Traceable self-evolution multi-agent high-speed train axle box bearing fault diagnosis method

Abstract

The invention discloses a traceable self-evolution multi-agent high-speed train axle box bearing fault diagnosis method, which comprises the following steps of: firstly, acquiring a bearing working condition parameter and a vibration signal, processing to generate an envelope spectrum and characteristic data, and packaging the envelope spectrum and the characteristic data; five types of agents are initialized, data are distributed by a coordinator agent, and each agent outputs a standardized evidence unit containing information such as a unique identifier and a reasoning track under the constraint of a cue word; the coordinator constructs an evidence graph, processes the confidence coefficient of the fault component through a weighted average or specific fusion strategy, and inquires a contradictory conclusion to obtain supplementary reasoning; generating a hierarchical structured detection report based on the multi-source information; and receiving engineer recheck information, updating the case knowledge base, and dynamically optimizing cue word and agent weight. According to the method, multi-source information is fused through multi-agent cooperation, so that diagnosis conclusion traceability and system self-evolution are realized, and the accuracy, adaptability and reliability of fault diagnosis under complex working conditions are improved.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology for high-speed train axle box bearings, and in particular to a multi-agent fault diagnosis method for high-speed train axle box bearings that is traceable to evolution. Background Technology

[0002] As a core transmission component, the operating status of axle box bearings in high-speed trains directly affects train safety and transportation efficiency. Under complex operating environments characterized by long-term high loads, variable conditions, and strong vibrations, bearings are prone to faults such as raceway wear, roller spalling, and cage damage. Failure to identify these faults promptly and accurately can lead to serious accidents such as bearing seizure and shaft breakage, resulting in significant economic losses and safety risks. With the increasing capacity and mileage of high-speed trains, bearing failure modes are becoming more diverse and complex. Traditional methods relying on manual inspections and single-sensor monitoring are no longer sufficient to meet the demands for real-time and accurate diagnosis. There is an urgent need to build an efficient and intelligent fault diagnosis system to achieve early warning, precise location, and source analysis of faults, providing scientific support for equipment maintenance decisions and ensuring the stable and reliable operation of high-speed transportation systems.

[0003] Existing technologies for fault diagnosis of axle box bearings in high-speed trains have two significant shortcomings: First, diagnostic methods often rely on single signal analysis models or fixed algorithms, lacking the ability to collaboratively integrate multi-source information. This makes it difficult to comprehensively capture the multi-dimensional characteristics of complex faults, resulting in poor adaptability to different operating conditions and fault types. Diagnostic biases are easily caused by signal interference and insufficient feature extraction. Furthermore, the reasoning process lacks traceability, making it impossible to clearly define the basis and logical chain for conclusion generation, which is detrimental to fault cause review and diagnostic model optimization. Second, existing diagnostic systems are mostly static architectures, lacking self-evolution capabilities. They cannot dynamically adjust diagnostic strategies based on new fault cases and expert experience feedback. As bearing operating conditions change and fault modes evolve, diagnostic accuracy tends to gradually decrease. At the same time, there is a lack of effective coordination mechanisms among various diagnostic models or algorithms, making it difficult to efficiently integrate multi-source reasoning results when faced with conflicting conclusions. This leads to low diagnostic efficiency and weak ability to handle uncertain conclusions. Summary of the Invention

[0004] In order to overcome the shortcomings and deficiencies of existing technologies, this invention provides a fault diagnosis method for high-speed train axle box bearings that is traceable to evolution and multi-agent systems.

[0005] The technical solution adopted in this invention is a traceable, evolution-derived multi-agent high-speed train axle box bearing fault diagnosis method, comprising the following steps: S1, collecting bearing model, operating speed, and sampling frequency related parameters; determining the optimal center frequency based on historical data and sending it to the data collection device; acquiring the original vibration signal through a timed trigger or vibration RMS mutation trigger mechanism; filtering and transforming the original vibration signal to generate an envelope spectrum; extracting significant frequency components and time-domain statistical features from the envelope spectrum; generating a unique diagnostic session ID as the root node; and encapsulating the envelope spectrum image, peak frequency list, and time-domain features for subsequent diagnosis; S2, initializing five types of agents: spectrum analysis, time-domain analysis, theoretical analysis, case matching, and coordinator. Each agent consists of a large language model, dedicated prompt words, and domain tools. Each agent generates structured and traceable preliminary reasoning results based on the prompt words and input data; S3, the coordinator agent receives and parses the diagnostic request packet, and separates the envelope spectrum image, time-domain features, bearing model, speed, peak frequency, and operating parameters into... The system sends data to the corresponding agents, each of which outputs standardized evidence units under the constraints of the prompt words, including a unique evidence node ID, observation description, action sequence, input-output pairs, knowledge references, intermediate results, and fault confidence. In S4, the coordinator agent collects all preliminary inference results and constructs an evidence graph, using measured peak values, theoretical frequencies, and case labels as nodes. Support or contradictory edges are established based on matching degree and conflict relationships. Confidence fusion is performed on each standard fault component using weighted averaging or evidence theory. Uncertain conclusions are marked, and contradictory agents are questioned to obtain supplementary inference. In S5, the coordinator agent generates a structured inspection report based on the evidence graph, questioning records, and results from each stage, including a summary layer, evidence layer, inference chain layer, and maintenance suggestion layer. In S6, the system receives confirmation information from engineers regarding the review conclusions, writes the confirmed or corrected diagnostic samples into the case knowledge base and records expert preferences, adds the inference log summaries of high-confidence, uncorrected samples to the example pool, dynamically concatenates them to the prompt words, and dynamically updates the weights of agents that perform poorly in identifying fault types.

[0006] Furthermore, the final confidence level of the faulty component without divergence in S4 is calculated using the following formula: ; in This represents the final confidence level of the faulty component C. These are the weight coefficients for the four types of agents. The confidence scores of fault components C output by the agent in spectral analysis, time-domain analysis, theoretical analysis, and case matching are respectively given. The theoretical analysis agent's successful matching is indicated by... The value is 1.0, and 0 if no match is found.

[0007] Furthermore, the divergent fault components in S4 are fused using Dempster-Shafer evidence theory, and the binary identification framework is defined as follows: ,in The identification framework for faulty component C is represented. Indicates a fault-free state. The formula for converting confidence levels to the BPA function to indicate a faulty state is as follows: ,in This represents the BPA function. This represents the confidence level of the faulty component C output by the corresponding agent. Theoretically, this indicates when the agent successfully matches the faulty component. The value is 1.0, and 0 if no match is found.

[0008] Furthermore, the Dempster synthesis rule in S4 is applied using the following formula: ; ; in Indicates the conflict coefficient. Let BPA functions represent the functions of the two agents respectively. This represents the synthesized BPA function value. Indicates a fault-free state. This indicates a faulty state. The fusion process is terminated and the corresponding conclusion is marked.

[0009] Furthermore, the expression for the agent to generate the preliminary structured reasoning result in S2 is: ,in This represents the initial structured reasoning result of the i-th type of intelligent agent. Denotes the i-th type of intelligent agent. This indicates a cue word specific to the i-th type of intelligent agent. This represents the data set input to the i-th type of intelligent agent, where the value of i corresponds to one of the five types of intelligent agents: spectrum analysis, time domain analysis, theoretical analysis, case matching, and coordinator.

[0010] Furthermore, the formula for generating the structured inspection report in S5 is as follows: ,in This indicates the final structured testing report. Represents the coordinator agent. These represent the structured reasoning results of the agent, which are derived from spectral analysis, temporal analysis, theoretical analysis, and case matching, respectively. This represents the constructed multi-source collaborative diagnostic evidence map.

[0011] Further, S3 includes the following sub-steps: S31, the coordinator agent receives the diagnostic request packet, analyzes and classifies the envelope spectrum image, time-domain features, bearing model, operating speed, peak frequency, and operating parameters, and clarifies the distribution objects corresponding to each type of data; S32, according to the preset data distribution rules, the envelope spectrum image and peak frequency data are sent to the spectrum analysis agent, the time-domain feature data is sent to the time-domain analysis agent, the bearing model and speed data are sent to the theoretical analysis agent, and the peak frequency, bearing model, and operating parameters are sent to the case matching agent; S33, after receiving the corresponding data, each agent performs targeted analysis and processing based on the reasoning logic and constraints set by the exclusive prompt words, forming a standardized evidence unit including a complete reasoning trajectory; S34, each agent embeds a unique evidence node ID, observation description, action sequence, input-output pair, knowledge reference, intermediate results, and fault confidence information into the standardized evidence unit and outputs it.

[0012] Further, S4 includes the following sub-steps: S41, the coordinator agent summarizes the preliminary reasoning results output by all agents, constructs evidence nodes with measured peak values, theoretical frequencies, and case labels as the core, establishes support edges based on data matching degree judgment, establishes contradictory edges based on conclusion conflict relationships, and forms an evidence graph with multi-source evidence structured association; S42, performs divergence judgment on each standard fault component, filters out fault components without divergence, and uses preset weight coefficients to weight the confidence of the output of various agents to obtain the final confidence of such fault components; S43, for fault components with divergence, initiates the evidence theory fusion process, converts the confidence of various agents into BPA functions, and gradually merges the BPA functions of all agents through synthesis rules to obtain the final fusion result; S44, judges the state of fault components based on the final confidence value range, marks the conclusions in the uncertainty interval and sends the evidence graph to the contradictory agent, and receives the supplementary reasoning results after processing by the agent.

[0013] Further, S5 includes the following sub-steps: S51, the coordinator agent extracts node association information, supporting and contradictory edge distribution from the evidence graph, organizes the preliminary reasoning results, questioning feedback and supplementary reasoning content of each agent, and forms a complete diagnostic data set; S52, based on the diagnostic data set, a summary layer is constructed to extract the core diagnostic conclusions, fault component judgment results and confidence distribution; S53, an evidence layer is constructed to integrate envelope spectrum images, peak frequencies, time-domain features, theoretical frequency calculation results and historical case matching information to label the evidence; S54, a reasoning chain layer is constructed to sort out the complete logical path from the original data input to the final conclusion output, clarify the data flow and reasoning basis at each stage, form a maintenance suggestion layer based on maintenance experience, and integrate information from each level to generate a structured test report.

[0014] A traceable, evolution-derived multi-agent high-speed train axle box bearing fault diagnosis method is implemented through different units, including: a multi-dimensional parameter and signal acquisition unit, used to acquire bearing model, operating speed, sampling frequency, operating condition parameters, and raw vibration signals; filtering and transforming the vibration signals to generate envelope spectra; extracting significant frequency components and time-domain statistical features and encapsulating them; an agent initialization and configuration unit, used to initialize five types of agents: spectrum analysis, time-domain analysis, theoretical analysis, case matching, and coordinator; configuring dedicated large language models, prompt words, and domain tools for each agent; and defining the agent inference output format; and a data distribution and inference execution unit, used to receive and parse diagnostic request packets; distribute various types of data to the corresponding agents according to preset rules; and drive each agent to operate within prompt word constraints. The system comprises several modules: an independent analysis and standardized evidence unit; a multi-source evidence fusion and graph construction unit, which collects reasoning results from various agents, constructs an evidence graph including supporting and contradictory edges, fuses the confidence levels of faulty components using weighted averaging or evidence theory, handles divergent conclusions, and questions contradictory agents; a structured report generation unit, which integrates the evidence graph, questioning records, and reasoning results from each stage, and generates a diagnostic report based on a structure of summary layer, evidence layer, reasoning chain layer, and maintenance suggestion layer; and a self-evolving closed-loop update unit, which receives engineer review and correction information, updates the case knowledge base and expert preference records, dynamically adjusts prompts and agent weights, and adaptively evolves the system's diagnostic capabilities. All units interact and collaborate via a data bus to ensure the orderly progress of the diagnostic process.

[0015] Beneficial Effects: This invention proposes a traceable, evolutionarily-based multi-agent method for diagnosing high-speed train axle box bearing faults. Through the collaborative efforts of five types of agents, it comprehensively integrates multi-source information such as spectrum, time domain, theory, and case studies. By constructing an evidence graph to establish data associations and conflict identification, and employing a targeted confidence fusion strategy to handle divergent conclusions, it solves the problems of poor adaptability of single models and insufficient fusion of multi-source information in traditional methods, significantly improving the accuracy of fault diagnosis under complex operating conditions. By embedding unique identifiers and complete trajectory records during the reasoning process, a hierarchical structured detection report is formed, enabling full traceability of diagnostic conclusions. This clarifies the logical chain and evidence supporting the conclusion generation, overcoming the shortcomings of opaque reasoning processes and difficulty in retrospection in existing technologies. Meanwhile, through a closed-loop mechanism of case knowledge base updates, dynamic optimization of prompt words, and adjustment of agent weights, the system can continuously absorb expert experience and new cases, adaptively adapt to the evolution of fault modes and changes in operating conditions, and significantly improve the diagnostic stability and adaptability during long-term operation. The distribution, questioning, and fusion functions of the coordinator agents efficiently integrate the reasoning results of various agents, strengthen the ability to handle divergent conclusions, and effectively improve diagnostic efficiency and reliability, providing strong support for the early identification, accurate location, and scientific maintenance of axle box bearing faults in high-speed trains. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the overall process of the method of the present invention. Figure 2 This is a flowchart of method step S3 of the present invention; Figure 3 This is a flowchart of method step S4 of the present invention; Figure 4 This is a flowchart of step S5 of the method of the present invention; Figure 5 This is a diagram showing the unit composition for implementing the method of the present invention. Detailed Implementation

[0017] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] like Figure 1 As shown, the multi-agent fault diagnosis method for axle box bearings in high-speed trains, traceable to evolution, includes the following steps: S1 collects bearing model, operating speed and sampling frequency related parameters, determines the optimal center frequency based on historical data and sends it to the data collection device, collects the original vibration signal through timed trigger or vibration RMS mutation trigger mechanism, filters and transforms the original vibration signal to generate envelope spectrum, extracts significant frequency components and time domain statistical features in the envelope spectrum, generates a unique diagnostic session ID as the root node, and encapsulates the envelope spectrum image, peak frequency list and time domain features for subsequent diagnosis. Specifically, step S1 is the data foundation construction stage of the diagnostic process, acquiring standardized, high-value raw data and feature information to provide reliable input for subsequent multi-agent inference. During implementation, the system parameter interaction module first collects key parameters such as the specific bearing model, real-time operating speed (range 500 rpm to 3000 rpm), and preset sampling frequency (selectable from 10kHz, 20kHz, and 50kHz). Based on over 100,000 diagnostic records of the same bearing model under different operating conditions in the historical diagnostic database, statistical analysis is used to determine the optimal center frequency (range 1kHz to 10kHz, dynamically adjusted according to the rule of increasing the center frequency by 500Hz for every 10mm increase in bearing diameter). This parameter is then sent to the triaxial accelerometer installed in the bearing housing of the axle box. Data acquisition employs a dual-mode approach: timed triggering and vibration RMS mutation triggering. The timed trigger interval can be set autonomously between 5 and 60 seconds, while the vibration RMS mutation trigger threshold is set to 1.5 to 3 times the average RMS value under normal operating conditions. When either trigger condition is met, the sensor acquires raw vibration signals for 2 to 10 seconds at a preset sampling frequency. After acquisition, high-frequency interference signals are filtered out using a low-pass filter with a cutoff frequency of 500Hz. Then, an envelope spectrum is generated using Hilbert transform. A peak detection algorithm is used to extract the top 20 significant frequency components in amplitude from the envelope spectrum. Simultaneously, six time-domain statistical features, including peak value, peak-to-peak value, mean, variance, kurtosis, and skewness, are calculated. The system automatically generates a unique diagnostic session ID composed of 32 random characters as the data root node. The envelope spectrum image (1024×768 pixels resolution), the peak frequency list (arranged in descending order of amplitude), and the time-domain feature values ​​are encapsulated into a standard diagnostic request package, ensuring a unified and traceable data structure and providing complete and standardized basic data support for subsequent agent analysis.

[0019] S2 initializes five types of intelligent agents: spectrum analysis, time domain analysis, theoretical analysis, case matching, and coordinator. Each type of intelligent agent consists of a large language model, dedicated prompt words, and domain tools. Each type of intelligent agent generates structured and traceable preliminary reasoning results based on prompt words and input data. Specifically, step S2 is the initialization and construction phase of the multi-agent diagnostic system. Its purpose is to build a well-defined, highly efficient intelligent diagnostic core architecture to ensure the accurate implementation of various diagnostic tasks. During implementation, five types of agents are initialized according to their functional positioning: spectrum analysis, time-domain analysis, theoretical analysis, case matching, and coordinator. Each type of agent consists of a basic large language model, dedicated prompt words, and domain tool modules. The basic large language model adopts a Transformer architecture model with a parameter scale of 1 billion to 10 billion, possessing powerful natural language understanding and logical reasoning capabilities. The spectrum analysis agent's dedicated prompts include core instructions such as envelope spectrum feature recognition, frequency component matching, and fault feature association, and is equipped with domain tools including spectrum peak extraction tools, frequency harmonic analysis tools, and spectrum similarity comparison tools. The time-domain analysis agent's prompts focus on time-domain statistical feature calculation, feature anomaly judgment, and fault degree quantification, and is equipped with time-domain feature filtering tools, anomaly threshold judgment tools, and feature trend analysis tools. The theoretical analysis agent's prompts include logic such as bearing structural parameter analysis, theoretical fault frequency calculation, and frequency matching verification, and is equipped with theoretical frequency calculation tools, parameter conversion tools, and matching degree judgment tools. The case matching agent's prompts include instructions such as case feature retrieval, similarity calculation, and conclusion mapping, and is equipped with case database retrieval tools, multi-dimensional similarity analysis tools, and conclusion adaptation tools. The coordinator agent's prompts focus on collaborative functions such as data distribution, result integration, conflict handling, and report generation, and is equipped with data parsing tools, evidence graph construction tools, confidence fusion tools, and questioning interaction tools. After the various intelligent agents are initialized, they connect with the data transmission module and storage module through the system's preset interface protocol to ensure that they can quickly receive input data. Based on the logical constraints of the exclusive prompt words and the functional support of domain tools, they generate preliminary reasoning results that include fixed fields, clear logic, and traceability, laying the foundation for subsequent multi-source evidence fusion.

[0020] S3, the coordinator agent receives and parses the diagnostic request packet, and distributes the envelope spectrum image, time domain features, bearing model, speed, peak frequency and operating parameters to the corresponding agents. Under the constraint of prompt words, each agent outputs a standardized evidence unit including a unique evidence node ID, observation description, action sequence, input-output pair, knowledge reference, intermediate results and fault confidence. Specifically, step S3 achieves efficient division of labor and orderly collaboration among various intelligent agents through precise data distribution and standardized inference output. During implementation, the coordinator agent first obtains the standard diagnostic request packet encapsulated in step S1 through the data receiving interface. It then calls a data parsing tool to classify and analyze the envelope spectrum image, six time-domain statistical features, bearing model, real-time operating speed, the top 20 peak frequencies, and current operating parameters (including load level, runtime, and ambient temperature) within the packet. Based on the correspondence between data types and agent functions, the target distribution objects for each type of data are determined. Subsequently, according to preset data distribution rules, the envelope spectrum image and peak frequency list are directionally sent to the spectrum analysis agent via a high-speed data bus. The six time-domain statistical features are sent completely to the time-domain analysis agent. The bearing model and real-time operating speed are sent to the theoretical analysis agent. The peak frequency list, bearing model, and complete operating parameters are sent to the case matching agent, ensuring that each intelligent agent obtains all the data required to complete inference. After receiving the corresponding data, each agent initiates the inference process under the logical constraints of specific prompts. The spectrum analysis agent analyzes the peak frequency distribution, harmonic components, and sideband characteristics in the envelope spectrum using domain tools. The time-domain analysis agent judges the degree of anomaly based on the numerical range and trend of time-domain statistical characteristics. The theoretical analysis agent calculates the theoretical fault frequency of the corresponding bearing and compares it with the measured data. The case matching agent searches for historical cases with similar characteristics in the case database. During the inference process, each agent automatically generates a standardized evidence unit, including a 32-bit unique evidence node ID, observation data description, tool call action sequence, complete input-output pair, referenced domain knowledge entries, key intermediate calculation results, and fault confidence in the 0-1 range. This unit is fed back to the coordinator agent through a data interface, ensuring that each inference result is traceable and verifiable, providing standardized data support for subsequent evidence fusion.

[0021] S4, after the coordinator agent collects all preliminary reasoning results, it constructs an evidence graph, uses measured peak values, theoretical frequencies, and case labels as nodes, establishes support or contradictory edges based on matching degree and conflict relationship, performs confidence fusion for each standard fault component using weighted average or evidence theory, marks uncertain conclusions and interrogates contradictory agents to obtain supplementary reasoning. Specifically, step S4 achieves accurate aggregation of diagnostic conclusions and effective resolution of uncertainties by constructing an evidence graph and employing targeted fusion strategies. During implementation, the coordinator agent first collects standardized evidence units output by all agents through a data interface. Based on the core information within these evidence units, evidence nodes are constructed using measured peak frequency, theoretical fault frequency, and case matching tags as core elements. Each node includes a unique identifier, data content, and information about the source agent. Subsequently, according to data matching rules, support edges are established when the data similarity between two nodes exceeds 85%, and conflict edges are established when the fault conclusions corresponding to two nodes contradict each other and both have a confidence level higher than 0.6. This forms a structured evidence graph with interconnected multi-source evidence, visually presenting the consistency and conflict relationships of the reasoning results from each agent. Next, a divergence judgment is performed on the inference conclusions of each standard faulty component (including inner ring, outer ring, roller, and cage). If the difference in fault confidence scores among all agents is less than 0.1, it is determined that there is no divergence. Preset agent weighting coefficients (0.3 for spectrum analysis agent, 0.25 for time domain analysis agent, 0.3 for theoretical analysis agent, and 0.15 for case matching agent) are used to weight the confidence scores of each agent, resulting in the final confidence score for the faulty component. If the difference in confidence scores between agents is greater than or equal to 0.1, it is determined that there is a divergence. The evidence-theory fusion process is then initiated, converting the fault confidence scores of each agent into corresponding trust functions and likelihood functions. The inference results of all agents are gradually fused using synthesis rules to obtain the final fused confidence score. Finally, based on the preset confidence level criteria (greater than 0.8 for high-confidence faults, 0.3 to 0.8 for uncertain faults, and less than 0.3 for no faults), the state of each faulty component is determined, conclusions in the uncertain range are marked, and evidence maps and divergence point information are sent to the agent that generates contradictions. The agent receives supplementary reasoning results after re-analysis based on supplementary data, further optimizing the accuracy of the fused conclusions and ensuring that the diagnostic results can comprehensively reflect the integrated judgment of multi-source evidence.

[0022] S5, the coordinator agent generates a structured inspection report based on the evidence graph, questioning records and results of each stage, including a summary layer, evidence layer, reasoning chain layer and maintenance suggestion layer; Specifically, step S5 is the structured presentation of diagnostic results. By integrating diagnostic information from the entire process, a well-structured and comprehensive test report is generated, providing comprehensive support for engineers' decision-making. During implementation, the coordinator agent first invokes a data integration tool to extract node association information and the distribution of supporting and contradictory edges from the evidence graph. The system then systematically reviews the preliminary reasoning results of each agent, the records of inquiries regarding divergent conclusions, and the supplementary reasoning content provided by the agents. All diagnostic-related data, results, and process information are summarized to form a complete diagnostic dataset, ensuring that all information required for report generation is included. Subsequently, based on this dataset, a layered detection report is constructed according to a pre-defined structure. First, a summary layer is built, extracting core diagnostic conclusions (including the presence of a fault, the faulty component, and the fault confidence level), key fault characteristics, and overall diagnostic conclusions, using concise and clear language to facilitate quick retrieval of core information. Next, an evidence layer is built, presenting key evidence such as envelope spectrum images, peak frequency lists, time-domain feature values, theoretical frequency calculation results, and historical case matching details by category. Each piece of evidence is labeled with its source and associated evidence node ID to ensure traceability of the conclusions. Then, a reasoning chain layer is built, using chronological order and logical relationships as the framework to outline the complete logical path from raw data input, independent reasoning by each agent, evidence graph construction, confidence fusion, to divergence handling, clarifying the data flow process, tool usage, and conclusion derivation basis at each stage, clearly demonstrating the generation process of the diagnostic conclusions. Finally, combining the maintenance specifications and historical maintenance experience of high-speed train axle box bearings, a maintenance recommendation layer is formed, including fault handling priorities, recommended maintenance measures, maintenance window suggestions, and subsequent monitoring priorities, providing specific guidance for equipment maintenance. Once each level is constructed, the coordinator agent integrates them into a structured test report in a unified format, which is then output to the terminal device through the system interface. This ensures that the report is comprehensive in information, logically clear, and highly practical, meeting the needs of engineers for diagnosis, review, and maintenance decisions.

[0023] S6 receives the engineer's confirmation of the review conclusion, writes the confirmed or corrected diagnostic sample into the case knowledge base and records the expert preferences, adds the inference log summary of the high-confidence uncorrected sample to the example pool and dynamically splices it into the prompt words, and dynamically updates the weights of agents that perform poorly in the fault type labeling.

[0024] Specifically, step S6 is the system's self-evolutionary closed-loop update process. By absorbing expert experience and practical data, it achieves continuous optimization and adaptive improvement of diagnostic capabilities. During implementation, the system first receives the engineer's review conclusions on the structured inspection report through a human-computer interaction interface. This includes confirmation information or correction suggestions for the diagnostic results. The correction suggestions must clearly point out the deviations in the original diagnostic conclusions and the basis for the correct conclusions. Subsequently, the system automatically writes the confirmed diagnostic samples into the case knowledge base according to a preset format (including raw data, diagnostic process, final conclusions, and operating condition information). Simultaneously, it records the expert preferences demonstrated by the engineer during the review process (including the focus on specific fault characteristics and the type of maintenance recommendations), enriching the diversity and practicality of the case library and providing a more comprehensive reference for subsequent case matching agents. For diagnostic samples with high confidence (confidence greater than 0.9) and no correction, the system extracts the core logic, tool call sequences, and key evidence relationships from their inference logs, generating a 100-200 word inference log summary. This summary is added to the agent's prompt word example pool and dynamically appended to the end of the corresponding agent's exclusive prompt word in the order of addition. This example guides the agent to optimize its inference logic, improving the diagnostic efficiency and accuracy for similar faults. Simultaneously, the system periodically analyzes the performance metrics of various agents in diagnosing different fault types, including diagnostic accuracy, confidence stability, and contribution to divergence handling. When the diagnostic accuracy of a certain type of agent for a specific fault type falls below 80% for 100 consecutive times, a dynamic agent weight update mechanism is activated. Based on a preset weight adjustment algorithm, the weight coefficient of poorly performing agents is reduced (maximum adjustment not exceeding 0.1), while the weights of other high-performing agents are correspondingly increased, ensuring that system diagnostic resources are tilted towards efficient modules. Through this closed-loop mechanism of case knowledge base updates, dynamic prompt word optimization, and agent weight adjustment, the system's diagnostic capabilities achieve adaptive evolution, continuously adapting to the evolution of fault modes and changes in operating conditions of high-speed train axle box bearings.

[0025] Preferably, the final confidence level of the faulty component without divergence in S4 is calculated using the following formula: ; in This represents the final confidence level of the faulty component C. These are the weight coefficients for the four types of agents. The confidence scores of fault components C output by the agent in spectral analysis, time-domain analysis, theoretical analysis, and case matching are respectively given. The theoretical analysis agent's successful matching is indicated by... The value is 1.0, and 0 if no match is found.

[0026] Specifically, the calculation method for the final confidence level of undisputed faulty components integrates the diagnostic results of four types of specialized intelligent agents through weighted averaging, ensuring the comprehensiveness and reliability of the conclusions. During implementation, for each standard faulty component (inner ring, outer ring, roller, cage, etc.), the fault confidence levels output by the four types of intelligent agents—spectrum analysis, time-domain analysis, theoretical analysis, and case matching—are collected. All confidence level values ​​are strictly limited to the range of 0 to 1, directly quantifying the degree to which each intelligent agent judges the component to be faulty. The system presets weight coefficients for the four types of intelligent agents, with the sum of all coefficients fixed at 1. The specific values ​​are determined by statistically analyzing the accuracy rates of each type of intelligent agent in nearly 1000 sets of historical diagnostic data. Specifically, the weight of the spectrum analysis agent is set to 0.3, the time-domain analysis agent to 0.25, the theoretical analysis agent to 0.3, and the case matching agent to 0.15. Quantifying the diagnostic contribution and reliability of different intelligent agents enhances the scientific rigor of the integration. The theoretical analysis agent accurately calculates the theoretical fault frequency based on bearing structural parameters and kinematic principles. When the measured frequency matches the theoretical frequency by more than 90%, the confidence level is directly set to 1.0; otherwise, it is set to 0, avoiding interference from invalid data. During implementation, the confidence level of each agent is multiplied by its corresponding weight coefficient, and all products are summed to obtain the final confidence level of the faulty component. This calculation method is suitable for consistent scenarios where the judgment difference between all agents is less than 0.1. It can efficiently integrate multi-source consistent information, improve the stability and reliability of diagnostic conclusions, and provide accurate quantitative basis for subsequent fault determination.

[0027] Preferably, the faulty components with discrepancies in S4 are fused using Dempster-Shafer evidence theory, and the binary identification framework is defined as follows: ,in The identification framework for faulty component C is represented. Indicates a fault-free state. The formula for converting confidence levels to the BPA function to indicate a faulty state is as follows: ,in This represents the BPA function. This represents the confidence level of the faulty component C output by the corresponding agent. Theoretically, this indicates when the agent successfully matches the faulty component. The value is 1.0, and 0 if no match is found.

[0028] Specifically, for the faulty components with discrepancies in step S4, the specific rules for confidence fusion using Dempster-Shafer evidence theory are clearly defined. The core is to convert the confidence of each agent into a unified trust function, laying the foundation for subsequent synthesis. In implementation, a binary recognition framework is first established for each faulty component. This framework includes only two mutually exclusive states: no fault and faulty, clearly defining the judgment boundaries of the fusion process. Then, the fault confidence output by each type of agent is converted into a basic probability allocation function. This function is used to accurately describe the trust allocation of each agent to the two states. The conversion logic is that the fault confidence output by the agent directly corresponds to its trust level in the faulty state of the component. The trust level in the no-fault state is obtained by subtracting the fault confidence from 1, ensuring the completeness and rationality of the trust allocation. The conversion rules of the theoretical analysis agent are consistent with those of other agents. When the matching degree between the measured frequency and the theoretical fault frequency is higher than 90%, the trust level for the faulty state is set to 1.0; when no match is found, it is set to 0, highlighting the core supporting role of theoretical basis in the fusion process. This conversion process standardizes the inference results of different agents, unifies the confidence data scattered in the 0 to 1 range into a standardized trust allocation format, effectively solves the problem of inconsistent formats of multi-source information and difficulty in direct integration in divergent scenarios, provides standardized and comparable data support for obtaining comprehensive conclusions through synthesis rules, and ensures the rigor of the fusion process.

[0029] Preferably, the Dempster synthesis rule in S4 uses the following formula: ; ; in Indicates the conflict coefficient. Let BPA functions represent the functions of the two agents respectively. This represents the synthesized BPA function value. Indicates a fault-free state. This indicates a faulty state. The fusion process is terminated and the corresponding conclusion is marked.

[0030] Specifically, in step S4, the specific synthesis rules of evidence theory in the divergence scenario, through calculating the conflict coefficient and scientifically processing conflict information, ultimately yield a unified trust allocation result after fusion. During implementation, firstly, for the basic probability allocation functions of the two agents to be fused, the conflict coefficient between them is calculated. This coefficient is obtained by multiplying one agent's trust level in the fault-free state by the other agent's trust level in the faulty state, and then adding the product of one agent's trust level in the faulty state and the other agent's trust level in the fault-free state. The sum is the conflict coefficient, with a value between 0 and 1, directly reflecting the degree of conflict between the reasoning results of the two agents. Then, it is determined whether the conflict coefficient is 1. If it is 1, it indicates that the judgments of the two agents are completely contradictory and there is no consensus. At this point, the fusion process is immediately terminated, and the conclusion is specially marked to avoid invalid fusion leading to erroneous results. If the conflict coefficient is not 1, the Dempster synthesis rule is used to fuse the two trust assignments. The trust levels for the faulty and fault-free states are calculated separately after fusion. The calculation logic involves multiplying the trust levels of the two agents for the same state, then dividing by the difference between 1 and the conflict coefficient to obtain the final trust assignment for that state. This synthesis process quantifies the degree of conflict and reasonably handles conflict information, resolving discrepancies while preserving the effective judgments of each agent. This makes the fusion result more objective and persuasive, and is suitable for complex fault diagnosis scenarios where the confidence difference between agents is greater than or equal to 0.1.

[0031] Preferably, the expression for the agent to generate the preliminary structured reasoning result in S2 is: ,in This represents the initial structured reasoning result of the i-th type of intelligent agent. Denotes the i-th type of intelligent agent. This indicates a cue word specific to the i-th type of intelligent agent. This represents the data set input to the i-th type of intelligent agent, where the value of i corresponds to one of the five types of intelligent agents: spectrum analysis, time domain analysis, theoretical analysis, case matching, and coordinator.

[0032] Specifically, step S2 involves the logic of various agents generating preliminary structured reasoning results, establishing the correspondence between agents, prompts, input data, and output results to ensure the standardization and traceability of the reasoning process. During implementation, the specific roles of five types of agents—spectral analysis, time-domain analysis, theoretical analysis, case matching, and coordinator—are first clarified. Each type of agent acts as an independent reasoning unit with its own functional boundaries and reasoning logic. Subsequently, a unique prompt is configured for each type of agent. This prompt is designed based on knowledge of the fault diagnosis domain and task requirements, including core content such as reasoning process specifications, output format requirements, and key feature focus areas. The character length is controlled between 500 and 800 characters to strictly constrain the agent's reasoning direction and behavioral patterns, ensuring that the reasoning process does not deviate from the diagnostic objective. Simultaneously, based on the functional positioning of each agent, a corresponding set of input data is assigned to it. The input data all originates from the diagnostic request package encapsulated in step S1. After parsing and classification by the coordinator agent, the data is distributed in a targeted manner. Specifically, the spectrum analysis agent acquires the envelope spectrum image and peak frequency list, while the time-domain analysis agent acquires six time-domain statistical features. This ensures that each agent only acquires the relevant data needed to complete its own reasoning task, avoiding data redundancy and interference. During implementation, after receiving the input data, the agents, guided by specific prompts, invoke their configured domain tools to perform analysis and reasoning. Through a series of logical operations and feature extraction, they generate structured preliminary reasoning results, including unique evidence node IDs, observation descriptions, and action sequences. This achieves a one-to-one correspondence between input data, reasoning process, and output results, providing a clear logical chain for subsequent evidence fusion and tracing.

[0033] Preferably, the formula for generating the structured inspection report in step S5 is: ,in This indicates the final structured testing report. Represents the coordinator agent. These represent the structured reasoning results of the agent, which are derived from spectral analysis, temporal analysis, theoretical analysis, and case matching, respectively. This represents the constructed multi-source collaborative diagnostic evidence map.

[0034] Specifically, the logic for generating the structured detection report in step S5 involves the coordinator agent integrating multi-source information to construct a complete and standardized diagnostic report according to a preset hierarchy. During implementation, the coordinator agent, as the core unit for report generation, first collects the structured reasoning results output by four types of specialized agents: spectrum analysis, time-domain analysis, theoretical analysis, and case matching. These results include core information such as each agent's fault judgment, confidence level in the 0-1 interval, key evidence, and complete reasoning process. Simultaneously, it extracts the multi-source collaborative diagnostic evidence graph constructed in step S4. This graph includes the relationships between all evidence nodes, supporting and contradictory edges, comprehensively reflecting the reasoning consensus and disagreements among the multiple agents. Subsequently, the coordinator agent initiates the report construction process, using the evidence graph as the core link to associate and match the reasoning results of the four types of agents with the node and edge information in the graph, clarifying the evidentiary support and logical relationships of each reasoning result. Based on this, the report is constructed according to a pre-set four-layer structure. First, the core conclusions are extracted to form the summary layer, with a character length controlled between 200 and 300 words. Then, various original evidences such as envelope spectrum images and peak frequency lists are integrated to form the evidence layer. Next, the complete reasoning logic is sorted out to form the reasoning chain layer, which records the data flow and conclusion derivation process in detail. Finally, the maintenance recommendation layer is formed by combining the maintenance specifications of high-speed train axle box bearings and historical experience. Throughout the entire construction process, the coordinator agent ensures the integrity and relevance of information at each level, realizing the full-process presentation from raw data and reasoning process to final conclusions and maintenance recommendations. The generated structured inspection report is not only highly logical and traceable, but also provides engineers with comprehensive and practical diagnostic references, effectively supporting the formulation of maintenance decisions.

[0035] Preferred, such as Figure 2 As shown, S3 includes the following sub-steps: S31, the coordinator agent receives the diagnostic request packet, analyzes and classifies the envelope spectrum image, time-domain features, bearing model, operating speed, peak frequency, and operating parameters, and clarifies the distribution objects corresponding to each type of data; S32, according to the preset data distribution rules, the envelope spectrum image and peak frequency data are sent to the spectrum analysis agent, the time-domain feature data is sent to the time-domain analysis agent, the bearing model and speed data are sent to the theoretical analysis agent, and the peak frequency, bearing model, and operating parameters are sent to the case matching agent; S33, after receiving the corresponding data, each agent performs targeted analysis and processing based on the reasoning logic and constraints set by the exclusive prompt words, forming a standardized evidence unit including a complete reasoning trajectory; S34, each agent embeds a unique evidence node ID, observation description, action sequence, input-output pair, knowledge reference, intermediate results, and fault confidence information into the standardized evidence unit and outputs it.

[0036] Specifically, step S3 achieves precise data distribution and standardized evidence unit output through step-by-step implementation, ensuring the orderliness and traceability of multi-agent collaborative reasoning. During implementation, in S31, the coordinator agent receives the standard diagnostic request packet encapsulated in step S1 via a preset high-speed interface. It then calls the built-in data parsing tool to classify and analyze the envelope spectrum image (1024×768 pixels resolution), six time-domain statistical features (peak value, peak-to-peak value, mean, variance, kurtosis, skewness), bearing model, operating speed (500 rpm to 3000 rpm), the top 20 peak frequencies, and current operating parameters (including load level, runtime, and ambient temperature), clarifying the target agent corresponding to each type of data to ensure accurate data distribution. In S32, according to the matching rules between agent functions and data types, targeted distribution is completed through the data bus. The envelope spectrum image and peak frequency list are sent to the spectrum analysis agent, the time-domain statistical features to the time-domain analysis agent, the bearing model and speed to the theoretical analysis agent, and the peak frequencies, bearing model, and operating parameters to the case matching agent, ensuring accurate data distribution. Each agent acquires only the specific data required to complete its own reasoning task. In S33, after receiving the corresponding data, each agent initiates domain tools to conduct targeted analysis under the constraint of specific prompts of 500 to 800 words. The spectrum analysis agent focuses on frequency distribution, harmonic components, and sideband characteristics; the time domain analysis agent focuses on the numerical range and trend of time domain characteristics; the theoretical analysis agent focuses on the calculation of theoretical fault frequencies and comparison with measured data; and the case matching agent focuses on the retrieval of historical case features and similarity analysis. In S34, during the reasoning process, each agent automatically generates standardized evidence units, including a 32-bit unique evidence node ID, observation data description, tool call action sequence, complete input-output pairs, referenced domain knowledge entries, key intermediate results, and fault confidence in the range of 0 to 1. These units are fed back to the coordinator agent through a preset interface to ensure that each reasoning result has traceable and verifiable characteristics, providing standardized data support for subsequent evidence fusion.

[0037] Preferred, such as Figure 3As shown, S4 includes the following sub-steps: S41, the coordinator agent summarizes the preliminary reasoning results output by all agents, constructs evidence nodes with measured peak values, theoretical frequencies, and case labels as the core, establishes support edges based on data matching degree judgment, establishes contradictory edges based on conclusion conflict relationships, and forms an evidence graph with multi-source evidence structured association; S42, performs divergence judgment on each standard fault component, filters out fault components without divergence, and uses preset weight coefficients to weight the confidence of the output of various agents to obtain the final confidence of such fault components; S43, for fault components with divergence, initiates the evidence theory fusion process, converts the confidence of various agents into BPA functions, and gradually merges the BPA functions of all agents through synthesis rules to obtain the final fusion result; S44, judges the state of fault components based on the final confidence value range, marks the conclusions in the uncertain interval and sends the evidence graph to the contradictory agent, and receives the supplementary reasoning results after processing by the agent.

[0038] Specifically, step S4 achieves precise aggregation of diagnostic conclusions by constructing an evidence graph, fusing confidence levels, and handling discrepancies step by step. In implementation, S41 involves the coordinator agent collecting standardized evidence units output by all agents through a data interface. Evidence nodes are constructed using measured peak frequency, theoretical fault frequency, and case matching tags as core elements. Each node includes a unique identifier, data content, and source agent information. Based on data matching rules, support edges are established when the similarity between two nodes is higher than 85%, and contradictory edges are established when the corresponding fault conclusions of two nodes contradict each other and both have confidence levels higher than 0.6, forming a multi-source evidence graph with structured associations. S42 involves discrepancy judgment for each standard fault component (inner ring, outer ring, roller, cage). If the difference in fault confidence levels output by all agents is less than 0.1, it is considered undisputed, and preset weighting coefficients are used (0.3 for spectral analysis, 0.25 for time-domain analysis, 0.3 for theoretical analysis, and 0.1 for case matching). 5) Calculate the confidence scores using weighted averages to obtain the final confidence score of the faulty component; S43 If the confidence score difference between agents is greater than or equal to 0.1, it is determined that there is a disagreement, and the evidence theory fusion process is initiated. The fault confidence scores of various agents are converted into corresponding trust functions and likelihood functions. The reasoning results of all agents are gradually fused through synthesis rules to obtain the final fused confidence score; S44 According to the preset confidence judgment criteria (greater than 0.8 is high confidence fault, 0.3 to 0.8 is uncertain, and less than 0.3 is no fault), the state of each faulty component is judged, the conclusions in the uncertain interval are marked, and the evidence map and disagreement point information are sent to the agents that produce contradictions. The supplementary reasoning results after the agents re-analyze the supplementary data are received to further optimize the accuracy of the fused conclusion and ensure that the diagnostic results fully reflect the comprehensive judgment of multi-source evidence.

[0039] Preferred, such as Figure 4 As shown, S5 includes the following sub-steps: S51, the coordinator agent extracts node association information, supporting and contradictory edge distribution from the evidence graph, organizes the preliminary reasoning results, questioning feedback and supplementary reasoning content of each agent, and forms a complete diagnostic data set; S52, based on the diagnostic data set, a summary layer is constructed to extract the core diagnostic conclusions, fault component judgment results and confidence distribution; S53, an evidence layer is constructed to integrate envelope spectrum images, peak frequencies, time-domain features, theoretical frequency calculation results and historical case matching information to calibrate evidence; S54, a reasoning chain layer is constructed to sort out the complete logical path from the original data input to the final conclusion output, clarify the data flow and reasoning basis at each stage, form a maintenance suggestion layer based on maintenance experience, and integrate information from each level to generate a structured test report.

[0040] Specifically, step S5 integrates diagnostic information step by step and constructs a layered report to generate a logically clear and complete diagnostic result. During implementation, S51 involves the coordinator agent calling a data integration tool to extract node association information and the distribution of supporting and contradictory edges from the evidence graph. The system then organizes the preliminary reasoning results of each agent, the records of inquiries regarding divergent conclusions, and the supplementary reasoning content provided by the agents, summarizing all diagnostic-related data, results, and process information to form a complete diagnostic data set, ensuring that no information required for report generation is omitted. S52 uses this data set to construct a summary layer, extracting core diagnostic conclusions (whether a fault exists, the faulty component, and the fault confidence level), key fault characteristics, and overall diagnostic conclusions. The character length is controlled to 200-300 characters to facilitate rapid retrieval of core information. S53 constructs an evidence layer, incorporating envelope spectrum images, peak frequency lists, time-domain feature values, and theoretical frequencies. Key evidence, such as calculation results and historical case matching details, is presented in a categorized manner. Each piece of evidence is labeled with its source and associated evidence node ID to ensure the traceability of conclusions. The S54 constructs a reasoning chain layer, using time sequence and logical relationships as the framework to outline the complete logical path from raw data input, independent reasoning by each agent, evidence graph construction, confidence fusion, to disagreement handling. It clarifies the data flow process, tool usage, and conclusion derivation basis at each stage. Combining high-speed train axle box bearing maintenance specifications and historical maintenance experience, a maintenance suggestion layer is formed, including fault handling priorities, recommended maintenance measures, maintenance window suggestions, and subsequent monitoring priorities. All levels are integrated into a structured inspection report in a unified format, which is output to terminal devices through the system interface to meet the needs of engineers for diagnosis review and maintenance decision-making.

[0041] like Figure 5As shown, a multi-agent high-speed train axle box bearing fault diagnosis method traceable to evolution is characterized by being implemented through different units, including: a multi-dimensional parameter and signal acquisition unit, used to acquire bearing model, operating speed, sampling frequency, operating condition parameters, and original vibration signals; filtering and transforming the vibration signals to generate an envelope spectrum; extracting significant frequency components and time-domain statistical features and encapsulating them; an agent initialization and configuration unit, used to initialize five types of agents: spectrum analysis, time-domain analysis, theoretical analysis, case matching, and coordinator; configuring exclusive large language models, prompt words, and domain tools for each type of agent; and defining the agent inference output format; and a data distribution and inference execution unit, used to receive and parse diagnostic request packets, distribute various types of data to the corresponding agents according to preset rules, and drive each agent to perform inference. The system comprises several modules: an independent analysis unit under prompt constraints, which outputs standardized evidence units; a multi-source evidence fusion and graph construction unit, which collects reasoning results from various agents, constructs an evidence graph including supporting and contradictory edges, fuses the confidence levels of faulty components using weighted averaging or evidence theory, handles divergent conclusions, and questions contradictory agents; a structured report generation unit, which integrates the evidence graph, questioning records, and reasoning results from each stage, and generates a diagnostic report according to a structure of summary layer, evidence layer, reasoning chain layer, and maintenance suggestion layer; and a self-evolving closed-loop update unit, which receives engineer review and correction information, updates the case knowledge base and expert preference records, dynamically adjusts prompts and agent weights, and adaptively evolves the system's diagnostic capabilities. All units interact and collaborate via a data bus to ensure the orderly progress of the diagnostic process.

[0042] A traceable, evolution-derived multi-agent high-speed train axle box bearing fault diagnosis method is proposed. This method divides intelligent agents into five categories: spectrum analysis, time-domain analysis, theoretical analysis, case matching, and a coordinator. This enables specialized processing of multi-dimensional data. The coordinator agent accurately distributes various parameters according to preset rules, allowing different agents to focus on their specific domains for targeted analysis. This comprehensive integration of multi-source information overcomes the limitations of traditional single-model approaches in capturing the multi-dimensional characteristics of complex faults. By embedding core information such as unique identifiers, action sequences, and knowledge references during the reasoning process, a hierarchical report is constructed, including summaries, evidence, reasoning chains, and maintenance recommendations. This makes the basis and logical chain for generating diagnostic conclusions clearly traceable, solving the problems of opaque reasoning processes and difficulty in retrospection in existing technologies.

[0043] Meanwhile, this method possesses dynamic self-evolution capabilities and an efficient disagreement handling mechanism, significantly improving diagnostic adaptability and reliability: by updating the case knowledge base with engineer review feedback, incorporating high-confidence sample reasoning logs into prompt words, and dynamically adjusting agent weights, the system can continuously absorb new experiences and adapt to fault mode evolution and operating condition changes, making up for the shortcomings of traditional static architecture diagnostic accuracy decay; the coordinator agent identifies data associations and conflicts by constructing evidence graphs, uses weighted averaging for undisputed conclusions and special fusion rules for disputed conclusions, and simultaneously queries contradictory agents to obtain supplementary reasoning, strengthening the ability to integrate multi-source results, solving the problems of weak disagreement handling and low diagnostic efficiency in existing technologies, and comprehensively improving the accuracy, stability, and practicality of high-speed train axle box bearing fault diagnosis.

[0044] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," "link," and "fix" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0045] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various equivalent changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A traceable, evolution-derived multi-agent method for diagnosing high-speed train axle box bearing faults, characterized in that: Includes the following steps: S1 collects bearing model, operating speed and sampling frequency related parameters, determines the optimal center frequency based on historical data and sends it to the data collection device, collects the original vibration signal through timed trigger or vibration RMS mutation trigger mechanism, filters and transforms the original vibration signal to generate envelope spectrum, extracts significant frequency components and time domain statistical features in the envelope spectrum, generates a unique diagnostic session ID as the root node, and encapsulates the envelope spectrum image, peak frequency list and time domain features for subsequent diagnosis. S2 initializes five types of intelligent agents: spectrum analysis, time domain analysis, theoretical analysis, case matching, and coordinator. Each type of intelligent agent consists of a large language model, dedicated prompt words, and domain tools. Each type of intelligent agent generates structured and traceable preliminary reasoning results based on prompt words and input data. S3, the coordinator agent receives and parses the diagnostic request packet, and distributes the envelope spectrum image, time domain features, bearing model, speed, peak frequency and operating parameters to the corresponding agents. Under the constraint of prompt words, each agent outputs a standardized evidence unit including a unique evidence node ID, observation description, action sequence, input-output pair, knowledge reference, intermediate results and fault confidence. S4, after the coordinator agent collects all preliminary reasoning results, it constructs an evidence graph, uses measured peak values, theoretical frequencies, and case labels as nodes, establishes support or contradictory edges based on matching degree and conflict relationship, performs confidence fusion for each standard fault component using weighted average or evidence theory, marks uncertain conclusions and interrogates contradictory agents to obtain supplementary reasoning. S5, the coordinator agent generates a structured inspection report based on the evidence graph, questioning records and results of each stage, including a summary layer, evidence layer, reasoning chain layer and maintenance suggestion layer; S6 receives the engineer's confirmation of the review conclusion, writes the confirmed or corrected diagnostic sample into the case knowledge base and records the expert preferences, adds the inference log summary of the high-confidence uncorrected sample to the example pool and dynamically splices it into the prompt words, and dynamically updates the weights of agents that perform poorly in the fault type labeling.

2. The traceable, evolutionarily derived multi-agent high-speed train axle box bearing fault diagnosis method according to claim 1, characterized in that, The final confidence level of the faulty component in S4 that has no divergence is calculated using the following formula: ； in This represents the final confidence level of the faulty component C. These are the weight coefficients for the four types of agents. The confidence scores of fault components C output by the agent in spectral analysis, time-domain analysis, theoretical analysis, and case matching are respectively given. The theoretical analysis agent's successful matching is indicated by... The value is 1.0, and 0 if no match is found.

3. The traceable, evolutionarily derived multi-agent high-speed train axle box bearing fault diagnosis method according to claim 1, characterized in that, The divergent fault components in S4 are fused using Dempster-Shafer evidence theory, and the binary identification framework is defined as follows: ,in The identification framework for faulty component C is represented. Indicates a fault-free state. The formula for converting confidence levels to the BPA function to indicate a faulty state is as follows: ,in This represents the BPA function. This represents the confidence level of the faulty component C output by the corresponding agent. Theoretically, this indicates when the agent successfully matches the faulty component. The value is 1.0, and 0 if no match is found.

4. The traceable, evolutionarily derived multi-agent high-speed train axle box bearing fault diagnosis method according to claim 1, characterized in that, The Dempster synthesis rule in S4 uses the following formula: ； ； in Indicates the conflict coefficient. Let BPA functions represent the functions of the two agents respectively. This represents the synthesized BPA function value. Indicates a fault-free state. This indicates a faulty state. The fusion process is terminated and the corresponding conclusion is marked.

5. The traceable, evolutionarily derived multi-agent high-speed train axle box bearing fault diagnosis method according to claim 1, characterized in that, The expression for the agent generating the preliminary structured reasoning result in S2 is as follows: ,in This represents the initial structured reasoning result of the i-th type of intelligent agent. Denotes the i-th type of intelligent agent. This indicates a cue word specific to the i-th type of intelligent agent. This represents the data set input to the i-th type of intelligent agent, where the value of i corresponds to one of the five types of intelligent agents: spectrum analysis, time domain analysis, theoretical analysis, case matching, and coordinator.

6. The traceable, evolutionarily derived multi-agent high-speed train axle box bearing fault diagnosis method according to claim 1, characterized in that, The formula for generating the structured inspection report in S5 is as follows: ,in This indicates the final structured testing report. Represents the coordinator agent. These represent the structured reasoning results of the agent, which are derived from spectral analysis, temporal analysis, theoretical analysis, and case matching, respectively. This represents the constructed multi-source collaborative diagnostic evidence map.

7. The traceable, evolutionarily derived multi-agent high-speed train axle box bearing fault diagnosis method according to claim 1, characterized in that, S3 includes the following sub-steps: S31, the coordinator agent receives the diagnostic request packet, analyzes and classifies the envelope spectrum image, time-domain features, bearing model, operating speed, peak frequency, and operating parameters, and clarifies the distribution objects corresponding to each type of data; S32, according to the preset data distribution rules, the envelope spectrum image and peak frequency data are sent to the spectrum analysis agent, the time-domain feature data is sent to the time-domain analysis agent, the bearing model and speed data are sent to the theoretical analysis agent, and the peak frequency, bearing model, and operating parameters are sent to the case matching agent; S33, after receiving the corresponding data, each agent performs targeted analysis and processing based on the reasoning logic and constraints set by the exclusive prompt words, forming a standardized evidence unit including a complete reasoning trajectory; S34, each agent embeds a unique evidence node ID, observation description, action sequence, input-output pair, knowledge reference, intermediate results, and fault confidence information into the standardized evidence unit and outputs it.

8. The traceable, evolutionarily derived multi-agent high-speed train axle box bearing fault diagnosis method according to claim 1, characterized in that, S4 includes the following sub-steps: S41, the coordinator agent summarizes the preliminary inference results output by all agents, constructs evidence nodes with measured peak values, theoretical frequencies, and case labels as the core, establishes support edges based on data matching degree judgment, establishes contradictory edges based on conclusion conflict relationships, and forms an evidence graph with structured association of multi-source evidence; S42, performs divergence judgment on each standard fault component, filters out fault components without divergence, and uses preset weight coefficients to weight the confidence of the output of various agents to obtain the final confidence of the fault component. S43, For faulty components with discrepancies, initiate the evidence theory fusion process, convert the confidence of various agents into BPA functions, and gradually fuse the BPA functions of all agents through synthesis rules to obtain the final fusion result; S44, determine the state of the faulty component based on the final confidence value range, mark the conclusions in the uncertainty interval and send the evidence map to the contradictory agents, and receive the supplementary reasoning results after processing by the agents.

9. The traceable, evolutionarily derived multi-agent high-speed train axle box bearing fault diagnosis method according to claim 1, characterized in that, S5 includes the following sub-steps: S51, the coordinator agent extracts node association information, supporting and contradictory edge distribution from the evidence graph, organizes the preliminary reasoning results, questioning feedback and supplementary reasoning content of each agent, and forms a complete diagnostic data set; S52, based on the diagnostic data set, a summary layer is constructed to extract the core diagnostic conclusions, fault component judgment results and confidence distribution; S53, an evidence layer is constructed to integrate envelope spectrum images, peak frequencies, time-domain features, theoretical frequency calculation results and historical case matching information to calibrate evidence; S54, a reasoning chain layer is constructed to sort out the complete logical path from the original data input to the final conclusion output, clarify the data flow and reasoning basis at each stage, form a maintenance suggestion layer based on maintenance experience, and integrate information from each level to generate a structured test report.

10. The method for diagnosing faults in high-speed train axle box bearings using a traceable, evolutionarily derived multi-agent system according to any one of claims 1-9, characterized in that... This method is implemented through different units, including: a multi-dimensional parameter and signal acquisition unit, used to acquire bearing model, operating speed, sampling frequency, operating condition parameters, and raw vibration signals; filtering and transforming the vibration signals to generate an envelope spectrum; extracting significant frequency components and time-domain statistical features and encapsulating them; an agent initialization and configuration unit, used to initialize five types of agents: spectrum analysis, time-domain analysis, theoretical analysis, case matching, and coordinator; configuring dedicated large language models, prompt words, and domain tools for each type of agent; and defining the agent's inference output format; and a data distribution and inference execution unit, used to receive and parse diagnostic request packets; distribute various types of data to the corresponding agents according to preset rules; drive each agent to conduct independent analysis under prompt word constraints; and output standardized data. The system comprises several units: an evidence unit; a multi-source evidence fusion and graph construction unit, used to collect reasoning results from various agents, construct an evidence graph including supporting and contradictory edges, fuse the confidence levels of faulty components using weighted averaging or evidence theory, handle divergent conclusions, and question contradictory agents; a structured report generation unit, used to integrate the evidence graph, questioning records, and reasoning results from each stage, generating diagnostic reports according to a structure of summary layer, evidence layer, reasoning chain layer, and maintenance suggestion layer; and a self-evolving closed-loop update unit, used to receive engineer review and correction information, update the case knowledge base and expert preference records, dynamically adjust prompts and agent weights, and adaptively evolve the system's diagnostic capabilities. All units interact and collaborate via a data bus to ensure the orderly progress of the diagnostic process.

Images