Elevator intelligent operation and maintenance and safety evaluation method based on large language model

Abstract

The invention discloses an elevator intelligent operation and maintenance and safety evaluation method based on a large language model, which comprises the following steps: establishing a multi-dimensional elevator knowledge base integrating a structured knowledge graph and unstructured historical cases, configuring the large language model as a core reasoning engine, and regularly collecting original operation state data of an elevator; after protocol analysis and feature extraction, converting into an equipment state parameter set, dynamically generating structured natural language state description, combining the state description with related background knowledge fragments retrieved from a knowledge base in real time, constructing a comprehensive query context, inputting the comprehensive query context into a large language model, and performing deep reasoning on the model according to the comprehensive query context; and generating a preliminary evaluation report, providing a human-computer interaction interface for operation and maintenance personnel to review and correct the report, and calibrating and generating a final decision report according to a correction instruction. Intelligent fusion and reasoning of multi-source heterogeneous data are achieved, domain knowledge and human experience are combined, and the perspectiveness, accuracy and operation and maintenance decision-making efficiency of elevator safety evaluation are improved.

Description

Technical Field

[0001] This application relates to the field of intelligent elevator control technology, and in particular to an intelligent elevator operation and maintenance and safety assessment method based on a large language model. Background Technology

[0002] Traditional elevator operation and maintenance and safety assessments primarily rely on the manual experience of on-site technicians and periodic planned maintenance. This method suffers from significant lag, typically allowing intervention only after a fault occurs or a component completely fails, making it difficult to achieve early warning and dynamic risk assessment. Furthermore, manual recording and analysis of faults are inefficient, and valuable maintenance experience is often scattered in unstructured text format, making it difficult to systematically accumulate, share, and reuse, resulting in difficulties in knowledge transfer and inconsistent operation and maintenance quality.

[0003] In recent years, with the popularization of IoT technology, elevator remote monitoring systems have been applied, enabling real-time collection of elevator operation data. However, most existing systems are limited to data visualization and threshold alarms, lacking the ability to deeply integrate and intelligently analyze multi-source heterogeneous data. Their analysis logic is often based on simple rule judgments, unable to understand the correlation between complex fault phenomena, let alone combine domain knowledge for deep reasoning, and unable to provide accurate fault root cause location and forward-looking maintenance decision support. The level of intelligence needs to be improved.

[0004] While artificial intelligence technologies, such as big data analytics and machine learning, offer new insights into predictive maintenance of equipment, their direct application in the elevator field still faces challenges. On one hand, these methods typically require massive amounts of high-quality labeled data for model training, while elevator fault samples are relatively scarce, resulting in insufficient model generalization ability. On the other hand, existing methods struggle to effectively integrate and jointly reason with structured real-time data and unstructured domain knowledge and historical cases. Their decision-making processes lack interpretability and cannot effectively incorporate the experience and judgment of human experts, making it difficult to build a trustworthy and reliable human-machine collaborative intelligent operation and maintenance system. Summary of the Invention

[0005] To achieve the above objectives, this application provides the following technical solution: According to a first aspect of the present invention, the present invention claims protection for an elevator intelligent operation and maintenance and safety assessment method based on a large language model, comprising: S1: Establish and maintain a multi-dimensional elevator knowledge base, which integrates a structured elevator domain knowledge graph and unstructured elevator fault history case texts, and configures a large language model pre-trained in a general domain as the core reasoning engine. S2: Periodically collect raw operating status data streams from the target elevator's monitoring system, including continuous electrical signal sequences, mechanical motion parameter sequences, and discrete system event logs; S3: Perform protocol parsing and feature extraction on the original operating status data stream, and transform it into a set of device status parameters with time-series correlation; S4: Based on the set of equipment status parameters, dynamically generate structured status description information for natural language processing to depict the overall operating status and key component behavior of the target elevator within a specific time window; S5: Combine the structured state description information with relevant background knowledge fragments obtained through real-time retrieval from the multi-dimensional elevator knowledge base to form a formatted comprehensive query context; S6: Input the comprehensive query context into the core inference engine to drive the engine to perform a coherent analysis task and generate a preliminary assessment report that includes a security quantitative score, potential failure mode inference and confidence level, and specific maintenance recommendations. S7: Provide a human-computer interaction interface to receive instructions from maintenance personnel to review and revise the preliminary assessment report, including confirmation, adjustment or supplementation of the assessment conclusions; S8: Based on the correction instructions, calibrate the preliminary assessment report to generate the final version of the elevator safety assessment and operation and maintenance decision report.

[0006] Furthermore, step S3 also includes: S31: Identify the communication protocol specifications corresponding to different data sources in the original running status data stream, and decompose and decode the data stream according to the protocol specifications to obtain the original data points with timestamps; S32: For an electrical signal sequence, calculate its statistical characteristics within a preset time interval, including mean, variance, peak value, and energy distribution in a specific frequency band; S33: For mechanical motion parameter sequences, detect their waveform morphology and identify abnormal fluctuation periods, out-of-limit points, and trend drifts. S34: For discrete system event logs, parse the event codes and associated parameters, and cluster and sort them according to event level and occurrence time; S35: Align and merge all the features and analysis results obtained in steps S32 to S34 according to a unified time axis to form the set of device state parameters.

[0007] Furthermore, step S4 also includes: S41: From the set of equipment status parameters, select a subset of key parameters that characterize the status of the core subsystem of the elevator, including the traction system, door system, control system and safety protection system. For each subsystem, extract the corresponding operating mode, main parameter values ​​and their changing trends from the subset of key parameters. S42: Construct a predefined state description template framework, which contains multiple structured paragraph slots for filling different subsystem states, and each paragraph slot is associated with a specific semantic description specification; S43: Based on the aforementioned subset of key parameters and the extracted operating modes, parameter values, and trends, and in accordance with the semantic description specifications corresponding to each subsystem, automatically generate descriptive text fragments that conform to natural language syntax and are used to fill the corresponding paragraph slots. S44: Organize the aforementioned subset of key parameters into a separate data appendix in the form of parameter name: numerical unit; S45: Combine the state description template framework filled with descriptive text fragments with the data appendix to form the complete structured state description information.

[0008] Furthermore, step S5 also includes: S51: Construct a multi-dimensional search request using key entities, fault phenomenon description terms, and abnormal parameter names in the structured state description information as search keywords; S52: Perform parallel retrieval in the multidimensional elevator knowledge base to obtain the first background knowledge fragment and the second background knowledge fragment; S53: Design a context formatting template, fill it in according to the specific format of the formatting template, and generate a complete text block containing role definition, task description, input information and output specifications, as the comprehensive query context.

[0009] Furthermore, step S6 also includes: S61: Analyze the instruction prompts and task list in the context to clarify the evaluation and analysis role to be performed and the specific output requirements; S62: Sequentially read and understand the current elevator status description, construct an internal semantic representation of the target elevator's current operating status, and identify the explicitly indicated anomalies, trend risks, and potential inter-system interactions. S63: Deeply integrate relevant background knowledge, check the compliance of the current situation with the normative standards in the knowledge graph, make analogous reasoning with patterns in similar historical cases, and use the domain logical relationships in the knowledge graph to make multi-step reasoning about the root causes of potential failures. S64: Generate a quantitative safety score according to the output format requirements, comprehensively consider the severity of the current failure, system redundancy and risk development trend, list the inferred potential failure modes, give a confidence assessment based on reasoning logic for each failure mode, and generate maintenance inspection suggestions or parameter adjustment suggestions for each potential failure mode or risk point. S65: Integrate the safety quantitative score, potential failure mode inference and confidence level, and specific maintenance recommendations to form a preliminary assessment report.

[0010] Furthermore, step S7 also includes: S71: Display the preliminary assessment report and the key parts of the structured state description information on which the report is based on side by side on the human-computer interaction interface; S72: Provide interactive controls next to the concluding paragraphs or recommendations in the preliminary assessment report; S73: Receive instructions input by maintenance personnel through the interactive control, and record all interactive operations and their contents as a complete record of the correction instructions.

[0011] Furthermore, step S8 also includes: S81: Traverse and analyze the correction instruction record to identify all content marked as confirmed, all original text and modified content that have been modified, and all supplementary information. S82: Confirmed information shall be retained in the final report; S83: For the modified content, directly replace the corresponding paragraphs or suggestions in the original report with the modified text entered by the operations and maintenance personnel; S84: For the supplementary content, integrate it into the relevant section of the original report in a clear format or as a separate note; S85: Perform a consistency check on all changes to ensure that the revised and supplemented reports are logically consistent and contain no contradictory statements; S86: Based on the calibrated content, regenerate the overall summary and key conclusions of the report to form a final version of the elevator safety assessment and operation and maintenance decision report with a unified format.

[0012] Furthermore, the method also includes a report distribution and tracking step S9: S91: Based on the urgency and components involved in the maintenance recommendations in the final version of the elevator safety assessment and operation and maintenance decision report, automatically determine the list of responsible positions or personnel who need to be notified; S92: Through the pre-integrated communication interface, the key content or full text of the final version report is pushed to the corresponding operation and maintenance execution system or the terminal device of the designated personnel in the form of a work order; S93: Provide a report status tracking view in the human-machine interface, and display the work order execution status corresponding to each report. The status includes at least dispatched, processing, pending verification, and completed. S94: Receives work order feedback information from the operation and maintenance execution system. When the work order status is updated to completed, it triggers the subsequent status data collection of the relevant elevators and starts a new round of evaluation process to verify the effectiveness of maintenance measures.

[0013] Furthermore, the interaction optimization between the core reasoning engine and the multi-dimensional elevator knowledge base is achieved through the following mechanism: Before each retrieval in step S52, a lightweight pre-query is performed on the core inference engine using the core entities and fault keywords extracted from the structured state description information to obtain the engine's preliminary prediction of the current situation's focus based on its parameter weights. Using the initial focus prediction as supplementary guiding information, the weight allocation of the multi-dimensional search requests is adjusted so that the search process focuses more on knowledge domains that are highly relevant to the predicted focus. When combining the search results into the comprehensive query context, background knowledge fragments that have a high degree of matching with the preliminary focus prediction are highlighted to guide the core reasoning engine to give them higher priority in subsequent reasoning.

[0014] According to a second aspect of the present invention, an electronic device is claimed, comprising: One or more processors; A memory that stores one or more programs, which, when executed by one or more processors, enable the one or more processors to implement the elevator intelligent operation and maintenance and safety assessment method based on a large language model.

[0015] This invention discloses an intelligent elevator operation and maintenance and safety assessment method based on a large language model. It establishes a multi-dimensional elevator knowledge base integrating a structured knowledge graph and unstructured historical cases, and configures a large language model as the core reasoning engine. The method periodically collects raw elevator operating status data, which, after protocol parsing and feature extraction, is transformed into a set of equipment status parameters and dynamically generates a structured natural language status description. This status description is combined with relevant background knowledge fragments retrieved in real-time from the knowledge base to construct a comprehensive query context input to the large language model. The model then performs deep reasoning based on this context, generating a preliminary assessment report. A human-computer interaction interface is provided for operation and maintenance personnel to review and revise the report, and based on the revision instructions, a final decision report is generated. This invention achieves intelligent fusion and reasoning of multi-source heterogeneous data, combining domain knowledge and human experience to improve the foresight, accuracy, and efficiency of elevator safety assessments and operation and maintenance decision-making. Attached Figure Description

[0016] Figure 1 A flowchart illustrating the workflow of an elevator intelligent operation and maintenance and safety assessment method based on a large language model, as claimed in an embodiment of the present invention. Figure 2 The second workflow diagram of an elevator intelligent operation and maintenance and safety assessment method based on a large language model, as claimed in the embodiments of the present invention; Figure 3The third workflow diagram of an elevator intelligent operation and maintenance and safety assessment method based on a large language model, as claimed in the embodiments of the present invention; Figure 4 The fourth flowchart of an elevator intelligent operation and maintenance and safety assessment method based on a large language model, as claimed in the embodiments of the present invention, is shown. Detailed Implementation

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the 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.

[0018] The terms "first," "second," and "third" in this application are for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first," "second," or "third" may explicitly or implicitly include at least one of those features. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified. All directional indications in the embodiments of this application, such as up, down, left, right, front, back, etc., are only used to explain the relative positional relationships and movements between components in a specific orientation as shown in the accompanying drawings. If the specific orientation changes, the directional indications will change accordingly. Furthermore, the terms "including" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0019] References to embodiments herein mean that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0020] According to the first embodiment of the present invention, referring to Figure 1 This invention claims protection for a method for intelligent elevator operation and maintenance and safety assessment based on a large language model, comprising: S1: Establish and maintain a multi-dimensional elevator knowledge base, which integrates a structured elevator domain knowledge graph and unstructured elevator fault history case texts, and configures a large language model pre-trained in a general domain as the core reasoning engine. S2: Periodically collect raw operating status data streams from the target elevator's monitoring system, including continuous electrical signal sequences, mechanical motion parameter sequences, and discrete system event logs; S3: Perform protocol parsing and feature extraction on the original operating status data stream, and transform it into a set of device status parameters with time-series correlation; S4: Based on the set of equipment status parameters, dynamically generate structured status description information for natural language processing to depict the overall operating status and key component behavior of the target elevator within a specific time window; S5: Combine the structured state description information with relevant background knowledge fragments obtained through real-time retrieval from the multi-dimensional elevator knowledge base to form a formatted comprehensive query context; S6: Input the comprehensive query context into the core inference engine to drive the engine to perform a coherent analysis task and generate a preliminary assessment report that includes a security quantitative score, potential failure mode inference and confidence level, and specific maintenance recommendations. S7: Provide a human-computer interaction interface to receive instructions from maintenance personnel to review and revise the preliminary assessment report, including confirmation, adjustment or supplementation of the assessment conclusions; S8: Based on the correction instructions, calibrate the preliminary assessment report to generate the final version of the elevator safety assessment and operation and maintenance decision report.

[0021] In this embodiment, a Schindler M-type elevator installed in a commercial building is used as the evaluation object. The system detects an anomaly in its door system and then triggers and completes a complete intelligent evaluation and operation and maintenance decision-making process.

[0022] The operations and maintenance center deployed this method and established a multi-dimensional elevator knowledge base. This knowledge base integrates two core components: first, a structured elevator domain knowledge graph, which includes national elevator safety standards and Schindler M-model technical manuals such as traction machine rated current, standard values ​​for door operator opening and closing force, and fault transmission relationships between components, such as poor door lock contact leading to the control system misjudging the door as not closed; second, an unstructured elevator fault history case text library, containing a large number of past maintenance reports, such as Case A: a certain elevator experienced high door opening and closing resistance due to dust accumulation on the door guide rails, causing periodic exceedances of the door operator current. Simultaneously, the system incorporates a large language model pre-trained with general-purpose text as the core inference engine, which possesses powerful natural language understanding and generation capabilities.

[0023] The target elevator's monitoring system operates continuously, collecting raw operational status data streams in real time during a morning work period. This data stream includes: continuous electrical signal sequences from the frequency converter, such as three-phase current and voltage waveforms; mechanical motion parameter sequences from encoders and sensors, such as car speed, door knife position, and door lock status; and discrete system event logs generated by the elevator's main controller, such as door zone sensing faults, record number XXX.

[0024] The raw data stream is processed to identify and parse the Modbus and CAN bus protocols, decoding the raw data points with timestamps. Then, for the electrical signals, the effective value and peak value trend of the door operator motor current during the most recent dozens of door opening and closing actions are calculated. For the mechanical parameters, abnormal fluctuations in the resistance curve at the end of the door closing phase are identified. For the event log, multiple warning records related to door zone sensing are parsed. Finally, these features are aligned along the time axis to form a set of equipment status parameters, for example: {Parameter: Door operator peak current, Value: High, Time: T1-T2; Parameter: Resistance fluctuation at the end of door closing phase, Value: Present, Time: T1-T2; Parameter: Door zone fault log, Value: Increased frequency, Time: T0-T2}.

[0025] Based on the above set, the system dynamically generates a structured state description. This information is organized in text format and contains the following: [Operating Status Summary] The target elevator, Schindler M model, number B1-02, has been running frequently in the past 2 hours. The operating parameters of the traction system and control system are stable, but there is a significant anomaly in the door system.

[0026] [Detailed Description of Door System] During the final stage of the door closing action, the peak drive current of the door operator consistently exceeds the historical average value for this model, and the resistance sensor shows irregular fluctuations. The door zone sensor triggered intermittent fault logs 3 times in the last 10 operations, and the system has automatically reset.

[0027]

Data Appendix

[0028] This description systematically portrays the overall state of the elevator and highlights key anomalies in the door system.

[0029] Using the keywords "Schindler M model," "high door operator current," "fluctuating closing resistance," and "door zone sensor fault," the system searches the knowledge base. From the knowledge graph, it retrieves first-level background knowledge fragments, including door system composition, door zone safety circuit principles, possible causes of excessive current such as increased resistance, and motor aging. From the historical case database, it finds three cases describing similar phenomena through semantic matching, extracting text indicating insufficient guide rail lubrication or worn door rollers as the cause of the fault, and suggesting cleaning the guide rails and adjusting the rollers as the second-level background knowledge fragment. Subsequently, the system formats this content according to a preset template, generating a comprehensive query context containing four main parts: instruction role-playing, current state description, relevant background knowledge, and output format requirements, ready to be input into the large language model.

[0030] The above context is input into the core inference engine. The engine first understands the instructions, analyzes the current state description, and identifies high current, high resistance, and intermittent faults as the core anomalies. Next, it integrates background knowledge for reasoning: the knowledge graph provides theoretical fault modes, and historical cases provide practical experience. The engine infers that the intermittent faults may be related to a slight door position shift caused by increased resistance. Finally, it generates a preliminary assessment report, including: a safety score (e.g., rated as medium risk), potential fault mode inferences ranked by probability: 1. Door rail obstruction or roller wear; 2. Slight aging of the door operator drive module, the confidence level of each inference, and specific maintenance recommendations such as prioritizing inspection and cleaning of the door rails, measuring roller clearance, and secondly, testing the output stability of the door operator drive module.

[0031] The preliminary assessment report, along with key status data, was displayed on the maintenance engineer's human-machine interface. Engineer Wang carefully read the report and, drawing on his experience, noticed a potential oversight: loose wiring in the door zone sensor itself could also cause intermittent malfunctions. He entered the suggestion to check the door zone sensor's plug connection using the supplementary text input box on the interface. Simultaneously, he clicked on the related case query, and the system displayed two historical cases of loose wiring for his reference. Engineer Wang then confirmed the rest of the report. All operations were recorded as correction instructions.

[0032] Calibration was performed according to Engineer Wang's correction instructions. The confirmed content from the original report was retained, and Engineer Wang's supplementary comments regarding wiring troubleshooting were added to the end of the maintenance recommendations section in a prominent format as supplementary suggestions from an operations and maintenance expert. After a consistency check, the system generated the final version of the elevator safety assessment and operations and maintenance decision report, which integrated model reasoning with expert experience.

[0033] The entire assessment process data, including the final report, the structured state description information that triggered the assessment, the core part of the comprehensive query context used for reasoning, and Wang's correction instructions, is automatically packaged into a new knowledge unit. This unit is assigned a unique ID and indexed by gate system anomalies such as high current and intermittent faults. It is then synchronously archived into the historical case text library of the knowledge base, enriching the system's experience in handling similar situations in the future.

[0034] Furthermore, referring to Figure 2 Step S3 further includes: S31: Identify the communication protocol specifications corresponding to different data sources in the original running status data stream, and decompose and decode the data stream according to the protocol specifications to obtain the original data points with timestamps; S32: For an electrical signal sequence, calculate its statistical characteristics within a preset time interval, including mean, variance, peak value, and energy distribution in a specific frequency band; S33: For mechanical motion parameter sequences, detect their waveform morphology and identify abnormal fluctuation periods, out-of-limit points, and trend drifts. S34: For discrete system event logs, parse the event codes and associated parameters, and cluster and sort them according to event level and occurrence time; S35: Align and merge all the features and analysis results obtained in steps S32 to S34 according to a unified time axis to form the set of device state parameters.

[0035] In this embodiment, the implementation details of the protocol parsing and feature extraction in step S3 are as follows: The system identifies that the elevator controller uses the CAN bus protocol, while the frequency converter uses the Modbus RTU protocol. Based on their respective protocol specifications, the data stream is unpacked, verified, and decoded to obtain a series of raw data points with precise millisecond-level timestamps, such as time T1: door lock signal = 1 (closed), current value = A.

[0036] For the three-phase current of the door operator motor electrical signal sequence obtained after decoding, the system takes a single door closing action as the time interval, calculates the mean of the current signal in the interval to reflect the average load, the variance to reflect the degree of fluctuation, and the peak value to reflect the maximum instantaneous load, and analyzes the energy distribution of its specific frequency band by performing a fast Fourier transform on the signal to detect whether there are abnormal harmonics.

[0037] For mechanical motion parameter sequences, such as the door closing displacement curve, the system detects its smoothness. In this example, the algorithm detected a non-standard sawtooth-shaped abnormal fluctuation period in the displacement curve during the last 10% of the door closing stroke, and at the same time, the resistance sensor reading momentarily exceeded the normal threshold.

[0038] For discrete system event logs, the system parses the code of each log entry, such as E45 representing a door zone sensing anomaly, and extracts related parameters such as the floor and car status at the time of the incident. Then, all E45 events are clustered by occurrence time, and their frequency of occurrence within the most recent time window is counted, revealing a clustering characteristic.

[0039] Finally, the current statistical characteristics obtained from the above steps, the waveform anomaly markers at the end of the closing phase, and the clustering analysis results of the E45 logs were time-aligned with time T1 as the reference. For example, the correlation shows that at the moment when the resistance fluctuation exceeded the limit (T1.5), the current peak also appeared simultaneously, and was recorded in the E45 log approximately 2 seconds later. All these correlated time attributes and information from the current sensor, displacement encoder, and controller logs together constitute the set of equipment status parameters used in this evaluation.

[0040] Furthermore, referring to Figure 3 Step S4 further includes: S41: From the set of equipment status parameters, select a subset of key parameters that characterize the status of the core subsystem of the elevator, including the traction system, door system, control system and safety protection system. For each subsystem, extract the corresponding operating mode, main parameter values ​​and their changing trends from the subset of key parameters. S42: Construct a predefined state description template framework, which contains multiple structured paragraph slots for filling different subsystem states, and each paragraph slot is associated with a specific semantic description specification; S43: Based on the aforementioned subset of key parameters and the extracted operating modes, parameter values, and trends, and in accordance with the semantic description specifications corresponding to each subsystem, automatically generate descriptive text fragments that conform to natural language syntax and are used to fill the corresponding paragraph slots. S44: Organize the aforementioned subset of key parameters into a separate data appendix in the form of parameter name: numerical unit; S45: Combine the state description template framework filled with descriptive text fragments with the data appendix to form the complete structured state description information.

[0041] In this embodiment, the process of generating structured state description information is detailed below: From a large set of equipment status parameters, key parameters were screened. The focus was first on core subsystems. For the traction system, current and speed feedback were screened to determine stable operation. For the door system, peak current, resistance curve, door lock opening / closing sequence, and E45 event log frequency were selected as key anomalies. For the control and safety protection systems, feedback signals and safety circuit status were checked, and no persistent anomalies were found. The system identified the door system's operating mode as frequent door opening and closing, with the main parameter being a consistently high peak current, whose fluctuations intensified with increasing operating frequency.

[0042] A predefined state description template framework is invoked. This framework contains five fixed structured paragraph slots: [Operating Status Summary], [Traction System Description], [Door System Description], [Control and Safety System Description], and [Data Appendix]. Each slot has a corresponding semantic description specification. For example, the [Door System Description] slot must include descriptions of whether the door opening and closing action is smooth, whether there are abnormal noises or vibrations, and the status of relevant sensors and logs. S43: Based on the specification, the system automatically generates descriptive text fragments to fill the slots. It compares the peak current value of the door operator with the normal threshold range of this model in the knowledge base to generate a qualitative description of the door operator's drive current being too high. It maps the event log encoding E45 to a standard statement reporting intermittent signal loss from the door zone sensor. It summarizes the sawtooth fluctuations of the resistance curve as a trend description of periodic oscillations and stagnation at the end of the door closing phase. These generation processes are based on logical judgments and string templates and do not involve complex model training.

[0043] Meanwhile, the key raw parameters selected, such as Door_Motor_Peak_Current: [numerical value] A, Close_Resistance_Max: [numerical value] N, etc., are organized into the data appendix according to the format of parameter name: numerical unit.

[0044] Finally, the filled descriptive text is combined into the template framework, and a data appendix is ​​attached at the end to form the final structured state description information. Although this information is text, its internal structure is clear, allowing the large language model to easily distinguish between a qualitative description of a gate system and the original data list of a gate system, greatly improving the accuracy and relevance of subsequent reasoning.

[0045] Furthermore, step S5 also includes: S51: Construct a multi-dimensional search request using key entities, fault phenomenon description terms, and abnormal parameter names in the structured state description information as search keywords; S52: Perform parallel retrieval in the multidimensional elevator knowledge base to obtain the first background knowledge fragment and the second background knowledge fragment; S53: Design a context formatting template, fill it in according to the specific format of the formatting template, and generate a complete text block containing role definition, task description, input information and output specifications, as the comprehensive query context.

[0046] In this embodiment, the process of constructing the comprehensive query context is specifically implemented as follows: The Schindler M model, door operator, high current, door closing obstruction, and E45 intermittent fault are extracted from the structured status description information as search keywords to construct a multi-dimensional search request covering equipment model, components, phenomena, and log codes.

[0047] Parallel retrieval is performed in the knowledge base. In the structured knowledge graph, the gantry crane entity is located through graph queries, and its attributes such as rated current and common fault modes are obtained, along with related entities and relationships such as the door guide rail, door rollers, and sensors. This extracts the first background knowledge fragment regarding the potential for increased current due to increased mechanical resistance. In the unstructured historical case database, the system calculates the semantic vector similarity between the current state description text and all cases, returning the three most similar cases. From these cases, the system precisely extracts plain text paragraphs detailing the root cause of the problem and the solution steps, forming the second background knowledge fragment. This fragment includes specific operational descriptions of guide rail cleaning and roller adjustment.

[0048] Use a carefully designed, fixed-context formatting template. The template clearly states: Part One instructs you, as an experienced elevator safety assessment expert, to complete a safety assessment, fault deduction, and recommendations based on the provided status and knowledge. Part Two, titled "##Current Elevator Status Description##", requires you to paste the complete structured status description information generated by S4. Part Three, titled "##Relevant Background Knowledge##", has two subheadings: "[Technical Specifications and Related Aspects]" and "[Historical Similar Cases]", which should be filled with the first and second background knowledge snippets, respectively. Part Four, titled "##Output Format Requirements##", explicitly requires the report to be generated in three parts: I. Safety Assessment (including scoring); II. Potential Failure Mode Analysis (ranked by probability with confidence level); and III. Maintenance Recommendations.

[0049] Fill in the above four parts strictly according to the template's format, headings, and order to generate a comprehensive query context text block of several thousand characters, with a clear structure and complete information. This text block not only provides facts but also, through structuring and instructions, rigorously constrains and guides the thinking and output framework of the large language model.

[0050] Furthermore, referring to Figure 4 Step S6 further includes: S61: Analyze the instruction prompts and task list in the context to clarify the evaluation and analysis role to be performed and the specific output requirements; S62: Sequentially read and understand the current elevator status description, construct an internal semantic representation of the target elevator's current operating status, and identify the explicitly indicated anomalies, trend risks, and potential inter-system interactions. S63: Deeply integrate relevant background knowledge, check the compliance of the current situation with the normative standards in the knowledge graph, make analogous reasoning with patterns in similar historical cases, and use the domain logical relationships in the knowledge graph to make multi-step reasoning about the root causes of potential failures. S64: Generate a quantitative safety score according to the output format requirements, comprehensively consider the severity of the current failure, system redundancy and risk development trend, list the inferred potential failure modes, give a confidence assessment based on reasoning logic for each failure mode, and generate maintenance inspection suggestions or parameter adjustment suggestions for each potential failure mode or risk point. S65: Integrate the safety quantitative score, potential failure mode inference and confidence level, and specific maintenance recommendations to form a preliminary assessment report.

[0051] In this embodiment, the model first reads and parses the instructions at the beginning of the context to clarify that it needs to play the role of an elevator safety assessment expert, and the task objective is to generate an assessment report containing three specific items.

[0052] The model begins to delve into the "Current Elevator Status Description" section. It doesn't simply identify keywords, but rather constructs an internal semantic representation: understanding that high door operator current and fluctuating door closing resistance are strongly correlated phenomena occurring simultaneously, and that intermittent fault logs are a possible outcome. It identifies the anomaly as a synchronous anomaly in current and resistance, the potential risk of this anomaly worsening with continued operation, and infers a potential interaction between the mechanical system guide rails / rollers, the electrical system door operator, and the sensor system sensors.

[0053] The model then deeply integrates relevant background knowledge. It checks the current current value against the rated current in the knowledge graph, confirming the high reading. Through analogy with historical cases, it finds a high similarity between the current phenomenon and a guide rail obstruction case. Utilizing the domain logic relationship in the knowledge graph—increased resistance -> increased current -> inaccurate door position -> sensor false alarm—it performs multi-step reasoning, connecting the surface phenomena into a reasonable chain of fault hypotheses.

[0054] Based on the above analysis, the model begins to generate output. First, it considers the potential impact of the fault, the intermittent redundancy of the current system, and the worsening trend, providing a quantitative safety score, such as a percentage or grade system. Second, it details the inferred fault modes: the primary possibility is increased mechanical resistance and guide rail / roller problems, given with high confidence; the secondary possibility is degraded performance of the door operator drive module, given with medium confidence. Finally, for the primary possibility, it generates specific maintenance recommendations for cleaning the guide rails and checking the wear of the rollers; for the secondary possibilities and sensor issues, it recommends testing the door operator's no-load current and checking the sensor wiring.

[0055] The model organizes all the above content into a well-structured, professionally written text according to the output format requirements, thus completing the generation of the preliminary evaluation report.

[0056] Furthermore, step S7 also includes: S71: Display the preliminary assessment report and the key parts of the structured state description information on which the report is based on side by side on the human-computer interaction interface; S72: Provide interactive controls next to the concluding paragraphs or recommendations in the preliminary assessment report; S73: Receive instructions input by maintenance personnel through the interactive control, and record all interactive operations and their contents as a complete record of the correction instructions.

[0057] In this embodiment, the specific implementation and interaction process of the human-computer interaction interface are as follows: After logging into the system, maintenance engineer Wang sees two main panels side-by-side on the main interface: the left panel presents the preliminary assessment report generated by the large language model, including scores, fault inferences, and recommendations; the right panel is collapsed to show key parts of the structured status description information on which the report was based, such as anomaly parameter summaries and major event logs, which Wang can click to expand and view details.

[0058] Next to each important conclusion in the assessment report on the left, there are interactive controls. For example, at the end of the paragraph on potential failure mode 1: increased mechanical resistance, there is a "Confirm" button, a "Modify" input box, a "Supplement" input box, and a "Query Similar Cases" link, which is an entry point for querying related cases.

[0059] Engineer Wang reviewed the report. He approved the analysis of fault mode 1 and clicked the "Confirm" button, changing the background color of that section to light green. Regarding the maintenance recommendations, he felt that while comprehensive, it might have overlooked a direct inspection of the sensor wiring. Therefore, he entered the following in the "Supplement" input box for the corresponding recommendation: "Supplementary recommendation: During on-site operations, the connectors of the door sensors should be checked for looseness and secured with anti-loosening agent." He also had questions about the root cause of the intermittent faults and clicked the "Query Similar Cases" link. The system then displayed two more historical case summaries emphasizing that loose wiring caused intermittent faults in a pop-up window below.

[0060] The backend system fully recorded all of Engineer Wang's actions: confirmation of fault mode 1, supplementary text content for maintenance recommendations, and triggered case query records. These records together constitute the set of corrective instructions for this assessment task.

[0061] Furthermore, step S8 also includes: S81: Traverse and analyze the correction instruction record to identify all content marked as confirmed, all original text and modified content that have been modified, and all supplementary information. S82: Confirmed information shall be retained in the final report; S83: For the modified content, directly replace the corresponding paragraphs or suggestions in the original report with the modified text entered by the operations and maintenance personnel; S84: For the supplementary content, integrate it into the relevant section of the original report in a clear format or as a separate note; S85: Perform a consistency check on all changes to ensure that the revised and supplemented reports are logically consistent and contain no contradictory statements; S86: Based on the calibrated content, regenerate the overall summary and key conclusions of the report to form a final version of the elevator safety assessment and operation and maintenance decision report with a unified format.

[0062] In this embodiment, the process of calibrating the report based on the human-computer interaction results is as follows: The system iterates through and analyzes the received correction instruction records. It identifies confirmed fault mode 1 and other related information, as well as supplementary text regarding checking sensor wiring.

[0063] For confirmed content, the system retains it when generating the final report without making any changes.

[0064] In this example, no modification command was triggered. If it had, the system would have replaced the old content with the new text.

[0065] Regarding Wang's additional suggestions, the system prefixes them with "Supplementary Suggestions from Operations Experts" and integrates them as a separate entry at the end of Section III, Operations Suggestions, to distinguish them from the original suggestions and make them easier to view.

[0066] Perform a consistency check to ensure that the newly added inspection wiring recommendations are logically consistent with the existing sensor intermittent fault descriptions in the report and that there are no contradictions.

[0067] Based on all the above-mentioned calibrated content, the system regenerates the report summary, updates the list of key conclusions, and ensures that the final report reflects the integration of model reasoning and expert experience, forming a final version report with standardized format and complete content.

[0068] Furthermore, the method also includes a report distribution and tracking step S9: S91: Based on the urgency and components involved in the maintenance recommendations in the final version of the elevator safety assessment and operation and maintenance decision report, automatically determine the list of responsible positions or personnel who need to be notified; S92: Through the pre-integrated communication interface, the key content or full text of the final version report is pushed to the corresponding operation and maintenance execution system or the terminal device of the designated personnel in the form of a work order; S93: Provide a report status tracking view in the human-machine interface, and display the work order execution status corresponding to each report. The status includes at least dispatched, processing, pending verification, and completed. S94: Receives work order feedback information from the operation and maintenance execution system. When the work order status is updated to completed, it triggers the subsequent status data collection of the relevant elevators and starts a new round of evaluation process to verify the effectiveness of maintenance measures.

[0069] In this embodiment, after completing the core evaluation process, the system executes an additional report distribution and tracking step S9: Analyzing the final report, the maintenance recommendations included both mechanical adjustments to the guide rails and rollers, and electrical checks of sensor wiring. According to pre-defined rules, this type of comprehensive task automatically requires notification to the foremen of both the mechanical and electrical maintenance teams, with a copy sent to the operations supervisor.

[0070] By using the API interface with WeChat Work / work order system, a maintenance work order is generated from the summary of this assessment, risk level, and specific maintenance recommendations, and then pushed to the work order list on the mobile terminal and computer of the responsible personnel.

[0071] On the human-machine interface of the operations and maintenance center, the report tracking view lists all generated reports. The work order associated with this report is initially in the status of "dispatched." After the maintenance team leader receives it, the status changes to "processing."

[0072] After the maintenance team completed their on-site work, they reported the completion status in the work order system and uploaded photos of the processed components. The work order status was updated to "Completed." This status update triggered the system to focus on monitoring and collecting relevant parameters of the door system of elevator B1-02 within the next 24 hours. The system automatically initiated a new round of status description and evaluation processes, starting from S2, generating a new evaluation report to verify whether measures such as cleaning the guide rails and tightening the wiring effectively eliminated the original abnormal state, thus forming a closed loop for verifying the effectiveness of the maintenance.

[0073] Furthermore, the interaction optimization between the core reasoning engine and the multi-dimensional elevator knowledge base is achieved through the following mechanism: Before each retrieval in step S52, a lightweight pre-query is performed on the core inference engine using the core entities and fault keywords extracted from the structured state description information to obtain the engine's preliminary prediction of the current situation's focus based on its parameter weights. Using the initial focus prediction as supplementary guiding information, the weight allocation of the multi-dimensional search requests is adjusted so that the search process focuses more on knowledge domains that are highly relevant to the predicted focus. When combining the search results into the comprehensive query context, background knowledge fragments that have a high degree of matching with the preliminary focus prediction are highlighted to guide the core reasoning engine to give them higher priority in subsequent reasoning.

[0074] In the long-term operation of this method system, the following mechanisms were implemented to optimize the targeting of knowledge retrieval: Before each knowledge retrieval, the core content of the structured state description information, such as high gantry current accompanied by fluctuations in closing resistance, is input into the large language model. However, only one simple question is posed: Based on the current description, which aspects of knowledge do you think require the most attention? The large language model, based on its built-in general knowledge and pattern recognition capabilities, quickly provides a preliminary prediction of areas of interest, such as the cause of the abnormal resistance in the mechanical transmission components and its impact on position sensing.

[0075] This predicted text is transformed into keyword enhancement signals, which are then used to adjust the weighting of subsequent multi-dimensional retrieval requests to the knowledge base. Specifically, when retrieving the knowledge graph, entities and relationships such as door rails, resistance, and wear are given higher weights; when retrieving historical cases, semantic similarity calculations are also biased towards cases containing these keywords.

[0076] Ultimately, when constructing the comprehensive query context, the system adds a "Highlighted Reference" marker to background knowledge fragments returned by the retrieval that highly match the two predicted concerns of mechanical resistance and sensing effects. When the large language model reads the context during the inference phase, this marked knowledge receives more attention resources, thereby guiding the inference process to be more efficient and focused on the most likely root causes of the problem.

[0077] According to a second embodiment of the present invention, the present invention claims protection for an electronic device comprising: One or more processors; A memory storing one or more programs, which, when executed by one or more processors, enable the processors to implement the aforementioned intelligent elevator operation and maintenance and safety assessment method based on a large language model. In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between apparatuses or units, and may be electrical, mechanical, or other forms.

[0078] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated units described above can be implemented in hardware or as software functional units. The above are merely embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made based on the description and drawings of this application, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

[0079] The specific embodiments of the invention have been described in detail above, but they are only examples, and this application is not limited to the specific embodiments described above. For those skilled in the art, any equivalent modifications or substitutions to the invention are also within the scope of this application. Therefore, all equivalent changes, modifications, and improvements made without departing from the spirit and principles of this application should be covered within the scope of this application.

Claims

1. An elevator intelligent operation and maintenance and safety evaluation method based on a large language model, characterized in that, Comprise: S1: Establish and maintain a multi-dimensional elevator knowledge base, integrating structured elevator domain knowledge graph and unstructured elevator fault history case text, and configuring a general domain pre-trained large language model as the core reasoning engine; S2: Collect raw running state data streams from the target elevator's monitoring system in a timely manner, including continuous electrical signal sequences, mechanical motion parameter sequences, and discrete system event logs; S3: Protocol analysis and feature extraction are performed on the raw running state data stream to convert it into a set of time-correlated device state parameters; S4: Based on the device state parameter set, dynamically generate structured state description information for natural language processing, depicting the overall running situation and key component behavior of the target elevator within a specific time window; S5: Combine the structured state description information and the relevant background knowledge fragments obtained from the multi-dimensional elevator knowledge base into a formatted comprehensive query context; S6: Input the comprehensive query context into the core reasoning engine to drive the engine to perform coherent analysis tasks and generate a preliminary evaluation report containing safety quantitative scores, potential fault mode inference and confidence, and specific maintenance recommendations; S7: Provide a human-computer interaction interface to receive the review and modification instructions of the preliminary evaluation report from the operation and maintenance personnel, including confirmation, adjustment or supplement of the evaluation conclusion; S8: According to the modification instructions, calibrate the preliminary evaluation report to generate the final version of the elevator safety evaluation and maintenance decision report.

2. The method of claim 1, wherein, The step S3 further comprises: S31: Identify the communication protocol specifications corresponding to different data sources in the raw running state data stream, and according to the protocol specifications, disassemble and decode the data stream to obtain raw data points with timestamps; S32: For electrical signal sequences, calculate their statistical features in a pre-set time interval, including mean, variance, peak value and energy distribution in a specific frequency band; S33: For mechanical motion parameter sequences, detect their waveform patterns, identify abnormal fluctuation periods, out-of-limit points and trend drifts; S34: For discrete system event logs, analyze event codes and associated parameters, cluster and sort them by event level and occurrence time; S35: Align and fuse all features and analysis results obtained in steps S32 to S34 according to a unified time axis to form the device state parameter set.

3. The method of claim 1, wherein, The step S4 further comprises: S41: From the device state parameter set, select a key parameter subset representing the state of the elevator core subsystem, including the traction system, door system, control system and safety protection system, and for each subsystem, extract the corresponding running mode, main parameter value and its change trend from the key parameter subset; S42: Construct a pre-defined state description template framework, containing multiple structured paragraph slots for filling in different subsystem conditions, each paragraph slot associated with a specific semantic description specification; S43: Based on the key parameter subset and the extracted operating mode, parameter value and trend, according to the semantic description specification corresponding to each subsystem, automatically generate a descriptive text segment in accordance with the natural language grammar for filling in the corresponding paragraph slot; S44: The key parameter subset itself is arranged as an independent data appendix part in the form of parameter name:value unit; S45: The state description template framework filled with descriptive text segments and the data appendix part are combined to form a complete structured state description information.

4. The method of claim 1, wherein, The step S5 further comprises: S51: Taking the key entities, fault phenomenon description words and abnormal parameter names in the structured state description information as retrieval keywords, a multi-dimensional retrieval request is constructed; S52: The first and second background knowledge segments are obtained by parallel retrieval in the multi-dimensional elevator knowledge base; S53: A context formatting template is designed, and a complete text block containing role definition, task description, input information and output specification is generated by filling in the specific format of the formatting template, serving as the comprehensive query context.

5. The method according to claim 1 or 4, characterized in that, The step S6 further comprises: S61: The instruction prompts and task lists in the context are parsed to determine the evaluation and analysis roles and specific output requirements to be performed by itself; S62: The current elevator state description part is sequentially read and understood to build an internal semantic representation of the current operating condition of the target elevator, and the abnormal points, trend risks and potential interactions between systems explicitly pointed out therein are identified; S63: The current condition is deeply combined with the relevant background knowledge part to perform conformity checking with the specification standards in the knowledge graph, analogical reasoning with the patterns in the historical similar cases, and multi-step reasoning on the root cause of the potential fault by using the domain logic relationships in the knowledge graph; S64: A safety quantification score is generated according to the output format requirements, the severity of the current fault, system redundancy and risk development trend are comprehensively considered, the inferred potential fault modes are listed, the confidence evaluation based on the reasoning logic is given for each fault mode, and maintenance check suggestions or parameter adjustment suggestions are generated for each potential fault mode or risk point; S65: The safety quantification score, potential fault mode inference and confidence, and specific maintenance suggestions are integrated to form a preliminary evaluation report.

6. The method of claim 1, wherein, The step S7 further comprises: S71: The preliminary evaluation report and the key part of the structured state description information on which the report is based are displayed side by side on the human-computer interaction interface; S72: An interactive control is provided beside the concluding paragraph or suggestion item of the preliminary evaluation report; S73: The instructions input by the operation and maintenance personnel through the interactive control are received, and all interactive operations and their contents are recorded as complete records of the correction instructions.

7. The method of claim 6, wherein, The step S8 further comprises: S81: The correction instruction records are analyzed to identify all confirmed contents, all modified original texts and modified contents, and all supplemented additional information; S82: The confirmed contents are retained in the final report; S83: For the modified content, directly replace the corresponding paragraph or suggestion item in the original report with the modified text input by the operation and maintenance personnel; S84: For the supplemented content, integrate it at the end of the relevant chapter of the original report or as an independent note section in a clear format; S85: Perform consistency check on all changes to ensure that the modified and supplemented report is logically self-consistent and has no contradictory statements; S86: Based on the calibrated content, regenerate the overall summary and key conclusions of the report to form the final version of the elevator safety evaluation and operation and maintenance decision report in a unified format.

8. The method of claim 1, wherein, The method further comprises a report distribution and tracking step S9: S91: According to the urgency and involved components of the maintenance suggestions in the final version of the elevator safety evaluation and operation and maintenance decision report, automatically determine the list of responsible positions or personnel who need to be notified; S92: Through the pre-integrated communication interface, push the key content or full text of the final version report to the corresponding operation and maintenance execution system or the terminal device of the designated personnel in the form of a work order; S93: Provide a report status tracking view on the human-computer interaction interface, and display the work order execution status corresponding to each report, including at least dispatched, in process, to be verified, and completed; S94: Receive work order feedback information from the operation and maintenance execution system, and when the work order status is updated to completed, trigger subsequent state data collection of the related elevator, and start a new round of evaluation process to verify the effectiveness of the maintenance measures.

9. The method of claim 1, wherein, The interaction optimization between the core reasoning engine and the multi-dimensional elevator knowledge base is achieved through the following mechanisms: Before each execution of step S52, use the core entities and fault keywords extracted from the structured state description information to perform a lightweight pre-query on the core reasoning engine, and obtain the preliminary focus point prediction of the engine based on its parameter weight for the current context; Use the preliminary focus point prediction as supplementary guidance information to adjust the weight distribution of the multi-dimensional retrieval request, so that the retrieval process focuses more on knowledge fields highly related to the predicted focus point; When combining the retrieval results into the comprehensive query context, highlight the background knowledge fragments that match the preliminary focus point prediction to guide the core reasoning engine to give them higher priority in subsequent reasoning.

10. An electronic device, comprising: Comprise: One or more processors; A memory having one or more programs stored thereon, when the one or more programs are executed by the one or more processors, cause the one or more processors to implement a large language model-based elevator intelligent operation and maintenance and safety evaluation method according to any one of claims 1 to 9.

Images