the static threat field data and the dynamic threat field data, and an individual intention prediction result of the target is obtained; and based on the individual intention prediction result of each target, carrying out fusion processing on the overall intentions of the multiple targets, and predicting the overall behavior intention. According to the global intention fusion mechanism in the application, through weighted average and weighted voting methods, prediction results of multiple targets are flexibly combined, and it is ensured that the overall behavior intention is accurately predicted in a complex environment in which multiple targets coexist.

Description

Technical Field

[0001] This application belongs to the technical field of intention prediction, and particularly relates to a behavior intention prediction method based on multi-object threat situation and GRU network. Background Art

[0002] In modern battlefields, the intention prediction of enemy targets has become a crucial part of military decision-making. Intention prediction can help commanders promptly identify potential actions of the enemy, reasonably adjust tactical deployments, and maximize combat effectiveness. Especially in a complex battlefield environment with high dynamics and multiple targets, how to accurately predict the action intentions of the enemy has become the key to enhancing combat effectiveness and decision-making efficiency.

[0003] Modern land battlefields are extremely dynamic and complex. Enemy targets on the battlefield usually have a high degree of uncertainty. At the same time, the acquisition of battlefield information is affected by sensor errors, environmental occlusion, and target camouflage, resulting in information incompleteness. In addition, there are often multiple enemy targets on modern battlefields, and there may be cooperative or adversarial relationships between these targets, increasing the difficulty of predicting enemy intentions. Therefore, in an environment of information incompleteness and high dynamic changes, how to accurately predict the action intentions of multiple enemy targets has become the main challenge of current technologies.

[0004] However, although existing deep learning methods have strong temporal modeling capabilities, the existing technologies still face some limitations. First, many methods rely on static threat assessments and fail to fully consider the dynamic changes in threat intensity with target movement and tactical adjustments. Second, most existing methods focus on the prediction of a single target or a single target type and ignore the cooperative combat behaviors of multiple targets, which often leads to a decrease in prediction accuracy when facing multiple enemy targets. Finally, information in the battlefield is often incomplete, and the loss, interference, and occlusion of sensor data make it impossible to obtain some behavior information of enemy targets in a timely manner. Existing methods have not effectively addressed this challenge.

[0005] Therefore, there is an urgent need for a new battlefield intention prediction method that can comprehensively consider the dynamic threat fields of multiple targets, overcome the problem of information incompleteness, and predict the action intentions of enemy targets in real time and accurately, providing timely and effective support for decision-makers. Summary of the Invention

[0006] This application provides a behavior intention prediction method based on multi-object threat situation and GRU network to solve the technical problems of information incompleteness, multi-object cooperative combat, and dynamic changes in battlefield threats existing in the prior art.

[0007] To solve the above technical problems, a technical solution adopted by this application is: A behavior intention prediction method based on multi-object threat situation and GRU network, including the following steps:

[0008] Based on multiple threat factors, establish a static threat field model and calculate the static threat field value of each target to us;

[0009] Based on the speed, acceleration, orientation, and distance of the target, as well as the static threat field value, establish a dynamic threat field model and calculate the dynamic threat field value of each target to us;

[0010] Based on historical trajectory data and target state prediction information, update the static threat field value and the dynamic threat field value;

[0011] Based on the GRU model, perform temporal modeling on historical trajectory data, static threat field, and dynamic threat field data to obtain the individual intention prediction results of the target;

[0012] Based on the individual intention prediction results of each target, perform fusion processing on the overall intentions of multiple targets to predict the overall behavior intention; among them, the overall behavior intention includes attack, retreat, and reconnaissance.

[0013] Furthermore, the threat factors include the distance between us and the enemy, range, firepower, defense ability, and mobility.

[0014] Furthermore, the method for establishing a dynamic threat field model includes:

[0015] Based on formula (1), obtain the dynamic threat field model; where formula (1) is:

[0016]

[0017] Where, is the static threat field value; C v 、C a 、C R 、C θ are the normalization factors of the target speed, acceleration, distance, and angle respectively, and α(t) is the time decay factor, which is described by an exponential decay function.

[0018] Furthermore, the method for updating the static threat field value and the dynamic threat field value based on historical trajectory data and target state prediction information includes:

[0019] When information is missing, based on the historical trajectory and the dynamic time warping algorithm, calculate the trajectory similarity of the target and provide a preliminary estimate for the initialization of particle filtering;

[0020] Based on the trajectory similarity of the target and the particle filtering method, obtain the target state prediction information and complete the missing information, and update the static threat field value and the dynamic threat field value.

[0021] Furthermore, the method for initializing particle filtering includes:

[0022] Based on formula (2), obtain the initialization formula of the particle filter; where formula (2) is:

[0023]

[0024] where T (i) is a partial trajectory of the current target, T (k) is a historical trajectory, and β is a regulation factor. Further, the method for obtaining the individual intention prediction result of the target includes:

[0025] Perform standardization processing on the historical trajectory data, splice the static threat field value and the dynamic threat field value into a joint vector, and map it to a threat feature vector through a fully connected layer.

[0026] Splice the historical trajectory data and the threat feature vector along the feature channel dimension to form a fused input vector:

[0027] Based on the input vector and the improved GRU model, perform temporal modeling and infer the individual intention prediction result of the target.

[0028] Further, the improved GRU model replaces the traditional tanh function with the Mish activation function and introduces a dual-field gating module, and calculates the weight ratio of the static threat field and the dynamic threat field through a fully connected layer, where the weight ratio is calculated through a gating vector.

[0029] Further, the method for performing the fusion processing of the overall intentions of multiple targets includes:

[0030] Based on formula (3), obtain the calculation weight of each target; where formula (3) is:

[0031]

[0032] where ω (i) is the calculation weight of the target, is the static threat field value, is the dynamic threat field value, i is the target serial number, and λ is the dynamic threat proportion coefficient; perform weighted averaging on the predicted results of the behavioral intentions of each target to obtain the probability distribution of the overall behavioral intention.

[0033] Further, the probability distribution of the overall behavioral intention is calculated through formula (4); where formula (4) is:

[0034]

[0035] where represents the predicted probability of target i for category c, and N is the total number of targets.

[0036] The beneficial effects of this application are as follows: This application takes into account the mutual influence between targets and the comprehensive effects of multiple targets, and realizes the efficient prediction of the overall behavior intention of the enemy. This application can dynamically adjust the influence of static and dynamic threat field information on target behavior prediction, and improve the prediction reliability in the case of incomplete information or missing sensor data. Among them, the global intention fusion mechanism flexibly combines the prediction results of multiple targets through weighted average and weighted voting methods, ensuring the accurate prediction of the overall behavior intention in a complex environment with multiple targets coexisting. It enhances the adaptability in a changing battlefield environment, can accurately infer the future behavior intention of the target, and thus effectively cope with the challenges of information asymmetry and dynamic changes in the modern battlefield environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 is a schematic flowchart of an embodiment of the method for predicting behavior intention based on multi-target threat situation and GRU network of this application;

[0038] Figure 2 is Figure 1 a schematic diagram of the threat field of a tank in an embodiment of step S1 in

[0039] Figure 3 is Figure 1 a schematic diagram of the threat field of an armored vehicle in an embodiment of step S1 in

[0040] Figure 4 is Figure 1 a schematic diagram of the threat field of an artillery vehicle in an embodiment of step S1 in

[0041] Figure 5 is Figure 1 a schematic diagram of threat field modeling in an embodiment of step S2 in

[0042] Figure 6 is a step diagram for processing information incompleteness of an embodiment of the method for predicting behavior intention based on multi-target threat situation and GRU network of this application;

[0043] Figure 7 is Figure 1 a GRU model framework diagram of a tank in an embodiment of step S4 in DETAILED DESCRIPTION OF THE EMBODIMENTS

[0044] To make the objectives, technical solutions and advantages of the present invention clearer, the following further elaborates on the present invention in detail with reference to specific embodiments.

[0045] Many specific details are set forth in the following description to facilitate a thorough understanding of the present invention. However, the present invention may be implemented in other ways different from those described herein. Therefore, the present invention is not limited by the specific embodiments disclosed in the following specification.

[0046] See also Figure 1 , Figure 1 This is a flow chart of an embodiment of a method for predicting behavioral intentions based on multi-target threat situations and GRU networks of the present application. The prediction method includes the following steps:

[0047] Step S1. Based on multiple threat factors, a static threat field model is established and the static threat field value of each target to our party is calculated.

[0048] Specifically, based on the five static threat factors of enemy-friendly distance d, range l, firepower A, defense capability D and mobility M, the static threat field model is used to calculate the threat intensity of each enemy target to our side. Considering the physical characteristics of different types of enemy targets, the static threat field model is adjusted according to the target characteristics.

[0049] like Figure 2 As shown in the figure, the tank's powerful firepower and defense capabilities make its threat field strongest at medium distances, and the threat decays faster as the distance increases. Specifically, the static threat field of the tank is expressed as:

[0050]

[0051] in, Indicates the threat intensity of a point (x, y), l optimal It is the most threatening mid-range position, σ tank is the threat diffusion coefficient of the tank.

[0052] like Figure 3 As shown in the figure, although the armored vehicle has strong mobility, its firepower and defense capabilities are weak, and the threat field of the armored vehicle has a greater impact within a short distance. Among them, the static threat field function of the armored vehicle is:

[0053]

[0054] in, Indicates the threat intensity of a point (x, y), σ armor is the threat diffusion coefficient of the armored vehicle.

[0055] like Figure 4 As shown in the figure, the artillery vehicle has strong long-range firepower, but weak mobility. The threat field of the artillery vehicle is strongest within the range and decays rapidly beyond the range. The static threat field function of the artillery vehicle is expressed as:

[0056]

[0057] in, Indicates the threat intensity of a point (x, y), l max is the maximum range of the artillery vehicle, σ cannon is the threat diffusion coefficient of the artillery vehicle.

[0058] Step S2. Based on the speed, acceleration, orientation, and distance of the target, as well as the static threat field value, establish a dynamic threat field model and calculate the dynamic threat field value of each target to our side.

[0059] Specifically, on the modern battlefield, the threat value of an enemy target changes dynamically with its motion state. For example Figure 5 As shown, through the dynamic threat field model, the threat intensity is adjusted by combining the four dynamic threat factors of the distance, speed, acceleration, and orientation of the enemy target. The calculation formula for the dynamic threat field value is:

[0060]

[0061] where is the static threat field value obtained from the above calculation; C v , C a , C R , C θ are the normalization factors of the target speed, acceleration, distance, and angle respectively, and α(t) is the time decay factor, which is described by an exponential decay function.

[0062] Step S3. Update the static threat field value and the dynamic threat field value based on historical trajectory data and target state prediction information.

[0063] Specifically, in the modern battlefield, target state information is often missing due to sensor failures, environmental occlusion, or enemy camouflage. To solve this problem, the particle filter method is used to dynamically update the threat field by combining historical trajectories and target state predictions, supplement missing information, and improve the accuracy of future behavior predictions.

[0064] Step S3 includes:

[0065] Step S31. When information is missing, based on historical trajectories and the dynamic time warping algorithm, calculate the trajectory similarity of the target and provide a preliminary estimate for the initialization of the particle filter.

[0066] Specifically, as Figure 6 shown, when the observed data is missing, first provide a preliminary estimate for the target state prediction through historical trajectory matching and target behavior pattern recognition. The initialization of the particle filter is completed through the following steps:

[0067] Retrieve the trajectory similar to the current target behavior pattern from the historical trajectory library, calculate the trajectory similarity using the dynamic time warping (DTW) algorithm, and match the behavior patterns of the current target and the historical trajectory:

[0068]

[0069] where T (i)is a partial trajectory of the current target, T (k) is the historical trajectory, and β is the adjustment factor; the k historical trajectories with the highest similarity are selected to initialize the particle set.

[0070] Step S32. Based on the trajectory similarity of the target and the particle filter method, obtain the target state prediction information, complete the missing information, and update the static threat field value and the dynamic threat field value.

[0071] Specifically, the particle filter predicts the future state of the target according to the historical data through the constant acceleration model, fills in the missing information, and provides a new state estimate for the update of the threat field. The update method of the threat field is realized through the weighted particle filter result:

[0072] The particle weight is calculated by the following formula:

[0073]

[0074] where Neighbor(i) = {j∣d ij <R th} is the set of neighboring targets of target i, and σ is the noise parameter. This formula considers the state difference between the target and its neighboring targets, dynamically adjusts the weight of the target, and reflects the relative behavior of the target and its surrounding environment.

[0075] Then, the cooperative particle filter is introduced to further complete the information. By the cooperative motion constraint of neighboring targets, the particle state is adjusted by correcting the target angle increment:

[0076]

[0077] where d ij represents the distance between target i and target j; R th is the neighborhood distance threshold, defined as the distance at which targets are considered to influence each other; γ is the cooperation coefficient, reflecting the intensity of the motion convergence between targets.

[0078] The dynamically predicted target state is used to update the threat field and supplement the missing part of the information. Based on the particle filter method, the threat field update formula based on the predicted state is:

[0079]

[0080] where is the threat field increment calculated based on the predicted trajectory, and α(t) is the time decay factor. This ensures that even if data is missing, the dynamic threat field of the target can still be reasonably estimated.

[0081] Step S4. Based on the GRU model, perform temporal modeling on the historical trajectory data, static threat field, and dynamic threat field data to obtain the individual intention prediction result of the target.

[0082] Specifically, as Figure 7 shown, by introducing a gated recurrent unit (GRU) model, the historical trajectory data, static threat field data, and dynamic threat field data of the target are modeled in time series to infer the future behavior intention of the target. Specifically, first, the historical trajectory data T t is normalized to eliminate the dimension difference, and the static threat field W s and the dynamic threat field W d (t) are concatenated into a joint vector [W s , W d (t)], and are mapped to a threat feature vector w t through a fully connected layer. The trajectory data and threat field data are concatenated along the feature channel dimension to form a fused input vector: x t = [T t , W s , W d (t)].

[0083] In the improved GRU model, the Mish activation function is used to replace the traditional tanh function to enhance the non-linear expression ability for complex threat patterns. At the same time, a dual-field gating module is introduced to calculate the weight ratio of the static threat field and the dynamic threat field through a fully connected layer, and adaptively adjust the fusion strength of the threat information and the historical trajectory data. The calculation formula for the threat weight gating vector is:

[0084] g t = σ(W g ·w t + b g ) (11);

[0085] where W g is the weight of the gating module, b g is the bias term, and σ(·) is the sigmoid function.

[0086] For the reset gate and the update gate, orthogonal initialization and LeCun adaptive initialization strategies are adopted to ensure the stability of the gradient and the convergence efficiency during the training process.

[0087] Through the improved GRU network, the input data sequence updates the hidden state h t layer by layer, and outputs the probability distribution of the target behavior intention through a fully connected layer and the Softmax function:

[0088]

[0089] where W o is the weight matrix of the output layer, and b o is the bias term.

[0090] Step S5. Based on the individual intention prediction results of each target, perform the fusion processing of the overall intentions of multiple targets to predict the overall behavior intention; wherein, the overall behavior intention includes attack, retreat, and reconnaissance.

[0091] Specifically, based on the individual intention prediction results of each target, perform the fusion processing of the overall intentions of multiple targets, comprehensively consider the mutual relationship between multiple targets, threat field information, and their behavior patterns, and predict the overall behavior intention.

[0092] First, for each target, generate the corresponding behavior intention probability distribution based on its static and dynamic threat field data and the behavior intention predicted by the historical trajectory. Then, assign a weight to each target, and this weight is calculated based on the intensity of its current threat field data and the threat contribution of the target to the overall battlefield. Specifically, the weight calculation formula of the target is as follows:

[0093]

[0094] where λ ∈ [0, 1] is the dynamic threat proportion coefficient, which is used to balance the influence of dynamic threats and static threats.

[0095] Then, through the weighted average method, fuse the behavior intention prediction result of each target with its corresponding weight ω (i) to obtain the probability distribution of the overall behavior intention

[0096] where represents the prediction probability of target i for category c, and N is the total number of targets. The global enemy intention selects the category with the highest probability as the final result.

[0097] To further enhance the prediction accuracy of the overall intention, when the intention categories of multiple targets are highly similar, the system will conduct a weighted vote on the categories and select the category with the most votes as the final overall intention. If no category significantly leads, the final behavior intention is determined according to the weighted vote result.

[0098] Next, take a specific embodiment as an example to further elaborate on this specific implementation manner.

[0099] Embodiment 1

[0100] To verify the effectiveness of the present invention, a simulation data set including 9 groups of battlefield scenarios is constructed. Each group simulates 100 time steps (1 time step per second), and involves 3 types of enemy targets (tanks, armored vehicles, and artillery vehicles).

[0101] Table 1 shows the specific parameters of each target

[0102]

[0103] The initial positions are randomly distributed within a 5000m×5000m battlefield, the speed v ∈ [5, 30] m / s, and the acceleration a ∈ [-2, 2] m / s 2 , and the orientation angle θ ∈ [0, 360°]. The trajectory simulation uses a random motion model, updates the state every 1 second, and adds Gaussian noise (σ = 0.1).

[0104] For each set of simulation data, 20% of the time steps are randomly selected to lose the observation data (position and speed information) to simulate sensor failure or occlusion scenarios.

[0105] Step 1: Static threat field modeling. Calculate the static threat field value of each target according to the formula.

[0106] Taking the tank as an example, the calculation formula for the static threat field value is:[[]]

[0107]

[0108] where d is the real-time distance between the enemy and us.

[0109] Step 2: Dynamic threat field modeling. Calculate the dynamic threat field value by combining the dynamic factor. Taking the time step t = 50 as an example:

[0110]

[0111] where the time decay factor α(t) uses α(t) = e -0.1t .

[0112] Step 3: Information incompleteness processing. For the missing data, use the DTW algorithm to match the historical trajectory library (including 100 historical trajectories) and calculate the similarity:

[0113]

[0114] Select the 5 trajectories with the highest similarity to initialize the particle filter, predict the future state through the constant acceleration model, fill in the missing position information, and update the threat field:

[0115]

[0116] Step 4: GRU time series modeling and intention prediction. After normalizing the trajectory data T t , concatenate it with the static threat field W s , and the dynamic threat field W d (t) to form x t = [T t , W s , Wd (t)] to construct an input vector, and concatenate the static threat field W s and the dynamic threat field W d (t) into a joint vector [W s , W d (t)], and map it to a threat feature vector w through a fully connected layer t .

[0117] In the improved GRU model, the Mish activation function is adopted. The hidden layer contains 64 units, the training period is 100 times, the learning rate is 0.001, and the batch size is 32.

[0118] Step 5: Multi-objective intention fusion. The GRU outputs the behavior probabilities (attack, retreat, reconnaissance) of each target. Set the dynamic threat proportion coefficient λ = 0.7, and the weight formula is:

[0119]

[0120] After weighted averaging, select the category with the highest probability as the final prediction result.

[0121]

[0122] Table 2 shows the accuracy comparison (%) of different models under complete data and missing data

[0123]

[0124] Table 3 shows the prediction accuracy (%) of different intention categories

[0125]

[0126] Table 4 shows the comparison of information completion effects (MSE) of different methods

[0127]

[0128] In summary, the above experiments verify the feasibility of the proposed behavior intention prediction method based on multi-objective threat situation and GRU network.

[0129] The above is only an embodiment of the present application, and does not limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made by using the content of the specification and drawings of the present application, or directly or indirectly applied in other related technical fields, shall be equally included in the patent protection scope of the present application.

Claims

1. A method for predicting behavioral intentions based on multi-objective threat postures and GRU networks, characterized in that The following steps are involved: Based on multiple threat factors, a static threat field model is established and the static threat field value of each target to our side is calculated; Based on the speed, acceleration, direction and distance of the target and the static threat field value, a dynamic threat field model is established and the dynamic threat field value of each target to our party is calculated; Based on historical trajectory data and target state prediction information, updating the static threat field value and the dynamic threat field value; Based on the GRU model, time series modeling is performed on the historical trajectory data, the static threat field data, and the dynamic threat field data to obtain individual intention prediction results of the target; Based on the individual intention prediction results of each of the targets, the overall intentions of multiple targets are fused to predict the overall behavioral intention; wherein the overall behavioral intention includes attack, retreat, and reconnaissance.

2. The method according to claim 1, wherein The threat factors include enemy and friendly distance, range, firepower, defensive capabilities and mobility.

3. The method according to claim 1, characterized in that Methods for establishing a dynamic threat landscape model include: Based on formula (1), the dynamic threat field model is obtained; wherein, formula (1) is: Among them, is the static threat field value; C v , C a , C R , C θ are the normalization factors of the target speed, acceleration, distance, and angle respectively. α(t) is the time decay factor, which is described by an exponential decay function.

4. The method according to claim 1, wherein The method for updating the static threat field value and the dynamic threat field value based on historical trajectory data and target state prediction information includes: When information is missing, based on historical trajectories and a dynamic time warping algorithm, the trajectory similarity of the target is calculated and a preliminary estimate is provided for particle filter initialization; Based on the trajectory similarity of the target and the particle filtering method, the target state prediction information is obtained and the missing information is supplemented, and the static threat field value and the dynamic threat field value are updated.

5. The method according to claim 4, wherein The particle filter initialization method includes: Based on formula (2), the initialization formula of the particle filter is obtained; wherein the formula (2) is: Among them, T (i) is a partial trajectory of the current target, T (k) is a historical trajectory, and β is a regulation factor.

6. The method according to claim 1, wherein The method for obtaining the individual intention prediction result of the target includes: The historical trajectory data is standardized, the static threat field value and the dynamic threat field value are concatenated into a joint vector, and mapped into a threat feature vector through a fully connected layer. The historical trajectory data and the threat feature vector are concatenated along the feature channel dimension to form a fusion input vector: Based on the input vector and the improved GRU model, time series modeling is performed and the individual intention prediction results of the target are inferred.

7. The method according to claim 6, wherein The improved GRU model replaces the traditional tanh function with the Mish activation function, introduces a dual-field gating module, and calculates the weight ratio of the static threat field and the dynamic threat field through a fully connected layer, wherein the weight ratio is calculated through a gating vector.

8. The method according to claim 7, characterized in that, The method for fusion processing of the overall intentions of multiple targets includes: Based on formula (3), the calculation weight of each target is obtained; wherein the formula (3) is: Among them, ω (i) is the target calculation weight, is the static threat field value, is the dynamic threat field value, i is the target serial number, and λ is the dynamic threat proportion coefficient; the prediction results of the behavioral intentions of each target are weighted and averaged to obtain the probability distribution of the overall behavioral intention.

9. The method according to claim 8, characterized in that, The probability distribution of the overall behavioral intention is calculated by formula (4); wherein, formula (4) is: Among them, represents the predicted probability of target i for category c, and N is the total number of targets.

Images