Rail transit health status assessment system, method, device and medium

Abstract

The present disclosure relates to the field of rail transit technology, and in particular to a rail transit health status assessment system, method, equipment and medium. The present disclosure builds a three-dimensional vertical decision-making intelligent body through the three-layer three-dimensional structure design of the terminal layer, the distribution layer and the centralized layer, and the interwoven structure of the multi-language model and the multi-expert model of the distribution layer, so as to simplify the parameters and input scale required by the final decision-making intelligent body, improve the operation efficiency, and solve the problem of high computing power requirements. In the vertical direction, by using a multi-language model and integrating multiple language models, the output credibility problem of a single model is solved. In the horizontal direction, by using a multi-expert model and integrating the strategy results under different optimization angles, the monotony problem of the algorithm is solved. The centralized layer's three-out-of-two judgment scheme and the expert weight judgment scheme, as well as the design of the boundary conditions of the large model, solve the problem that the output of a single model is uncertain and the results of each calculation are different, thereby improving the authority and credibility of the rail transit system model.

Description

Technical Field

[0001] The present disclosure relates to the field of rail transit technology, and in particular to a rail transit health status assessment system, method, device, and medium. Background Art

[0002] The rail transit industry involves passenger transportation and is responsible for personnel safety, making reliable system operation crucial. Due to the complexity of the system, failure of any key component can trigger a chain reaction, leading to systemic safety incidents. Health status assessments enable the establishment of a full-lifecycle equipment status monitoring system, identifying potential single points of failure. This allows for timely maintenance intervention before system-level failures occur, preventing safety incidents while ensuring operational efficiency.

[0003] Traditional health status assessments rely primarily on expert scoring systems. Based on their personal knowledge, domain experts assign weights or ratings to various indicators, such as fault type and frequency. This weighted calculation ultimately yields a rough assessment of the equipment's health status. This health status assessment method provides only a rough assessment of the system's status and cannot accurately determine whether maintenance is necessary or to what extent. Furthermore, different experts have different expertise and differing risk assessments, making it difficult to establish an objective, unified quantitative benchmark for scoring, severely limiting the credibility of the assessment results.

[0004] With the rapid development of big model technology, various industries are actively applying it. The goal is to use algorithms and computer systems to simulate complex intelligent logical decision-making, enabling machines to handle complex problems and self-improve, thereby more efficiently outputting final solutions. The performance of big models is influenced by both the training data and the model structure and algorithm. Big models learn patterns, regularities, and features from data to perform various tasks, so data directly influences their performance. The richer the data, the more input information the big model receives, and the better the training results. However, this requires more computing power. Big models have a vast number of internal parameters and complex neural networks. Even with the same input data, different model structures and algorithms can lead to different outputs, making it difficult to intuitively judge the quality of different model outputs. Even with the exact same model structure and algorithm, the output is inherently uncertain, with each calculation yielding different results.

[0005] System equipment health assessment involves evaluating the equipment's status to determine its current operating status, performance, and reliability. This assessment can assess past equipment failures, identify potential faults and issues, and implement appropriate maintenance measures to improve equipment reliability and efficiency.

[0006] In engineering applications, large models rely entirely on data correlation mining and network calculations to fit equipment states, which cannot directly correspond to the direct causes of equipment failure. Therefore, the evaluation conclusions cannot provide a verifiable physical basis for maintenance decisions, making the model output difficult to fully trust. In addition, because the output of large models requires repeated iterative debugging, for scenarios requiring optimization from multiple angles, such as high maintenance efficiency, low economic costs, and high equipment reliability, calculations in different dimensions cannot be optimized simultaneously if the computing power is fixed. To improve the optimization effect, multiple iterations or increased resource consumption are required.

[0007] The existing large model technology has the following problems:

[0008] 1. Limited credibility. Due to the complex nonlinear structure of neural networks, the decision-making process of a single model is hidden within multiple layers of parameters. This makes it difficult to visually display the basis for the final result and trace the calculation process. Therefore, it is impossible to intuitively judge whether the logic behind the conclusion is completely correct. Furthermore, even if an error in the conclusion is manually discovered, the high degree of coupling of parameters in the neural network makes it impossible to infer the location of the deviation from the erroneous result. Therefore, it is impossible to trace back and adjust the algorithm to correct the model and reach the correct conclusion.

[0009] 2. Monotonous decision-making algorithms. Single, large-model agents achieve general capabilities through pre-training on massive amounts of data, but their architectural design is often optimized for specific tasks. After data is input, there are clear analysis premises and objectives, making them inadequate for fully open application scenarios. For example, when inputting fault data, the agent is required to provide data analysis results, which should be accompanied by specific conditions: whether a particular fault phenomenon will lead to system failure; the impact of maintenance measures taken after a particular fault phenomenon occurs on the subsequent failure rate; and whether the repair or replacement cost is higher after a particular fault phenomenon. However, actual applications require that after inputting fault data, the agent must perform calculations to comprehensively determine the optimal maintenance method. For phenomena that do not cause system failure, the lowest-cost maintenance method is used. For phenomena that do cause system failure, a compromise maintenance plan is adopted based on the maintenance cost and the subsequent predicted failure rate. The above example is just a minimal problem; in practice, there are countless similar comprehensive calculation problems that require an autonomous decision-making system to complete these comprehensive calculation tasks.

[0010] 3. High computing power requirements. The cost of training existing large-model intelligent agents is already very high. To improve performance, it is necessary to increase the input volume or add parameters, which will lead to rising costs. A single machine or a simple multi-machine network cannot meet the computing power requirements.

[0011] (1) Increasing the input of an existing single large model agent will lead to an imbalance between input and output. On the one hand, the model needs to have data fusion capabilities to adapt to the expansion of data sources, and the original network without structural adjustment is difficult to effectively extract related features, resulting in insufficient information utilization of the newly added data; on the other hand, if the input volume and training time are increased to compensate for the inherent defects of the model architecture, the cost required to improve the unit accuracy will increase exponentially.

[0012] (2) Increasing the number of parameters for a single large model agent. Due to the fixed architecture of a single model, it is difficult to increase the number of parameters without retraining. Structural adjustments to the entire network are required to increase the number of parameters. These retraining processes not only generate additional expenses, but also increase resource consumption in storage, deployment, and inference due to model expansion.

[0013] 4. Limited system availability. Within a single, large-scale intelligent agent computing architecture, the successful execution of health assessments and maintenance decisions depends on the agent's own health. If the agent experiences any hardware or software failure, the health of the monitored system loses oversight, leading to operational risks.

[0014] In summary, there is an urgent need for a more optimized large model structure that has high computing power, low resource consumption, high reliability and comprehensive decision-making capabilities to obtain the best evaluation results and maintenance strategies. Summary of the Invention

[0015] To address the above issues, the present disclosure provides a rail transit health status assessment system, method, equipment and medium, which are used to complete comprehensive health status assessment and maintenance decision-making problem solving under limited resource conditions, improve the credibility of decision-making solutions, ensure the safety and reliability of rail transit operations, and at the same time improve operation and maintenance efficiency and reduce operation and maintenance costs.

[0016] In a first aspect, a rail transit health status assessment system includes:

[0017] The terminal layer, distribution layer and centralization layer are connected in sequence;

[0018] The terminal layer includes several independent information extraction agents; the distribution layer includes several distribution layer agents; the centralized layer includes a decision-making agent;

[0019] The information extraction agent is used to obtain raw data input from the terminal, classify and extract the raw data to complete the preprocessing of structured equipment information and unstructured operation and maintenance information to obtain preprocessed data;

[0020] The distributed layer agent uses a deep neural network architecture based on preprocessed data and configured language models and expert models to assess the health status of system equipment. It performs machine learning on the input information and outputs quantitative assessment results of the system health status, system maintenance strategies, expected health status, and parameter packages.

[0021] The decision-making agent is used to coordinate the analysis results of all distribution layer agents based on the quantitative assessment results of the system health status, system maintenance strategy, expected health status and parameter package, and generate the final health status assessment results and final maintenance strategy through logical judgment and priority sorting.

[0022] Furthermore, the information extraction agent includes: a data acquisition module, a protocol analysis module and an information encapsulation module;

[0023] The data acquisition module integrates CAN, Ethernet, and serial multi-format protocol communication units, used to communicate with the system equipment to be evaluated according to the configuration and receive structured equipment logs and warning information; it also integrates a human-computer interaction communication unit to receive unstructured manually input fault information and maintenance information; it adopts dual modes of time-series data stream processing and batch data processing to receive real-time data or batch imported data;

[0024] The protocol parsing module parses structured equipment logs and warning information based on known protocol formats, extracting valid text or values ​​into standardized formatted information. For unstructured manually input fault information and maintenance information, the module performs intelligent semantic analysis on the text based on a pre-set database, extracting fault time and fault symptoms into standardized formatted information.

[0025] The information encapsulation module is used to encapsulate information with a standardized format, add authenticity, integrity and confidentiality coding to obtain encoded information; and transmit the encoded information to the distribution layer intelligent agent through a pre-defined external communication protocol.

[0026] Furthermore, the distribution layer includes several distribution layer agents, specifically including:

[0027] Each distributed layer agent contains a language model and an expert model;

[0028] When j language model agents and k expert model agents are deployed, j×k distribution layer agents are dynamically combined; the j×kth agent B jk The language of uses the jth language model and the calculation strategy uses the kth expert model; among them, the jth language model comes from the training results of the jth language model agent, and the kth expert model comes from the training results of the kth expert model agent;

[0029] The language model agent is used to calculate the optimal strategy under the current language, and the expert model agent is used to calculate the optimal strategy under the current expert role.

[0030] Furthermore, language model agents include: DeepSeek, Wenxin, Kimi, Inseption, GPT-3, BERT and DQN.

[0031] Furthermore, the expert model agent includes the following dimensions: hardware, risk assessment, cost, and maintenance.

[0032] Furthermore, the hardware assessment agent prioritizes hardware status recovery. During training, the agent assigns the highest weight to hardware-related information in the input data. The agent assesses objective defects in the hardware status and develops the optimal maintenance decision plan for restoring hardware functionality. Hardware-related information includes abnormal parameters in device logs, alarm devices, alarm times, alarm phenomena, as well as fault and maintenance information in warning messages.

[0033] The risk assessment agent prioritizes reducing operational and safety risks. During training, operational-related information in the input data is weighted the most heavily. It comprehensively assesses the operational and safety risks posed by equipment status and maintenance methods, providing a more pessimistic health assessment from a risk perspective and formulating robust maintenance decision plans. Operational-related information includes the risk of operational efficiency loss due to prolonged maintenance time and the risk of recurring equipment failures due to insufficient maintenance coverage.

[0034] The cost assessment agent prioritizes reducing the overall financial costs of the rail transit system. During training, the input data for financial cost-related information is weighted the most heavily. This agent prioritizes cost accounting and analysis, providing cost-based health status assessments and formulating maintenance decision plans with the lowest cost. Financial cost-related information includes equipment hardware costs, maintenance personnel labor costs, maintenance tool costs, and production loss costs caused by equipment downtime.

[0035] Furthermore, the distribution layer agent includes the following modules:

[0036] The input communication module is used to receive communication data from the information extraction agent, strip the communication layer protocol, decrypt the valid data, and verify the authenticity and integrity of the data. All data that passes the verification is regarded as valid data input and handed over to the health status assessment module and strategy generation module for calculation. Data that fails the verification is judged as invalid input and notified to the decision-making layer through communication.

[0037] The health status assessment module is used to extract objective information from device logs, warnings, faults, and maintenance information, integrate it by source, and link information from the same device. It also integrates information from different devices. It uses the configured language model and expert model to iteratively train the data to complete the trained health status assessment model for quantitative assessment of system health status.

[0038] The strategy generation module extracts predictive information from device logs, warnings, faults, and maintenance information, integrates it by source, and forms links between information from the same device. It also integrates information from different devices. It iteratively trains data using configured language models and expert models, and generates a large model from the trained strategy. This model is used to calculate system maintenance strategies and the expected health status achieved through maintenance.

[0039] The parameter output module is used to insert an independent supervision function into the health status assessment module and the strategy generation module, collect the weight parameters and threshold parameters in the neural network calculation, and combine them into a parameter package;

[0040] The output communication module is used to encapsulate the quantitative assessment results of the system health status, system maintenance strategy, expected health status and parameter package, add authenticity, integrity and confidentiality coding to obtain coded information; and transmit the coded information to the centralized layer through a pre-defined external communication protocol.

[0041] Furthermore, the decision-making agent includes the following modules:

[0042] The interface module is used to receive communication data from the distribution layer agent, decrypt valid data, and verify the authenticity and integrity of the data. All data that passes the verification, if it is the calculation result of the distribution layer, is regarded as valid data input and handed over to the information fusion module and quality supervision module for calculation. If the distribution layer inputs invalid data alarms to the communication module, it is forwarded to the quality supervision module. Data that fails the verification is judged as invalid input and notified to the quality supervision module.

[0043] The information fusion module is used to perform fusion calculations on the calculation results of the distributed layer intelligent agents. The fusion calculation uses large model technology, and the boundary conditions of the fusion calculation come from the quality supervision module and the human-machine interface module.

[0044] The quality supervision module is used to perform legitimacy checks. If a legitimacy check fails, it will provide feedback to the information fusion module, prompting it to disable illegal input. The legitimacy check is completed using large model technology. The legitimacy check includes the legitimacy of data input to the distribution layer communication module, the legitimacy of data in the centralized layer interface module, and the legitimacy of parameter package data of all distribution layer agents.

[0045] The human-machine interface module is used to realize the interaction between the intelligent agent and humans, and output the calculation results according to the specified boundary conditions. The interaction content includes: receiving manual input commands, textual principles, and textual feedback results after the actual application of strategies; the output calculation results include: health status assessment results, maintenance strategies and health status improvement values.

[0046] Furthermore, the information fusion module performs fusion calculations in the following ways:

[0047] Assume that the distributed layer agent data verified by the interface module uses n different language models and m different expert models to form n×m combinations of language models and expert models, where n and m are positive integers;

[0048] The health status assessment results of the distributed layer agents using the first expert model are judged according to the "two out of three" principle, that is, if there are n results, the closest [n / 2]+1 results are calculated, and [] represents the integer part; the results of the n distributed layer agents using the first expert model are judged to obtain the first primary judgment result;

[0049] The health status assessment results of the n distributed layer agents using the second expert model are judged according to the "three out of two" principle to obtain the second primary judgment result;

[0050] The health status assessment results of the n distributed layer agents using the mth expert model are judged in turn according to the "three out of two" principle to obtain the mth primary judgment result;

[0051] For the m primary assessment results, the “two out of three” principle is used to make the final assessment. The closest [m / 2]+1 results are calculated, where [] represents the integer part, to obtain the final health status assessment result.

[0052] For the maintenance strategy results of the m distributed layer agents using the first language model, the calculation weights are set according to the boundary conditions, and the judgment is made according to the "three out of two" principle to obtain the first primary maintenance strategy with weights and the health status improvement value;

[0053] For the maintenance strategy results of the m distributed layer agents using the second language model, the calculation weights are set according to the boundary conditions, and the judgment is made according to the "three-out-two" principle to obtain the second primary maintenance strategy with weights and the health status improvement value;

[0054] The maintenance strategy results of the m distributed layer agents using the nth language model are calculated in turn according to the boundary conditions, and the "three out of two" principle is used to make a judgment, thus obtaining the nth primary maintenance strategy with weights and the health status improvement value;

[0055] The above n primary maintenance strategies and health status improvement values ​​are judged according to the "three out of two" principle, and the closest [n / 2]+1 results are calculated, where [] represents the integer part, to obtain the final maintenance strategy and health status improvement value.

[0056] In a second aspect, a rail transit health status assessment method includes:

[0057] Obtain raw data input from the terminal, classify and extract the raw data to complete the preprocessing of structured equipment information and unstructured operation and maintenance information to obtain preprocessed data;

[0058] Based on pre-processed data, a deep neural network architecture is used to assess the health status of system equipment according to the configured language model and expert model. Machine learning is performed on the input information to output quantitative assessment results of system health status, system maintenance strategy, expected health status and parameter package;

[0059] Based on the quantitative assessment results of the system health status, system maintenance strategy, expected health status and parameter package, all analysis results are coordinated, and the final health status assessment results and final maintenance strategy are generated through logical judgment and priority sorting.

[0060] Furthermore, raw data input from the terminal is obtained, and the raw data is classified and extracted to complete the preprocessing of structured equipment information and unstructured operation and maintenance information, and obtain preprocessed data, including:

[0061] Adopting CAN, Ethernet and serial port multi-format protocol communication units, it communicates with the system equipment to be evaluated according to the configuration and receives structured equipment logs and warning information; adopting human-computer interaction communication units, it receives unstructured manually input fault information and maintenance information; adopting time series data stream processing and batch data processing dual modes, it receives real-time data or batch imported data;

[0062] For structured equipment logs and warning information, the system parses the received data based on known protocol formats and extracts valid text or values ​​into information with a standardized format. For unstructured manually input fault information and maintenance information, the system performs intelligent semantic analysis on the text based on a pre-set database and extracts fault time and fault symptoms into information with a standardized format.

[0063] The information with a standard format is encapsulated and encoded with authenticity, integrity and confidentiality to obtain encoded information.

[0064] According to a third aspect, an electronic device includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other via the communication bus;

[0065] a memory storing a computer program;

[0066] The processor is used to implement the above-mentioned rail transit health status assessment method when executing the computer program stored in the memory.

[0067] In a fourth aspect, a computer-readable storage medium stores a computer program, which, when executed by a processor, implements the above-mentioned rail transit health status assessment method.

[0068] The present disclosure includes at least the following beneficial effects:

[0069] This disclosure utilizes large-scale model technology to conduct comprehensive health status assessments and maintenance decision-making, addressing the issues of expert scoring systems relying on human experience, failing to accurately evaluate maintenance strategies, and lacking standardized quantitative benchmarks. Through a three-layer, three-dimensional design encompassing terminal, distribution, and centralized layers, and the interweaving of multilingual and multi-expert models at the distribution layer, a three-dimensional vertical decision-making agent is constructed. This simplifies the parameters and input size required by the final decision-making agent, improves computational efficiency, and addresses the issue of high computational power requirements. Vertically, the use of multilingual models fuses multiple language models, addressing the output credibility of a single model. Horizontally, the use of multi-expert models fuses strategy results from different optimization perspectives, addressing algorithmic monotonicity. A centralized, two-out-of-three decision-making scheme and expert weighted decision-making scheme, along with the design of large-scale model boundary conditions, address the issues of uncertainty in the output of a single model and the varying results of each computation, thereby enhancing the authority and credibility of the rail transit system model. The multi-model, three-dimensional structure addresses the issue of system functionality relying on a single model for normal operation and improves the redundancy of the assessment system.

[0070] The present disclosure adopts a three-dimensional intelligent agent architecture design to solve the output credibility problem of a single model by using a multi-language model and integrating multiple language models in the vertical direction. The three-dimensional intelligent agent architecture design also solves the problem of algorithm monotony by using a multi-expert model and integrating the strategy results under different optimization angles in the horizontal direction. The three-dimensional intelligent agent architecture design is used to build a three-dimensional vertical decision-making intelligent agent, simplify the parameters and input scale required for the final decision-making intelligent agent, improve computing efficiency, and solve the problem of high computing power requirements. The three-dimensional intelligent agent architecture design and multi-agent composite computing solve the problem of paralysis of the entire system function after the failure of a single intelligent agent.

[0071] Other features and advantages of the present disclosure will be described in the following description, and in part will become apparent from the description, or will be understood by practicing the present disclosure. The purpose and other advantages of the present disclosure can be achieved and obtained through the structures indicated in the description and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] In order to more clearly illustrate the embodiments of the present disclosure or the technical solutions in the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present disclosure. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0073] Figure 1 This is a schematic diagram of the evaluation system architecture of an embodiment of the present disclosure;

[0074] Figure 2 This is a schematic diagram of the information flow of the evaluation system according to an embodiment of the present disclosure;

[0075] Figure 3 This is a schematic diagram of the structure of an electronic device according to an embodiment of the present disclosure. DETAILED DESCRIPTION

[0076] To make the objectives, technical solutions, and advantages of the embodiments of the present disclosure more clear, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present disclosure, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present disclosure without making any creative efforts shall fall within the scope of protection of the present disclosure.

[0077] like Figure 1 As shown, a rail transit health status assessment system includes:

[0078] The terminal layer, distribution layer and centralization layer are connected in sequence;

[0079] The terminal layer includes several independent information extraction agents; the distribution layer includes several distribution layer agents; the centralized layer includes a decision-making agent;

[0080] The information extraction agent is used to obtain raw data input from the terminal, classify and extract the raw data to complete the preprocessing of structured equipment information and unstructured operation and maintenance information to obtain preprocessed data;

[0081] The distributed layer agent uses a deep neural network architecture based on preprocessed data and configured language models and expert models to assess the health status of system equipment. It performs machine learning on the input information and outputs quantitative assessment results of the system health status, system maintenance strategies, expected health status, and parameter packages.

[0082] The decision-making agent is used to coordinate the analysis results of all distribution layer agents based on the quantitative assessment results of the system health status, system maintenance strategy, expected health status and parameter package, and generate the final health status assessment results and final maintenance strategy through logical judgment and priority sorting.

[0083] The specific implementation is as follows:

[0084] like Figure 2 As shown, the technical solution consists of three layers: terminal layer, distribution layer, and centralized layer. The terminal layer consists of several information extraction agents. The distribution layer consists of several language model agents and several expert model agents. The centralized layer consists of a decision-making agent.

[0085] 1. Terminal layer:

[0086] 1. Information Extraction Agent: This agent directly receives raw data input from terminals, such as device logs and warnings, as well as manually entered fault and maintenance information. It then preprocesses structured device information and unstructured operation and maintenance information. Each agent, operating as an independent service unit, performs simple classification and extraction of raw data.

[0087] 2. Terminal layer structure:

[0088] The terminal layer consists of i independent information extraction agents (1 to i). Usually, one information extraction agent is deployed for each terminal within the evaluation range. When there is communication interaction between terminals and the amount of information is small, multiple terminals can share one information extraction agent.

[0089] The information extraction agent performs small amounts of computation and does not require high computing resources. It is suitable for deployment in existing systems and does not require additional hardware equipment.

[0090] 3. The information extraction agent includes the following core processing modules and processes:

[0091] 1) Data Acquisition Module: This module integrates multi-format protocol communication units such as CAN, Ethernet, and serial ports. It can communicate with the system equipment to be evaluated based on the configuration and receive structured equipment logs and warning information. It also integrates a human-computer interaction communication unit to receive unstructured manually input fault information and maintenance information. It supports both time-series data stream processing and batch data processing modes, and can receive real-time data or batch imported data.

[0092] 2) Protocol Parsing Module: For structured device logs and warning messages, this module parses the received data based on known protocol formats, stripping away the communication layer protocols and extracting valid text or numerical values. For unstructured, manually entered fault and maintenance information, this module performs intelligent semantic analysis based on a pre-built database, stripping away operator information irrelevant to health status assessments and other invalid statements, extracting valid text such as fault time and symptoms. This module extracts raw information from all levels and devices of the assessed system into standardized formats.

[0093] 3) Information Encapsulation Module: This module encapsulates the standardized information and adds authenticity, integrity, and confidentiality codes. The encoded information is then transmitted to the intelligent agents in the distribution layer via a predefined external communication protocol.

[0094] 2. Distribution layer:

[0095] 1. Language Model Agent: The role of the language model agent is to use a deep neural network architecture to perform machine learning on the input information based on the configured language model to form a language model. The goal is to evaluate the health status of system equipment, propose the final maintenance strategy under the current language architecture, and output the neural network calculation parameters and the final health status assessment results and maintenance strategy.

[0096] Language model agents 1-j are learning systems based on j heterogeneous language models, trained and constructed. These include, but are not limited to, DeepSeek, Wenxin, Kimi, Inseption, GPT-3, BERT, and DQN. Due to the differences in model architecture, decoding strategies, and randomness control strategies across large model algorithms, the outputs of different language model agents are not necessarily identical. While shared components are more reliable, differences require further analysis and computation.

[0097] 2. Expert module agent: The role of the expert model agent is to use a deep neural network architecture to perform machine learning on the input information based on the configured expert role to form an expert model. The goal is to evaluate the health status of system equipment, propose the final maintenance strategy under the current expert role, and output the neural network calculation parameters and the final health status assessment results and maintenance strategy.

[0098] Expert model agents 1 to k are a learning system based on k heterogeneous expert models, trained to cover various dimensions including but not limited to hardware, software, risk assessment, security, reliability, availability, cost, and maintenance. The following three agents are used as examples:

[0099] 1) Hardware Assessment Agent: Prioritizing hardware status recovery, the agent assigns the highest weight to hardware-related information in the input data during training, such as abnormal parameters in device logs, alarm devices, alarm times, and alarm phenomena in warning messages, fault information, and repair methods in maintenance information. Using machine learning algorithms, the agent assesses objective defects in the hardware status and develops the optimal maintenance decision plan for restoring hardware functions.

[0100] 2) Risk Assessment Agent: Prioritizing operational and safety risk reduction, this agent prioritizes operational-related information within its training data, such as alarm levels and repair times. It comprehensively assesses operational and safety risks posed by equipment status and maintenance methods, such as the risk of operational efficiency loss due to prolonged repair times or the risk of recurring equipment failures due to insufficient repair coverage. This agent, taking a risk-based approach, generates more pessimistic health assessments and develops more robust maintenance decision-making plans.

[0101] 3) Cost Assessment Agent: Prioritizing reducing the overall financial costs of the rail transit system, the agent prioritizes financial cost information within its training inputs, such as equipment hardware costs, maintenance personnel costs, maintenance tool costs, and production losses caused by equipment downtime. This agent prioritizes cost accounting and analysis, providing cost-based health status assessments and developing cost-minimizing maintenance decisions.

[0102] 3. Distribution layer structure:

[0103] 1) Assuming that there are j language model agents deployed in 1~j and k expert model agents deployed in 1~k, the distribution layer actually has j×k agents. B 11 The language uses the first language model, the calculation strategy uses the first expert model, and the j×kth agent B jk The language of the j-th language model is used, and the calculation strategy is based on the k-th expert model. The j×k agents are independent of each other.

[0104] 2) Each independent distribution layer agent generally adopts a distributed architecture to improve computing speed and reduce hardware resource consumption. In special application scenarios, a centralized architecture can also be used to achieve physical layer information confidentiality and reduce interaction costs.

[0105] 4. Processing flow

[0106] Each distributed layer agent contains the following core processing modules and processes:

[0107] 1) Input Communication Module: Receives communication data from i information extraction agents (i, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 67, 68, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 90, 91, 92, 93, 94, 95, 96, 97, 108, 119, 120, 121, 122, 123, 124, 125, 126, 127, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 1

[0108] 2) Health Assessment Module: This module first extracts objective information from device logs, warnings, faults, and maintenance records (e.g., elapsed operating time, faults encountered, and maintenance operations performed). This information is then integrated by source, linking information from the same device. Information from different devices is then integrated to enhance diversity. After this data processing, the module iteratively trains the data using the configured language model and expert model. The training, validation, and test sets used for calculations are divided differently for each distribution layer agent based on the language model. The trained model is then used to calculate a quantitative system health assessment score.

[0109] 3) Strategy Generation Module: This module first extracts predictive information from device logs, warnings, faults, and maintenance information (e.g., required operating time and required maintenance specifications). This information is then integrated by source, linking information from the same device. Information from different devices is then integrated to enhance diversity. After this data processing, the health assessment module's calculation results are added. The data is iteratively trained using the configured language model and expert model. The training, validation, and test sets used for the calculations are divided differently for each distributed layer agent based on the language model. The trained large model is then used to calculate system maintenance strategies and the health improvements achieved through maintenance.

[0110] 4) Parameter output module: Insert an independent supervision function into the health status assessment module and the strategy generation module to collect and combine all key parameters such as weight parameters and threshold parameters in the neural network calculation into a parameter package without interfering with the calculation, and output it to the outside.

[0111] 5) Output Communication Module: This module encapsulates the calculated system health status quantitative assessment score, system maintenance strategy, expected health status after maintenance, and parameter package, adding authenticity, integrity, and confidentiality codes. This encoded information is transmitted to the centralized layer via a predefined external communication protocol.

[0112] 3. Centralized layer:

[0113] 1. Decision-making agent: Responsible for coordinating the analysis results of all distributed layer agents, generating the final health status assessment results and final maintenance strategy based on the system health status quantitative assessment score, system maintenance strategy, health status improvement, and parameter package through logical judgment and priority sorting.

[0114] 2. Centralized layer structure:

[0115] The centralized layer consists of 1 decision-making agent.

[0116] The decision-making agent uses different architectures depending on the number of distributed layer agents. If the number is small, a centralized architecture is used to ensure information confidentiality at the physical layer and reduce interaction costs. If the number is large, a distributed architecture is adopted to increase computing speed and reduce hardware resource consumption.

[0117] 3. Processing flow:

[0118] The decision-making agent includes the following core processing modules and processes:

[0119] 1) Interface Module: Receives communication data from j×k distributed layer agents (1 to j×k), strips the communication layer protocol, decrypts valid data, and verifies the authenticity and integrity of the data. Data that passes all verification and is the result of distribution layer calculations is considered valid data input and transferred to the information fusion module and quality supervision module for further calculations. Invalid data from the distribution layer input communication module is reported to the quality supervision module. Data that fails verification is considered invalid input and notified to the quality supervision module.

[0120] 2) Information Fusion Module: This module fuses the computational results of the j×k distributed layer agents according to the following principles. The fusion calculation uses large model technology, and the boundary conditions come from the quality supervision module and the human-machine interface module:

[0121] Assume that the distributed layer agent data verified by the interface module uses n (n≤j) different language models and m (m≤k) different expert models.

[0122] a) The health status assessment results of the n distributed layer agents using the first expert model are judged according to the "three-to-two" principle, that is, if there are three results, the two closest results are calculated; if there are four results, the three closest results are calculated; if there are five results, the three closest results are calculated; if there are n results, the closest [n / 2]+1 results are calculated, where [] represents the integer part; thus, the first primary judgment result is obtained after judgment using the results of the n distributed layer agents using the first expert model;

[0123] The health status assessment results of the n distributed layer agents using the second expert model are judged according to the "three out of two" principle to obtain the second primary judgment result;

[0124] The health status assessment results of the n distributed layer agents using the mth expert model are judged in turn according to the "three out of two" principle to obtain the mth primary judgment result;

[0125] According to the above principles, m (m ≤ k) primary assessment results are obtained. For the above m primary assessment results, the "two out of three" principle is used to make assessments, and the closest [m / 2] + 1 results are calculated, where [] represents the integer part. This step obtains the final health status assessment result.

[0126] b) For the maintenance strategy results of the m distributed layer agents using the first language model, calculate the weights according to the boundary conditions, make a judgment according to the "three out of two" principle, and obtain the first weighted maintenance strategy and health status improvement value;

[0127] For the maintenance strategy results of the m distributed layer agents using the second language model, the calculation weights are set according to the boundary conditions, and the judgment is made according to the "three out of two" principle to obtain the second maintenance strategy with weights and the health status improvement value;

[0128] The maintenance strategy results of the m distributed layer agents using the nth language model are calculated in turn, with weights set according to the boundary conditions, and judged according to the "three out of two" principle to obtain the nth maintenance strategy with weights and health status improvement value;

[0129] According to the above principles, n (n≤j) maintenance strategy results are obtained. For the above n evaluation results, the "three out of two" principle is used to determine the closest [n / 2]+1 results, where [] represents the integer part. This step obtains the final maintenance strategy and health status improvement value.

[0130] 3) Quality Supervision Module: This module performs a validity check on the following content. If the validity check fails, it provides feedback to the information fusion module, prompting it to disable the illegal input. The validity check is completed using large model technology.

[0131] a) The validity of data received from the distribution layer input communication module.

[0132] b) The legitimacy of the data received from the centralized layer interface module.

[0133] c) The validity of parameter package data of all distributed layer agents.

[0134] 4) Human-computer interface module: This module implements the interaction between the agent and the human, receives the boundary conditions specified by the human, and outputs the calculation results. The main interaction contents include:

[0135] a) Receive manual input commands.

[0136] b) Accept the principle of textualization.

[0137] c) Receive textual feedback values ​​after actual application.

[0138] d) Output the final calculation results, including health status assessment results, maintenance strategies, and health status improvement values.

[0139] A rail transit health status assessment method, comprising:

[0140] Obtain raw data input from the terminal, classify and extract the raw data to complete the preprocessing of structured equipment information and unstructured operation and maintenance information to obtain preprocessed data;

[0141] Based on pre-processed data, a deep neural network architecture is used to assess the health status of system equipment according to the configured language model and expert model. Machine learning is performed on the input information to output quantitative assessment results of system health status, system maintenance strategy, expected health status and parameter package;

[0142] Based on the quantitative assessment results of the system health status, system maintenance strategy, expected health status and parameter package, the analysis results of all distributed layer intelligent agents are coordinated, and the final health status assessment results and final maintenance strategy are generated through logical judgment and priority sorting.

[0143] like Figure 3 As shown, the present disclosure provides an electronic device, including a processor 301, a communication interface 302, a memory 303 and a communication bus 304, wherein the processor 301, the communication interface 302 and the memory 303 communicate with each other through the communication bus 304;

[0144] Memory 303, storing computer programs;

[0145] The processor 301 is configured to implement the above method when executing the computer program stored in the memory 303 .

[0146] The present disclosure provides a computer-readable storage medium storing a computer program, which implements the above method when executed by a processor.

[0147] The computer-readable storage medium may be included in the device / apparatus described in the above embodiments, or may exist independently without being incorporated into the device / apparatus. The computer-readable storage medium carries one or more programs, which, when executed, implement the method according to the embodiments of the present disclosure.

[0148] According to embodiments of the present disclosure, a computer-readable storage medium may be a non-volatile computer-readable storage medium, such as, but not limited to, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In the present disclosure, a computer-readable storage medium may be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0149] Although the present disclosure has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present disclosure.

Claims

1. A rail transit health status assessment system, characterized in that: include: The terminal layer, distribution layer and centralization layer are connected in sequence; The terminal layer includes several independent information extraction agents; the distribution layer includes several distribution layer agents; the centralized layer includes a decision-making agent; The information extraction agent is used to obtain raw data input from the terminal, classify and extract the raw data to complete the preprocessing of structured equipment information and unstructured operation and maintenance information to obtain preprocessed data; The distributed layer agent uses a deep neural network architecture based on preprocessed data and configured language models and expert models to assess the health status of system equipment. It performs machine learning on the input information and outputs quantitative assessment results of the system health status, system maintenance strategies, expected health status, and parameter packages. The decision-making agent, based on the quantitative assessment results of the system health status, the system maintenance strategy, the expected health status, and the parameter package, coordinates the analysis results of all distributed layer agents and generates the final health status assessment results and the final maintenance strategy through logical judgment and priority sorting; The distribution layer includes several distribution layer agents, including: Each distributed layer agent contains a language model and an expert model; The language model agent is used to calculate the optimal strategy under the current language, and the expert model agent is used to calculate the optimal strategy under the current expert role; The decision-making agent includes an information fusion module, which is used to perform fusion calculations on the calculation results of the distribution layer agent. The fusion calculation uses large model technology, and the boundary conditions of the fusion calculation come from the quality supervision module and the human-machine interface module. The information fusion module performs fusion calculations in the following ways: Assume that the distributed layer agent data verified by the interface module uses n different language models and m different expert models to form n×m combinations of language models and expert models, where n and m are positive integers; The health status assessment results of the n distributed layer agents using the mth expert model are judged in turn according to the "three out of two" principle to obtain the mth primary judgment result; For the m primary judgment results, the "three out of two" principle is used to judge and obtain the final health status assessment result; The maintenance strategy results of the m distributed layer agents using the nth language model are calculated based on the boundary conditions and weights are set. The "two out of three" principle is used to determine the nth primary maintenance strategy and health status improvement value with weights. The above n primary maintenance strategies and health status improvement values ​​are judged according to the "three out of two" principle to obtain the final maintenance strategy and health status improvement value.

2. A rail transit health status assessment system according to claim 1, characterized in that: Information extraction agent, including: data acquisition module, protocol analysis module and information encapsulation module; The data acquisition module integrates CAN, Ethernet, and serial multi-format protocol communication units, used to communicate with the system equipment to be evaluated according to the configuration and receive structured equipment logs and warning information; it also integrates a human-computer interaction communication unit to receive unstructured manually input fault information and maintenance information; it adopts dual modes of time-series data stream processing and batch data processing to receive real-time data or batch imported data; The protocol parsing module parses structured equipment logs and warning information based on known protocol formats, extracting valid text or values ​​into standardized formatted information. For unstructured manually input fault information and maintenance information, the module performs intelligent semantic analysis on the text based on a pre-set database, extracting fault time and fault symptoms into standardized formatted information. The information encapsulation module is used to encapsulate information with a standardized format, add authenticity, integrity and confidentiality coding to obtain encoded information; and transmit the encoded information to the distribution layer intelligent agent through a pre-defined external communication protocol.

3. A rail transit health status assessment system according to claim 1, characterized in that: The distribution layer includes several distribution layer agents, including: When j language model agents and k expert model agents are deployed, j×k distribution layer agents are dynamically combined; the j×kth agent B jk The language uses the j-th language model and the calculation strategy uses the k-th expert model; among them, the j-th language model comes from the training results of the j-th language model agent, and the k-th expert model comes from the training results of the k-th expert model agent.

4. A rail transit health status assessment system according to claim 1, characterized in that: Language model agents, including: DeepSeek, Wenxin, Kimi, Inseption, GPT-3, BERT, and DQN.

5. A rail transit health status assessment system according to claim 1, characterized in that: Expert model agent, including the following dimensions: hardware, risk assessment, cost, and maintenance.

6. A rail transit health status assessment system according to claim 5, characterized in that: The hardware assessment agent prioritizes hardware status recovery. During training, hardware-related information in the input data is given the highest weight. This agent is used to assess objective defects in the hardware status and develop the optimal maintenance decision plan to restore hardware functionality. Hardware-related information includes abnormal parameters in device logs, alarm devices, alarm times, alarm phenomena, fault information, and maintenance information. The risk assessment agent prioritizes reducing operational and safety risks. During training, operational-related information in the input data is weighted the most heavily. This agent is used to comprehensively assess operational and safety risks posed by equipment status and maintenance methods, and to develop robust maintenance decision plans. This operational-related information includes the risk of operational efficiency loss due to prolonged maintenance time and the risk of recurring equipment failures due to insufficient maintenance coverage. The cost assessment agent prioritizes reducing the overall financial cost of the rail transit system. During training, the financial cost-related information in the input data is given the highest weight. It is used to provide health status assessment results based on cost and formulate maintenance decision plans with the lowest cost. Financial cost-related information includes: equipment hardware costs, maintenance personnel labor costs, maintenance tool costs, and production loss costs caused by equipment downtime.

7. A rail transit health status assessment system according to claim 1, characterized in that: The distribution layer agent includes the following modules: The input communication module is used to receive communication data from the information extraction agent, strip the communication layer protocol, decrypt the valid data, and verify the authenticity and integrity of the data. All data that passes the verification is regarded as valid data input and handed over to the health status assessment module and strategy generation module for calculation. Data that fails the verification is judged as invalid input and notified to the decision-making layer through communication. The health status assessment module is used to extract objective information from equipment logs, warning information, fault information, and maintenance information, integrate them according to the source, and form a link relationship between the information of the same equipment; Fuse information from different devices; iteratively train the data using the configured language model and expert model to complete the trained health status assessment model for quantitative assessment of system health status; The strategy generation module is used to extract prediction information that has not yet occurred from equipment logs, warning information, fault information, and maintenance information, and integrate them according to the source to form a link relationship between the information of the same equipment; Fuse information from different devices; iteratively train the data using the configured language model and expert model, and generate a large model based on the trained strategy to calculate the system maintenance strategy and the expected health status achieved through maintenance; The parameter output module is used to insert an independent supervision function into the health status assessment module and the strategy generation module, collect the weight parameters and threshold parameters in the neural network calculation, and combine them into a parameter package; The output communication module is used to encapsulate the quantitative assessment results of the system health status, system maintenance strategy, expected health status and parameter package, add authenticity, integrity and confidentiality coding, and obtain coded information; The encoded information is transmitted to the centralized layer through a predefined external communication protocol.

8. A rail transit health status assessment system according to claim 1, characterized in that: The decision-making agent also contains the following modules: The interface module is used to receive communication data from the distribution layer agent, decrypt valid data, and verify the authenticity and integrity of the data. If all verified data is the calculation result of the distribution layer, it is regarded as valid data input and handed over to the information fusion module and quality supervision module for calculation. If the distribution layer inputs invalid data to the communication module, an alarm is issued and forwarded to the quality supervision module. Data that fails verification is judged as invalid input and notified to the quality supervision module; The quality supervision module is used to perform legality checks. If the legality check fails, it will be fed back to the information fusion module. The legality check is completed using large model technology. The legality check content includes: the legality of the data input to the communication module of the distribution layer, the legality of the data of the interface module of the centralized layer, and the legality of the parameter package data of all distributed layer agents. The human-machine interface module is used to realize the interaction between the intelligent agent and humans, and output the calculation results according to the specified boundary conditions. The interaction content includes: receiving manual input commands, textual principles, and textual feedback results after the actual application of strategies; the output calculation results include: health status assessment results, maintenance strategies and health status improvement values.

9. A rail transit health status assessment method, characterized in that: include: Through the information extraction agent, the original data input from the terminal is obtained, and the original data is classified and extracted to complete the preprocessing of structured equipment information and unstructured operation and maintenance information to obtain preprocessed data; Through the distributed layer agent, based on pre-processed data, a deep neural network architecture is used to evaluate the health status of system equipment according to the configured language model and expert model. Machine learning is performed on the input information to output quantitative assessment results of the system health status, system maintenance strategy, expected health status and parameter package; Through the decision-making intelligent agent, based on the quantitative assessment results of the system health status, system maintenance strategy, expected health status and parameter package, all analysis results are coordinated, and the final health status assessment results and final maintenance strategy are generated through logical judgment and priority sorting; The terminal layer includes several independent information extraction agents; the distribution layer includes several distribution layer agents; the centralized layer includes a decision-making agent; The terminal layer, distribution layer, and centralization layer are connected in sequence; The distribution layer includes several distribution layer agents, including: Each distributed layer agent contains a language model and an expert model; The language model agent is used to calculate the optimal strategy under the current language, and the expert model agent is used to calculate the optimal strategy under the current expert role; The decision-making agent includes an information fusion module, which is used to perform fusion calculations on the calculation results of the distribution layer agent. The fusion calculation uses large model technology, and the boundary conditions of the fusion calculation come from the quality supervision module and the human-machine interface module. The information fusion module performs fusion calculations in the following ways: Assume that the distributed layer agent data verified by the interface module uses n different language models and m different expert models to form n×m combinations of language models and expert models, where n and m are positive integers; The health status assessment results of the n distributed layer agents using the mth expert model are judged in turn according to the "three out of two" principle to obtain the mth primary judgment result; For the m primary judgment results, the "three out of two" principle is used to judge and obtain the final health status assessment result; The maintenance strategy results of the m distributed layer agents using the nth language model are calculated based on the boundary conditions and weights are set. The "two out of three" principle is used to determine the nth primary maintenance strategy and health status improvement value with weights. The above n primary maintenance strategies and health status improvement values ​​are judged according to the "three out of two" principle to obtain the final maintenance strategy and health status improvement value.

10. A rail transit health status assessment method according to claim 9, characterized in that: Obtain raw data input from the terminal, classify and extract the raw data to complete the preprocessing of structured equipment information and unstructured operation and maintenance information, and obtain preprocessed data, including: Adopting CAN, Ethernet and serial port multi-format protocol communication units, it communicates with the system equipment to be evaluated according to the configuration and receives structured equipment logs and warning information; adopting human-computer interaction communication units, it receives unstructured manually input fault information and maintenance information; adopting time series data stream processing and batch data processing dual modes, it receives real-time data or batch imported data; For structured equipment logs and warning information, the system parses the received data based on known protocol formats and extracts valid text or values ​​into information with a standardized format. For unstructured manually input fault information and maintenance information, the system performs intelligent semantic analysis on the text based on a pre-set database and extracts fault time and fault symptoms into information with a standardized format. The information with a standard format is encapsulated and encoded with authenticity, integrity and confidentiality to obtain encoded information.

11. An electronic device, characterized in that: The processor, the communication interface, the memory and the communication bus are connected to each other via the communication bus. a memory storing a computer program; The processor is used to implement a rail transit health status assessment method as described in claim 9 or 10 when executing a computer program stored in the memory.

12. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, a rail transit health status assessment method according to claim 9 or 10 is implemented.

Images