Cognitive decision security conversion and risk regulation and control method for intelligent autonomous unmanned system

Abstract

The invention provides a cognitive decision security conversion and risk regulation and control method for an intelligent autonomous unmanned system, belongs to the technical field of artificial intelligence, and aims to solve the problems of low confidence and inaccurate risk assessment result caused by the fact that the decision process of a current machine learning system is completely not externally controlled and cannot be objectively understood. The method comprises the following steps: S1, building a super real-time parallel simulation platform; s2, constructing a decision risk quantitative evaluation model; and S3, designing a real-time accurate risk regulation and control mechanism. According to the method, super-real-time parallel simulation software is adopted for large-scale comparison tests, the unmanned aerial vehicles successfully complete tasks or safely return voyage and evacuate through no less than secondary tests, and it is proved that the cognitive decision safe conversion and risk regulation and control method for the machine learning system has excellent performance and is suitable for popularization and application. And the safety of task execution of the unmanned aerial vehicle in the simulated battlefield is greatly improved.

Description

Technical Field

[0001] This invention relates to a method for safe conversion and risk control of cognitive decision-making in intelligent autonomous unmanned systems, belonging to the field of artificial intelligence technology. Background Technology

[0002] With the development of artificial intelligence technology, machine learning systems have been widely applied in many complex and dynamic fields, including unmanned aerial vehicles (UAVs) performing corresponding tasks. When UAVs perform missions, machine learning systems face complex battlefield environments. The safety and reliability of their cognitive decisions directly affect the operational safety of the UAV and the success or failure of the mission. UAVs also need to cope with changing weather and geographical conditions (wind, sunshine, day / night, altitude, terrain, etc.), constantly changing internal conditions (remaining fuel, current speed, etc.), and abruptly changing complex mission environments. In this highly adversarial environment, the flexibility of traditional rule-based algorithms is greatly limited, while machine learning systems, by learning and analyzing large amounts of flight data and environmental information, can control UAVs to perform multiple target tasks. However, machine learning systems still face many risks and challenges in controlling unmanned systems: (1) The uninterpretability of machine learning system decisions.

[0003] Because machine learning decision-making models are a kind of "black box decision-making," their decision results depend more on the training and inference of the internal network. This decision-making process is completely uncontrolled by the outside and cannot be objectively understood. This also leads to the problem that the output decision results of machine learning decision-making models are not very credible, and this problem indirectly threatens the operational safety of drones.

[0004] (2) Inaccuracy of risk assessment The accuracy of existing machine learning systems in risk assessment still needs improvement. Due to limited risk awareness and a lack of effective risk quantification methods, it is difficult to accurately assess the risk level of current decisions. For example, when a drone encounters multiple target threats, a machine learning decision model may still classify it as low-risk and issue a command to continue, potentially resulting in the drone being destroyed by the targets.

[0005] (3) Instability of risk control In complex and ever-changing execution environments, traditional risk control models struggle to adapt to rapid changes in the environment, tasks, and risk assessment levels. The control outputs of these models often become highly unstable due to these rapid environmental shifts. Therefore, a more precise risk control mechanism suitable for complex and dynamic environments is needed. Summary of the Invention

[0006] This invention addresses the problems of current machine learning systems' decision-making processes being completely uncontrolled by external factors and unable to be objectively understood, leading to low confidence levels, inaccurate risk assessment results, and unstable risk control. It proposes a method for safe conversion and risk control of cognitive decision-making in intelligent autonomous unmanned systems.

[0007] The technical solution adopted by the present invention to solve the above problems is as follows: The present invention includes the following steps: Step 1: Build an ultra-real-time parallel simulation platform and obtain key parameters of the UAV real-time scene based on the ultra-real-time parallel simulation platform; Step 2: Collect relevant data during the drone's mission execution process and historical data of the drone's mission execution in real time; Step 3: Construct a decision risk quantification assessment model based on relevant and historical data collected during the drone's mission execution process, and input the historical data of the drone's mission execution to calculate risk weights; Step 4: Obtain key parameters of relevant data from the real-time acquisition of UAV mission execution process using the ultra-real-time parallel simulation platform, and calculate the risk outcome of the current machine learning decision by matching Euclidean distance with historical data of UAV mission execution. Risk Based on risk outcome Risk Obtain the risk assessment level of the current machine learning system's decision-making process; Step 5: Based on the baseline security system, develop a corresponding real-time risk control mechanism for each risk assessment level to achieve precise real-time risk control.

[0008] Furthermore, step 1, building the ultra-real-time parallel simulation platform, specifically includes: Step 1.1: Set initial parameters, which include at least the initial number of drones, drone detection range, drone maximum speed, drone mission boundary, drone minimum flight altitude, drone maximum / initial fuel level, weather conditions, terrain factors, drone maximum acceleration, drone position, and drone speed; Step 1.2: Set the parameters for changes during UAV operation. These parameters should include at least the target position, the target range and the location of the task to be performed, the target radar position, the target moving speed, the UAV moving speed, weather changes, the current detection radius of the UAV, the current fuel level of the UAV, and the size / location of the navigation blind spot. Step 1.3: Based on the set initial parameters and the parameters that change during the operation of the UAV, construct the key parameters of the real-time UAV scene; Step 1.4 Set up the human-computer interaction page to visualize the key parameters of the drone's real-time scene.

[0009] Furthermore, step 2 specifically includes: Historical data on UAV missions includes mission execution results, environmental situation parameters at each moment, UAV's own state parameters at each moment, UAV mission execution parameters at each moment, and machine learning decision results at the corresponding moment.

[0010] Furthermore, step 3 specifically includes: Step 3.1: Normalize the environmental situation parameters, UAV state parameters, and UAV mission execution parameters in the historical data of UAV mission execution; Step 3.2: Introduce a time decay factor The Pearson coefficients between the normalized environmental situation parameters, the UAV's own state parameters, the UAV's mission execution parameters, and the mission execution results were calculated respectively. Step 3.3: Based on the Pearson coefficient and time decay factor between the corresponding parameters and the task execution results. Calculate the risk weights of the corresponding parameters, and sort the corresponding historical environmental situation parameters, UAV self-state parameters, and UAV mission execution parameters according to the magnitude of the risk weights; The expression for normalization is: (1); In formula (1), Indicates the first j The first historical test task in the group i The parameter of the first k The normalized value at each moment. and They represent the first i The maximum and minimum values ​​of each parameter in historical data; The formula for calculating the Pearson coefficient is: (2); In formula (2), For the first i The correlation coefficients between each parameter and the task results For the first j The parameter in the first... k The normalized value at each moment. For the first j The result of each task execution, if the task is successfully executed... If the task fails, then , and The first i The average of each parameter and the task result. For the first j The number of time slots for each task This is the time decay factor, used to measure the time difference between the current parameter in historical data and the end of the task; The formula for calculating risk weights is: (3); In formula (3), For the first i The parameters at time... t The weight, It is a moment t The time decay factor.

[0011] Furthermore, step 4 specifically includes: Step 4.1: Obtain the key parameters of the relevant data of the UAV in real time during the mission execution process from the ultra-real-time parallel simulation platform, repeat step 3 to obtain the risk weights of the current key parameters and sort them; Step 4.2: Use Euclidean distance to match the sorted current key parameters with historical data, select the historical data with the smallest Euclidean distance to the current key parameter, and calculate the risk result of the current machine learning system's decision based on the risk weights of the corresponding historical data parameters. Risk ,in, The value range is 0-1; Step 4.3: When At this point, the risk assessment level is A, and the decision made by the current machine learning system is considered safe. At that time, the risk assessment level is B, indicating that the current machine learning system's decision-making carries a certain probability of danger. At that time, the risk assessment level was C, indicating that the decision-making of the current machine learning system had a serious probability of danger; The formula for calculating Euclidean distance is: (4); In formula (4), For historical data, These are the current key parameters. n This represents the number of current key parameters; Risk outcomes of current machine learning decision-making Risk The calculation formula is: (5); In formula (5), Let t be the risk outcome value of the current machine learning decision at time t.

[0012] Furthermore, step 5 specifically includes: When the risk assessment level is A, there is no need to trigger the real-time risk precision control mechanism; When the risk assessment level is B, a decay coefficient is set. And calculate the decision weights of the machine learning system. According to the decision weights of the machine learning system Calculate the decision weights of the baseline security system Over time, gradually reduce the decision weights of the machine learning system. Increase the decision weight of the baseline security system ,when When the value is 1, the drone's task execution is entirely controlled by the machine learning system. At that time, with the attenuation coefficient The cumulative product of decision weights in a machine learning system Gradually reduce the decision weight of the baseline security system Gradually increasing, the drone's task execution is still controlled by the machine learning system; when If the machine learning system's decisions cause the drone to operate in a dangerous state for an extended period of time, the baseline safety system will take over control of the drone instead of the machine learning system, and the drone will remain stationary and wait for the next mission instructions from the ground personnel. When the risk assessment level is C, the baseline safety system replaces the machine learning system to control the drone, controlling the drone to perform a return-to-home operation or fly to a designated location. Decision weights of machine learning systems The calculation formula is: (6); Decision weights of baseline security systems The calculation formula is: (7).

[0013] The beneficial effects of this invention are: 1. This invention addresses the problems of low decision-making reliability and difficulty in risk assessment in machine learning systems by constructing a quantitative assessment index system for decision risk. It comprehensively considers multiple dimensions, including the accuracy of decision results, the current state of the unmanned system, environmental adaptability, and task completion rate, determining the weight and calculation method for each index. Using historical and real-time data, the decision risk of the machine learning system is quantitatively assessed, calculating the risk level of the current decision. A weighted summation method is used to weight each risk index, obtaining a comprehensive risk value. Based on the risk level, corresponding risk warning and control mechanisms are triggered.

[0014] 2. To address the instability of risk control models in machine learning systems, this invention dynamically adjusts the decision weights of the machine learning system and the safety baseline system based on risk assessment results, system operating status, and changes in the environment and tasks. When the system is operating stably and the risk is low, the decision weight of the machine learning system is increased to fully leverage its intelligent learning and adaptive capabilities. When the system faces high risk or instability, the decision weight of the safety baseline system is increased to ensure system security. This patent designs a smooth transition strategy based on time-progressive weight adjustment to achieve a smooth switch between the machine learning system and the safety baseline system, avoiding system instability caused by sudden decision changes.

[0015] 3. This invention employs ultra-real-time parallel simulation software for large-scale comparative testing, undergoing trials of no less than [number missing] hours. In the test, all UAVs successfully completed their missions or returned safely to base, demonstrating that the machine learning system cognitive decision-making safety conversion and risk control method proposed in this invention has excellent performance and significantly improves the safety of UAVs performing missions in simulated battlefields. Attached Figure Description

[0016] Figure 1 A flowchart illustrating a method for safe conversion and risk control of cognitive decision-making in an intelligent autonomous unmanned system provided by the present invention; Figure 2 This is a schematic diagram of the ultra-real-time parallel simulation platform established by the present invention; Figure 3 A schematic diagram of the human-computer interaction interface of the ultra-real-time parallel simulation platform provided by the present invention; Figure 4 A flowchart illustrating the process of collecting historical data provided by this invention; Figure 5 A schematic diagram of the decision risk quantification assessment model and risk precision control mechanism provided by this invention; Figure 6 This is a schematic diagram illustrating the simulated drone mission failure when only a machine learning system is used to control the operation of drones in a dangerous scenario with a large number of targets, as provided by the present invention. Figure 7 A schematic diagram illustrating the safe return of a drone to its destination using the cognitive decision-making safety conversion and risk control algorithm provided by this invention. Detailed Implementation Specific implementation method one: For drones in battlefield environments, the decision-making results of their internal machine learning systems are directly related to their survival. The safe conversion and precise risk control of the cognitive decision-making of machine learning systems can maximize the safety of machine learning system decisions, thereby ensuring the safe flight of drones.

[0018] Combination Figure 1-5 This embodiment describes a method for safe transition and risk control in cognitive decision-making of an intelligent autonomous unmanned system, comprising the following steps: S1: Build an ultra-real-time parallel simulation platform; This embodiment is as follows: Figure 2 The ultra-real-time parallel simulation platform shown depicts a battlefield UAV performing a mission and its actual flight environment. Some of the constructed parameters are listed in Table 1 below. These parameters include initial settings and parameters that change during operation, comprehensively considering natural environmental factors such as weather and terrain, including day / night cycles, wind and rain, and mountainous / plain terrain; considering the characteristics of the UAV itself, including key initial parameters such as the number of UAVs; and setting the mission requirements. Various real-time changing parameters were also designed for when the simulation platform starts running. The comprehensive simulation of these parameters accurately reflects a realistic high-intensity adversarial environment and provides data support for the subsequent construction of a machine learning-based risk quantification assessment model.

[0019] Meanwhile, in order to more intuitively quantify the changes in various parameters and decision-making of the simulation platform, this implementation method builds a human-computer interaction interface for visualizing the corresponding results.

[0020] Table 1

[0021] S2: Construction of a quantitative assessment model for decision-making risks; Quantifying decision-making risks requires comprehensive consideration of the simulated drone's own state, environmental information, and other factors. By combining historical data and real-time parameters, a risk assessment of the current state and environment of the drone can be provided for subsequent precise risk decision-making and control.

[0022] S201: The decision risk quantitative assessment model needs to be constructed by combining historical data and real-time collected data.

[0023] To ensure that historical data can cover a wide range of normal and extreme mission scenarios, test scenarios were first constructed by orthogonally coupling multiple types of parameters, and then large-scale automated parallel testing was conducted. Relevant information about the UAV's mission execution was recorded, as shown in Table 2 below.

[0024] Table 2

[0025] S202: Risk Weight Calculation To output the subsequent risk assessment level, it is necessary to conduct a comprehensive analysis of historical data, identify the high-threat influencing factors in the simulated UAV's own state parameters, environmental parameters, and mission parameters, and give the corresponding parameter risk weights.

[0026] S20201: To ensure comparability of different dimensions of multiple parameters, it is necessary to normalize the parameters. The normalization formula is shown in equation (1) below: (1); In formula (1), Indicates the first j The first historical test task in the group i The parameter of the first k The normalized value at each moment. and They represent the first i The maximum and minimum values ​​of each parameter in historical data; S20202: Taking the success or failure of the final UAV mission as the main guide, the correlation between each parameter and the mission result is calculated. This implementation method adopts the Pearson correlation method, as shown in the following formula (2): (2); In formula (2), For the first i The correlation coefficients between each parameter and the task results For the first j The parameter in the first... k The normalized value at each moment. For the first j The result of each task execution, if the task is successfully executed... If the task fails, then , and The first i The average of each parameter and the task result. For the first j The number of time slots for each task This is the time decay factor, used to measure the time difference between the current parameter in historical data and the end of the task; By introducing a time decay factor It can accurately control the relationship between parameters of various drone statuses, environmental conditions, and mission modes and different time stages of mission execution.

[0027] S20203: Calculate the risk weights for each parameter. The calculation of risk weights takes into account the time factor of the UAV's mission execution, scene information parameters, and mission success or failure. The resulting risk weights are used to subsequently evaluate the risk level of the test scenario.

[0028] The formula for calculating risk weights is shown in equation (3) below: (3); In formula (3), For the firsti The parameters at time... t The weight, It is a moment t The time decay factor.

[0029] Examples of the key impact parameters in the above history are given. The key impact parameters at each moment are calculated and sorted according to formulas (1), (2) and (3). The key impact parameters are affected by the success or failure of the task and the execution stage (time). They are divided into early, middle and late stages according to time, and are summarized in Table 3 below: Table 3

[0030] S203: Scenario matching and risk level assessment; The above content is all about the construction of a risk quantification assessment model for historical test data. In order to use the above model to conduct a decision risk quantification assessment of the real-time running simulated drone, it is necessary to match the real-time scenario of the simulated drone with a large number of historical scenarios. The key parameters of the real-time test scenario of the drone can be obtained directly from ultra-real-time parallel simulation software or human-computer interaction interface. The key parameters are sorted according to their weights, sorted according to 1...n, and matched with the historical dataset in turn. The matching method uses Euclidean distance to find approximate test samples. The calculation of Euclidean distance is shown in the following formula (4): (4); In formula (4), For historical data, These are the current key parameters. n This represents the number of current key parameters; The set of samples with the smallest Euclidean distance is selected as the matched historical samples. The risk assessment result of a certain task at a certain moment based on the matched historical data is shown in the following formula (5): (5); In formula (5), Let t be the risk outcome value of the current machine learning decision at time t.

[0031] When Risk is no more than 0.3, the risk assessment level is A, indicating that the current machine learning system's decision-making safety is relatively high; when Risk is greater than 0.3 but not greater than 0.7, the risk assessment level is B, indicating that the current machine learning system's decision-making results have a certain probability of danger, and this probability has reached a level that cannot be ignored; when Risk is greater than 0.7, the risk assessment level is C, indicating that the simulated drone is facing serious operational safety issues, and based on historical experience, the machine learning system's decisions are very likely to lead the drone into an even more dangerous situation.

[0032] S3: Design of a real-time risk precision control mechanism; The risk level output by the pre-decision risk quantification assessment model determines the operation of the real-time risk precision control mechanism, and the function of real-time risk precision control is controlled and implemented by the baseline security system.

[0033] When the risk rating is A, the decision output of the machine learning system is considered to have high security. In this case, it is not necessary to trigger the real-time risk precision control mechanism of the baseline safety system, and the machine learning system is still used to control the mission execution of the simulated drone.

[0034] When the risk rating is B, it is considered that the current machine learning system's decision-making results have a certain probability of being dangerous, and this probability has reached a level that cannot be ignored. Over time, the machine learning system is highly likely to lead the drone into a dangerous state. This implementation method sets an attenuation coefficient. As time goes on, the decision weights of the machine learning system are gradually reduced and the decision weights of the baseline security system are increased. The weight calculation formulas are shown in (6) and (7) below.

[0035] Decision weights of machine learning systems The calculation formula is: (6); Decision weights of baseline security systems The calculation formula is: (7).

[0036] when When the value is 1, the drone's task execution is entirely controlled by the machine learning system. As the decay factor accumulates, the weight of the machine learning system gradually decreases, while the weight of the baseline safety system gradually increases. At this point, the machine learning system still controls the unmanned task execution. If the machine learning system's decisions cause the drone to operate in a dangerous state for an extended period, the baseline safety system will take over control of the drone, which will then hover in place awaiting further instructions from ground personnel.

[0037] When the risk assessment is Level C, it indicates that the drone is in an extremely dangerous operating state, and the machine learning system is unable to guide it to a safe operating state. In this case, the baseline safety system forcibly takes over control of the drone. Under the control of the baseline safety system, the drone executes either the command to immediately return to base or to fly to a safe point.

[0038] In summary, this invention addresses the issues of low decision-making reliability and difficulty in risk assessment in machine learning systems by constructing a quantitative assessment index system for decision risk. This system comprehensively considers multiple dimensions, including the accuracy of decision results, the current state of the unmanned system, environmental adaptability, and task completion rate, determining the weights and calculation methods for each index. Using historical and real-time data, the decision-making risk of the machine learning system is quantitatively assessed, calculating the risk level of the current decision. A weighted summation method is used to weight each risk index, yielding a comprehensive risk value. Based on the risk level, corresponding risk warning and control mechanisms are triggered.

[0039] The process of the decision-making risk quantitative assessment model and the real-time risk precision control mechanism is attached. Figure 5 As shown. In summary, to address the instability of risk control models in machine learning systems, this invention dynamically adjusts the decision weights of the machine learning system and the safety baseline system based on risk assessment results, system operating status, and changes in the environment and tasks. When the system is operating stably and the risk is low, the decision weight of the machine learning system is increased to fully leverage its intelligent learning and adaptive capabilities; when the system faces high risk or instability, the decision weight of the safety baseline system is increased to ensure system security. This implementation design employs a time-based, gradual weight adjustment smooth transition strategy to achieve a smooth switch between the machine learning system and the safety baseline system, avoiding system instability caused by sudden decision changes.

[0040] Specific Implementation Method Two: Combining Figure 6 and Figure 7 This embodiment is described in detail. In order to verify the effectiveness of the intelligent autonomous unmanned system's cognitive decision-making safety conversion and risk control proposed in this invention, this embodiment uses ultra-real-time parallel simulation software to conduct large-scale comparative tests.

[0041] Assuming consistent initial parameter configurations, a machine learning system will be used to control the drone's mission execution. A comparative test will be conducted using cognitive decision-making safety transition and risk control methods with and without the machine learning system.

[0042] In hazardous scenarios involving a large number of targets, when only a machine learning system is used to control the drones, the simulated drones cannot recognize the current dangerous state, ultimately leading to mission failure. Figure 6 As shown.

[0043] In hazardous scenarios involving a large number of deployed targets, when using a machine learning system to control the drone's operation, and simultaneously employing the machine learning system's cognitive decision-making safety conversion and risk control algorithms, the simulated drone, upon detecting a large number and dense distribution of targets, assesses the risk as Level C. The baseline safety model quickly takes over control of the drone, executing a return-to-home command, and ultimately the drone safely withdraws. Figure 7As shown.

[0044] After no less than The experimental results show that the machine learning system cognitive decision safety conversion and risk control method proposed in this invention has excellent performance and significantly improves the safety of simulated drones performing tasks.

[0045] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent substitutions, and improvements made to the above embodiments without departing from the scope of the present invention, based on the technical essence of the present invention and within the spirit and principles of the present invention, shall still fall within the protection scope of the present invention.

Claims

1. A method for safe transition and risk control in cognitive decision-making of an intelligent autonomous unmanned system, characterized in that, include: Step 1: Build an ultra-real-time parallel simulation platform and obtain key parameters of the UAV real-time scene based on the ultra-real-time parallel simulation platform; Step 2: Collect relevant data during the drone's mission execution process and historical data of the drone's mission execution in real time; Step 3: Construct a decision risk quantification assessment model based on relevant and historical data collected during the drone's mission execution process, and input the historical data of the drone's mission execution to calculate risk weights; Step 4: Obtain key parameters of relevant data from the real-time acquisition of UAV mission execution process using the ultra-real-time parallel simulation platform, and calculate the risk outcome of the current machine learning decision by matching Euclidean distance with historical data of UAV mission execution. Risk Based on risk outcome Risk Obtain the risk assessment level of the current machine learning system's decision-making process; Step 5: Based on the baseline security system, develop a corresponding real-time risk control mechanism for each risk assessment level to achieve precise real-time risk control.

2. The method for safe transition and risk control of cognitive decision-making in an intelligent autonomous unmanned system according to claim 1, characterized in that, Step 1, building the ultra-real-time parallel simulation platform, specifically includes: Step 1.1: Set initial parameters, which include at least the initial number of drones, drone detection range, drone maximum speed, drone mission boundary, drone minimum flight altitude, drone maximum / initial fuel level, weather conditions, terrain factors, drone maximum acceleration, drone position, and drone speed; Step 1.2: Set the parameters for changes during UAV operation. These parameters should include at least the target position, the target range and the location of the task to be performed, the target radar position, the target moving speed, the UAV moving speed, weather changes, the current detection radius of the UAV, the current fuel level of the UAV, and the size / location of the navigation blind spot. Step 1.3: Based on the set initial parameters and the parameters that change during the operation of the UAV, construct the key parameters of the real-time UAV scene; Step 1.4: Set up the human-computer interaction page to visualize the key parameters of the drone's real-time scene.

3. The method for safe transition and risk control of cognitive decision-making in an intelligent autonomous unmanned system according to claim 1, characterized in that, Step 2 specifically includes: Historical data on UAV missions includes mission execution results, environmental situation parameters at each moment, UAV's own state parameters at each moment, UAV mission execution parameters at each moment, and machine learning decision results at the corresponding moment.

4. The method for safe transition and risk control of cognitive decision-making in an intelligent autonomous unmanned system according to claim 1, characterized in that, Step 3 specifically includes: Step 3.1: Normalize the environmental situation parameters, UAV state parameters, and UAV mission execution parameters in the historical data of UAV mission execution; Step 3.2: Introduce a time decay factor The Pearson coefficients between the normalized environmental situation parameters, the UAV's own state parameters, the UAV's mission execution parameters, and the mission execution results were calculated respectively. Step 3.3: Based on the Pearson coefficient and time decay factor between the corresponding parameters and the task execution results. Calculate the risk weights of the corresponding parameters, and sort the corresponding historical environmental situation parameters, UAV self-state parameters, and UAV mission execution parameters according to the magnitude of the risk weights; The expression for normalization is: (1)； In formula (1), Indicates the first j The first historical test task in the group i The parameter of the first k The normalized value at each moment. and They represent the first i The maximum and minimum values ​​of each parameter in historical data; The formula for calculating the Pearson coefficient is: (2)； In formula (2), For the first i The correlation coefficients between each parameter and the task results For the first j The parameter in the first... k The normalized value at each moment. For the first j The result of each task execution, if the task is successfully executed... If the task fails, then , and The first i The average of each parameter and the task result. For the first j The number of time slots for each task This is the time decay factor, used to measure the time difference between the current parameter in historical data and the end of the task; The formula for calculating risk weights is: (3)； In formula (3), For the first i The parameters at time... t The weight, It is a moment t The time decay factor.

5. The method for safe transition and risk control of cognitive decision-making in an intelligent autonomous unmanned system according to claim 4, characterized in that, Step 4 specifically includes: Step 4.1: Obtain the key parameters of the relevant data of the UAV in real time during the mission execution process from the ultra-real-time parallel simulation platform, repeat step 3 to obtain the risk weights of the current key parameters and sort them; Step 4.2: Use Euclidean distance to match the sorted current key parameters with historical data, select the historical data with the smallest Euclidean distance to the current key parameter, and calculate the risk result of the current machine learning system's decision based on the risk weights of the corresponding historical data parameters. Risk ,in, The value range is 0-1; Step 4.3: When At this point, the risk assessment level is A, and the decision made by the current machine learning system is considered safe. At that time, the risk assessment level is B, indicating that the current machine learning system's decision-making carries a certain probability of danger. At that time, the risk assessment level was C, indicating that the decision-making of the current machine learning system had a serious probability of danger; The formula for calculating Euclidean distance is: (4)； In formula (4), For historical data, These are the current key parameters. n This represents the number of current key parameters; Risk outcomes of current machine learning decision-making Risk The calculation formula is: (5)； In formula (5), Let t be the risk outcome value of the current machine learning decision at time t.

6. The method for safe transition and risk control of cognitive decision-making in an intelligent autonomous unmanned system according to claim 5, characterized in that, Step 5 specifically includes: When the risk assessment level is A, there is no need to trigger the real-time risk precision control mechanism; When the risk assessment level is B, a decay coefficient is set. And calculate the decision weights of the machine learning system. According to the decision weights of the machine learning system Calculate the decision weights of the baseline security system Over time, gradually reduce the decision weights of the machine learning system. Increase the decision weight of the baseline security system ,when When the value is 1, the drone's task execution is entirely controlled by the machine learning system. At that time, with the attenuation coefficient The cumulative product of decision weights in a machine learning system Gradually reduce the decision weight of the baseline security system Gradually increasing, the drone's task execution is still controlled by the machine learning system; when If the machine learning system's decisions cause the drone to operate in a dangerous state for an extended period of time, the baseline safety system will take over control of the drone instead of the machine learning system, and the drone will remain stationary and wait for the next mission instructions from the ground personnel. When the risk assessment level is C, the baseline safety system replaces the machine learning system to control the drone, controlling the drone to perform a return-to-home operation or fly to a designated location. Decision weights of machine learning systems The calculation formula is: (6)； Decision weights of baseline security systems The calculation formula is: (7)。

Images