Modularized management method, system, equipment and product based on intelligent storage warehouse-in and warehouse-out

Abstract

The invention discloses a modular management method, system, equipment and product based on intelligent storage warehouse-in and warehouse-out, and relates to the technical field of intelligent storage and emergency logistics. The method comprises the steps of firstly analyzing a disaster situation text into a disaster situation vector through a natural language processing model, then mapping the vector into at least one fire-fighting function module identifier based on a mapping rule, and then determining an initial fire-fighting equipment set according to the identifier, and generating a complete fire-fighting equipment set with comprehensive optimal internal function and space through a community discovery algorithm by combining an equipment association graph with fire-fighting equipment as nodes and function association degree and goods allocation proximity weighted value as edge weights, and then generating a simulated ex-warehouse task in a storage digital twin model based on equipment real-time goods allocation. And a multi-agent reinforcement learning algorithm is utilized to plan a global conflict-free collaborative goods taking instruction set for a plurality of AGVs, and finally, when simulation verification is passed, an instruction is issued to an AGV controller for execution, so that the fire emergency response efficiency can be greatly improved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent warehousing and emergency logistics technology, specifically relating to a modular management method, system, equipment and products based on intelligent warehousing inbound and outbound operations. Background Technology

[0002] The effectiveness of emergency rescue operations for disasters such as fires highly depends on the speed of the initial response and the scientific nature of resource allocation. As the physical hub of emergency resources, the level of intelligence in the management model of fire equipment storage directly determines the efficiency of rescue force deployment.

[0003] Currently, fire equipment storage management has mainly gone through the following stages of development: Manual management mode: In the early stage, equipment management and allocation relied entirely on paper ledgers and manual memory. That is, when the equipment was issued, it was necessary to search and count it on the spot, which was inefficient and prone to errors or omissions of key equipment in emergency situations. Furthermore, it was impossible to effectively monitor the real-time status of the equipment (such as power and pressure). Information-based management model: With the application of technologies such as barcodes and RFID (Radio Frequency Identification), equipment information can be digitally entered and tracked. Equipment inventory and location can be quickly queried through the Warehouse Management System (WMS). However, this model is essentially the management of "static inventory". Outbound decisions (i.e., "what to take and how much to take") still rely entirely on the commander's human experience and on-site judgment. The system only provides list support and cannot provide intelligent assistance in disaster analysis and resource recommendation. Automated warehousing model: In recent years, some advanced fire-fighting reserve warehouses have begun to introduce hardware equipment such as automated storage and retrieval systems (AS / RS) and automated guided vehicles (AGVs). For example, public information shows that fire-fighting equipment warehouses in Wuhan and other places have been built using AGVs to achieve automatic "goods-to-person" picking. Related technical solutions (such as CN121329287A) have also proposed a warehouse management method that combines path planning and AR display. These solutions have significantly improved the automation level of material handling, but have not solved the core intelligent decision-making problem of emergency dispatch. Their operation logic is usually to execute a pre-set and fixed outbound instruction, or to perform simple zoning picking by category. Therefore, their core limitation is that the system cannot understand "why outbound", and thus cannot actively generate the "optimal outbound plan".

[0004] Specifically, existing technical solutions suffer from the following bottlenecks that urgently need to be addressed: (1) The response is rigid and cannot be dynamically adapted to the disaster situation. That is, the smallest unit of system scheduling is "single equipment" or "pre-packaged fixed material box". When faced with complex and ever-changing disaster situations, it is impossible to dynamically combine fully functional and precisely configured equipment packages according to the specific characteristics of the disaster (such as building height, presence of hazardous chemicals and / or the scale of trapped personnel, etc.). The fixed configuration of existing technical solutions may either be insufficient or may cause redundancy, reducing transportation and deployment efficiency. (2) The scheduling and path planning are disconnected, that is, the equipment scheduling scheme (what to pick up) and the AGV execution scheme (how to pick up) are two separate links. Existing path planning is mostly aimed at the efficiency of a single AGV or a simple task queue. It lacks the ability to perform global collaborative picking path planning for a dynamically generated and unprecedented equipment combination in emergency multi-task concurrent scenarios. It is easy to generate traffic conflicts and bottlenecks within the warehouse and cannot achieve the optimal overall outbound efficiency of the system. (3) Lack of closed-loop verification capability, that is, the system lacks the ability to quickly simulate and rehearse the overall scheme (especially the multi-AGV collaborative path) in a virtual environment before the scheme is executed; in addition, if a fault is found in a certain equipment during the execution process, the system is difficult to quickly and automatically start the backup scheme, and the fault tolerance is poor.

[0005] Therefore, the current field of intelligent fire protection warehousing urgently needs an integrated solution that can understand disaster situations, intelligently assemble resources, collaboratively plan, and perform closed-loop verification, in order to break through the bottleneck from "automated execution" to "intelligent decision-making". Summary of the Invention

[0006] The purpose of this invention is to provide a modular management method, system, computer equipment, computer-readable storage product, and computer program product based on intelligent warehouse inbound and outbound operations, in order to solve the problems of rigid response, inability to dynamically adapt to disaster situations, disconnect between scheduling and path planning, and lack of closed-loop verification capabilities in existing fire equipment warehouse management solutions.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: Firstly, a modular management method based on intelligent warehouse inbound and outbound operations is provided, executed by a management server communicating with each automated guided vehicle (AGV) in the intelligent warehouse inbound and outbound operations, including: Obtain disaster information in the form of emergency fire-fighting commands; A pre-trained natural language processing model is invoked to parse the disaster text information into a structured disaster situation vector, wherein the disaster situation vector contains at least two items from the following: disaster type, altitude of the disaster location, status of people at the disaster location, and type of hazardous materials at the disaster location; Based on pre-built scenario and function mapping rules, the disaster scenario vector is mapped to a unique identifier of at least one fire protection function module. The scenario and function mapping rules are used to define the correspondence between different disaster scenario vectors and multiple preset fire protection function modules. Each fire protection function module is associated with at least one mandatory fire protection equipment and at least one optional fire protection device. Based on the unique identifier of the at least one fire-fighting functional module, determine the initial set of fire-fighting equipment for responding to the emergency fire-fighting command; Based on the initial set of fire-fighting equipment and the pre-constructed fire-fighting equipment association graph, a complete set of fire-fighting equipment for responding to the emergency fire-fighting command is generated through a community detection algorithm. The fire-fighting equipment association graph uses all fire-fighting equipment in the warehouse as nodes, and uses the weighted value of the functional co-occurrence correlation degree and physical storage location proximity between any two fire-fighting equipment as the weight of the edge connecting the two nodes that correspond one-to-one with any two fire-fighting equipment. The optimization objective of the community detection algorithm is to make the combined functional correlation and storage location proximity within the generated equipment set optimal. Based on the real-time location of each fire-fighting equipment in the complete fire-fighting equipment set, a simulated outbound task containing multiple picking sub-tasks is triggered and generated in the warehouse digital twin model of the intelligent warehouse inbound and outbound. Based on a multi-agent reinforcement learning algorithm, a globally conflict-free collaborative picking instruction set is planned for multiple AGVs responsible for executing the simulated outbound task. The collaborative picking instruction set includes motion control instructions generated independently for each of the multiple AGVs. The planned collaborative picking instruction set is simulated and verified. After the verification is successful, each motion control instruction in the collaborative picking instruction set is sent to the corresponding AGV controller for execution.

[0008] Based on the above-mentioned invention, an integrated fire-fighting and warehousing emergency solution is provided, capable of understanding disaster situations, intelligent assembly, collaborative planning, and closed-loop verification. First, the disaster situation text information of the emergency fire-fighting command is acquired and parsed into a structured disaster situation vector using a pre-trained natural language processing model. Then, based on pre-built mapping rules, the vector is mapped to at least one fire-fighting functional module identifier. Next, an initial set of fire-fighting equipment is determined based on the identifiers. Combined with an equipment association graph with fire-fighting equipment as nodes and edge weights based on functional relevance and location proximity, a community detection algorithm generates a complete set of fire-fighting equipment with optimal internal functionality and space. Then, based on the real-time location of the equipment, a simulated outbound task is generated in a warehouse digital twin model. A multi-agent reinforcement learning algorithm is used to plan a globally conflict-free collaborative picking instruction set for multiple AGVs. Finally, upon successful simulation verification, instructions are sent to the AGV controller for execution. This achieves full-process automation and intelligence from intelligent disaster situation analysis to dynamic optimization and assembly of equipment and collaborative scheduling of multiple AGVs, greatly improving the efficiency of fire emergency response and facilitating practical application and promotion.

[0009] In one possible design, based on the unique identifier of the at least one fire-fighting functional module, an initial set of fire-fighting equipment for responding to the emergency fire-fighting command is determined, including: Based on the unique identifier of the at least one fire protection functional module, locate all the required fire protection equipment associated with the at least one fire protection functional module in the warehouse, and use the search results as the initial set of fire protection equipment for responding to the emergency fire protection command; For each fire protection functional module in the at least one fire protection functional module, if all optional fire protection equipment associated with the corresponding module can be found in the warehouse according to the corresponding unique identifier, then the specific equipment model and quantity are selected from all optional fire protection equipment according to at least one specific quantitative parameter in the disaster situation vector and the predefined equipment adaptation rule, and the selection result is added to the initial fire protection equipment set. The specific quantitative parameter includes at least one of the following: the height value of the disaster location, the estimated number of trapped people, and the type of hazardous chemicals. The equipment adaptation rule is used to define the mapping relationship between different quantitative parameter ranges and the model and quantity of optional fire protection equipment.

[0010] In one possible design, based on the initial set of fire-fighting equipment and a pre-built fire-fighting equipment association graph, a complete set of fire-fighting equipment for responding to the emergency fire-fighting command is generated using a community detection algorithm, including: Using all equipment in the initial fire-fighting equipment set as target nodes, a community detection algorithm is run on a pre-constructed fire-fighting equipment association graph to find a connected subgraph that satisfies the following conditions: it contains all the target nodes and maximizes the ratio of the sum of the weights of the edges inside the subgraph to the sum of the weights of all possible edges in the subgraph. The fire-fighting equipment association graph uses all fire-fighting equipment in the warehouse as nodes, and uses the weighted value of the functional co-occurrence correlation degree and physical storage location proximity between any two fire-fighting equipment as the weight of the edge connecting the two nodes that correspond one-to-one with the two fire-fighting equipment. The optimization objective of the community detection algorithm is to make the combined functional correlation and storage location proximity within the generated equipment set optimal. The fire-fighting equipment corresponding to the non-target nodes contained in the found connected subgraph is added to the initial fire-fighting equipment set, thus forming a complete fire-fighting equipment set for responding to the emergency fire-fighting command.

[0011] In one possible design, a globally conflict-free set of cooperative picking instructions is planned for the multiple AGVs responsible for executing the simulated outbound task based on a multi-agent reinforcement learning algorithm, including: Using the aforementioned warehouse digital twin model as a simulation environment, each AGV in the smart warehouse inbound and outbound processes is modeled as an intelligent agent. Construct a multi-agent deep reinforcement learning model, wherein the multi-agent deep reinforcement learning model includes a centralized value critic network and a distributed policy enforcer network corresponding one-to-one with each agent; The multi-agent deep reinforcement learning model is trained offline and fine-tuned online using an actor-critic training framework with a composite reward function to guide agent cooperation, as follows: During training, the centralized value critic network evaluates the value of the global state based on the joint state of all agents and global map information, and guides each distributed policy enforcer network to optimize its execution policy to maximize the expected long-term cumulative reward. The composite reward function is obtained by weighted summation using the following formula:

[0012] In the formula, The serial number of the intelligent agent is indicated. Indicates a time step. Indicates the first An agent at time step Compound rewards, Indicates the first The agents at the time step The task reward items, and if the first The agents at the time step If the target location is successfully reached and the virtual pickup action is completed, the task reward item will receive a positive fixed value reward; otherwise, the task reward item will be zero. Indicates the first The agents at the time step The collision penalty term, and at the time step If the first If the bounding box of an agent in the simulation scene overlaps with the bounding box of any other agent or static obstacle, the collision penalty term receives a negative fixed value reward; otherwise, the collision penalty term is set to zero. Indicates the first The agents at the time step The collaborative efficiency reward item, and at the time step Statistically analyze the travel behavior of all the aforementioned agents at the shared resource point. If the first... If an agent engages in alternating or following traffic without conflict with other agents within a pre-defined competitive area, the collaborative efficiency reward item receives a positive, non-fixed value that is dynamically calculated based on the smoothness of local traffic flow. Otherwise, the collaborative efficiency reward item is set to zero. , and These represent the preset positive weighting coefficients, and have... ; For each AGV, the trained and corresponding distributed policy executor network independently generates corresponding motion control commands based on the real-time local observation information of the corresponding AGV. The motion control commands of each AGV are combined to form a globally conflict-free collaborative picking instruction set planned by the multi-agent reinforcement learning algorithm for the multiple AGVs responsible for executing the simulated outbound task.

[0013] In one possible design, the planned collaborative pickup instruction set is simulated and verified, including: In the warehouse digital twin model, the entire task process is accelerated through simulation based on the planned collaborative picking instruction set and the preset AGV motion dynamics model. If a deadlock, collision, or task timeout is predicted during the simulation, the verification is deemed unsuccessful; otherwise, the verification is deemed successful.

[0014] In one possible design, after issuing the various motion control commands to the corresponding AGV controllers for execution, the method further includes: During the process of physical AGV performing a picking task, the electrical quantity, pressure value or last maintenance time of the target fire-fighting equipment is verified to be within the available threshold range by reading the RFID tag on the target fire-fighting equipment or the IoT sensor on the target fire-fighting equipment connected to the communication. If the verification fails, other fire-fighting equipment will be selected nearby to replace the target fire-fighting equipment according to the equipment association diagram, and the retrieval route of the affected AGV will be replanned in real time.

[0015] Secondly, a modular management system based on intelligent warehouse entry and exit is provided, which is suitable for deployment in the management server of each automated guided vehicle (AGV) in the intelligent warehouse entry and exit. It includes a disaster text acquisition unit, a text parsing and processing unit, a functional module mapping unit, an equipment set initial determination unit, an equipment set final determination unit, a simulation task generation unit, a collaborative picking planning unit, and an instruction verification and issuance unit that are connected in sequence. The disaster text acquisition unit is used to acquire disaster text information of emergency fire fighting orders; The text parsing and processing unit is used to call a pre-trained natural language processing model to parse the disaster text information into a structured disaster situation vector, wherein the disaster situation vector contains at least two items from the following: disaster type, altitude of the disaster location, status of people at the disaster location, and type of hazardous materials at the disaster location. The functional module mapping unit is used to map the disaster scenario vector to a unique identifier of at least one fire protection functional module based on pre-built scenario and function mapping rules. The scenario and function mapping rules are used to define the correspondence between different disaster scenario vectors and multiple preset fire protection functional modules. Each fire protection functional module is associated with at least one mandatory fire protection equipment and at least one optional fire protection device. The equipment set initial determination unit is used to determine the initial fire-fighting equipment set for responding to the emergency fire-fighting command based on the unique identifier of the at least one fire-fighting functional module. The equipment set finalization unit is used to generate a complete fire equipment set for responding to the emergency fire command based on the initial fire equipment set and the pre-constructed fire equipment association graph through a community detection algorithm. The fire equipment association graph uses all fire equipment in the warehouse as nodes, and uses the weighted value of the functional co-occurrence correlation degree and physical storage location proximity between any two fire equipment as the weight of the edge connecting the two nodes that correspond one-to-one with any two fire equipment. The optimization goal of the community detection algorithm is to make the combined functional correlation and storage location proximity within the generated equipment set optimal. The simulation task generation unit is used to trigger the generation of a simulated outbound task containing multiple picking sub-tasks in the warehouse digital twin model of the intelligent warehouse inbound and outbound based on the real-time location of each fire-fighting equipment in the complete fire-fighting equipment set. The collaborative picking planning unit is used to plan a globally conflict-free collaborative picking instruction set for multiple AGVs responsible for executing the simulated outbound task based on a multi-agent reinforcement learning algorithm. The collaborative picking instruction set includes motion control instructions generated independently for each of the multiple AGVs. The instruction verification and issuance unit is used to simulate and verify the planned collaborative picking instruction set, and after the verification is passed, to issue each motion control instruction in the collaborative picking instruction set to the corresponding AGV controller for execution.

[0016] Thirdly, the present invention provides a computer device comprising a storage module, a processing module, and a transceiver module connected in sequence for communication, wherein the storage module is used to store a computer program, the transceiver module is used to send and receive messages, and the processing module is used to read the computer program and execute the modular management method as described in the first aspect or any possible design in the first aspect.

[0017] Fourthly, the present invention provides a computer-readable storage product having instructions stored thereon that, when executed on a computer, perform the modular management method as described in the first aspect or any possible design in the first aspect.

[0018] Fifthly, the present invention provides a computer program product, including a computer program or instructions, wherein the computer program or instructions, when executed by a computer, implement the modular management method as described in the first aspect or any possible design in the first aspect.

[0019] The beneficial effects of the above scheme are: (1) This invention creatively provides an integrated fire-fighting and warehousing emergency solution that can understand disaster situations, intelligently assemble and coordinate, collaboratively plan and verify closed loops. First, the disaster situation text information of the emergency fire-fighting command is obtained and parsed into a structured disaster situation vector through a pre-trained natural language processing model. Then, based on the pre-built mapping rules, the vector is mapped to at least one fire-fighting functional module identifier. Then, the initial fire-fighting equipment set is determined according to the identifier. Combined with the equipment association graph with fire-fighting equipment as nodes and functional correlation and location proximity as edge weights, a community detection algorithm is used to generate a complete fire-fighting equipment set with optimal internal functions and space. Then, based on the real-time location of the equipment, a simulated outbound task is generated in the warehouse digital twin model. A multi-agent reinforcement learning algorithm is used to plan a globally conflict-free collaborative picking instruction set for multiple AGVs. Finally, when the simulation verification is passed, the instruction is sent to the AGV controller for execution. Thus, the entire process from intelligent disaster situation analysis to dynamic optimization and assembly of equipment and collaborative scheduling of multiple AGVs can be automated and intelligent, which greatly improves the efficiency of fire emergency response. (2) It has realized a paradigm leap from "static execution" to "dynamic intelligent allocation" in emergency resource scheduling. On the one hand, through natural language understanding and rule mapping, the ambiguous disaster text is transformed into modular tasks with clear functions. Then, based on specific disaster parameters (such as altitude and number of people), the available equipment models and quantities are accurately matched, which solves the inherent contradiction of mismatch between fixed material packages and variable disasters and avoids insufficient configuration or waste of resources. On the other hand, it innovatively introduces "equipment association graph" and uses community discovery algorithm for optimization. It not only ensures that the core equipment is complete, but also automatically and intelligently adds equipment with highly related functions and adjacent storage locations, so that the final equipment set generated achieves the best overall performance in terms of functional synergy and outbound picking efficiency, realizing data-driven intelligent packaging that surpasses human experience. (3) A highly efficient execution system from "single-machine planning" to "multi-agent global collaboration" was constructed. On the one hand, a multi-agent reinforcement learning model of "centralized training and distributed execution" was adopted. Through a carefully designed composite reward function (including task, collision and collaboration efficiency terms), multiple AGVs were driven to learn advanced collaboration strategies such as avoidance and alternating passage. This enabled the planning of a globally conflict-free and time-shortest collaborative path for dynamically generated complex task sets, fundamentally solving the congestion and efficiency problems when multiple AGVs operate in parallel. On the other hand, the collaborative instruction set was simulated and verified based on a digital twin model. This allowed for the pre-execution simulation and discovery of potential risks such as deadlock and collision, ensuring the safety and reliability of the solution and realizing a decision-making closed loop of "computation-verification-execution". (4) A reliable guarantee mechanism has been formed from "one-time planning" to "full-process flexible fault tolerance". On the one hand, during the execution of physical AGVs, the equipment status is checked in real time through IoT sensors. Once a fault is detected, the closest alternative equipment can be selected intelligently and quickly based on the "equipment association diagram". Only the affected AGVs are replanned locally, realizing second-level self-healing of abnormal situations and ensuring the final success rate of the rescue mission. On the other hand, the whole solution constitutes a complete intelligent closed loop: "disaster understanding → intelligent assembly → collaborative planning → simulation verification → execution fault tolerance". It integrates the decision-making ability of artificial intelligence into the entire chain of emergency response, and upgrades the traditional warehouse management system into an intelligent emergency brain with perception, decision-making, optimization and fault tolerance capabilities. It significantly improves the overall speed, scientificity and reliability of fire emergency response, and is convenient for practical application and promotion. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 A flowchart illustrating the modular management method for intelligent warehouse inbound and outbound operations provided in this application embodiment.

[0022] Figure 2 This is a schematic diagram of the modular management system for intelligent warehouse inbound and outbound operations provided in an embodiment of this application.

[0023] Figure 3 A schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the present invention will be briefly introduced below in conjunction with the accompanying drawings and descriptions of the embodiments or the prior art. Obviously, the following description of the structure of the accompanying drawings is only some embodiments of the present invention. For those skilled in the art, other embodiments can be obtained based on these embodiments without creative effort. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention.

[0025] It should be understood that although the terms "first" and "second", etc., may be used herein to describe various objects, these objects should not be limited by these terms. These terms are only used to distinguish one object from another. For example, the first object may be referred to as the second object, and similarly, the second object may be referred to as the first object, without departing from the scope of the exemplary embodiments of the invention.

[0026] It should be understood that the term "and / or" that may appear in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, B exists alone, or A and B exist simultaneously. Another example is A, B and / or C, which can mean that any one of A, B, and C or any combination thereof exists. The term " / and" that may appear in this document describes another relationship between related objects, indicating that two relationships can exist. For example, A / and B can mean: A exists alone or A and B exist simultaneously. In addition, the character " / " that may appear in this document generally indicates that the related objects before and after it are in an "or" relationship.

[0027] Example like Figure 1 As shown, the modular management method based on intelligent warehouse inbound and outbound operations provided in the first aspect of this embodiment can be executed, but is not limited to, by a management server of each automated guided vehicle (AGV) with certain computing resources and communication connections within the intelligent warehouse inbound and outbound operations, for example, by a management platform server for the intelligent warehouse inbound and outbound operations. Figure 1 As shown, the modular management method includes, but is not limited to, the following steps S1 to S8.

[0028] S1. Obtain disaster information in the form of emergency fire-fighting commands.

[0029] In step S1, the emergency fire command is typically generated and issued by the fire command center through a dedicated system. Essentially, it is a structured dispatch instruction, containing, but not limited to, core fields such as time, location, response level, and unstructured disaster description. It can be directly pushed to the management server in this embodiment via a network interface. The disaster text information is the unstructured disaster description carried in the emergency fire command, which can be obtained before the command is generated through, but is not limited to, the following three methods: (A) alarm entry, i.e., key text compiled by the operator based on the caller's statement, such as "A fire has occurred on the 20th floor of XX Building, people are trapped, and there is heavy smoke"; (B) on-site feedback, i.e., a brief report transmitted back by the commander or frontline team members via mobile terminal; (C) automatic sensing, i.e., linkage with the city's Internet of Things system to obtain auxiliary text data about hazardous materials and building structures.

[0030] S2. Call a pre-trained natural language processing model to parse the disaster text information into a structured disaster situation vector, wherein the disaster situation vector includes, but is not limited to, at least two of the following: disaster type, altitude of the disaster location, status of people at the disaster location, and type of hazardous materials at the disaster location.

[0031] In step S2, the natural language processing model is a sequence understanding model based on deep learning (especially the Transformer architecture), such as BERT, RoBERTa, or their domain-optimized variants. The core capability of this model is to understand the semantics of free text and accurately extract predefined key information entities and their attributes from unstructured sentences. For example, if the original disaster text information is "Report received: A fire has occurred in the Innovation Building at No. 88 Chuangye Road, High-tech Zone. The building is a 30-story office building. It has been confirmed that people are trapped on the 15th floor and above. The burning materials are suspected to include plastic and cables," then through model analysis, a structured disaster situation vector can be obtained as follows: {"Disaster type": "Fire", "Disaster location height": "High-rise building (30 floors)", "Disaster location personnel status": "Yes, trapped on the 15th floor and above", "Disaster location hazardous material type": ["Plastic", "Cable"]}. Furthermore, by performing routine pre-training on the natural language processing model, the model can acquire a foundation in general language understanding. By further fine-tuning it on specialized corpora such as fire alarm records and disaster assessment reports, the model can accurately identify entities in fields such as "high-rise buildings," "trapped," and "hazardous chemical names," ensuring the professionalism and reliability of the parsing process.

[0032] S3. Based on the pre-built scenario and function mapping rules, the disaster scenario vector is mapped to a unique identifier of at least one fire protection function module. The scenario and function mapping rules are used to define the correspondence between different disaster scenario vectors and multiple preset fire protection function modules. Each fire protection function module is associated with at least one mandatory fire protection equipment and at least one optional fire protection device.

[0033] In step S3, the scenario-function mapping rule is not a single rule table, but a decision-making system that supports complex logical judgments. Its construction and implementation mainly include the following two complementary methods: (A) An explicit mapping library based on a rule engine. The construction process specifically includes: based on national fire emergency rescue regulations, typical case reviews, and the experience of domain experts, formalizing the correspondence between disaster scenario characteristics and functional modules into a series of IF-THEN judgment rules; for example: IF Disaster type includes "fire" AND building height belongs to "high-rise" THEN Mapping identifier is MOD_001 (i.e., the unique identifier of the high-rise fire attack module) and MOD_002 (i.e., the unique identifier of the high-rise water supply module); IF Disaster type includes "hazardous chemical leak" THEN The mapping identifiers are MOD_005 (i.e., the unique identifier of the chemical decontamination module) and MOD_006 (i.e., the unique identifier of the personal protective equipment module); (B) Intelligent mapping based on knowledge graph and machine learning model, the construction process specifically includes: establishing a knowledge graph in the field of fire protection, associating conceptual entities such as disaster type, building structure, hazardous substances, disposal measures and equipment functions, and using historical outbound records and corresponding disaster reports as training data to train a classification or recommendation model; thus, the "disaster situation vector" is input into the model or queried and reasoned in the graph, and by analyzing the multidimensional association between entities, the most matching functional module identifiers can be recommended; this method can handle more complex and atypical disaster combinations and has self-learning optimization capabilities.

[0034] In step S3, each preset fire protection function module is a logical task package, such as "high-rise water supply module", "personnel search and rescue module" and "hazardous materials leak sealing module"; each module is associated with two parts of equipment: (1) mandatory fire protection equipment, that is, the basic equipment necessary to complete the core function of the module, which must be included in any case when the module is called; for example, the mandatory equipment of the "personnel search and rescue module" may include life detectors, demolition tool sets and rescue tripods; (2) optional fire protection equipment, that is, equipment selectively equipped according to the specific detailed parameters of the disaster (such as the scale of the disaster and the complexity of the environment, etc.), whether it is ultimately selected and the specific model will be determined by subsequent steps according to the "specific quantitative parameters" and "equipment adaptation rules"; for example, for the same "high-rise water supply module", depending on the building height, the optional equipment may be a 65mm or 80mm diameter high-pressure water hose. In addition, the fire protection function module may be associated with only one mandatory fire protection equipment, or it may be associated with one mandatory fire protection equipment and one optional fire protection equipment.

[0035] In step S3, the specific mapping process can be exemplified as follows: Assume the disaster scenario vector is: {Disaster type: fire, building height: 30 floors, personnel status: trapped, hazardous material: cable}; the rule engine matches the features "high-rise fire" and "personnel trapped"; finally, based on these features, it outputs the corresponding unique identifiers of functional modules, such as: MOD_001 (the unique identifier of the high-rise fire attack module), MOD_002 (the unique identifier of the high-rise water supply module), and MOD_003 (the unique identifier of the personnel search and rescue and lighting module). This yields a clear list of logical task units that need to be invoked for this response, laying the foundation for determining the specific equipment set in the next step.

[0036] S4. Determine the initial set of fire-fighting equipment for responding to the emergency fire-fighting command based on the unique identifier of the at least one fire-fighting functional module.

[0037] In step S4, since each fire-fighting functional module is associated with at least one mandatory fire-fighting equipment and at least one optional fire-fighting equipment, the initial set of fire-fighting equipment can be determined based on these relationships. Addressing the contradiction between "basic support" and "precise adaptation" in emergency response, this embodiment, to ensure that core rescue capabilities are not lacking, first identifies the mandatory equipment as basic support, and then intelligently selects from the optional equipment based on disaster quantification parameters and equipment adaptation rules, achieving precise matching and efficient utilization of resources. Specifically, preferably, the initial set of fire-fighting equipment for responding to the emergency fire-fighting command is determined based on the unique identifier of the at least one fire-fighting functional module, including but not limited to the following steps S41-S42.

[0038] S41. Based on the unique identifier of the at least one fire-fighting functional module, locate all mandatory fire-fighting equipment associated with the at least one fire-fighting functional module in the warehouse, and use the search results as the initial set of fire-fighting equipment for responding to the emergency fire-fighting command.

[0039] In step S41, the warehouse is used to store all fire-fighting equipment. It can maintain a pre-built database or configuration table, and use the unique identifier of each fire-fighting functional module as the primary key to clearly associate one or more "mandatory fire-fighting equipment" with their numbers, names, and standard storage locations within the warehouse. Thus, once the unique identifier of at least one fire-fighting functional module is obtained, a quick database query operation can retrieve all associated mandatory equipment information, forming the initial set of fire-fighting equipment. For example, after mapping the unique identifier of the "high-rise water supply module," it can automatically query and aggregate its associated mandatory equipment such as "high-pressure fire pump (P001)" and "65mm water hose (H005)."

[0040] S42. For each fire-fighting functional module in the at least one fire-fighting functional module, if all optional fire-fighting equipment associated with the corresponding module can be found in the warehouse according to the corresponding unique identifier, then according to at least one specific quantitative parameter in the disaster scenario vector and the predefined equipment adaptation rule, the model and quantity of specific equipment are selected from all optional fire-fighting equipment, and the selection result is added to the initial fire-fighting equipment set. The specific quantitative parameter includes, but is not limited to, at least one of the following: the height value of the disaster location, the estimated number of trapped people, and the type of hazardous chemicals. The equipment adaptation rule is used to define the mapping relationship between different quantitative parameter ranges and the model and quantity of optional fire-fighting equipment.

[0041] In step S42, the objective is to intelligently supplement and refine the equipment configuration based on the specific quantitative characteristics of the disaster, while ensuring basic functions (mandatory equipment), thereby generating a final equipment plan that is highly compatible with the disaster situation and highly operable. First, determining whether the current functional module is configured with optional fire-fighting equipment is also done by querying the pre-built "module-equipment" association database: if a list of optional equipment associated with the module identifier exists in the database, the refined selection process begins; otherwise, this step for this module is skipped. Secondly, regarding the intelligent selection based on quantitative parameters and adaptation rules, its input and logic can be designed as follows: First, extract quantifiable decision parameters from the disaster scenario vector generated upstream (these parameters are objective, specific numerical or enumerated values, such as: the height of the disaster location: such as "80 meters", "30 floors"; the estimated number of trapped people: such as "about 10 people", "a large number"; the types of hazardous chemicals: such as "liquefied petroleum gas", "cyanide"). Then, complete the selection task of optional equipment through predefined equipment adaptation rules (i.e., a built-in rule base, which contains judgment logic formulated by domain experts to map the above quantitative parameters to specific equipment models and quantities. The rules usually adopt the IF-THEN form and support range judgment). Rule example 1 (for height): IF module is "high-rise water supply module" AND height value > 100 meters THEN Select "high-pressure fire pump with a head of more than 90 meters (model: P-90), quantity: 2 units" from the optional equipment; IF module is "high-rise water supply module" AND 50 meters < height value ≤ 100 meters THEN: Select "70-meter head high-pressure fire pump (model: P-70), quantity: 2 units"; Rule Example 2 (for the number of people): IF module is "personnel search and rescue module" AND estimated number of trapped people > 20 people THEN: Add "multi-functional rescue stretcher (model: S-3), quantity: add 3 sets" from the optional equipment; Finally, after performing rule matching, the selected specific equipment model and quantity are used as the selection result, and these newly added equipment entries are added to the previously formed initial fire-fighting equipment set. Thus, the initial fire-fighting equipment set has evolved from a list that only meets basic functions into a list of equipment to be issued that is deeply adapted to the specific disaster situation.

[0042] S5. Based on the initial fire-fighting equipment set and the pre-constructed fire-fighting equipment association graph, a complete fire-fighting equipment set for responding to the emergency fire-fighting command is generated through a community detection algorithm. The fire-fighting equipment association graph uses all fire-fighting equipment in the warehouse as nodes, and uses the weighted value of the functional co-occurrence correlation degree and physical location proximity between any two fire-fighting equipment as the weight of the edge connecting the two nodes that correspond one-to-one with the two fire-fighting equipment. The optimization objective of the community detection algorithm is to make the combined functional correlation and location proximity within the generated equipment set optimal.

[0043] In step S5, the purpose is to, based on the initial set of fire-fighting equipment directly related to the disaster, further utilize the inherent relationships between equipment in the warehouse to intelligently and automatically add a batch of related equipment that, while not on the original list, can significantly improve overall picking efficiency or functional completeness. This generates a final outbound list optimized in both "functional logic" and "space efficiency." The fire-fighting equipment relationship graph is a weighted undirected graph that quantifies the inherent connections between all equipment in the warehouse from both functional and spatial dimensions. Each independent piece of fire-fighting equipment in the warehouse (such as a water pump, a hose, or a detector) is a node in the graph, connecting any two equipment nodes. weight of the edge It is determined by a weighted combination of the following two core indicators: (1) Functional co-occurrence correlation It is based on historical outbound data statistics; if two pieces of equipment are frequently called up at the same time in many past rescue missions, their functions are considered to be highly correlated; this correlation can be quantified by calculating the co-occurrence frequency of the two in historical missions (such as the Jaccard coefficient), and corrected by combining the functional description of the equipment in the product manual; (2) Physical location proximity This is calculated based on the digital layout of the warehouse; the Euclidean distance or picking path distance between the two pieces of equipment is calculated according to their actual storage location coordinates. The closer the distance, the higher the proximity; this value is usually normalized. The final edge weight is the weighted sum of the two, i.e. ,in, and These are the preset weighting coefficients (specifically...) This is used to adjust the relative importance of functional association and spatial proximity in the optimization objective; for example, if faster outbound processing is emphasized, it can be increased. The value of .

[0044] In step S5, each piece of equipment in the initial fire-fighting equipment set is considered a "seed" node that must be within the same "community" (i.e., the final equipment set). Then, on the constructed fire-fighting equipment association graph, a community discovery algorithm (such as the existing Louvain algorithm or Leiden algorithm) is run. The optimization objective of the community discovery algorithm is set to find a connected subgraph (i.e., a "community") that contains all the "seed" nodes, maximizing the modularity within the subgraph. (This general graph theory metric is given a specific business meaning: the higher the modularity, the higher the overall level of functional co-occurrence association and physical proximity between equipment within the subgraph; that is, the algorithm automatically finds an equipment group that not only contains all the necessary initial equipment, but also where other equipment within the group is highly functionally synergistic with the initial equipment and very close in location.) After the algorithm runs, equipment nodes contained in this connected subgraph that meets the "overall optimal" condition but are not in the initial set are automatically identified as "associated additional equipment" and added to the initial set, thus forming the complete fire-fighting equipment set. For example, suppose the initial set is determined based on the "high-rise fire" disaster situation: high-pressure water pump A, water hose B, and demolition tool C; in the association graph, the algorithm may find two nodes, fire hose adapter D and mobile lighting E, which frequently co-occur with nodes A, B, and C in historical tasks (high functional correlation), and their locations are near B and C (high spatial proximity). In this case, the algorithm will automatically add D and E to the final set; in this way, the generated complete fire-fighting equipment set not only meets the core fire-fighting requirements, but also pre-equips the necessary adapters and lighting equipment that may be urgently needed. Furthermore, since D and E are located close to each other, the AGV can pick them up along the way in one trip, greatly improving the overall picking efficiency.

[0045] In step S5, specifically, based on the initial fire-fighting equipment set and the pre-constructed fire-fighting equipment association graph, a complete fire-fighting equipment set for responding to the emergency fire-fighting command is generated through a community detection algorithm, including but not limited to the following steps S51-S52: S51. Using all equipment in the initial fire-fighting equipment set as target nodes, the community detection algorithm is run on the pre-constructed fire-fighting equipment association graph to find a connected subgraph that satisfies the following conditions: it contains all the target nodes, and the ratio of the sum of the weights of the edges inside the subgraph to the sum of the weights of all possible edges in the subgraph is maximized. The fire-fighting equipment association graph uses all fire-fighting equipment in the warehouse as nodes, and uses the weighted value of the functional co-occurrence correlation and physical location proximity between any two fire-fighting equipment as the weight of the edge connecting the two nodes corresponding to each of the two fire-fighting equipment. The optimization objective of the community detection algorithm is to optimize the combined functional correlation and location proximity within the generated equipment set; S52. The fire-fighting equipment corresponding to the non-target nodes contained in the found connected subgraph is added to the initial fire-fighting equipment set, ultimately forming a complete fire-fighting equipment set for responding to the emergency fire-fighting command.

[0046] S6. Based on the real-time location of each fire-fighting equipment in the complete fire-fighting equipment set, a simulated outbound task containing multiple picking sub-tasks is generated in the warehouse digital twin model of the intelligent warehouse inbound and outbound.

[0047] In step S6, the specific task generation process can be briefly described as follows: (1) Data preparation: Based on the unique number of each piece of equipment in the complete fire-fighting equipment set, query the warehouse management database in real time to obtain its current accurate location coordinates (e.g., A-3 row-5 floor-2); at the same time, obtain the real-time layout of the warehouse, the channel status and the dynamic information of AGV position from the digital twin model; (2) Task modeling: In the virtual environment of the warehouse digital twin model synchronized with the physical warehouse, take the real-time location of each piece of equipment as the target point and automatically create a picking sub-task. This sub-task is a structured data object, which at least includes the target equipment ID, location coordinates and task priority; (3) Task package generation: Aggregate the picking sub-tasks corresponding to all equipment to form a structured simulated outbound task data package. This task package fully defines all physical picking requirements of this response and serves as the direct input of the next multi-agent path planning algorithm to calculate the optimal operation route of the AGV. Thus, through the digital twin environment, it can be ensured that the task generation is based on the real-time and real warehouse status, providing accurate input for subsequent collaborative path planning and simulation verification.

[0048] S7. Based on a multi-agent reinforcement learning algorithm, a globally conflict-free collaborative picking instruction set is planned for the multiple AGVs responsible for executing the simulated outbound task, wherein the collaborative picking instruction set includes, but is not limited to, motion control instructions independently generated for each of the multiple AGVs.

[0049] In step S7, the multi-agent reinforcement learning algorithm is an artificial intelligence method used to enable multiple AGVs (i.e., "agents") to learn optimal collaborative strategies through continuous interaction with the environment (i.e., the warehouse digital twin model). The correspondence between the multiple AGVs and the multiple picking sub-tasks is a dynamic, optimized many-to-many allocation relationship, rather than a fixed one-to-one allocation. For example, for a simulated outbound task containing 9 picking sub-tasks (i.e., sub-tasks A to I), with 3 AGVs, the algorithm may plan: AGV1 is responsible for sub-tasks A, B, and C; AGV2 is responsible for sub-tasks D, E, and F; and AGV3 is responsible for sub-tasks G, H, and I. Furthermore, the algorithm precisely arranges the passage timing of AGV1 and AGV2 at narrow passage entrances to ensure that they do not arrive simultaneously and cause deadlock. All of these are planned automatically and in one go. Specifically, based on the multi-agent reinforcement learning algorithm, a globally conflict-free collaborative picking instruction set is planned for the multiple AGVs responsible for executing the simulated outbound task, including but not limited to the following steps S71 to S75.

[0050] S71. Using the aforementioned warehouse digital twin model as the simulation environment, each AGV in the smart warehouse inbound and outbound processes is modeled as an intelligent agent.

[0051] In step S71, the core elements of each agent can be defined as follows: (1) State space (S): for each agent At time step Observed local state It is a vector containing its normalized two-dimensional coordinates in the warehouse map. Speed ​​magnitude Orientation Angle Coordinates of the current target storage location Self-power , and the relative distance and direction of the nearest obstacle or other AGV within a 3m radius around it, as perceived by LiDAR or depth camera; (2) Action space (A): The actions of each agent are discrete, including: {forward, backward, left turn (in place), right turn (in place), stop}; (3) In the continuous control scenario, the action can be defined as normalized forward acceleration [-1, 1] and turning angular velocity [-1, 1]. In addition, in digital twins, the motion of AGVs follows a preset dynamic model; for example, using a differential drive model, its state at the next moment is determined by the current state, the action performed, and the collision detection results with map elements.

[0052] S72. Construct a multi-agent deep reinforcement learning model, wherein the multi-agent deep reinforcement learning model includes a centralized value critic network and a distributed policy enforcer network corresponding one-to-one with each agent.

[0053] In step S72, each agent It has an independent policy network (i.e., the distributed policy executor network), which is a multilayer perceptron (MLP) whose input is the agent's own local state. After passing through two hidden layers (128 and 64 neurons respectively, using the ReLU activation function), the output is either a probability distribution in the action space A (discrete actions) or a direct output of the action value (continuous actions). The input to the centralized value critic network is a time step... all The joint state of (a positive integer) agents It also includes a grid map encoding a global static map and dynamic obstacle information; its network structure contains a fully connected branch for processing joint states and a convolutional neural network branch for processing the grid map. The features of the two branches are fused in the intermediate layer, and finally output a scalar value. , representing the global state Below is an estimate of the total cumulative rewards expected to be obtained in the future.

[0054] S73. The multi-agent deep reinforcement learning model is trained offline and fine-tuned online using an actor-critic training framework with a composite reward function to guide agent cooperation, as follows: During training, the centralized value critic network evaluates the value of the global state based on the joint state of all agents and global map information, and guides each distributed policy enforcer network to optimize its respective execution policy to maximize the expected long-term cumulative reward. The composite reward function is obtained by weighted summation using the following formula:

[0055] In the formula, The serial number of the intelligent agent is indicated. Indicates a time step. Indicates the first An agent at time step Compound rewards, Indicates the first The agents at the time step The task reward items, and if the first The agents at the time step If the target location is successfully reached and the virtual pickup action is completed, the task reward item will receive a positive fixed value reward; otherwise, the task reward item will be zero. Indicates the first The agents at the time step The collision penalty term, and at the time step If the first If the bounding box of an agent in the simulation scene overlaps with the bounding box of any other agent or static obstacle, the collision penalty term receives a negative fixed value reward; otherwise, the collision penalty term is set to zero. Indicates the first The agents at the time step The collaborative efficiency reward item, and at the time step Statistically analyze the travel behavior of all the aforementioned agents at the shared resource point. If the first... If an agent engages in alternating or following traffic without conflict with other agents within a pre-defined competitive area, the collaborative efficiency reward item receives a positive, non-fixed value that is dynamically calculated based on the smoothness of local traffic flow. Otherwise, the collaborative efficiency reward item is set to zero. , and These represent the preset positive weighting coefficients, and have... .

[0056] In step S73, the specific implementation examples of each item are as follows: (1) Task reward item When the intelligent agent When the navigation reaches within 0.5m of its current target location and triggers the "virtual pickup" signal, a fixed value of +10 is obtained; otherwise, it is 0. (2) Collision penalty item : Continuous bounding box collision detection in digital twins; if the agent If the bounding box of an agent overlaps with the bounding box of any other agent or static obstacle, a huge negative reward of -20 is obtained and the training round is terminated immediately; otherwise, it is 0. (3) Cooperative efficiency reward item To achieve "efficient traffic flow," it is necessary to quantify the "smoothness of local traffic flow." One specific implementation method in this embodiment is: in each preset competitive area (such as a junction of channels), an intelligent agent... Centered on the target, count all items within a 2m radius around it. The average velocity of (a positive integer) agents (including themselves) over the last 10 simulation steps And calculate the variance of these average velocities. The smaller the variance, the more orderly and smooth the traffic flow in the area; this reward can be designed as follows:

[0057] In the formula, Represented as a reward coefficient (e.g.) ), This represents the preset maximum allowable variance threshold, which is applied only when the agent... Located in a competitive area and This reward will only be calculated at that time.

[0058] In step S73, the specific details of the offline training and online fine-tuning include, but are not limited to, the following steps: (1) Data collection: In the digital twin environment, each agent collects data according to the current strategy. Interact with the environment to generate a large number of state-action-reward sequence trajectories, which are stored in a shared experience replay buffer; (2) Centralized training: sample a batch of data from the buffer, and first use a centralized critic network. Estimate the state value and calculate the advantage function. Then, using Guide the updating of all distributed executor networks The parameters are updated to maximize the policy objective function with pruning mechanism to ensure the stability of training; (3) Online fine-tuning: After the offline trained model is deployed to the actual system, it can continue to be fine-tuned online in the real digital twin environment using the actual generated scheduling task data to adapt to the subtle changes in warehouse layout or new operation mode; (4) Key hyperparameters: training discount factor γ=0.99, generalized advantage estimation (GAE) parameter λ=0.95, policy learning rate 3e-4, value function learning rate 1e-3, mini-batch size for each iteration is 1024, and the total number of training steps is not less than 1e7.

[0059] S74. For each AGV, the trained and corresponding distributed policy executor network independently generates the corresponding motion control command based on the real-time local observation information of the corresponding AGV.

[0060] S75. The motion control instructions of each AGV are combined to form a globally conflict-free collaborative picking instruction set planned by the multi-agent reinforcement learning algorithm for the multiple AGVs responsible for executing the simulated outbound task.

[0061] S8. Perform simulation verification on the planned collaborative picking instruction set, and after the verification is passed, send each motion control instruction in the collaborative picking instruction set to the corresponding AGV controller for execution.

[0062] In step S8, specifically, the planned collaborative pickup instruction set is simulated and verified, including but not limited to the following steps S81 to S82.

[0063] S81. In the warehouse digital twin model, the entire task process is accelerated through simulation based on the planned collaborative picking instruction set and the preset AGV motion dynamics model.

[0064] In step S81, specifically within a virtual environment, the corresponding twin model is driven to move strictly according to the path point sequence, speed, and action commands set for each AGV in the instruction set. This process is based on the preset AGV motion dynamics model, which precisely defines the physical characteristics of the AGV, such as the impact of maximum speed, acceleration, deceleration, turning radius, and load on motion, thereby ensuring that the simulation can realistically reflect the motion behavior and spatial and temporal occupancy of the AGV in the actual warehouse. The accelerated simulation refers to calculation and rendering at a speed much faster than real-time (e.g., a time ratio of 10:1 or higher), thereby completing a simulation of an actual work process that may take several minutes within seconds or tens of seconds, providing the possibility for rapid verification and iterative optimization.

[0065] S82. If deadlock, collision or task timeout is predicted during the simulation, the verification is deemed unsuccessful; otherwise, the verification is deemed successful.

[0066] In step S82, the following real-time monitoring and detection are performed during the accelerated simulation process: (1) Collision detection, i.e., continuously calculating whether the bounding boxes or geometric models between each AGV twin and between the AGV and static obstacles (shelves, walls) overlap. Once an overlap is detected, it is determined to be a collision; (2) Deadlock detection, i.e., monitoring the movement status of the AGV group. If two or more AGVs are detected to be stuck in a long-term (exceeding a preset threshold, such as 30 seconds) stagnation state in areas such as passages and intersections due to mutual obstruction, and it is predicted that this state cannot be resolved on its own, it is determined to be a deadlock; (3) Task timeout detection, i.e., setting a preset maximum allowable completion time threshold for the entire simulated outbound task or key sub-task. If any task is not completed within this threshold when the simulation ends, it is determined to be a timeout. As long as any of the above negative events are triggered during the simulation process, it is automatically determined that the verification fails; otherwise, if no event is triggered throughout the process, it is determined to pass. In addition, the result of verification failure will automatically trigger the re-planning of the path or the reassignment of tasks, i.e., return to the execution of steps S6 to S7.

[0067] After step S8, considering that in the traditional mode, the unavailability of equipment (such as power depletion or insufficient pressure) is often only discovered after arriving at the scene, resulting in delays in rescue, it is necessary to perform real-time status verification during picking and automatically, nearby, and quickly select alternatives based on the equipment association diagram knowledge base to achieve a closed loop from static planning to dynamic and reliable execution, ensuring the final effectiveness of the outbound plan. That is, preferably, after issuing each motion control command to the corresponding AGV controller for execution, the method also includes, but is not limited to, the following steps S91 to S92.

[0068] S91. During the process of the physical AGV performing the picking task, by reading the RFID tag on the picked target fire-fighting equipment or the Internet of Things sensor connected to the picked target fire-fighting equipment, it is verified whether the electrical quantity, pressure value or last maintenance time of the target fire-fighting equipment is within the available threshold range.

[0069] In step S91, the specific implementation includes, but is not limited to, the following: When the AGV arrives at the target location and performs a physical pickup, its onboard RFID reader or dedicated communication module (such as Bluetooth or ZigBee module) will automatically read the RFID tag attached to the equipment or establish a connection with the IoT sensor built into the equipment. The RFID tag pre-stores the equipment's basic identification information and the timestamp of the last maintenance. The IoT sensor directly monitors and reports the equipment's real-time status parameters, such as the remaining battery power (electrical quantity), the internal pressure value of the fire extinguisher, and the oil pressure value of the hydraulic equipment. After receiving this data, the management server immediately compares it with the available threshold range of that model of equipment pre-stored in the database (for example, the life detector's battery power must be >50%, and the hydraulic demolition tool's pressure value must be between 20-30 MPa). Verification is successful only if all verified parameters are within the threshold range; otherwise, it fails. This step ensures the immediate availability of the outgoing equipment.

[0070] S92. If the verification fails, other fire-fighting equipment shall be selected nearby to replace the target fire-fighting equipment according to the equipment association diagram, and the picking route of the affected AGV shall be replanned in real time.

[0071] In step S92, if the verification fails, an emergency replacement process is immediately initiated: First, using the faulty equipment as the starting point, a search is performed in the constructed fire equipment association graph. Based on the functional co-occurrence correlation and physical location proximity information implied by the edge weights in the graph, other available equipment with the most similar function and closest physical location to the faulty equipment is prioritized as a replacement. Then, only the affected AGV performing this retrieval task undergoes local replanning: In the warehouse digital twin model, using the location of the replacement equipment as the new target point, a pre-trained multi-agent decision-making model is used to quickly regenerate an optimal movement instruction from the current location to the new target point for the affected AGV, ensuring it does not conflict with other AGV paths. The new movement instruction is issued immediately, and the affected AGV proceeds to retrieve the replacement equipment, while the tasks of other AGVs remain unaffected. This step ensures that a single node failure does not lead to the interruption or significant delay of the entire outbound task, achieving a dynamic self-healing and resilient response mechanism in the event of unexpected failures.

[0072] Therefore, based on the modular management method for intelligent warehouse inbound and outbound operations described in steps S1 to S8 above, an integrated fire-fighting and warehousing emergency solution is provided, capable of understanding disaster situations, intelligent assembly, collaborative planning, and closed-loop verification. First, the disaster situation text information of the emergency fire-fighting command is acquired and parsed into a structured disaster situation vector using a pre-trained natural language processing model. Then, based on pre-built mapping rules, the vector is mapped to at least one fire-fighting functional module identifier. Next, an initial set of fire-fighting equipment is determined based on the identifier. Combined with an equipment association graph with fire-fighting equipment as nodes and functional relevance and location proximity as edge weights, a community detection algorithm generates a complete set of fire-fighting equipment with optimal internal functionality and space. Then, based on the real-time location of the equipment, a simulated outbound task is generated in a warehouse digital twin model. A multi-agent reinforcement learning algorithm is used to plan a globally conflict-free collaborative picking instruction set for multiple AGVs. Finally, when simulation verification is successful, instructions are sent to the AGV controller for execution. This achieves full-process automation and intelligence from intelligent disaster situation analysis to dynamic optimization and assembly of equipment and collaborative scheduling of multiple AGVs, greatly improving the efficiency of fire emergency response and facilitating practical application and promotion.

[0073] like Figure 2 As shown, the second aspect of this embodiment provides a virtual system for implementing the modular management method described in the first aspect. It is suitable for deployment in a management server that is connected to each automated guided vehicle (AGV) in an intelligent warehouse. The system includes a disaster text acquisition unit, a text parsing and processing unit, a functional module mapping unit, an equipment set initial determination unit, an equipment set final determination unit, a simulation task generation unit, a collaborative picking planning unit, and an instruction verification and issuance unit that are connected in sequence. The disaster text acquisition unit is used to acquire disaster text information of emergency fire fighting orders; The text parsing and processing unit is used to call a pre-trained natural language processing model to parse the disaster text information into a structured disaster situation vector, wherein the disaster situation vector contains at least two items from the following: disaster type, altitude of the disaster location, status of people at the disaster location, and type of hazardous materials at the disaster location. The functional module mapping unit is used to map the disaster scenario vector to a unique identifier of at least one fire protection functional module based on pre-built scenario and function mapping rules. The scenario and function mapping rules are used to define the correspondence between different disaster scenario vectors and multiple preset fire protection functional modules. Each fire protection functional module is associated with at least one mandatory fire protection equipment and at least one optional fire protection device. The equipment set initial determination unit is used to determine the initial fire-fighting equipment set for responding to the emergency fire-fighting command based on the unique identifier of the at least one fire-fighting functional module. The equipment set finalization unit is used to generate a complete fire equipment set for responding to the emergency fire command based on the initial fire equipment set and the pre-constructed fire equipment association graph through a community detection algorithm. The fire equipment association graph uses all fire equipment in the warehouse as nodes, and uses the weighted value of the functional co-occurrence correlation degree and physical storage location proximity between any two fire equipment as the weight of the edge connecting the two nodes that correspond one-to-one with any two fire equipment. The optimization goal of the community detection algorithm is to make the combined functional correlation and storage location proximity within the generated equipment set optimal. The simulation task generation unit is used to trigger the generation of a simulated outbound task containing multiple picking sub-tasks in the warehouse digital twin model of the intelligent warehouse inbound and outbound based on the real-time location of each fire-fighting equipment in the complete fire-fighting equipment set. The collaborative picking planning unit is used to plan a globally conflict-free collaborative picking instruction set for multiple AGVs responsible for executing the simulated outbound task based on a multi-agent reinforcement learning algorithm. The collaborative picking instruction set includes motion control instructions generated independently for each of the multiple AGVs. The instruction verification and issuance unit is used to simulate and verify the planned collaborative picking instruction set, and after the verification is passed, to issue each motion control instruction in the collaborative picking instruction set to the corresponding AGV controller for execution.

[0074] The working process, working details and technical effects of the aforementioned system provided in the second aspect of this embodiment can be found in the modular management method described in the first aspect, and will not be repeated here.

[0075] like Figure 3As shown, the third aspect of this embodiment provides a computer device for executing the modular management method described in the first aspect, including a storage module, a processing module, and a transceiver module connected in sequence for communication. The storage module stores a computer program, the transceiver module sends and receives messages, and the processing module reads the computer program and executes the modular management method described in the first aspect. Specifically, the storage module may include, but is not limited to, random-access memory (RAM), read-only memory (ROM), flash memory, first-in-first-out (FIFO) memory, and / or first-in-last-out (FILO) memory, etc.; the processing module may, but is not limited to, use a microprocessor of the STM32F105 series. Furthermore, the computer device may also include, but is not limited to, a power supply module, a display screen, and other necessary components.

[0076] The working process, working details and technical effects of the aforementioned computer equipment provided in the third aspect of this embodiment can be found in the modular management method described in the first aspect, and will not be repeated here.

[0077] This fourth aspect of the embodiment provides a computer-readable storage product that stores instructions comprising the modular management method described in the first aspect. Specifically, the computer-readable storage product stores instructions that, when executed on a computer, perform the modular management method described in the first aspect. The computer-readable storage product refers to a data storage medium, which may include, but is not limited to, computer-readable storage media such as floppy disks, optical disks, hard disks, flash memory, USB flash drives, and / or Memory Sticks. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices.

[0078] The working process, working details and technical effects of the aforementioned computer-readable storage product provided in the fourth aspect of this embodiment can be found in the modular management method described in the first aspect, and will not be repeated here.

[0079] This fifth aspect of the embodiment provides a computer program product, including a computer program or instructions, which, when executed by a computer, implements the modular management method described in the first aspect. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device.

[0080] Finally, it should be noted that the above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A modular management method for inbound and outbound operations based on intelligent warehousing, characterized in that, This is executed by the management server of each Automated Guided Vehicle (AGV) in the smart warehouse's inbound and outbound operations, including: Obtain disaster information in the form of emergency fire-fighting commands; A pre-trained natural language processing model is invoked to parse the disaster text information into a structured disaster situation vector, wherein the disaster situation vector contains at least two items from the following: disaster type, altitude of the disaster location, status of people at the disaster location, and type of hazardous materials at the disaster location; Based on pre-built scenario and function mapping rules, the disaster scenario vector is mapped to a unique identifier of at least one fire protection function module. The scenario and function mapping rules are used to define the correspondence between different disaster scenario vectors and multiple preset fire protection function modules. Each fire protection function module is associated with at least one mandatory fire protection equipment and at least one optional fire protection device. Based on the unique identifier of the at least one fire-fighting functional module, determine the initial set of fire-fighting equipment for responding to the emergency fire-fighting command; Based on the initial set of fire-fighting equipment and the pre-constructed fire-fighting equipment association graph, a complete set of fire-fighting equipment for responding to the emergency fire-fighting command is generated through a community detection algorithm. The fire-fighting equipment association graph uses all fire-fighting equipment in the warehouse as nodes, and uses the weighted value of the functional co-occurrence correlation degree and physical storage location proximity between any two fire-fighting equipment as the weight of the edge connecting the two nodes that correspond one-to-one with any two fire-fighting equipment. The optimization objective of the community detection algorithm is to make the combined functional correlation and storage location proximity within the generated equipment set optimal. Based on the real-time location of each fire-fighting equipment in the complete fire-fighting equipment set, a simulated outbound task containing multiple picking sub-tasks is triggered and generated in the warehouse digital twin model of the intelligent warehouse inbound and outbound. Based on a multi-agent reinforcement learning algorithm, a globally conflict-free collaborative picking instruction set is planned for multiple AGVs responsible for executing the simulated outbound task. The collaborative picking instruction set includes motion control instructions generated independently for each of the multiple AGVs. The planned collaborative picking instruction set is simulated and verified. After the verification is successful, each motion control instruction in the collaborative picking instruction set is sent to the corresponding AGV controller for execution.

2. The modular management method according to claim 1, characterized in that, Based on the unique identifier of the at least one fire-fighting functional module, an initial set of fire-fighting equipment for responding to the emergency fire-fighting command is determined, including: Based on the unique identifier of the at least one fire protection functional module, locate all the required fire protection equipment associated with the at least one fire protection functional module in the warehouse, and use the search results as the initial set of fire protection equipment for responding to the emergency fire protection command; For each fire protection functional module in the at least one fire protection functional module, if all optional fire protection equipment associated with the corresponding module can be found in the warehouse according to the corresponding unique identifier, then the specific equipment model and quantity are selected from all optional fire protection equipment according to at least one specific quantitative parameter in the disaster situation vector and the predefined equipment adaptation rule, and the selection result is added to the initial fire protection equipment set. The specific quantitative parameter includes at least one of the following: the height value of the disaster location, the estimated number of trapped people, and the type of hazardous chemicals. The equipment adaptation rule is used to define the mapping relationship between different quantitative parameter ranges and the model and quantity of optional fire protection equipment.

3. The modular management method according to claim 2, characterized in that, Based on the initial set of fire-fighting equipment and the pre-constructed fire-fighting equipment association graph, a complete set of fire-fighting equipment for responding to the emergency fire-fighting command is generated using a community detection algorithm, including: Using all equipment in the initial fire-fighting equipment set as target nodes, a community detection algorithm is run on a pre-constructed fire-fighting equipment association graph to find a connected subgraph that satisfies the following conditions: it contains all the target nodes and maximizes the ratio of the sum of the weights of the edges inside the subgraph to the sum of the weights of all possible edges in the subgraph. The fire-fighting equipment association graph uses all fire-fighting equipment in the warehouse as nodes, and uses the weighted value of the functional co-occurrence correlation degree and physical storage location proximity between any two fire-fighting equipment as the weight of the edge connecting the two nodes that correspond one-to-one with the two fire-fighting equipment. The optimization objective of the community detection algorithm is to make the combined functional correlation and storage location proximity within the generated equipment set optimal. The fire-fighting equipment corresponding to the non-target nodes contained in the found connected subgraph is added to the initial fire-fighting equipment set, thus forming a complete fire-fighting equipment set for responding to the emergency fire-fighting command.

4. The modular management method according to claim 1, characterized in that, Based on a multi-agent reinforcement learning algorithm, a globally conflict-free collaborative picking instruction set is planned for the multiple AGVs responsible for executing the simulated outbound task, including: Using the aforementioned warehouse digital twin model as a simulation environment, each AGV in the smart warehouse inbound and outbound processes is modeled as an intelligent agent. Construct a multi-agent deep reinforcement learning model, wherein the multi-agent deep reinforcement learning model includes a centralized value critic network and a distributed policy enforcer network corresponding one-to-one with each agent; The multi-agent deep reinforcement learning model is trained offline and fine-tuned online using an actor-critic training framework with a composite reward function to guide agent cooperation, as follows: During training, the centralized value critic network evaluates the value of the global state based on the joint state of all agents and global map information, and guides each distributed policy enforcer network to optimize its execution policy to maximize the expected long-term cumulative reward. The composite reward function is obtained by weighted summation using the following formula: In the formula, The serial number of the intelligent agent is indicated. Indicates a time step. Indicates the first An agent at time step Compound rewards, Indicates the first The agents at the time step The task reward items, and if the first The agents at the time step If the target location is successfully reached and the virtual pickup action is completed, the task reward item will receive a positive fixed value reward; otherwise, the task reward item will be zero. Indicates the first The agents at the time step The collision penalty term, and at the time step If the first If the bounding box of an agent in the simulation scene overlaps with the bounding box of any other agent or static obstacle, the collision penalty term receives a negative fixed value reward; otherwise, the collision penalty term is set to zero. Indicates the first The agents at the time step The collaborative efficiency reward item, and at the time step Statistically analyze the travel behavior of all the aforementioned agents at the shared resource point. If the first... If an agent engages in alternating or following traffic without conflict with other agents within a pre-defined competitive area, the collaborative efficiency reward item receives a positive, non-fixed value that is dynamically calculated based on the smoothness of local traffic flow. Otherwise, the collaborative efficiency reward item is set to zero. , and These represent the preset positive weighting coefficients, and have... ; For each AGV, the trained and corresponding distributed policy executor network independently generates corresponding motion control commands based on the real-time local observation information of the corresponding AGV. The motion control commands of each AGV are combined to form a globally conflict-free collaborative picking instruction set planned by the multi-agent reinforcement learning algorithm for the multiple AGVs responsible for executing the simulated outbound task.

5. The modular management method according to claim 1, characterized in that, The collaborative pickup instruction set obtained from the planning was simulated and verified, including: In the warehouse digital twin model, the entire task process is accelerated through simulation based on the planned collaborative picking instruction set and the preset AGV motion dynamics model. If a deadlock, collision, or task timeout is predicted during the simulation, the verification is deemed unsuccessful; otherwise, the verification is deemed successful.

6. The modular management method according to claim 1, characterized in that, After issuing each motion control command to the corresponding AGV controller for execution, the method further includes: During the process of physical AGV performing a picking task, the electrical quantity, pressure value or last maintenance time of the target fire-fighting equipment is verified to be within the available threshold range by reading the RFID tag on the target fire-fighting equipment or the IoT sensor on the target fire-fighting equipment connected to the communication. If the verification fails, other fire-fighting equipment will be selected nearby to replace the target fire-fighting equipment according to the equipment association diagram, and the retrieval route of the affected AGV will be replanned in real time.

7. A modular management system based on intelligent warehouse inbound and outbound operations, characterized in that, It is suitable for deployment in the management server of each Automated Guided Vehicle (AGV) in the smart warehouse, which is connected to the AGVs in the inbound and outbound processes. It includes a disaster text acquisition unit, a text parsing and processing unit, a functional module mapping unit, an equipment set initial determination unit, an equipment set final determination unit, a simulation task generation unit, a collaborative picking planning unit, and an instruction verification and issuance unit that are connected in sequence. The disaster text acquisition unit is used to acquire disaster text information of emergency fire fighting orders; The text parsing and processing unit is used to call a pre-trained natural language processing model to parse the disaster text information into a structured disaster situation vector, wherein the disaster situation vector contains at least two items from the following: disaster type, altitude of the disaster location, status of people at the disaster location, and type of hazardous materials at the disaster location. The functional module mapping unit is used to map the disaster scenario vector to a unique identifier of at least one fire protection functional module based on pre-built scenario and function mapping rules. The scenario and function mapping rules are used to define the correspondence between different disaster scenario vectors and multiple preset fire protection functional modules. Each fire protection functional module is associated with at least one mandatory fire protection equipment and at least one optional fire protection device. The equipment set initial determination unit is used to determine the initial fire-fighting equipment set for responding to the emergency fire-fighting command based on the unique identifier of the at least one fire-fighting functional module. The equipment set finalization unit is used to generate a complete fire equipment set for responding to the emergency fire command based on the initial fire equipment set and the pre-constructed fire equipment association graph through a community detection algorithm. The fire equipment association graph uses all fire equipment in the warehouse as nodes, and uses the weighted value of the functional co-occurrence correlation degree and physical storage location proximity between any two fire equipment as the weight of the edge connecting the two nodes that correspond one-to-one with any two fire equipment. The optimization goal of the community detection algorithm is to make the combined functional correlation and storage location proximity within the generated equipment set optimal. The simulation task generation unit is used to trigger the generation of a simulated outbound task containing multiple picking sub-tasks in the warehouse digital twin model of the intelligent warehouse inbound and outbound based on the real-time location of each fire-fighting equipment in the complete fire-fighting equipment set. The collaborative picking planning unit is used to plan a globally conflict-free collaborative picking instruction set for multiple AGVs responsible for executing the simulated outbound task based on a multi-agent reinforcement learning algorithm. The collaborative picking instruction set includes motion control instructions generated independently for each of the multiple AGVs. The instruction verification and issuance unit is used to simulate and verify the planned collaborative picking instruction set, and after the verification is passed, to issue each motion control instruction in the collaborative picking instruction set to the corresponding AGV controller for execution.

8. A computer device, characterized in that, It includes a storage module, a processing module, and a transceiver module that are sequentially connected in communication. The storage module is used to store computer programs, the transceiver module is used to send and receive messages, and the processing module is used to read the computer programs and execute the modular management method as described in any one of claims 1 to 6.

9. A computer-readable storage product, characterized in that... The computer-readable storage product stores instructions that, when executed on a computer, perform the modular management method as described in any one of claims 1 to 6.

10. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or the instructions are executed by the computer, they implement the modular management method as described in any one of claims 1 to 6.

Images