Bus station multi-role intelligent service and autonomous inspection system and method based on generative large model

Abstract

The application discloses a kind of based on generative large model's bus station multi-role intelligent service and autonomous inspection system and method, including intelligent mobile service terminal, sensing and data acquisition module, business knowledge and task management module, generative large model workflow decision module, task state control module, multi-role intelligent interaction module, task execution module and event response and closed-loop control module, through the multi-stage decision workflow of generative large model as high-level control core, in combination with task state control and the architecture of separation of bottom execution, the automatic switching and closed-loop execution of driver guide task, station inspection task and multi-role service are realized on the same intelligent mobile service terminal, significantly improve the operation efficiency of bus station, shorten the driver preparation time, improve the abnormal event discovery timely rate and inspection coverage, through the state switching mechanism and abnormal response closed-loop control of role driving, enhance the safety and stability of bus station operation.

Description

Technical Field

[0001] This invention belongs to the field of intelligent service technology for public transportation stations, specifically relating to a multi-role intelligent service and autonomous inspection system and method for public transportation stations based on a generative large model. Background Technology

[0002] With the continuous expansion of urban public transportation systems, bus stations have evolved into complex operational scenarios integrating vehicle dispatching, personnel organization, safety management, and facility support.

[0003] However, the following problems still exist in the operation of existing bus depots: 1. Low efficiency in driver preparation: After entering the depot, drivers need to sign in, confirm their shifts for the day, and go to the corresponding bus. The existing system relies heavily on static schedules, vehicle number lookups, or manual dispatching to provide information. Drivers need to rely on experience to find their vehicles, especially in large or multi-area depots, where the average time to find a vehicle is as long as 6-15 minutes, which can easily lead to delays and reduced on-time departure rates; 2. Low utilization rate and slow response of depot inspection resources: Inspection work mainly relies on manual timed patrols or fixed monitoring equipment. The inspection process is independent of the driver service process, making it difficult to achieve 24-hour service. 1. 24 / 7 coverage is insufficient for timely detection of abnormal events such as fire smoke, water accumulation, oil leaks, structural anomalies, improper vehicle parking, and leftover foreign objects, which can easily lead to safety hazards or operational disruptions. 2. Lack of unified control over multi-role services and task execution: Existing intelligent devices have limited functionality and cannot dynamically switch task modes based on the service recipient's role (driver, station manager, passenger), resulting in high equipment idle rates, frequent manual intervention, and failure to achieve coordinated scheduling of service and inspection tasks. 3. Ineffective integration of multi-source heterogeneous data: Scheduling data, real-time vehicle location, monitoring images, environmental sensor data, etc., are stored in a scattered manner, lacking a unified business knowledge system to support high-level semantic decision-making and control state switching, making it difficult to fully realize the value of the data. To address these issues, we propose a multi-role intelligent service and autonomous inspection system and method for bus stations based on a generative large model. Summary of the Invention

[0004] The purpose of this invention is to provide a multi-role intelligent service and autonomous inspection system and method for public transportation stations based on a generative large model, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a multi-role intelligent service and autonomous inspection system for bus depots based on a generative large model, comprising: an intelligent mobile service terminal deployed within the bus depot for autonomously moving within the depot area and performing service and inspection tasks; a perception and data acquisition module for collecting operational data generated by the intelligent mobile service terminal during operation; a business knowledge and task management module for structuring the operational data, constructing a bus depot business knowledge graph, and inputting the bus depot business knowledge graph and real-time operational data as retrieval sources into the generative large model to form a semantic knowledge system with retrieval-enhanced generation capabilities; a generative large model workflow decision module for performing semantic understanding, service object role identification, and task status decision on the current bus depot operation status based on the semantic knowledge system and according to a preset multi-stage decision workflow, and outputting structured task status decision results; and a task status control module for receiving the generative large model workflow decision results. The workflow decision module outputs the task status decision result, and uses the task status decision result as the control input to drive the same intelligent mobile service terminal to switch between different task states; the multi-role intelligent interaction module is used to determine the corresponding service object role according to the task status decision result, and generate interaction instructions that match the task state; the task execution module is used to execute the corresponding driver guidance task, station inspection task, or information service task according to the task status decision result and interaction instructions; the event response and closed-loop control module is used to trigger the generative large model workflow decision module to re-execute the decision workflow when an abnormal event is detected, generate an abnormal response task state, and drive the system to return to the preset normal task state after the abnormal event is handled; wherein, the system uses the task status decision result output by the generative large model workflow decision module as the unified control basis to realize the automatic switching and closed-loop execution of driver guidance tasks, station inspection tasks, and multi-role intelligent service tasks within the same intelligent mobile service terminal.

[0006] Preferably, the station operation data collected by the sensing and data acquisition module includes at least one or more of the following: driver check-in data, station environment image data, environmental status data, bus location information, and station task scheduling data.

[0007] Preferably, the multi-stage decision workflow of the generative large model workflow decision module includes at least the following stages: Operational data and bus station business knowledge graph fusion stage: used to perform structured processing and correlation modeling on real-time collected station operational data, current task status, and historical task execution results, mapping the data into a unified semantic representation, and retrieving business rules and historical handling information related to the current operational status from the station business knowledge graph based on a retrieval-enhanced generation mechanism, constructing a model context for generative large model reasoning; Service object role determination stage: used to determine the service object role of the current interaction object based on the model context; Task status decision stage: using... The system generates a target task state based on the service object role and predefined task state transition rules, and outputs a clear task type state. The structured execution instruction generation stage generates structured task instructions to drive task execution based on the target task state. The target task state output in the task state decision stage serves as a core control signal, driving the task state control module to switch task states and execute subsequent tasks. The generative large-scale model workflow decision module does not directly output execution action control signals, but instead outputs structured task state decision results, including task state decisions and task instructions, which are parsed and executed by the task state control module and the task execution module, respectively.

[0008] Preferably, the service target roles include at least driver roles, station management personnel roles, passenger roles, and unmanned interaction states, which are automatically determined by the generative big model workflow decision module based on the real-time operating context of the bus station; when the service target is determined to be a passenger role, the multi-role intelligent interaction module provides services such as bus timetable query, route consultation, ticket information notification, or travel guidance based on the generative big model.

[0009] Preferably, the task status includes at least driver guidance task status, station inspection task status, multi-role information service task status, and abnormal response task status; the triggering conditions for task status switching include at least one or more of the following: driver check-in event, interaction object change event, and abnormal detection event.

[0010] Preferably, when in the driver guidance task state, the task execution module queries the driver's daily shift information and the real-time parking location of the corresponding vehicle, and the intelligent mobile service terminal guides the driver to the target vehicle along the planned path in an autonomous navigation manner; and after the perception and data acquisition module receives the driver's check-in data, it triggers the generative large model workflow decision module to enter the task state.

[0011] Preferably, the intelligent mobile service terminal is equipped with a visible light camera and an infrared thermal imaging camera. In the unmanned interaction state or the station inspection task state, the generative large model workflow decision module analyzes the collected station environment images to identify abnormal events such as fire smoke, ground water accumulation, oil leakage, abnormal facility structure, abnormal vehicle parking, and foreign objects left behind.

[0012] Preferably, after the event response and closed-loop control module detects an abnormal event, it feeds back the abnormal information to the generative large model workflow decision module, generates an abnormal response task status and corresponding handling instructions, and drives the system to revert to the task status before the abnormal event occurred or the default inspection task status.

[0013] Preferably, it also includes a cloud-based scheduling and model optimization platform, which is used to aggregate task execution logs, anomaly handling results and interaction records, and to continuously iterate and optimize the generative large model and the bus station business knowledge graph.

[0014] A multi-role intelligent service and autonomous inspection method for bus depots based on a generative large model is applied to a multi-role intelligent service and autonomous inspection system for bus depots based on a generative large model, including the following steps: S1: Collect bus depot operation data through an intelligent mobile service terminal; S2: Perform structured processing on the operation data to construct a bus depot business knowledge graph and form a business knowledge set for generative large model reasoning; S3: Input the business knowledge set and real-time operation data into the generative large model, execute a multi-stage decision workflow, and output structured task state decision results; S4: Use the structured task state decision results as the control basis to drive the same intelligent... The mobile service terminal switches between different task states and executes corresponding driver guidance, station inspection, or multi-role information service tasks. The task state switching is based on a predefined set of task states and state transition rules. The triggering conditions for the state transition rules include at least driver check-in events, interaction object change events, and abnormal events. S5: When an abnormal event is detected, the multi-stage decision-making workflow is re-executed to generate an abnormal response task state and handling instructions. After the abnormal event is handled, the task state is rolled back to form a closed loop. S6: Based on the task execution log and results, the bus station business knowledge graph is updated and the generative big model is optimized to achieve system self-learning.

[0015] Compared with the prior art, the beneficial effects of the present invention are:

[0016] This invention utilizes a multi-stage decision-making workflow based on a generative large model as the high-level control core. Combined with an architecture that separates task state control from low-level execution, it achieves automatic switching and closed-loop execution of driver guidance tasks, station inspection tasks, and multi-role services on the same intelligent mobile service terminal. This significantly improves the operational efficiency of bus stations, shortens driver preparation time, increases the timeliness of anomaly detection and inspection coverage, while reducing equipment redundancy and manual intervention costs. Through a role-driven state switching mechanism and anomaly response closed-loop control, it enhances the safety and stability of station operations, demonstrating significant economic and social benefits, as detailed below:

[0017] 1. Significantly improve operational efficiency and reduce time costs: Addressing the issue of low driver preparation efficiency, this invention utilizes a fully intelligent guidance system encompassing sign-in triggering, real-time vehicle positioning, and autonomous navigation. This reduces the average time drivers spend finding their vehicles from the traditional 6-15 minutes to 2-4 minutes, a reduction of 60%-73%, effectively reducing the risk of departure delays. Simultaneously, the on-time departure rate increases from 92% to 98.5%, significantly optimizing the timeliness of public transportation services and enhancing the passenger travel experience. 2. Strengthen inspection coverage and anomaly response, improving station safety: This invention utilizes intelligent mobile service terminals for 24-hour autonomous inspections, combined with the multi-dimensional data collection capabilities of visible light and infrared thermal imaging dual cameras, increasing daily inspection coverage from 60%-70% to 100%. The system enables comprehensive on-site inspections, and the generative large-scale model accurately identifies abnormal events such as fire smoke and ground water accumulation, increasing the timeliness of anomaly detection from 65% to 96%, an improvement of 48%. Combined with a closed-loop control mechanism for anomaly identification, response, handling, and rollback, it significantly reduces the probability of safety hazards evolving into accidents, enhancing the stability of on-site operations. 3. Optimized resource allocation and reduced operating and equipment costs: This invention achieves automatic switching between driver guidance, inspection, and multi-role information services through a single intelligent mobile service terminal, increasing equipment utilization from 40% to 88%, an improvement of 120%. It avoids redundant procurement and maintenance costs of multiple sets of single-function equipment. Simultaneously, autonomous inspections replace manual scheduled inspections, and intelligent guidance reduces manual intervention in dispatching, significantly reducing the on-site's manual operating costs. 4. Maximize data value and improve decision-making accuracy: This invention enhances generation capabilities through the construction and retrieval of business knowledge graphs, integrating scattered scheduling data, vehicle location, environmental sensor data, etc., into a unified semantic knowledge system. This provides accurate decision support for generative large-scale models. The multi-stage decision-making workflow further ensures the accuracy of role recognition and task switching. For example, the accuracy rate of passenger role semantic understanding and the success rate of driver guidance tasks both reach over 98%, avoiding decision-making bias and service errors caused by data fragmentation. 5. Enhance system security and controllability, and ensure stable operation: This invention adopts a decoupled architecture for decision-making and execution. The generative large-scale model only outputs structured decision results and does not directly control hardware actions. The task status control module and execution module parse and execute the results, avoiding the risk of misoperation at the architectural level. At the same time, the highest priority design for abnormal response states and the automatic rollback mechanism after handling ensure a seamless connection between rapid response in emergency scenarios and regular operations. The probability of continuous system operation without failure reaches 99%.5% or higher; 6. Possesses continuous self-learning capabilities to adapt to dynamic operational needs: Through a cloud-based scheduling and model optimization platform, this invention can continuously iterate and optimize the generative large model and business knowledge graph based on task execution logs, anomaly handling results, and interaction records. For example, it can optimize the semantic understanding capability of interactions through high-frequency passenger consultation data, and update the association rules of the knowledge graph through anomaly handling feedback. This allows the system to continuously improve service accuracy and inspection efficiency as the station's operational scenarios change, with long-term operational effects continuously optimized without frequent secondary development. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall system structure of the present invention; Figure 2 This is a schematic diagram of the multi-stage decision-making workflow of the generative large model of the present invention; Figure 3 This is a schematic diagram of the task state switching process of the present invention; Figure 4 This is a schematic diagram of the driver guidance task process of the present invention; Figure 5 This is a schematic diagram of the site inspection and anomaly response process of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figure 1The present invention provides a multi-role intelligent service and autonomous inspection system for bus depots based on a generative large model, comprising: an intelligent mobile service terminal deployed within the bus depot for autonomous movement and execution of service and inspection tasks within the depot area; a perception and data acquisition module for collecting operational data generated by the intelligent mobile service terminal during operation; a business knowledge and task management module for structuring the operational data, constructing a bus depot business knowledge graph, and inputting the bus depot business knowledge graph and real-time operational data as retrieval sources into the generative large model to form a semantic knowledge system with retrieval-enhanced generation capabilities; a generative large model workflow decision module for performing semantic understanding, service object role identification, and task status decision-making based on the semantic knowledge system and a preset multi-stage decision-making workflow, outputting structured task status decision results; and a task status control module for receiving the output from the generative large model workflow decision module. The system comprises a task status decision module, which uses the task status decision results as control input to drive the same intelligent mobile service terminal to switch between different task states; a multi-role intelligent interaction module, which determines the corresponding service object role based on the task status decision results and generates interaction instructions that match the task state; a task execution module, which executes the corresponding driver guidance task, station inspection task, or information service task based on the task status decision results and interaction instructions; and an event response and closed-loop control module, which triggers the generative large model workflow decision module to re-execute the decision workflow when an abnormal event is detected, generates an abnormal response task state, and drives the system to return to the preset normal task state after the abnormal event is handled. The system uses the task status decision results output by the generative large model workflow decision module as the unified control basis to achieve automatic switching and closed-loop execution of driver guidance tasks, station inspection tasks, and multi-role intelligent service tasks within the same intelligent mobile service terminal.

[0021] In this embodiment, the station operation data collected by the sensing and data acquisition module includes at least one or more of the following: driver check-in data, station environment image data, environmental status data, bus location information, and station task scheduling data.

[0022] In this embodiment, the multi-stage decision-making workflow of the generative large model workflow decision module includes at least the following stages: Operational data and bus station business knowledge graph fusion stage: This stage is used to perform structured processing and correlation modeling on real-time collected station operational data, current task status, and historical task execution results, mapping the data into a unified semantic representation. Based on a retrieval-enhanced generation mechanism, it retrieves business rules and historical handling information related to the current operational status from the station business knowledge graph, constructing a model context for generative large model reasoning; Service object role determination stage: This stage is used to determine the service object role of the current interaction object based on the model context; Task status decision. The process is divided into two phases: Phase 1: Generates the target task state based on the service object role and predefined task state transition rules, and outputs a clear task type state; Phase 2: Generates structured task instructions based on the target task state to drive task execution; The target task state output by the task state decision phase serves as the core control signal, driving the task state control module to switch task states and execute subsequent tasks; The generative large model workflow decision module does not directly output execution action control signals, but instead outputs structured task state decision results, including task state decisions and task instructions, which are parsed and executed by the task state control module and the task execution module, respectively.

[0023] In this embodiment, the service target roles include at least driver roles, station management personnel roles, passenger roles, and unmanned interaction states, which are automatically determined by the generative big model workflow decision module based on the real-time operating context of the bus station. When the service target is determined to be a passenger role, the multi-role intelligent interaction module provides services such as bus timetable query, route consultation, ticket information notification, or travel guidance based on the generative big model.

[0024] In this embodiment, the task status includes at least the driver guidance task status, the station inspection task status, the multi-role information service task status, and the abnormal response task status.

[0025] The triggering conditions for task state switching include at least one or more of the following: driver check-in event, interaction object change event, and anomaly detection event.

[0026] In this embodiment, when in the driver guidance task state, the task execution module queries the driver's daily shift information and the real-time parking location of the corresponding vehicle. The intelligent mobile service terminal guides the driver to the target vehicle along the planned path using autonomous navigation. After the perception and data acquisition module receives the driver's check-in data, it triggers the generative large model workflow decision module to enter the task state.

[0027] In this embodiment, the intelligent mobile service terminal is equipped with a visible light camera and an infrared thermal imaging camera. In the unmanned interaction state or the station inspection task state, the generative large model workflow decision module analyzes the collected station environment images to identify abnormal events such as fire smoke, ground water accumulation, oil leakage, abnormal facility structure, abnormal vehicle parking, and foreign objects left behind.

[0028] In this embodiment, after the event response and closed-loop control module detects an abnormal event, it feeds back the abnormal information to the generative large model workflow decision module, generates an abnormal response task status and corresponding handling instructions, and drives the system to revert to the task status before the abnormal event occurred or the default inspection task status.

[0029] This embodiment also includes a cloud-based scheduling and model optimization platform, which is used to aggregate task execution logs, anomaly handling results, and interaction records to continuously iterate and optimize the generative large model and the bus station business knowledge graph.

[0030] This invention also provides a method for multi-role intelligent service and autonomous inspection of bus depots based on generative large models, applied to a multi-role intelligent service and autonomous inspection system for bus depots based on generative large models, including the following steps: S1: Collect bus depot operation data through an intelligent mobile service terminal; S2: Perform structured processing on the operation data to construct a bus depot business knowledge graph and form a business knowledge set for generative large model reasoning; S3: Input the business knowledge set and real-time operation data into the generative large model, execute a multi-stage decision workflow, and output structured task state decision results; S4: Use the structured task state decision results as the control basis to drive the same... The intelligent mobile service terminal switches between different task states and executes corresponding driver guidance, station inspection, or multi-role information service tasks. The task state switching is based on a predefined set of task states and state transition rules. The triggering conditions for the state transition rules include at least driver check-in events, interaction object change events, and abnormal events. S5: When an abnormal event is detected, the multi-stage decision-making workflow is re-executed to generate an abnormal response task state and handling instructions. After the abnormal event is handled, the task state is rolled back to form a closed loop. S6: Based on the task execution log and results, the bus station business knowledge graph is updated and the generative big model is optimized to achieve system self-learning.

[0031] This embodiment uses a bus station in a city with an area of ​​approximately 50,000 square meters and an average of 200 bus trips per day as a pilot scenario. The system is deployed on a wheeled intelligent mobile service terminal (robot). The terminal is equipped with LiDAR, UWB positioning, a high-definition visible light camera, an infrared thermal imaging camera, a microphone, a speaker, and a 5G communication module. The generative large model is based on a fine-tuned open-source Transformer architecture and integrates the RAG mechanism. The knowledge graph is built using Neo4j, specifically including:

[0032] Intelligent mobile service terminal: The intelligent mobile service terminal is the physical execution body of the system. It adopts a differential wheel chassis with a maximum moving speed of 1.5m / s. It supports autonomous navigation in the entire area of ​​the work station. The terminal integrates a multi-sensor fusion positioning system (UWB+IMU+laser SLAM) with a positioning accuracy of ≤10cm, ensuring reliable obstacle avoidance and path planning in the complex environment of the bus station. The terminal also undertakes voice interaction, image acquisition and task execution functions, realizing the hardware unification of service and inspection tasks.

[0033] The perception and data acquisition module, integrated into the intelligent mobile service terminal hardware, includes a high-definition visible light camera (1920×1080 resolution), an infrared thermal imaging camera (640×512 resolution), a microphone array, and environmental sensors (temperature, smoke, humidity). During operation, it collects driver check-in data (voice / face), station environment images, vehicle location information (obtained via vehicle GPS or station positioning base station), and task scheduling data (interfaced with the station scheduling system) in real time. The collected data is published in the form of ROS messages, with a sampling frequency of 10Hz for images and 1Hz for environmental data to ensure real-time performance.

[0034] The business knowledge and task management module runs on a hybrid architecture of edge computing unit and cloud in intelligent mobile service terminal. Real-time running data is structured and then used to build a dynamic knowledge graph. For example, entity nodes include "driver ID", "vehicle number", "parking space number" and "facility point", and relationship edges include "daily allocation", "real-time parking" and "inspection required". The knowledge graph is updated once a day and is used together with real-time running data as a vector retrieval source input to the generative large model to form a RAG-enhanced semantic knowledge system. This module ensures that the large model can accurately retrieve station-specific business knowledge when making decisions, such as the association between driver scheduling and vehicle location.

[0035] Generative large model workflow decision modules, such as Figure 2 As shown, this module is the core decision-making unit, executing a pre-defined multi-stage workflow:

[0036] The operational data and knowledge graph fusion stage: This stage is used to perform structured processing and correlation modeling on real-time collected station operational data, current task status, and historical task execution results. The data is mapped into a unified semantic representation. Based on a retrieval-enhanced generation mechanism, business rules and historical handling information related to the current operational status are retrieved from the station's business knowledge graph to construct a model context for generative large-scale model reasoning. In the system's multi-stage decision-making workflow, the "operational data and knowledge graph fusion stage" serves as the starting point for constructing the context of the generative large-scale model. Its core purpose is to integrate real-time operational data (such as driver check-in data, station environment images, vehicle locations, etc.), current task status (such as during inspection or interaction), and historical task execution results (such as past anomaly handling logs and inspection coverage records) into a unified semantic context for use in subsequent stages. This stage is implemented based on a retrieval-enhanced generation (RAG) mechanism.

[0037] Main processing logic:

[0038] (1) Structured processing of real-time operation data: The real-time operation data acquired by the perception and data acquisition module is first processed in a structured manner: 1) Site environment image data: target features or abnormal feature vectors are extracted through computer vision model (YOLO); 2) Vehicle location and terminal location information: converted into spatial coordinate entities or area identifiers; 3) Driver check-in data and task scheduling data: parsed into event entities with timestamps to represent the current interaction context;

[0039] (2) Organization of business knowledge graph: The business knowledge graph of bus station is pre-constructed as an entity-relationship-attribute model, wherein: entity nodes include: vehicles, drivers, abnormal events, station areas, task types, etc.; relations include: location, association, trigger, handling, membership, etc.; attributes are used to describe business rules, priorities and constraints.

[0040] (3) Incremental integration of historical task execution results: Historical task execution results (such as anomaly handling logs, inspection completion rate, and task time) are regarded as dynamic incremental knowledge. The processing methods include: parsing historical execution logs into structured task result entities; vectorizing them based on graph embedding algorithm (GraphSAGE); and updating them as new nodes or attributes in the business knowledge graph.

[0041] Integration of core mechanisms:

[0042] (1) Combination of vector retrieval and graph structure reasoning: In the context construction process: 1) Vectorize the real-time running data and the current task status; 2) Perform similarity retrieval in the vector index space of the knowledge graph; 3) Based on the retrieved relevant nodes, trigger graph traversal (such as depth-first or limited-level traversal) to expand related entities and historical handling experience.

[0043] (2) Quantification of historical task execution results: Time decay weighting mechanism: Set decay weights based on the time difference of task execution to make recent tasks have a greater impact on current decisions; Execution effect weighting: Give higher weights to historical tasks with higher success rates and lower false alarm rates; Selection of Top-k related historical items: Select only a limited number of historical records that are most relevant to the current scenario to participate in context construction.

[0044] Comparative Verification of Decision-Making Effects Through Fusion of Operational Data and Business Knowledge Graph: To verify the performance advantages of the "fusion stage of operational data and business knowledge graph" in this invention, the system of this invention was compared with two baseline methods through simulation testing:

[0045] Simulation test environment: Based on historical operation data of bus depots, a Transformer-based generative large model (parameter scale 7B) was used. The fusion method adopted a GraphRAG-like mechanism (vector retrieval + graph traversal). The baselines included: "plainRAG": only vector retrieval of real-time operation data (such as check-in, images), without knowledge graph; "graph-only": only static graph query, without real-time data fusion. Test dataset: 1000 decision queries (including driver guidance, anomaly detection, and inspection routes), based on historical logs and real-time simulation. Indicators calculated: decision accuracy (the proportion of correct task status outputs, manually labeled and verified); response speed (end-to-end latency, ms). The results are shown in Table 1 below. Table 1: ;

[0046] Service object role determination stage: Based on context, semantic analysis is performed to determine the role (driver, manager, passenger) or the state of no interaction;

[0047] Task status decision stage: Generate the target task status based on the role and state transition rules (such as sign-in event → driver guidance state); Structured execution instruction generation stage: Output task instructions in JSON format (such as {"task_type":"driver_guide","vehicle_position":"Area A, No. 12"});

[0048] This module does not directly control the hardware; it only outputs high-level decision results. The task status control module and the multi-role intelligent interaction module, such as... Figure 3As shown, this is the collaborative working process between the task status control module and the multi-role intelligent interaction module: the task status decision result output by the generative large model workflow decision module is transmitted to the task status control module as a control input. The task status control module analyzes and switches the current task status based on the preset task status set and state transition rules. The task status includes at least the idle inspection status, driver guidance status, inspection execution status and abnormal response status.

[0049] Table 2 below shows the parameter table for task status and service object role: ;

[0050] After the task status is determined, the task status control module sends the current task status and the corresponding service object role information to the multi-role intelligent interaction module. The multi-role intelligent interaction module generates a natural language interaction command that matches the current task status and outputs the interaction command to the task execution module to drive the intelligent mobile service terminal to perform the corresponding driver guidance task, station inspection task or information service task.

[0051] When an abnormal event is detected in any task state, the task state control module triggers the system to enter an abnormal response state. Through event response and closed-loop control mechanisms, it drives the generative large model workflow decision module to re-execute the decision process. After the abnormal event is handled, the system returns to the preset normal task state, thus achieving closed-loop control based on task state. Through this structure, the system uses the task state decision results output by the generative large model as a unified control basis, enabling automatic switching between multiple task states and collaborative execution of multi-role intelligent interactions within the same intelligent mobile service terminal. For example, in driver guidance mode, it might announce, "Mr. Zhang, you are driving vehicle XX today, located at position 15 in section B. Please follow me." The task execution module, under the unified scheduling of the task state control module, parses the task state decision results and corresponding interaction instructions output by the generative large model workflow decision module, and drives the same intelligent mobile service terminal to execute corresponding task behaviors in different task states.

[0052] The task execution module supports at least the following task execution methods:

[0053] (1) Driver guidance task execution method, such as Figure 4As shown: When the task status decision result indicates a driver guidance task status, the task execution module queries the driver's daily shift information and the real-time parking location of the corresponding bus, generates a guidance path based on the location information, and controls the intelligent mobile service terminal to move autonomously within the station in a follow-guide mode, maintaining a preset safe distance (e.g., 1-2 meters) from the driver, and guides the driver to the target bus along the planned path.

[0054] (2) The execution method of station inspection tasks, such as Figure 5 As shown: When the task status decision result indicates that the inspection task status is, the task execution module controls the intelligent mobile service terminal to move along the preset or dynamically generated inspection path, and calls the perception and data acquisition module in real time to collect the site environment images and status data during the movement, for subsequent anomaly identification and analysis.

[0055] (3) Information service task execution mode: When the task status decision result indicates the information service task status, the task execution module calls the multi-role intelligent interaction module, and provides information services such as bus time query, route information, ticket consultation or travel guidance to the service object in the form of voice or text based on the interaction instructions generated by the generative big model; through the above method, the task execution module can realize the flexible switching and collaborative execution of driver guidance task, station inspection task and multi-role information service task under the unified control of the task status decision result output by the generative big model workflow;

[0056] Event response and closed-loop control module:

[0057] like Figure 5 As shown, if an anomaly is detected in any state, the large model workflow is immediately re-executed, generating an anomaly response state (alarm, notification to administrators). After the handling is completed (if confirmed by administrators), it automatically reverts to the state before the anomaly, forming a closed loop; the specific process is as follows:

[0058] When the task status decision result output by the generative large model workflow decision module indicates an inspection task status or a system in an unmanned interactive state, the task status control module drives the system to enter the station inspection task execution process. In this state, the task execution module controls the intelligent mobile service terminal to move autonomously within the bus station along a preset inspection path or a dynamically generated inspection path based on the station's operating conditions. During the movement, the terminal continuously calls the perception and data acquisition module to collect station environmental image data and environmental status data. The collected image data and environmental status data are transmitted to the generative large model workflow decision module in real time or periodically to identify abnormal events such as fire smoke, ground water accumulation, oil leaks, abnormal facility structures, abnormal vehicle parking, or foreign objects left behind. When the generative large model workflow decision module determines that an abnormal event exists, it outputs an abnormal response task status, and the task status control module drives the system to switch from the inspection task status to the abnormal response status, entering the abnormal handling process. After the abnormal handling is completed, the system returns to the inspection task status or the preset normal task status, forming a closed-loop inspection and abnormal response mechanism.

[0059] According to the above embodiments, the multi-role intelligent service and autonomous inspection system for bus depots based on generative large models of the present invention was deployed in a bus depot for pilot operation. The pilot period was three consecutive months. During the pilot period, the system operated stably and was compared with the traditional manual management method without the deployment of the present invention under the same bus depot conditions.

[0060] Through statistical analysis of system operation logs, task execution records, and site operation data, the quantitative effect comparison results shown in Table 3 below were obtained:

[0061] Table 3: ;

[0062] As shown in Table 3, the average time drivers spend finding their vehicles has been significantly reduced from 6-15 minutes using the traditional method to 2-4 minutes, a reduction of 60%-73%. This effectively reduces drivers' preparation time and provides crucial support for on-time departures. The timely detection rate of station anomalies has increased from 65% to 96%, an improvement of 48%, ensuring that safety hazards such as fire smoke, water accumulation, and oil leaks can be quickly identified. The daily inspection coverage rate has increased from 60%-70% to 100%, achieving comprehensive inspection within the station, an improvement of 43%. Equipment utilization has doubled from 40% to 88%, completely changing the problem of traditional intelligent equipment having limited functionality and high idle rates. The on-time departure rate has increased from 92% to 98.5%, an improvement of 7%, significantly improving the quality of public transportation services.

[0063] The event response and closed-loop control module in this invention can significantly reduce the exposure time of safety risks within the site by automatically identifying and handling abnormal events such as fire smoke, water accumulation on the ground, oil leaks, abnormal facility structures, abnormal vehicle parking, and foreign objects left behind. This reduces the causes of accidents from a technical perspective. For example, fire smoke detection can trigger a fire extinguishing response in a timely manner, reducing the occurrence of fires; water accumulation / oil leak detection can prevent people from slipping; and abnormal vehicle parking or foreign objects left behind can prevent the risk of collisions or tripping.

[0064] During the pilot phase, the system improved the timeliness of anomaly detection from 65% to 96% (an increase of 48%), which directly correlated with a reduction in safety incidents. According to statistics from the pilot phase, the total number of anomaly samples was approximately 460, with an average of approximately 5 anomaly samples per day; among them: water / oil spills accounted for approximately 38%; foreign object residue accounted for approximately 27%; abnormal vehicle parking accounted for approximately 21%; and facility and other anomalies accounted for approximately 14%.

[0065] All of the above anomalies entered a closed-loop process of "anomaly response-handling-state rollback". The above results show that the present invention effectively shortens the exposure time of safety risks and reduces the probability of engineering conditions for safety accidents by early identification and closed-loop handling of abnormal events.

[0066] Based on a three-month field pilot operation, the system of this invention underwent a six-month continuous simulation test. The simulation adopted the Monte Carlo method, using the monthly historical task execution results to feed back an iterative generative large model and optimize the RAG semantic knowledge system. Key indicators included fault-free operation rate, anomaly identification accuracy, and task execution success rate. The specific results are shown in Table 4.

[0067] Table 4: ;

[0068] This invention utilizes a multi-stage decision-making workflow based on a generative large model as the high-level control core. Combined with an architecture that separates task state control from low-level execution, it achieves automatic switching and closed-loop execution of driver guidance tasks, station inspection tasks, and multi-role services on the same intelligent mobile service terminal. This significantly improves the operational efficiency of bus stations, shortens driver preparation time, increases the timeliness of anomaly detection and inspection coverage, while reducing equipment redundancy and manual intervention costs. Through a role-driven state switching mechanism and anomaly response closed-loop control, it enhances the safety and stability of station operations, demonstrating significant economic and social benefits, as detailed below:

[0069] I. Significantly improve operational efficiency and reduce time costs: Addressing the issue of low driver preparation efficiency, this invention provides intelligent guidance throughout the entire process, including check-in triggering, real-time vehicle positioning, and autonomous navigation. This reduces the average time drivers spend finding their vehicles from the traditional 6-15 minutes to 2-4 minutes, a reduction of 60%-73%. This effectively reduces the risk of departure delays. At the same time, the on-time departure rate increases from 92% to 98.5%, significantly optimizing the timeliness of public transportation services and improving the passenger travel experience.

[0070] II. Enhanced Inspection Coverage and Anomaly Response to Improve Site Safety: This invention utilizes intelligent mobile service terminals for 24-hour autonomous inspections, combined with the multi-dimensional data acquisition capabilities of visible light and infrared thermal imaging dual cameras, increasing the daily inspection coverage from 60%-70% to 100%, achieving comprehensive site inspections. The generative large-scale model accurately identifies abnormal events such as fire smoke and ground water accumulation, increasing the timeliness of anomaly detection from 65% to 96%, an improvement of 48%. Coupled with a closed-loop control mechanism for anomaly identification, response, handling, and rollback, it significantly reduces the probability of safety hazards evolving into accidents, enhancing the operational stability of the site.

[0071] 3. Optimize resource allocation and reduce operating and equipment costs: This invention realizes automatic switching of driver guidance, inspection and multi-role information services through the same intelligent mobile service terminal, which increases the equipment utilization rate from 40% to 88%, an increase of 120%. It avoids the redundant procurement and maintenance costs of multiple sets of single-function equipment. At the same time, autonomous inspection replaces manual timed inspection and intelligent guidance reduces dispatching manual intervention, which significantly reduces the station's manual operation costs.

[0072] IV. Maximizing data value and improving decision-making accuracy: This invention enhances generation capabilities through the construction and retrieval of business knowledge graphs, integrating scattered scheduling data, vehicle location, environmental sensor data, and other data into a unified semantic knowledge system. This provides accurate decision support for generative large models, and the multi-stage decision-making workflow further ensures the accuracy of role recognition and task switching. For example, the accuracy rate of passenger role semantic understanding and the success rate of driver guidance tasks both reach over 98%, avoiding decision-making bias and service errors caused by data fragmentation.

[0073] V. Enhanced System Security and Controllability, Ensuring Stable Operation: This invention adopts a decoupled decision-making and execution architecture. The generative large model only outputs structured decision results and does not directly control hardware actions. The task status control module and execution module parse and execute the results, avoiding the risk of misoperation at the architectural level. At the same time, the highest priority design for abnormal response status and the automatic rollback mechanism after handling ensure seamless connection between rapid response in emergency scenarios and regular operations. The probability of continuous system operation without failure reaches over 99.5%. VI. Continuous Self-Learning Capability, Adapting to Dynamic Operational Needs: Through cloud scheduling and model optimization platforms, this invention can continuously iterate and optimize the generative large model and business knowledge graph based on task execution logs, abnormal handling results, and interaction records. For example, it can optimize the semantic understanding capability of interactions through high-frequency passenger consultation data and update the association rules of the knowledge graph through abnormal handling feedback. This allows the system to continuously improve service accuracy and inspection efficiency as the station operation scenario changes, with continuous optimization of long-term operation results, eliminating the need for frequent secondary development.

[0074] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multi-role intelligent service and autonomous inspection system for bus depots based on a generative large model, characterized in that, include: Intelligent mobile service terminals are deployed within bus depots to autonomously move within the depot area and perform service and inspection tasks. The sensing and data acquisition module is used to collect the operational data generated by the intelligent mobile service terminal during operation; The business knowledge and task management module is used to perform structured processing on the operational data, construct a bus station business knowledge graph, and use the bus station business knowledge graph and real-time operational data as retrieval source inputs to a generative large model to form a semantic knowledge system with retrieval enhancement and generation capabilities. The generative large model workflow decision module is used to perform semantic understanding, service object role identification and task status decision on the current bus station operation status based on the semantic knowledge system and according to the preset multi-stage decision workflow, and output structured task status decision results. The task status control module is used to receive the task status decision results output by the generative large model workflow decision module, and use the task status decision results as control input to drive the same intelligent mobile service terminal to switch between different task states. The multi-role intelligent interaction module is used to determine the corresponding service object role based on the task status decision result, and generate interaction instructions that match the task status. The task execution module is used to execute corresponding driver guidance tasks, station inspection tasks, or information service tasks based on the task status decision results and interaction instructions. The event response and closed-loop control module is used to trigger the generative large model workflow decision module to re-execute the decision workflow when an abnormal event is detected, generate an abnormal response task state, and drive the system to return to the preset normal task state after the abnormal event is handled. The system uses the task status decision results output by the generative large model workflow decision module as the unified control basis, and realizes automatic switching and closed-loop execution of driver guidance tasks, station inspection tasks and multi-role intelligent service tasks within the same intelligent mobile service terminal. The multi-stage decision workflow of the generative large model workflow decision module includes at least the following stages: The stage of integrating operational data with the knowledge graph of bus station business is used to perform structured processing and correlation modeling on the real-time collected station operation data, current task status and historical task execution results, map the data into a unified semantic representation, and retrieve business rules and historical disposal information related to the current operation status from the station business knowledge graph based on the retrieval enhancement generation mechanism, and construct a model context for generative large model reasoning. Service object role determination stage: used to determine the service object role of the current interaction object based on the model context; Task status decision stage: This stage is used to generate the target task status based on the service object role and predefined task status transition rules, and output a clear task type status. Structured execution instruction generation stage: used to generate structured task instructions to drive task execution based on the target task state; The target task state output in the task state decision-making stage serves as the core control signal, which drives the task state control module to switch task states and execute subsequent tasks. The generative large model workflow decision module does not directly output execution action control signals, but outputs structured task state decision results, including task state decisions and task instructions, which are parsed and executed by the task state control module and the task execution module, respectively. The intelligent mobile service terminal is equipped with a visible light camera and an infrared thermal imaging camera. In the unmanned interactive state or the station inspection task state, the generative large model workflow decision module analyzes the collected station environment images to identify abnormal events such as fire smoke, ground water accumulation, oil leakage, abnormal facility structure, abnormal vehicle parking, and foreign objects left behind. After detecting an abnormal event, the event response and closed-loop control module feeds back the abnormal information to the generative large model workflow decision module, generates an abnormal response task status and corresponding handling instructions, and drives the system to revert to the task status before the abnormal event occurred or the default inspection task status.

2. The multi-role intelligent service and autonomous inspection system for bus depots based on generative large models according to claim 1, characterized in that: The station operation data collected by the sensing and data acquisition module includes at least one or more of the following: driver check-in data, station environment image data, environmental status data, bus location information, and station task scheduling data.

3. The multi-role intelligent service and autonomous inspection system for bus stations based on a generative large model as described in claim 2, characterized in that: The service recipient roles include at least driver roles, station management personnel roles, passenger roles, and unmanned interaction states, which are automatically determined by the generative large model workflow decision module based on the real-time operating context of the bus station. When the service recipient is determined to be a passenger, the multi-role intelligent interaction module provides services such as shuttle bus timetable query, route consultation, ticket information notification, or travel guidance based on a generative big data model.

4. The multi-role intelligent service and autonomous inspection system for bus depots based on generative large models according to claim 1, characterized in that: The task status includes at least driver guidance task status, station inspection task status, multi-role information service task status, and abnormal response task status. The triggering conditions for task state switching include at least one or more of the following: driver check-in event, interaction object change event, and anomaly detection event.

5. The multi-role intelligent service and autonomous inspection system for bus stations based on a generative large model as described in claim 4, characterized in that: When in driver guidance task state, the task execution module queries the driver's daily shift information and the real-time parking location of the corresponding vehicle, and the intelligent mobile service terminal guides the driver to the target vehicle along the planned path in an autonomous navigation manner. Furthermore, after the perception and data acquisition module receives the driver's sign-in data, it triggers the generative large model workflow decision module to enter the task state.

6. The multi-role intelligent service and autonomous inspection system for bus depots based on a generative large model as described in claim 1, characterized in that: It also includes a cloud-based scheduling and model optimization platform, which aggregates task execution logs, anomaly handling results, and interaction records to continuously iterate and optimize generative large models and bus station business knowledge graphs.

7. A method for multi-role intelligent service and autonomous inspection of bus depots based on generative large models, applied to the multi-role intelligent service and autonomous inspection system for bus depots based on generative large models as described in any one of claims 1-6, characterized in that, Includes the following steps: S1: Collect bus station operation data through intelligent mobile service terminals; S2: Perform structured processing on the operational data, construct a business knowledge graph of bus stations, and form a set of business knowledge for generative large model reasoning; S3: Input business knowledge sets and real-time operational data into a generative large model, execute a multi-stage decision-making workflow, and output structured task status decision results; S4: Based on the structured task state decision results, drive the same intelligent mobile service terminal to switch between different task states and execute corresponding driver guidance, station inspection or multi-role information service tasks. Task state switching is based on a predefined set of task states and state transition rules. The triggering conditions for the state transition rules include at least driver check-in events, interaction object change events, and abnormal events. S5: When an abnormal event is detected, the multi-stage decision-making workflow is re-executed to generate the abnormal response task status and handling instructions. After the abnormal event is handled, the task status is rolled back to form a closed loop. S6: Based on task execution logs and results, update the bus station business knowledge graph and optimize the generative large model to achieve system self-learning.

Images