A three-dimensional parking garage autonomous operation and maintenance system and method based on a large language model

Abstract

This invention discloses an autonomous operation and maintenance system and method for automated parking garages based on a large language model. The system includes a central intelligent agent layer, a toolset and workflow layer, and a knowledge enhancement layer. The central intelligent agent layer includes a system cognition and decision-making hub, a chain-based reasoning and decision generator, and a formalized safety constraint verifier. The system cognition and decision-making hub is a large language model with tool invocation capabilities. The toolset and workflow layer includes a perception and data fusion pipeline module, a prediction and health management toolset, a planning and scheduling module, and an execution and monitoring workflow module, all communicating with the central intelligent agent layer. The knowledge enhancement layer includes a knowledge enhancement retrieval module and a vector database, also communicating with the central intelligent agent layer. Through the above setup, an upgrade from automatic early warning to safe and compliant autonomous operation and maintenance is achieved, significantly improving the intelligence level, operational efficiency, and safety reliability of automated parking garages.

Description

Technical Field

[0001] This invention belongs to the field of autonomous operation and maintenance technology for intelligent parking lots, specifically relating to an autonomous operation and maintenance system and method for three-dimensional parking garages based on a large language model. Background Technology

[0002] With the continuous and rapid growth of car ownership in my country, the shortage of ground-level parking spaces in urban core areas is widening. Automated parking garages, with their high space utilization and flexible deployment, have become a core solution to the urban parking problem. Currently, mainstream static traffic equipment such as lift-and-slide and vertical circulation automated parking garages are widely used in various scenarios. However, existing monitoring and maintenance systems for static traffic automated parking garages still face technical bottlenecks in long-term operation: traditional automated parking garages rely on manual inspections or preset threshold alarms, making it difficult to detect hidden faults such as wire rope wear, motor bearing aging, and sensor drift in real time. According to industry statistics, approximately 62% of automated parking garage safety accidents are caused by untimely fault diagnosis. While some existing remote monitoring solutions have achieved basic data transmission and alarm functions, their decision-making logic is rigid and unable to make dynamic judgments based on complex contexts (such as equipment historical status, environmental factors, and usage frequency), resulting in high false alarm rates and a high risk of missed alarms.

[0003] In recent years, domestic static traffic research has proposed fault diagnosis methods for three-dimensional parking garages based on supporting SVM vector machines, GRU deep learning, or knowledge graphs. For example, classification models are trained by collecting sensor data such as vibration and current, or fault knowledge graphs are constructed to assist maintenance personnel in querying. However, such solutions have the following key defects: (1) Lack of semantic understanding and dynamic context adaptation: Such models cannot understand natural language instructions such as "there have been many rainstorms recently, and non-emergency faults should be handled with priority for waterproof equipment", nor can they integrate external dynamic information such as weather and traffic flow to dynamically adjust maintenance priorities. (2) Decision-making is disconnected from physical safety constraints: Existing prediction models only output fault probabilities or provide maintenance suggestions, but cannot autonomously generate maintenance plans that meet hard safety constraints such as parking garage PLC logic interlock, anti-fall device timing, and motor thermal overload. For example, without confirming that the vehicle platform is unloaded and the anti-fall device is locked, no intelligent agent may issue a "lateral movement" command. Currently, all solutions rely on manual completion of this safety verification, which is inefficient and prone to errors. (3) Semantic fusion of multi-source heterogeneous data has not been achieved: There is still a semantic gap between vibration spectrum, current waveform and maintenance text record. The model cannot, like a senior engineer, directly infer "current drop and increase in high frequency vibration components" as "drive key wear".

[0004] Existing monitoring systems typically only store parameters locally, lacking effective integration between data from different sensors, between equipment operation data and maintenance records, and between the parking garage system and external environmental information. The system cannot form a holistic understanding of equipment health status, let alone support preventative maintenance decisions based on full lifecycle data. Furthermore, expanding the functionality of existing systems (such as adding new sensors, changing alarm rules, or integrating with third-party parking platforms) usually requires modifying PLC programs or underlying code, resulting in long deployment cycles and high costs. Additionally, the interaction between maintenance personnel and the system is limited to fixed dashboards and preset operation buttons, hindering complex queries, analyses, and command issuance via natural language, thus limiting further improvements in maintenance efficiency.

[0005] In summary, existing automated parking garage operation and maintenance systems are still in a semi-automatic stage, lacking an intelligent central hub with advanced cognitive capabilities, the ability to autonomously understand complex contexts, perform chain-like reasoning, and simultaneously meet underlying physical safety constraints. How to introduce large language model intelligent agent technology into the field of industrial equipment operation and maintenance, and construct a closed-loop autonomous operation and maintenance system encompassing perception, diagnosis, decision-making, safety verification, and execution, is a pressing technical problem that needs to be solved in this field. Summary of the Invention

[0006] In order to overcome the shortcomings of the existing technology, the purpose of this invention is to provide an autonomous operation and maintenance system and method for three-dimensional parking garages based on a large language model.

[0007] The objective of this invention is achieved through the following technical solution:

[0008] An autonomous operation and maintenance system for a three-dimensional parking garage based on a large language model includes a central intelligent agent layer, a toolset and workflow layer, and a knowledge enhancement layer.

[0009] The central intelligent agent layer includes a system cognition and decision-making hub, a chain reasoning and decision generator, and a formal safety constraint verifier. The system cognition and decision-making hub is a large language model with tool invocation capabilities.

[0010] The toolset and workflow layer includes a perception and data fusion pipeline module, a prediction and health management toolset, a planning and scheduling module, and an execution and monitoring workflow module that are communicatively connected to the central intelligent agent layer.

[0011] The knowledge enhancement layer includes a knowledge enhancement retrieval module and a vector database that are communicatively connected to the central intelligent agent layer;

[0012] The infrastructure layer includes a parking space terminal layer and an execution layer;

[0013] The chain reasoning and decision generator responds to abnormal events or user commands by calling the prediction and health management toolset for quantitative assessment, calling the knowledge enhancement retrieval module to retrieve relevant background knowledge, and inputting the assessment results and background knowledge into the system's cognition and decision-making center for chain reasoning to generate a preliminary decision plan.

[0014] The reasoning process of the chain reasoning and decision generator adopts an explicit thought chain prompting strategy, specifically including the following steps:

[0015] T1: Target parsing, which parses abnormal events or user commands into structured task targets G=(type, constraints, priority), where type belongs to the set {emergency repair, preventive maintenance, information query, scheduling optimization}, and constraints include time windows, security levels, resource limitations, etc. priority is the task urgency indicator, with a value range of predefined priority levels (high, medium, low), used to determine the response order and resource allocation weights among multiple pending tasks.

[0016] T2: Tool orchestration. Based on the task type, it retrieves matching tools from the tool registry and generates an ordered sequence of tool calls, with the sorting based on the data dependencies between tools.

[0017] T3: Evidence collection, sequentially executing tool calls, with the result of each call serving as context input for the next call, collecting quantitative evaluation results.

[0018] T4: Multi-objective trade-off. The collected evidence E is input into the large language model, which is required to explicitly calculate the comprehensive score for each candidate strategy s: score(s) = α × safety(s) + β × efficiency(s) + γ × satisfaction(s), where α, β, and γ are configurable weight coefficients, with preset values ​​of α = 0.5, β = 0.3, and γ = 0.2. safety(s) ∈ [0,1] represents the probability of failure deterioration and the probability of remaining life assessment, efficiency(s) is based on the impact of expected maintenance time and vehicle retrieval efficiency, and satisfaction(s) is based on the peak-hour impact coefficient.

[0019] T5: Strategy generation. Select the strategy output with the highest comprehensive score as the initial decision plan, while retaining the top-3 candidate plans as alternatives.

[0020] The formal safety constraint verifier integrates a time-series safety rule base. It receives preliminary decision schemes and converts them into structured action sequences, checking each structured action sequence to see if it satisfies the relevant constraints in the time-series safety rule base. If all constraints are met, it outputs a compliance signal and transmits the decision scheme to the planning and scheduling module. If any clause is violated, it generates violation feedback information with the reason for the violation and correction suggestions, returning it to the chain reasoning and decision generator to trigger it to re-perform reasoning to generate a corrected decision scheme, until a compliant decision scheme is obtained or the preset number of retries is reached.

[0021] Unlike traditional model checking methods, this invention does not treat formal verification as a separate post-processing step, but rather embeds it within the decision-making loop of the large language model agent: the violation feedback information output by the formal safety constraint verifier, containing the reasons for violations and correction suggestions, is directly injected into the context of the large language model's prompt words in a structured form, triggering chain-like reasoning and the decision generator to perform targeted corrections instead of reasoning from scratch. This closed-loop mechanism of verification-feedback-correction differs from the traditional binary model of post-generation verification and rejection upon violation detection, and can automatically generate compliant correction schemes without relying on human intervention.

[0022] Optionally, the source of the timing safety rule base is: the hardware interlock logic of the garage PLC controller, the mandatory safety clauses of the general safety requirements for parking equipment in the national standard for automated parking garages (GB / T26559-2011), and the safety operating procedures in the technical manuals provided by the equipment manufacturers. These three types of knowledge are uniformly converted into linear timing logic formulas for storage.

[0023] The timing safety rule base contains no fewer than 27 core safety rules, covering all safety interlock conditions for three types of actions: lifting, lateral movement, and rotation.

[0024] The time-series security rule base supports incremental updates: when a new device type is added or a new security vulnerability is discovered, the administrator can describe the new security rule in natural language, which will be automatically converted into the corresponding time-series logic formula by the formal security constraint verifier and stored in the rule base.

[0025] The temporal security rule base uses linear temporal logic formulas to store security constraints, and the transformation rules include:

[0026] a) "B must be satisfied before A can be executed" is mapped to □(A→(B));

[0027] b) "Prohibit the execution of A under condition C" is mapped to □(A→(¬C));

[0028] c) “Executing A must be followed by executing B” is mapped to □(A→O (B));

[0029] Where □ represents “always”, O represents “next moment”, and ¬ represents “logical NOT”.

[0030] Furthermore, the structured action sequence of the formal security constraint verifier is represented using a six-tuple structure:

[0031] Action=(action id, action type, target device, preconditions, postconditions, timestamp);

[0032] Wherein, action id is the unique identifier of the action, action type is the enumerated value of the action type, targetdevice is the identifier of the target device, preconditions is the set of preconditions, postconditions is the set of postconditions, and timestamp is the planned execution time.

[0033] The formal security constraint verifier employs a model verification algorithm to perform security verification on the decision scheme: 1) Security rules from the time-series security rule base and the system state transition model are input into the verification engine. The system state model consists of the current actual state plus all possible states that can be reached after the execution action sequence. 2) Within a preset step size (default setting 10 steps, configurable), the verifier exhaustively searches for execution paths that violate security rules. 3) If a rule-violation path is found, the first violating action and the reason for the violation are immediately extracted from the path, and a structured violation feedback is output. Initially, the maximum number of reachable states is set to 30, the rule base size is 27, and the single verification time is less than 100 milliseconds, meeting real-time operation and maintenance requirements. When the rule base size increases, the verifier supports incremental verification: only newly added or modified rules are re-verified, and existing rule results are cached and reused.

[0034] The formal security constraint verifier performs a step-by-step verification process: it analyzes the preliminary decision scheme to extract the corresponding structured action sequence, obtains the current system state from the perception and data fusion pipeline module, matches the relevant rules in the temporal security rule base for each action, and outputs a compliance signal for the structured action sequence when all rules pass; otherwise, it outputs feedback that the structured action sequence violates the rules.

[0035] The planning and scheduling module is used to generate an execution plan based on the compliance decision-making scheme;

[0036] The execution and monitoring workflow module is used to issue and track execution plans, and trigger replanning when deviations are detected.

[0037] Furthermore, the perception and data fusion pipeline module is used to access and standardize multi-source heterogeneous operating data from the parking space terminal layer to form raw numerical features. The multi-source heterogeneous operating data accessed by the perception and data fusion pipeline module includes at least two of the following: wire rope tension, motor current, vibration spectrum, photoelectric limit switch status, fall arrestor status, vehicle overlength / overweight detection signals, panoramic video stream, and environmental temperature and humidity data. The perception and data fusion pipeline module is also equipped with a semantic feature extraction unit, which is used to convert numerical sensor data into semantic labels describing the health status of equipment through a pre-trained mapping model.

[0038] The pre-trained mapping model in the semantic feature extraction unit is a supervised learning model. It uses historical sensor data and corresponding maintenance record annotations as training samples. The model outputs semantic labels and confidence score pairs, where the semantic labels are predefined. These labels include three dimensions: component name, fault mode, and severity level. When multiple semantic labels conflict, the semantic state fusion engine performs weighted fusion based on the confidence scores of each label and combines this with causal constraints in the fault propagation knowledge graph for consistency verification, generating a robust estimate of the equipment's health status.

[0039] Furthermore, the central intelligent agent layer also includes a semantic state fusion engine, which receives semantic tags and raw numerical features from the perception and data fusion pipeline module, combines them with the fault propagation knowledge graph obtained from the knowledge enhancement retrieval module, and constructs a dynamic causal graph of the current garage system. The semantic state fusion engine can identify inconsistencies between multi-sensor data and use the implicit reasoning of the large language model of the system's cognition and decision-making center to generate robust estimates of the equipment health status.

[0040] Furthermore, the prediction and health management toolset includes at least two of the following tools: a fault prediction tool based on support vector machines, a component remaining life prediction tool based on deep learning, and an unsupervised anomaly detection tool based on autoencoders.

[0041] Furthermore, the central intelligent agent layer also includes a context memory and interaction management unit, which is used to maintain a short-term state during dialogue or task execution, record historical decisions, user feedback and execution results, support multi-turn interaction and complex task decomposition, and be used for subsequent decision self-optimization.

[0042] An autonomous operation and maintenance method for automated parking garages based on large language models, applied to the aforementioned autonomous operation and maintenance system for automated parking garages based on large language models, includes the following steps:

[0043] S1: When the perception and data fusion pipeline module collects operational data, generates abnormal events, or receives user instructions, it generates structured event descriptions and pushes them to the central intelligent agent layer.

[0044] S2: The central intelligent agent parses the structured event description, calls the prediction and health management toolset and knowledge enhancement retrieval module to perform chain reasoning, and generates a preliminary decision plan with fault root cause judgment and maintenance strategy;

[0045] S3: Convert the preliminary decision plan into a structured action sequence, check whether each action meets all relevant constraints in the time-series safety rule base, and if a violation is found, generate a violation feedback and return to S2 to regenerate;

[0046] S4: The central intelligent agent layer sends the compliance decision-making scheme to the planning and scheduling module, generating an execution plan with personnel, time, spare parts and scheduling strategies;

[0047] S5: The execution and monitoring workflow module issues execution plans and tracks their status. If deviations occur, a replanning is triggered.

[0048] S6: After the task is completed, the central intelligent agent layer generates an analysis report and stores the processing process in the vector database.

[0049] Furthermore, in step S3, the violation feedback information adopts a structured JSON format, which includes a unique violation identifier, the corresponding security rule number, a natural language description of the rule, the violation action steps, violation action details, the current system status, the expected conditions to be met, and correction suggestions.

[0050] The present invention has the following beneficial effects:

[0051] 1. Most existing automated parking garage operation and maintenance systems remain at the stage of automated monitoring plus manual decision-making. Predictive solutions only output failure probabilities or maintenance suggestions, while safety verification and planning still rely on manual completion. This invention uses a large language model intelligent agent as the core of cognition and decision-making, and integrates a formal safety constraint verifier to construct a complete closed loop from perception, diagnosis, decision-making, verification, execution to learning. The system can autonomously complete root cause analysis, strategy generation, safety verification, resource coordination, and task distribution. All action sequences are executed only after passing the temporal safety rule verification, achieving a qualitative leap from automatic early warning to safe and compliant autonomous operation and maintenance.

[0052] 2. This invention utilizes the semantic understanding and chain reasoning capabilities of a large language model to enable the system to parse implicit constraints in natural language instructions, integrate external dynamic information for comprehensive judgment, and automatically generate the globally optimal processing strategy by balancing multiple dimensions such as equipment security, operational efficiency, and user satisfaction through a chain reasoning generator.

[0053] 3. This invention embeds a formal safety constraint verifier into the decision-making link of the intelligent agent, transforms the safety operating procedures into linear temporal logic formulas to form an executable rule base, and performs strict temporal safety model verification on the action sequence. Once a violation is found, it triggers re-reasoning and correction, thus eliminating safety accidents caused by model reasoning deviations from the architecture.

[0054] 4. This invention adopts an architecture design in which a central intelligent agent communicates and calls with various modules such as the perception and data fusion pipeline module, the prediction and health management toolset, the knowledge enhancement retrieval module, and the planning and scheduling workflow module through a standardized tool call interface. When adding a new sensor or prediction algorithm, only a new tool module needs to be developed and registered to the toolset layer. There is no need to modify the core code of the central intelligent agent or the underlying PLC logic. When connecting to a new external system, only the corresponding API adapter needs to be developed, which greatly reduces the complexity and cost of system upgrades and cross-scenario deployments.

[0055] 5. This invention uses a semantic feature extraction unit in the perception and data fusion pipeline to uniformly convert numerical sensor data into semantic tags that describe the health status of equipment. At the same time, through a knowledge-enhanced retrieval module, real-time perception data is deeply integrated with equipment manuals, historical maintenance records, fault case libraries, and fault propagation knowledge graphs in a vector database, forming a global understanding of the health status of equipment.

[0056] 6. This invention tracks the task execution status in real time through the execution and monitoring workflow module. When a deviation is detected, it automatically triggers a replanning process, generates a corrective decision plan, and passes security verification again, ensuring the integrity of the entire closed loop from plan generation to execution. Attached Figure Description

[0057] Figure 1 This is a schematic diagram of the overall architecture of the autonomous operation and maintenance system for three-dimensional parking garages based on a large language model, as described in this invention.

[0058] Figure 2 This is a flowchart of the autonomous operation and maintenance method for automated parking garages based on a large language model, as described in this invention.

[0059] Figure 3 This is a timing diagram of the interaction and processing of the system of the present invention in the scenario of Embodiment 4. Detailed Implementation

[0060] 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 this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0061] Before providing a detailed description of the autonomous operation and maintenance system and method for automated parking garages based on a large language model, this invention first defines some terms used in this invention:

[0062] ① Chain reasoning refers to the process by which a large language model, within a given context, decomposes an external input problem into several sub-problems and solves them step by step through multi-step logical reasoning.

[0063] ② Semantic labels refer to human-understandable device status descriptions obtained by converting numerical sensor data through a pre-trained mapping model. The format is semantic category and confidence score.

[0064] ③Structured action sequences refer to the conversion of natural language decision-making schemes into action data objects composed of predefined fields that can be parsed by computer programs.

[0065] Example 1

[0066] An autonomous operation and maintenance system for automated parking garages based on a large language model, such as Figure 1 and Figure 2 As shown, it includes a user layer, a central intelligent agent layer, a knowledge enhancement layer, a toolset and workflow layer, and an infrastructure layer.

[0067] The user layer includes an operations and maintenance management module and a third-party management system module. The operations and maintenance management module communicates with the central intelligent agent through natural language interaction. The third-party management system module includes meteorological monitoring data, urban parking management system, intelligent building system, etc., and exchanges data and controls with the central intelligent agent through API gateway.

[0068] The central intelligent agent layer is the core of the system's cognition and decision-making. It includes a system cognition and decision-making hub, a semantic state fusion engine, a chained reasoning and decision generator, a formal safety constraint verifier, and a context memory and interaction management unit. The system cognition and decision-making hub is a pre-trained large language model configured as an intelligent agent with tool invocation capabilities. It invokes and coordinates the semantic state fusion engine, chained reasoning and decision generator, formal safety constraint verifier, and context memory and interaction management unit to complete tasks. Specifically, the central intelligent agent layer adopts a large language model based on the Qwen-2.5-72B-Instruct model with fine-tuned instructions, configured with a context memory length of 8k tokens, and a tool invocation timeout of 30 seconds.

[0069] The knowledge enhancement layer directly connects to the system's cognitive and decision-making center. It includes a knowledge enhancement retrieval module and a vector database, both communicating with the central intelligent agent layer. The knowledge enhancement retrieval module connects to a pre-built vector database, which stores equipment technical manuals, historical maintenance records, a fault case library, safe operating procedures, and a fault propagation knowledge graph. The fault propagation knowledge graph records the causal chains of faults between equipment components in the form of triples, for example: wire rope wear → tension reduction → fall arrestor false triggering. The knowledge enhancement retrieval module is configured to receive query vectors, retrieve background knowledge relevant to the current context from the vector database, and return search results containing fault propagation paths, historical similar cases, and safety constraint rules.

[0070] The toolset and workflow layer includes a perception and data fusion pipeline module, a planning and scheduling module, an execution and monitoring workflow module, and a predictive and health management toolset, all communicating with the central intelligent agent layer. The infrastructure layer of a multi-level parking garage typically includes a parking space terminal layer and an execution layer; the parking space terminal layer typically consists of PLCs, various sensors, and cameras, while the execution layer typically includes maintenance personnel terminals, garage PLC controllers, and parking guidance screens.

[0071] The perception and data fusion pipeline module is used to access and standardize multi-source heterogeneous operational data from the parking space terminal layer in real time to form raw numerical features. This multi-source heterogeneous operational data includes at least two of the following: wire rope tension, motor current, vibration spectrum, photoelectric limit switch status, fall arrestor status, vehicle overlength / overweight detection signals, panoramic video stream, and environmental temperature and humidity data. The perception and data fusion pipeline module is also equipped with a semantic feature extraction unit, which converts numerical sensor data (such as the frequency band energy distribution of the vibration spectrum, the distortion rate of the current waveform, etc.) into semantic labels describing the health status of the equipment through a pre-trained mapping model, such as "mid-stage bearing wear," "insufficient guide rail lubrication," and "excessive transmission key clearance," thereby eliminating the semantic gap between raw data and high-level inference.

[0072] The pre-trained mapping model in the semantic feature extraction unit is a supervised learning model. It uses historical sensor data and corresponding maintenance record annotations as training samples. The model outputs semantic labels and confidence score pairs, where the semantic labels are predefined. These labels include three dimensions: component name, fault mode, and severity level. When multiple semantic labels conflict, the semantic state fusion engine performs weighted fusion based on the confidence scores of each label and combines this with causal constraints in the fault propagation knowledge graph for consistency verification, generating a robust estimate of the equipment's health status.

[0073] The predictive and health management toolset includes at least two of the following tools: a fault prediction tool based on support vector machines, a component remaining life prediction tool based on deep learning, and an unsupervised anomaly detection tool based on autoencoders.

[0074] The semantic state fusion engine receives raw numerical features and semantic labels from the perception and data fusion pipeline module, and combines them with a fault propagation knowledge graph obtained from the knowledge enhancement retrieval module to construct a dynamic causal graph of the current garage system. The semantic state fusion engine can also identify inconsistencies between multi-sensor data and utilize implicit reasoning from a large language model to generate robust estimates of equipment health status.

[0075] The chain-based reasoning and decision generator autonomously orchestrates the invocation sequence based on the objective of the event or instruction: first, it invokes one or more prediction tools to obtain quantitative assessment results; then, it invokes the knowledge-enhanced retrieval module to retrieve fault propagation paths and historical cases; finally, it inputs the aforementioned multi-source information (i.e., numerical prediction, semantic labels, causal paths, environmental context, etc.) into a large language model for multi-step chain-based reasoning, generating a preliminary decision plan that includes fault root cause identification, maintenance strategy selection, and priority setting. The maintenance strategy selection can be immediate repair, delayed repair, and / or preventative replacement. During the reasoning process, the generator explicitly considers at least three dimensions of quantitative indicators: equipment safety indicators, operational efficiency indicators, and user satisfaction indicators.

[0076] The reasoning process of the chain-based reasoning and decision generator employs an explicit thought chain prompting strategy, specifically including the following steps:

[0077] T1: Target parsing, which parses abnormal events or user commands into structured task targets G=(type, constraints, priority), where type belongs to the set {emergency repair, preventive maintenance, information query, scheduling optimization}, and constraints include time windows, security levels, resource limitations, etc. priority is the task urgency indicator, with a value range of predefined priority levels (high, medium, low), used to determine the response order and resource allocation weights among multiple pending tasks.

[0078] T2: Tool orchestration. Based on the task type, it retrieves matching tools from the tool registry and generates an ordered sequence of tool calls, with the sorting based on the data dependencies between tools.

[0079] T3: Evidence collection, sequentially executing tool calls, with the result of each call serving as context input for the next call, collecting quantitative evaluation results.

[0080] T4: Multi-objective trade-off. The collected evidence E is input into the large language model, which is required to explicitly calculate the comprehensive score for each candidate strategy s: score(s) = α × safety(s) + β × efficiency(s) + γ × satisfaction(s), where α, β, and γ are configurable weight coefficients, with preset values ​​of α = 0.5, β = 0.3, and γ = 0.2. safety(s) ∈ [0, 1] represents the probability of failure deterioration and the probability of remaining life assessment, efficiency(s) is based on the impact of expected maintenance time and vehicle retrieval efficiency, and satisfaction(s) is based on the peak-hour impact coefficient.

[0081] T5: Strategy generation. Select the strategy output with the highest comprehensive score as the initial decision plan, while retaining the top-3 candidate plans as alternatives.

[0082] The formal safety constraint verifier integrates a timing safety rule library extracted from garage PLC logic, safety operating procedures, and equipment technical manuals. The timing safety rule library stores multiple inviolable safety-related constraints using LTL linear timing logic formulas, including but not limited to: before executing the "lateral movement" command, it must be ensured that the anti-fall device of the target vehicle platform is in the "closed" state and the lateral movement limit switch is not triggered; before executing the "lifting" command, it must be ensured that there are no foreign objects under the vehicle platform and the predicted current value of the lifting motor does not exceed 120% of the rated value; when any maintenance operation makes a parking space unavailable, a guidance strategy must be automatically generated to prevent vehicles from being assigned to that parking space; it is prohibited to simultaneously execute lateral movement operations on adjacent parking spaces without confirmation of safety interlocks.

[0083] The formal safety constraint verifier is configured to receive a preliminary decision scheme generated by the chained reasoning and decision generator, convert the scheme into a structured action sequence, and then check whether the structured action sequence satisfies all safety-related constraints in the temporal safety rule base. If all are satisfied, a "pass" signal is output, and the decision scheme is passed to the planning and scheduling module; if any rule is violated, a "fail" signal is output, and violation feedback information is generated that clearly indicates which rule was violated and at which action step, and then returned to the chained reasoning and decision generator; the latter explicitly adds the violation feedback information to the prompt, re-performs reasoning to generate a corrected compliant decision scheme, until a compliant decision scheme is obtained or the preset number of retries is reached.

[0084] Unlike traditional model checking methods, this invention does not treat formal verification as a separate post-processing step, but rather embeds it within the decision-making loop of the large language model agent: the violation feedback information output by the formal safety constraint verifier, containing the reasons for violations and correction suggestions, is directly injected into the context of the large language model's prompt words in a structured form, triggering chain-like reasoning and the decision generator to perform targeted corrections instead of reasoning from scratch. This closed-loop mechanism of verification-feedback-correction differs from the traditional binary model of post-generation verification and rejection upon violation detection, and can automatically generate compliant correction schemes without relying on human intervention.

[0085] The timing safety rule base is sourced from three sources: the hardware interlock logic of the parking garage PLC controller, the mandatory safety clauses of the national standard for parking equipment (GB / T26559-2011) for automated parking systems, and the safety operating procedures in the technical manuals provided by equipment manufacturers. These three types of knowledge are uniformly converted into linear timing logic formulas for storage. The timing safety rule base contains no fewer than 27 core safety rules, covering all safety interlock conditions for lifting, lateral movement, and rotation. The timing safety rule base supports incremental updates: when a new equipment type is added or a new safety hazard is discovered, the administrator can describe the new safety rule using natural language. This rule is then automatically converted into the corresponding timing logic formula by the formal safety constraint verifier and stored in the rule base.

[0086] The specific method for implementing temporal safety rule verification using a formal safety constraint verifier is as follows:

[0087] 1) Construct an executable temporal security rule base. Convert security rules described in natural language into linear temporal logic formulas, following these mapping transformation rules:

[0088] a) The natural language pattern “Before executing A, B must be satisfied” is mapped to □(A→B), for example, “lateral movement command→(O fall arrestor status=closed)”;

[0089] b) The natural language pattern “prohibit execution of A under C” is mapped to □(A→(¬C)), for example, “adjacent parking space moving laterally → ¬current parking space moving laterally”;

[0090] c) The natural language pattern “Executing A must be followed by executing B” is mapped to □(A→O(B)), for example, “Power off → O insulation test”;

[0091] Where □ represents "always", O represents "next moment", and ¬ represents "logical NOT".

[0092] 2) The structured action sequence of the formal safety constraint verifier is represented using a six-tuple structure:

[0093] Action=(action id, action type, target device, preconditions, postconditions, timestamp);

[0094] Where: action id is the unique identifier of the action; action type is the enumeration value of the action type, including detection, execution, issuance, locking, etc.; target device is the identifier of the target device, such as parking space number or motor number; preconditions is the set of preconditions; postconditions is the set of postconditions, indicating the expected state change after the action is executed; timestamp is the planned execution time.

[0095] The formal safety constraint verifier employs a model verification algorithm to perform safety verification on decision-making schemes: 1) Safety rules from the time-series safety rule base and the system state transition model are input into the verification engine. The system state model consists of the current actual state plus all possible states that can be reached after the execution action sequence. 2) Within a preset step size (default setting 10 steps, configurable), the verifier exhaustively searches for execution paths that violate safety rules. 3) If a rule-violation path is found, the first violating action and the reason for the violation are immediately extracted from the path, and a structured violation feedback is output. Initially, the maximum number of reachable states is 30, the rule base size is 27, and the single verification time is less than 100 milliseconds, meeting real-time operation and maintenance requirements. When the rule base size increases, the verifier supports incremental verification: only newly added or modified rules are re-verified, and existing rule results are cached and reused.

[0096] 3) The formal safety constraint verifier performs the following temporal safety rule verification: parses the decision scheme output by the central agent layer, and extracts the action sequence A=[a1, a2, ..., a...]. n The current system state S0 is obtained from the perception and data fusion pipeline module. The state is represented as a set of key-value pairs, such as {(fall arrester #5, closed), (motor current #3, 12.5A), ...}. For i from 1 to n, the following steps are executed in a loop: the current state S0 is... i-1 With action a i Input the LTL model validator; iterate through each rule Ф in the time-series safety rule base. j Determine if in state S i-1 Execute a i Does it satisfy Ф? j Preconditions; if all relevant rules are met, then execute a. i According to a i Update the state with postconditions to obtain S i If rule Ф exists kIf the requirements are not met, the verification process will immediately terminate, and the violation information will be recorded. If all actions pass the verification, a "compliant" signal will be output, and the final state S will be changed. n Output all at once; if a violation is found, output a "failed" signal and violation feedback information.

[0097] The violation feedback information adopts a structured JSON format and includes at least the following fields: unique identifier of violation, corresponding security rule number, rule description in natural language, violation action steps, violation action details, current system status, expected conditions to be met, and correction suggestions.

[0098] The context memory and interaction management unit is used to maintain a short-term state during dialogue or task execution, record historical decisions, user feedback and execution results, support multi-turn interactions and complex task decomposition, and be used for subsequent decision self-optimization.

[0099] The planning and scheduling module is configured to generate an execution plan that includes resource allocation, time windows, personnel assignments, and spare parts lists based on task requirements. The execution and monitoring workflow module is configured to distribute the approved plan to the execution layer, continuously track the execution status, and report to the central intelligent agent layer when deviations are detected.

[0100] The architecture of the autonomous operation and maintenance system for multi-level parking garages based on a large language model in this invention is constructed through unified scheduling of the cognition and decision-making center of the central intelligent agent layer, combined with the knowledge retrieval support of the knowledge enhancement layer, the perception, prediction, planning and execution capabilities of the toolset and workflow layer, and the security verification closed loop of the embedded formal security constraint verifier. This builds a full-link autonomous operation and maintenance system from perception to execution, thereby achieving a technological leap from automatic early warning to safe and compliant autonomous operation and maintenance. It solves the technical problems of traditional multi-level parking garage operation and maintenance systems that rely on manual decision-making, cannot handle complex contexts, and have poor scalability.

[0101] Example 2

[0102] An autonomous operation and maintenance method for automated parking garages based on a large language model, such as... Figure 1 and Figure 2 As shown, an autonomous operation and maintenance system for an automated parking garage based on a large language model, applied in Example 1, includes the following steps:

[0103] Step S1: The perception and data fusion pipeline module of the toolset and workflow layer collects real-time operational data of the automated parking garage from the infrastructure layer (e.g., wire rope tension, motor current, vibration spectrum, etc.) and sends it to the central intelligent agent layer. The system cognition and decision-making center of the central intelligent agent layer performs data cleaning, spatiotemporal alignment, and multimodal feature extraction on the data. Then, it calls the semantic state fusion engine to extract semantic features from the quantified features through a pre-trained mapping model, converting key numerical features into semantic labels (e.g., "mid-term bearing wear" or "insufficient guide rail lubrication"). When a preset key indicator anomaly is detected, a warning signal is received from the prediction tool, or abnormal behavior is identified from the video stream (e.g., personnel illegally entering the movement area), a structured abnormal event is generated. This structured abnormal event typically contains the original value, semantic label, time and location information, and is pushed to the central intelligent agent layer along with contextual data. At the same time, the perception and data fusion pipeline module also sends the received user instructions to the central intelligent agent layer, and the system also receives user instructions input by the user through natural language.

[0104] Step S2: The central intelligent agent layer analyzes abnormal events or user commands to determine the task objective. Then, the semantic state fusion engine combines real-time data and knowledge graphs to construct a dynamic causal graph. The semantic state fusion engine identifies inconsistencies between multi-sensor data and uses implicit reasoning from a large language model to generate a robust estimate of the device's health status. Next, the chained reasoning and decision generator autonomously orchestrates the calling sequence: first, it calls one or more prediction and health management tools to obtain quantitative assessment results, and then it calls the knowledge-enhanced retrieval module to retrieve relevant background knowledge. Based on the reasoning capabilities of the large language model, the central intelligent agent layer performs chained fusion of the above multi-source information to form a preliminary decision scheme with fault root cause judgment and maintenance strategies.

[0105] During the reasoning process, the chain reasoning and decision generator first considers at least three quantitative indicators: equipment safety indicators related to the probability of failure deterioration and expected remaining lifespan, operational efficiency indicators related to expected maintenance time and the impact on vehicle access efficiency, and user satisfaction indicators related to the impact coefficient during peak hours. Through implicit weighting or explicitly defined utility functions of the large language model, a comprehensive preliminary decision scheme is generated.

[0106] Step S3: The formal security constraint verifier performs offline model verification on the preliminary decision scheme. The decision scheme is converted into a structured action sequence, and each action is checked to ensure it meets all relevant constraints in the temporal security rule base. The temporal security rule base uses LTL (Linear Temporal Logic) formulas to store security constraints, and the action sequence is represented using a six-tuple structure. The verification process involves parsing the decision scheme to extract the structured action sequence, obtaining the current system state from the perception and data fusion pipeline module, matching each action against relevant rules in the temporal security rule base, and outputting a compliance signal if all rules pass, otherwise outputting structured violation feedback. If all rules are met, the decision scheme is marked as "compliant" and proceeds to step S4. If any rule is violated, violation feedback information is generated and returned to the chain reasoning and decision generator in step S2. The generator explicitly adds violation constraints to the prompts based on the feedback information and regenerates a corrected decision scheme. This verification process is repeated until a compliant decision scheme is obtained or the preset maximum number of retries is reached. If a compliant scheme still cannot be obtained, the administrator is notified for intervention.

[0107] Step S4: The central intelligent agent layer sends the compliance decision plan to the planning and scheduling module. Based on the task type, time requirements, and resource needs in the compliance decision plan, the planning and scheduling module invokes path optimization algorithms, resource allocation algorithms, and conflict detection mechanisms to generate a specific execution plan. The execution plan includes at least: designated maintenance personnel, work time windows, a list of required spare parts, and a garage scheduling strategy.

[0108] Step S5: The central intelligent agent layer distributes the execution plan to the corresponding execution layers through the execution and monitoring workflow module. The execution layers include maintenance personnel terminals, garage PLC controllers, or parking guidance screens. The execution and monitoring workflow module tracks the task execution status in real time. If an execution deviation is detected, such as personnel exceeding the time limit without confirmation, spare parts being out of stock, or new sensor alarms occurring during execution, it automatically reports to the central intelligent agent layer, triggering the replanning process starting from step S2.

[0109] Step S6: After the task is completed, the central intelligent agent layer automatically summarizes the perception data, diagnostic basis, decision reasoning chain, security verification records, execution results, and other information throughout the entire event process, generates an analysis report in natural language format, and pushes it to the designated user. Simultaneously, the handling process of this event is stored as positive / negative samples in the vector database for subsequent optimization of knowledge retrieval or fine-tuning of the intelligent agent model. The security verifier records each violation feedback and correction process, and periodically analyzes high-frequency violation types to update or refine the security rule base.

[0110] The steps S1 to S6 of the autonomous operation and maintenance method for automated parking garages based on a large language model of the present invention, through unified scheduling and chain reasoning of the central intelligent agent layer, combined with the security verification closed loop of the formal security constraint verifier, and the task execution closed loop of the planning and scheduling module and the execution monitoring module, together constitute a complete autonomous operation and maintenance process from event perception to task completion and feedback learning. This solves the technical problems of traditional automated parking garage operation and maintenance systems that rely on manual decision-making, cannot handle complex contexts, and have poor scalability.

[0111] Based on the above descriptions of the autonomous operation and maintenance system and method for automated parking garages based on large language models in Embodiments 1 and 2, the practical application of the present invention will be described in detail below through actual examples:

[0112] Example 3: Application Case of Intelligent Transformation of Static Traffic Automated Parking Garages in Urban Commercial Complexes

[0113] This embodiment uses the autonomous operation and maintenance system for automated parking garages based on large language models from Embodiment 1 and the autonomous operation and maintenance method for automated parking garages based on large language models from Embodiment 2 as an example for illustration. The garage is a lift-and-slide structure with 236 parking spaces and has been in service for 6 years. Before the renovation, the garage experienced an average of 17 malfunctions per year, with an average annual operation and maintenance cost of approximately 800,000 yuan, of which manual inspection and troubleshooting accounted for more than 60%.

[0114] The solution is to deploy the system of this invention on a cloud server, such as... Figure 1 and Figure 2 As shown, the autonomous operation and maintenance system for automated parking garages based on a large language model is divided into four layers from top to bottom: user layer, central intelligent agent layer, knowledge enhancement layer, toolset and workflow layer, and infrastructure layer.

[0115] The user layer includes an operations and maintenance management module and a third-party management system module. The operations and maintenance management module communicates with the central intelligent agent through natural language interaction. The third-party management system module includes meteorological monitoring data, urban parking management system, intelligent building system, etc., and exchanges data and controls with the central intelligent agent through API gateway.

[0116] The central intelligent agent layer comprises a system cognition and decision-making hub, a semantic state fusion engine, a chained reasoning and decision generator, a formal safety constraint verifier, and a context memory and interaction management unit. The system cognition and decision-making hub is a pre-trained large language model configured as an intelligent agent with tool-calling capabilities. It invokes and coordinates the semantic state fusion engine, chained reasoning and decision generator, formal safety constraint verifier, and context memory and interaction management unit to complete tasks. Specifically, the central intelligent agent layer uses a large language model fine-tuned based on the Qwen-2.5-72B-Instruct model, with a context memory length of 8k tokens and a tool-call timeout of 30 seconds.

[0117] The knowledge enhancement layer directly connects to the system's cognitive and decision-making center, including a knowledge enhancement retrieval module and a vector database. The knowledge enhancement retrieval module connects to a pre-built vector database, which stores equipment technical manuals, historical maintenance records, a fault case library, safe operating procedures, and a fault propagation knowledge graph. Specifically, the knowledge enhancement layer imports the garage's historical maintenance records from the past 6 years (342 records in total), equipment manuals (PDF format, approximately 120 pages), and fault cases of the automated parking system (2300 records after anonymization). Using OpenAI's text-embedding-3-small model, text blocks are converted into 768-dimensional vectors and stored in the Milvus vector database. Simultaneously, Neo4j is used to construct a fault propagation knowledge graph containing 47 fault modes and 132 causal edges.

[0118] The toolset and workflow layer includes a perception and data fusion pipeline module, a planning and scheduling module, an execution and monitoring workflow module, and a prediction and health management toolset. The toolset and workflow layer retains the original SVM fault prediction model, which performs binary classification of motor overload and wire rope wear based on current and vibration characteristics. It also adds a Transformer-based vibration trend prediction model, which takes the vibration spectrum of the past 168 hours as input and outputs the probability of failure in the next 24 hours. The prediction and health management toolset includes a support vector machine-based fault prediction tool, a deep learning-based component remaining life prediction tool, and an autoencoder-based unsupervised anomaly detection tool. The planning and scheduling module interfaces with a WeChat mini-program for maintenance personnel containing GPS location and work schedules, a parking garage PLC control interface (Siemens S7-1200, via OPC UA protocol), and a parking guidance screen system. The execution and monitoring module supports the issuance of OPC UA protocol PLC control commands.

[0119] The infrastructure layer includes the parking space terminal layer and the execution layer.

[0120] The perception and data fusion pipeline module is configured to receive and standardize multi-source heterogeneous operational data from the parking space terminal layer in real time. This multi-source heterogeneous operational data includes, but is not limited to: wire rope tension, motor current, vibration spectrum, photoelectric limit switch status, fall arrestor status, vehicle overlength / overweight detection signals, panoramic video streams, and environmental temperature and humidity data. This module also includes a semantic feature extraction unit, used to convert numerical sensor data (such as the frequency band energy distribution of vibration spectra and the distortion rate of current waveforms) into semantic labels describing the health status of equipment through a pre-trained mapping model, such as "mid-stage bearing wear," "insufficient guide rail lubrication," and "excessive transmission key clearance," thereby eliminating the semantic gap between raw data and high-level inference. In practical implementation, the sensing and data fusion pipeline module, based on the PLC signal, is equipped with 8 vibration transmitters (model CTC AC192, sensitivity 100 mV / g, sampling frequency 2kHz), 12 current transmitters (model Mik-DZI, 4-20mA output), and 2 1080P panoramic cameras (Hikvision brand, RTSP stream). Data acquisition frequency: every 200ms for digital signals, every 10 seconds for analog signals, and every 2 seconds for video keyframes. Deployed on an edge gateway (NVIDIA Jetson AGX Orin), the sensing fusion pipeline runs, and the standardized data is uploaded to the cloud via MQTT.

[0121] The semantic feature extraction unit pre-trains a random forest regression model. The input feature vector is 10-dimensional: energy in three frequency bands of the vibration spectrum (1-3kHz, 3-5Hz, 5-8Hz) (3-dimensional), current waveform distortion rate (1-dimensional), temperature change rate (1-dimensional), motor torque fluctuation coefficient (1-dimensional), and the first-order difference of the above features (4-dimensional). The output consists of three semantic labels: bearing_wear_probability (bearing wear probability, 0-1), lubrication_score (lubrication status score, 0-100), and transmission_gap_level (transmission gap level, 0-3). The model training data comes from the garage's historical maintenance records over the past 6 years (342 records in total) and corresponding sensor data, using 5-fold cross-validation. The average F1-score on the test set is 0.87. For example, when the 1-2kHz energy exceeds the threshold and THD>5%, the output semantic label is "mid-stage bearing wear, confidence level 0.87".

[0122] The semantic state fusion engine receives semantic labels and raw numerical features from the perception and data fusion pipeline module, and combines them with a fault propagation knowledge graph obtained from the knowledge enhancement retrieval module to construct a dynamic causal graph of the current garage system. This engine can identify inconsistencies between multi-sensor data and utilize implicit reasoning from a large language model to generate robust estimates of equipment health status.

[0123] The chain-based reasoning and decision generator autonomously orchestrates the invocation sequence based on the objective of the event or instruction: first, it invokes one or more predictive tools to obtain quantitative assessment results; then, it invokes the knowledge-enhanced retrieval module to retrieve fault propagation paths and historical cases; finally, it inputs the aforementioned multi-source information into a large language model for multi-step reasoning, generating a preliminary decision plan that includes fault root cause identification, maintenance strategy selection, and priority setting. During the reasoning process, the generator explicitly considers at least three dimensions of quantitative indicators: equipment safety indicators, operational efficiency indicators, and user satisfaction indicators.

[0124] The formal safety constraint verifier integrates a timing safety rule base extracted from garage PLC logic, safety operating procedures, and equipment technical manuals. This timing safety rule base stores multiple inviolable safety constraints using LTL (Linear Time Logic) formulas, including but not limited to:

[0125] Rule 1: Before executing the "lateral movement" command, it must be ensured that the fall arrestor of the target vehicle platform is in the "closed" state and the lateral movement limit switch is not triggered;

[0126] Rule 2: Before executing the "lift" command, ensure that there are no foreign objects under the vehicle platform and that the predicted current of the lifting motor does not exceed 120% of the rated value;

[0127] Rule 3: Whenever a maintenance operation renders a parking space unavailable, a guidance strategy must be automatically generated to prevent vehicles from being assigned to that parking space;

[0128] Rule 4: It is prohibited to simultaneously perform lateral movement between adjacent parking spaces without confirming the safety interlock.

[0129] The formal safety constraint verifier is configured to: receive a preliminary decision scheme generated by the chained reasoning and decision generator, convert the scheme into a structured action sequence, and then check whether the action sequence satisfies all relevant constraints in the temporal safety rule base. If all constraints are satisfied, a "pass" signal is output, and the decision scheme is passed to the planning and scheduling module; if any rule is violated, a "fail" signal is output, and violation feedback information is generated and returned to the chained reasoning and decision generator; the latter explicitly adds the violation feedback information to the prompt and re-performs the reasoning to generate a corrected compliant decision scheme.

[0130] The context memory and interaction management unit is used to maintain a short-term state during dialogue or task execution, record historical decisions, user feedback and execution results, support multi-turn interactions and complex task decomposition, and be used for subsequent decision self-optimization.

[0131] The planning and scheduling module is configured to generate a detailed execution plan based on task requirements, including resource allocation, time windows, personnel assignments, and spare parts lists. The execution and monitoring workflow module is configured to distribute the approved plan to the execution layer, continuously track the execution status, and report to the central intelligent agent layer when deviations are detected.

[0132] This invention employs a cloud-edge collaborative deployment architecture: the perception and data fusion pipeline module is deployed on the edge gateway (NVIDIA Jetson AGXOrin), responsible for real-time data acquisition and semantic tag generation, with a single processing latency of less than 50ms; the central intelligent agent layer, knowledge enhancement layer, and planning and scheduling module are deployed on the cloud server, communicating with the edge gateway via the MQTT protocol. When cloud communication is interrupted, the edge gateway initiates a degradation strategy: 1) continuing to execute the currently issued work order; 2) caching newly generated perception data; 3) for emergency safety events (such as abnormal triggering of the fall arrestor), the edge gateway directly sends an emergency stop command to the PLC, without relying on cloud decision-making. End-to-end latency budget: perception acquisition → semantic tag generation less than 50 milliseconds, anomaly detection → cloud push less than 100 milliseconds, LLM inference and verification less than 2 seconds, total latency less than 3 seconds, meeting the real-time requirements of automated parking garage operation and maintenance.

[0133] The following combination Figure 2 The method flow shown is illustrated using an actual fault handling process as an example:

[0134] S1 Event Awareness: At 10:23 on March 15, 2025, the sensing pipeline pushed a structured early warning event: the No. 5 carrier plate motor vibrates at 1-2kHz with an energy of 0.32 and a current distortion rate of 7.1%, with the semantic tag being "mid-term bearing wear" and an SVM failure probability of 0.87; at the same time, meteorological information was obtained through the external interface: continuous heavy rain in the next 3 days, with relative humidity >90%.

[0135] S2 Autonomous Diagnosis and Decision-Making: The agent's autonomous orchestration tool is invoked: First, the SVM details tool is called to confirm the fault as "bearing wear at the motor drive end." Then, the lifespan prediction tool is called to output a remaining lifespan of 96 hours. Finally, the knowledge base is searched for historical cases: In 2024, a similar fault resulted in a 72-hour delay in repair, leading to bearing seizure, a 240% increase in costs, and a 4-hour downtime for the entire system. The preliminary decision is to perform preventative replacement from 22:00 on March 15, 2025 to 02:00 the following day.

[0136] S3 Formal Security Verification and Compliance: The initial security verification failed due to two rule violations: ① Failure to check the fall arrestor status before locking the parking space; ② Failure to include a power disconnection step in motor maintenance. After the verifier returned correction suggestions, the agent added the pre-step "check fall arrestor closure → disconnect motor power" to the action sequence, and the second verification passed.

[0137] S4 Task Planning and Resource Coordination: The planning module automatically matches the manual maintenance team, schedules SKF-6205 bearing spare parts, and generates an execution plan: During maintenance, parking space No. 5 is locked, and the guidance screen prompts the user to use parking spaces No. 6 and No. 7.

[0138] S5 Execution and Closed-Loop Monitoring: After the system pushes a notification to the administrator for confirmation, it automatically executes the power-off command at 22:00. The replacement is completed after manual scanning of the code for confirmation. The actual time taken is 2.5 hours, and there is no impact on the user's vehicle retrieval.

[0139] S6 Reporting and Learning: Upon completion of a task, a report is automatically generated, recording information such as "root cause of failure is pitting of bearing rolling elements, 1 violation of safety verification interception, 1 SKF-6205 bearing consumed as a spare part," and stored in the vector database as a reference for subsequent decision-making.

[0140] This embodiment fully demonstrates the entire process from sensor data acquisition to the closed loop of maintenance tasks, verifying the feasibility and effectiveness of the system and method of the present invention in actual operation and maintenance scenarios of automated parking garages.

[0141] Example 4: System Message Interaction Sequence and JSON Data Format

[0142] This embodiment uses the application scenario of embodiment 3 to further illustrate the message interaction sequence between the modules of the system, and provides a complete JSON data format example for each stage.

[0143] like Figure 3 As shown, Figure 3 The sequence diagram illustrates the message interaction order between the user, central intelligent agent module, perception and data fusion pipeline, prediction and health management toolset, knowledge enhancement retrieval module, planning and scheduling module, and execution and monitoring workflow module. The timing diagram clearly distinguishes between synchronous calls and asynchronous notifications; synchronous calls are represented by solid arrows, while asynchronous notifications are represented by dashed arrows.

[0144] The specific process includes: the sensing pipeline detects an 87% probability of bearing failure on plate 5 and pushes an early warning event, while simultaneously obtaining meteorological information on heavy rain for the next three days from the API; the central agent sequentially calls the SVM detail query tool and lifespan prediction tool in the prediction and health management tool module to obtain quantitative assessment results; then it calls the knowledge enhancement retrieval module to query relevant background knowledge; the central agent generates a decision plan after chain reasoning; it calls the planning module to generate a detailed plan; it pushes a notification to the user and waits for confirmation; after confirmation, it issues a work order to the execution module; it receives execution status feedback; and finally, it pushes a completion report to the user.

[0145] The following are complete JSON data format examples for each stage:

[0146] (a) Event perception stage

[0147] The system triggers event detection by detecting abnormal data in the sensing pipeline: At 10:23 on a certain day, the sensing pipeline pushes a structured abnormal event to the central intelligent agent.

[0148] {

[0149] "event id": "EVT-2305-01",

[0150] "type": "predictive alert",

[0151] "device": "No. 5 vehicle carrier motor"

[0152] "raw features": {"vibration 1 2kHz": 0.32, "current thd": 7.1%, "temp rate": 0.8°C / min},

[0153] "semantic labels": ["Mid-stage bearing wear", "Normal lubrication condition"]

[0154] "svm output": { "fault type": "bearing wear", "probability": 0.87},

[0155] "timestamp": "2025-03-15T10:23:00Z"

[0156] }

[0157] Meanwhile, the central intelligent agent obtained information from the meteorological service through the API gateway: "Heavy rain will continue for the next three days, with relative humidity >90%".

[0158] The central agent invoked the `get SVM details` tool, confirming the fault type as "bearing wear on the motor drive end," with a historical accuracy of 86%. It then invoked a deep learning model tool to predict remaining life, inputting a vibration sequence from the past 24 hours, which output a remaining effective lifespan of 96 hours (90% confidence interval: 72-110 hours), with the note that "for every 10% increase in humidity, lifespan degradation accelerates by 15%." The knowledge retrieval module was invoked, querying the vector "bearing fault, damp repair window safety rules," returning historical cases (similar faults where a 72-hour delay in repair led to bearing seizure, increased repair costs by 240%, and caused a 4-hour shutdown of the entire garage) and safety rules (Garage Safety Operation Procedure 4.2: "Before repairing motor bearings in a damp environment, the main power supply must be disconnected and a 'Do Not Operate' sign must be displayed; insulation testing must be performed after repair"). Garage usage statistics showed that during heavy rain (6:00 PM to 8:00 AM the next day), vehicle retrieval frequency decreased by 31% compared to normal days.

[0159] The chained reasoning and decision generator outputs preliminary decision solutions in JSON format:

[0160] {

[0161] "decision id": "D-2305-01",

[0162] "Root cause": "Pitting of rolling elements in motor bearings (vibration spectrum characteristic matching degree 92%)".

[0163] "strategy": "preventative replacement"

[0164] "priority": "high"

[0165] "scheduled window": {"start": "2025-03-15T22:00:00", "end": "2025-03-16T02:00:00"},

[0166] "action sequence": [

[0167] { "action": "lock parking space", "params": { "space id": 5, "reason": "maintenance"}},

[0168] { "action": "send work order", "params": { "worker": "manual group", "task": "replace SKF-6205 bearing"}}

[0169] { "action": "update guidance", "params": { "message": "Parking space #5 is under maintenance, please use spaces #6 or #7"}} ]

[0171] }

[0172] (ii) Formal security verification phase

[0173] The action sequence is input into the formal safety constraint validator. The validator loads the temporal safety rule base (27 rules in total, stored as LTL formulas). A check of each rule reveals the following violations: Rule R09 is violated—"Before locking the parking space, the check safety device action must be performed and the result must pass"—the current sequence lacks confirmation of the fall arrestor status and motor power disconnection; Rule R12 is violated—"When maintenance involves the motor, the action sequence must include a disconnect power step, and the steps must precede locking the parking space." The validator generates violation feedback information in JSON format.

[0174] {

[0175] "violation id": "VF-001",

[0176] "rule ids": ["R09", "R12"],

[0177] "description": "The parking space was locked without checking the status of the fall arrestor and the motor being de-energized, violating safety operating procedures."

[0178] "Suggested fix": "Please add the following before 'lock parking space': ①check fallarrestor(space id) ②disconnect motor power(space id)"

[0179] }

[0180] (III) Decision-making revision stage

[0181] Feedback information is injected into the prompts of the chained reasoning and decision generator (as a system message). The central agent regenerates the revised action sequence in JSON format:

[0182] {

[0183] "action sequence": [

[0184] { "action": "check fall arrestor", "params": { "space id": 5, "expected": "closed"}},

[0185] { "action": "disconnect motor power", "params": { "space id": 5}},

[0186] { "action": "lock parking space", "params": { "space id": 5, "reason": "maintenance"}},

[0187] { "action": "send work order", "params": { "worker": "manual group", "task": "replace SKF-6205 bearing"}}

[0188] { "action": "update guidance", "params": { "message": "Parking space #5 is under maintenance, please use spaces #6 or #7"}} ]

[0190] }

[0191] Submit to the security verifier again; all rules are approved. The verifier outputs "Compliant" and records the violation and the correction process.

[0192] (iv) User confirmation and execution phase

[0193] The central intelligent agent pushes a natural language notification to the administrator, explaining the fault risk, the established compliant maintenance plan, the coordinated maintenance team and spare parts information, and prompting the administrator to confirm the safety verification of the fall arrestor and power-off procedures. The administrator replies with "Confirm." The system automatically sends the work order to the maintenance personnel's terminal and simultaneously sends a pre-locking command to the PLC. At the scheduled time, the system automatically executes the power-off command, and the maintenance personnel scan the code to confirm before replacing the bearing. After the maintenance is completed, the system automatically restores power and unlocks the parking space.

[0194] (V) Reporting and Learning Phase

[0195] After maintenance is completed, the central intelligent agent system's cognition and decision-making center generates a report containing information such as event identifier, root cause of failure, decision basis, safety verification record, actual time consumption, and spare parts consumption, which is stored in the vector database as a reference for subsequent decision-making.

[0196] This embodiment clearly demonstrates the data flow and message interaction process at each stage, from event perception, autonomous diagnosis, security verification, corrective decision-making to execution confirmation, through a complete JSON data format example. It verifies the consistency of data formats and the integrity of the interaction process among the modules of the system described in this invention.

[0197] Example 5: Root Cause Analysis Driven by Natural Language

[0198] This embodiment uses the system described in Embodiment 1 and the method described in Embodiment 2 to demonstrate the system's ability to process natural language commands.

[0199] One week, the administrator asked the system via WeChat mini-program: "In the past week, there have been more complaints about strange noises at night in the B area garage. Please help me find out the reason."

[0200] The central agent first parses the user command through the semantic state fusion engine, determining the task objective as "analyzing the root cause of nighttime abnormal noises in the B parking garage." Then, the chain-based reasoning and decision generator autonomously arranges the calling sequence: it calls the query timeseries tool to retrieve vibration sensor data from InfluxDB for all parking spaces in B over the past 7 days (one sample per parking space every 10 seconds, totaling approximately 18,000 data points). It then calls the spectrum analyzer tool to perform an FFT on the vibration data, finding that the abnormal noise peaks are concentrated in the 500-800Hz range (the dry friction characteristic frequency band), with a cosine similarity of 0.92 to the historical dry friction data of the guide rails. Next, it calls the knowledge-enhancing retrieval module to search the knowledge base: the dry friction pattern intensifies at temperatures below 5℃, the average nighttime temperature last week was 2℃, and the guide rails in parking spaces 12, 15, and 18 were last lubricated 90 days ago (standard cycle 60 days). Based on the reasoning capabilities of the large language model, the central agent performs chain-based fusion of the above multi-source information to generate a preliminary decision scheme including fault root cause identification and maintenance strategy selection. The final report recommends applying low-temperature grease to these three parking spaces. After administrator confirmation, the system automatically generates a work order.

[0201] This embodiment verifies that the system described in this invention can understand natural language instructions and perform complex root cause analysis, demonstrating the advantages of large language model agents in semantic understanding and context awareness.

[0202] Example 6: Simultaneous Early Warning from Multiple Devices and Global Priority Decision Making

[0203] This embodiment uses the system described in Embodiment 1 and the method described in Embodiment 2 to demonstrate the system's ability to perform multi-objective trade-offs and global optimization in complex scenarios.

[0204] At 08:30 one day (weekday morning rush hour), the system received three warnings simultaneously:

[0205] A: The lifting motor of parking space No. 3 has an abnormal current, and the SVM outputs "overload fault probability 91%";

[0206] B: The tension of the wire rope at parking space No. 8 has decreased by 12%, and the knowledge search indicates "wire rope wear is possible";

[0207] C: The photoelectric switch for parking space No. 15 keeps flickering, possibly due to dust obstruction.

[0208] The chain-based reasoning and decision generator of the central intelligent agent layer performs multi-objective trade-offs: calculating a comprehensive score for each fault, explicitly considering three quantitative indicators: equipment safety indicators, operational efficiency indicators, and user satisfaction indicators. After sorting, an optimal processing order is generated: B (B is due to wire rope wear potentially affecting safety, so it is processed first) → A (A is due to overload potentially affecting peak-hour vehicle retrieval) → C (C is a low-priority order, which can be processed later). The formal safety constraint verifier confirms that the three maintenance plans have no resource conflicts and then schedules them for execution at different time periods.

[0209] This embodiment verifies that the system described in this invention can automatically weigh and generate the globally optimal processing order when multiple devices issue warnings simultaneously.

[0210] Example 7: Handling Complex Spatiotemporal Constraints in User Natural Language Commands

[0211] This embodiment uses the system described in Embodiment 1 and the method described in Embodiment 2 to demonstrate the system's ability to process natural language instructions with complex spatiotemporal constraints.

[0212] The user typed: "Please arrange for the steel cable to be replaced next week, so as not to affect picking up the car during the day. Ideally, it would be on Monday or Wednesday evening. If it rains, it will be postponed."

[0213] The semantic state fusion engine of the central intelligent agent layer resolves the constraints: the time window is Monday / Wednesday evening; weather conditions (postponed if it rains); the goal is "steel cable replacement" (corresponding to the steel cable in parking space 8). The chain reasoning and decision generator calls the weather forecast tool to obtain the weather forecast for next Monday and Wednesday: sunny on Monday, light rain on Wednesday. Monday is automatically selected. The planning module is called to check if there are any other maintenance tasks conflicting on Monday evening (no), and a plan is generated. The formal safety constraint verifier is called: confirming that the action sequence for replacing the steel cable includes safety steps such as "fall arrestor locking", "motor power off", and "post-maintenance tension test", and passes. Reply: "The steel cable replacement for parking space 8 has been scheduled for 21:00-24:00 on March 20th (Monday). The weather is sunny and suitable for the operation. The safety steps have been verified. Please reply 'Confirm' to confirm."

[0214] This embodiment verifies that the system described in this invention can process natural language commands containing complex spatiotemporal constraints, obtain real-time information by calling external tools and perform multi-objective trade-offs, and generate decision schemes that both conform to user intent and meet security constraints.

[0215] Example 8: Deviations during execution triggering replanning

[0216] This embodiment uses the maintenance scenario of Embodiment 3 to demonstrate the closed-loop capability of the system to automatically trigger replanning after detecting deviations during execution, reflecting the complete implementation of step S5 in the method described in Embodiment 2.

[0217] After the maintenance work order was issued, a manual scan revealed that the SKF-6205 bearing was out of stock in the spare parts inventory (the actual inventory was updated late in the system). The manual operator reported the "spare parts out of stock" through a mini-program.

[0218] The execution and monitoring workflow module detected a deviation (the work order status changed to "Spare Parts Missing")

[0219] The system immediately reports to the central intelligent agent layer. The central intelligent agent layer triggers a replanning process: it searches the spare parts inventory, queries nearby warehouses, and finds the nearest warehouse with stock 25 kilometers away, with an estimated delivery time of 2 hours. A new decision is made: the repair window is postponed to 00:00-04:00 the following day, and the user is notified. Simultaneously, the planning module is invoked to adjust the guidance strategy. A message is pushed to the administrator: "Due to a shortage of SKF-6205 bearings, repair is postponed to March 16th, 00:00-04:00. Confirm?" After the administrator confirms, the work order is updated.

[0220] This embodiment fully demonstrates the closed-loop process from plan issuance to execution monitoring and deviation replanning, verifying that even if unexpected situations occur during execution, the invention can automatically trigger the replanning process and generate a correction plan without manual intervention, reflecting the improved robustness and automation level of the invention.

[0221] The embodiments of the present invention are not limited thereto. Based on the above description of the present invention, and using common technical knowledge and conventional means in the field, the present invention can be modified, replaced or combined in various other forms without departing from the basic technical idea of ​​the present invention, and all such modifications, replacements or combinations fall within the scope of protection of the present invention.

Claims

1. An autonomous operation and maintenance system for an automated parking garage based on a large language model, characterized in that, include: The central intelligent agent layer includes a system cognition and decision-making hub, a chain reasoning and decision generator, and a formal safety constraint verifier. The system cognition and decision-making hub is a large language model with tool invocation capabilities. The toolset and workflow layer includes a perception and data fusion pipeline module, a prediction and health management toolset, a planning and scheduling module, and an execution and monitoring workflow module that are communicatively connected to the central intelligent agent layer. The knowledge enhancement layer includes a knowledge enhancement retrieval module and a vector database that are communicatively connected to the central intelligent agent layer. The infrastructure layer includes the parking space terminal layer and the execution layer; The chain reasoning and decision generator responds to abnormal events or user commands by calling the prediction and health management toolset for quantitative assessment, calling the knowledge enhancement retrieval module to retrieve relevant background knowledge, and inputting the assessment results and background knowledge into the system's cognition and decision-making center for chain reasoning to generate a preliminary decision plan. The formal safety constraint verifier integrates a time-series safety rule base. It receives preliminary decision schemes and converts them into structured action sequences, then checks each structured action sequence to see if it satisfies the relevant constraints in the time-series safety rule base. If all constraints are met, it outputs a compliance signal and transmits the decision scheme to the planning and scheduling module. If any clause is violated, a violation feedback message with the reason for the violation and a correction suggestion will be generated and returned to the chain reasoning and decision generator, triggering it to re-reason and generate a corrected decision solution until a compliant decision solution is obtained or the preset number of retries is reached. The planning and scheduling module is used to generate an execution plan based on the compliance decision-making scheme; The execution and monitoring workflow module is used to issue and track execution plans, and trigger replanning when deviations are detected.

2. The autonomous operation and maintenance system for automated parking garages based on a large language model as described in claim 1, characterized in that: The perception and data fusion pipeline module is used to access and standardize multi-source heterogeneous operating data from the parking space terminal layer to form raw numerical features. The multi-source heterogeneous operating data accessed by the perception and data fusion pipeline module includes at least two of the following: wire rope tension, motor current, vibration spectrum, photoelectric limit switch status, fall arrestor status, vehicle overlength / overweight detection signals, panoramic video stream, and environmental temperature and humidity data. The perception and data fusion pipeline module is also equipped with a semantic feature extraction unit, which is used to convert numerical sensor data into semantic labels describing the health status of equipment through a pre-trained mapping model.

3. The autonomous operation and maintenance system for automated parking garages based on a large language model as described in claim 2, characterized in that, The central intelligent agent layer also includes a semantic state fusion engine, which receives semantic labels and raw numerical features from the perception and data fusion pipeline module, combines them with the fault propagation knowledge graph obtained from the knowledge enhancement retrieval module, and constructs a dynamic causal graph of the current garage system. The semantic state fusion engine can identify inconsistencies between multi-sensor data and use the implicit reasoning of the large language model of the system's cognition and decision-making center to generate robust estimates of the equipment health status.

4. The autonomous operation and maintenance system for automated parking garages based on a large language model as described in claim 1, characterized in that, The temporal security rule base uses linear temporal logic formulas to store security constraints, and the transformation rules include: a) "B must be satisfied before A can be executed" is mapped to □(A→(B)); b) "Prohibit the execution of A under condition C" is mapped to □(A→(¬C)); c) "Executing A must be followed by executing B" is mapped to □(A→O (B)); Where □ represents "always", O represents "next moment", and ¬ represents "logical NOT".

5. The autonomous operation and maintenance system for automated parking garages based on a large language model as described in claim 1 or 4, characterized in that, The structured action sequence of the formal security constraint verifier is represented using a six-tuple structure: Action=(action id, action type, target device, preconditions, postconditions, timestamp); Wherein, action id is the unique identifier of the action, action type is the enumerated value of the action type, target device is the identifier of the target device, preconditions is the set of preconditions, postconditions is the set of postconditions, and timestamp is the planned execution time.

6. The autonomous operation and maintenance system for automated parking garages based on a large language model as described in claim 5, characterized in that, The formal security constraint verifier performs the following step-by-step verification process: it parses the preliminary decision scheme to extract the corresponding structured action sequence, obtains the current system state from the perception and data fusion pipeline module, matches the relevant rules in the temporal security rule base for each action, and outputs a compliance signal for the structured action sequence when all rules pass; otherwise, it outputs feedback that the structured action sequence violates the rules.

7. The autonomous operation and maintenance system for automated parking garages based on a large language model as described in claim 1, characterized in that, The prediction and health management toolset includes at least two of the following tools: a fault prediction tool based on support vector machines, a component remaining life prediction tool based on deep learning, and an unsupervised anomaly detection tool based on autoencoders.

8. The autonomous operation and maintenance system for automated parking garages based on a large language model as described in claim 1, characterized in that, The central intelligent agent layer also includes a context memory and interaction management unit, which is used to maintain a short-term state during dialogue or task execution, record historical decisions, user feedback and execution results, support multi-turn interaction and complex task decomposition, and be used for subsequent decision self-optimization.

9. A method for autonomous operation and maintenance of an automated parking garage based on a large language model, applied to the automated operation and maintenance system for an automated parking garage based on a large language model as described in any one of claims 1-8, characterized in that, Includes the following steps: S1: The perception and data fusion pipeline module collects operational data, generates abnormal events or receives user instructions, and pushes them to the central intelligent agent layer. S2: The central intelligent agent layer parses events or instructions, calls the prediction and health management toolset and knowledge enhancement retrieval module to perform chain reasoning, and generates a preliminary decision plan with fault root cause judgment and maintenance strategy. S3: Convert the preliminary decision plan into a structured sequence of actions, check whether each action meets all relevant constraints in the time-series safety rule base, and if a violation is found, generate a violation feedback and return to S2 to regenerate; S4: The central intelligent agent layer sends the compliance decision-making scheme to the planning and scheduling module, generating an execution plan with personnel, time, spare parts and scheduling strategies; S5: The execution and monitoring workflow module issues execution plans and tracks their status. If deviations occur, replanning is triggered. S6: After the task is completed, the central intelligent agent layer generates an analysis report and stores the processing process in the vector database.

10. The autonomous operation and maintenance method for automated parking garages based on a large language model as described in claim 9, characterized in that, In step S3, the violation feedback information is in structured JSON format. The structured JSON format includes a unique violation identifier, the corresponding security rule number, a natural language description of the rule, the violation action steps, violation action details, the current system status, the expected conditions to be met, and correction suggestions.

Images