An adaptive intelligent decision system and method for complex scenarios

Abstract

The application relates to a complex scene-oriented adaptive intelligent decision system and method, belonging to the technical field of intelligent decision, which comprises an environment perception module, a task evaluation module, a decision planning module, an algorithm adaptation module and an execution control module; multi-source perception data of a target complex scene are acquired through the above modules, environment features are extracted, an environment model is constructed, the priority of a target task is calculated to obtain a task priority sequence, candidate decision strategies are retrieved from a decision strategy library to obtain an optimal decision strategy, a decision planning result is obtained, the comprehensive performance scores of all algorithms are calculated based on an established algorithm evaluation index system, an optimal algorithm is selected, control instructions are generated based on the decision planning result and the optimal algorithm, and a corresponding target task is executed by driving an execution mechanism to generate an execution result, multi-source information fusion, task conflict arbitration and resource risk adaptive balance in a complex scene are realized, and the decision robustness, execution safety and evolution capability of the system are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent decision-making technology, and specifically to an adaptive intelligent decision-making system and method for complex scenarios. Background Technology

[0002] Currently, in complex scenarios characterized by dynamic and ever-changing environments, intertwined tasks, heterogeneous information, and stringent resource constraints, achieving reliable, efficient, and adaptive intelligent decision-making has become crucial for enhancing the core capabilities of various autonomous systems. Systems must be able to understand the environment in real time amidst uncertainty, assess their own state, coordinate conflicting objectives, and plan safe actions. This places extremely high demands on the adaptability, robustness, and real-time performance of decision-making technologies.

[0003] Existing intelligent decision-making systems typically follow a linear process of perception, decision-making, and execution. They process data through fixed rules or limited algorithms, rely on preset models to understand the environment, and control execution through basic feedback mechanisms. When faced with information from multiple heterogeneous sensors, they usually adopt simple data-level fusion or one-to-one trust methods. When handling concurrent multi-tasks, they schedule according to static priorities or empirical rules. Their core algorithms and parameters are mostly set offline before system deployment.

[0004] However, in existing technologies, static environment models and rules are difficult to match the rapid changes in dynamic scenarios, leading to decision lag or failure; the lack of a quantitative coordination mechanism for multi-task and system constraints easily causes resource conflicts and execution bottlenecks; open-loop architectures and fixed algorithms cannot learn to optimize from execution results, resulting in long-term performance stagnation of the system, and their robustness and security are often insufficient when facing sudden interference or extreme working conditions. Summary of the Invention

[0005] In view of the shortcomings of related technologies, the purpose of this invention is to provide an adaptive intelligent decision-making system and method for complex scenarios, so as to solve the technical problems of existing technologies, which adopt static models, fixed rules and open-loop architecture, making it difficult to adapt to dynamic environmental changes, unable to coordinate multi-task conflicts, and lacking online learning capabilities, resulting in lag in decision-making, insufficient robustness and inability to continuously optimize performance in complex scenarios.

[0006] This invention provides an adaptive intelligent decision-making system for complex scenarios, comprising: The environmental perception module is used to acquire multi-source perception data of the target complex scene, extract environmental features from the multi-source perception data, identify scene types based on the environmental features, and detect obstacle motion states and sudden events to construct an environmental model. The task evaluation module is used to receive and parse the target task instructions, obtain the task parsing results, calculate the priority of each target task based on the task parsing results, obtain the task priority sequence, and generate the task evaluation results. The decision planning module is used to retrieve candidate decision strategies from a preset decision strategy library based on the scenario type and the task priority sequence, optimize the candidate decision strategies in combination with the environment model and the task evaluation results to obtain the optimal decision strategy, and perform path planning and multi-task scheduling based on the optimal decision strategy to obtain the decision planning result. The algorithm adaptation module is used to establish an algorithm evaluation index system based on the scenario type and the task parsing result, calculate the comprehensive performance score of each algorithm in the preset algorithm library based on the algorithm evaluation index system, and select the optimal algorithm based on the comprehensive performance score. The execution control module is used to generate control instructions based on the decision planning results and the optimal algorithm, drive the execution mechanism to perform the corresponding target task based on the control instructions, and generate execution results.

[0007] By setting up environmental perception, task evaluation, decision planning, algorithm adaptation, and execution control modules, a complete technical closed loop covering perception, cognition, decision-making, and execution is constructed. The responsibilities and data flow of each link are clearly defined, ensuring that the system can perform coordinated and consistent adaptive processing in response to changes in complex external scenarios and the needs of multiple internal tasks. This solves the problems of fragmented modules and slow response in traditional systems, and provides a system-level solution for intelligent agents to achieve stable and reliable autonomous operation in dynamic and unknown environments.

[0008] In some embodiments of the present invention, the environment perception module is specifically configured as follows: The original heterogeneous data of the target complex scene is acquired by multi-source heterogeneous sensors, and the original heterogeneous data is preprocessed to obtain the multi-source sensing data. Environmental features are extracted from the multi-source sensing data, and the local decision-making results built into the multi-source heterogeneous sensors are fused with the environmental features to obtain environmental sensing results, thereby constructing the environmental model. By clarifying the specific process by which the environmental perception module must first preprocess the raw data from multi-source heterogeneous sensors before performing feature extraction and decision-level fusion, the high quality and consistency of the raw perception data are ensured. This effectively eliminates the negative impacts of different sensors on data format, noise characteristics, and local decision conflicts, providing a clean and unified data foundation for the subsequent construction of accurate and reliable environmental models, thereby improving the system's robustness and accuracy in understanding complex scenarios.

[0009] In some embodiments of the present invention, the environment sensing module is further configured to: Extract static and dynamic environmental feature parameters from the environmental perception results to construct an environmental feature database; Based on the environmental feature database, scene types are identified using a deep learning model; Based on the dynamic environmental characteristic parameters, the extended Kalman filter algorithm is used to establish a kinematic model of the obstacle and predict the motion state of the obstacle. Calculate the rate of change of the dynamic environmental feature parameters within a preset sliding window, compare the rate of change with a mutation threshold, and identify mutation events; The environment model is constructed by combining the static environment feature parameters, the scene type, the obstacle motion state, and the mutation event.

[0010] By specifying that the environmental perception module must perform feature parameter extraction, scene type recognition, obstacle motion prediction, and sudden event detection, and finally comprehensively construct an environmental model, a progressive processing chain from raw perception to structured environmental cognition is established. This enables the environmental model to not only include static layout information, but also integrate the predicted trajectory of dynamic obstacles and the real-time judgment of sudden events, thereby providing the system with a refined environmental representation that supports forward-looking decision-making with both spatial and temporal dimensions.

[0011] In some embodiments of the present invention, the environment sensing module is further configured to: Based on the obstacle's motion state and the static environmental characteristic parameters, the dynamic obstacle density index and the rate of change of environmental characteristics are calculated respectively. The dynamic obstacle density index and the rate of change of environmental characteristics are weighted and calculated to obtain the environmental dynamic assessment index; The environmental dynamics assessment index is compared with the index threshold to obtain the comparison result. Based on the comparison result, the time step and covariance matrix value in the extended Kalman filter algorithm are adjusted to adjust the update frequency and modeling accuracy of the environmental model.

[0012] By introducing an environmental dynamics assessment index based on dynamic obstacle density and environmental feature change rate, and adjusting the model update frequency and accuracy according to the index, the environmental model is endowed with self-adjustment capabilities. This enables the system to quantify the severity of environmental changes and adaptively allocate computing resources: saving computing power in stable scenarios and improving the agility of perception and modeling in drastic scenarios, thereby optimizing the overall system's energy efficiency while ensuring the timeliness of the environmental model.

[0013] In some embodiments of the present invention, the task evaluation module is specifically configured as follows: Real-time monitoring of the running status of hardware components and software processes, and acquisition of multiple running parameters; Based on the parameter characteristics of each of the aforementioned operating parameters, the corresponding normalized value is calculated. The normalized values ​​corresponding to each of the aforementioned operating parameters are then weighted to obtain the health index. By requiring the task evaluation module to monitor the operating status of the underlying hardware and software of the system in real time and calculate the health index, the system can achieve continuous quantitative self-diagnosis of its own status. Integrating operational parameters into a unified health metric provides the system with an objective basis for assessing its current execution capabilities and resource adequacy. This is a prerequisite for subsequent task scheduling and resource allocation based on available resources, enhancing the pragmatism and reliability of system decision-making.

[0014] In some embodiments of the present invention, the task evaluation module is further configured as follows: The target parameters, execution requirements, time windows, and priority information of each target task in the target task instruction are parsed to obtain the task parsing result; Based on the health index and the preset performance requirements of each target task, calculate the performance satisfaction rate of each target task; Based on the operating parameters and the preset energy consumption requirements of each target task, calculate the energy sufficiency rate of each target task; Based on the time window, the time sufficiency of each target task is calculated; By combining the performance satisfaction rate, the energy sufficiency rate, and the time sufficiency, the constraint satisfaction rate of each target task is calculated, and a constraint satisfaction report for each target task is obtained. By stipulating that the task evaluation module should calculate the constraint satisfaction rate of the task in the three dimensions of performance, energy and time based on the health index, operating parameters and task time window, a multi-dimensional and refined evaluation of task feasibility is achieved. Matching abstract task requirements with concrete real-time system status and resource levels can identify tasks that may fail due to insufficient system capacity or limited resources in advance. This provides crucial risk warnings and constraint boundary information for decision-making and planning, avoiding unrealistic decisions.

[0015] In some embodiments of the present invention, the task evaluation module is further configured as follows: The urgency and importance of each target task are calculated based on the priority information. The priority of each target task is obtained by weighting the urgency and importance of each target task and the constraint satisfaction of each target task. The priorities of each target task are sorted to obtain the task priority sequence; The task evaluation result is obtained by combining the task priority sequence and the constraint satisfaction report of each target task. By combining task urgency, importance, and constraint satisfaction to quantify priority calculation, and sorting and conflict marking of the results, the basis for multi-task scheduling has been transformed from empirical rules to scientific calculation. By comprehensively considering the inherent attributes of the task, external time constraints, and the current execution capacity of the system, a priority sequence that reasonably reflects the value and urgency of the task can be generated, providing a clear, fair, and efficient decision-making basis for task arbitration and execution order arrangement under subsequent resource competition.

[0016] In some embodiments of the present invention, the decision planning module is specifically configured as follows: Based on the task evaluation results and the environment model, optimization objectives are set for the candidate decision strategies. Based on the optimization objectives, a multi-objective optimization model is established. The multi-objective optimization model is solved by a non-dominated sorting genetic algorithm, and the optimal decision strategy is output. Based on the optimal decision-making strategy and the environment model, a path search is performed using an improved A* algorithm to generate a path planning scheme. Based on the task priority sequence and the health index, the execution order and resource allocation strategy of each target task are formulated to obtain a task scheduling scheme. The decision planning result is output by combining the optimal decision-making strategy, the path planning scheme, and the task scheduling scheme. By stipulating that the decision planning module must use a multi-objective optimization model to solve the optimization problem that integrates the environmental model and task evaluation results, and collaboratively generate strategies, paths and scheduling schemes, the decision-making process has been elevated from single-point optimization to global collaboration. This ensures that the final planning results can comprehensively weigh the benefits of task completion and the execution costs while meeting environmental safety constraints and system resource limitations, and output decision planning results that are consistent in spatial paths, action strategies and time series.

[0017] In some embodiments of the present invention, the adaptive intelligent decision-making system further includes an iterative optimization module, which is configured to: The execution results are compared with the preset decision expectation targets, decision evaluation indicators are calculated, and decision weights are set for the decision evaluation indicators; The decision evaluation indicators are weighted based on the decision weights to obtain the decision evaluation results. Based on the decision evaluation results, the model parameters in the environment model, the strategy rules in the preset decision strategy library, the algorithm parameters in the preset algorithm library, and the decision weights are iteratively updated. Through the iterative optimization module, an evaluation and feedback mechanism based on the comparison between the execution results and the expected goals can be established, and the model parameters, policy rules and algorithm parameters can be updated accordingly. This gives the system the ability to continuously learn and optimize from historical experience, breaks the limitations of traditional open-loop decision-making systems, and forms a complete learning closed loop of decision-making, execution, evaluation and optimization. This enables the system to automatically correct model biases, refine effective strategies and optimize algorithm performance, thereby achieving autonomous evolution and long-term improvement of the overall intelligence level.

[0018] Some embodiments of the present invention further provide an adaptive intelligent decision-making method for complex scenarios, comprising the following steps: The environmental perception step involves acquiring multi-source perception data of the target complex scene, extracting environmental features from the multi-source perception data, identifying the scene type based on the environmental features, and detecting obstacle motion states and abrupt events to construct an environmental model. The task evaluation step involves receiving and parsing the target task instructions to obtain the task parsing results, calculating the priority of each target task based on the task parsing results, obtaining the task priority sequence, and generating the task evaluation results. The decision planning step involves retrieving candidate decision strategies from a preset decision strategy library based on the scenario type and the task priority sequence, optimizing the candidate decision strategies in conjunction with the environment model and the task evaluation results to obtain the optimal decision strategy, and performing path planning and multi-task scheduling based on the optimal decision strategy to obtain the decision planning result. The algorithm adaptation steps include: establishing an algorithm evaluation index system based on the scenario type and the task parsing results; calculating the comprehensive performance score of each algorithm in the preset algorithm library based on the algorithm evaluation index system; and selecting the optimal algorithm based on the comprehensive performance score. The execution control steps involve generating control commands based on the decision planning results and the optimal algorithm, driving the actuator to perform the corresponding target task based on the control commands, and generating execution results.

[0019] By mapping the module functions in the system to sequentially executed steps of environmental perception, task evaluation, decision planning, algorithm adaptation, and execution control, a complete operational process for implementing this adaptive intelligent decision-making system is provided. This process clarifies the orderly connection and data transmission relationship between each technical stage, ensuring that the system can be fully implemented through specific methods and procedures. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Figure 1 This is a schematic diagram of the structure of an adaptive intelligent decision-making system for complex scenarios provided in an embodiment of the present invention; Figure 2 A schematic diagram of another adaptive intelligent decision-making system for complex scenarios provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a multi-source sensing fusion module provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of a dynamic environment modeling module provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of a task evaluation module provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of the structure of a decision planning module provided in an embodiment of the present invention; Figure 7 This is a flowchart of an adaptive intelligent decision-making method for complex scenarios provided in an embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this application clearer, the application is described and illustrated below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application. All other embodiments obtained by those skilled in the art based on the embodiments provided in this application without inventive effort are within the scope of protection of this application. It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations according to this application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. Furthermore, it should be understood that the terms “comprising” and “having”, and any variations thereof, are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such process, method, product, or apparatus. Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other. The technical solution of the present invention will be described in detail below with reference to specific embodiments and accompanying drawings.

[0022] like Figure 1 and Figure 2As shown, this invention provides an adaptive intelligent decision-making system for complex scenarios, which includes an environmental perception module. The environmental perception module acquires multi-source perception data of the target complex scenario, extracts environmental features from the multi-source perception data, identifies the scenario type based on the environmental features, and detects obstacle movement states and abrupt events to construct an environmental model.

[0023] In some embodiments, the environment sensing module is specifically configured as follows: The raw heterogeneous data of the target complex scene is obtained by acquiring raw heterogeneous data through multi-source heterogeneous sensors, and the raw heterogeneous data is preprocessed to obtain multi-source sensing data. Environmental features are extracted from multi-source sensing data, and the local decision-making results built into the multi-source heterogeneous sensors are fused with the environmental features to obtain environmental sensing results, so as to construct an environmental model. By clarifying the specific process by which the environmental perception module must first preprocess the raw data from multi-source heterogeneous sensors before performing feature extraction and decision-level fusion, the high quality and consistency of the raw perception data are ensured. This effectively eliminates the negative impacts of different sensors on data format, noise characteristics, and local decision conflicts, providing a clean and unified data foundation for the subsequent construction of accurate and reliable environmental models, thereby improving the system's robustness and accuracy in understanding complex scenarios.

[0024] Furthermore, such as Figure 3 As shown, the environmental perception module specifically includes a sensor access adaptation unit, a heterogeneous data preprocessing unit, a feature-level fusion unit, and a decision-level fusion unit; The sensor access adapter unit is used to connect multi-source heterogeneous sensors with different interface types and data types; Multi-source heterogeneous sensors include vision sensors, lidar, millimeter-wave radar, ultrasonic sensors, inertial measurement units (IMUs), positioning modules, and IoT sensors; The positioning module can use GPS positioning or Beidou positioning. Specifically, the sensor access adapter unit has a built-in sensor type identification library and data format parsing rules. The sensor type identification library can automatically identify the type of the accessed sensor, and the data format parsing rules can convert the raw heterogeneous data output by the multi-source heterogeneous sensors into a unified data format of the system. It also establishes a sensor data index table to record the number, type, data dimension and sampling period of each multi-source heterogeneous sensor. The heterogeneous data preprocessing unit is used to receive data output from the sensor access adapter unit and sequentially perform noise reduction processing, time synchronization processing, missing data completion processing and data normalization processing. Among them, the noise reduction process adopts an adaptive algorithm for different types of data; For example, median filtering is used to remove salt-and-pepper noise from image data; statistical filtering is used to remove outliers from point cloud data; and Kalman filtering is used to suppress random noise from numerical data. Time synchronization processing is based on a unified system timestamp and uses linear interpolation to align the sampling data of different multi-source heterogeneous sensors, ensuring that the multi-source data at the same time correspond and match. The missing data completion process uses an adaptive interpolation method based on historical data trends. For data missing for a short period of time, it is completed by linear interpolation of data from adjacent time points. For data missing for a long period of time, it is completed by multinomial interpolation combined with sensor type and environmental characteristics. Data normalization is a process that maps data of different dimensions to the [0,1] interval to eliminate the impact of dimensional differences on subsequent fusion processing.

[0025] The feature-level fusion unit is used to extract environmental features from preprocessed multi-source sensing data, construct a multi-dimensional feature set, and perform feature fusion using an improved DS evidence theory algorithm. Specifically, the system extracts target contour features, color features, and texture features from data collected by visual sensors; it extracts obstacle distance features, velocity features, and contour features from data collected by lidar; it extracts target velocity features, distance features, and azimuth features from data collected by millimeter-wave radar; it extracts attitude angle features, acceleration features, and angular velocity features from data collected by IMU inertial measurement unit; and it extracts position features and velocity features from data collected by positioning module. The feature-level fusion unit normalizes and assigns weights to environmental features, handles feature conflicts based on the improved DS evidence theory, and outputs a unified multi-dimensional feature set, which includes static environmental features and dynamic environmental features.

[0026] The decision-level fusion unit is used to receive the local decision results output by each multi-source heterogeneous sensor based on its own data, and to perform decision fusion with the environmental features in the multi-dimensional feature set using a Bayesian inference algorithm. Among them, the local decision results of the visual sensor can be the type of obstacle identified; the local decision results of the lidar can be the level of obstacle threat determined; and the local decision results of the positioning module can be the location reliability. The decision-level fusion unit establishes a decision reliability evaluation model for each multi-source heterogeneous sensor, and statistically analyzes the decision accuracy of each multi-source heterogeneous sensor in different complex scenarios based on historical data, which is used as the weight for decision fusion. Based on the weights of decision fusion, the Bayesian inference algorithm is used to fuse environmental features in the multi-dimensional feature set with local decision results, eliminate decision conflicts, and output globally consistent environmental perception results. The environmental perception results include the type, location, status, environmental scene characteristics, and potential risks of environmental targets.

[0027] In some embodiments, the environment sensing module is further configured to: Extract static and dynamic environmental feature parameters from the environmental perception results and construct an environmental feature database; Based on an environmental feature database, scene types are identified using a deep learning model. Based on dynamic environmental characteristic parameters, the extended Kalman filter algorithm is used to establish a kinematic model of the obstacle and predict the obstacle's motion state. Calculate the rate of change of dynamic environmental feature parameters within a preset sliding window, compare the rate of change with a mutation threshold, and identify mutation events; An environment model is constructed by combining static environment characteristic parameters, scene type, obstacle motion state, and sudden events.

[0028] By specifying that the environmental perception module must perform feature parameter extraction, scene type recognition, obstacle motion prediction, and sudden event detection, and finally comprehensively construct an environmental model, a progressive processing chain from raw perception to structured environmental cognition is established. This enables the environmental model to not only include static layout information, but also integrate the predicted trajectory of dynamic obstacles and the real-time judgment of sudden events, thereby providing the system with a refined environmental representation that supports forward-looking decision-making with both spatial and temporal dimensions.

[0029] Furthermore, such as Figure 4 As shown, the environmental perception module also includes an environmental feature extraction unit, a dynamic obstacle modeling unit, an environmental change detection unit, and a scene type recognition unit; The environmental feature extraction unit is used to receive the environmental feature set output by the multi-source sensing fusion module and further extract key environmental feature parameters, including static environmental feature parameters and dynamic environmental feature parameters. Static environmental characteristic parameters include road width, road curvature, number of lanes, location and size of fixed obstacles, terrain slope, and environmental boundary range; Dynamic environmental characteristic parameters include the number, location, speed, acceleration, trajectory, and rate of change of environmental parameters of moving obstacles; The environmental feature extraction unit uses a feature clustering algorithm to classify features of the same type and establish an environmental feature database.

[0030] The dynamic obstacle modeling unit establishes the motion equations and observation equations of obstacles based on the position, velocity, and acceleration of obstacles in the environmental perception results. It updates the motion state of obstacles in real time through the extended Kalman filter algorithm and predicts the motion trajectory of obstacles in the future time T. Among them, T can be dynamically adjusted according to the scene, and the default T is 1s. Obstacles may include moving objects or people; Simultaneously, an obstacle behavior decision-making model is established to determine the obstacle's movement intention based on its historical movement state and environmental characteristics, thereby improving the accuracy of trajectory prediction.

[0031] The environmental mutation detection unit detects mutation events in complex scenarios based on time-series changes in environmental characteristics, using the sliding window method and mutation threshold method. The environmental change detection unit presets a sliding window length and calculates the rate of change of dynamic environmental characteristic parameters within the sliding window; wherein, the rate of change of dynamic environmental characteristic parameters includes the rate of change of position, the rate of change of velocity, and the rate of change of environmental parameters; Based on scenario requirements and historical data, an offline mutation threshold is calibrated. When the rate of change of dynamic environmental characteristic parameters exceeds the mutation threshold and the duration is greater than or equal to 2 sampling periods, it is determined to be a mutation event, and the mutation type, location, and scope of influence of the mutation event are output. These mutation events include sudden obstacles, road closures, abrupt changes in environmental parameters, and sensor malfunctions leading to sudden changes in perception data.

[0032] The scene type recognition unit uses a deep learning model to identify scene types based on extracted environmental features; the deep learning model can be a CNN-LSTM hybrid model. The scene type recognition unit pre-builds a scene type sample library; the scene type sample library includes urban road scenes, highway scenes, indoor scenes, outdoor complex terrain scenes, emergency rescue scenes, and intelligent manufacturing workshop scenes; The CNN-LSTM hybrid model extracts spatial features of the environment through CNN and extracts temporal features of the environment through LSTM. It combines spatial and temporal features to classify and identify scene types and outputs scene type identification results and confidence scores. A confidence score greater than or equal to 0.8 is considered a valid identification result.

[0033] Finally, the dynamic environment modeling module combines static environment feature parameters, scene type, obstacle motion state, and abrupt events to construct an environment model; The environment model includes a static environment sub-model, a dynamic obstacle sub-model, an environment mutation sub-model, and a scene type sub-model. The static environment sub-model stores static environment feature parameters and represents the spatial distribution of the environment using a rasterization method; the dynamic obstacle sub-model updates the motion state and predicted trajectory of moving obstacles in real time; the environmental mutation sub-model records relevant information about mutation events; and the scene type sub-model stores scene type and confidence level. The environmental model uses an incremental update mechanism, updating once after each sampling period to ensure consistency between the model and the actual environment.

[0034] In some embodiments, the environment sensing module is further configured to: Based on the obstacle motion state and static environmental characteristic parameters, the dynamic obstacle density index and the rate of change of environmental characteristics are calculated respectively. The dynamic environmental assessment index is obtained by weighting the dynamic obstacle density index and the rate of change of environmental characteristics. The environmental dynamic assessment index is compared with the index threshold to obtain the comparison results. Based on the comparison results, the time step and covariance matrix value in the extended Kalman filter algorithm are adjusted to adjust the update frequency and modeling accuracy of the environmental model.

[0035] By introducing an environmental dynamics assessment index based on dynamic obstacle density and environmental feature change rate, and adjusting the model update frequency and accuracy according to the index, the environmental model is endowed with self-adjustment capabilities. This enables the system to quantify the severity of environmental changes and adaptively allocate computing resources: saving computing power in stable scenarios and improving the agility of perception and modeling in drastic scenarios, thereby optimizing the overall system's energy efficiency while ensuring the timeliness of the environmental model.

[0036] Furthermore, to quantify the degree of change in the dynamic environment in complex scenarios and provide a quantitative basis for adjusting decision-making strategies, a calculation formula for the environmental dynamics assessment index is introduced:

[0037] in, It is an environmental dynamic assessment index, with a value range of [0,1]. The larger the value, the stronger the environmental dynamics, and the more flexible the decision-making strategy needs to be. The dynamic obstacle density index is calculated as the ratio of the number of obstacles in a unit area to the total area of ​​the area, with a value range of [0,1]. The calculation process for the dynamic obstacle density index is as follows: Define an obstacle density index centered on the current system position with a radius of... The circular area is the statistical region. Within this statistical region, the movement speed exceeds a preset threshold. The obstacles are classified as dynamic obstacles, and their number is recorded as follows: Total area of ​​the region Furthermore, a maximum reference density is defined. This is used for normalization. The dynamic obstacle density index is then used for normalization. The calculation formula is:

[0038] When the actual density exceeds the maximum reference density, the index takes the upper limit of 1. The rate of change of environmental characteristics is calculated as the ratio of the number of changes in static environmental characteristic parameters per unit time to the total number of static environmental characteristic parameters, with a value range of [0,1]. The calculation process for the rate of change of environmental characteristics is as follows: assuming the system is within a time window... Internal continuous monitoring Key static environmental characteristic parameters, such as lane curvature, fixed sign location, and road boundary coordinates. For the first... Calculate the static environmental characteristic parameters at the current time. value Compared with the previous sampling time value relative change Set a threshold for feature changes. ,like If the value changes within a given time period, then the characteristic is considered to have changed during that period. Statistical time window Total number of characteristic parameters that change internally Then the rate of change of environmental characteristics The calculation formula is:

[0039] , The weighted coefficients are divided into the dynamic obstacle density index and the rate of change of environmental characteristics, satisfying the following conditions: It is used to adjust the degree of influence of each environmental assessment indicator on environmental dynamics, and is determined by offline training combined with scenario requirements; The environmental dynamic assessment index is compared with the index threshold. If the environmental dynamic assessment index exceeds the index threshold, a fast response mechanism is triggered to reduce the time step in the extended Kalman filter algorithm, update the motion state of the obstacle at a higher frequency, and increase the value of the covariance matrix in the extended Kalman filter algorithm by dynamically adjusting it to improve the decision frequency and flexibility. If the environmental dynamics assessment index does not exceed the index threshold, the time step in the extended Kalman filter algorithm is increased to dynamically adjust and reduce the value of the covariance matrix, thereby reducing computational resource consumption and maintaining stable system operation.

[0040] In the dynamic environment modeling module, the environmental dynamics assessment index serves as an important parameter, which can calibrate the update frequency and modeling accuracy of the environmental model, ensuring that the environmental model can adapt to environmental changes in real time.

[0041] In some embodiments, the adaptive intelligent decision-making system further includes a task evaluation module. The task evaluation module receives and parses target task instructions to obtain task parsing results, calculates the priority of each target task based on the task parsing results, obtains a task priority sequence, and generates task evaluation results.

[0042] In some embodiments, the task evaluation module is specifically configured as follows: Real-time monitoring of the running status of hardware components and software processes, and acquisition of multiple running parameters; The normalized value corresponding to each operating parameter is calculated based on its parameter characteristics. The normalized values ​​corresponding to each operating parameter are then weighted and calculated to obtain the health index. By requiring the task evaluation module to monitor the operating status of the underlying hardware and software of the system in real time and calculate the health index, the system can achieve continuous quantitative self-diagnosis of its own status. Integrating operational parameters into a unified health metric provides the system with an objective basis for assessing its current execution capabilities and resource adequacy. This is a prerequisite for subsequent task scheduling and resource allocation based on available resources, enhancing the pragmatism and reliability of system decision-making.

[0043] Furthermore, such as Figure 5 As shown, the task evaluation module specifically includes a system self-status monitoring unit; The system's own status monitoring unit monitors the operating status of the system's hardware components and software processes in real time and obtains multiple operating parameters; The hardware components include sensors, processors, memory, actuators, and power modules. The parameters monitored include processor load rate, remaining memory space, actuator response speed, remaining power supply capacity, and hardware temperature. The running status of software components includes the running status of algorithms; Set corresponding parameter thresholds in the system's own status monitoring unit. When any operating parameter exceeds its corresponding parameter threshold, it is determined that the system status is abnormal, and the abnormality type, location and scope of impact information are output. Simultaneously, a system health assessment model is established, and a health index is calculated based on the operating parameters of each hardware component and software process. Health Index The calculation model is as follows:

[0044] in, Indicates the first One operating parameter, The first The normalized value corresponding to each running parameter is defined differently depending on the characteristics of the parameter, for example: The normalized values ​​for processor load rate, remaining memory space, actuator response speed, and remaining power supply are: ; in, For maximum processor load, Total storage space of the memory. For the fastest response speed of the actuator, Maximum battery capacity; For hardware temperature Adopting a design based on the optimal operating temperature range The piecewise function scores close to 1 when the hardware temperature is within the optimal range, and the score decreases as the temperature deviates from the range.

[0045] in, This is the lowest temperature limit that the system can withstand; below this temperature, the score drops to 0. This is the highest temperature limit that the system can withstand; if the temperature exceeds this limit, the score drops to 0.

[0046] Algorithm running status Based on the algorithm error rate, let the maximum acceptable error rate of the algorithm be... ,when =0, ;when , ;when , 0.

[0047] It is the first The parameter weights of each operating parameter represent their contribution to the overall health index, satisfying the following conditions: Parameter weights Determined based on expert experience and system design requirements.

[0048] The health index ranges from [0,1]. When the health index is greater than or equal to 0.7, the system is considered to be in good condition; when the health index is greater than or equal to 0.3 and less than 0.7, the system is considered to be in fair condition; when the health index is less than 0.3, the system is considered to be in poor condition.

[0049] In some embodiments, the task evaluation module is further configured as follows: The target parameters, execution requirements, time windows, and priority information of each target task in the target task instruction are parsed to obtain the task parsing results; Based on the health index and the preset performance requirements of each target task, the performance satisfaction rate of each target task is calculated. Based on the operating parameters and the preset energy consumption requirements of each target task, calculate the energy sufficiency rate of each target task. Calculate the time sufficiency of each objective task based on the time window; By combining performance satisfaction rate, energy adequacy rate, and time sufficiency, the constraint satisfaction rate of each objective task is calculated, and a report on the constraint satisfaction status of each objective task is obtained. By stipulating that the task evaluation module should calculate the constraint satisfaction rate of the task in the three dimensions of performance, energy and time based on the health index, operating parameters and task time window, a multi-dimensional and refined evaluation of task feasibility is achieved. Matching abstract task requirements with concrete real-time system status and resource levels can identify tasks that may fail due to insufficient system capacity or limited resources in advance. This provides crucial risk warnings and constraint boundary information for decision-making and planning, avoiding unrealistic decisions.

[0050] Furthermore, the task evaluation module also includes a multi-task requirement analysis unit and a constraint evaluation unit; The multi-task requirement parsing unit receives target task instructions from external input, such as "follow the vehicle in front, avoid pedestrians, and keep the lane" for autonomous driving, and "search and rescue targets, transport supplies, and avoid dangerous areas" for emergency rescue robots. The target parameters, execution requirements, time windows, and priority information of each target task instruction in the target task instruction are parsed to obtain the task parsing results; The target parameters include the endpoint location, target status, and output result requirements of the task execution; the execution requirements include execution accuracy, execution speed, and reliability requirements; the time window includes the task start time, the latest completion time, and the execution duration; and the priority information includes the task urgency, importance, and reward value. The multi-task requirement parsing unit stores the task parsing results in the task information database, providing basic data for task priority calculation.

[0051] The constraint evaluation unit is used to evaluate the impact of system constraints in terms of energy, time, and performance on the execution of each objective task; The assessment of energy constraints is based on the remaining power supply in the operating parameters and the estimated energy consumption of each target task in the energy consumption requirements of each target task, and the energy sufficiency rate for executing each target task is calculated. Energy adequacy ratio is used to assess whether the current power supply has sufficient remaining power to support the completion of a specific task. Its calculation process is as follows: The system first retrieves the estimated energy consumption value of the target task from the pre-stored task knowledge base, based on the target task type, the organization to be executed, and historical energy consumption data. Then, the system divides the current remaining power of the power supply by the estimated energy consumption value of the target task, and the resulting ratio is the energy sufficiency rate.

[0052] If the calculation result is greater than 1, it indicates that the current power is very sufficient to complete the task, and the energy sufficiency rate is treated as 1. If the calculation result is less than 1, it is calculated according to the actual ratio. The smaller the value, the tighter the power supply. The time constraint assessment is based on the current time and the time window, calculating the time sufficiency for executing each objective task; Time adequacy is used to assess whether the system has enough time to complete the task before the deadline, starting from the current moment. Its calculation process is as follows: The system first determines the latest completion time of the target task, and then subtracts the current time from the latest completion time to obtain the available time for executing the task.

[0053] At the same time, the system obtains the estimated time to complete the task from historical statistical data or task models; Finally, based on the ratio of available time to estimated time, the time sufficiency is obtained. If the ratio is greater than 1, it means that there is enough time, and the sufficiency is calculated as 1; if it is less than 1, it means that there is tight time, and the actual ratio is used for calculation. The smaller the value, the tighter the time is. The performance constraint assessment is based on the health index and the performance requirements of each target task, and the performance satisfaction rate of each target task is calculated. The performance satisfaction rate is used to quantify whether the current health status of the system meets the specific requirements of the task for various performance indicators. Its calculation process is as follows: The system organizes the various performance requirements of a task into a clear list of performance requirements; Based on the real-time health index, the system calculates the actual capability values ​​of various performance indicators through a preset mapping relationship. For example, when the health index is 0.9, the positioning error capability is 0.05 meters and the latency capability is 15 milliseconds. Next, the system compares each performance requirement with its corresponding actual system capabilities one by one; If the performance indicator is an upper limit requirement, then determine whether the system capability value is less than or equal to the required value. If yes, then this item gets 1 point; if not, then divide the required value by the capability value to get a score less than 1. If the performance indicator is a lower limit requirement, then determine whether the system capability value is greater than or equal to the required value. If yes, then this item gets 1 point; if not, then divide the capability value by the required value to get a score less than 1. The constraint assessment unit calculates the constraint satisfaction rate of each target task by weighting the performance satisfaction rate, energy adequacy rate, and time adequacy rate, and generates a report on the constraint satisfaction status of each target task. It also marks target tasks with insufficient constraints, providing a constraint basis for decision-making and planning. Constraint Satisfaction The calculation formula is:

[0054] in, For performance satisfaction rate; For energy sufficiency rate; For sufficient time; , , These are the weighting coefficients corresponding to performance satisfaction rate, energy adequacy rate, and time adequacy, and The weighting coefficients can be preset according to the task type or system operation strategy. The calculated... The range of values ​​is A higher value indicates that the target task is more feasible to execute under the current system state.

[0055] The constraint evaluation unit is based on the calculated constraint satisfaction of each objective task. Generate a constraint satisfaction report, which lists the constraints for each objective task. Values ​​and their components, and right Target tasks that fall below a preset alarm threshold or whose components fall below their individual thresholds are marked as having insufficient constraints, providing key constraint boundaries and risk warning information for subsequent decision-making and planning modules.

[0056] In some embodiments, the task evaluation module is further configured as follows: The urgency and importance of each target task are calculated based on priority information; The priority of each objective task is obtained by weighting the urgency and importance of each objective task and the constraint satisfaction of each objective task. The priorities of each objective task are sorted to obtain a task priority sequence; By combining the task priority sequence and the constraint satisfaction report for each objective task, the task evaluation results are obtained. By combining task urgency, importance, and constraint satisfaction to quantify priority calculation, and sorting and conflict marking of the results, the basis for multi-task scheduling has been transformed from empirical rules to scientific calculation. By comprehensively considering the inherent attributes of the task, external time constraints, and the current execution capacity of the system, a priority sequence that reasonably reflects the value and urgency of the task can be generated, providing a clear, fair, and efficient decision-making basis for task arbitration and execution order arrangement under subsequent resource competition.

[0057] Furthermore, the task evaluation module also includes a task priority calculation unit; The task priority calculation unit calculates the priority of each target task based on the task parsing results using a quantitative formula, providing a basis for multi-task scheduling; Considering the three core factors of urgency, importance, and constraint satisfaction, a task priority calculation model is established:

[0058] in, For the first The priority of each target task, with a value range of [0,1]. The larger the value, the higher the priority of the target task. For the first The urgency of each target task ranges from [0,1] and is calculated based on the difference between the latest completion time of the target task and the current time. The smaller the difference, the higher the urgency. Urgency is used to measure how pressing a task is in terms of time, and it is calculated based on the task's deadline. The system first calculates the available time for the task, which is the latest completion time of the task minus the current time. If the calculation result is negative, it means that the timeout has occurred, and the available time is treated as 0.

[0059] The system then compares this available time with a preset time reference value, which is a benchmark time unit used to measure the degree of urgency. Finally, the urgency value is set such that the shorter the available time, the higher the urgency; the longer the available time, the lower the urgency. If the available time exceeds the preset reference value, the urgency is set to 0, indicating that the task is not urgent at all. For the first The importance of each target task, ranging from [0,1], is determined based on its contribution weight to the overall system goal; the greater the contribution, the higher the importance. The system pre-sets an importance mapping table to map target tasks to their respective importance values. Its range is ; For example, for life-saving or emergency evacuation missions, importance The importance rating is 1.0; for tasks involving the transportation of core materials or critical instructions, the importance rating is... The importance score is 0.8; for routine inspections or data collection tasks, the importance score is... The importance score is 0.5; for tasks involving background data backup or log upload, the importance score is... It is 0.3; For the first The constraint satisfaction of each objective task, with a value range of [0,1]. , , The weighting coefficients corresponding to urgency, importance, and constraint satisfaction are respectively satisfied. The determination is made through offline optimization based on the needs of the application scenario; After calculating the priority of each target task, the task priority calculation unit sorts the target tasks, outputs the task priority sequence, and marks the target tasks that conflict. Among them, conflicting target tasks can be two high-priority target tasks that need to occupy the same execution resources.

[0060] In some embodiments, the adaptive intelligent decision-making system further includes a decision planning module. The decision planning module retrieves candidate decision strategies from a pre-defined decision strategy library based on scenario type and task priority sequence, optimizes the candidate decision strategies by combining the environmental model and task evaluation results to obtain the optimal decision strategy, and performs path planning and multi-task scheduling based on the optimal decision strategy to obtain the decision planning result.

[0061] In some embodiments, the decision planning module is specifically configured as follows: Based on the task evaluation results and the environmental model, optimization objectives are set for candidate decision strategies. Based on the optimization objectives, a multi-objective optimization model is established. The multi-objective optimization model is solved by a non-dominated sorting genetic algorithm to output the optimal decision strategy. Based on the optimal decision-making strategy and the aforementioned environment model, a path search is performed using an improved A* algorithm to generate a path planning scheme. Based on the task priority sequence and the health index, the execution order and resource allocation strategy of each target task are formulated to obtain a task scheduling scheme. The decision planning results are output by combining the optimal decision-making strategy, path planning scheme, and task scheduling scheme. By stipulating that the decision planning module must use a multi-objective optimization model to solve the optimization problem that integrates the environmental model and task evaluation results, and collaboratively generate strategies, paths and scheduling schemes, the decision-making process has been elevated from single-point optimization to global collaboration. This ensures that the final planning results can comprehensively weigh the benefits of task completion and the execution costs while meeting environmental safety constraints and system resource limitations, and output decision planning results that are consistent in spatial paths, action strategies and time series.

[0062] Furthermore, such as Figure 6 As shown, the decision planning module specifically includes a decision strategy library management unit, a multi-objective optimization decision unit, a dynamic path planning unit, and a task scheduling unit; The decision strategy library management unit stores decision strategies under different scenarios and different task combinations to build a structured decision strategy library; Decision-making strategies include basic strategies and composite strategies. Basic strategies are for single tasks and simple scenarios, such as "single obstacle avoidance strategy" and "straight path tracking strategy". Composite strategies are for multi-task conflicts and complex scenarios, such as "multi-obstacle avoidance + path tracking composite strategy" and "high-priority task priority and low-priority task delayed execution composite strategy". Each strategy includes information such as the applicable scenarios, applicable task types, constraints, execution steps, and parameter configurations. The decision strategy library management unit supports adding, deleting, modifying and querying strategies, and retrieves suitable candidate decision strategies based on scenario type and task priority sequence. The multi-objective optimization decision-making unit selects the optimal strategy from candidate decision strategies or generates new composite strategies based on the environment model of the dynamic environment modeling module and the task evaluation results of the task evaluation module. A multi-objective optimization decision-making unit is established with the optimization objectives of maximizing decision benefits, minimizing execution costs, and maximizing constraint satisfaction. Among them, the benefits of decision-making include task completion rate, target achievement accuracy, and system security; the execution costs include decision-making energy consumption, time consumption, and resource utilization. The task completion rate is used to evaluate the proportion of tasks that a decision strategy is expected to successfully complete. It is calculated by the system making a feasibility judgment on each task to be executed based on its current health index, remaining power, time window, and resource allocation plan in the decision strategy. If a target task is considered achievable under the current conditions, it is marked as achievable; otherwise, it is marked as inachiable. Finally, the percentage obtained by dividing the number of achievable tasks by the total number of tasks is the expected task completion rate of this strategy. Target achievement accuracy is used to measure the degree of closeness between the plan planned by the decision-making strategy and the ideal target of the task. The calculation method is as follows: for tasks with clear quantitative targets, such as "reaching coordinate point A" or "moving the end of the robotic arm to position B", the target value planned by the decision-making strategy, such as the planned end point coordinates, is compared with the ideal target value in the task instructions, and the absolute deviation is calculated. Then, divide this deviation by the maximum error threshold allowed for the task to obtain the relative deviation of the task. Calculate the relative deviation for all such tasks and take the average value. Then subtract this average relative deviation from 1 to obtain the normalized target achievement accuracy. The closer this value is to 1, the more accurate the decision scheme is. System security is used to assess the overall risk level faced by the system when executing a certain decision strategy. The lower the value, the higher the risk. The calculation method is to assess the risk based on three dimensions of risk index and assign different weights to them for comprehensive assessment. Based on the path or action sequence of the decision-making plan, combined with the predicted trajectory of dynamic obstacles in the environmental model, the probability of a collision is assessed to obtain a collision risk index. Assess whether the planned actions approach or exceed the various safety boundaries preset by the system to obtain the boundary crossing risk index; When the decision-making strategy is executed, the estimated utilization of computing resources such as processor and memory is assessed to determine whether it will cause system performance overload, so as to obtain the system overload risk index. The system calculates the weighted sum of the three risk indices and then subtracts the sum from 1 to obtain the system safety score. The closer the score is to 1, the safer the strategy is. Decision energy consumption is used to estimate the total energy required to execute a certain decision strategy. The calculation method is to pre-determine the estimated energy consumption of the target task based on the task type, the executing agency and historical data, summarize the estimated energy consumption of all tasks planned to be completed by the strategy, and add the static energy consumption necessary for the system to maintain basic operation during execution to obtain the decision energy consumption of the strategy. Time consumption is used to estimate the total time required to execute a certain decision strategy. The calculation method is to determine the order and parallel relationship of each task under the task scheduling scheme of the decision planning, find the execution path with the longest time consumption, obtain the estimated execution time of the target task based on historical data, add up the estimated execution time of all tasks on the path, and add fixed time costs such as task switching and system response, which is the estimated time consumption. Resource utilization rate is used to assess the degree of occupancy of critical system resources when implementing a certain decision strategy. It is calculated by calculating the ratio of the estimated usage of each critical resource to the total capacity of that resource, and taking the maximum value among all resource utilization rates as the resource utilization rate of the strategy.

[0063] The multi-objective optimization model is solved by non-dominated sorting genetic algorithm, and the optimal decision strategy is output, which clarifies the execution order of each objective task, resource allocation scheme, and alternative solutions to cope with environmental changes. The dynamic path planning unit generates a dynamically adjustable path planning scheme based on the environment model and the optimal decision-making strategy for target tasks that require spatial movement. The dynamic path planning unit uses the improved A* algorithm as the basic path search algorithm, and combines the environmental dynamics evaluation index and path safety margin to optimize the path search cost function. The path safety margin is used to evaluate the safety of the planned path, taking into account obstacle dynamics and system constraints. Its calculation formula is as follows:

[0064] in, For the first The path safety margin of each path ranges from [0,1], with a larger value indicating a safer path. For the first The first path The distance from each path point to the nearest obstacle; The preset safe distance threshold for the system is determined based on the scenario and the performance of the actuator. For the first The total number of path points on the path; This is the dynamic correction coefficient for obstacles, with a value range of [0.8, 1.2]. The higher the speed of the dynamic obstacle, the greater the correction coefficient. The smaller the value, the better it is used to correct the impact of dynamic obstacles on path safety; For the first The rate of change of curvature of a path reflects the smoothness of the path; The maximum permissible rate of change of curvature is determined based on the kinematic constraints of the actuator; The dynamic path planning unit searches for an initial path that meets the path safety margin requirements by improving the A* algorithm. Combined with obstacle prediction trajectory, the rolling window method is used to update the path in real time. When the environment changes suddenly or the movement state of the obstacle changes, path replanning is triggered to ensure the safety and feasibility of the path. The task scheduling unit formulates a task scheduling plan for each target task. The task scheduling plan includes the task execution order, execution time allocation, resource allocation, and task switching rules. For tasks without conflicts, they are executed sequentially or in parallel according to priority. For tasks with conflicts, scheduling strategies such as prioritizing high-priority tasks, pausing or delaying the execution of low-priority tasks, splitting tasks and executing them in time, and dynamically allocating resources are adopted. The task scheduling unit outputs a task scheduling scheme, providing a basis for subsequent algorithm adaptation and execution control.

[0065] In some embodiments, the adaptive intelligent decision-making system further includes an algorithm adaptation module. The algorithm adaptation module is used to establish an algorithm evaluation index system based on scenario type and task analysis results, calculate the comprehensive performance score of each algorithm in a preset algorithm library based on the algorithm evaluation index system, and select the optimal algorithm based on the comprehensive performance score.

[0066] In some embodiments, the algorithm adaptation module includes an algorithm library unit, an algorithm performance evaluation unit, an adaptive algorithm selection unit, and an online algorithm parameter tuning unit; The algorithm library unit stores a variety of intelligent decision-making and control algorithms, covering multiple categories such as reinforcement learning, fuzzy logic, model predictive control, and adaptive dynamic programming. Among them, reinforcement learning algorithms include DQN, PPO, and SAC algorithms; fuzzy logic algorithms include traditional fuzzy control algorithms and adaptive fuzzy control algorithms; model predictive control algorithms include linear model predictive control algorithms and nonlinear model predictive control algorithms; and adaptive dynamic programming algorithms include value-based ADP algorithms and policy-based ADP algorithms. Each algorithm includes information such as algorithm principle, applicable scenarios, parameter range, computational complexity, and performance indicators; The algorithm library unit supports the expansion and updating of algorithms to meet the algorithm requirements of different scenarios and tasks.

[0067] The algorithm performance evaluation unit establishes an algorithm evaluation index system based on scenario type, preset task requirements, and operating parameters; The algorithm evaluation index system includes five core algorithm evaluation indicators: real-time performance, accuracy, robustness, algorithm energy consumption, and resource utilization. Real-time performance measures the speed at which an algorithm executes, directly impacting system response time. It is calculated through performance testing, recording the total time taken for the algorithm to process data from input to output; this time is typically called the single execution time. Under the same computational conditions, algorithms with shorter single execution times receive higher real-time performance scores. Accuracy is used to measure the accuracy of the algorithm's output. It is calculated by running the algorithm in a test set or simulation environment with a clear standard answer, comparing the algorithm's output with the standard expected target value, and calculating the magnitude of the deviation between them. Common indicators include mean absolute error, root mean square error, or classification accuracy. The smaller the deviation or the higher the accuracy, the higher the accuracy score of the algorithm. Robustness is used to measure the stability and adaptability of an algorithm under non-ideal conditions. The calculation method is to artificially introduce various disturbances such as input data noise, sensor signal abrupt changes, large changes in environmental parameters, and model parameter drift in a standard test environment, and then observe and compare the performance of the algorithm under these disturbance conditions with that under standard conditions. The smaller the performance drop and the stronger the retention ability of the algorithm, the higher its robustness score. It can be quantified by multiple disturbance experiments and statistically analyzing the average retention rate of performance. Algorithm energy consumption measures the system energy consumed by an algorithm during runtime. It is calculated by monitoring the total system power consumption during algorithm execution and subtracting the baseline power consumption when the system is idle. The resulting energy consumption increment is typically the average of multiple runs. All other things being equal, an algorithm that consumes less energy receives a higher energy consumption score. Resource utilization rate measures the degree to which an algorithm consumes system computing resources during runtime. It is calculated by monitoring the usage of key hardware resources during algorithm execution, and is usually expressed as a percentage. Examples include the average CPU utilization, peak or average memory usage during algorithm execution. A lower resource utilization rate means a lighter computational burden on the system, and a higher resource utilization rate score. The algorithm performance evaluation unit uses the analytic hierarchy process (AHP) to determine the weights of each algorithm evaluation index and calculates the comprehensive performance score of each algorithm based on the weighted summation method. The comprehensive performance score ranges from [0,1], and the higher the score, the better the algorithm's performance in the current scenario.

[0068] The adaptive algorithm selection unit dynamically selects the optimal algorithm based on the comprehensive performance score of each algorithm and the preset decision planning requirements; When a single algorithm has a comprehensive performance score greater than or equal to 0.8, that algorithm is directly selected as the optimal algorithm. When the overall performance score of no single algorithm is greater than or equal to 0.8, an algorithm fusion strategy is adopted. Two to three algorithms with better performance are selected, and a weighted fusion mechanism is used to assign fusion weights based on the overall performance scores of each algorithm. Then, the algorithms are fused to obtain the optimal algorithm. For example, reinforcement learning algorithms can be fused with model predictive control algorithms, or fuzzy logic algorithms can be fused with adaptive dynamic programming algorithms. The adaptive algorithm selects the unit outputting the optimal algorithm, providing algorithmic support for execution control.

[0069] The online parameter tuning unit for the optimal algorithm uses gradient descent combined with reinforcement learning to achieve online tuning of algorithm parameters. First, determine the key parameters of the optimal algorithm, such as the learning rate and discount factor of the reinforcement learning algorithm, the membership function parameters of the fuzzy control algorithm, and the prediction time domain and control time domain of the model predictive control algorithm. Then, a parameter optimization objective function is established, with the goal of maximizing the overall performance score of the algorithm. The parameter adjustment direction is calculated based on the gradient descent method, and the parameter adjustment step size is dynamically adjusted in combination with the reinforcement learning strategy to avoid the parameters getting stuck in local optima. The online algorithm parameter tuning unit monitors the algorithm execution effect in real time and updates the algorithm parameters once after each optimization cycle to ensure that the algorithm performance remains at its optimal state.

[0070] In some embodiments, the adaptive intelligent decision-making system further includes an execution control module. The execution control module generates control commands based on the decision planning results and the optimal algorithm, drives the actuator to perform the corresponding target task based on the control commands, and generates execution results.

[0071] Furthermore, the execution control module includes a control command generation unit, a multi-mode execution adaptation unit, an interference suppression unit, and a safety boundary control unit; The control command generation unit generates standardized control commands based on the decision planning results of the decision planning module and the optimal algorithm selected by the algorithm adaptation module. For mobile tasks such as navigation and trajectory tracking, generate speed commands, position commands, and attitude commands; for operational tasks such as robotic arm operations and equipment control, generate force / torque commands, displacement commands, and motion sequence commands. The format and parameters of the control commands are determined based on the type of actuator and the interface protocol to ensure that the actuator can accurately parse them; The control instruction generation unit simultaneously generates the time sequence of instruction execution, specifying the execution order, execution time, and duration of each control instruction.

[0072] The multi-mode execution adaptation unit is used to adapt to different types of actuators; the execution structure includes mobile actuators, operational actuators, and equipment control actuators. For example, mobile actuators may include wheeled chassis, tracked chassis, and multi-rotor aircraft; operating actuators may include robotic arms and grippers; and equipment control actuators may include motors, cylinders, and solenoid valves. The multi-mode execution adaptation unit has built-in drive protocols and control logic for various actuators, and can convert standardized control instructions into actuator-specific drive instructions according to the type of actuator. For example, for a wheeled differential drive chassis, speed commands are converted into left and right wheel speed commands; for a robotic arm, displacement commands are converted into angle commands for each joint. Meanwhile, the multi-mode execution adaptation unit supports collaborative control of multiple execution agencies, ensuring that each execution agency works synchronously according to the task scheduling plan; The interference suppression unit employs a sliding mode control algorithm to suppress the impact of external interference and system uncertainties on the execution process; External disturbances include vibrations caused by uneven road surfaces, wind disturbances, load fluctuations, and signal interference; system uncertainties include actuator parameter drift and model errors. Sliding mode control algorithms, by designing switching functions and control laws, enable the system's state trajectory to converge to the sliding surface within a finite time, thereby achieving robust suppression of disturbances and uncertainties. By combining environmental change information from the decision planning module, when strong interference or environmental change is detected, the switching gain of the sliding mode control is automatically adjusted to enhance the interference suppression capability. The safety boundary control unit establishes the execution safety boundary based on the system's preset safety constraints and risk information in the environmental model; Safety boundaries include, but are not limited to, position boundaries, velocity boundaries, attitude boundaries, and force or torque boundaries; Among them, the position boundary includes the prohibited dangerous area; the speed boundary includes the maximum and minimum permissible speeds; the attitude boundary includes the maximum permissible tilt angle; and the force or torque boundary includes the maximum permissible output torque of the robotic arm. The safety boundary control unit monitors the actual execution status of the actuator in real time. When it detects that the execution status is close to or exceeds the safety boundary, it automatically triggers the safety control strategy. The safety control strategy includes deceleration, stopping, and emergency avoidance to ensure the safety of the execution process. To achieve adaptive and stable control during the execution process, the execution control module employs an adaptive PID control algorithm. Combining disturbance suppression and safety boundary control requirements, it calculates the control quantity using the following formula:

[0073]

[0074]

[0075]

[0076] in, for Control quantity at any given moment; for Tracking error at any given moment; for The rate of change of the tracking error at any given time; for Adaptive scaling at any given time; for Adaptive integral at time step; for The adaptive differential coefficients at time t; , , These are the initial values ​​for the coefficients; , , This represents the coefficient update rate.

[0077] Based on the aforementioned adaptive intelligent decision-making system, by setting up an environmental perception module, task evaluation module, decision planning module, algorithm adaptation module, and execution control module, a complete technical closed loop covering perception, cognition, decision-making, and execution is constructed. The responsibilities and data flow of each link are clearly defined, ensuring that the system can perform coordinated and consistent adaptive processing in response to changes in complex external scenarios and the needs of multiple internal tasks. This solves the problems of fragmented modules and slow response in traditional systems, and provides a system-level solution for intelligent agents to achieve stable and reliable autonomous operation in dynamic and unknown environments.

[0078] Preferably, taking the specific scenario of a wheeled mobile robot performing material handling tasks in a complex factory environment as an example, the core actuator of the wheeled mobile robot is a differential drive wheeled chassis. The task is to transport materials from point A in the warehouse area to point B on the assembly line within a dynamic factory workshop. The workshop environment is complex, with other moving AGVs, occasional personnel passing by, and fixed shelves and equipment.

[0079] Upon system startup, the multi-source perception fusion module constructs an environmental model encompassing static shelves and the movement trajectories of other AGVs using LiDAR, visual cameras, and ultrasonic sensors. The task evaluation module receives the instruction to "transport materials from point A to point B" and parses out the core task objective.

[0080] The decision-making and planning module, based on the environment model, plans a safe and efficient path from point A to point B. Simultaneously, the task evaluation module calculates "avoiding dynamic obstacles" as a high-priority subtask. After evaluating the current scenario, the algorithm adaptation module selects model predictive control from the algorithm library as the optimal algorithm for accurate trajectory tracking and dynamic obstacle avoidance.

[0081] The system receives the desired path and desired velocity curve output by the decision planning module, and combines them with the MPC algorithm to generate a set of standardized control commands, such as: "At time t1, the target's forward speed is 0.8 m / s and the target's heading angle is 30 degrees."

[0082] Since the robot's actuator is a differential drive wheeled chassis, this unit converts the aforementioned standardized instructions into dedicated instructions that the chassis drive motors can recognize. Specifically, it uses a kinematic model to calculate "forward speed 0.8m / s" and "heading angle 30 degrees" into low-level motor control instructions of "target speed of left wheel X revolutions per minute, target speed of right wheel Y revolutions per minute".

[0083] When the robot travels on an uneven surface, the robot body experiences bumps and vibrations, and the wheels may slip slightly, constituting external interference. At this point, the sliding mode controller built into the interference suppression unit activates. It compares the robot's actual motion state with the desired state given by the MPC algorithm in real time. When it detects a deviation from the trajectory caused by bumps, the sliding mode controller quickly generates an additional compensation control variable, superimposed on the motor commands, forcing the system state to "slide" towards and maintain the desired trajectory surface. This suppresses interference from uneven road surfaces, ensuring smooth travel and accurate trajectory tracking.

[0084] The safety boundary control unit defines dynamic safety boundaries for the robot based on pre-set rules and real-time environmental model information. For example, the maximum speed boundary is set to 5 km / h, and a virtual electronic fence is set up outside the human activity area as the position boundary. During the robot's task execution, this unit monitors its actual speed and position in real time. If, for any reason, the robot's speed exceeds 4.5 km / h or its trajectory deviates from the designated restricted area, the safety boundary control unit will immediately trigger safety strategies. It may directly send a "sudden deceleration" command to the motors or request a completely new emergency avoidance path from the decision-making and planning module, thereby ensuring that collisions or speeding hazards are absolutely avoided.

[0085] After the task is completed, the iterative optimization module analyzes the execution data throughout the process. For example, it might discover that the MPC algorithm performs poorly on a certain curve, or that the gain of the sliding mode controller can be better adjusted on a certain road surface. Based on this data, the system automatically updates the algorithm parameter library and controller parameters, resulting in better overall performance when executing the task in a similar scenario next time.

[0086] In some embodiments, the adaptive intelligent decision-making system further includes an iterative optimization module. The iterative optimization module is configured as follows: The execution results are compared with the preset decision-making expectations, decision evaluation indicators are calculated, and decision weights are set for the decision evaluation indicators; The decision evaluation indicators are weighted based on the decision weights to obtain the decision evaluation results. Based on the decision evaluation results, the model parameters in the environmental model, the policy rules in the preset decision policy library, the algorithm parameters in the preset algorithm library, and the decision weights are iteratively updated. Through the iterative optimization module, an evaluation and feedback mechanism based on the comparison between the execution results and the expected goals can be established, and the model parameters, policy rules and algorithm parameters can be updated accordingly. This gives the system the ability to continuously learn and optimize from historical experience, breaks the limitations of traditional open-loop decision-making systems, and forms a complete learning closed loop of decision-making, execution, evaluation and optimization. This enables the system to automatically correct model biases, refine effective strategies and optimize algorithm performance, thereby achieving autonomous evolution and long-term improvement of the overall intelligence level.

[0087] Furthermore, the iterative optimization module includes an execution status monitoring unit, a decision effect evaluation unit, a data storage and analysis unit, and a system parameter iterative update unit; The execution status monitoring unit is used to monitor the actual execution status of the execution control module in real time, collect the position, speed, attitude, force or torque of the actuator, obtain the actual output data, and collect the execution status data during the execution process; The execution status monitoring unit uses a data verification algorithm to ensure the accuracy of the collected data, and uses timestamp synchronization to achieve the correspondence and matching between execution status data and control commands, and outputs an execution status monitoring report; the execution status monitoring report includes execution results, execution errors, and abnormal event information during the execution process; The decision effectiveness evaluation unit is used to compare the execution results with the decision-making expectations in the decision planning results, calculate decision evaluation indicators, and evaluate the decision effectiveness. Among them, the decision evaluation indicators include task completion rate, target achievement accuracy, execution efficiency, energy consumption control effect, and safety performance; The overall score of decision-making effectiveness is calculated using the fuzzy comprehensive evaluation method. By mimicking the thinking process of human experts in comprehensive evaluation, it combines precise data with subjective judgment rules. The calculation process consists of the following steps: The first step is to define the evaluation items and levels. The system defines five aspects as evaluation items: task completion rate, target achievement accuracy, execution efficiency, energy consumption control effect, and safety performance. For each item, four evaluation levels are set: "Excellent," "Good," "Average," and "Poor." Each level has a corresponding numerical standard. For example, "Excellent" for "task completion rate" may mean that the completion rate reaches more than 95%. The second step is to determine the importance of each item. Based on the core objectives of the current task scenario, the system assigns different weights to the five evaluation items. For example, in a scenario prioritizing security, "security performance" has the highest weight; in a scenario prioritizing efficiency, "execution efficiency" has the highest weight, and the sum of the weights of the five items is 1.

[0088] The third step involves performing fuzzy scoring for each item. The system compares the five data points after actual execution with the level standards set in the first step. For example, a completion rate of 92% might be judged as 80% "Excellent" and 20% "Good". After performing this calculation for all five items, a table is generated that includes the degree of belonging of all items to different levels.

[0089] The fourth step is to calculate the comprehensive fuzzy evaluation. The system performs specific mathematical calculations on the weights from the second step and the membership table obtained in the third step, and combines the fuzzy scores of each item according to their importance to finally obtain an overall evaluation vector, which describes the probability that the overall effect of this decision belongs to the four levels of "excellent, good, average, and poor". The fifth step is to obtain the specific score. The result of the fourth step is still a probability distribution. The system assigns a representative score to each of the four levels, "Excellent," "Good," "Average," and "Poor," using methods such as weighted averaging. For example, 100, 80, 60, and 40 points respectively. Then, based on the probability proportion of each level in the overall evaluation vector, a precise final score is calculated.

[0090] Step 6: Result Determination and Application. The system will process this final score, typically a value between 0 and 1, and compare it with a preset threshold to determine whether the decision-making effect is "excellent," "average," or "poor." This determination and score will be recorded for subsequent analysis of the effectiveness of the decision-making strategy and optimization of system parameters.

[0091] When the overall decision performance score is greater than or equal to 0.8, the decision performance is considered excellent; when the overall decision performance score is greater than or equal to 0.6 and less than 0.8, the decision performance is considered average; when the overall decision performance score is less than 0.6, the decision performance is considered poor. At the same time, the reasons for poor decision-making results were analyzed, including environmental modeling errors, task evaluation biases, unreasonable decision-making strategies, improper algorithm adaptation, and execution interference. The data storage and analysis unit is used to store various types of data during system operation, including multi-source environmental perception data, data in the environmental model, task evaluation results, decision planning results, algorithm parameters, execution status data, and decision evaluation results; it adopts a distributed storage architecture to support efficient storage and retrieval of massive amounts of data; The data storage and analysis unit uses big data analytics to uncover patterns in the data and obtain data mining conclusions. Among them, the correlation patterns in the data include the matching patterns between environmental characteristics and decision-making strategies, the correlation patterns between algorithm parameters and execution performance, and the mapping patterns between system state and task completion rate; data mining is used to identify the optimization space of the system and provide data support for iterative updates.

[0092] The system parameter iterative update unit updates the parameters of the environment model, the policy rules in the decision policy library, the algorithm parameters in the algorithm library, and the decision weights based on the decision evaluation results and data mining conclusions. The updating of environmental model parameters includes the correction of static environmental feature parameters, the optimization of parameters of dynamic obstacle sub-models, and the adjustment of mutation thresholds. Updating the policy rules in the decision policy library includes revising the applicable conditions of policies, adding new policies, and deleting inefficient policies. The algorithm parameters in the algorithm library are updated based on the optimal parameters obtained from online tuning, and the default parameters of the corresponding algorithms in the algorithm library are also updated. The decision weights are updated based on the feedback of decision effects in different scenarios, and the weight coefficients of each decision evaluation indicator are adjusted accordingly. The system parameter iterative update unit adopts an incremental update mechanism, performing an iterative update once after each set of tasks is completed or after a preset time period, so that the system performance continues to improve over time.

[0093] like Figure 7 As shown, this embodiment of the invention also provides an adaptive intelligent decision-making method for complex scenarios, including the following steps: In the environmental perception step S1, multi-source perception data of the target complex scene is acquired, environmental features of the multi-source perception data are extracted, scene type is identified based on environmental features, and obstacle motion state and sudden events are detected to build an environmental model. Task evaluation step S2: Receive and parse the target task instructions to obtain the task parsing results, calculate the priority of each target task based on the task parsing results, obtain the task priority sequence, and generate the task evaluation results; In decision planning step S3, based on the scenario type and the task priority sequence, candidate decision strategies are retrieved from the preset decision strategy library. The candidate decision strategies are optimized by combining the environment model and task evaluation results to obtain the optimal decision strategy. Based on the optimal decision strategy, path planning and multi-task scheduling are performed to obtain the decision planning result. In algorithm adaptation step S4, an algorithm evaluation index system is established based on the scenario type and task parsing results. The comprehensive performance score of each algorithm in the preset algorithm library is calculated based on the algorithm evaluation index system, and the optimal algorithm is selected based on the comprehensive performance score. Execution control step S5: Based on the decision planning results and the optimal algorithm, control instructions are generated. Based on the control instructions, the actuator is driven to perform the corresponding target task and the execution result is generated.

[0094] By mapping the module functions in the system to sequentially executed steps of environmental perception, task evaluation, decision planning, algorithm adaptation, and execution control, a complete operational process for implementing this adaptive intelligent decision-making system is provided. This process clarifies the orderly connection and data transmission relationship between each technical stage, ensuring that the system can be fully implemented through specific methods and procedures.

[0095] In some embodiments, the adaptive intelligent decision-making method further includes an iterative optimization step S6; Specifically, iterative optimization step S6 is as follows: The execution results are compared with the preset decision-making expectations, decision evaluation indicators are calculated, and decision weights are set for the decision evaluation indicators; The decision evaluation indicators are weighted based on the decision weights to obtain the decision evaluation results. Based on the decision evaluation results, the model parameters in the environmental model, the policy rules in the preset decision policy library, the algorithm parameters in the preset algorithm library, and the decision weights are iteratively updated.

[0096] Furthermore, if there are unfinished target tasks, the environmental perception step S1, task evaluation step S2, decision planning step S3, algorithm adaptation step S4, execution control step S5, and iterative optimization step S6 are repeated. If all target tasks are completed, or a system termination command is received, a task completion report is output, and the system is reset to standby mode.

[0097] It should be noted that the above is a reference method for an adaptive intelligent decision-making system and method for complex scenarios, and the present invention is not limited thereto.

[0098] The embodiments of the present invention achieve efficient fusion of multi-source information in complex scenarios, scientific arbitration of multi-task conflicts, and adaptive balancing of system resources and execution risks, thereby comprehensively improving the decision robustness, execution security, and long-term evolution capability of intelligent systems in uncertain environments. It solves the technical problems of existing technologies, which use static models, fixed rules, and open-loop architectures, making it difficult to adapt to dynamic environmental changes, unable to coordinate multi-task conflicts, and lacking online learning capabilities, resulting in lagging decision-making, insufficient robustness, and inability to continuously optimize performance in complex scenarios.

[0099] Finally, it should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other. The above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them; although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications can still be made to the specific implementation of the present invention or equivalent substitutions can be made to some technical features without departing from the spirit of the technical solutions of the present invention, and all such modifications and substitutions should be covered within the scope of the technical solutions claimed in the present invention.

Claims

1. An adaptive intelligent decision-making system for complex scenarios, characterized in that, include: The environmental perception module is used to acquire multi-source perception data of the target complex scene, extract environmental features from the multi-source perception data, identify scene types based on the environmental features, and detect obstacle motion states and sudden events to construct an environmental model. The task evaluation module is used to receive and parse the target task instructions, obtain the task parsing results, calculate the priority of each target task based on the task parsing results, obtain the task priority sequence, and generate the task evaluation results. The decision planning module is used to retrieve candidate decision strategies from a preset decision strategy library based on the scenario type and the task priority sequence, optimize the candidate decision strategies in combination with the environment model and the task evaluation results to obtain the optimal decision strategy, and perform path planning and multi-task scheduling based on the optimal decision strategy to obtain the decision planning result. The algorithm adaptation module is used to establish an algorithm evaluation index system based on the scenario type and the task parsing result, calculate the comprehensive performance score of each algorithm in the preset algorithm library based on the algorithm evaluation index system, and select the optimal algorithm based on the comprehensive performance score. The execution control module is used to generate control instructions based on the decision planning results and the optimal algorithm, drive the execution mechanism to perform the corresponding target task based on the control instructions, and generate execution results.

2. The adaptive intelligent decision-making system for complex scenarios according to claim 1, characterized in that, The environment perception module is specifically configured as follows: The original heterogeneous data of the target complex scene is acquired by multi-source heterogeneous sensors, and the original heterogeneous data is preprocessed to obtain the multi-source sensing data. Environmental features are extracted from the multi-source sensing data, and the local decision-making results built into the multi-source heterogeneous sensors are fused with the environmental features to obtain environmental sensing results, thereby constructing the environmental model.

3. The adaptive intelligent decision-making system for complex scenarios according to claim 2, characterized in that, The environment perception module is also configured to: Extract static and dynamic environmental feature parameters from the environmental perception results to construct an environmental feature database; Based on the environmental feature database, scene types are identified using a deep learning model; Based on the dynamic environmental characteristic parameters, the extended Kalman filter algorithm is used to establish a kinematic model of the obstacle and predict the motion state of the obstacle. Calculate the rate of change of the dynamic environmental feature parameters within a preset sliding window, compare the rate of change with a mutation threshold, and identify mutation events; The environment model is constructed by combining the static environment feature parameters, the scene type, the obstacle motion state, and the mutation event.

4. The adaptive intelligent decision-making system for complex scenarios according to claim 3, characterized in that, The environment perception module is also configured to: Based on the obstacle's motion state and the static environmental characteristic parameters, the dynamic obstacle density index and the rate of change of environmental characteristics are calculated respectively. The dynamic obstacle density index and the rate of change of environmental characteristics are weighted and calculated to obtain the environmental dynamic assessment index; The environmental dynamics assessment index is compared with the index threshold to obtain the comparison result. Based on the comparison result, the time step and covariance matrix value in the extended Kalman filter algorithm are adjusted to adjust the update frequency and modeling accuracy of the environmental model.

5. The adaptive intelligent decision-making system for complex scenarios according to claim 1, characterized in that, The task evaluation module is specifically configured as follows: Real-time monitoring of the running status of hardware components and software processes, and acquisition of multiple running parameters; Based on the parameter characteristics of each of the aforementioned operating parameters, the corresponding normalized value is calculated. The normalized values ​​corresponding to each of the aforementioned operating parameters are then weighted to obtain the health index.

6. The adaptive intelligent decision-making system for complex scenarios according to claim 5, characterized in that, The task evaluation module is further configured as follows: The target parameters, execution requirements, time windows, and priority information of each target task in the target task instruction are parsed to obtain the task parsing result; Based on the health index and the preset performance requirements of each target task, calculate the performance satisfaction rate of each target task; Based on the operating parameters and the preset energy consumption requirements of each target task, calculate the energy sufficiency rate of each target task; Based on the time window, the time sufficiency of each target task is calculated; By combining the performance satisfaction rate, the energy sufficiency rate, and the time sufficiency, the constraint satisfaction rate of each target task is calculated, and a constraint satisfaction report for each target task is obtained.

7. The adaptive intelligent decision-making system for complex scenarios according to claim 6, characterized in that, The task evaluation module is further configured as follows: The urgency and importance of each target task are calculated based on the priority information. The priority of each target task is obtained by weighting the urgency and importance of each target task and the constraint satisfaction of each target task. The priorities of each target task are sorted to obtain the task priority sequence; The task evaluation result is obtained by combining the task priority sequence and the constraint satisfaction report of each target task.

8. The adaptive intelligent decision-making system for complex scenarios according to claim 7, characterized in that, The decision-making and planning module is specifically configured as follows: Based on the task evaluation results and the environment model, optimization objectives are set for the candidate decision strategies. Based on the optimization objectives, a multi-objective optimization model is established. The multi-objective optimization model is solved by a non-dominated sorting genetic algorithm, and the optimal decision strategy is output. Based on the optimal decision-making strategy and the environment model, a path search is performed using an improved A* algorithm to generate a path planning scheme. Based on the task priority sequence and the health index, the execution order and resource allocation strategy of each target task are formulated to obtain a task scheduling scheme. The decision planning result is output by combining the optimal decision-making strategy, the path planning scheme, and the task scheduling scheme.

9. The adaptive intelligent decision-making system for complex scenarios according to any one of claims 1 to 8, characterized in that, The adaptive intelligent decision-making system further includes an iterative optimization module, which is configured as follows: The execution results are compared with the preset decision expectation targets, decision evaluation indicators are calculated, and decision weights are set for the decision evaluation indicators; The decision evaluation indicators are weighted based on the decision weights to obtain the decision evaluation results. Based on the decision evaluation results, the model parameters in the environment model, the strategy rules in the preset decision strategy library, the algorithm parameters in the preset algorithm library, and the decision weights are iteratively updated.

10. An adaptive intelligent decision-making method for complex scenarios, characterized in that, Includes the following steps: The environmental perception step involves acquiring multi-source perception data of the target complex scene, extracting environmental features from the multi-source perception data, identifying the scene type based on the environmental features, and detecting obstacle motion states and abrupt events to construct an environmental model. The task evaluation step involves receiving and parsing the target task instructions to obtain the task parsing results, calculating the priority of each target task based on the task parsing results, obtaining the task priority sequence, and generating the task evaluation results. The decision planning step involves retrieving candidate decision strategies from a preset decision strategy library based on the scenario type and the task priority sequence, optimizing the candidate decision strategies in conjunction with the environment model and the task evaluation results to obtain the optimal decision strategy, and performing path planning and multi-task scheduling based on the optimal decision strategy to obtain the decision planning result. The algorithm adaptation steps include: establishing an algorithm evaluation index system based on the scenario type and the task parsing results; calculating the comprehensive performance score of each algorithm in the preset algorithm library based on the algorithm evaluation index system; and selecting the optimal algorithm based on the comprehensive performance score. The execution control steps involve generating control commands based on the decision planning results and the optimal algorithm, driving the actuator to perform the corresponding target task based on the control commands, and generating execution results.

Images