Unmanned vehicle trajectory prediction method based on KAN
By combining the KAN deep learning framework with threat distribution maps for feature extraction and data preprocessing, the interpretability and accuracy issues of neural network trajectory prediction methods in multi-agent adversarial scenarios are solved, achieving interpretability and efficient prediction of target trajectories.
Patent Information
- Application Number
- CN202511699966.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-19
- Publication Date
- 2026-02-03
AI Technical Summary
Existing neural network-based trajectory prediction methods are difficult to interpret in multi-agent adversarial scenarios and cannot provide an explicit expression of the target trajectory, and they rely on a large amount of historical data.
The KAN deep learning framework is adopted. By replacing the weight parameters with learnable univariate functions, feature extraction and data preprocessing are performed in conjunction with threat distribution maps. Dual KANs are used to predict target trajectories, and fitting errors and visualization comparisons are performed to evaluate model accuracy and interpretability.
It improves the accuracy and interpretability of trajectory prediction, can extract effective data features with relatively small amounts of data, has few-sample learning capabilities, and can predict the future position of the target and obtain an explicit expression of the target's motion trajectory.
Smart Images
Figure CN121459286A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of multi-agent adversarial systems, and specifically designs a method for predicting the trajectory of unmanned vehicles based on KAN. Background Technology
[0002] Traditional target trajectory prediction methods often rely on large amounts of historical data and complex model parameters. Neural network-based prediction models represent the latest approach to trajectory prediction, but due to the "black box" nature of neural networks, the resulting solutions are difficult to interpret. Although various mature methods have been proposed for trajectory prediction both domestically and internationally, most known neural network-based methods can only predict the target's position within a certain future timeframe and cannot provide an explicit expression of the target's trajectory. KAN (Kolmogorov-Arnold Network), as a novel deep learning framework, replaces the weight parameters with learnable univariate functions, offering higher performance and better interpretability compared to traditional multilayer perceptron (MLP) models. In multi-agent adversarial scenarios, drone swarms searching for targets often generate relevant threat distribution maps based on historical information. These threat distribution maps can be used to predict target trajectories. KAN, through training, can obtain both the target's position within a certain future timeframe and potentially an explicit expression of the target's trajectory. Summary of the Invention
[0003] The purpose of this invention is to provide a KAN-based method for predicting the trajectory of unmanned vehicles, which can effectively improve the accuracy and interpretability of trajectory prediction.
[0004] Step 1: Perform a secondary screening on the threat distribution map generated by the drone swarm based on the search information to obtain a threat distribution map that is available for predicting target locations and trajectories;
[0005] Step 2: Extract features from the selected threat distribution map to obtain the approximate location of the target determined by the search from the dataset;
[0006] Step 3: Perform data preprocessing on the confirmed target locations to avoid gradient explosion during KAN training;
[0007] Step 4: Select appropriate KAN activation functions, basis functions, and neural network sizes for each direction of motion, and use dual KANs to predict the target trajectory respectively;
[0008] Step 5: Normalize the original data and align it with the predicted data. Visualize and compare it with the target theoretical position. Evaluate the model accuracy and interpretability from both the fitting error and visualization comparison aspects.
[0009] Furthermore, step one includes:
[0010] The drone swarm is first filtered to select the threat distribution map corresponding to drones that are still on missions and have not been disturbed, ensuring the accuracy and timeliness of the data. Then, a second filtering is performed to select the distribution map with high-threat areas. In other words, the threat distribution map containing the potential location information of the target can be obtained based on the search information during the drone swarm search process.
[0011] Furthermore, step two includes:
[0012] Target location features are extracted using either the high-threat region weighted midpoint method or a clustering-based location extraction method. The high-threat region weighted midpoint method extracts all high-threat region locations and performs a weighted average to comprehensively consider the influence of each high-threat region point, making it suitable for single-target scenarios. In multi-target scenarios, where it's not guaranteed that different targets can be found simultaneously, the clustering-based location extraction method uses a density-based spatial clustering algorithm to cluster high-threat points, making it more suitable for multi-target scenarios. If the extracted features are unstable in the early stages of the task, data truncation can be performed.
[0013] Furthermore, step three includes:
[0014] The target position is normalized within the search area, and the same normalization process is performed when calculating the theoretical trajectory of the target for subsequent visualization and comparison. Normalization can effectively avoid the gradient explosion problem. After the operations of "mesh refinement", "pruning" and "symbolic fitting", the training error and test error are significantly reduced, making each operation in the model more meaningful.
[0015] Furthermore, step four includes:
[0016] By trying different network structures and appropriately reducing the number of symbolic functions, the network structure of [1,3,5,1] was finally selected. The prediction task was split into two KANs that focus on one coordinate. The explicit expressions of x=f(t) and y=g(t) were trained respectively, which can reduce the complexity of model learning. Since the trained expression is the target trajectory in polar coordinates, the symbolic functions were selected as the linear function 'x' and the trigonometric functions 'cos' and 'sin'.
[0017] Furthermore, step five includes:
[0018] The raw data is normalized and aligned with the predicted data. The fitting error is used to evaluate the fitting effect of KAN on the target position obtained from the search information of the UAV swarm. The predicted trajectory and the theoretical trajectory are visualized and compared to evaluate the fitting effect of the trajectory expression predicted by KAN on the target theoretical trajectory, and to demonstrate the interpretability of its trajectory features.
[0019] Compared with existing technologies, the advantages of this invention are:
[0020] This invention addresses the "black box" effect of current neural network-based trajectory prediction algorithms by proposing a trajectory prediction method with partial interpretability. Theoretical analysis and simulation results show that this method fully utilizes data features, and even with relatively small amounts of data, it can still extract effective data features reflecting the target's partial motion characteristics in locally unstable and non-convergent situations. This small-sample learning ability demonstrates the strong inductive ability of this invention. The predicted target trajectory has predictive effectiveness in a short time, and partial motion characteristics of the target trajectory can be judged based on the fitted terms, exhibiting a certain degree of interpretability. Attached Figure Description
[0021] Figure 1 This is a basic framework diagram of an autonomous vehicle trajectory prediction model based on KAN.
[0022] Figure 2 This is a flowchart of the KAN model training process.
[0023] Figure 3 This is a visualization comparing the predicted and theoretical trajectories of three drones searching for one unmanned vehicle (P3C1) without data truncation. The theoretical trajectory is a circle with a diameter of 100 meters.
[0024] Figure 4 This is a visualization comparing the predicted and theoretical trajectories of four drones searching for one unmanned vehicle (P4C1) without data truncation. The theoretical trajectory is a circle with a diameter of 100 meters.
[0025] Figure 5 This is a visualization comparing the predicted and theoretical trajectories of 5 drones searching for 1 unmanned vehicle (P5C1) without data truncation. The theoretical trajectory is a circle with a diameter of 100 meters.
[0026] Figure 6 This is a visualization comparing the predicted and theoretical trajectories of six drones searching for one unmanned vehicle (P6C1) without data truncation. The theoretical trajectory is a circle with a diameter of 100 meters.
[0027] Figure 7This is a visualization comparing the predicted and theoretical trajectories of 7 drones searching for 1 unmanned vehicle (P7C1) without data truncation. The theoretical trajectory is a circle with a diameter of 100 meters.
[0028] Figure 8 This is a visualization comparing the predicted and theoretical trajectories of 8 drones searching for 1 unmanned vehicle (P8C1) without data truncation. The theoretical trajectory is a circle with a diameter of 100 meters.
[0029] Figure 9 This is a visualization comparing the predicted and theoretical trajectories of six drones searching for one unmanned vehicle (P6C1), with the data truncated. The theoretical trajectory is a circle with a diameter of 100 meters.
[0030] Figure 10 This is a visualization comparing the predicted and theoretical trajectories of seven drones searching for one unmanned vehicle (P7C1). The theoretical trajectory is a circle with a diameter of 100 meters.
[0031] Figure 11 This is a visualization comparing the predicted and theoretical trajectories of 8 drones searching for 1 unmanned vehicle (P8C1) using data truncation. The theoretical trajectory is a circle with a diameter of 100 meters.
[0032] Figure 12 This is a visualization comparing the long-term predicted trajectory and the theoretical trajectory of 7 drones searching for 1 unmanned vehicle (P7C1). The theoretical trajectory is a circle with a diameter of 100 meters. Detailed Implementation
[0033] The present invention will now be described in detail with reference to the embodiments shown in the accompanying drawings. However, it should be noted that these embodiments are not intended to limit the present invention. Equivalent changes or substitutions in function, method, or structure made by those skilled in the art based on these embodiments are all within the scope of protection of the present invention.
[0034] This embodiment provides a KAN-based method for predicting the trajectory of autonomous vehicles, including:
[0035] Step 1: Perform a secondary screening on the threat distribution map generated by the drone swarm based on the search information to obtain a threat distribution map that is available for predicting target locations and trajectories;
[0036] Step 2: Extract features from the selected threat distribution map to obtain the approximate location of the target determined by the search from the dataset;
[0037] Step 3: Perform data preprocessing on the confirmed target locations to avoid gradient explosion during KAN training;
[0038] Step 4: Select appropriate KAN activation functions, basis functions, and neural network sizes for each direction of motion, and use dual KANs to predict the target trajectory respectively;
[0039] Step 5: Normalize the original data and align it with the predicted data. Visualize and compare it with the target theoretical position. Evaluate the model accuracy and interpretability from both the fitting error and visualization comparison aspects.
[0040] This KAN-based autonomous vehicle trajectory prediction method fully utilizes data features to achieve short-term prediction of target trajectories. The predicted target trajectory exhibits a certain degree of interpretability and has the following advantages:
[0041] Compared to trajectory prediction tasks, the KAN-based autonomous vehicle trajectory prediction method makes full use of data features. With a relatively small amount of data, it can extract effective data features that reflect the motion features of the target. This small-sample learning ability demonstrates the strong inductive ability of this invention.
[0042] Compared to conventional trajectory prediction algorithms, the KAN-based autonomous vehicle trajectory prediction method can not only predict the future position of the target, but also train an explicit expression of the target's motion trajectory based on historical data, thus possessing a certain degree of interpretability.
[0043] The algorithm is convergent and can predict the trend of target change for locally unstable and non-convergent data. Therefore, even on "bad" datasets, the results predicted by the algorithm can still be close to the theoretical position of the target.
[0044] Therefore, a trajectory prediction algorithm with partial interpretability is proposed based on KAN for autonomous vehicle trajectory prediction.
[0045] In this embodiment, step one includes:
[0046] The drone swarm is first filtered to select the threat distribution map corresponding to drones that are still on missions and have not been disturbed, ensuring the accuracy and timeliness of the data. Then, a second filtering is performed to select the distribution map with high-threat areas. In other words, the threat distribution map containing the potential location information of the target can be obtained based on the search information during the drone swarm search process.
[0047] In this embodiment, step two includes:
[0048] Target location features are extracted using either the high-threat region weighted midpoint method or a clustering-based location extraction method. The high-threat region weighted midpoint method extracts all high-threat region locations and performs a weighted average to comprehensively consider the influence of each high-threat region point, making it suitable for single-target scenarios. In multi-target scenarios, where it's not guaranteed that different targets can be found simultaneously, the clustering-based location extraction method uses a density-based spatial clustering algorithm to cluster high-threat points, making it more suitable for multi-target scenarios. If the extracted features are unstable in the early stages of the task, data truncation can be performed.
[0049] In this embodiment, step three includes:
[0050] The target position is normalized within the search area, and the same normalization process is performed when calculating the theoretical trajectory of the target for subsequent visualization and comparison. Normalization can effectively avoid the gradient explosion problem. After the operations of "mesh refinement", "pruning" and "symbolic fitting", the training error and test error are significantly reduced, making each operation in the model more meaningful.
[0051] In this embodiment, step four includes:
[0052] By trying different network structures and appropriately reducing the number of symbolic functions, the network structure of [1,3,5,1] was finally selected. The prediction task was split into two KANs that focus on one coordinate. The explicit expressions of x=f(t) and y=g(t) were trained respectively, which can reduce the complexity of model learning. Since the trained expression is the target trajectory in polar coordinates, the symbolic functions were selected as the linear function 'x' and the trigonometric functions 'cos' and 'sin'.
[0053] In this embodiment, step five includes:
[0054] The raw data is normalized and aligned with the predicted data. The fitting error is used to evaluate the fitting effect of KAN on the target position obtained from the search information of the UAV swarm. The predicted trajectory and the theoretical trajectory are visualized and compared to evaluate the fitting effect of the trajectory expression predicted by KAN on the target theoretical trajectory, and to demonstrate the interpretability of its trajectory features.
[0055] The KAN-based autonomous vehicle trajectory prediction method provided in this embodiment has the following technical effects:
[0056] Compared to trajectory prediction tasks, the KAN-based autonomous vehicle trajectory prediction method makes full use of data features. With a relatively small amount of data, it can extract effective data features that reflect the motion features of the target. This small-sample learning ability demonstrates the strong inductive ability of this invention.
[0057] Compared to conventional trajectory prediction algorithms, the KAN-based autonomous vehicle trajectory prediction method can not only predict the future position of the target, but also train an explicit expression of the target's motion trajectory based on historical data, thus possessing a certain degree of interpretability.
[0058] The algorithm is convergent and can predict the trend of target change for locally unstable and non-convergent data. Therefore, even on "bad" datasets, the results predicted by the algorithm can still be close to the theoretical position of the target.
[0059] The detailed descriptions listed above are merely specific descriptions of feasible embodiments of the present invention, and are not intended to limit the scope of protection of the present invention. All equivalent embodiments or modifications made without departing from the spirit of the present invention should be included within the scope of protection of the present invention.
[0060] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, it is intended that all variations falling within the meaning and scope of equivalents of the claims be included within the present invention.
Claims
1. A trajectory prediction method based on KAN, characterized by: Step 1: Perform a secondary screening on the threat distribution map generated by the drone swarm based on the search information to obtain a threat distribution map that is available for predicting target locations and trajectories; Step 2: Extract features from the selected threat distribution map to obtain the approximate location of the target determined by the search from the dataset; Step 3: Perform data preprocessing on the confirmed target locations to avoid gradient explosion during KAN training; Step 4: Select appropriate KAN activation functions, basis functions, and neural network sizes for each direction of motion, and use dual KANs to predict the target trajectory respectively; Step 5: Normalize the original data and align it with the predicted data. Visualize and compare it with the target theoretical position. Evaluate the model accuracy and interpretability from both the fitting error and visualization comparison aspects.
2. The trajectory prediction method based on KAN according to claim 1, characterized in that, Step one includes: The drone swarm is first filtered to select the threat distribution map corresponding to drones that are still on missions and have not been disturbed, ensuring the accuracy and timeliness of the data. Then, a second filtering is performed to select the distribution map with high-threat areas. In other words, the threat distribution map containing the potential location information of the target can be obtained based on the search information during the drone swarm search process.
3. The trajectory prediction method based on KAN according to claim 1, characterized in that, Step two includes: Target location features can be extracted using either the high-threat region weighted midpoint method or a clustering-based location extraction method. The high-threat region weighted midpoint method extracts all high-threat region locations and performs a weighted average to comprehensively consider the influence of each high-threat region point, making it suitable for single-target scenarios. In multi-target scenarios, where it's not guaranteed that different targets can be simultaneously located, the clustering-based location extraction method uses a density-based spatial clustering algorithm to cluster high-threat points, making it more suitable for multi-target scenarios. If the extracted features are unstable in the early stages of the task, data truncation can be performed.
4. The trajectory prediction method based on KAN according to claim 1, characterized in that, Step three includes: The target position is normalized within the search area, and the same normalization process is performed when calculating the theoretical trajectory of the target for subsequent visualization and comparison. Normalization can effectively avoid the gradient explosion problem. After the operations of "mesh refinement", "pruning" and "symbolic fitting", the training error and test error are significantly reduced, making each operation in the model more meaningful.
5. The trajectory prediction method based on KAN according to claim 1, characterized in that, Step four includes: By trying different network structures and appropriately reducing the number of symbolic functions, the network structure of [1,3,5,1] was finally selected. The prediction task was split into two KANs that focus on one coordinate. The explicit expressions of x=f(t) and y=g(t) were trained respectively, which can reduce the complexity of model learning. Since the trained expression is the target trajectory in polar coordinates, the symbolic functions were selected as the linear function 'x' and the trigonometric functions 'cos' and 'sin'.
6. The trajectory prediction method based on KAN according to claim 1, characterized in that, Step five includes: The raw data is normalized and aligned with the predicted data. The fitting error is used to evaluate the fitting effect of KAN on the target position obtained from the search information of the UAV swarm. The predicted trajectory and the theoretical trajectory are visualized and compared to evaluate the fitting effect of the trajectory expression predicted by KAN on the target theoretical trajectory, and to demonstrate the interpretability of its trajectory features.