Emergency resource cooperative scheduling optimization method and system

Abstract

The invention discloses an emergency resource collaborative scheduling optimization method and system, and belongs to the technical field of emergency management and intelligent scheduling. The knowledge graph is used for carrying out structured modeling and storage on emergency resources, emergencies, environment information and attributes and relationships among the emergency resources, the emergencies and the environment information; in response to a monitored emergency, acquiring real-time multi-source data, and performing state perception and logical reasoning by using the knowledge graph to generate a feasible scheduling scheme candidate set conforming to a preset business rule; taking the feasible scheduling scheme candidate set as an action constraint, and inputting the feasible scheduling scheme candidate set into a pre-trained multi-agent reinforcement learning model; and the multi-agent reinforcement learning model outputs an optimal cooperative scheduling scheme based on the current environment state and the action constraint. According to the method, the optimal resource allocation scheme can be quickly generated in a complex emergency scene, and the intelligent level and rescue efficiency of cross-department emergency response are improved.

Description

Technical Field

[0001] This invention relates to the field of emergency management and intelligent dispatching technology, specifically to an emergency resource collaborative dispatching optimization method and system. Background Technology

[0002] In emergency scenarios such as natural disasters and public health emergencies, the efficiency of emergency supplies and transportation capacity scheduling directly affects the effectiveness of rescue efforts and the protection of people's livelihoods. The current emergency dispatch system suffers from significant pain points: departments often operate independently, material reserve data is not shared, and transportation resources are managed in a decentralized manner, creating "information silos," preventing the command center from grasping the real-time status of available resources. For example, after an emergency occurs, the tent reserve data of the civil affairs department and the freight vehicle dispatch system of the transportation department cannot be linked in real time, leading to resource waste such as "supplies available but no transportation capacity" or "transport capacity running empty." Manually calculating the optimal dispatch plan is time-consuming, and delayed decisions can lead to missed opportunities for crucial rescue efforts. Dispatch instructions are often unidirectional, lacking the ability for collaborative negotiation and dynamic adjustment among multiple rescue units en route. For example, dispatched rescue forces may be stuck on the road due to congestion, failing to achieve collective intelligence such as "if one road is blocked, another should be taken." Manual decision-making struggles to quickly quantify and weigh multiple conflicting objectives such as "fastest arrival," "covering the most injured," and "lowest resource consumption."

[0003] Existing cross-departmental emergency resource dispatch methods mostly rely on manual coordination or fixed process-based dispatch, facing three core challenges: First, resource information is fragmented, with personnel, vehicle, and material data from various departments (public security, fire, medical, etc.) stored separately, lacking a unified and interconnected view; second, they have poor adaptability to dynamic environments, with traditional operations optimization methods struggling to cope with real-time scenarios such as sudden changes in road conditions and updates to resource status; and third, cross-departmental collaboration is inefficient, making it difficult to dynamically balance conflicting objectives such as "response speed - resource utilization - coverage," such as prioritizing rapid arrival for firefighters versus prioritizing resource preservation for medical personnel.

[0004] While existing technologies such as knowledge graphs can construct resource association networks, they are often used for static resource management and lack dynamic decision-making capabilities. Multi-agent reinforcement learning (MARL) has the potential for collaborative decision-making, but its convergence speed is slow under complex resource constraints, and the interpretability of its decisions is poor. Therefore, there is an urgent need to build an intelligent scheduling mechanism that can break down departmental barriers, dynamically adapt to complex disaster situations, and achieve multi-objective collaborative optimization. Summary of the Invention

[0005] The technical objective of this invention is to address the above-mentioned shortcomings by providing an emergency resource collaborative scheduling optimization method and system. This method and system can solve the problems of information silos, decision-making delays, lack of collaboration, and multi-objective trade-offs in existing emergency scheduling. It can quickly generate optimal resource allocation schemes in complex emergency scenarios, improve the intelligence level and rescue efficiency of cross-departmental emergency response, and provide key technical support for the efficient handling of major emergencies.

[0006] The technical solution adopted by this invention to solve its technical problem is:

[0007] An optimization method for collaborative scheduling of emergency resources, the implementation of which includes the following steps:

[0008] Step S1: Construct a knowledge graph in the emergency response domain. The knowledge graph is used to structure and store emergency resources, emergencies, environmental information, and their attributes and relationships.

[0009] Step S2: In response to the detection of a sudden event, real-time multi-source data is acquired, and the knowledge graph is used for state perception and logical reasoning to generate a set of feasible scheduling scheme candidates that conform to preset business rules;

[0010] Step S3: The set of feasible scheduling scheme candidates is used as action constraints and input into the pre-trained multi-agent reinforcement learning model; the multi-agent reinforcement learning model takes minimizing the overall emergency response time as the global optimization objective, and multiple agents participate in the decision-making, with each agent representing a rescue unit;

[0011] Step S4: The multi-agent reinforcement learning model outputs an optimal cooperative scheduling scheme based on the current environmental state and the action constraints;

[0012] Step S5: The optimal collaborative scheduling scheme is sent to the corresponding rescue unit for execution, and environmental changes are continuously monitored during the execution process to trigger dynamic rescheduling.

[0013] This method constructs an intelligent scheduling and decision-making framework based on "knowledge graph + multi-agent reinforcement learning." Knowledge graphs provide structured environmental cognition, while multi-agent reinforcement learning enables dynamic collaborative decision-making. By combining the symbolic approach (precise knowledge representation) of knowledge graphs with the behaviorist approach (learning optimal strategies through interaction) of multi-agent reinforcement learning, it achieves precise, rapid, and collaborative scheduling of cross-departmental emergency resources.

[0014] Furthermore, the knowledge graph constructed in step S1 includes the following entity types: rescue units, personnel, supplies, facilities, events, and environment;

[0015] The knowledge graph uses relation edges to represent the membership, location, need, and influence relationships between entities.

[0016] Furthermore, in step S2, the state perception and logical reasoning are performed using the knowledge graph, specifically including:

[0017] Based on the type and severity of the event, and using predefined business rules in the knowledge graph, the type, quantity, and skill requirements of the necessary rescue resources are inferred.

[0018] Based on the real-time status and location attributes of rescue units, spatial and status queries are performed in the knowledge graph to filter out a set of available and reachable rescue units;

[0019] Based on the above reasoning and query results, a candidate set of feasible scheduling schemes is generated.

[0020] Furthermore, the multi-agent reinforcement learning model described in step S3 adopts a centralized training and distributed execution architecture:

[0021] During the training phase, an Actor-Critic-based multi-agent reinforcement learning algorithm is adopted, in which a centralized critic network provides global value evaluation for the agent network of each agent to update the network parameters.

[0022] During the execution phase, each agent makes distributed decisions based solely on its own local observations and a global state summary from the knowledge graph.

[0023] Furthermore, the reward function R of the multi-agent reinforcement learning model is designed as a multi-objective weighted form:

[0024] R=ω1·R time +ω2·R coverage +ω3·R cost +ω4·R coordination

[0025] Among them, R time For reward items that are negatively correlated with total response time, R coverage For rewards that are positively correlated with the area of ​​a successful rescue, R cost For reward items that are negatively correlated with resource consumption costs, R coordination The reward is positively correlated with the efficiency of cooperation between agents, and ω1 to ω4 are configurable weight coefficients.

[0026] Furthermore, the step S4 of outputting an optimal collaborative scheduling scheme specifically refers to the multi-agent reinforcement learning model searching within the feasible solution space defined by the action constraints and outputting a Pareto optimal solution, which simultaneously specifies the travel path, task allocation, and execution order of each rescue unit.

[0027] This invention also claims an emergency resource collaborative scheduling and optimization system, comprising:

[0028] The dynamic resource knowledge graph construction and updating module is used to build knowledge graphs in the emergency response field.

[0029] A multi-agent reinforcement learning model construction module, wherein the multi-agent reinforcement learning model is used to output the optimal cooperative scheduling scheme based on the current environmental state and the action constraints;

[0030] The online execution module is used to issue and execute the optimal collaborative scheduling scheme, and continuously monitor environmental changes during the execution process to trigger dynamic rescheduling;

[0031] The system specifically achieves optimized collaborative scheduling of emergency resources through the methods described above.

[0032] Furthermore, the dynamic resource knowledge graph construction and update module includes the following: entity types, including rescue units, personnel, materials, facilities, events, and environment; relationship types between entities; attribute types between entities; and the knowledge graph enables real-time query, rule reasoning, and state awareness functions.

[0033] The multi-agent reinforcement learning model construction module includes:

[0034] (1) Definition of intelligent agent: Each rescue unit (such as each ambulance or each fire brigade) in the knowledge graph is abstracted into an intelligent agent, and the entities and relationships such as road network, event point, and resource point in the knowledge graph are transformed into a multi-agent reinforcement learning environment;

[0035] (2) State Space: The agent's observed state is derived from the real-time query results of the dynamic resource knowledge graph, including:

[0036] Its own status, including location, speed, and remaining resources;

[0037] Event information, including type, level, and location;

[0038] Environmental information, including traffic flow, weather, and road closures;

[0039] Other agent information, including the location and status of nearby rescue units;

[0040] (3) Action Space: The actions that each agent can choose include:

[0041] Movement direction: Select the next section of road to travel on;

[0042] Task selection: Choose to go to a specific sub-task location;

[0043] Resource allocation: Deciding whether to carry or unload certain supplies;

[0044] (4) Reward function: Designed as a multi-objective weighted function:

[0045] (5) Algorithm selection: The MAPPO (Multi-Agent Proximal Policy Optimization) algorithm is adopted to learn how to collaboratively select the best path and task; the reasoning ability of the dynamic resource knowledge graph is used to provide expert prior knowledge for multi-agent reinforcement learning; for example, the dynamic resource knowledge graph infers that "event type = chemical leak" requires "professional chemical defense troops", and the multi-agent reinforcement learning model will use this as a constraint, which will greatly reduce the action exploration space and accelerate training convergence;

[0046] The online execution module includes the following execution process:

[0047] Phase 1, Initialization and Perception: Sensors detect events (such as traffic accidents), data is transmitted to the system, and the knowledge graph updates event information and its associated entities in real time;

[0048] Phase Two, Knowledge Graph Reasoning: The dynamic resource knowledge graph first performs rapid reasoning to generate a set of feasible scheduling schemes that conform to business rules. This set of schemes serves as the candidate action space for multi-agent reinforcement learning, significantly reducing the search dimensions.

[0049] Phase 3, MARL Collaborative Decision Making: Each agent (rescue unit) obtains the current state (from the dynamic resource knowledge graph), and the multi-agent reinforcement learning model receives the "candidate action set". Based on this, it performs fast and refined multi-objective optimization, calculates an optimal or Pareto optimal scheduling scheme, balances multiple objectives, and outputs the final scheduling scheme, including which units to assign, along which path, and what tasks to perform.

[0050] Phase 4, Dynamic Execution and Adjustment: The plan is distributed to various terminals (in-vehicle navigation, mobile APP), and each intelligent agent begins execution. The system continuously monitors environmental changes. If changes occur, the dynamic resource knowledge graph (KG) is updated in real time, and the multi-agent reinforcement learning model quickly re-decides and generates adjustment instructions to achieve dynamic rescheduling.

[0051] The present invention also claims an emergency resource collaborative scheduling optimization device, comprising: at least one memory and at least one processor;

[0052] The at least one memory is used to store a machine-readable program;

[0053] The at least one processor is used to call the machine-readable program to implement the above method.

[0054] The present invention also claims a computer-readable medium storing computer instructions that, when executed by a processor, implement the above-described method.

[0055] Compared with the prior art, the emergency resource collaborative scheduling optimization method and system of the present invention have the following advantages:

[0056] The proposed emergency resource collaborative scheduling optimization method based on knowledge graphs and multi-agent reinforcement learning demonstrates significant benefits in terms of decision-making speed and quality.

[0057] 1. Significantly improved decision-making speed: From "minute-level" manual decision-making to "second-level" automatic decision-making, seizing the golden rescue time.

[0058] 2. Scientifically optimal decision quality: Simultaneously optimize multiple objectives to find the global optimal solution, rather than the local optimal or empirical solution.

[0059] 3. Decision interpretability: Dynamic resource knowledge graph (KG) provides a logical chain for decision-making. By tracing the entity relationships in the knowledge graph, the decision-making of multi-agent reinforcement learning (MARL) is transformed from a "black box" to a "white box," making it understandable and trustworthy for human commanders.

[0060] 4. Dynamic environment adaptation capability: Multi-agent reinforcement learning (MARL) enables the system to handle new scenarios not seen during training and has the ability to continuously learn and evolve.

[0061] 5. Achieving True Collaboration: Knowledge graphs break down information barriers, and multi-agent reinforcement learning (MARL) enables goal collaboration. Each rescue unit is no longer an independent pawn, but an intelligent agent capable of dynamically adjusting based on the actions of others and the overall state, avoiding resource clustering or gap coverage. Attached Figure Description

[0062] Figure 1 This is a flowchart illustrating the emergency resource collaborative scheduling optimization method provided in an embodiment of the present invention. Detailed Implementation

[0063] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0064] This invention provides an emergency resource collaborative scheduling optimization method, the implementation of which includes the following steps:

[0065] Step S1: Construct a knowledge graph in the emergency response domain. The knowledge graph is used to structure and store emergency resources, emergencies, environmental information, and their attributes and relationships.

[0066] Step S2: In response to the detection of a sudden event, real-time multi-source data is acquired, and the knowledge graph is used for state perception and logical reasoning to generate a set of feasible scheduling scheme candidates that conform to preset business rules;

[0067] Step S3: The set of feasible scheduling scheme candidates is used as action constraints and input into the pre-trained multi-agent reinforcement learning model; the multi-agent reinforcement learning model takes minimizing the overall emergency response time as the global optimization objective, and multiple agents participate in the decision-making, with each agent representing a rescue unit;

[0068] Step S4: The multi-agent reinforcement learning model outputs an optimal cooperative scheduling scheme based on the current environmental state and the action constraints;

[0069] Step S5: The optimal collaborative scheduling scheme is sent to the corresponding rescue unit for execution, and environmental changes are continuously monitored during the execution process to trigger dynamic rescheduling.

[0070] The knowledge graph constructed in step S1 includes the following entity types: rescue units, personnel, supplies, facilities, events, and environment.

[0071] The knowledge graph uses relation edges to represent the membership, location, need, and influence relationships between entities.

[0072] In step S2, state perception and logical reasoning are performed using the knowledge graph, specifically including:

[0073] Based on the type and severity of the event, and using predefined business rules in the knowledge graph, the type, quantity, and skill requirements of the necessary rescue resources are inferred.

[0074] Based on the real-time status and location attributes of rescue units, spatial and status queries are performed in the knowledge graph to filter out a set of available and reachable rescue units;

[0075] Based on the above reasoning and query results, a candidate set of feasible scheduling schemes is generated.

[0076] The multi-agent reinforcement learning model described in step S3 adopts a centralized training and distributed execution architecture:

[0077] During the training phase, an Actor-Critic-based multi-agent reinforcement learning algorithm is adopted, in which a centralized critic network provides global value evaluation for the agent network of each agent to update the network parameters.

[0078] During the execution phase, each agent makes distributed decisions based solely on its own local observations and a global state summary from the knowledge graph.

[0079] The reward function R of the multi-agent reinforcement learning model is designed as a multi-objective weighted form:

[0080] R=ω1·R time +ω2·R coverage +ω3·R cost +ω4·R coordination

[0081] Among them, R time For reward items that are negatively correlated with total response time, R coverage For rewards that are positively correlated with the area of ​​a successful rescue, R cost For reward items that are negatively correlated with resource consumption costs, R coordination The reward is positively correlated with the efficiency of cooperation between agents, and ω1 to ω4 are configurable weight coefficients.

[0082] The step S4 of outputting an optimal collaborative scheduling scheme specifically refers to the multi-agent reinforcement learning model searching within the feasible solution space defined by the action constraints and outputting a Pareto optimal solution, which simultaneously specifies the travel path, task allocation, and execution order of each rescue unit.

[0083] This method constructs an intelligent scheduling and decision-making framework based on "knowledge graph + multi-agent reinforcement learning." Knowledge graphs provide structured environmental cognition, while multi-agent reinforcement learning enables dynamic collaborative decision-making. By combining the symbolic approach (precise knowledge representation) of knowledge graphs with the behaviorist approach (learning optimal strategies through interaction) of multi-agent reinforcement learning, it achieves precise, rapid, and collaborative scheduling of cross-departmental emergency resources.

[0084] The specific implementation of this method is as follows:

[0085] 1. Steps for constructing and updating a dynamic resource knowledge graph (KG), including:

[0086] (1) Entity type, for example:

[0087] Rescue units: ambulances, fire trucks, police cars, tow trucks, emergency supply depots, etc.;

[0088] Personnel: Doctors, firefighters, traffic police, etc.

[0089] Supplies: snow melting agent, anti-skid sand, first aid medicines, etc.

[0090] Facilities: Hospitals, gas stations, highway exits, etc.;

[0091] Events: Traffic accidents, landslides, etc.;

[0092] Environment: roads, bridges, weather, etc.

[0093] (2) Relationship type, for example:

[0094] Located at: Rescue Unit → Location;

[0095] Handling: Fire truck → Fire;

[0096] Required: Event → Resource Type;

[0097] Congestion: Road → Real-time Status;

[0098] Carrying: Vehicles → Supplies.

[0099] (3) Attribute type, for example:

[0100] Rescue units: real-time location, capacity, skills, status, etc.;

[0101] Personnel: Department, professional skills, current status, location coordinates, etc.;

[0102] Material milestones: category, quantity, storage location, shelf life, allocation permissions, etc.

[0103] Event: type, level, location, number of casualties, etc.

[0104] The functions of the knowledge graph:

[0105] Real-time query: such as finding all fire departments that can handle hazardous chemical leaks and are reachable within 10 minutes;

[0106] Rule-based reasoning: Resource capability reasoning is implemented based on SWRL rules, such as an ambulance equipped with ECMO equipment ∩ within 5km of the incident site → can handle the transfer of critically ill patients;

[0107] Status awareness: Real-time collection of resource status changes through IoT terminals and dynamic updating of entity attributes.

[0108] 2. Construction of Multi-Agent Reinforcement Learning (MARL) models, including:

[0109] (1) Definition of Agent:

[0110] Abstract each rescue unit in the knowledge graph (such as each ambulance or fire brigade) into an intelligent agent, and transform the entities and relationships in the knowledge graph, such as road networks, event points, and resource points, into a MARL environment.

[0111] (2) State space:

[0112] The agent's observed state is derived from the real-time query results of the KG, including:

[0113] Self-status: position, speed, remaining resources;

[0114] Event information: type, level, location;

[0115] Environmental information: traffic flow, weather, road closures;

[0116] Other agent information: Location and status of nearby rescue units.

[0117] (3) Action space:

[0118] Each agent can choose from the following actions:

[0119] Movement direction: Select the next section of road to travel on;

[0120] Task selection: Choose to go to a specific sub-task location;

[0121] Resource allocation: Decide whether to carry or unload a certain type of material.

[0122] (4) Reward function:

[0123] Designed as a multi-objective weighted function:

[0124] R=ω1·R time +ω2·R coverage +ω3·R cost +ω4·R coordination ;

[0125] R time = - Total rescue time

[0126] R coverage = + Number of people successfully rescued

[0127] R cost = - Total resource consumption

[0128] R coordination = - Number of conflicts between rescue units.

[0129] (5) Algorithm selection:

[0130] The MAPPO (Multi-Agent Proximal Policy Optimization) algorithm is employed to learn how to collaboratively select the optimal path and task. The reasoning capabilities of the KG (Knowledge Group) are leveraged to provide MARL with expert prior knowledge. For example, if the KG infers that "event type = chemical leak" requires "specialized chemical defense forces," MARL will use this as a constraint, significantly narrowing the action exploration space and accelerating training convergence.

[0131] 3. Execute online:

[0132] (1) Phase 1: Initialization and Perception

[0133] When a sensor detects an event (such as a traffic accident), the data is transmitted to the system, and the knowledge graph updates the event information and its associated entities in real time.

[0134] (2) Phase Two: Knowledge Graph Reasoning

[0135] KG first performs rapid reasoning to generate a set of feasible scheduling schemes that conform to business rules. This set of schemes serves as the candidate action space for MARL, significantly reducing the search dimensions.

[0136] (3) Phase Three: MARL Collaborative Decision Making

[0137] Each agent (rescue unit) obtains its current state (from KG). MARL receives the "candidate action set" and performs rapid and refined multi-objective optimization based on it, calculating an optimal or Pareto optimal scheduling scheme, balancing multiple objectives, and outputting the final scheduling scheme: which units to assign, along which path, and what task to perform.

[0138] (4) Phase Four: Dynamic Execution and Adjustment

[0139] The solution is distributed to various terminals (in-vehicle navigation, mobile APP), and each intelligent agent begins execution. The system continuously monitors environmental changes. If changes occur, the KG updates in real time, and the MARL quickly re-decides and generates adjustment instructions to achieve dynamic rescheduling.

[0140] This method constructs an emergency resource knowledge graph that integrates multi-source data: It builds an emergency domain knowledge graph that integrates cross-departmental static knowledge and dynamic real-time data. This graph provides structured modeling of entities such as rescue units, personnel, materials, facilities, events, and the environment, along with their complex relationships, and incorporates a built-in business rule reasoning engine. Based on the graph, a multi-agent reinforcement learning simulation environment is constructed: each rescue unit is abstracted as an agent, and a multi-agent reinforcement learning model is built using the knowledge graph as a virtual environment. This model is driven by a multi-objective reward function that integrates time efficiency, rescue effectiveness, resource cost, and coordination. Before decision-making, the knowledge graph performs rapid logical reasoning, generating a candidate set of feasible scheduling schemes that conform to business rules. This guides and constrains the search space of the multi-agent reinforcement learning, transforming reinforcement learning from time-consuming blind exploration into efficient targeted optimization. Based on this, the multi-agent reinforcement learning model performs rapid and refined multi-objective collaborative decision-making, outputting the optimal scheduling instruction. The two form a hybrid intelligent decision-making paradigm of "knowledge reasoning first, collaborative decision optimization." By using knowledge graph reasoning to guide the reinforcement learning training and decision-making process, it achieves the rapid generation of optimal scheduling schemes that balance time, effectiveness, and cost. It also senses emergencies in real time during the decision-making and execution process and updates the knowledge graph, dynamically rescheduling according to environmental changes.

[0141] This method solves the problems of information silos, decision-making delays, lack of coordination and multi-objective balancing in existing emergency dispatch. It can quickly generate optimal resource allocation plans in complex emergency scenarios, improve the intelligence level and rescue efficiency of cross-departmental emergency response, and provide key technical support for the efficient handling of major emergencies.

[0142] This invention also provides an emergency resource collaborative scheduling optimization system, comprising:

[0143] The dynamic resource knowledge graph construction and updating module is used to build knowledge graphs in the emergency response field.

[0144] A multi-agent reinforcement learning model construction module, wherein the multi-agent reinforcement learning model is used to output the optimal cooperative scheduling scheme based on the current environmental state and the action constraints;

[0145] The online execution module is used to issue and execute the optimal collaborative scheduling scheme, and continuously monitor environmental changes during the execution process to trigger dynamic rescheduling;

[0146] Specifically, the system achieves emergency resource collaborative scheduling optimization through the emergency resource collaborative scheduling optimization method described in the above embodiments.

[0147] The dynamic resource knowledge graph construction and update module includes the following: Knowledge graph construction:

[0148] (1) Entity type, for example:

[0149] Rescue units: ambulances, fire trucks, police cars, tow trucks, emergency supply depots, etc.;

[0150] Personnel: Doctors, firefighters, traffic police, etc.

[0151] Supplies: snow melting agent, anti-skid sand, first aid medicines, etc.

[0152] Facilities: Hospitals, gas stations, highway exits, etc.;

[0153] Events: Traffic accidents, landslides, etc.;

[0154] Environment: roads, bridges, weather, etc.

[0155] (2) Relationship type, for example:

[0156] Located at: Rescue Unit → Location;

[0157] Handling: Fire truck → Fire;

[0158] Required: Event → Resource Type;

[0159] Congestion: Road → Real-time Status;

[0160] Carrying: Vehicles → Supplies.

[0161] (3) Attribute type, for example:

[0162] Rescue units: real-time location, capacity, skills, status, etc.;

[0163] Personnel: Department, professional skills, current status, location coordinates, etc.;

[0164] Material milestones: category, quantity, storage location, shelf life, allocation permissions, etc.

[0165] Event: type, level, location, number of casualties, etc.

[0166] The functions of the knowledge graph:

[0167] Real-time query: such as finding all fire departments that can handle hazardous chemical leaks and are reachable within 10 minutes;

[0168] Rule-based reasoning: Resource capability reasoning is implemented based on SWRL rules, such as an ambulance equipped with ECMO equipment ∩ within 5km of the incident site → can handle the transfer of critically ill patients;

[0169] Status awareness: Real-time collection of resource status changes through IoT terminals and dynamic updating of entity attributes.

[0170] The multi-agent reinforcement learning (MARL) model building module includes:

[0171] (1) Definition of Agent:

[0172] Abstract each rescue unit in the knowledge graph (such as each ambulance or fire brigade) into an intelligent agent, and transform the entities and relationships in the knowledge graph, such as road networks, event points, and resource points, into a MARL environment.

[0173] (2) State space:

[0174] The agent's observed state is derived from the real-time query results of the KG, including:

[0175] Self-status: position, speed, remaining resources;

[0176] Event information: type, level, location;

[0177] Environmental information: traffic flow, weather, road closures;

[0178] Other agent information: Location and status of nearby rescue units.

[0179] (3) Action space:

[0180] Each agent can choose from the following actions:

[0181] Movement direction: Select the next section of road to travel on;

[0182] Task selection: Choose to go to a specific sub-task location;

[0183] Resource allocation: Decide whether to carry or unload a certain type of material.

[0184] (4) Reward function:

[0185] Designed as a multi-objective weighted function:

[0186] R=ω1·R time +ω2·R coverage +ω3·R cost +ω4·R coordination ;

[0187] R time = - Total rescue time

[0188] R coverage = + Number of people successfully rescued

[0189] R cost = - Total resource consumption

[0190] R coordination = - Number of conflicts between rescue units.

[0191] (5) Algorithm selection:

[0192] The MAPPO (Multi-Agent Proximal Policy Optimization) algorithm is employed to learn how to collaboratively select the optimal path and task. The reasoning capabilities of the KG (Knowledge Group) are leveraged to provide MARL with expert prior knowledge. For example, if the KG infers that "event type = chemical leak" requires "specialized chemical defense forces," MARL will use this as a constraint, significantly narrowing the action exploration space and accelerating training convergence.

[0193] The online execution module, the execution process includes:

[0194] (1) Phase 1: Initialization and Perception

[0195] When a sensor detects an event (such as a traffic accident), the data is transmitted to the system, and the knowledge graph updates the event information and its associated entities in real time.

[0196] (2) Phase Two: Knowledge Graph Reasoning

[0197] KG first performs rapid reasoning to generate a set of feasible scheduling schemes that conform to business rules. This set of schemes serves as the candidate action space for MARL, significantly reducing the search dimensions.

[0198] (3) Phase Three: MARL Collaborative Decision Making

[0199] Each agent (rescue unit) obtains its current state (from KG). MARL receives the "candidate action set" and performs rapid and refined multi-objective optimization based on it, calculating an optimal or Pareto optimal scheduling scheme, balancing multiple objectives, and outputting the final scheduling scheme: which units to assign, along which path, and what task to perform.

[0200] (4) Phase Four: Dynamic Execution and Adjustment

[0201] The solution is distributed to various terminals (in-vehicle navigation, mobile APP), and each intelligent agent begins execution. The system continuously monitors environmental changes. If changes occur, the KG updates in real time, and the MARL quickly re-decides and generates adjustment instructions to achieve dynamic rescheduling.

[0202] This invention also provides an emergency resource collaborative scheduling optimization device, comprising: at least one memory and at least one processor;

[0203] The at least one memory is used to store a machine-readable program;

[0204] The at least one processor is used to call the machine-readable program to implement the emergency resource collaborative scheduling optimization method described in the above embodiments.

[0205] This invention also provides a computer-readable medium storing computer instructions, which, when executed by a processor, implement the emergency resource collaborative scheduling optimization method described in the above embodiments. Specifically, a system or apparatus equipped with a storage medium may be provided, on which software program code implementing the functions of any of the embodiments described above is stored, and the computer (or CPU or MPU) of the system or apparatus can read and execute the program code stored in the storage medium.

[0206] In this case, the program code read from the storage medium can itself implement the function of any of the above embodiments, and therefore the program code and the storage medium storing the program code constitute part of the present invention.

[0207] Examples of storage media used to provide program code include floppy disks, hard disks, magneto-optical disks, optical disks (such as CD-ROM, CD-R, CD-RW, DVD-ROM, DVD-RAM, DVD-RW, DVD+RW), magnetic tapes, non-volatile memory cards, and ROMs. Alternatively, program code can be downloaded from a server computer via a communication network.

[0208] Furthermore, it should be clear that not only can the program code read by the computer be executed, but also the operating system or other components operating on the computer can be instructed based on the program code to perform some or all of the actual operations, thereby realizing the function of any of the embodiments described above.

[0209] Furthermore, it is understood that the program code read from the storage medium is written to the memory set in the expansion board inserted into the computer or to the memory set in the expansion unit connected to the computer. Then, based on the instructions of the program code, the CPU or other components installed on the expansion board or expansion unit execute some and all of the actual operations, thereby realizing the function of any of the embodiments described above.

[0210] The present invention has been shown and described in detail above with reference to the accompanying drawings and preferred embodiments. However, the present invention is not limited to these disclosed embodiments. Based on the above multiple embodiments, those skilled in the art will know that more embodiments of the present invention can be obtained by combining the code review methods in the different embodiments above. These embodiments are also within the protection scope of the present invention.

Claims

1. An emergency resource collaborative scheduling optimization method, characterized in that, The method comprises the following steps: Step S1: constructing a knowledge graph of the emergency field, which is used for structurally modeling and storing emergency resources, emergencies, environmental information, and their attributes and relationships; Step S2: in response to monitoring an emergency, acquiring real-time multi-source data, and using the knowledge graph for state perception and logical reasoning to generate a set of feasible scheduling scheme candidates conforming to preset business rules; Step S3: inputting the set of feasible scheduling scheme candidates as action constraints into a pre-trained multi-agent reinforcement learning model; the multi-agent reinforcement learning model takes minimizing the overall emergency response time as a global optimization goal, and multiple agents jointly participate in decision-making, with each agent representing a rescue unit; Step S4: the multi-agent reinforcement learning model outputs an optimal collaborative scheduling scheme based on the current environmental state and the action constraints; Step S5: the optimal collaborative scheduling scheme is issued to the corresponding rescue units for execution, and environmental changes are continuously monitored during the execution process to trigger dynamic rescheduling.

2. The method of claim 1, wherein, The knowledge graph constructed in step S1 comprises the following entity types: rescue units, personnel, materials, facilities, events, and environments; The knowledge graph represents the membership, location, need, and influence relationships between entities through relationship edges.

3. The method of claim 1, wherein, In step S2, the knowledge graph is used for state perception and logical reasoning, which specifically comprises: According to the type and level of the event, the type, quantity, and skill requirements of the required rescue resources are inferred based on the pre-defined business rules in the knowledge graph; According to the real-time state and location attributes of the rescue units, spatial and state queries are performed in the knowledge graph to filter out a set of available and reachable rescue units; The above reasoning and query results are integrated to generate the set of feasible scheduling scheme candidates.

4. The method of claim 1, wherein, The multi-agent reinforcement learning model in step S3 adopts a centralized training and distributed execution architecture: In the training phase, an Actor-Critic-based multi-agent reinforcement learning algorithm is used, and a centralized critic network provides global value evaluation for the executor network of each agent to update the network parameters; In the execution phase, each agent only relies on its local observation and global state summary from the knowledge graph for distributed decision-making.

5. The method of claim 1, wherein, The reward function R of the multi-agent reinforcement learning model is designed in a multi-objective weighted form: R = ω1 · R time + ω2 · R coverage + ω3 · R cost + ω4 · R coordination wherein R time is a reward term negatively correlated with the total response time, R coverage is a reward term positively correlated with the successful rescue range, R cost is a reward term negatively correlated with the resource consumption cost, R coordination is a reward term positively correlated with the inter-agent coordination efficiency, and ω1 to ω4 are configurable weight coefficients.

6. The method of claim 1, wherein, In step S4, the output of the optimal collaborative scheduling scheme specifically refers to that the multi-agent reinforcement learning model searches within the feasible solution space defined by the action constraints and outputs a Pareto optimal solution that specifies the travel path, task allocation, and execution order of each rescue unit.

7. An emergency resource collaborative scheduling optimization system, comprising: It comprises: a dynamic resource knowledge graph construction and update module for constructing a knowledge graph of the emergency field, a multi-agent reinforcement learning model construction module, which is used for outputting an optimal collaborative scheduling scheme based on the current environmental state and the action constraints; an online execution module for issuing and executing the optimal collaborative scheduling scheme, and continuously monitoring environmental changes during the execution process to trigger dynamic rescheduling. The system specifically realizes the emergency resource collaborative scheduling optimization through the method of any one of claims 1 to 6.

8. An emergency resource collaborative scheduling optimization system according to claim 7, characterized in that, The dynamic resource knowledge graph construction and updating module includes: entity types, including rescue units, personnel, materials, facilities, events, and environment; relationship types between entities; attribute types between entities; the knowledge graph realizes real-time query, rule inference, and state perception functions; The multi-agent reinforcement learning model construction module includes: (1) Agent definition: each rescue unit in the knowledge graph is abstracted as an agent, and the entities and relationships in the knowledge graph are converted into a multi-agent reinforcement learning environment; (2) State space: the observation state of the agent is derived from the real-time query results of the dynamic resource knowledge graph, including: Self-state, including position, speed, and resource capacity; Event information, including type, level, and location; Environmental information, including traffic flow, weather, and road closure; Other agent information, including the location and state of adjacent rescue units; (3) Action space: the actions available to each agent include: Moving direction: selecting the next road segment to travel on; Task selection: selecting a sub-task point to go to; Resource allocation: deciding whether to carry or unload a certain material; (4) Reward function: designed as a multi-objective weighted function: (5) Algorithm selection: MAPPO algorithm is used to learn how to collaboratively select the best path and task; the reasoning ability of the dynamic resource knowledge graph is used to provide expert prior knowledge for multi-agent reinforcement learning; The online execution module includes the following execution processes: Phase one, initialization and perception: the sensor detects the event, the data is transmitted into the system, and the knowledge graph is updated in real time with the event information and associated entities; Phase two, knowledge graph reasoning: the dynamic resource knowledge graph first performs rapid reasoning to generate a set of feasible scheduling solutions that meet the business rules, which is used as the candidate action space for multi-agent reinforcement learning; Phase three, MARL collaborative decision-making: each agent obtains the current state, the multi-agent reinforcement learning model receives the candidate action set, and based on this, multi-objective optimization is performed to calculate the optimal or Pareto optimal scheduling solution, balance multiple objectives, and output the final scheduling solution, including the assigned unit, the path to follow, and the tasks to perform; Phase four, dynamic execution and adjustment: the solution is issued to each terminal, and each agent starts execution, the system continuously monitors environmental changes, and if changes occur, the dynamic resource knowledge graph is updated in real time, the multi-agent reinforcement learning model is re-decided quickly, and adjustment instructions are generated to realize dynamic rescheduling.

9. An emergency resource collaborative scheduling optimization apparatus, characterized in that, It includes: At least one memory and at least one processor; The at least one memory is used to store machine-readable programs; The at least one processor is used to call the machine-readable programs to realize the method of any one of claims 1 to 7.

10. A computer readable medium characterized by The computer readable medium stores computer instructions, which when executed by a processor, realize the method of any one of claims 1 to 7. The computer readable medium stores computer instructions, which when executed by a processor, realize the method of any one of claims 1 to 7.

Images