An AI intelligent collaborative control method based on IQL algorithm

Abstract

The present invention discloses an AI intelligent collaborative control method based on the IQL algorithm, which relates to the field of artificial intelligence control technology and includes the following steps: obtaining the local observation state sequence of the target agent within a continuous control cycle, dividing the local observation state sequence into multiple time windows of equal time length, extracting observation feature variables in each time window, and forming a basic observation vector group. The present invention realizes the joint quantitative expression of local state disturbance and environmental behavior complexity by constructing a disturbance index system and introducing a time decoupling control mechanism, and effectively suppresses multi-agent synchronization conflicts and strategy resonance problems through dynamic rhythm adjustment and slightly staggered execution, reduces the risk of path blockage and behavioral disorder, and significantly improves the collaborative consistency and operational stability of the multi-agent system in a dynamic interference environment.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence control technology, and in particular to an AI intelligent collaborative control method based on an IQL algorithm. Background Art

[0002] AI intelligent collaborative control based on the IQL algorithm refers to a collaborative decision-making control method that uses the Independent Q-Learning (IQL) algorithm to implement collaborative decision-making and control of each agent in an independent learning state in a multi-agent system. This method does not rely on global state information or centralized strategies. Instead, each agent maintains its own Q-value function and independently updates its strategy based on its own observations and environmental feedback, thereby gradually achieving collaborative optimization of the entire system while ensuring local optimal behavior. In collaborative control tasks, multiple AI agents can use the IQL algorithm to effectively learn how to achieve collective goals in a changing environment, such as multi-robot path collaboration, vehicle-road collaborative control, and distributed energy scheduling, under conditions of limited communication and partially observable states. It has the characteristics of low computational overhead, strong adaptability, and easy scalability.

[0003] Existing technologies have the following shortcomings: In existing multi-agent collaborative control methods based on the IQL (Independent Q-Learning) algorithm, each agent relies on its own local observation information to update its strategy during training and decision-making. It lacks direct perception of the global state and a mechanism to infer the strategic intentions of other agents. Therefore, when faced with dense dynamic obstacle scenarios, the system is prone to severe local information bias. In particular, when multiple collaborative robots simultaneously perform obstacle avoidance behaviors and dynamically interfere with each other, the local state information perceived by each agent may fluctuate dramatically in a very short period of time, leading to the instantaneous collapse of the local observation space. This phenomenon occurs when the observation space, which originally had a clear state-policy mapping, is perturbed into a "blind spot state" where it is unrecognizable or has insufficient policy coverage over consecutive moments. This results in the reinforcement learning model being unable to generate effective action outputs based on the existing Q-value function. Once the observation space collapses, the intelligent agent is often unable to accurately judge the optimal behavior in the current state, and thus outputs actions outside the strategy, or even degenerates into nearly random behavior or meaningless repetitive actions, such as aimless rotation, frequent pauses, or entering high-risk areas. In severe cases, it will cause system-level failures such as path blockage, collision loss of control, and task interruption.

[0004] The above information disclosed in this Background section is only for enhancement of understanding of the background of the present disclosure and therefore it may contain information that does not form the prior art that is already known to a person of ordinary skill in the art. Summary of the Invention

[0005] The purpose of the present invention is to provide an AI intelligent collaborative control method based on the IQL algorithm. By constructing a disturbance index system and introducing a time decoupling control mechanism, it can realize the joint quantitative expression of local state disturbances and environmental behavior complexity, and through dynamic rhythm adjustment and slightly staggered execution, effectively suppress multi-agent synchronization conflicts and strategy resonance problems, reduce the risks of path blockage and behavioral disorder, and significantly improve the collaborative consistency and operational stability of the multi-agent system in a dynamic interference environment, so as to solve the problems in the above-mentioned background technology.

[0006] In order to achieve the above object, the present invention provides the following technical solution: an AI intelligent collaborative control method based on the IQL algorithm, comprising the following steps:

[0007] Obtain the local observation state sequence of the target agent in the continuous control cycle, and divide the local observation state sequence into multiple time windows of equal time length. Extract the observation feature variables in each time window to form a basic observation vector group;

[0008] Perform dimension-by-dimension difference calculation on the basic observation vectors in adjacent time windows, and apply the perturbation gain weight function to amplify the response intensity of the change value of the mutation feature dimension to generate the primary perturbation trend vector;

[0009] Perform normalization and normal deviation compression operations on the primary disturbance trend vector, build a reference disturbance baseline model using historical stable state samples, and calculate the disturbance offset distance index based on the Euclidean distance between the current primary disturbance trend vector and the reference disturbance baseline model;

[0010] The disturbance offset distance indicator is fused with the local spatial dynamic entropy factor within the current position area of ​​the target agent, and the fusion result is output as a local disturbance response value with a single dimension;

[0011] A time sliding window is constructed for multiple continuous local disturbance response values, and extreme value capture and change rate weighted averaging operations are performed within the time sliding window to generate a local state disturbance index;

[0012] When the local state disturbance index is greater than the preset reference threshold, the space-time decoupling control mechanism is triggered to dynamically adjust the action execution cycle of the target agent, introduce a slight time delay difference in the control cycle between multiple agents, form an asynchronous interleaved action execution window, and suppress the synchronous disturbance response between agents.

[0013] Preferably, obtaining the local observation state sequence of the target agent within the continuous control period and dividing it into multiple time windows of equal time length comprises the following steps:

[0014] Based on the target agent observation data stream collected during the continuous control cycle, time synchronization processing is performed at preset time intervals to align the observation data at each control moment, ensuring the integrity and consistency of the data sequence in the time dimension;

[0015] After completing the time division, the short-term trajectory trend estimation is performed on the observation data in each time window, and the position change value, the distance value of the adjacent obstacles, the local density factor value and the number of feasible paths of the target agent in the time period are combined.

[0016] The observation feature variables extracted in each time window are normalized and combined into a basic observation vector group in a unified format.

[0017] Preferably, the specific steps of performing dimension-by-dimension difference calculation on basic observation vectors in adjacent time windows to generate a primary disturbance trend vector are as follows:

[0018] For the basic observation vectors in adjacent time windows, the dimension-by-dimension difference calculation is performed in the order of the corresponding feature dimensions to establish the basic state disturbance vector, which is used to reflect the local state fluctuation trend of the target agent in each continuous time period;

[0019] To enhance the detection sensitivity of sudden disturbances, a mutation recognition threshold mechanism is constructed based on the historical disturbance statistical distribution model. When the difference in any characteristic dimension exceeds the corresponding disturbance anomaly judgment threshold, it is marked as a mutation characteristic dimension, and the disturbance gain weight function is called to assign a nonlinear response amplification coefficient according to the magnitude of the mutation, thereby amplifying the response intensity of the change value of the mutation dimension.

[0020] The original difference results of the mutated feature dimension that has completed the response amplification processing and the unmutated feature dimension are fused to form a complete primary disturbance trend vector.

[0021] Preferably, the specific steps of performing normalization processing and normal deviation compression operation on the primary disturbance trend vector and calculating the disturbance offset distance index are as follows:

[0022] For each characteristic dimension of the primary disturbance trend vector, the maximum and minimum normalization processing is performed according to the distribution range of the dimension in the historical observation data, and all dimensions are mapped to a unified numerical range;

[0023] After standardization, a normal deviation compression operation is performed on each dimension of the normalized vector. Based on the mean and standard deviation obtained from historical stable state sample statistics, the Z-score standard deviation of the current eigenvalue is calculated. The degree of deviation is then mapped to a compressed disturbance intensity value using an exponential decay function to attenuate minor deviations and amplify significant deviations, thereby highlighting structural abnormalities.

[0024] Taking the historical stable state sample set as the reference disturbance baseline model, the Euclidean distance between the current compressed disturbance trend vector and the reference disturbance baseline is calculated to form a single-value disturbance offset distance indicator.

[0025] Preferably, the specific steps of fusing the disturbance offset distance indicator with the local spatial dynamic entropy factor within the current position area of ​​the target agent to output the local disturbance response value are as follows:

[0026] A local observation grid is constructed based on the current location of the target agent. The rate of change of the number of neighboring agents is counted in real time within the current area. The rate of change of the number of neighboring agents is calculated using the trend of the number of neighboring agents in a continuous time window.

[0027] For all neighboring agents within the same local space, the action direction vectors in continuous time segments are extracted, their angle change rate and direction fluctuation amplitude are calculated, and the local space dynamic entropy factor is constructed in combination with the change rate of the number of neighbors.

[0028] The disturbance offset distance index and the local spatial dynamic entropy factor are input into the fusion function model, and a weighted normalization fusion algorithm is used to generate a local disturbance response value with a single dimension.

[0029] Preferably, the specific steps of constructing a time sliding window to process multiple continuous local disturbance response values ​​to generate a local state disturbance index are as follows:

[0030] Within the preset time scale range, a time sliding window covering the continuous control cycle of the target agent is constructed with a fixed length and sliding step size, and the corresponding local disturbance response value sequence in each time window is organized into structured time series data;

[0031] The extreme value capture operation is performed in each time sliding window to extract the maximum, minimum and average values ​​of the local disturbance response value in the window, and the disturbance change rate is calculated for the adjacent response values ​​at each time step. The corresponding dynamic weighting factors are assigned to different time points according to the fluctuation amplitude of the change rate.

[0032] All disturbance response values ​​within the window are fused with the weighted values ​​of the rate of change according to the exponentially weighted moving average method to output the local state disturbance index.

[0033] Preferably, when the local state disturbance index is greater than a preset reference threshold, the space-time decoupling control mechanism is triggered, and a normalized deviation amplitude factor is defined according to the degree of deviation between the local state disturbance index and the reference threshold. The calculation expression is as follows:

[0034] , where is the local state perturbation index, is the reference threshold of the local state disturbance index, is a logarithmic function, is the normalized deviation amplitude factor;

[0035] For each target agent participating in the current local cooperative control, the corresponding slight time delay adjustment factor is calculated. The calculation expression is as follows:

[0036] , where is the delay control reference amplitude constant, is the agent's number, is the total number of agents, is the mathematical constant, pi, It is The nonlinear distribution weights of the agents, It is an intelligent agent At the moment Action execution time delay adjustment factor;

[0037] After obtaining the action execution time delay adjustment factor for each target agent Afterwards, the action execution cycle is dynamically modified to calculate the actual action execution cycle of each agent at the current moment. The calculation expression is as follows:

[0038] , where It is An agent at time The actual action execution cycle, is the basic control cycle.

[0039] In the above technical solution, the technical effects and advantages provided by the present invention are:

[0040] By introducing disturbance trend modeling, dynamic disturbance response assessment, and a sliding window aggregation mechanism, a spatiotemporally continuous disturbance index system was constructed. This system achieves a joint quantitative expression of the intensity of local state changes and the complexity of environmental behavior, improving the system's perception resolution of key disturbance events. Furthermore, a disturbance-driven temporal decoupling control mechanism was combined to dynamically adjust the agent's action execution rhythm and introduce a slight time staggering strategy. This successfully overcomes the inherent "synchronous decision-making-synchronous conflict" flaw of the original IQL structure, effectively suppressing policy resonance and information interference between agents. This significantly reduces system-level risks such as path blockage, action disorder, and policy collapse, and improves the behavioral consistency and overall stability of collaborative control tasks in highly dynamic, multi-interference, and partially observable environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction to the drawings required for use in the embodiments will be given below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0042] Figure 1 This is a method flow chart of an AI intelligent collaborative control method based on the IQL algorithm of the present invention. DETAILED DESCRIPTION

[0043] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, these example embodiments are provided so that the description of this disclosure will be thorough and complete and will fully convey the concepts of the example embodiments to those skilled in the art.

[0044] The present invention provides Figure 1 An AI intelligent collaborative control method based on the IQL algorithm is shown, comprising the following steps:

[0045] In a multi-agent system, the local observation state sequence of the target agent in a continuous control cycle is obtained and divided into multiple time windows of equal length. In each time window, the observation feature variables are extracted, including the position change value, the distance value of the adjacent obstacle, the local density factor value, and the number of feasible paths, to form a basic observation vector group.

[0046] In a multi-agent system, the local observation state sequence of the target agent within a continuous control cycle is obtained and divided into multiple time windows of equal length. Within each time window, observation feature variables, including position change, distance to nearby obstacles, local density factor, and number of feasible paths, are extracted to form a basic observation vector set. This vector set primarily provides input data with temporal continuity, spatial structure, and behavioral sensitivity for subsequent state perturbation trend modeling and policy dynamic control. This step, through the introduction of a time window mechanism, segments the original unstructured continuous observation sequence into statistically meaningful time blocks, enabling accurate identification and quantification of local behavioral changes at the microscale. Furthermore, the feature extraction process not only captures the target agent's motion characteristics (e.g., position change) but also incorporates multi-level information from environmental factors (e.g., obstacle proximity), system structure (e.g., local density), and decision channels (e.g., number of paths). This enables the constructed basic observation vector to capture the system's local dynamic complexity. Standardized modeling of these observation vectors not only helps to improve the sensitivity of state disturbance perception, but also effectively enhances the robustness of the model in the face of multi-agent interference and environmental disturbances, providing an accurate and reliable data basis for subsequent disturbance trend calculation, state stability judgment and dynamic control strategy triggering, and ultimately ensuring the operational stability and behavioral coordination of the collaborative system in a dynamic, intensive and partially observable environment.

[0047] Obtaining the local observation state sequence of the target agent within a continuous control cycle and dividing it into multiple time windows of equal length includes the following steps:

[0048] Based on the target agent observation data stream collected during the continuous control cycle, time synchronization processing is performed at preset time intervals to align the observation data at each control moment, ensuring the integrity and consistency of the data sequence in the time dimension;

[0049] To achieve standardized acquisition of observation data for the target agent, a raw observation sequence consisting of multi-dimensional data such as sensor input, position state, and velocity vector must be recorded within each control cycle. By embedding a time synchronization module, the collected data is sorted and calibrated based on a unified timestamp, resolving timing misalignments caused by sensor latency and acquisition frequency variations. Time synchronization ensures that different types of observation data accurately reflect the system state at each time point, providing fundamental temporal consistency for subsequent time window partitioning and state trend analysis.

[0050] After completing the time division, the short-term trajectory trend estimation is performed on the observation data in each time window, and the position change value of the target agent in the time period, the distance value of the neighboring obstacles, the local density factor value, and the number of feasible paths are extracted. The distance value of the neighboring obstacles is calculated by analyzing the spatial distribution of obstacles in the local coordinate system of the target agent. The local density factor value is obtained by counting the number of neighboring agents and their distribution range in the time window. The number of feasible paths is determined by the number of orientation channels after matching the current motion direction.

[0051] After completing time synchronization and window division, feature extraction is performed on the data within each time window. The position change value is obtained by calculating the difference in the target agent's position coordinates at the start and end of the current window, reflecting its motion trend. The distance value to the neighboring obstacle is based on lidar, ultrasonic, or image perception data, extracting the average or minimum distance to the nearest obstacle in the agent's local coordinate system. The local density factor value is obtained by counting the number of neighboring agents within a specific radius within the window and combining it with distribution uniformity modeling. The number of feasible paths is determined by the environment map and the current orientation of the passable path search algorithm to determine the existence and availability of path channels. These features together constitute the observation vector that reflects the environment structure and local dynamic changes, and are the input core for subsequent disturbance modeling and policy control.

[0052] The observation feature variables extracted in each time window are normalized and combined into a basic observation vector group in a unified format, providing a standardized input basis for subsequent state disturbance trend modeling.

[0053] The position change values, distance values ​​to adjacent obstacles, local density factors, and number of feasible paths extracted within each time window are normalized to ensure that the numerical dimensions of each feature are comparable and to prevent interference with model learning due to a large range of values ​​in a particular dimension. Normalization methods can adopt minimum-maximum scaling, Z-score normalization, or empirical distribution mapping to adapt to the statistical distribution of features in different environments. The standardized feature values ​​are combined into a basic observation vector group of unified dimensions, which serves as the input unit for subsequent state perturbation trend calculations and provides the system with the ability to resist scale perturbations, thereby improving the robustness and generalization performance of the entire multi-agent collaborative system in dynamic environments.

[0054] Perform dimension-by-dimension difference calculation on the basic observation vectors in adjacent time windows to measure the trend of state disturbance changes in continuous time series. At the same time, apply the disturbance gain weight function to amplify the response intensity of the change value of the mutation feature dimension to generate the primary disturbance trend vector.

[0055] Dimension-by-dimensional interpolation is performed on the basic observation vectors within adjacent time windows to measure the trend of state perturbations in a continuous time series. A perturbation gain weighting function is applied to amplify the response strength of changes in the sudden feature dimensions to generate a primary perturbation trend vector. This step accurately captures the magnitude of state changes and their dynamic evolution characteristics within a local timeframe, providing a quantitative basis for subsequent assessment of system stability and triggering of dynamic control mechanisms. In multi-agent systems, small changes in local states can rapidly evolve into system-wide instability. This is particularly true in dense environments, where mutual interference between multiple agents can lead to frequent fluctuations in the observed state. Dimension-by-dimensional interpolation effectively quantifies the rate of change of each feature dimension (such as position change, obstacle distance, density change, and number of paths) between consecutive time periods, constructing a basic data representation of state perturbations. Furthermore, by introducing a perturbation gain weighting function for the sudden feature dimensions to amplify the response strength to unexpected changes, the system's sensitivity to nonlinear state changes is significantly enhanced, enabling the model to prioritize perturbations that have a significant impact on behavioral decisions and preventing key features from being masked by averaging. The final generated primary disturbance trend vector integrates the basic change trend and mutation response amplification information. As the core intermediate data of the disturbance analysis chain, it not only has dynamic responsiveness but also has high-resolution recognition capabilities. It can provide data support for the intelligent agent to achieve more forward-looking dynamic obstacle avoidance, path adjustment and collaborative strategy planning. It is a key link in realizing the dynamic robustness and environmental adaptability of the intelligent collaborative control system.

[0056] The specific steps for performing dimension-by-dimension difference calculation on the basic observation vectors in adjacent time windows to generate the primary perturbation trend vector are as follows:

[0057] For the basic observation vectors in adjacent time windows, the dimension-by-dimension difference calculation is performed in the order of the corresponding feature dimensions. The continuous time difference method is used to extract the change of each feature dimension in the time evolution process, and the basic state disturbance vector is established to reflect the local state fluctuation trend of the target agent in each continuous time period.

[0058] For the basic observation vector extracted within the continuous time window, a first-order difference operation is performed on the corresponding dimension data in the two previous and next time windows according to the arrangement order of the fixed feature dimensions. That is, the feature value of the current window is subtracted from the feature value of the same dimension in the previous window, thereby obtaining the instantaneous rate of change of each dimension during the time evolution process. This dimension-by-dimension difference calculation method can be implemented through a simple vector difference operation, with the advantages of high computational efficiency and real-time operation. The difference result constitutes the basic state disturbance vector, which reflects the fluctuation trend of each key state variable of the target agent within a specific time period. It is the first quantitative step in the perception of state variation in a dynamic environment. Its role is to provide an accurate basis for the subsequent determination of the nature and intensity of the disturbance.

[0059] To enhance the detection sensitivity of sudden disturbances, a mutation recognition threshold mechanism is constructed based on the historical disturbance statistical distribution model. When the difference in any characteristic dimension exceeds the corresponding disturbance anomaly judgment threshold, it is marked as a mutation characteristic dimension, and the disturbance gain weight function is called to assign a nonlinear response amplification coefficient according to the magnitude of the mutation, thereby amplifying the response intensity of the change value of the mutation dimension.

[0060] To identify sudden change feature dimensions that significantly impact system behavioral stability from the underlying disturbance vector, a mutation identification threshold mechanism is required. Based on a historical disturbance statistical model, this mechanism establishes the mean and standard deviation distribution of disturbance differences for each feature dimension and sets a dynamic threshold to determine whether the current difference significantly deviates from the normal range. When the difference in a dimension exceeds the threshold for that dimension, it is identified as a sudden change feature dimension. For these dimensions, a predefined disturbance gain weighting function (such as an exponential amplification function or a sigmoid amplification curve) is applied to automatically assign nonlinear amplification coefficients based on the degree of sudden change, thereby increasing the response weight of these key disturbances in subsequent analysis. This step effectively improves the sensitivity of detecting abnormal state changes, enabling the model to more accurately identify potentially risky behaviors.

[0061] The original difference results of the mutation feature dimension that has completed response amplification processing and the non-mutation feature dimension are fused to form a complete primary disturbance trend vector. This vector not only retains the basic disturbance information, but also significantly enhances the key mutation signal, providing more sensitive and differentiated data input for subsequent disturbance fluctuation detection and state stability assessment.

[0062] The difference between the mutated feature dimensions, which have undergone response intensity amplification, is fused dimensionally with the original difference between the remaining feature dimensions that have not undergone mutations to construct a complete primary disturbance trend vector. This process requires maintaining the consistency of the vector dimensions and strict alignment of their order to ensure that each dimension is accurately restored. The fused primary disturbance trend vector retains the original information of the underlying state changes while enhancing the expressiveness of high-risk state characteristics, providing excellent abnormal response identification capabilities. This vector serves as direct input for subsequent disturbance offset modeling, state stability assessment, and control strategy triggering, laying the data foundation for refined environmental perception and dynamic action decision-making.

[0063] The primary disturbance trend vector is normalized and compressed, and a reference disturbance baseline model is constructed using historical stable state samples. The disturbance offset distance index is calculated based on the Euclidean distance between the current primary disturbance trend vector and the reference disturbance baseline model to characterize the relative magnitude of the current state disturbance.

[0064] The primary disturbance trend vector is normalized and compressed using normal deviations. A reference disturbance baseline model is constructed using historical stable state samples. The disturbance offset distance index is calculated based on the Euclidean distance between the current primary disturbance trend vector and the reference disturbance baseline model. Its core purpose is to establish a mechanism for assessing the stability of multidimensional disturbance states and quantifying the degree of deviation, thereby enabling accurate perception of the magnitude of changes in the current local environment of the multi-agent system and building a foundation for dynamic response. In multi-agent collaborative control systems, the local state changes of agents are often nonlinear, abrupt, and multidimensionally interactive. Relying solely on the original disturbance trend vector makes it difficult to determine whether the current state constitutes an anomaly. Therefore, standardization is used to normalize data from different feature dimensions, eliminating weight interference caused by different scales and making the features comparable. Furthermore, normal deviation compression is performed, using the statistical characteristics of historical stable states as a reference, to enhance the system's sensitivity to deviations from the central trend, making the model more easily able to identify significant abnormal disturbances. On this basis, a reference perturbation baseline model is introduced as a statistical mapping of the system's "normal state perturbation behavior." By calculating the Euclidean distance from the current state, a perturbation offset distance index is generated, achieving a single-value quantification of the degree of deviation between the current state and the normal state. This index not only reveals the severity of state changes in the current local observation space but also serves as the core judgment basis for triggering control strategy adjustments. This is of great significance for ensuring that the system maintains behavioral coordination and strategy robustness in dynamic and non-stationary environments.

[0065] The specific steps for performing normalization and normal deviation compression on the primary disturbance trend vector and calculating the disturbance offset distance index are as follows:

[0066] For each characteristic dimension of the primary perturbation trend vector, the maximum and minimum normalization processing is performed according to the distribution range of the dimension in the historical observation data, and all dimensions are mapped to a unified numerical range to ensure the dimensional consistency and weight balance between the characteristic dimensions in subsequent calculations, and prevent the model bias caused by inconsistent characteristic scales;

[0067] Each feature dimension in the primary perturbation trend vector is normalized using the maximum-minimum normalization method. This method uses the minimum and maximum values ​​of that dimension in the historical steady-state data as boundaries to linearly map the current feature value to the interval [0, 1]. This processing step not only unifies the numerical scales across different feature dimensions, preventing a single dimension from dominating the overall perturbation assessment due to its large numerical range, but also improves the ability of subsequent algorithms to discern subtle changes. Normalization is crucial for the unified modeling of multidimensional feature vectors. Its core role is to lay a foundation for numerical consistency in subsequent deviation compression and distance measurement, ensuring that each dimension has equivalent weight in the calculation and preventing judgment bias caused by feature imbalance.

[0068] After standardization, a normal deviation compression operation is performed on each dimension of the normalized vector. Based on the mean and standard deviation obtained from historical stable state sample statistics, the Z-score standard deviation of the current eigenvalue is calculated. The degree of deviation is then mapped to a compressed disturbance intensity value using an exponential decay function to attenuate minor deviations and amplify significant deviations, thereby highlighting structural abnormalities.

[0069] After completing the maximum and minimum normalization, a normal deviation compression operation is performed on each standardized feature dimension. This operation is based on the mean and standard deviation of the corresponding dimension in the historical steady-state samples. By calculating the Z-score of the current feature value (i.e., (current value - mean) / standard deviation), the deviation of the feature from the statistical distribution is obtained. To enhance the response to significant anomalies while preventing the system from overreacting to minor perturbations, an exponential decay function (e.g., exp(-|Z|)) can be used to convert the Z-score into a compressed perturbation value. This compresses normal fluctuations while retaining and amplifying extreme deviations. The key benefit of this process is that it can highlight the potential threats of nonlinear changes to the system when constructing perturbation trends, while filtering out low-impact perturbations and enhancing the model's ability to identify anomalies in the actual state.

[0070] Taking the historical stable state sample set as the reference disturbance baseline model, the Euclidean distance between the current compressed disturbance trend vector and the reference disturbance baseline is calculated to form a single-value disturbance offset distance index. This index is used to measure the overall deviation of the current observation state from the stable state of the system, and is an important basis for subsequent judgment of state stability and triggering of control mechanisms.

[0071] The Euclidean distance between the compressed disturbance trend vector and a reference disturbance baseline model constructed from historical stable state samples is calculated. This reference model can be constructed by averaging multiple sets of typical stable state disturbance vectors and serves as a concentrated expression of the disturbance pattern under the system's "normal" state. By calculating the Euclidean distance between the current disturbance vector and the reference baseline, a single-value disturbance offset distance index can be derived. The larger the index, the more serious the deviation of the current state from the stable state, and vice versa. The core function of this index is to provide a quantifiable and comparable state fluctuation measurement result, which is used to determine whether the current system has entered an unstable operating range and serve as a criterion for triggering dynamic control strategies (such as control cycle adjustment, behavior delay insertion, etc.), thereby improving the response accuracy and operational robustness of the entire multi-agent system in dynamic and complex environments.

[0072] The disturbance offset distance indicator is fused with the local spatial dynamic entropy factor within the target agent's current location area. The local spatial dynamic entropy factor is constructed by evaluating the rate of change of the number of neighboring agents and the amplitude of change of the neighboring agent's movement direction. The fusion result outputs a local disturbance response value with a single dimension.

[0073] The disturbance offset distance metric is combined with the local spatial dynamic entropy factor within the target agent's current location to output a single-dimensional local disturbance response value. Its core function is to achieve a joint quantitative assessment of the target agent's current local state stability and spatial behavioral complexity, providing a more comprehensive, accurate, and real-time risk assessment basis for dynamic collaborative decision-making in multi-agent systems. In a multi-agent collaborative control environment, state disturbances often arise not only from abnormal fluctuations in the agent's own state but are also closely related to behavioral changes in neighboring agents. While relying solely on the disturbance offset distance metric can reflect the degree of deviation of the current state from a stable baseline, it is insufficient to capture the dynamic uncertainty of the agent's environment. By introducing the local spatial dynamic entropy factor, a quantitative representation of behavioral diversity and uncertainty is established based on the rate of change in the number of neighboring agents and the magnitude of changes in their movement directions. This allows the system to perceive potential issues such as significant behavioral conflicts, collaborative interference, or confusion of intent in the surrounding environment. By integrating these two metrics, a unified local perturbation response value can be generated that integrates state deviation and behavioral disorder. This allows the model to not only focus on individual state fluctuations but also consider environmental complexity and collaborative uncertainty, thereby more accurately identifying the triggering conditions of high-risk states. This response value serves as a key input for the system to trigger dynamic control strategies (such as control cycle adjustment, behavioral priority adjustment, or path avoidance), helping to improve the decision-making robustness and responsiveness of the entire multi-agent system in dynamic, incompletely observable, and multi-source interference environments.

[0074] The specific steps for fusing the disturbance offset distance indicator with the local spatial dynamic entropy factor within the target agent's current location area to output the local disturbance response value are as follows:

[0075] A local observation grid is constructed based on the current location of the target agent. The rate of change of the number of neighboring agents is counted in real time within the current area. The rate of change of the number of neighboring agents is calculated using the trend of the change in the number of neighboring agents within a continuous time window to reflect the aggregation or sparseness trend of the neighbors in the local space.

[0076] In a multi-agent system, to capture the local dynamic environmental state of a target agent at its current location, a local observation grid with a fixed or adaptive radius centered on its current coordinates must first be constructed. This grid can be divided into two-dimensional polar coordinates or equidistant squares to define the analysis scope of the current local space. Subsequently, within each control cycle, the perception module detects and counts the number of neighboring agents within the local observation grid and records the change in their number within a continuous time window to form a time series. By performing first-order differences or sliding average derivatives on this time series, the rate of change of the number of neighboring agents within the continuous control cycle is obtained. The rate of change per unit time is further calculated to characterize whether neighbors are rapidly converging around the target agent or sparsely diffusing from the target area. This step aims to identify the density dynamics of group behavior in the local space. Specifically, aggregation trends or sparse trends reflect potential increases in collaborative behavior or environmental interference. This provides a density perturbation input foundation for the subsequent construction of the dynamic entropy factor, thereby enhancing the system's real-time perception of the evolving local behavior patterns in complex behavioral scenarios.

[0077] For all neighboring agents within the same local space, the action direction vectors in continuous time segments are extracted, and their angle change rate and direction fluctuation amplitude are calculated. The dynamic entropy factor of the local space is constructed in combination with the change rate of the number of neighbors. The dynamic entropy factor reflects the degree of chaos and uncertainty of the distribution of local spatial behavior through the information entropy calculation model.

[0078] To further enhance understanding of the behavioral dynamics of neighboring agents, the system extracts the action direction vector of each neighboring agent within a continuous time window and calculates the angular amplitude of its directional changes. By constructing a direction sequence for each agent and calculating the change in direction angle between adjacent moments, the fluctuation rate of its action direction can be derived. After integrating the directional fluctuation data of all neighbors with the rate of change of their number, the information entropy model is used to model the chaotic degree of behavioral changes in the local space, resulting in a local space dynamic entropy factor. The higher the entropy factor, the more uncertain and chaotic the neighbor's behavior. The key role of this step is to provide a precise numerical basis for judging the dynamic complexity of the current local spatial environment. It is a core parameter for evaluating the safety and behavioral predictability of the local area.

[0079] The disturbance offset distance index and the local spatial dynamic entropy factor are input into the fusion function model, and a weighted normalized fusion algorithm is used to generate a local disturbance response value with a single dimension. The fusion output is used to comprehensively characterize the local state disturbance intensity and spatial behavior complexity, which is an important decision-making indicator for determining whether the current behavior environment of the target intelligent agent is in a high-risk situation.

[0080] The calculated disturbance offset distance indicator and the local spatial dynamic entropy factor are used as input and processed using a weighted normalization fusion function to generate a local disturbance response value with a unified dimension. The fusion function can adopt linear weighting or a joint valuation function based on a Bayesian model. During the fusion process, weight coefficients are assigned based on experimental data or environmental importance, so that the output value reflects both the intensity of the local state disturbance and the complexity of the spatial behavior. The fusion result is a single numerical value, which is easy to compare with the preset risk threshold and serves as a criterion for whether the intelligent agent enters a high-risk decision-making state. The generation of this value not only enhances the multidimensional fusion expression capabilities of environmental modeling but also provides a more robust, sensitive, and real-time input basis for triggering subsequent behavior control strategies.

[0081] A time sliding window is constructed for multiple continuous local disturbance response values, and extreme value capture and change rate weighted averaging operations are performed within the time sliding window to generate a local state disturbance index. The local state disturbance index is used to judge the stability level of the current local observation space.

[0082] A sliding time window is constructed for multiple continuous local disturbance response values. Within this window, extreme value capture and rate-of-change weighted averaging are performed to generate a local state disturbance index. Its core purpose is to integrate the temporal trends of local disturbances to form an indicator system for dynamically quantifying the stability of the local observation space. In multi-agent collaborative control systems, local disturbances often occur in a continuous and fluctuating manner. The disturbance value at a single point in time cannot fully reflect the system's true state. By constructing a sliding time window and organizing the disturbance response values ​​within a certain time range into a time series, the dynamic variation of the disturbance can be preserved. Furthermore, extracting extreme values ​​within the window, such as the maximum, minimum, and average values, helps characterize the overall amplitude range and central trend of the disturbance, reflecting the activity of the local state in the short term. Furthermore, by calculating the rate of change of the response value and introducing a weighting mechanism, the contribution of periods of high disturbance fluctuation to the overall judgment is emphasized, thereby enhancing sensitivity to abnormal disturbances. Ultimately, the local state disturbance index, generated by integrating this information using a weighted average, not only maintains temporal continuity and real-time updating capabilities but also unifies various disturbance indicators into a single numerical expression, facilitating comparisons with preset thresholds and enabling automated stability assessment. This index, a core state perception parameter in the agent's decision-making engine, can be used to trigger dynamic response mechanisms such as control frequency adjustment, path avoidance, and behavior suppression. It serves as a key basis for improving the system's safety, adaptability, and robustness in complex dynamic environments.

[0083] The specific steps of constructing a time sliding window to process multiple continuous local disturbance response values ​​to generate a local state disturbance index are as follows:

[0084] Within a preset time scale, a time sliding window covering the continuous control cycle of the target agent is constructed with a fixed length and sliding step size. The corresponding local disturbance response value sequence within each time window is organized into structured time series data to ensure the temporal integrity and continuous expression of continuous disturbance information.

[0085] Within the continuous control cycle of the target agent, a time sliding window structure is constructed based on a set time length (e.g., 1 second) and a sliding step size (e.g., 0.2 seconds). This structure can cover the sequence of local disturbance response values ​​generated by multiple adjacent control cycles. During each sliding process, the system extracts the set of disturbance response values ​​within the current window and organizes them into structured time series input data in chronological order. The purpose of constructing a time sliding window is to transform the originally discrete disturbance values ​​into a data structure with continuous time semantics, enabling the system to analyze disturbance trends over longer time spans, avoiding misjudgments caused by single-point anomalies, and improving the robustness and temporal integrity of state analysis.

[0086] An extreme value capture operation is performed in each time sliding window to extract the maximum, minimum, and average values ​​of the local disturbance response values ​​within the window. The disturbance change rate is calculated for adjacent response values ​​at each time step. According to the fluctuation amplitude of the change rate, corresponding dynamic weighting factors are assigned to different time points to enhance the system's response sensitivity to severe disturbance stages.

[0087] Within each constructed sliding window, the system needs to perform extreme value capture and change rate analysis. Extreme value capture includes extracting the maximum, minimum, and average values ​​within the window to determine the fluctuation range and central trend of the local disturbance response value; while the disturbance change rate analysis reflects the disturbance speed and change intensity by calculating the difference in disturbance response values ​​between adjacent time steps. In order to highlight the impact of abnormal disturbances, the system assigns a higher dynamic weighting factor to time points with a larger rate of change, so that the stage of drastic changes occupies a higher weight in subsequent fusion calculations. The role of this step is to extract key disturbance features from continuous disturbance data, improve the system's ability to identify sudden unstable states, and enhance the system's risk judgment accuracy in dynamically changing scenarios.

[0088] All disturbance response values ​​and change rate weighted values ​​in the window are fused according to the exponentially weighted moving average method to output a local state disturbance index with continuous values, dynamic updates and uniform scale. This index is used to quantitatively judge the stability level of the current local observation space and is an important criterion for subsequently triggering dynamic control strategies and switching intelligent agent execution modes.

[0089] The sequence of disturbance response values ​​and the weighted rate of change within each time sliding window are input into an exponentially weighted moving average function, and a dynamic attenuation mechanism is used to calculate the final local state disturbance index. This exponential weighting approach assigns higher weights to recent disturbances, thereby enhancing the model's response to current environmental changes, while retaining historical disturbance information to reduce judgment noise. The output local state disturbance index is a single continuous value with dynamic update capabilities and a uniform scale, facilitating real-time comparison with a set threshold. This index serves as the core basis for determining whether the current local observation space is unstable. It can be used to trigger behavioral adjustments, control strategy switching, or path replanning of the intelligent agent, and is a key control variable in the dynamic adaptation mechanism of the entire system.

[0090] When the local state disturbance index exceeds the preset reference threshold, the space-time decoupling control mechanism is triggered to dynamically adjust the action execution cycle of the target agent. A slight time delay difference in the control cycle is introduced between multiple agents to form an asynchronous staggered action execution window, suppressing the synchronous disturbance response between agents, reducing the probability of local observation space collapse, and enhancing the execution stability and dynamic environment adaptability of the multi-agent collaborative control strategy.

[0091] When the local state disturbance index is greater than the preset reference threshold, the space-time decoupling control mechanism is triggered. According to the degree of deviation between the local state disturbance index and the reference threshold, the normalized deviation amplitude factor is defined. The calculation expression is as follows:

[0092] , where It is the local state disturbance index, which is used to reflect the overall stability level of the local observation space where the agent is located at the current moment. The larger the value, the more unstable the environment and the more severe the disturbance. is the reference threshold of the local state disturbance index, Is a logarithmic function, which is the ratio of the disturbance The growth of is smoothly compressed to avoid the explosive rise of the control parameter when the maximum value is reached, while ensuring that a positive response can be generated even if the value exceeds the limit slightly. The addition of 1 is to ensure that the function value is 0 when the disturbance value is 0, which has mathematical stability and continuity. is the normalized deviation amplitude factor, which represents the deviation intensity of the current local state disturbance index relative to the system reference threshold. It is converted into a continuous and smooth control coefficient after logarithmic compression. It is a dimensionless positive real number used to control the amplification ratio of the agent's time delay factor.

[0093] Normalized deviation amplitude factor It is used to logarithmically compress and amplify the disturbance amplitude, which can not only increase the sensitivity to high-amplitude disturbances, but also avoid unnecessary strategy changes caused by low-amplitude disturbances.

[0094] For each target agent participating in the current local cooperative control, the corresponding slight time delay adjustment factor is calculated. The calculation expression is as follows:

[0095] , where It is the delay control reference amplitude constant, which is the maximum time delay upper limit set by the system and controls the overall adjustment amplitude in seconds. Its function is to limit the maximum peak offset range. Usually, it takes a small value (such as 0.05-0.5 seconds) to ensure delicate adjustment and soft rhythm without causing system rhythm collapse. is the agent's number, is the total number of agents, is the mathematical constant, pi, It is The nonlinear distribution weights of the agents are numbered Map to The sine function value within the interval is used to control the differential distribution of delay intensity between different agents. It is an intelligent agent At the moment The action execution time delay adjustment factor is in seconds, indicating that in the current control cycle, the system is The slight action delay introduced by each agent directly determines the action trigger point of the agent on the time axis and plays a decoupling role;

[0096] The introduction of a sinusoidal function ensures that all agents exhibit a nonlinear, staggered distribution of their action execution times, avoiding repeated interference caused by linear synchronization adjustments. This step leverages the degree of disturbance to fine-tune the timing of the agents' actions, establishing a dynamic execution foundation with asynchronous control characteristics.

[0097] After obtaining the action execution time delay adjustment factor for each target agent Afterwards, the action execution cycle is dynamically modified to calculate the actual action execution cycle of each agent at the current moment. The calculation expression is as follows:

[0098] , where It is An agent at time The actual action execution cycle of each agent is dynamically adjusted to make its execution cycle staggered due to the degree of disturbance rather than synchronization, thus forming an interleaved control structure in the time domain. It is the basic control period, which is the standard action execution period shared by all agents when the disturbance level is stable.

[0099] By superimposing different delay factors onto a unified basic control cycle, the execution time points of each agent are actively stretched apart, constructing an interleaved asynchronous execution window in the time domain, thereby suppressing the behavioral pattern of multiple agents making synchronous responses in the same state.

[0100] This mechanism can significantly reduce the local observation space collapse caused by synchronization disturbances, and improve the operational stability, response robustness and effectiveness of the coordination strategy of the entire system in dynamic and dense environments.

[0101] The AI-powered collaborative control method based on the IQL algorithm effectively mitigates the instantaneous collapse of the observation space caused by drastic fluctuations in local observation information, enhancing the stable perception and robustness of multi-agent systems in dynamic obstacle-intensive scenarios. By incorporating disturbance trend modeling, dynamic disturbance response evaluation, and a sliding window aggregation mechanism, this method constructs a spatiotemporally continuous disturbance index system, achieving a joint quantitative representation of the intensity of local state changes and the complexity of environmental behavior, thereby improving the system's perception resolution of critical disturbance events. Furthermore, combined with a disturbance-driven temporal decoupling control mechanism, it dynamically adjusts the agents' action execution rhythm and introduces a slight time staggering strategy. This successfully overcomes the inherent "synchronous decision-making-synchronous conflict" flaw of the original IQL structure, effectively suppressing policy resonance and information interference between agents. This significantly reduces system-level risks such as path blockage, action disorder, and policy collapse, and improves the behavioral consistency and overall stability of collaborative control tasks in highly dynamic, multi-disturbing, and partially observable environments.

[0102] The above formulas are all dimensionless and numerical calculations. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters in the formulas are set by technicians in this field according to actual conditions.

[0103] The above description is merely illustrative of certain exemplary embodiments of the present invention. It goes without saying that those skilled in the art will be able to modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the above drawings and description are illustrative in nature and should not be construed as limiting the scope of protection of the claims.

[0104] It should be noted that, in this document, if there are relational terms such as first and second, etc., they are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "comprises", "comprising" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device that includes a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprising a ..." does not exclude the presence of other identical elements in the process, method, article or device that includes the element.

[0105] It should be understood that in the various embodiments of the present application, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0106] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0107] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0108] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0109] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0110] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

[0111] The above description is merely illustrative of certain exemplary embodiments of the present invention. It goes without saying that those skilled in the art will be able to modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the above drawings and description are illustrative in nature and should not be construed as limiting the scope of protection of the claims.

Claims

1. An AI intelligent collaborative control method based on IQL algorithm, characterized in that: The following steps are involved: Obtain the local observation state sequence of the target agent in the continuous control cycle, and divide the local observation state sequence into multiple time windows of equal time length. Extract the observation feature variables in each time window to form a basic observation vector group; Perform dimension-by-dimension difference calculation on the basic observation vectors in adjacent time windows, and apply the perturbation gain weight function to amplify the response intensity of the change value of the mutation feature dimension to generate the primary perturbation trend vector; Perform normalization and normal deviation compression operations on the primary disturbance trend vector, build a reference disturbance baseline model using historical stable state samples, and calculate the disturbance offset distance index based on the Euclidean distance between the current primary disturbance trend vector and the reference disturbance baseline model; The disturbance offset distance indicator is fused with the local spatial dynamic entropy factor within the current position area of ​​the target agent, and the fusion result is output as a local disturbance response value with a single dimension; A time sliding window is constructed for multiple continuous local disturbance response values, and extreme value capture and change rate weighted averaging operations are performed within the time sliding window to generate a local state disturbance index; When the local state disturbance index is greater than the preset reference threshold, the space-time decoupling control mechanism is triggered to dynamically adjust the action execution cycle of the target agent, introduce a slight time delay difference in the control cycle between multiple agents, form an asynchronous interleaved action execution window, and suppress the synchronous disturbance response between agents.

2. The AI ​​intelligent collaborative control method based on the IQL algorithm according to claim 1 is characterized in that: Obtaining the local observation state sequence of the target agent within a continuous control cycle and dividing it into multiple time windows of equal length includes the following steps: Based on the target agent observation data stream collected during the continuous control cycle, time synchronization processing is performed at preset time intervals to align the observation data at each control moment, ensuring the integrity and consistency of the data sequence in the time dimension; After completing the time division, the short-term trajectory trend estimation is performed on the observation data in each time window, and the position change value, the distance value of the adjacent obstacles, the local density factor value and the number of feasible paths of the target agent in the time period are combined. The observation feature variables extracted in each time window are normalized and combined into a basic observation vector group in a unified format.

3. The AI ​​intelligent collaborative control method based on the IQL algorithm according to claim 1 is characterized in that: The specific steps for performing dimension-by-dimension difference calculation on the basic observation vectors in adjacent time windows to generate the primary perturbation trend vector are as follows: For the basic observation vectors in adjacent time windows, the dimension-by-dimension difference calculation is performed in the order of the corresponding feature dimensions to establish the basic state disturbance vector, which is used to reflect the local state fluctuation trend of the target agent in each continuous time period; To enhance the detection sensitivity of sudden disturbances, a mutation recognition threshold mechanism is constructed based on the historical disturbance statistical distribution model. When the difference in any characteristic dimension exceeds the corresponding disturbance anomaly judgment threshold, it is marked as a mutation characteristic dimension, and the disturbance gain weight function is called to assign a nonlinear response amplification coefficient according to the magnitude of the mutation, thereby amplifying the response intensity of the change value of the mutation dimension. The original difference results of the mutated feature dimension that has completed the response amplification processing and the unmutated feature dimension are fused to form a complete primary disturbance trend vector.

4. The AI ​​intelligent collaborative control method based on the IQL algorithm according to claim 1 is characterized in that: The specific steps for performing normalization and normal deviation compression on the primary disturbance trend vector and calculating the disturbance offset distance index are as follows: For each characteristic dimension of the primary disturbance trend vector, the maximum and minimum normalization processing is performed according to the distribution range of the dimension in the historical observation data, and all dimensions are mapped to a unified numerical range; After standardization, a normal deviation compression operation is performed on each dimension of the normalized vector. Based on the mean and standard deviation obtained from historical stable state sample statistics, the Z-score standard deviation of the current eigenvalue is calculated. The degree of deviation is then mapped to a compressed disturbance intensity value using an exponential decay function to attenuate minor deviations and amplify significant deviations, thereby highlighting structural abnormalities. Taking the historical stable state sample set as the reference disturbance baseline model, the Euclidean distance between the current compressed disturbance trend vector and the reference disturbance baseline is calculated to form a single-value disturbance offset distance indicator.

5. The AI ​​intelligent collaborative control method based on the IQL algorithm according to claim 1 is characterized in that: The specific steps for fusing the disturbance offset distance indicator with the local spatial dynamic entropy factor within the target agent's current location area to output the local disturbance response value are as follows: A local observation grid is constructed based on the current location of the target agent. The rate of change of the number of neighboring agents is counted in real time within the current area. The rate of change of the number of neighboring agents is calculated using the trend of the number of neighboring agents in a continuous time window. For all neighboring agents within the same local space, the action direction vectors in continuous time segments are extracted, their angle change rate and direction fluctuation amplitude are calculated, and the local space dynamic entropy factor is constructed in combination with the change rate of the number of neighbors. The disturbance offset distance index and the local spatial dynamic entropy factor are input into the fusion function model, and a weighted normalization fusion algorithm is used to generate a local disturbance response value with a single dimension.

6. The AI ​​intelligent collaborative control method based on the IQL algorithm according to claim 1 is characterized in that: The specific steps of constructing a time sliding window to process multiple continuous local disturbance response values ​​to generate a local state disturbance index are as follows: Within the preset time scale range, a time sliding window covering the continuous control cycle of the target agent is constructed with a fixed length and sliding step size, and the corresponding local disturbance response value sequence in each time window is organized into structured time series data; The extreme value capture operation is performed in each time sliding window to extract the maximum, minimum and average values ​​of the local disturbance response value in the window, and the disturbance change rate is calculated for the adjacent response values ​​at each time step. The corresponding dynamic weighting factors are assigned to different time points according to the fluctuation amplitude of the change rate. All disturbance response values ​​within the window are fused with the weighted values ​​of the rate of change according to the exponentially weighted moving average method to output the local state disturbance index.

7. The AI ​​intelligent collaborative control method based on the IQL algorithm according to claim 1 is characterized in that: When the local state disturbance index is greater than the preset reference threshold, the space-time decoupling control mechanism is triggered. According to the degree of deviation between the local state disturbance index and the reference threshold, the normalized deviation amplitude factor is defined. The calculation expression is as follows: , where is the local state perturbation index, is the reference threshold of the local state disturbance index, is a logarithmic function, is the normalized deviation amplitude factor; For each target agent participating in the current local cooperative control, the corresponding slight time delay adjustment factor is calculated. The calculation expression is as follows: , where is the delay control reference amplitude constant, is the agent's number, is the total number of agents, is the mathematical constant, pi, It is The nonlinear distribution weights of the agents, It is an intelligent agent At the moment Action execution time delay adjustment factor; After obtaining the action execution time delay adjustment factor for each target agent Afterwards, the action execution cycle is dynamically modified to calculate the actual action execution cycle of each agent at the current moment. The calculation expression is as follows: , where It is An agent at time The actual action execution cycle, is the basic control cycle.

Images