Unmanned aerial vehicle autonomous strike decision adaptive adjustment system based on multi-source information fusion

Abstract

The invention discloses an unmanned aerial vehicle autonomous strike decision self-adaptive adjustment system based on multi-source information fusion, and particularly relates to the technical field of unmanned aerial vehicle autonomous control and self-adaptive decision, the system comprises an information fusion and decision inner ring and a meta-cognitive self-adaptive adjustment outer ring, the inner ring executes multi-source information fusion and generates a strike decision, and the outer ring executes a meta-cognitive self-adaptive adjustment. And the outer ring evaluates the environment and information quality in real time, drives the adjustment strategy generator based on reinforcement learning by comparing the expected efficiency and the actual efficiency of the decision, and dynamically adjusts the fusion weight and the decision parameter of the inner ring or switches the decision model. The problems that an existing system is poor in environmental adaptability and cannot achieve online self-optimization are solved, autonomous, closed-loop and continuous optimization of decision-making performance in a dynamic battlefield environment is achieved, and the autonomous combat effectiveness, the environmental adaptability and the task robustness of the unmanned aerial vehicle are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of unmanned aerial vehicle autonomous control and adaptive decision-making, more particularly, the present application relates to a kind of unmanned aerial vehicle autonomous attack decision-making adaptive adjustment system based on multi-source information fusion. BACKGROUND

[0002] In prior art, unmanned aerial vehicle autonomous attack decision-making system generally relies on multi-source information fusion and preset decision-making rules or offline training intelligent model, such system usually works under fixed or preset fusion weight, its decision logic or model parameters remain unchanged after deployment, lack of real-time adaptive ability to environmental dynamic changes and information quality fluctuations, when facing complex and changeable battlefield environment, such as target rapid maneuvering, sudden strong electromagnetic interference or partial sensor performance degradation, the decision accuracy and robustness of existing system will be significantly reduced.

[0003] Although there are some researches introducing adaptive filtering or simple feedback mechanism, but its adjustment mode is mostly local, open-loop or based on fixed rules, cannot carry out closed-loop, online global optimization from the overall decision efficiency of system, also does not have the ability of self-learning and evolution from continuous task execution.

[0004] Therefore, how to make unmanned aerial vehicle decision-making system have high environmental adaptability, online self-optimization ability and strong robustness to deal with uncertainty, is still a key technical problem to be solved in the field. SUMMARY

[0005] In order to overcome the above-mentioned defects of prior art, the present application provides a kind of unmanned aerial vehicle autonomous attack decision-making adaptive adjustment system based on multi-source information fusion, to solve the problems raised in the above background art.

[0006] To achieve the above-mentioned purpose, the present application provides the following technical scheme: a kind of unmanned aerial vehicle autonomous attack decision-making adaptive adjustment system based on multi-source information fusion, including information fusion and decision-making inner loop and metacognition adaptive adjustment outer loop;

[0007] The information fusion and decision-making inner loop includes:

[0008] Multi-source information perception module, for acquiring environment perception data from multiple heterogeneous sensors in real time;

[0009] Adaptive information fusion module, for fusing environment perception data according to dynamic fusion weight strategy, outputting comprehensive situation information;

[0010] Core decision-making module, for generating attack decision-making instructions according to dynamic decision-making parameters and comprehensive situation information;

[0011] Decision execution and feedback module, for executing instructions and collecting battle result feedback data;

[0012] The metacognition adaptive adjustment outer loop comprises:

[0013] An environment and information quality assessment unit is configured to calculate an environment dynamic index representing the complexity of environmental changes and a confidence index of each information source in real time;

[0014] An effectiveness prediction and verification unit is configured to predict an expected effectiveness value before decision making and calculate an actual effectiveness value based on battle feedback data after decision execution;

[0015] An adjustment strategy generator is configured to compare the expected and actual effectiveness values, and generate adjustment instructions for the dynamic fusion weight strategy and / or dynamic decision parameters in combination with the changes in the environment dynamic index and the confidence index, to optimize the decision performance of the inner loop in a closed loop.

[0016] Preferably, the adjustment strategy generator is further configured to generate a model switching instruction when the environment dynamic index exceeds a preset complexity threshold and the confidence index of the key information source is lower than a preset reliability threshold;

[0017] The core decision module pre-stores a decision model library and switches from a first decision model currently used to a second decision model with lower computational complexity or stronger robustness in response to the model switching instruction.

[0018] Preferably, the expected and actual effectiveness values are calculated based on the same multi-dimensional effectiveness index system, which at least includes two of the predicted damage probability, the predicted task completion time and the estimated risk value of the own platform.

[0019] Preferably, the dynamic fusion weight strategy adopted by the adaptive information fusion module specifically assigns variable weighting coefficients to data from visible light, infrared, radar and electronic support measurement sensors, and the weighting coefficients are adjusted in real time by the adjustment instructions output by the adjustment strategy generator.

[0020] Preferably, the dynamic decision parameters include at least one of the weight factor used in target threat assessment, the risk aversion coefficient used in attack path planning, or the confidence threshold used in shooting solution in the core decision module.

[0021] Preferably, the environment dynamic index is quantified by calculating at least one of the target motion state change rate per unit time, the frequency of new threat appearance, and the electromagnetic signal density in the preset airspace.

[0022] Preferably, the adjustment strategy generator is constructed based on a reinforcement learning algorithm, and the environment dynamic index and the confidence index are combined as state input, the adjustment action on the fusion weight or the decision parameter is output, and the actual effectiveness value is taken as a reward signal for online policy optimization.

[0023] Preferably, the decision model library comprises at least one expert system model based on preset rules and one intelligent model based on a deep neural network, and the expert system model is set to be enabled when the second decision model.

[0024] Preferably, the system is deployed on a UAV platform in the form of an on-board embedded system, and the optimization process of the meta-cognitive adaptive adjustment outer loop is autonomously completed online within a single task cycle.

[0025] Technical effects and advantages of the present application:

[0026] 1. The present application realizes online and closed-loop optimization of the attack decision process by introducing a double-loop collaborative architecture composed of a meta-cognitive adaptive adjustment outer loop and an information fusion and decision inner loop. The system can evaluate the environmental dynamics and information quality in real time, and dynamically adjust the fusion strategy and decision parameters of the inner loop based on the performance feedback mechanism of prediction-verification, thereby significantly improving the autonomous decision adaptability and overall task performance of the UAV in a high-dynamic and strongly uncertain battlefield environment.

[0027] 2. The system has the ability to continuously learn and evolve from historical experience through the online optimization mechanism of the adjustment strategy generator based on reinforcement learning. The system can automatically optimize its adjustment strategy according to actual task feedback, realizing the transition from static and fixed rules to dynamic and data-driven decision adjustment mode, thereby showing stronger intelligence and scene generalization ability when facing novel or complex scenes not covered by training data.

[0028] 3. The integrated decision model library and model switching mechanism effectively enhance the robustness and survivability of the system. When detecting that extreme environment causes key information sources to fail, the system can automatically trigger a quick switch from a complex intelligent model to a lightweight and highly robust rule-based model, and simultaneously adjust the fusion strategy, ensuring that the system can still maintain basic reliable decision-making functions under the condition of partial sensor performance degradation or information loss, avoiding the collapse of overall performance and improving the fault tolerance and success rate of task execution. BRIEF DESCRIPTION OF DRAWINGS

[0029] Figure 1 The system architecture diagram provided for the embodiments of the present application.

[0030] Figure 2 The adaptive adjustment workflow diagram provided for the embodiments of the present application.

[0031] The figure marks are: 100, information fusion and decision inner loop; 101, multi-source information perception module; 102, adaptive information fusion module; 103, core decision module; 104, decision execution and feedback module;

[0032] 200, meta-cognition adaptive adjustment outer loop; 201, environment and information quality assessment unit; 202, performance prediction and verification unit; 203, adjustment strategy generator. DETAILED DESCRIPTION

[0033] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application.

[0034] Embodiment 1,

[0035] As shown in the accompanying drawings, Figure 1 The embodiment provides an unmanned aerial vehicle autonomous attack decision adaptive adjustment system based on multi-source information fusion, which comprises an information fusion and decision inner loop 100 and a meta-cognition adaptive adjustment outer loop 200.

[0036] The information fusion and decision inner loop 100 is a basic execution layer of the system, comprising:

[0037] The multi-source information perception module 101 integrates multiple heterogeneous sensors such as visible light cameras, infrared thermal imagers, millimeter wave radars, electronic support measurement sensors and inertial navigation systems, which collect real-time perception data of the environment where the unmanned aerial vehicle is located, including image sequences, point cloud data, radio frequency signals and positioning information, etc.

[0038] The adaptive information fusion module 102 adopts an improved weighted fusion algorithm, and the initial weight is set as the default confidence of each sensor. The module receives the original data from the multi-source information perception module 101, and performs fusion processing on the spatio-temporal aligned data to generate comprehensive situation information containing target position, velocity, attribute and threat level.

[0039] The core decision module 103 internally has an intelligent decision model based on a deep neural network as a first decision model. The module receives the comprehensive situation information, combines the currently loaded dynamic decision parameters including target threat evaluation weight and path planning risk coefficient, etc., and generates specific attack decision instructions including target allocation, attack path, attack opportunity and weapon selection through inference calculation.

[0040] The decision execution and feedback module 104 converts the attack decision instructions into control instructions to drive the unmanned aerial vehicle flight control system, optoelectronic pod and weapon launching device to execute the attack task. At the same time, the module collects damage evaluation images and target state changes after attack, etc. battle result feedback data through airborne sensors.

[0041] The meta-cognition adaptive adjustment outer loop 200 is an intelligent optimization layer of the system, comprising:

[0042] The environment and information quality evaluation unit 201: real-time monitoring of the running state of the inner ring and the change of the environment, specifically, by calculating the change rate of the target machine acceleration in unit time to quantify the environmental dynamic index, by analyzing the signal-to-noise ratio, data update rate and consistency with other sensors of each sensor data to quantify the confidence index of each information source, for example, when the contrast of infrared image is reduced due to smoke interference, its confidence index is correspondingly reduced;

[0043] The performance prediction and verification unit 202: contains two functional subunits, the performance prediction subunit estimates the expected performance value of the decision based on the current environmental dynamic index, the confidence index of each sensor and the historical combat data of similar scenes when the core decision module 103 generates the attack decision instruction, the value is calculated by the weighted calculation of the predicted damage probability, the predicted task completion time and the risk value of the enemy, and the performance verification subunit analyzes the battle result feedback data after the attack task is executed to calculate the actual performance value of this decision;

[0044] The adjustment strategy generator 203: as the core of the outer ring, the adjustment strategy generator 203 is realized by a deep reinforcement learning model, which takes the environmental dynamic index and the confidence index of each information source as input, takes the adjustment instruction of the inner ring parameter as output, and takes the actual performance value as training reward, for example, in a typical task, the generator compares and finds that the actual damage probability of a certain attack is significantly lower than the expected value, and the confidence index of the radar sensor suddenly drops during the attack stage, combined with these information, the generator outputs the adjustment instruction: reduce the weight of radar data in fusion, and fine-tune the decision threshold of attack opportunity in the core decision model, the adjustment is real-time issued to the corresponding module of the inner ring through the data bus.

[0045] The working process of the system is shown in the accompanying Figure 2 figure, forming a closed loop of perception-decision-execution-evaluation-adjustment, in a complete task, the outer ring can complete multiple adjustment cycles to realize online self-optimization in a single task cycle.

[0046] Embodiment 2,

[0047] This embodiment further elaborates the cooperative working mode of the adjustment strategy generator 203 and the core decision module 103 under specific conditions based on the system architecture of embodiment 1, that is, the model switching function.

[0048] When the unmanned aerial vehicle enters a strong electromagnetic interference area, the environment and information quality evaluation unit 201 monitors that the environmental dynamic index calculated by the electromagnetic signal density increases sharply and exceeds the preset threshold, and the confidence index of GPS and radar and other key information sources decreases below the reliability threshold, the adjustment strategy generator 203 generates a model switching instruction according to the preset emergency rule.

[0049] The core decision module 103 pre-stores a decision model library, in addition to the default deep neural network model as the first decision module, also contains an expert system model based on preset rule conditions as the second decision model, which does not rely on accurate coordinate information and continuous sensor data stream, but makes decisions based on relative motion logic and discrete observations of photoelectric sensors, although the decision accuracy decreases, but the calculation speed is fast, and the anti-interference ability is strong;

[0050] After receiving the switching instruction, the core decision module 103 completes the model switching within milliseconds, and the unmanned aerial vehicle uses the expert system model for decision-making, at the same time, the adaptive information fusion module 102 also switches the fusion strategy from weighted fusion to selective fusion based on photoelectric data according to the instruction, so that the unmanned aerial vehicle can still maintain basic combat capability in extreme environment, successfully escape from the interference area or complete emergency attack.

[0051] Embodiment 3,

[0052] This embodiment is based on embodiment 1 and embodiment 2, and details the online optimization process and implementation mechanism of the adjustment strategy generator 203 based on the reinforcement learning algorithm:

[0053] The adjustment strategy generator 203 uses the proximal policy optimization algorithm framework to realize the generation and iteration of its intelligent adjustment strategy, and its core is to establish a closed-loop learning framework with system state as input, adjustment action as output, and decision effectiveness as optimization target. The core elements of the online optimization process are defined as follows:

[0054] 1. State definition: at each decision adjustment time, the state of the system is represented by a state vector, which is composed of the environmental dynamic index and the real-time confidence index of each main information source , specifically, the state vector is used to fully represent the current environmental complexity and information source reliability of the system, and its expression is:

[0055] ;

[0056] wherein,

[0057] : environmental dynamic index, quantified by the weighted sum of the target motion state change rate, the new threat appearance frequency, and the electromagnetic signal density in the preset airspace within a unit time (weighting coefficients are 0.4, 0.3, and 0.3 respectively), the value range is [0, 1], the larger the value, the more intense the environmental change;

[0058] : Visible light sensor confidence index, calculated based on sensor data signal-to-noise ratio (weight 0.5), data update rate (weight 0.3), and consistency with other sensor data (weight 0.2), with a value range of [0, 1], and a larger value indicating higher data reliability.

[0059] : Infrared sensor confidence index, with the same calculation logic as the visible light sensor, with a value range of [0, 1].

[0060] : Radar sensor confidence index, with the same calculation logic as the visible light sensor, with a value range of [0, 1].

[0061] : Electronic support measurement sensor confidence index, with the same calculation logic as the visible light sensor, with a value range of [0, 1].

[0062] 2. Action definition: Adjust the action output by the regulation strategy generator 203 is a regulation instruction vector, directly corresponding to the adjustment amount of the adjustable parameters of the information fusion and decision-making inner loop 100, a typical action vector corresponds to the adjustment instruction for the inner loop fusion weight and decision-making parameter, and its expression is: ;

[0063] Among them,

[0064] : Visible light sensor fusion weight adjustment increment, with a value range of [-0.2, 0.2], a positive value indicating an increase in weight, and a negative value indicating a decrease in weight.

[0065] : Infrared sensor fusion weight adjustment increment, with a value range of [-0.2, 0.2], with the same logic as above.

[0066] : Radar sensor fusion weight adjustment increment, with a value range of [-0.2, 0.2], with the same logic as above.

[0067] : Electronic support measurement sensor fusion weight adjustment increment, with a value range of [-0.2, 0.2], with the same logic as above.

[0068] : Core decision parameter adjustment increment, adjusting key parameters such as target threat assessment weight factor, attack path planning risk avoidance coefficient, and shooting data calculation confidence threshold, with a value range of [-0.1, 0.1], and the specific adjustment object is determined by the current system state.

[0069] 3. Reward calculation: Reward signal According to the matching degree of actual performance and expected performance after decision execution, and combined with task efficiency, the reward function is calculated for evaluating the pros and cons of the adjustment action For evaluating the pros and cons of the adjustment action, the core goal is to maximize the balance of decision performance and task efficiency, and its expression is:

[0070] ;

[0071] Among them,

[0072] : Actual damage probability after decision execution, calculated based on battle feedback data, with a value range of [0, 1];

[0073] : Expected damage probability before decision generation, predicted based on historical similar scene data, with a value range of [0, 1];

[0074] : Performance matching degree index, the smaller the deviation between actual and expected damage probability, the larger the index value, and the higher the reward;

[0075] : Actual task time of this decision execution;

[0076] : Preset standard task time, preset according to task type and battlefield environment;

[0077] : Performance matching degree weight coefficient, with a value of 1, which can be adjusted according to actual combat requirements;

[0078] : Task timeliness weight coefficient, with a value of 0.5, which can be adjusted according to actual combat requirements;

[0079] Reward function The value range of the reward function is [-1.0, 1.0], and a positive value indicates that the adjustment action is effective, and a negative value indicates that the adjustment action needs to be optimized.

[0080] 4. Training and deployment process:

[0081] Offline training phase: First, in a simulated simulation environment containing various typical and extreme battlefield scenarios, drive the system to perform millions of task cycles, and the adjustment strategy generator 203 updates its internal policy network parameters through interaction with the environment, such as observing the state , executing action , obtaining reward , and constantly updating its internal policy network parameters, finally learning the mapping relationship from complex state to optimal adjustment action;

[0082] Online application and fine-tuning phase: The well-trained policy network model is loaded into the UAV's onboard computing unit. During actual mission execution, the generator uses the real-time acquired state vectors... Generate adjustment action vectors This enables real-time, online adjustment of inner-loop parameters. Simultaneously, the system can utilize state-action-reward sequence data generated in real tasks to perform small-scale online fine-tuning of the policy network, making it better adaptable to the specific characteristics of the actual deployment environment.

[0083] Through the above-mentioned online optimization mechanism based on reinforcement learning, the policy generator 203 enables the entire system to learn from experience, ultimately achieving the goal of autonomously and continuously optimizing its decision-making performance in a dynamic and uncertain battlefield environment.

[0084] In summary:

[0085] This invention constructs a two-layer intelligent system comprising an inner loop 100 for information fusion and decision-making and an outer loop 200 for metacognitive adaptive adjustment. The inner loop is responsible for generating and executing strike decisions based on multi-source sensor data fusion, while the outer loop monitors environmental complexity and information quality in real time. By comparing the predictive effectiveness of the decisions with the actual combat results, it drives a reinforcement learning-based adjustment strategy generator 203 to dynamically and in a closed loop adjust the fusion weights, decision parameters, and even switch decision models of the inner loop. This enables the entire system to autonomously and continuously optimize its decision-making performance in a complex and dynamic battlefield environment.

[0086] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A UAV autonomous strike decision-making adaptive adjustment system based on multi-source information fusion, characterized in that: It includes an inner loop of information fusion and decision-making (100) and an outer loop of metacognitive adaptive regulation (200). The information fusion and decision-making inner loop (100) includes: A multi-source information sensing module (101) is used to acquire environmental sensing data from multiple heterogeneous sensors in real time; The adaptive information fusion module (102) is used to fuse environmental perception data according to a dynamic fusion weight strategy and output comprehensive situation information; The core decision module (103) is used to generate strike decision commands based on dynamic decision parameters and comprehensive situation information; The decision execution and feedback module (104) is used to execute instructions and collect battle results feedback data; The metacognitive adaptive regulation outer loop (200) includes: The Environmental and Information Quality Assessment Unit (201) is used to calculate in real time the environmental dynamic index, which characterizes the complexity of environmental changes, and the confidence index of each information source. The performance prediction and verification unit (202) is used to predict the expected performance value before the decision is generated and to calculate the actual performance value based on the battle result feedback data after the decision is executed. The adjustment strategy generator (203) is used to compare the expected and actual performance values ​​and, in combination with the changes in the environmental dynamic index and confidence index, generate adjustment instructions for the dynamic fusion weight strategy and / or dynamic decision parameters to optimize the decision performance of the inner loop in a closed loop.

2. The UAV autonomous strike decision adaptive adjustment system based on multi-source information fusion according to claim 1, characterized in that: The adjustment strategy generator (203) is further configured to generate a model switching instruction when the environmental dynamic index exceeds a preset complexity threshold and the confidence index of the key information source is lower than a preset reliability threshold. The core decision module (103) has a pre-stored decision model library and, in response to a model switching command, switches from the currently used first decision model to a second decision model with lower computational complexity or stronger robustness.

3. The UAV autonomous strike decision adaptive adjustment system based on multi-source information fusion according to claim 1, characterized in that: The expected performance value and the actual performance value are calculated based on the same multi-dimensional performance index system, which includes at least two of the following: predicted damage probability, expected mission completion time, and estimated risk value of the platform.

4. The UAV autonomous strike decision adaptive adjustment system based on multi-source information fusion according to claim 1, characterized in that: The adaptive information fusion module (102) adopts a dynamic fusion weighting strategy, which specifically assigns variable weighting coefficients to data from visible light, infrared, radar and electronic support measurement sensors. The weighting coefficients are adjusted in real time by adjustment instructions output by the adjustment strategy generator (203).

5. The adaptive adjustment system for autonomous strike decision-making of unmanned aerial vehicles based on multi-source information fusion according to claim 1, characterized in that: The dynamic decision parameters include at least one of the following: the weighting factor used in the core decision module (103) for target threat assessment, the risk avoidance coefficient for attack path planning, or the confidence threshold for firing parameters calculation.

6. The UAV autonomous strike decision adaptive adjustment system based on multi-source information fusion according to claim 1, characterized in that: The environmental dynamic index is quantified by calculating at least one of the following: the rate of change of the target's motion state per unit time, the frequency of new threats, and the electromagnetic signal density in a preset airspace.

7. The adaptive adjustment system for autonomous strike decision-making of unmanned aerial vehicles based on multi-source information fusion according to claim 1, characterized in that: The adjustment strategy generator (203) is built based on the reinforcement learning algorithm. It takes the combination of the environmental dynamic index and the confidence index as the state input, the adjustment action of the fusion weight or decision parameters as the output, and the actual performance value as the reward signal to perform online strategy optimization.

8. The UAV autonomous strike decision adaptive adjustment system based on multi-source information fusion according to claim 2, characterized in that: The decision model library includes at least one expert system model based on preset rules and one intelligent model based on deep neural networks. The expert system model is set to be enabled when a second decision model is used.

9. The adaptive adjustment system for autonomous strike decision-making of unmanned aerial vehicles based on multi-source information fusion according to claim 1, characterized in that: The system is deployed on the UAV platform as an airborne embedded system, and the optimization process of the metacognitive adaptive adjustment outer loop (200) is completed autonomously online within a single mission cycle.

Images