Storage AGV dynamic path planning system based on multi-objective optimization

Abstract

The invention discloses a storage AGV dynamic path planning system based on multi-objective optimization, and relates to the technical field of storage logistics, and the system comprises a multi-source sensing and data collection module which is used for collecting the operation state, operation environment and external traffic information of an AGV and generating a standardized feature vector; and the cross-modal digital twinning and risk simulation module is used for constructing a virtual twinning body of a warehouse and external traffic and carrying out risk prediction and simulation under the driving of cross-modal sensing data. According to the invention, through the multi-source sensing and data acquisition module, the system can comprehensively acquire the AGV operation state, the operation environment and the external traffic information, and through combination with an advanced data fusion technology, a high-dimensional standardized feature vector is generated, so that an accurate and comprehensive data basis is provided for subsequent path planning and risk prediction; the cross-modal digital twinning and risk simulation module constructs a virtual twinning body of a warehouse and external traffic, and can reflect the dynamic change of the physical world in real time.

Description

Technical Field

[0001] This invention relates to the field of warehousing and logistics technology, and in particular to a dynamic path planning system for warehousing AGVs based on multi-objective optimization. Background Technology

[0002] With the rapid development of e-commerce and intelligent manufacturing, warehousing and logistics systems are demanding increasingly higher levels of automation and efficiency. Automated Guided Vehicles (AGVs), as key equipment in warehouse automation, directly impact the operating costs, operational efficiency, and safety of the entire logistics system through their path planning and scheduling efficiency.

[0003] A published patent, "A Transportation Route Planning System for a Material Storage Center" (Publication No.: CN119204899A), relates to the field of transportation route planning technology, specifically to a transportation route planning system for a material storage center. The system includes an order generation module, an in-warehouse route analysis module, an out-of-warehouse route analysis module, an in-warehouse route planning module, an out-of-warehouse route planning module, a server, a data communication module, a processor, and a display terminal. This invention acquires outbound orders from the material storage center, extracts the storage location information and destination location information of the materials, and sends this information to the server for storage. It then performs real-time monitoring and analysis of the in-warehouse and out-of-warehouse route parameters of the target transport vehicles, generating a comprehensive evaluation coefficient for both routes. Based on this comprehensive evaluation coefficient, it dynamically plans the optimal in-warehouse and out-of-warehouse travel routes. This not only reduces ineffective vehicle travel within the warehouse but also optimizes external transportation routes in real time, effectively improving transportation efficiency and reducing logistics costs.

[0004] The aforementioned patents have the following shortcomings: Inside the warehouse, only passage value, congestion value, and turning value are used as parameters, failing to comprehensively consider factors such as dynamic obstacles, vehicle energy consumption, and path safety, and lacking a congestion prediction mechanism, resulting in insufficient path adaptability; outside the warehouse, parameter evaluation relies on short-term monitoring, lacks traffic flow evolution modeling, has a single approach to handling weather factors, and has limited accuracy in path selection; in terms of optimization methods, they generally rely on single-objective comprehensive coefficients with fixed weights, lacking multi-objective collaborative optimization and adaptive learning capabilities, making it difficult to cope with the dynamic needs of different scenarios; at the system level, the multi-vehicle collaboration and anomaly handling mechanisms are imperfect, lacking digital twin modeling and robust design, resulting in insufficient overall operational efficiency and safety. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing a dynamic path planning system for warehouse AGVs based on multi-objective optimization.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a warehouse AGV dynamic path planning system based on multi-objective optimization, comprising a multi-source perception and data acquisition module for collecting AGV operating status, working environment, and external traffic information and generating standardized feature vectors; a cross-modal digital twin and risk simulation module for constructing a virtual twin of the warehouse and external traffic, and performing risk prediction and simulation driven by cross-modal perception data; a multi-objective intelligent optimization module for generating candidate path schemes under multiple objectives such as transportation time, energy consumption, safety, equipment lifespan, and task priority; an AGV collaborative scheduling module for realizing dynamic collaborative scheduling of multiple AGVs through game theory and reinforcement learning mechanisms; a self-evolutionary learning and transfer module for improving the system's adaptability in different scenarios through learning and transfer mechanisms; an anomaly detection and robust protection module for detecting hardware failures, communication anomalies, path congestion, and emergencies, and executing graded response strategies; and a visualization and operation and maintenance support module for providing digital twin visualization, causal heatmaps, and natural language generation explanations.

[0007] As a further description of the above technical solution:

[0008] The multi-source sensing and data acquisition module collects data through lidar, vision, RFID, millimeter-wave radar and acoustic sensors, and combines adaptive Kalman filtering and Bayesian optimization to achieve data fusion and compensation.

[0009] As a further description of the above technical solution:

[0010] The cross-modal digital twin and risk simulation module utilizes dynamic graph convolutional networks for risk propagation modeling, combines Transformer-XL and spatiotemporal graph attention mechanisms to achieve multi-scale risk prediction, and uses generative adversarial networks for extreme condition simulation.

[0011] As a further description of the above technical solution:

[0012] The multi-objective intelligent optimization module uses hierarchical multi-objective reinforcement learning and evolutionary algorithms to generate Pareto optimal solution sets, and combines causal constraints and fuzzy multi-objective optimization to improve the rationality and robustness of the solutions.

[0013] As a further description of the above technical solution:

[0014] The AGV collaborative scheduling module uses Stackelberg game, coalition game and signaling game mechanisms to allocate paths and tasks, and introduces Shapley value allocation and cooperative transfer game in the scheduling to achieve fairness and congestion self-healing.

[0015] As a further description of the above technical solution:

[0016] The self-evolutionary learning and transfer module includes meta-learning, graph embedding transfer, few-shot learning, cross-scenario knowledge distillation, lifelong learning, and self-supervised comparative learning mechanisms.

[0017] As a further description of the above technical solution:

[0018] The anomaly detection and robust protection module detects anomalies by monitoring message latency, jitter, and packet loss rate when detecting communication delays. Under moderate anomalies, it performs task reallocation, and under severe anomalies, it triggers redundant AGV takeover.

[0019] As a further description of the above technical solution:

[0020] The visualization and operation and maintenance support module adopts a logical structure of current situation-cause-measure-effect when generating natural language explanations, and displays the causal explanation chain and risk heat map through an interactive interface.

[0021] The present invention has the following beneficial effects:

[0022] 1. In this invention, firstly, through the multi-source sensing and data acquisition module, the system can comprehensively acquire AGV operating status, working environment, and external traffic information. Combined with advanced data fusion technology, it generates high-dimensional standardized feature vectors, providing an accurate and comprehensive data foundation for subsequent path planning and risk prediction. The cross-modal digital twin and risk simulation module constructs a virtual twin of the warehouse and external traffic, which can reflect the dynamic changes of the physical world in real time. Through risk prediction and simulation driven by cross-modal sensing data, the system can identify potential risks in advance and evaluate different solutions in the virtual environment, significantly improving the system's ability to cope with emergencies. The multi-objective intelligent optimization module comprehensively optimizes multiple key objectives such as transportation time, energy consumption, safety, equipment lifespan, and task priority, generating a Pareto optimal solution set. This allows the system to flexibly select the path solution that best meets the current business objectives according to actual needs, avoiding the limitations of single-objective optimization and improving overall operational efficiency.

[0023] 2. In this invention, the AGV collaborative scheduling module introduces game theory and reinforcement learning mechanisms, which can effectively solve path conflicts and task allocation problems among multiple AGVs. Through intelligent scheduling strategies, it realizes dynamic collaboration of the AGV cluster, reduces congestion, improves overall transportation efficiency, and takes into account the fairness of task allocation. The self-evolutionary learning and transfer module enables the system to quickly learn and adapt in different scenarios, reducing manual intervention. The anomaly detection and robust protection module can promptly detect and handle hardware failures, communication anomalies, path congestion, and emergencies, and ensures stable operation of the system in complex and uncertain environments through hierarchical response strategies, significantly improving the system's robustness. The visualization and operation and maintenance support module provides operation and maintenance personnel with intuitive and comprehensive system operation status and decision-making basis through digital twin visualization, causal heatmaps, and natural language generation explanations. In particular, the natural language explanation function can present complex optimization decision-making processes and anomaly causes in an easy-to-understand way, greatly reducing the difficulty of operation and maintenance and improving decision-making efficiency. Attached Figure Description

[0024] Figure 1 This is a system architecture diagram of the present invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] Reference Figure 1 This invention provides an embodiment of a warehouse AGV dynamic path planning system based on multi-objective optimization, comprising: a multi-source perception and data acquisition module for collecting AGV operating status, working environment, and external traffic information and generating standardized feature vectors; a cross-modal digital twin and risk simulation module for constructing a virtual twin of the warehouse and external traffic, and performing risk prediction and simulation driven by cross-modal perception data; a multi-objective intelligent optimization module for generating candidate path schemes under multiple objectives such as transportation time, energy consumption, safety, equipment lifespan, and task priority; an AGV collaborative scheduling module for achieving dynamic collaborative scheduling of multiple AGVs through game theory and reinforcement learning mechanisms; a self-evolutionary learning and transfer module for improving the system's adaptability in different scenarios through learning and transfer mechanisms; an anomaly detection and robust protection module for detecting hardware failures, communication anomalies, path congestion, and emergencies, and executing graded response strategies; and a visualization and operation and maintenance support module for providing digital twin visualization, causal heatmaps, and natural language generation explanations.

[0027] The multi-source sensing and data acquisition module collects data through LiDAR, vision, RFID, millimeter-wave radar, and acoustic sensors, and combines adaptive Kalman filtering and Bayesian optimization to achieve data fusion and compensation. The cross-modal digital twin and risk simulation module utilizes dynamic graph convolutional networks for risk propagation modeling, combines Transformer-XL and spatiotemporal graph attention mechanisms to achieve multi-scale risk prediction, and uses generative adversarial networks for extreme condition simulation. The multi-objective intelligent optimization module employs hierarchical multi-objective reinforcement learning and evolutionary algorithms to generate Pareto optimal solution sets, and combines causal constraints and fuzzy multi-objective optimization to improve the rationality and robustness of the solutions. The AGV collaborative scheduling module uses Stackelberg game, coalition game, and signaling game mechanisms for path and task allocation, and introduces Shapley value allocation and cooperative transfer game in scheduling to achieve fairness and congestion self-healing. The self-evolutionary learning and transfer module includes meta-learning, graph embedding transfer, few-shot learning, cross-scene knowledge distillation, lifelong learning, and self-supervised comparative learning mechanisms. The anomaly detection and robust protection module identifies anomalies by monitoring message latency, jitter, and packet loss rate when detecting communication delays. Under moderate anomalies, it performs task reallocation, and under severe anomalies, it triggers redundant AGV takeover. The visualization and operation and maintenance support module uses a logical structure of current situation—cause—measure—effect when generating natural language explanations, and displays the causal explanation chain and risk heatmap through an interactive interface.

[0028] The multi-source sensing and data acquisition module is used to comprehensively sense and accurately collect the operating status, working environment and external traffic factors of AGVs in the warehouse environment. It supplements information through various types of sensors and achieves data consistency and standardization by means of fusion and compensation algorithms, thereby providing stable input for downstream digital twin modeling and multi-objective optimization. It not only ensures the real-time nature of sensing information, but also achieves robustness under complex working conditions through redundant design and algorithm enhancement.

[0029] Data acquisition is conducted using multimodal sensors: LiDAR is used to acquire 3D point cloud data of the warehouse interior space, accurately depicting shelf locations, aisle contours, and the distribution of static obstacles; visual sensors are used to acquire warehouse image data, and dynamic obstacles, including workers, forklifts, and other mobile devices, are identified through image segmentation and target recognition; RFID tags are used to quickly calibrate material locations, establishing a correspondence between goods and their locations; millimeter-wave radar is used to measure relative speed and distance, enhancing the detection capability in low-light, high-reflection areas; acoustic sensors are used to acquire echo information, supplementing detection in areas where LiDAR and vision have blind spots or are obstructed; edge environment sensors collect real-time data on warehouse temperature, humidity, and dust concentration to provide environmental input for subsequent path safety and energy consumption assessments; and external road traffic flow, weather conditions, and emergency information are collected simultaneously to ensure that path planning takes into account the overall operating environment inside and outside the warehouse.

[0030] The collected data undergoes fusion and calibration: Adaptive Kalman filtering is used to fuse data from LiDAR, vision, millimeter-wave radar, and acoustic sensors to effectively suppress noise interference and maintain the smoothness of sensor data in dynamic environments. Since the noise characteristics of each sensor differ, the weights of observation noise and process noise need to be dynamically adjusted using adaptive Kalman filtering to achieve optimal fusion. The steps are: State modeling: Define the motion state vector of the AGV: x t =[p x ,p y ,v x ,v y ] T p x ,p y : The position coordinates of the AGV in the warehouse coordinate system; v x ,v y : The velocity components of the AGV in the x and y directions. Prediction process: x t =Fx t-1 +w t F: State transition matrix (defined according to the kinematic model); w t Process noise, with a mean of 0 and a covariance of Q. t The Gaussian distribution of z. Observation process: z t =Hx t +v t , z t H: Observation vector (position and velocity measured by multiple sensors); H: Observation matrix (describing the mapping relationship from state to observation); v t Observational noise follows a pattern with a mean of 0 and a covariance of R. tThe noise follows a Gaussian distribution. Adaptive adjustment: Traditional Kalman filtering assumes Q and R are fixed, but actual noise varies with the environment (e.g., lighting, obstacles). Online update of observation noise: α: Forgetting factor (0-1), controls historical weights; The predicted state vector; Prediction residuals (the difference between observations and predictions). Fusion output: Multi-sensor data according to the updated R... t Weighted fusion; to obtain the optimal state estimate This involves fusing the AGV's position and speed; using Bayesian optimization to compensate for deviations from various sensors during long-term operation, ensuring stable and accurate output; establishing a cross-sensor clock synchronization and spatial alignment mechanism to ensure consistency of data from different sources in both time and space; Bayesian optimization utilizes a Gaussian process (GP) to model the error function and iteratively selects the optimal parameters for online correction. Implementation steps: Deviation modeling: z i : The observation value of sensor i; True value; δ i : Systematic bias of the sensor; ∈ i Random noise. Objective function: f(δ) = ||z|| i -z ref ||, z ref : High-precision reference sensor or fused value; f(δ): The difference between the sensor observation and the reference value. Gaussian process modeling: f(δ)~GP(μ(δ),k(δ,δ) ′ μ(δ): mean function; k(δ,δ ′ Kernel function: Characterizes the correlation of errors under different biases. Acquisition function selection: Expected improvement: EI(δ) = E[max(0, f...] best -f(δ))],f best The current optimal observation error; a larger EI indicates that the bias value is more likely to improve performance. Iterative optimization: continuously select the optimal δ, update the sensor bias parameters; output the corrected observation values.

[0031] Anomaly detection and redundancy correction for sensor data: For sensor output anomalies, such as data delay, drift, or failure, the system can quickly detect them by comparing the outputs of redundant sensors. When the primary sensor source malfunctions, a backup sensor source takes over the corresponding data channel to avoid data loss due to single-point failure. Sudden anomalies are identified through continuous detection and statistical feature comparison, and a correction mechanism ensures the integrity of the feature sequence. When a sensor fails or outputs anomalies, compensation is provided by other redundant sensor sources. Implementation steps: Anomaly detection: If the observed residual... If the threshold θ is exceeded, sensor i is deemed abnormal. Redundancy replacement: If the LiDAR fails, a camera + millimeter-wave radar is used for spatial location estimation; if the camera fails due to insufficient lighting, the LiDAR point cloud is used as a replacement. Weight adjustment: The contribution of each sensor source is controlled by the confidence factor γi: z fused =∑ i γ i z i ,∑ i γ i =1,z fused : The merged observations; γ i The weight of sensor i is dynamically adjusted based on its health status.

[0032] All collected information is uniformly encoded into standardized feature vectors, which include location features, velocity features, energy consumption status features, environmental features, and external traffic features. Normalization and robustness enhancement processing are performed on the feature vectors to ensure that subsequent cross-modal digital twin and optimization modules can directly call the module's output. The output standardized feature vectors serve as a unified input data interface for the entire system, ensuring data compatibility and scalability between different modules. Data collected by different sensors has different formats and must be converted into a unified feature vector before it can be used as input for the optimization and learning modules. Conversion steps: Spatial parameters: The LiDAR point cloud is processed through voxel meshing and extracted into sparse coordinate features [x, y, z]. Motion parameters: Velocity information obtained from camera optical flow estimation is converted into [v...]. x ,v y Energy consumption parameters: Battery state is converted to SOC (State of Charge) and SOH (State of Health), and normalized to [0,1]. Environmental parameters: Temperature, humidity, and dust concentration are standardized to form an environmental vector. Unified splicing: The final feature vector is formed: φ = [p x ,p y ,v x ,v y ,SOC,SOH,Temp,Humid,Dust,…],p x ,p y : location; v x ,v y : Speed; SOC: Battery State of Charge; SOH: Battery Health; Temp: Temperature; Humid: Humidity; Dust: Dust Concentration. Vectors serve as a unified input, ensuring that downstream path optimization, twin modeling, and reinforcement learning can all process data within the same dimensional space.

[0033] During system operation, the multi-source sensing and data acquisition module works continuously and updates the sensing results in real time. Its output robust standardized feature vector provides a reliable foundation for downstream cross-modal digital twin and risk simulation, multi-objective intelligent optimization and AGV collaborative scheduling. Through the combination of sensor diversification and redundancy mechanisms, the multi-source sensing and data acquisition module achieves data integrity and high-precision sensing in complex warehousing scenarios, ensuring that AGV path planning has real-time performance, stability and robustness.

[0034] The cross-modal digital twin and risk simulation module constructs a virtual twin of the warehouse and couples it with a twin of the external traffic environment to achieve integrated risk simulation inside and outside the warehouse. It uses cross-modal sensor input as the driving source and combines dynamic graph convolutional networks, spatiotemporal prediction models and generative adversarial mechanisms to perform multi-scale prediction and simulation verification of potential congestion, collisions and extreme events, thereby providing a forward-looking decision-making basis for path planning and task scheduling.

[0035] Constructing a Virtual Twin: A virtual twin of the warehouse is constructed, including the warehouse road topology, the 3D layout of the shelving, and the AGV dynamics model. A warehouse road topology map is established: warehouse aisles and intersections are extracted as directed graphs, where nodes represent intersections or workstations, edges represent passable passageways, and edge weights include length, speed limit, width, slope, and safety factor. Formula: This is a map showing the passageway; Let ε be the set of nodes; let ε be the set of edges. For a weighted adjacency matrix, element A ij This represents the channel weight from node i to node j. The weight can be obtained by combining a function of length, speed limit, width, slope, and safety factor. A 3D model of the shelving is established: the 3D bounding box of the shelving is reconstructed based on point cloud and vision, and the coordinates of the storage locations, shelf height, and restricted buffer zones are recorded. A unified coordinate formula (using sensor points p) is used. s (Go to the global repository system): p G p is the global coordinate point; s The sensor coordinates; It is a rotation matrix; A dynamic model of the AGV is established using the translation vector: a discrete single-track model is adopted to represent the evolution of pose and velocity for path simulation and energy consumption assessment. State update formula: x t+1 =f(x) t ,u t ,η t ), x t =[p x ,p y ,ψ,v] T u t =[a,ω] T x tLet p be the state vector at time t, containing position p. x ,p y , heading angle ψ, velocity v; u t The control inputs include longitudinal acceleration *a* and angular velocity *ω*; η t This is the disturbance term, reflecting uncertainties such as friction, gradient, and load changes. Energy consumption estimation formula (travel segment e): E e This represents the energy consumption of this segment; P drv (v,a) represents the driving power, which is a function of velocity v and acceleration a; P aux [t] is the auxiliary power constant; e ,t e +Δt] represents the time interval; a virtual twin of external traffic is constructed, including the road network, traffic flow dynamic evolution model, and traffic signal control parameters. The external traffic twin is established by abstracting the park and surrounding roads into a road network map, incorporating a basic traffic map and weather influences to estimate capacity and travel time. Speed-density relationship (Greenshields model): v(k) represents the traffic flow velocity; k represents the density; V f k represents the free-flow velocity. jam This refers to congestion density. Capacity is affected by weather conditions. Let C0 be the effective external traffic capacity at time t; C0 be the baseline capacity; R t V is the normalized value of rainfall; t F is the normalized value of wind speed. t α is the normalized visibility value. r ,α v ,α f To correspond to the influence coefficient; parametric modeling is used to ensure that the twin can reflect the structural characteristics and operational patterns of the warehouse and external roads.

[0036] The update is driven by cross-modal perception: Visual sensors input image features to update the position and movement trends of dynamic obstacles; LiDAR inputs 3D point cloud features to correct the distribution of roads and obstacles within the cabin in real time; millimeter-wave radar inputs speed and distance information to supplement perception in low-light and occluded scenarios; acoustic sensors input echo information to improve detection coverage of weakly reflective objects. Through the above multimodal data-driven approach, dynamic updates and redundant perception of the twin model are achieved, ensuring high consistency between the virtual twin and the real environment. All observations are unified to the same time reference and transferred to the global coordinate system, with coordinate transformation as described above. For each mode, observations such as position, velocity, classification, and obstacle probability are extracted, and the observation noise covariance R is established for each mode. mSymbol convention: m is the modality index, encompassing lidar, vision, millimeter wave, acoustic, and RFID. Stable fusion of information matrices and information vectors yields the twin-state posterior. Information fusion formula:

[0037] Λ prior With η prior For the prior information matrix and information vector; Λ post With η post The fused posterior information matrix and information vector; H m R is the observation mapping (or Jacobian) for mode m; m Let z be the observation noise covariance of mode m; m Let (Λ) be the observation vector for mode m. post ,η post The optimal state estimate can be obtained by solving the problem, thereby updating the twin in real time.

[0038] Risk propagation modeling: A dynamic graph convolutional network is used to dynamically model the road topology within the warehouse; events such as local congestion and increased vehicle density are used as input features to simulate their propagation process throughout the warehouse environment; potential risk regions at the global level are output for path avoidance in subsequent optimization modules, defining the risk state on edges or nodes in the graph. The risk vector (congestion degree, blocking probability, security risk) is defined for each edge or node. Let r be the risk at the edge level. t Risk is propagated in a rolling manner over discrete time, and external sources (peak tasks, crowds) are injected. Propagation formula: r t+1 =σ(W0r t +W1Ar t +b+u t ), Let A be the risk vector of each edge at time t; A be the matrix of edge adjacency or mapping from node adjacency to edges; W0, W1 be the learnable weights; b be the bias; u t σ is the external activation vector; σ(·) is the nonlinear activation function. Node-level equivalent GCN form (optional): H (l) Features of the l-th layer nodes; To add a self-connected adjacency matrix; W is the degree of correspondence matrix. (l) φ is the weight of the l-th layer; φ(·) is the activation function.

[0039] Multi-scale risk prediction: Utilizing Transformer-XL to model short-time series data, achieving second-level micro-prediction is used to identify potential AGV collision risks. Second-level micro-prediction (Transformer-XL collision avoidance): Inputting the local time series and local topology of the next K steps, outputting the proximity collision probability and short-term channel availability for the next h steps. Formula: For the future h-step microscopic prediction vector; [s t-K+1：t [This represents a local state sequence of approximately K steps (velocity, spacing, rotation angle, braking signal);] This describes the local neighborhood topology; TXL(·) is the Transformer-XL encoder-decoder mapping; combining the spatiotemporal graph attention mechanism to model long-term time-series data, it achieves minute-level macro-prediction, used to identify road congestion trends, regional bottlenecks, and the evolution of work-intensive areas. Minute-level macro-prediction (Spatiotemporal Graph Attention ST-GAT): Input a longer time window and full-graph features to predict global congestion and queuing trends in the next H steps. Formula: For the future H-step macroscopic forecast tensor; X t-L+1：t The feature sequence of the entire graph in the near L steps (flow, occupancy rate, arrival and departure rate, and historical risk per side); This is the internal traffic map; ST-GAT(·) is the spatiotemporal attention network, and micro-macro fusion forms a synthetic risk score: second-level and minute-level predictions are fused according to a trade-off coefficient for direct use by the optimization and scheduling modules. Formula:

[0040] ρ t λ represents the composite risk score vector; λ∈[0,1] represents the micro-to-macro tradeoff coefficient; Agg(·) is the local aggregation function (maximum, mean, or Top-k); Pool(·) is the global pooling function (partition average or attention convergence).

[0041] Extreme Condition Robust Simulation: Extreme operating conditions are generated using a Robust Generative Adversarial Network (GAN), including complete road closures, extreme weather effects, and sudden equipment failures. These scenarios are run in a virtual twin to perform extreme stress tests on the system's path planning and scheduling algorithms. The AGV's response behavior under extreme scenarios is recorded to evaluate the robustness of the path planning strategy. A perturbation space is defined, and extreme operating conditions are generated. Perturbations include a joint distribution of traffic accidents, road closures, low adhesion coefficients, heavy rainfall, and sensor failures. The generator produces the "most difficult scenario." The stability of the planning and scheduling strategy is tested under the most difficult scenario, using a min-maximum adversarial objective. The outer loop seeks more difficult scenarios, while the inner loop optimizes the planning strategy. Formula: θ P For planning / scheduling strategy parameters; θ Gξ represents the generator parameters; ξ represents the disturbance scene vector (accident location, blockade length, friction coefficient, rainfall intensity, sensor failure rate). The scene distribution output by the generator; This represents the comprehensive cost (weighted sum of time, energy consumption, and risk penalties) under scenario ξ. Robustness assessment is performed using quantile risk, as shown in the formula: The τ quantile risk of strategy π (typically τ = 0.95 or 0.99); Let be the probability under the perturbation distribution; r be the cost threshold. Use metrics to screen robust strategies that pass stress tests.

[0042] Two-way coupling simulation between warehouse and external environments: The warehouse twin and the external traffic twin interact bidirectionally. When peak logistics activity within the warehouse causes congestion at the exit, the external traffic twin can sense this and adjust traffic flow predictions. When external roads become congested due to accidents or traffic control, the warehouse twin re-evaluates AGV scheduling and path planning based on input, resulting in a globally consistent risk simulation outcome. Coupling boundaries and stepping strategies are defined, with warehouse entrances and exits serving as boundary nodes. The internal and external simulations use a unified step size or sub-step size for synchronous exchange of boundary flows and delays. The external → internal boundary inbound cycle time is mapped using the following formula: Let t be the inbound arrival rate; For external vehicle traffic entering the park; κ1 represents the saturation level of the external road section (0 to 1); κ1 is the port capacity conversion factor. The formula for writing back the external load from the inner to outer boundary is: This is the next step to determine the total traffic volume of external road sections; For external background traffic; κ2 represents the internal outbound release rate; κ2 is the amplification factor related to the number of entrances / exits and lanes. Weather-related adjustments to external capacity are introduced and linked to internal scheduling, as shown in the formula (same as before): External effective passage capacity; C0 is the baseline capacity; R t Normalize the rainfall; V t For wind speed normalization; F t For visibility normalization; α r ,α v ,α f This represents the impact coefficient. Capacity reduction will trigger internal flow control, delayed release, and rerouting strategies. The integrated risk aggregation output merges micro-risks, macro-trends, and boundary conditions into a single risk score, which can be directly used by the path planning and scheduling modules. Formula (same as the previous fusion formula):

[0043] ρ tLet λ be the integrated risk score; λ be the tradeoff coefficient; Agg(·) be the local aggregation function; and Pool(·) be the global pooling function. This score is directly used as a constraint or penalty input for multi-objective optimization, achieving "closed-loop" risk control that coordinates internal and external operations.

[0044] The cross-modal digital twin and risk simulation module plays a continuous role in the system's operation. First, it relies on cross-modal perception data to drive real-time twin updates, ensuring synchronization between the virtual environment and the real warehouse scenario. Then, it identifies potential risks through dynamic graph convolutional networks and spatiotemporal prediction models, and verifies robustness by combining extreme operating conditions generated by GANs. Finally, it couples the in-warehouse and out-of-warehouse twins to form an integrated risk simulation system, providing reliable and forward-looking data support for multi-objective intelligent optimization and game-driven AGV collaborative scheduling.

[0045] The multi-objective intelligent optimization module abstracts the AGV path planning task into a multi-objective optimization problem, comprehensively balancing five dimensions: transportation time, energy consumption, safety, equipment lifespan, and task priority. Its core is the combination of hierarchical multi-objective reinforcement learning and evolutionary algorithms to generate a Pareto optimal solution set while satisfying causal logic constraints and uncertainty robustness. Furthermore, through classification and dynamic update mechanisms, it forms optimal candidate path schemes adapted to different operating scenarios.

[0046] Optimization index construction: Transportation time index: Calculate the total travel time of the AGV from the task start point to the destination, reflecting the efficiency level, and the time cost T(π) (based on the sum of the travel time of the path π): T(π): Total time for the entire path; π: Path consisting of ordered edges e; l e : Length of edge e; v e (k e ): The effective speed of side e; k e Traffic density at edge e (vehicles / meter); V f : Free-flow velocity; k jam Congestion density threshold. The first equation sums the "length / speed" of each edge; the second equation is the Greenshields basic graph, using density to approximate speed; Energy consumption index: Combining battery SOC and SOH states, the energy consumption required for different paths is estimated, reflecting energy use efficiency, energy cost E(π) (based on discrete power integral approximation): E(π): Path energy consumption; t: Discrete-time index; N: Number of steps; m: Total vehicle mass; a t Longitudinal acceleration; v t Speed; c r Rolling resistance coefficient; g: acceleration due to gravity; ρ: air density; Cd : Drag coefficient; A: Frontal area; η: Transmission efficiency; Δt: Sampling interval. These three items correspond to acceleration work, rolling resistance work, and wind resistance work; Safety indicators: quantified through factors such as path conflict probability, obstacle risk value, and collision risk prediction to reflect operational safety. Safety indicator R(π) (based on second-level collision risk synthesis):

[0047] R(π): Path safety risk (lower is safer); p col (t): Collision probability at time t output by the micro-prediction model; N: Number of time steps. The formula represents the synthesis of parallel risks that "occur at least once"; Equipment life index: Based on the wear and tear of mechanical components and battery degradation caused by frequent acceleration, sudden braking and high load operation, the degree of life loss is estimated, life loss L(π) (combined with battery SOH and mechanical shock): L(π): Total lifetime loss; β1, β2, β3: Trade-off coefficients; a t : acceleration; j t =(a t -a t-1 ) / Δt represents the jerk (impact); ΔSOH represents the decrease in battery health over the cycle (estimated by the coulomb flux and temperature stress model). The first three terms penalize frequent acceleration / deceleration, impact, and chemical aging, respectively; Task priority index: scores the urgency of orders, assigning higher priority tasks with greater weight during optimization, task priority P(π) (based on delivery deadline and urgency):

[0048] P(π): Priority penalty (the smaller the better); The path is expected to arrive at the specified time; d: order deadline; u∈[0,1]: task urgency (1 indicates the most urgent); γ1, γ2: tradeoff coefficients. Exceeding the deadline will increase the penalty, with non-urgent tasks incurring a larger penalty, thus allocating resources to high-urgency tasks. Subsequent optimization will use a minimization vector objective: F(π)=[T(π), E(π), R(π), L(π), P(π)].

[0049] Optimization Method Design: Hierarchical Multi-Objective Reinforcement Learning (HMORL) is used to model different optimization objectives hierarchically. Through training, a policy that can consider multiple dimensions of performance is obtained. A hierarchical structure and MDP are defined, along with a high-level policy. Select target g ("corridor / area / critical segment"), low-level strategy Select a specific edge / control given g. Formula: State space (position, speed, battery level, local congestion, remaining time); Motion space (edge ​​selection, speed settings); Transition probability; The vector reward corresponds to a negative form of [-T, -E, -R, -L, -P] so that "the larger the reward, the better"; γ∈(0,1): discount factor. Scalar training of the weight vector sampling is performed to cover the Pareto surface, sampling the weight vector w∈Δ. 5 (5-dimensional simplex) scalarizes the vector reward. Formula:

[0050] J(θ; w): Expected return given weight w; θ = {θ H ,θ L}: High-level and low-level strategy parameters; r t Vector reward at time t. PPO-style policy update (once every w):

[0051] Importance ratio; The advantage function is based on scalar return; ∈: truncation threshold controls the update magnitude; π θ Hierarchical strategy joint distribution; combining evolutionary algorithms (NSGA-II and NSGA-III) to quickly generate multi-objective Pareto optimal solution sets, avoiding single-objective bias; population initialization and objective evaluation; random generation of multiple candidate paths; calculation of F(π). Non-dominated sorting (Pareto advantages and disadvantages): and π i Not worse than π on all targets j And at least one is better; m: target index (five targets); F m The m-th objective value. Crowding distance (NSGA-II for maintaining diversity): Individual π i Crowded distance; The target values ​​of adjacent individuals on target m; The objective is to find the maximum and minimum values ​​within the current population. Reference direction projection (NSGA-II for high-dimensional objectives): Individuals are projected onto a preset reference direction to uniformly cover the Pareto surface, preserving representative solutions close to different reference directions. Crossover and mutation: Consistent crossover using path encoding and small-perturbation mutation are used to generate a new generation, repeating until convergence or the number of generations is exhausted.

[0052] Causal constraint introduction: Through causal modeling, the logical chain between different optimization objectives is identified, and path solutions that do not conform to causal logic are eliminated; when a path shortens transportation time but causes an extreme increase in energy consumption and triggers rapid degradation of battery state of harmonics (SOH), the solution is excluded, and a causal graph is constructed. Nodes include T, E, R, L, P, and exogenous variables (load, gradient, speed limit, weather, traffic density). Directed edges represent causal directions, such as "higher speed → higher energy consumption and higher risk". Counterfactual testing and average causal effects are used to intervene in candidate paths by "shortening time δ" and assess whether energy consumption has abnormally increased beyond a threshold. Formula: ACE E←T : Average causal effect of time on energy consumption; do(·): Pearl's exogenous interference inference; δ>0: Time reduction. Elimination criterion: If ACE E←T >τ E If the threshold is reached, it is considered an irrational path of "time-reduced detonation energy consumption" and is eliminated.

[0053] Fuzzy robust optimization processing: For uncertainties such as traffic flow and obstacle movement trajectories, a fuzzy multi-objective optimization method is employed to construct fuzzy constraint boundaries. This ensures that the path solution remains feasible and stable even under fluctuating actual operating conditions. Uncertain traffic volume / density is modeled as fuzzy numbers; for example, "the maximum acceptable time for edge e" is a trapezoidal fuzzy number. Opportunity-based fuzzy constraints, for constraints Require, Π(·): Probability measure; ξ: Traffic uncertainty; α∈[0,1]: Credibility threshold. When When the number is a trapezoidal fuzzy number, the above formula is equivalent to: T e (π, ξ)≤b+(10-α)(cb), where a,b,c,d are the trapezoidal fuzzy number support and kernel; b+(1-α)(cb) is the soft upper bound under a given confidence level. For capacity and density constraints, the same α-cut transformation can be used to obtain interval parallel testing.

[0054] Candidate Solution Set Aggregation and Classification: The generated Pareto optimal solution set is clustered and summarized to form three categories of candidate solutions: Fast-optimal solution (shortest transportation time); Stable-optimal solution (optimal balance between safety and lifespan loss); Energy-efficient-optimal solution (minimum energy consumption). These three categories of solutions can be selected by the system or manually according to different scenario requirements. Normalized target vector:

[0055] y i : Standardized features of the i-th candidate solution; Z∈{T, E, R, L, P}; minZ, maxZ: Minimum and maximum values ​​of the objective in the candidate set. K-means clustering (k=3): Cluster of type j; μ j : The center of the j-th cluster. Labeling rule: If the cluster center μ j exist Minimum priority, marked as "fast and efficient"; in and Smaller sizes are the primary focus, marked as "stable and superior"; in Minimum priority, marked as "Energy Saving Advantage".

[0056] Rolling optimization and dynamic updates: Employing a rolling optimization mechanism using Model Predictive Control (MPC), the optimal path is recalculated and updated at regular time windows. When warehouse conditions, traffic status, or task requirements change, the system can quickly correct the path, maintaining the real-time nature and adaptability of the optimization results. A predictive model and rolling window are established: in the current state x... t Next, plan the future H-step control sequence {u t:t+H-1}, only the first step is executed, and then the clock is moved forward by Δt seconds to solve again. {u t:t+H-1}: Future H-step control (edge ​​selection, speed); c = [T k E k ,R k ,L k ,P k [] represents the cost vector at each step; w * Weights selected from the Pareto solution or policy library; Φ(·) terminal potential (e.g., heuristic distance to the target); λ term Trade-off coefficients; H prediction step size. Constraints and dynamics: x t+k+1 =f(x) t+k u t+k ), g(x t+k u t+k )≤0, f(·): AGV dynamics; g(·): Constraints such as collision, speed, turning radius, and fuzzy α-cut. Scrolling execution: Only executes u. t As time progresses to t+Δt, the prediction is updated, and the above problem is solved again.

[0057] During system operation, the multi-objective intelligent optimization module continuously receives input information from the multi-source sensing and data acquisition module, as well as the cross-modal digital twin and risk simulation module. By establishing a multi-dimensional optimization index system and generating candidate path solutions under the protection of causal constraints and fuzzy robust mechanisms, it ensures that the AGV can obtain the comprehensive optimal path in terms of time, energy consumption, safety, lifespan, and task priority under different task conditions. The path solutions are provided to the game-driven AGV collaborative scheduling module in the form of classification and aggregation, realizing globally efficient collaborative transportation scheduling.

[0058] The AGV collaborative scheduling module achieves dynamic collaborative scheduling of multiple AGVs in complex warehousing scenarios by combining game theory modeling and reinforcement learning. It abstracts the multi-vehicle path and task allocation problem into a distributed game process, using mechanisms such as Stackelberg game, coalition game, signaling game, and cooperative transition game to solve problems of information asymmetry, resource conflicts, and path congestion. Simultaneously, it employs multi-agent reinforcement learning to train the AGV group to form long-term effective interaction strategies, and a fairness mechanism ensures reasonable distribution of benefits and resources, improving overall scheduling efficiency and robustness.

[0059] Distributed Stackelberg game mechanism: Some AGVs are designated as leaders, and the rest as followers. Participants, strategies, and payoffs are defined. Several AGVs in the same work area or adjacent road networks are considered as a single game, denoted as [the set of AGVs]. A portion of these paths are selected by the Leader first, while the rest are followed by lower-level paths that make the optimal response based on the Leader's decision. Each individual i's strategy is either a path or a control sequence, denoted by π. i The system aims to improve overall throughput and timeliness while minimizing congestion / conflict costs. The master vehicle makes advance decisions in route selection and task execution, and slave vehicles dynamically adjust their routes based on the master vehicle's choices. This mechanism reduces the probability of global conflict through hierarchical game theory, enabling overall scheduling to maintain autonomy while also possessing coordination. Leader route selection (upper-level optimization): The leader first selects a route within its feasible domain to maximize its own utility, while also considering externalities to the environment and followers. Formula: U L (·): Leader's utility (which can be composed of a negative sum of time, energy consumption, and conflict penalty); π L : The path to the Leader; Fol lower's optimal response; Π L The set of feasible paths for the Leader; Leader constraints (speed, turning radius, collision safety distance, time window, etc.). Follower optimal response (lower-level optimization): Given the Leader's path decision, the Follower finds the optimal response path within its own feasible region. Formula:

[0060] π F : The path to the Follower; Π F :Fol lower the set of feasible paths; U F (·): The effect of lowering the fold; Follow lower constraints (including avoidance constraints on the Leader path). Distributed optimal response iteration and convergence criteria: Distributed optimal response iteration using nearest neighbor communication: The Leader gives... Each lowercase is solved in parallel. Leader updates based on Follower response. formula: Shutdown conditions: BR F (·),BR L (·): Optimal response operator for Follower / Leader; dist(·,·): Distance metric between paths (can be total time difference, edge set symmetry difference, or weighted edit distance); ε: Convergence threshold; k: Number of iterations; Leader score selection and proportion control: Use interpretable metrics for scoring, select Top-K as Leaders, and the rest as Followers, allowing rotation. Formula: s i =w1c i +w2b i +w3k i +w4χ i s i AGVi's Leader rating; c i : Current mission payload (weight or volume normalized); b i ∈[0,1]: State of Charge (SOC); κ i : Network betweenness centrality (the degree of "criticality" of a path to overall traffic); χ i Historical on-time performance / strategy stability and other capability indicators; w1~w4: weights (which can be obtained from experience or learning). The highest-scoring candidate becomes the Leader. Leader rotation mechanism: To avoid long-term stagnation, a rotation occurs every H... L Each scheduling cycle, for s i Reassess and apply cooling-off factor penalties based on contributions within the window to promote fair rotation. Formula: The score includes a cooling factor; δ>0: cooling coefficient; Near H L The number of times you serve as Leader within a period.

[0061] Multi-level alliance game mechanism: When encountering large-scale material handling or high-priority tasks, multiple AGVs can temporarily form an alliance; within the alliance, each AGV determines its division of labor and path through cooperative game theory to jointly complete the large task; the alliance automatically disbands after the task is completed, ensuring scheduling flexibility. Alliance value function and merging / splitting rules: Multiple AGVs can temporarily form an alliance. Jointly transporting large quantities of materials; the value of the alliance is measured in the total time or cost saved. Formula: v(S) = ΔTS +λ E ΔE S -λ R ΔR S v(S): The value of the alliance S; ΔT S ΔE: Total time saved compared to independent execution S Total energy savings; ΔR S : Reduced total risk; λ E ,λ R >0: Energy consumption versus risk trade-off coefficient. Merge-split criterion: If there exist two alliances S1 and S2 such that v(S1∪S2)≥v(S1)+v(S2), then merge; if v(S)<∑ j v(S j Then it is split into S j Within the alliance, task and dispatch synchronization is implemented: Batch task splitting and synchronized dispatch window optimization are performed within the alliance to steadily improve throughput. Formula (Synchronization Window Optimization): t i AGVi's departure times; The alliance expects to synchronize its actions at the same time. Each AGV can be dispatched via a window.

[0062] Signal game mechanism: During the scheduling process, AGV actively broadcasts its own load status and energy consumption level as a game signal; other AGVs adjust their own paths and task selections based on the received signals, reducing resource competition caused by information asymmetry.

[0063] This mechanism improves scheduling transparency and global optimization efficiency. Broadcast message structure and frequency: per Δt s State vector broadcast once per second: Formula: s i (t)=[l i (t), b i (t), η i (t), q i (t)],s i (t): The common signal of AGVi at time t; l i (t): Load factor (0-1); b i (t): Electricity (sOC, 0-1); η i (t): Estimated remaining travel time (ETA); q i (t): Local congestion score of the current channel, receiver's belief update and cost correction: other AGVs on s i Perform exponential smoothing, update the state estimate for i, and incorporate its own conflict and yield costs. Formula (smoothing):

[0064] Smoothing estimate of i; α∈[0,1]: forgetting factor; Δt s Broadcast interval. Formula (edge ​​cost correction): c e (t)=c e (t)+μφ e (t)+vψ e (t), c e (t): The corrected cost of edge e; c e (t): Basic cost (time / energy consumption); φ e (t): According to the neighboring AGV The estimated congestion externalities; ψ e (t): Conflict probability estimate; μ, v>0: Penalty coefficients. The adjusted cost is used for subsequent shortest path or MPC solutions.

[0065] Multi-agent reinforcement learning training: Multiple AGVs are modeled as agents, and their long-term interaction strategies are trained using multi-agent reinforcement learning methods. Conflict penalties and cooperation rewards are introduced during training to guide the AGV group to gradually form a stable scheduling strategy with low conflict and high cooperation. The trained strategy can be adaptively applied in dynamic environments, reducing frequent conflicts. CTDE training framework and joint strategy: "Centralized training, distributed execution (CTDE)" is adopted: global information is visible during training, while each AGV uses only local observations during execution. Objective: θ: Multi-agent joint policy parameters; Y∈(0,1): Reward discount; T t Time cost; E t Energy consumption cost; C t : Conflict event count / probability; I t : Cost of empty runs / waiting; Coop t : Cooperation metrics (e.g., successful relay, yielding); α, β, γ c δ, ζ > 0: cost / reward weights, for Actor-Critic-based policy updates: algorithms such as MADDPG / QMIX can be used; the general policy gradient form is given below. Formula: θ i : Policy parameters of agent i; In observation o i Take action a i The probability of; Advantage function estimation (provided by a centralized Critic). The trained policy depends only on local operations during execution. i This reduces conflict.

[0066] Fairness Guarantee Mechanism: A Shapley value allocation method is introduced into the scheduling results. Revenue and path resources are allocated based on the marginal contribution of each AGV to the overall task completion rate. This ensures that different AGVs obtain reasonable task revenue and energy consumption allocation in long-term operation, avoiding some AGVs from being overloaded for extended periods. Task revenue is allocated using Shapley values, and the value v(S) of the alliance S (see step 7) is allocated according to the Shapley value, ensuring symmetry, fairness, and budget balance. Formula:

[0067] φ i (v): AGVi's Shapley value (share of earnings due); Participants: the entire set; S: any subset excluding i; |·|: set size; v(·): alliance value function. Payment rule: by φ i (v) Settlement incentives or scoring, with long-term accumulation to ensure fairness.

[0068] Congestion Self-Healing Cooperation Mechanism: When large-scale congestion occurs in the warehouse, a cooperative transfer game is triggered. The congestion trigger criterion is that when the density or travel delay of local channel e exceeds the threshold, "self-healing" is triggered. Formula (effective upon any trigger):

[0069] or k e (t): Vehicle density at edge e at time t; Density threshold; τ e (t): Current travel delay; Free-flow delay; ρ τ >1: Delay ratio threshold; AGVs achieve load transfer through task exchange and path reconstruction, gradually alleviating congestion. Task exchange (reassignment): This involves solving a bipartite graph matching problem (Hungarian or linear programming) at the cost of "total delay + energy consumption + switching cost". Formula:

[0070] σ(·): Assigning the AGV to the task matching; C i,j The comprehensive cost of AGVi undertaking task j (considering current location, detour cost, estimated completion time, and switching loss) is calculated. A new task—vehicle matching—is obtained. During this process, alliance game and signal game mechanisms work together to ensure the speed and stability of congestion resolution. Local path reconstruction (rolling optimization) involves constrained shortest path / MPC replanning for affected AGVs within their neighborhood subgraph, changing only necessary segments to reduce oscillations. MPC form (subgraph): {u t:t+H-1}: Future H-step control / edge selection sequence; T k E k C k: Time, energy consumption, and conflict cost of step k; α ′ ,β ′ ,γ c ′ Local weights; x k+1 =f(x) k ,u k ): Dynamics; g(·)≤0: Safety and traffic constraints. After executing the first step, roll forward until the congestion is relieved.

[0071] The AGV collaborative scheduling module is responsible for path coordination and task allocation among multiple AGVs during system operation. When the warehousing environment is relatively stable, AGVs achieve autonomous path and task coordination through Stackelberg game and signaling game mechanisms. When facing large or high-priority tasks, the alliance game mechanism ensures efficient collaboration among multiple vehicles. During long-term operation, the multi-agent reinforcement learning strategy ensures that the AGV group develops adaptive collaborative habits. When the system faces congestion risks, the cooperative transfer game mechanism enables AGVs to achieve congestion self-healing through task exchange and path adjustment. Ultimately, this module ensures the fairness, efficiency, and robustness of the AGV group in complex environments.

[0072] The self-evolutionary learning and transfer module introduces methods such as meta-learning, graph embedding transfer, few-shot learning, knowledge distillation, lifelong learning, and self-supervised comparative learning to enable the system to rapidly adapt and evolve under different warehousing environments and task conditions. Its core objective is to ensure that the system can quickly transfer existing experience and continuously optimize itself when facing changes in warehouse layout, increased task types, and evolving operational scenarios, thereby avoiding performance degradation caused by environmental differences and insufficient data.

[0073] Meta-learning (MAML) mechanism: Utilizing model-independent meta-learning methods, an initial model is trained that can quickly adapt to various warehouse layouts. "Warehouse layout / operating conditions" are defined as meta-tasks, with each warehouse or different shifts / flow intensities treated as a task. Collect a small "support set" S from this task environment. i With the "validation set" Q i When the system is deployed to a new repository, only a few iterations are needed to adjust the model parameters, enabling rapid adaptation of path planning strategies. This mechanism significantly shortens the model tuning cycle in new environments, with an inner loop (performing one or more rapid updates on the support set). θ: Shared initial parameters; θ′ i In the task The parameters after adaptation; α>0: inner loop learning rate; f θ Strategy / value or planning scoring model; In support set S iThe loss on the upper loop (which can be a weighted sum of time, energy consumption, and collision risk), and the outer loop (evaluating and updating meta-parameters using the validation set): β>0: Outer loop learning rate; Loss on the validation set using the adapted parameters.

[0074] Graph embedding migration mechanism: Transform the road topology, shelf layout, and task distribution of different warehouses into a graph structure, representing the warehouse as a graph with attributes: Node set (intersection / workstation); Edge assembly (passage / ramp); Node attributes (speed limit, width, slope, hazard factor, etc.); A: Adjacency matrix (including length / traffic weight); A general representation of the warehouse structure is extracted using graph embedding methods to achieve topological information migration between warehouses; When the system switches between different warehouses, it can quickly map and reuse existing scheduling experience. Graph neural network encoding nodes / graph-level embedding: Node embedding; Add self-connects and normalized adjacencies; Correspondence matrix; GNN(·): Graph convolution / attention encoder; h: Embedding dimension. Graph-level embedding: g = READOUT(Z) (mean / attention convergence), cross-warehouse distribution alignment (aligning source domain embeddings to target domain): Θ: Parameters of GNN and downstream head; Task loss (e.g., path evaluation error); MMD(·,·): maximum mean difference, used to measure the source domain graph embedding g. (s) With the target domain g (t) The distribution difference; λ>0: alignment weight. Structural alignment can also be performed.

[0075] P: Node-corresponding matrix; A transport polyhedron with row and column boundaries a, b; d s ,d t : Shortest path or topological distance between node pairs in the source / target graph. This formula is Gromov–Wasserstein aligned and can be used for node-level migrations.

[0076] Few-shot learning mechanism: To address situations where a large amount of training data is lacking for new tasks or environments, a few-shot learning method is introduced. By completing effective model updates with limited sample data, the system maintains its path planning performance even in data-scarce scenarios. This mechanism effectively solves the cold start problem in the initial deployment of new warehouse scenarios, using metric learning (prototype network) to stabilize generalization with few samples. c k : Prototype center of category / intent k; S k: Supports samples within the set that belong to k; f θ (x): Representation of sample x; d(·,·): Distance (Euclidean / cosine); τ>0: Temperature; p(y=k|x): Classification or intent selection probability. A prototype can be estimated with a small number of samples, used for intent / segment selection under new layouts, with fast fine-tuning via regularization. φ: Fine-tuned parameters; θ: Initial parameters (can be derived from MAML); Few-sample loss (cross-entropy / regression); λ reg : Regularity coefficients close to the prior.

[0077] Cross-scenario knowledge distillation mechanism: Complex training models are run on a simulation platform to obtain large-scale task experience and path planning strategies; the knowledge from these complex models is distilled and compressed into a lightweight model, which is then deployed in the AGV edge controller; this enables the path planning model to run efficiently under real-time and computationally resource-constrained conditions. Classification / scoring distillation (temperature soft label): z T ,z S Teacher / student logits; σ(·): Softmax; y: Truth label (if any); T>0: Distillation temperature; λ∈[0,1]: Distillation weights; CE: Cross-entropy; KL: KL divergence, policy distillation (reinforcement learning scenario):

[0078] π T ,π S : Teacher / student strategy distribution; s: state; V T V S ξ: Value estimation of teachers / students; ξ>0: Value distillation weight.

[0079] Lifelong Learning (EWC) Mechanism: Through the elastic weight retention algorithm, constraints are imposed on key parameters when learning new tasks; this prevents the model from forgetting the performance of old tasks during updates, ensuring the stability of the system in long-term evolution; it enables the system to accumulate learning and maintain performance under different storage environments, and statistically analyzes important parameters of old tasks (Fisher information).

[0080] F i : parameter θ i The Fisher information is approximate; Old task data; π θ The current strategy is to add a penalty for "moving away from the old optimal" to new tasks.

[0081] New mission losses; The optimal parameters learned from the previous task; λ ewc>0: Stability-Plasticity tradeoff coefficient. This ensures that key parameters are not drastically altered when learning new scenarios, preventing catastrophic forgetting.

[0082] Self-supervised contrastive learning mechanism: During system operation, unlabeled data generated by AGVs is used for contrastive learning; by constructing contrastive relationships between positive and negative samples, the model continuously improves its representational ability under unsupervised conditions; ensuring the system possesses self-optimization capabilities during long-term operation, without relying on large amounts of manually labeled data. It generates two views (time-series jitter, sensor occlusion simulation, noise, etc.) for the same short-window trajectory as positive sample pairs, and the rest as negative samples. InfoNCE contrastive loss is used to bring positive pairs closer and push negative pairs further apart.

[0083] z i , Two enhanced embeddings of the same trajectory segment; z k : Negative sample embedding; sim(·,·): Similarity (commonly cosine) );τ>0: Temperature. After training, it means that "robust features" that are insensitive to noise / occlusion have been learned, which can be used as upstream inputs or initialization parameters to improve overall robustness.

[0084] The self-evolutionary learning and transfer modules play a continuous role during system operation. When the system faces a new warehouse layout, meta-learning and graph embedding transfer mechanisms ensure rapid adaptation; when data is insufficient, few-shot learning mechanisms ensure basic performance; when computing resources are limited, knowledge distillation mechanisms ensure efficient model operation at the edge; in long-term evolution, lifelong learning mechanisms prevent catastrophic forgetting, while self-supervised comparative learning mechanisms enable the system to continuously improve performance based on operational data. Through the synergistic effect of these mechanisms, the system possesses self-evolutionary capabilities across warehouses, scenarios, and time dimensions, thereby significantly improving the generalization and robustness of AGV path planning.

[0085] The anomaly detection and robust protection module addresses sudden anomalies and operational risks of warehouse AGVs in complex and dynamic environments. Combining statistical analysis, causal reasoning, and a tiered response mechanism, it enables real-time identification and protective measures against hardware failures, communication anomalies, path congestion, and other emergencies. Without compromising overall system stability, it ensures the continuity and safety of AGV group operation through tiered response strategies and redundant takeover mechanisms, and enhances the understandability and traceability of anomaly handling through causal explanation chains.

[0086] Anomaly Detection Mechanism: Hardware Fault Detection: Continuously monitor the health status of the AGV's motor, sensors, communication modules, and battery. Utilize time-series statistical characteristics and causal dependencies to determine the presence of fault signals. Mahalanobis Distance Threshold Test (Instantaneous Anomaly): Identify instantaneous anomalies in sensors / actuators using the Mahalanobis distance of the residuals. Formula: like Then it is determined to be abnormal. R is the Mahalanobis distance; t The observation noise covariance; p is the residual dimension; The chi-square threshold at significance level α, EWMA / CUSUM drift detection (slow degradation): suitable for detecting sensor zero drift and execution efficiency degradation. Formula (EWMA): S t =λ||r t ||+(1-λ)S t-1 If S t >τ ewma Detect abnormality, S t The exponential moving average statistic is λ; λ∈(0,1) is the forgetting factor; ||r|| t || represents the residual norm; τ ewma For the threshold. Formula (one-sided CUSUM): C t =max(0, C t-1 +(||r t ||-μ0)-k), if C t >h indicates an anomaly, C t For cumulative sum statistics; μ0 is the normal mean; k is the drift sensitivity; h is the alarm threshold; Communication delay detection: Real-time evaluation of the delay and packet loss rate of the wireless communication link. Once the set threshold is exceeded, an anomaly marker is triggered. Delay and packet loss statistical thresholds are used to monitor the round-trip delay D of messages. t Interval jitter J t Window packet loss rate P t . formula:

[0087] If D t >μ D +κ D σ D or P t >τ P Detect abnormality, J t =|(t) k -t k-1 )-(t k-1 -t k-2 )|,μ D , σ D κ represents the mean and standard deviation of the time delay (sliding window estimation); D τ is a multiplier (e.g., 3); P t is the packet loss threshold; kJ represents the arrival time of the k-th message; t For jitter; Path congestion detection: Through multi-source sensing input and digital twin feedback, it identifies phenomena such as excessively high vehicle density or decreased movement speed in local areas, travel delay ratio and queue length threshold: simultaneously using "travel"

[0088] The dual metrics of "line delay ratio" and "queue length" are more robust. Formula: or τ e (t) represents the current travel delay of edge e; For free flow delay; ρ τ >1 is the multiplier threshold; L q (t) is the approximate queue length of M / M / 1; ρ(t) = λ(t) / μ(t) is the utilization rate (arrival rate / service rate); τ L Queuing threshold; Emergency detection: Identify emergencies in warehousing scenarios such as accidental entry of personnel, falling goods, and equipment failure, and immediately mark them as high-priority anomalies. Use Bayesian network posterior regression to locate the root cause of the failure, and construct variables: hardware failure F, communication anomaly C, congestion G, and observational evidence. formula: This represents the posterior probability of a hardware failure. The likelihood is given by P(F=f), and the prior is given by P(F=f). Similarly, the posterior is calculated for communication (C) and congestion (G), followed by counterfactual testing (excluding spurious congestion / anomalies). Intervention is applied to "traffic flow" to observe whether travel delays decrease. Formula: Δτ represents the time delay change after reducing the arrival rate; λ represents the arrival rate; do(·) represents exogenous intervention. If Δτ≈0 but the residual remains large, it tends to indicate sensor anomaly; if Δτ<0 is significant, it tends to indicate actual congestion.

[0089] Tiered Response Strategy: Mild Anomaly Handling: When mild hardware fluctuations or minor congestion are detected, the system replans local paths through a path fine-tuning mechanism to avoid task interruption. This involves local cost correction + shortest path / MPC fine-tuning, increasing penalties for affected edges, and calculating alternative local paths; or performing short-window MPC fine-tuning to adjust speed / edge selection. Formula (Cost Correction): c′ e (t)=c e (t)+μφ e (t)+vψ e (t), c′ e (t) represents the cost of correcting the following values; c e (t) represents the basic cost (time / energy consumption); φ e (t) represents the congestion penalty; ψ e (t) represents the conflict penalty; μ,ν>0 are coefficients. Formula (Local MPC): {u t:t+H-1} is for future H-step control; ΔT k,ΔE k R represents the time / energy consumption increment relative to the original plan. k Risk is represented by α, β, γ > 0; weights are α, β, γ > 0; f represents dynamics; g represents safety / traffic constraints; moderate anomaly handling: when communication delays, partial AGV failures, or regional congestion are detected, the system triggers a task reassignment mechanism to dynamically transfer tasks to other normally operating AGVs. Bipartite graph assignment minimizes the total cost, using a cost matrix constructed from "estimated completion time, additional detour, switching loss, and energy consumption" to solve for the optimal vehicle-to-order matching. Formula (linear assignment): ∑ i x ij =1,x ij ∈{0,1}, A set of available AGVs; C is the set of tasks to be assigned. ij The total cost of executing j for i; x ij This is a 0 / 1 variable indicating whether an assignment is made; Severe anomaly handling: In the event of a serious hardware failure, large-scale communication interruption, or complete blockage of the critical path, the system triggers a redundant AGV takeover mechanism. A backup vehicle automatically replaces the faulty AGV to perform the task, ensuring the continuity of global scheduling. Takeover candidate selection and safe relay: The system selects the vehicle with the lowest "takeover cost" from the redundant fleet and performs a safe relay to isolate the faulty vehicle. Formula (candidate scoring):

[0090] J k The cost of taking over candidate k; The estimated time to reach the take-off point; dist(k,site) is the distance from the current location to the take-off point; b k ∈[0,1] represents the charge; (1-b k ) + =max(0,1-b) k ) is a low battery penalty; ω t ,ω d ,ω b >0 indicates a weight. Reliability verification (optional):

[0091] Avail k Availability; MTBF k Mean Time Between Failures (MTTR) k The average repair time is calculated. Vehicles with high availability are prioritized. A comprehensive risk score and grading threshold are used to weight the posterior probabilities of the three types of anomalies into a total score, and a dual-threshold grading system is set. Formula: S represents the total score for abnormalities; w F ,w C ,w G>0 represents the weight (can be calibrated offline or learned online). Grading rules: Mild: S∈[θ1,θ2), Moderate: S∈[θ2,θ3), Severe: S≥θ3, where θ1<θ2<θ3 is a threshold obtained through experience or learning (with hysteresis to avoid jitter).

[0092] Robust protection mechanism: Establish a multi-layer redundancy strategy to ensure that when a single device or a single communication link fails, the task can still be completed through the backup node; through cross-vehicle communication and game-theoretic scheduling feedback, task reconstruction is realized in the case of local anomalies to prevent small-scale failures from spreading to global paralysis; use causal reasoning model to determine the root cause of the anomaly to avoid unnecessary scheduling overhead caused by misjudgment of appearance.

[0093] Causal Explanation Chain Generation: While detecting and handling anomalies, the system generates a causal explanation chain for the cause of the anomaly. The explanation chain includes the antecedents, triggering mechanisms, handling measures, and result evaluation, ensuring that operations and maintenance personnel can quickly understand the system's response logic. This function not only supports manual review and operations and maintenance decisions but also serves as input for subsequent system optimization and model updates, constructing causal graphs and evidence path scoring. Causal graph nodes include "hardware state F, communication state C, congestion G, and observational evidence." Action A. Select the causal path that best explains the observations and quantify its contribution. Formula (Path Explanation Score): Score(π) = ∑ e∈π IG(e), IG(e) = KL(P(Y|do(X)) e ))||P(Y)), where π is a directed path from the "root cause" to the "abnormal manifestation"; IG(e) is the information gain of edge e; KL(·||·) is the KL divergence; X e Y is the starting variable of this edge; Y is the set of terminal observation variables. The higher the score, the more explanatory the path. A readable explanation template is generated, and the posterior probability, threshold breach items, and handling actions are filled into the fixed template to automatically generate audit text.

[0094] The visualization and operation and maintenance support module works continuously during system operation, monitoring the AGV's operating status and environmental feedback in real time. Once an anomaly is detected, the system immediately handles it according to a tiered response strategy, progressing step by step from path fine-tuning to redundancy takeover, ensuring the continuity and robustness of scheduling. Simultaneously, the causal explanation chain makes the processing transparent, providing operation and maintenance personnel with clear decision-making basis. Through the operation of this module, the warehouse AGV system can maintain a high level of stability, safety, and controllability when facing complex environments and unexpected events.

[0095] The visualization and operation and maintenance support module provides system operation and maintenance personnel with an intuitive, transparent, and explainable support interface through 3D digital twin visualization, causal heatmap presentation, and natural language generation technology. It outputs complex path planning, risk prediction, and anomaly handling logic in a visual and natural language manner, enabling operation and maintenance personnel to quickly understand the AGV's operating status and the system's decision-making basis in a non-technical context, thereby improving system controllability and operation and maintenance efficiency.

[0096] Digital Twin Visualization: Constructing a 3D digital twin interface to present the warehouse's internal road topology, shelf layout, and AGV's operating trajectory and location distribution in real time; dynamically rendering path planning results in the visualization interface, allowing maintenance personnel to intuitively see the task execution progress of each AGV; integrating energy consumption and efficiency visualization indicators to display the energy consumption level of a single vehicle and the overall system, task completion rate, and operating efficiency in the interface; data access and spatiotemporal alignment (status → screen), obtaining the AGV pose, path, and risk grid for each frame from the digital twin; aligning with a unified timestamp and mapping it to the warehouse's global coordinate system; camera view and projection (turning 3D points into screen pixels); any 3D point p W First, transform (world coordinates) to camera coordinates, then project them into pixel coordinates using perspective. Formula (World → Camera): p C Camera coordinates; The camera's rotation relative to the world; This represents the camera's translation in world coordinates. Formula (Camera → Normalized Device Coordinates, NDC): K is the perspective projection matrix (including focal length, principal point, and view frustum parameters); Formula (NDC → pixel): (u,v) represent pixel coordinates; W,H represent the canvas width and height in pixels. Path rendering (smooth trajectory and visible corners): Discrete road network nodes are interpolated into smooth curves to avoid jagged edges and visually display turning radii. Formula (uniform B-spline):

[0097] p(u) is a point on the curve; P i For control points (road network nodes); N i，k (u) represents the k-th degree B-spline basis function; n is the number of control points minus one; u is the curve parameter. Optional index (curvature is used for coloring or marking dangerous corners): κ(s) is the curvature at arc length s; p′(s) and p″(s) are the first and second derivatives, respectively.

[0098] Risk and Anomaly Visualization: A risk warning layer is overlaid on the interface to display potential localized congestion areas, collision points, and restricted passages. Different risk levels are marked with color gradients for quick identification of high-risk areas. The results output from the anomaly detection module are displayed synchronously, including hardware fault markers, communication delay alerts, and path congestion warnings. Risk heatsinking (color = risk intensity, transparency = confidence level) overlays the twin risk field (from prediction / propagation) onto the ground or path with a semi-transparent heatsink layer. Formula (risk field kernel density estimation):

[0099] R(x) represents the risk intensity at location x; x i Let σ be the location of the i-th risk source; i The diffusion scale (grid smoothing radius); w i is the weight; M is the number of risk sources. Formula (color and transparency mapping): C represents color (such as the perceptually consistent Viridis color band); α represents transparency; For normalized risk; α min ,α max This represents the upper and lower limits of transparency. Level of Detail (LOD) and frustum clipping (for smoother high frame rates): distant objects are simplified, while nearby objects are detailed; objects outside the frustum are directly culled. Formula (bounded sphere frustum test): Let c be the enclosing sphere; c be the center of the sphere; r be the radius; n be the radius of the sphere. j ,d j Let be the normal and offset of the j-th view frustum; ∧ represents the logical AND. Formula (LOD selection): l represents the LOD level; s(l) represents the expected pixel size of the model at that level on the screen; s target Given the desired pixel size. Timeline playback and comparison (accident review / solution comparison), for the state sequence {x t The interpolation is smooth and supports drag-and-drop playback. Formula (linear interpolation): x(t) represents the interpolated state at time t; t k ,t k+1 The time of adjacent keyframes.

[0100] Causal Heatmap Presentation: A causal heatmap is generated based on causal inference results, displaying the causal relationships between different path choices, energy consumption changes, and abnormal events in heatmap form. The heatmap highlights key nodes and causal strength in the causal chain, enabling operations personnel to understand the logical basis for the system's choice of a particular path or execution of a specific anomaly handling. Interactive functionality is provided, allowing operations personnel to click on a node to view a detailed causal explanation chain and corresponding processing procedure. A causal graph is constructed (variables → directed edges → adjacency weights): nodes represent key indicators and exogenous factors (such as T, E, R, L, P and weather, load, speed limits, etc.), and edges represent causal directions. Formula (edge ​​importance = absolute value of average causal effect):

[0101] e i→j Let be the edge weight from variable i to j; ACE j←i For average causal effect; X i δ is the causal variable; δ is the coefficient of performance on X. i The extent of intervention; Y j For the outcome variable, the thermal matrix and color mapping (matrix = heatmap), let e i→j Represent it as a matrix H, and color it according to strength. Formula (normalization and coloring):

[0102] The edge strength is normalized to 0–1; e min ,e max C represents the minimum / maximum edge strength. ij The color of the block represents the path explanation score (which decision chain is the most "critical"). The total contribution of the causal path from "strategy parameters" to "outcome metrics" is calculated and marked with a thick line and a bright color. Formula (path score): π represents a directed path from the starting point to the ending point; This represents the normalized strength of the edge. Node layout (clear arrangement, no knots): Use spectral layout to map nodes to the 2D plane, avoiding intersections and crowding. Formula (Graph Laplacian and Eigenvector Embedding): L = DA, x = u2, y = u3, where L is the graph Laplacian matrix; D is the degree matrix; A is the (symmetric) adjacency matrix; u2 and u3 are the second and third smallest eigenvectors of L (the first is an all-1 vector). Interactive filtering (only show "important" relationships): Only display based on the threshold τ. The edge; the slider is adjusted in real time.

[0103] Natural Language Generation Explanation: Utilizing natural language generation technology, complex path planning logic, risk prediction results, and anomaly handling processes are converted into human-readable explanatory text. When an AGV detours to avoid a certain channel, the system can generate an explanation stating that the channel has a high probability of congestion in the next few minutes, thus choosing an alternative path. Regarding anomaly handling, the system can explain in natural language what tiered response measures were triggered, such as "Communication delay detected, task transferred to backup AGV." Content selection (choosing the "most useful" fact within the word budget): Candidate facts k (e.g., "Channel C congestion 1.8×", "Rerouting reduces energy consumption by 12%) have relevance scores and word cost, solving a small selection problem. Formula (0 / 1 selection, word budget B): y k To determine whether to select the k-th fact; s k Assign it an importance score (weighted by risk, return, and user attention); k B represents the word count for transcribing this fact; B is the total word count budget. Document micro-planning (information order and coherence): adopts a rhetorical structure of "current situation → cause → measure → effect," penalizing jumps and repetitions. Formula (sequential cost):

[0104] {i follows j}, For the set of dependency pairs that should "priorize" (e.g., "cause precedes measure"); ω ij The penalty for violating this order; 1{·} is the indicator function. Sentence generation (template + numerical standardization): Use an auditable template to "slot" variables, and standardize percentages and units. Formula (rounding and unit unification): Pct(x)=round(100·x,1)% Pct(x) is the percentage of x converted to one decimal place; Unit(q) normalizes the physical quantity q to the reference unit u. base Include the unit symbol u (e.g., m, s, kWh). Describe the uncertainty and confidence level (a more honest output), providing the mean ± confidence interval for the predicted value, or using quantiles. Formula (confidence interval for a normal approximation): To estimate the mean; z is the standard deviation; α / 2 Confidence coefficient (e.g., 1.96 for 95%). Overall quality objective (high coverage, few contradictions, length compliance): Optimized using the comprehensive loss during training / selection. Formula (NLG loss): Losses due to the failure to cover material facts; Punishment for discrepancies with the data; Penalties are applied for excessively long / short sentences; weights are λ1, λ2, λ3 > 0. Evidence is traceable (click text → locate evidence): The alignment relationship between sentences and evidence is stored in an index table (click to highlight risk cells / path segments). Formula (Sentence-Evidence Matching Score): Align(s, e) = sim(φ(s), ψ(e)), where Align(s, e) is the similarity between sentence s and evidence item e; φ and ψ are the embedding functions of the sentence / evidence (such as TF-IDF or a small encoder); sim is the similarity (cosine / dot product).

[0105] Operation and Maintenance Interaction and Support: Provides an interactive operation interface, allowing operation and maintenance personnel to monitor tasks, review risks, and adjust policies through a visual interface; the system can record the intervention operations of operation and maintenance personnel and combine them with the causal explanation chain to form a traceable operation and maintenance log; supports multi-user permission management to ensure that operation and maintenance personnel at different levels can obtain the corresponding visual data and natural language reports according to their permission scope.

[0106] The visualization and operation and maintenance support module works continuously during system operation, collaborating with the cross-modal digital twin and risk simulation module, and the anomaly detection and robust protection module to output AGV operating status, path planning decisions, and anomaly explanations in real time. Through 3D digital twin visualization, causal heatmaps, and natural language generation, this module transforms complex technical logic into intuitive, interactive, and understandable information, significantly improving operation and maintenance efficiency and decision controllability, enabling non-technical personnel to participate in system management and operation.

[0107] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A warehouse AGV dynamic path planning system based on multi-objective optimization, characterized in that: include, The multi-source sensing and data acquisition module is used to collect AGV operating status, working environment and external traffic information and generate standardized feature vectors; The cross-modal digital twin and risk simulation module is used to construct a virtual twin of the warehouse and external transportation, and to perform risk prediction and simulation driven by cross-modal perception data; the multi-objective intelligent optimization module is used to generate candidate route solutions under multiple objectives such as transportation time, energy consumption, safety, equipment life and task priority. The AGV collaborative scheduling module is used to achieve dynamic collaborative scheduling of multiple AGVs through game theory and reinforcement learning mechanisms; the self-evolutionary learning and transfer module is used to improve the system's adaptability in different scenarios through learning and transfer mechanisms; the anomaly detection and robust protection module is used to detect hardware failures, communication anomalies, path congestion and sudden events, and execute hierarchical response strategies. The visualization and operation and maintenance support module is used to provide digital twin visualization, causal heatmaps, and natural language generation and explanation.

2. The warehouse AGV dynamic path planning system based on multi-objective optimization according to claim 1, characterized in that: The multi-source sensing and data acquisition module collects data through lidar, vision, RFID, millimeter-wave radar and acoustic sensors, and combines adaptive Kalman filtering and Bayesian optimization to achieve data fusion and compensation.

3. The warehouse AGV dynamic path planning system based on multi-objective optimization according to claim 1, characterized in that: The cross-modal digital twin and risk simulation module utilizes dynamic graph convolutional networks for risk propagation modeling, combines Transformer-XL and spatiotemporal graph attention mechanisms to achieve multi-scale risk prediction, and uses generative adversarial networks for extreme condition simulation.

4. The warehouse AGV dynamic path planning system based on multi-objective optimization according to claim 1, characterized in that: The multi-objective intelligent optimization module uses hierarchical multi-objective reinforcement learning and evolutionary algorithms to generate Pareto optimal solution sets, and combines causal constraints and fuzzy multi-objective optimization to improve the rationality and robustness of the solutions.

5. The warehouse AGV dynamic path planning system based on multi-objective optimization according to claim 1, characterized in that: The AGV collaborative scheduling module uses Stackelberg game, coalition game and signaling game mechanisms to allocate paths and tasks, and introduces Shapley value allocation and cooperative transfer game in the scheduling to achieve fairness and congestion self-healing.

6. The warehouse AGV dynamic path planning system based on multi-objective optimization according to claim 1, characterized in that: The self-evolutionary learning and transfer module includes meta-learning, graph embedding transfer, few-shot learning, cross-scenario knowledge distillation, lifelong learning, and self-supervised comparative learning mechanisms.

7. The warehouse AGV dynamic path planning system based on multi-objective optimization according to claim 1, characterized in that: The anomaly detection and robust protection module detects anomalies by monitoring message latency, jitter, and packet loss rate when detecting communication delays. Under moderate anomalies, it performs task reallocation, and under severe anomalies, it triggers redundant AGV takeover.

8. The warehouse AGV dynamic path planning system based on multi-objective optimization according to claim 1, characterized in that: The visualization and operation and maintenance support module adopts a logical structure of current situation-cause-measure-effect when generating natural language explanations, and displays the causal explanation chain and risk heat map through an interactive interface.

Images