Agent-based fault diagnosis methods, electronic devices and storage media

Abstract

This application provides an agent-based fault diagnosis method, electronic device, and storage medium. It relates to the field of intelligent fault diagnosis for industrial systems. The method includes: acquiring first text describing a device fault; determining a starting node for fault diagnosis based on the first text; obtaining a first tool to be invoked based on a pre-trained first detection model, starting from the starting node, and obtaining first local detection evidence based on the first detection model and the first tool to be invoked; calibrating the credibility of the first local detection evidence based on the first detection model, and determining a first closure state of the calibrated evidence; and determining the fault diagnosis result based on the first closure state. This improves the accuracy of fault diagnosis in multi-level industrial systems.

Description

Technical Field

[0001] This application relates to the field of intelligent fault diagnosis in industrial systems, and in particular to a fault diagnosis method, electronic device and storage medium based on intelligent agents. Background Technology

[0002] Industrial systems typically consist of multi-level equipment, components, sensors, and control loops, with physical connections, upstream and downstream dependencies, media flow, or functional coupling relationships between these components. During the operation of industrial systems, rapid fault location and root cause analysis are crucial for ensuring production safety and efficiency.

[0003] Currently, fault classification is typically based on sensor data or historical data to diagnose faults in industrial systems. However, the accuracy of fault diagnosis is low when the fault manifests as abnormal signals from multiple downstream components. Summary of the Invention

[0004] This application provides a fault diagnosis method, electronic device, and storage medium based on intelligent agents to improve the accuracy of fault diagnosis.

[0005] In a first aspect, embodiments of this application provide a fault diagnosis method based on an intelligent agent, including:

[0006] Retrieve the first text used to describe the device malfunction;

[0007] Based on the first text, determine the starting node for fault diagnosis;

[0008] Starting from the starting node, a first tool to be called is obtained based on a pre-trained first detection model, and a first local detection evidence is obtained based on the first detection model and the first tool to be called. The first detection model is trained based on multiple fault samples, the expert diagnosis trajectory corresponding to each fault sample, and the confidence of each of the multiple nodes corresponding to each fault sample.

[0009] Based on the first detection model, the credibility of the first local detection evidence is calibrated, and the first closure state of the calibrated evidence is determined. The first closure state is used to indicate whether the calibrated evidence is sufficient to lock the root cause of the failure and whether there is evidence conflict.

[0010] Based on the first closed state, the fault diagnosis result is determined.

[0011] In one possible implementation, the first tool to be invoked includes one or more of the following: a topology query tool, a node fault diagnosis tool, and a sensor data reading tool;

[0012] The topology query tool is used to query the topological connection relationships between devices, components, and sensors within the system. The node fault diagnosis tool is used to output the fault probability and local faults of nodes. The sensor data reading tool is used to collect the sensing information and time-series operation data of the corresponding measurement points.

[0013] In one possible implementation, obtaining first local detection evidence based on a first detection model and a first tool to be invoked includes:

[0014] When the first tool to be invoked includes a topology query tool, the local observation topology and upstream and downstream related node information corresponding to the node are determined based on the first tool to be invoked, and the first local detection evidence is determined based on the local observation topology and upstream and downstream related node information.

[0015] When the first tool to be invoked includes a node fault diagnosis tool, the fault judgment result of the node is determined based on the first tool to be invoked, and the first local detection evidence is determined based on the fault judgment result;

[0016] When the first tool to be invoked includes a sensor data reading tool, the sensing time series operation data of the target measurement point corresponding to the time period is determined based on the first tool to be invoked, and the first local detection evidence is determined based on the sensing time series operation data.

[0017] In one possible implementation, determining the fault diagnosis result based on the first closed state includes:

[0018] When the first closed state indicates that the calibrated evidence is sufficient to pinpoint the root cause of the fault and there is no conflict of evidence, the fault diagnosis results, including the root cause node, the fault propagation path across devices, and the tracing evidence chain, are determined based on the first local detection evidence.

[0019] When the first closed state indicates that the evidence after calibration is insufficient to pinpoint the root cause of the fault or that there is a conflict of evidence, a second tool to be invoked is obtained, the second tool to be invoked is invoked to obtain second local detection evidence, and the fault diagnosis result is determined based on the first local detection evidence and the second local detection evidence.

[0020] In one possible implementation, determining the first closure state of the calibrated evidence includes:

[0021] Based on the calibrated evidence, a first candidate root cause set and a first candidate root cause evidence distribution are determined. The first candidate root cause evidence distribution is used to indicate the degree of fault support of the calibrated evidence for each candidate root cause in the first candidate root cause set.

[0022] Align the first candidate root cause set and determine the first evidence credibility information after alignment. The first evidence credibility information is used to indicate the failure confidence of each candidate root cause in the first candidate root cause set.

[0023] Based on the first evidence credibility information and the first candidate root cause evidence distribution, the second evidence credibility information is determined. The second evidence credibility information is used to indicate the failure confidence of each candidate root cause obtained after the evidence is fused and calibrated.

[0024] Based on the credible information of the second piece of evidence, the first closed state is determined.

[0025] In one possible implementation, the training method for the first detection model includes:

[0026] The system acquires multiple fault samples, multiple first expert diagnostic trajectories corresponding to each fault sample, and the confidence scores of multiple nodes corresponding to each fault sample. The first expert diagnostic trajectories sequentially include the nodes to be investigated and the first tool to be called when investigating each node.

[0027] Based on multiple fault samples, multiple first-expert diagnostic trajectories, and the confidence levels of multiple nodes corresponding to each fault sample, a second detection model is determined.

[0028] Based on multiple fault samples and the second detection model, the first detection model is determined.

[0029] In one possible implementation, a second detection model is determined based on multiple fault samples, multiple first expert diagnostic trajectories, and the confidence levels of multiple nodes corresponding to each fault sample, including:

[0030] Based on multiple fault samples, multiple first-expert diagnostic trajectories, and the confidence levels of multiple nodes corresponding to each fault sample, structured samples are determined. The structured samples are used to indicate the local topology, single-step diagnostic action, single-step tool observation data, and subsequent detection status updated after the action is executed for a single detection.

[0031] The second detection model is obtained by performing supervised fine-tuning training based on structured samples.

[0032] In one possible implementation, the first detection model is determined based on multiple fault samples and a second detection model, including:

[0033] For each fault sample, a fault diagnosis trajectory is determined based on the second detection model;

[0034] For each fault diagnosis trajectory, determine the tool usage information. The tool usage information is used to indicate the resource consumption cost incurred by calling the tool during the process of determining the fault diagnosis trajectory.

[0035] For each fault diagnosis trajectory, the credibility calibration is performed on each observation evidence obtained using multiple tools to be invoked within the trajectory, and the second candidate root cause set and the distribution of the second candidate root cause evidence are determined after calibration.

[0036] For each fault diagnosis trajectory, a second closure state is determined based on the second candidate root cause set and the distribution of evidence for the second candidate root cause.

[0037] Based on the second closed state of each fault diagnosis trajectory and the tool usage information, the second detection model is iteratively optimized to obtain the first detection model.

[0038] Secondly, embodiments of this application provide a fault diagnosis device based on an intelligent agent, comprising: an acquisition module, a first determination module, a first processing module, a second processing module, and a second determination module, wherein...

[0039] The acquisition module is used to acquire the first text describing the device malfunction.

[0040] The first determining module is used to determine the starting node for fault diagnosis based on the first text;

[0041] The first processing module is used to obtain the first tool to be called based on the pre-trained first detection model, starting from the starting node, and to obtain the first local detection evidence based on the first detection model and the first tool to be called. The first detection model is trained based on multiple fault samples, the expert diagnosis trajectory corresponding to each fault sample, and the confidence of each of the multiple nodes corresponding to each fault sample.

[0042] The second processing module is used to perform credibility calibration on the first local detection evidence based on the first detection model, and determine the first closure state of the calibrated evidence. The first closure state is used to indicate whether the calibrated evidence is sufficient to lock the root cause of the fault and whether there is evidence conflict.

[0043] The second determining module is used to determine the fault diagnosis result based on the first closed state.

[0044] In one possible implementation, the first tool to be invoked includes one or more of the following: a topology query tool, a node fault diagnosis tool, and a sensor data reading tool;

[0045] The topology query tool is used to query the topological connection relationships between devices, components, and sensors within the system. The node fault diagnosis tool is used to output the fault probability and local faults of nodes. The sensor data reading tool is used to collect the sensing information and time-series operation data of the corresponding measurement points.

[0046] In one possible implementation, the first processing module is specifically used for:

[0047] When the first tool to be invoked includes a topology query tool, the local observation topology and upstream and downstream related node information corresponding to the node are determined based on the first tool to be invoked, and the first local detection evidence is determined based on the local observation topology and upstream and downstream related node information.

[0048] When the first tool to be invoked includes a node fault diagnosis tool, the fault judgment result of the node is determined based on the first tool to be invoked, and the first local detection evidence is determined based on the fault judgment result;

[0049] When the first tool to be invoked includes a sensor data reading tool, the sensing time series operation data of the target measurement point corresponding to the time period is determined based on the first tool to be invoked, and the first local detection evidence is determined based on the sensing time series operation data.

[0050] In one possible implementation, the second determining module is specifically used for:

[0051] When the first closed state indicates that the calibrated evidence is sufficient to pinpoint the root cause of the fault and there is no conflict of evidence, the fault diagnosis results, including the root cause node, the fault propagation path across devices, and the tracing evidence chain, are determined based on the first local detection evidence.

[0052] When the first closed state indicates that the evidence after calibration is insufficient to pinpoint the root cause of the fault or that there is a conflict of evidence, a second tool to be invoked is obtained, the second tool to be invoked is invoked to obtain second local detection evidence, and the fault diagnosis result is determined based on the first local detection evidence and the second local detection evidence.

[0053] In one possible implementation, the second processing module is specifically used for:

[0054] Based on the calibrated evidence, a first candidate root cause set and a first candidate root cause evidence distribution are determined. The first candidate root cause evidence distribution is used to indicate the degree of fault support of the calibrated evidence for each candidate root cause in the first candidate root cause set.

[0055] Align the first candidate root cause set and determine the first evidence credibility information after alignment. The first evidence credibility information is used to indicate the failure confidence of each candidate root cause in the first candidate root cause set.

[0056] Based on the first evidence credibility information and the first candidate root cause evidence distribution, the second evidence credibility information is determined. The second evidence credibility information is used to indicate the failure confidence of each candidate root cause obtained after the evidence is fused and calibrated.

[0057] Based on the credible information of the second piece of evidence, the first closed state is determined.

[0058] In one possible implementation, the agent-based fault diagnosis device further includes a third processing module, which is used for:

[0059] The system acquires multiple fault samples, multiple first expert diagnostic trajectories corresponding to each fault sample, and the confidence scores of multiple nodes corresponding to each fault sample. The first expert diagnostic trajectories sequentially include the nodes to be investigated and the first tool to be called when investigating each node.

[0060] Based on multiple fault samples, multiple first-expert diagnostic trajectories, and the confidence levels of multiple nodes corresponding to each fault sample, a second detection model is determined.

[0061] Based on multiple fault samples and the second detection model, the first detection model is determined.

[0062] In one possible implementation, the third processing module is specifically used for:

[0063] Based on multiple fault samples, multiple first-expert diagnostic trajectories, and the confidence levels of multiple nodes corresponding to each fault sample, structured samples are determined. The structured samples are used to indicate the local topology, single-step diagnostic action, single-step tool observation data, and subsequent detection status updated after the action is executed for a single detection.

[0064] The second detection model is obtained by performing supervised fine-tuning training based on structured samples.

[0065] In one possible implementation, the third processing module is specifically used for:

[0066] For each fault sample, a fault diagnosis trajectory is determined based on the second detection model;

[0067] For each fault diagnosis trajectory, determine the tool usage information. The tool usage information is used to indicate the resource consumption cost incurred by calling the tool during the process of determining the fault diagnosis trajectory.

[0068] For each fault diagnosis trajectory, the credibility calibration is performed on each observation evidence obtained using multiple tools to be invoked within the trajectory, and the second candidate root cause set and the distribution of the second candidate root cause evidence are determined after calibration.

[0069] For each fault diagnosis trajectory, a second closure state is determined based on the second candidate root cause set and the distribution of evidence for the second candidate root cause.

[0070] Based on the second closed state of each fault diagnosis trajectory and the tool usage information, the second detection model is iteratively optimized to obtain the first detection model.

[0071] Thirdly, embodiments of this application provide an electronic device, including: at least one processor and a memory; the memory stores computer-executable instructions; the at least one processor executes the computer-executable instructions stored in the memory, causing the at least one processor to perform the agent-based fault diagnosis method as described in the first aspect above and various possible designs of the first aspect.

[0072] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions. When a processor executes the computer-executable instructions, it implements the above-mentioned first aspect and various possible designs of the first aspect, a fault diagnosis method based on an intelligent agent.

[0073] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the agent-based fault diagnosis method as described in the first aspect and various possible designs of the first aspect.

[0074] In a sixth aspect, embodiments of this application provide a chip, the chip including at least one processor, the processor being configured to execute program instructions to implement the agent-based fault diagnosis method as described in the first aspect above and various possible designs of the first aspect.

[0075] This application provides a fault diagnosis method, electronic device, and storage medium based on intelligent agents. By employing a first detection model trained by fusing fault samples, expert diagnostic trajectories, and fault confidence of each node, the method retrieves the first local detection evidence step by step from the starting node, performs credibility calibration on the collected first local detection evidence, and determines the closed state based on the completeness of the evidence and the situation of evidence conflict before outputting the detection conclusion. This method can trace the source layer by layer along the upstream and downstream coupling relationship of the industrial system, distinguish the original fault root cause from the abnormal signals generated downstream, improve the defect of easy misjudgment when multiple components have abnormalities at the same time, and improve the accuracy of fault diagnosis in multi-level industrial systems. Attached Figure Description

[0076] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0077] Figure 1 This is a schematic diagram of the architecture of a fault diagnosis system provided in an embodiment of this application;

[0078] Figure 2 A flowchart illustrating a fault diagnosis method based on an intelligent agent provided in an embodiment of this application;

[0079] Figure 3 A flowchart illustrating a method for determining a first closed state provided in an embodiment of this application;

[0080] Figure 4 A flowchart illustrating a training method for a first detection model provided in an embodiment of this application;

[0081] Figure 5This is a schematic diagram of the structure of a fault diagnosis device based on an intelligent agent provided in an embodiment of this application;

[0082] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0083] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0084] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0085] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of the relevant data all comply with the relevant laws, regulations, and standards of the relevant regions, have taken necessary confidentiality measures, do not violate public order and good morals, and provide corresponding operation portals for users to choose to authorize or refuse.

[0086] Furthermore, the technical solution involved in this application, which involves big data analysis of user information (including but not limited to personal biometrics, identity data, consumption data, asset data, electronic terminal operation data, etc.) and the use of artificial intelligence technology for automated decision-making, and makes decisions that have a significant impact on personal rights based on the results of automated decision-making, provides users with corresponding operation entry points for users to choose to agree to or reject the results of automated decision-making; if the user chooses to reject, the process will proceed to the expert decision-making process.

[0087] It should be noted that the fault diagnosis method, electronic device and storage medium based on intelligent agents provided in this application can be used in the field of intelligent fault diagnosis of industrial systems, and can also be used in fields other than the field of intelligent fault diagnosis of industrial systems. The application field of the fault diagnosis method, electronic device and storage medium based on intelligent agents in this application is not limited.

[0088] It should be noted that in the embodiments of this application, certain software, components, models and other existing solutions in the industry may be mentioned. These should be regarded as exemplary and are only intended to illustrate the feasibility of implementing the technical solution of this application. However, it does not mean that the applicant has used or necessarily used the solution.

[0089] Industrial systems typically consist of multi-level equipment, components, sensors, and control loops, with physical connections, upstream and downstream dependencies, media flow, or functional coupling relationships between these components. In the operation of industrial systems (such as energy, manufacturing, and aerospace), rapid fault location and root cause analysis are crucial for ensuring production safety and efficiency.

[0090] Currently, fault classification typically relies on sensor data or historical data to diagnose faults in industrial systems. However, these methods tend to identify abnormal symptoms rather than accurately depicting the propagation of anomalies between different nodes. When multiple nodes experience anomalies sequentially or simultaneously, secondary anomaly nodes are easily misidentified as root cause nodes, resulting in low accuracy in fault diagnosis.

[0091] To address the aforementioned issues, this application provides an agent-based fault diagnosis method. The method involves: acquiring first text describing a device fault; determining a starting node for fault diagnosis based on the first text; using the starting node as the starting point, acquiring a first tool to be invoked based on a pre-trained first detection model, and acquiring first local detection evidence based on the first detection model and the first tool to be invoked. The first detection model is trained based on multiple fault samples, expert diagnostic trajectories corresponding to each fault sample, and the confidence levels of multiple nodes corresponding to each fault sample; calibrating the credibility of the first local detection evidence based on the first detection model, and determining a first closure state of the calibrated evidence. The first closure state indicates whether the calibrated evidence is sufficient to pinpoint the root cause of the fault and whether there are evidence conflicts; and determining the fault diagnosis result based on the first closure state.

[0092] In the above method, a first detection model is trained by fusing fault samples, expert diagnostic trajectories, and fault confidence of each node. Starting from the initial node, the first local detection evidence is obtained by calling the tool step by step. The credibility of the collected first local detection evidence is calibrated, and the closed state is determined based on the completeness of the evidence and the situation of evidence conflict before the detection conclusion is output. This method can trace the source layer by layer along the upstream and downstream coupling relationship of the industrial system, distinguish the original fault root cause from the abnormal signals generated downstream, improve the defect of easy misjudgment when multiple components have abnormalities at the same time, and improve the accuracy of fault diagnosis of multi-level industrial systems.

[0093] First, based on Figure 1The architecture of the agent-based fault diagnosis system provided in the embodiments of this application will be described, for example, Figure 1 This is a schematic diagram of the architecture of an agent-based fault diagnosis system provided in an embodiment of this application. Please refer to [link / reference]. Figure 1 As shown, the fault diagnosis system may include: an industrial system, a first detection model, and an integrated tool environment, wherein the first detection model is integrated into the intelligent agent.

[0094] The industrial system comprises a system data source and a system topology. The system topology includes the topology of subsystem A and the topology of subsystem B. The system data source provides operational and maintenance information such as sensor data, maintenance data, and maintenance manuals. The system topology organizes the subsystems, equipment nodes, component nodes, sensor nodes, and their physical connections, upstream and downstream dependencies, and media flow directions. The industrial system also includes runtime sequence data and sensor measurement point data corresponding to each node, used to provide evidence for local fault diagnosis and operational status.

[0095] The first detection model is used to initiate multiple rounds of tool calls based on the fault phenomenon, gradually obtain local topology and diagnostic evidence, complete the root cause location of the fault and the construction of the evidence chain, so as to determine the fault diagnosis result.

[0096] The integrated tool environment encapsulates the topology query tool, node diagnostic model, and sensor data query tool into callable tools. During the fault diagnosis process, based on the call to the first detection model, it only returns partial observation content to the first detection model, supporting the step-by-step investigation and verification of the first detection model.

[0097] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0098] Figure 2 This is a flowchart illustrating a fault diagnosis method based on an intelligent agent, provided in an embodiment of this application. Please refer to... Figure 2 As shown, the method may include the following steps:

[0099] S201. Obtain the first text used to describe the equipment malfunction.

[0100] The execution subject of this application embodiment can be an electronic device or an agent-based fault diagnosis device installed in an electronic device. The agent-based fault diagnosis device can be implemented through software or through a combination of software and hardware.

[0101] In one possible implementation, the first text refers to the text containing abnormal phenomena of the carrying equipment, changes in operating status, alarm content, maintenance records, or text compiled by maintenance personnel. For example, the first text is "cooling water pump vibration value exceeds the standard".

[0102] In one possible implementation, the first text can be obtained through the input box of the monitoring interface, the alarm summary pushed by the alarm management system, the maintenance description field in the work order system, or the text result output by the speech-to-text module. After obtaining the first text, the text can be standardized to obtain the standardized first text. The standardization process includes, for example, removing duplicate spaces, unifying device name aliases, converting time expressions into standard timestamps, extracting fault objects, abnormal phenomenon words, and status change words.

[0103] S202. Based on the first text, determine the starting node for fault diagnosis.

[0104] In one possible implementation, the starting node is the first device node to be investigated, determined based on the first text and the system topology.

[0105] In one possible implementation, the starting node corresponds to a device node, component node, sensor node, or process connection node in the industrial system operating environment. The industrial system operating environment consists of devices, components, sensors, and their connection relationships, which may include physical connections, media flow direction, control dependencies, and upstream and downstream process associations.

[0106] In one possible implementation, determining the starting node for fault diagnosis may include: inputting the first text into the first detection model, performing semantic parsing on the first text, extracting keywords from the first text, retrieving the system topology stored in the industrial system, and combining the expert trajectory logic learned during the training phase of the first detection model to determine the starting node among multiple device nodes in the system topology.

[0107] The first detection model extracts keywords from the first text, which may include: performing word segmentation, entity recognition and fault semantic matching on the first text, and extracting equipment entity words, component entity words, sensor measurement point words, equipment anomaly description words, operating condition words and medium flow direction related words from the first text as keywords.

[0108] For a detailed description of the first detection model and its training method, please refer to the subsequent embodiments; they will not be repeated here.

[0109] S203. Starting from the starting node, obtain the first tool to be called based on the pre-trained first detection model, and obtain the first local detection evidence based on the first detection model and the first tool to be called. The first detection model is trained based on multiple fault samples, the expert diagnosis trajectory corresponding to each fault sample, and the confidence of each of the multiple nodes corresponding to each fault sample.

[0110] In one possible implementation, the first detection model is used to determine the first tool to be invoked, local detection evidence, and calibration confidence to obtain fault diagnosis results.

[0111] In one possible implementation, the first tool to be invoked is the tool determined by the first detection model within each step of the detection process.

[0112] In one possible implementation, the first local detection evidence refers to the raw local observation data returned by the integrated tool environment after a single step invocation of the first tool to be invoked. This data has not undergone confidence calibration, has not been integrated with historical diagnostic information, and is unenhanced basic evidence.

[0113] In one possible implementation, a fault sample refers to a complete historical fault case used to indicate the fault phenomenon; an expert diagnostic trajectory refers to the complete steps of manual troubleshooting by an operations and maintenance expert, including the tools to be called in each step, the nodes queried in each step, and the observation results of each step; multiple nodes corresponding to a fault sample refer to multiple nodes that the expert may detect in the next step in a single step; and the confidence level of each of the multiple nodes corresponding to a fault sample refers to the probability of each of the multiple nodes that the expert may detect in the next step in a single step.

[0114] In some implementations, the first tool to be invoked includes one or more of the following: a topology query tool, a node fault diagnosis tool, and a sensor data reading tool. The topology query tool is used to query the topological connections between devices, components, and sensors within the system; the node fault diagnosis tool is used to output the fault probability and local faults of nodes; and the sensor data reading tool is used to collect sensing information and time-series operational data from corresponding measurement points.

[0115] In another possible implementation, the first tool to be invoked may also include: an upstream and downstream tracing tool and a cross-system correlation query tool. The upstream and downstream tracing tool is used to query the correlation information of upstream supply equipment and downstream load equipment of the target node along the media flow direction and control dependency relationship. The cross-system correlation query tool is used to retrieve the observation information of equipment and components that have a coupling relationship between different subsystems.

[0116] In this way, by setting up multi-type hierarchical callable tools, the first detection model can autonomously match single or combined tools to retrieve local observation data step by step according to the current diagnostic status. This avoids wasting computing power by loading massive amounts of data from the entire system at once, while fully covering the information required for multi-dimensional diagnosis, including topology, raw sensor data, node fault determination, upstream and downstream propagation links, and cross-system coupling correlation.

[0117] It should be noted that within each step, one or more tools can be called in sequence, and the first tool to be called only returns local observation data.

[0118] For example, in the first step of fault diagnosis, only the topology query tool can be invoked; in the second step of fault diagnosis, only the node fault diagnosis tool can be invoked; in the third step of fault diagnosis, only the sensor data reading tool can be invoked; in the fourth step of fault diagnosis, the sensor data reading tool and the node fault diagnosis tool can be invoked in sequence; and in the fifth step of fault diagnosis, the topology query tool, the node fault diagnosis tool, and the sensor data reading tool can be invoked in sequence.

[0119] In one possible implementation, obtaining the first tool to be invoked based on the pre-trained first detection model may include: inputting the identifiers of the nodes detected in each step of the fault diagnosis process (the node detected in the first step of the fault diagnosis is the starting node) into the first detection model to obtain the first tool to be invoked.

[0120] In one possible implementation, the first detection model can send an instruction to the integrated tool environment carrying the identifier of the first tool to be invoked and the identifier of the node being detected in the current step. The integrated tool environment can access the system data source and system structure topology in the industrial system and return the local observation raw data corresponding to the instruction according to the instruction. This local observation raw data is the first local detection evidence.

[0121] In some implementations, first local detection evidence is obtained based on a first detection model and a first tool to be invoked, including:

[0122] When the first tool to be invoked includes a topology query tool, the local observation topology and upstream and downstream related node information corresponding to the node are determined based on the first tool to be invoked, and the first local detection evidence is determined based on the local observation topology and upstream and downstream related node information.

[0123] When the first tool to be invoked includes a node fault diagnosis tool, the fault judgment result of the node is determined based on the first tool to be invoked, and the first local detection evidence is determined based on the fault judgment result;

[0124] When the first tool to be invoked includes a sensor data reading tool, the sensing time series operation data of the target measurement point corresponding to the time period is determined based on the first tool to be invoked, and the first local detection evidence is determined based on the sensing time series operation data.

[0125] In another possible implementation, when the first tool to be invoked includes a sensor data reading tool and a node fault diagnosis tool in sequence, the sensor timing data and the fault judgment result of the node corresponding to the target measurement point are determined in sequence based on the first tool to be invoked, and the first local detection evidence is determined based on the sensor timing data and the fault judgment result of the node.

[0126] In another possible implementation, when the first tool to be invoked sequentially includes a topology query tool, a sensor data reading tool, and a node fault diagnosis tool, based on the first tool to be invoked, the local observation topology corresponding to the node, the upstream and downstream related node information, the sensing time series operation data of the target measuring point for the corresponding time period, and the fault judgment result of the node are determined sequentially. Based on the local observation topology, the upstream and downstream related node information, the sensing time series operation data, and the fault judgment result, the first local detection evidence is determined.

[0127] Specifically, based on the first tool to be invoked, determining the local observation topology and upstream and downstream related node information corresponding to the node may include: passing the unique identifier of the current node to be detected and the query depth threshold to the topology query tool; the tool traverses the system topology structure, extracts the direct upstream supply node, downstream load node, medium flow direction, physical connection edge, and coupling control relationship of the node, and outputs the local topology primitive set and the list of upstream and downstream nodes.

[0128] Based on the first tool to be invoked, the fault judgment result of the node can be determined, which may include: passing the target device node number and fault time interval to the node fault diagnosis tool, calling the node's dedicated built-in fault diagnosis model to calculate, and outputting the original occurrence probability, fault category label, and fault trigger judgment threshold corresponding to various faults of the node.

[0129] Based on the first tool to be invoked, the sensor timing data for the corresponding time period of the target measurement point can be determined. This may include: passing the measurement point identifier, fault start and end timestamps to the sensor data reading tool, pulling the raw sensor values ​​continuously collected during that time period, and simultaneously extracting the maximum value, minimum value, mean, variance, and out-of-limit time markers of the data.

[0130] Specifically, determining the first local detection evidence based on the local observation topology and upstream and downstream related node information may include: performing structured encapsulation of local topology node numbers, media flow direction, and upstream and downstream device affiliation relationships to generate an indexable topology observation message as the first local detection evidence.

[0131] Based on the fault judgment results, the first local detection evidence is determined, which may include: the original fault probability of the storage node, the fault type label, and the local fault triggering conditions. The complete fault judgment array is then encapsulated as the first local detection evidence.

[0132] Based on the sensor time-series operation data, the first local detection evidence can be determined, which may include: extracting continuous sensor values, extreme values, and fluctuation ranges during the fault period, and then encapsulating them into the first local detection evidence after standardizing the time-series data format.

[0133] Based on the sensor timing data and the fault judgment results of the nodes, the first local detection evidence can be determined, which may include: binding and associating standardized sensor timing data with the original fault judgment array of the nodes, synchronously storing the corresponding mapping relationship between the measurement points and the fault results, and integrating and encapsulating them into a composite first local detection evidence.

[0134] Based on local observation topology, upstream and downstream related node information, sensor timing data, and fault judgment results, the first local detection evidence can be determined. This can include: establishing a correlation index between structured topology messages, standardized sensor timing data, and node fault judgment arrays, and uniformly packaging them to generate multi-dimensional composite first local detection evidence.

[0135] In this way, by supporting two evidence collection modes—single-tool and multi-tool sequential combination—different types of observation information are standardized and structured and encapsulated separately, and single-dimensional or multi-dimensional composite original local detection evidence is generated as needed. This can completely preserve the original observation information of the entire diagnostic dimension, which is convenient for subsequent evidence calibration, conflict comparison and fault propagation path tracing.

[0136] S204. Based on the first detection model, the credibility of the first local detection evidence is calibrated, and the first closure state of the calibrated evidence is determined. The first closure state is used to indicate whether the calibrated evidence is sufficient to lock down the root cause of the fault and whether there is evidence conflict.

[0137] In one possible implementation, credibility calibration refers to combining the calibration function built into the first detection model to make reliability corrections to the first local detection evidence, so as to eliminate data deviations caused by sensor drift and node model false alarms, and output the fault distribution of the nodes and the reliability of the evidence after calibration.

[0138] In one possible implementation, when the first closed state indicates that the evidence after calibration is sufficient to pinpoint the root cause of the fault and there is no conflict of evidence, it means that the evidence is closed, that is, there is a fault node with leading confidence among the current calibration evidence, and the observation data returned by each tool are not contradictory.

[0139] In one possible implementation, when the first closed state indicates that the evidence after calibration is insufficient to pinpoint the root cause of the fault, it means that the evidence is not closed, that is, the confidence values ​​of each candidate fault node are close, there is no clear leading root cause, the existing evidence is insufficient to complete the root cause determination, and it is necessary to continue to retrieve more observation data from the tool to be called.

[0140] In one possible implementation, when the first closed state is used to indicate the existence of a conflict of evidence, it indicates that the observation conclusions output by different tools for the same device are contradictory, the credibility of the existing evidence is insufficient, and a new verification tool is needed for secondary verification.

[0141] For example, a scenario where the evidence is closed could be: the confidence level of the water pump root cause after calibration is 86%, and the confidence level of the second-best candidate chiller unit is 12%, with no conflict, and the result can be directly output at this time; a scenario where the evidence is not closed could be: the confidence level of the water pump is 42%, and the confidence level of the other devices is 30%, with no obvious leading root cause, and the tool to be called needs to be retrieved; a scenario where the evidence is conflicting could be: the water pump diagnostic tool outputs a high failure probability, but the water pump vibration sensor reading is completely normal, the two types of evidence contradict each other, and the tool to be called needs to be retrieved.

[0142] In one possible implementation, the calibrated evidence may include the calibrated node state distribution and the reliability of the calibrated evidence. The calibrated node state distribution is used to indicate the probability range of various faults occurring in the currently detected node after correction. The calibrated node state distribution refers to the standardized probability array of each fault type corresponding to the node after eliminating sensor drift and model-inherent false alarm bias. The reliability of the calibrated evidence is used to indicate the degree of trustworthiness of the observation results of this tool.

[0143] In one possible implementation, the credibility calibration of the first local detection evidence based on the first detection model may include: inputting the first local detection evidence into a preset calibration function to obtain the calibrated node state distribution, and inputting the first local detection evidence into a preset evidence reliability calculation function to obtain the reliability of the calibrated evidence.

[0144] For example, when the first tools to be invoked sequentially include a sensor data reading tool and a node fault diagnosis tool, the evidence of calibration can be represented as follows:

[0145]

[0146] in, This indicates the node being detected in the current step. This indicates the first local test evidence. This refers to sensor data acquired through a sensor data reading tool. This represents the node fault model built into the node fault diagnosis tool. This indicates the preset calibration function. This represents a pre-defined function for calculating the reliability of evidence. This represents the calibrated node state distribution. This indicates the reliability of the calibrated evidence.

[0147] in, The temperature calibration parameters are calculated using a validation set and are used to correct overfitting or undercalibration of the model. These temperature calibration parameters are fixed correction coefficients obtained by fine-tuning the calibration function using a validation set labeled with real faults. They are used to scale and offset the original probability values ​​output by the node fault model, offsetting the overestimation or underestimation of probabilities caused by long-term sensor drift and model training. This ensures that the probabilities of various faults closely match the actual fault patterns of the equipment, improving the accuracy of the node state distribution after calibration. It should be noted that the method for obtaining the validation set can be found in subsequent embodiments and will not be elaborated here.

[0148] in, ,in This represents the maximum predicted probability after calibration (representing the model's current output confidence). Give the inherent reliability score (such as historical accuracy) of the diagnostic tool for this node on the validation set. , which is the balancing weight coefficient between the two.

[0149] Alternatively, the maximum predicted probability after calibration can be directly determined as the reliability of the calibrated evidence.

[0150] It should be noted that the method for determining the first closed state of the calibrated evidence can be referred to in subsequent embodiments, and will not be repeated here.

[0151] S205. Based on the first closed state, determine the fault diagnosis result.

[0152] In one possible implementation, the fault diagnosis result refers to the final output of the detection, which may include the root cause node of the fault, the fault propagation path across devices, and a complete traceability evidence chain.

[0153] In some implementations, the fault diagnosis result is determined based on the first closed state, including:

[0154] When the first closed state indicates that the calibrated evidence is sufficient to pinpoint the root cause of the fault and there is no evidence conflict, the fault diagnosis result, including the root cause node, the fault propagation path across devices, and the traceable evidence chain, is determined based on the first local detection evidence. When the first closed state indicates that the calibrated evidence is insufficient to pinpoint the root cause of the fault or there is evidence conflict, a second tool to be invoked is obtained, the second tool to be invoked is invoked to obtain the second local detection evidence, and the fault diagnosis result is determined based on the first local detection evidence and the second local detection evidence.

[0155] In one possible implementation, the second tool to be invoked refers to a new round of query tools regenerated by the first detection model when the evidence following the first closed state indication determined in this step is insufficient to pinpoint the root cause of the fault or when there are conflicting evidence. This tool is used to supplement the collection of second local detection evidence.

[0156] In one possible implementation, based on the first local detection evidence, the fault diagnosis results, including the root cause node, the cross-device fault propagation path, and the traceability evidence chain, are determined. This may include: integrating the calibrated node state distribution and the reliability of each node's evidence; for each node, summing the highest fault confidence value indicated by the calibrated node state distribution with the corresponding evidence reliability to obtain a comprehensive fault score, thereby selecting the node with the highest comprehensive fault score as the root cause node; combining the local observation topology, upstream and downstream media flow direction, and equipment control correlation to sort out the complete link of the fault propagation from the root cause to the downstream equipment as the cross-device fault propagation path; and connecting the identifier of the first tool to be called used in each step of fault detection, the queried node, multiple first local detection evidences determined in the process of each step of node detection, the calibrated node state distribution, and the reliability of each node's evidence to form a complete traceability evidence chain with a time-series index.

[0157] In one possible implementation, obtaining a second tool to be invoked, invoking the second tool to obtain second local detection evidence, and determining the fault diagnosis result based on the first and second local detection evidence can include: determining the second tool to be invoked based on the first detection model, and sending an instruction to the second tool to be invoked to collect newly added second local detection evidence to the integrated tool environment; then, based on the first detection model, calibrating the credibility of the second local detection evidence, and determining the second closure state of the calibrated evidence; when the second closure state indicates that the calibrated evidence is sufficient to lock the root cause of the fault and there is no evidence conflict, fusing the first and second local detection evidence to determine the fault diagnosis result; when the second closure state indicates that the calibrated evidence is insufficient to lock the root cause of the fault or there is evidence conflict, continuing to acquire newly added local detection evidence until the closure state of the newly added local detection evidence indicates that the evidence is sufficient to lock the root cause of the fault and there is no evidence conflict, thus determining the fault diagnosis result.

[0158] In this embodiment, a first text describing a device fault is obtained; based on the first text, a starting node for fault diagnosis is determined; starting from the starting node, a first tool to be invoked is obtained based on a pre-trained first detection model, and first local detection evidence is obtained based on the first detection model and the first tool to be invoked. The first detection model is trained based on multiple fault samples, expert diagnosis trajectories corresponding to each fault sample, and the confidence levels of multiple nodes corresponding to each fault sample; based on the first detection model, the credibility of the first local detection evidence is calibrated, and a first closure state of the calibrated evidence is determined. The first closure state is used to indicate whether the calibrated evidence is sufficient to lock down the root cause of the fault and whether there is evidence conflict; based on the first closure state, the fault diagnosis result is determined.

[0159] In the above method, a first detection model is trained by fusing fault samples, expert diagnostic trajectories, and fault confidence of each node. Starting from the initial node, the first local detection evidence is obtained by calling the tool step by step. The credibility of the collected first local detection evidence is calibrated, and the closed state is determined based on the completeness of the evidence and the situation of evidence conflict before the detection conclusion is output. This method can trace the source layer by layer along the upstream and downstream coupling relationship of the industrial system, distinguish the original fault root cause from the abnormal signals generated downstream, improve the defect of easy misjudgment when multiple components have abnormalities at the same time, and improve the accuracy of fault diagnosis of multi-level industrial systems.

[0160] Next, combined Figure 3 The method for determining the first closed state is described in detail. For example, Figure 3 A flowchart illustrating a method for determining a first closed state provided in an embodiment of this application is shown below. Figure 3 As shown, the method includes:

[0161] S301. Based on the calibrated evidence, determine the first candidate root cause set and the first candidate root cause evidence distribution, wherein the first candidate root cause evidence distribution is used to indicate the degree of fault support of the calibrated evidence for each candidate root cause in the first candidate root cause set.

[0162] In one possible implementation, the first candidate root cause set refers to the set of all nodes with the potential for failure, selected based on the semantics of the fault text and the upstream and downstream topology of the system.

[0163] In one possible implementation, the first candidate root cause evidence distribution refers to the fault support strength of the calibration evidence for each candidate root cause within the first candidate root cause set in this detection step. The value is obtained from local observation and topology upstream and downstream compatibility calculation, and is used to indicate the corroborating strength of the calibration evidence for each node fault in this detection step.

[0164] In one possible implementation, determining the first candidate root cause set based on calibrated evidence may include: obtaining the node state distribution and evidence reliability of the nodes detected in the current step and the nodes detected in steps prior to the current step; screening devices, components, and sensor nodes in the calibrated evidence whose fault probability indicated by the node distribution state is greater than or equal to a first threshold and which have a physical connection, medium flow direction, or control coupling relationship with the fault phenomenon described in the first text; and obtaining the first candidate root cause set after deduplication.

[0165] In one possible implementation, determining the first candidate root cause evidence distribution based on calibrated evidence may include: traversing each candidate root cause in the candidate root cause set of the next detection step, calculating the propagation compatibility between the node detected in the current step and each candidate root cause in the visible local topology, combining the support strength of each candidate root cause with the calibrated node state distribution, and then introducing a preset smoothing term for normalization to obtain the fault support probability corresponding to each candidate root cause, combining them to form the first candidate root cause evidence distribution. For example, the fault support probability corresponding to each candidate root cause can be expressed as:

[0166]

[0167] Where c represents each candidate root cause in the first candidate root cause set. Let t represent the probability that the evidence after calibration at step t supports the failure of candidate root cause c, and t be a preset smoothing term (which can be a very small positive number, for example, 10). -6 (), used to avoid the numerator or denominator being zero, This indicates the node v being detected in the current step. t The compatibility of the candidate root cause c with fault propagation in the local topology, with a value of 0 or 1. A value of 1 indicates that the candidate root cause is related to the node detected in the current step, and that the fault is more likely to propagate from the candidate root cause c to the node v detected in the current step. t , Indicates the calibrated node distribution status Fault support strength for candidate root cause c Indicates the calibrated node distribution status For the fault support strength of each candidate root cause in the (t+1)th detection step, C t+1 Let represent the set of candidate root causes at step t+1, and let c' represent each candidate root cause in the set of candidate root causes at step t+1. The set of candidate root causes at step t+1 can be obtained by performing structured diagnostic actions on the set of candidate root causes at step t.

[0168] For example, It can be represented as: ,in, Indicates an indicator function, This indicates whether there exists a path in the system topology that points from candidate root cause c to node v being detected in the current step. t A directed path, if it exists, =1; if it does not exist, It is 0.

[0169] For example, It can be represented as: ,in, The fault probability calibration matrix is ​​formed by the actual probability of various faults occurring in historical data, and T represents the transpose.

[0170] It should be noted that in the current detection step, before determining the first candidate root cause set and the first candidate root cause evidence distribution, a structured diagnostic action can be performed on the candidate root cause set of the current detection step to obtain the candidate root cause set and candidate root cause evidence distribution for the next step, and based on this, the first candidate root cause set and the first candidate root cause evidence distribution are determined.

[0171] S302. Align the first candidate root cause set and determine the first evidence credibility information after alignment. The first evidence credibility information is used to indicate the failure confidence level accumulated by each candidate root cause in the first candidate root cause set.

[0172] In one possible implementation, aligning the first candidate root cause set means matching and merging the nodes corresponding to the newly added evidence in this round with the candidate root cause nodes retained from historical diagnoses, eliminating duplicate nodes, and supplementing the newly added potential fault nodes in this round, so that the dimensions of the root cause sets before and after are unified, and new and old evidence can be fused and calculated.

[0173] In one possible implementation, the first evidence confidence information refers to the original fault confidence of each candidate root cause obtained based on the accumulation of historical evidence over multiple rounds after root cause alignment is completed. It is used to indicate the inherent judgment of each node fault before the fusion of newly added calibrated evidence.

[0174] In one possible implementation, determining the credible information of the first evidence after alignment may include: retrieving the candidate root cause set and candidate root cause evidence distribution cached in the previous round of diagnostic process; performing dimensional mapping based on the aligned first candidate root cause set: for nodes that exist in both the historical candidate root cause set and the current round's first candidate root cause set, directly reusing the historically accumulated fault confidence and adding the fault confidence to the fault confidence array; for newly added candidate root cause nodes in the current round, initializing their fault confidence to a preset global average device fault probability (e.g., 2%), and adding the fault confidence to the fault confidence array; for irrelevant nodes that existed historically but have been removed in the current round, directly removing them from the fault confidence array; finally, outputting a fault confidence array that corresponds one-to-one with the aligned first candidate root cause set, and determining the fault confidence array as the credible information of the first evidence.

[0175] S303. Based on the first evidence credibility information and the first candidate root cause evidence distribution, determine the second evidence credibility information. The second evidence credibility information is used to indicate the fault confidence of each candidate root cause obtained after the evidence is fused and calibrated.

[0176] In one possible implementation, the second evidence credible information refers to the new fault confidence level after weighted fusion and normalization update of the first evidence credible information using the distribution and reliability of the first candidate root cause evidence in this round. It is the latest node fault determination score after superimposing the observation evidence in this round.

[0177] In one possible implementation, determining the second evidence credibility information based on the first evidence credibility information and the first candidate root cause evidence distribution may include: performing a weighted fusion operation on the first evidence credibility information and the first candidate root cause evidence distribution element by element in the previous detection step, and then performing a global normalization process on the fused combination to map the confidence of each candidate root cause to the [0,1] interval and combine them into 1. The normalized array is the second evidence credibility information.

[0178] For example, the second evidence credibility information in the (t+1)th detection step can be represented as:

[0179]

[0180] Where Norm represents normalization. e represents the first piece of credible evidence in the t-th detection step. t This represents the distribution of candidate root cause evidence in the t-th detection step. This indicates the reliability of the evidence in the t-th detection step.

[0181] S304. Based on the credible information of the second piece of evidence, determine the first closed state.

[0182] In one possible implementation, determining the first closure state based on the second evidence credibility information may include: first, determining the candidate root cause with the highest support based on the second evidence credibility information of the next detection step; then, determining the evidence credibility information interval between the best and second-best candidate root causes based on the candidate root cause; determining the normalized evidence credibility information entropy; determining the reliability-weighted evidence conflict strength based on the second evidence credibility information; and finally, determining the first evidence closure score based on the evidence credibility information interval, the reliability-weighted evidence conflict strength, and the normalized evidence credibility information entropy, and determining the first closure state based on the first evidence closure score and the reliability-weighted evidence conflict strength.

[0183] It should be noted that in the current detection step, before determining the first closure state, a structured diagnostic action can be performed on the candidate root cause set of the current detection step to obtain the candidate root cause set and candidate root cause evidence distribution for the next step. This allows the fault confidence of each candidate root cause in the next detection step to be obtained, and based on this, the first closure state of the current step can be determined.

[0184] For example, the candidate root cause with the highest support can be represented as:

[0185]

[0186] Where argmax represents the index of the element corresponding to the maximum value in the array, used to select the optimal candidate root cause with the highest confidence from the second piece of evidence. C t Let represent the set of candidate root causes in step t. This represents the failure confidence of candidate root cause c in the next detection step (i.e., the second piece of evidence credibility information for the next detection step). Let represent the set of candidate root causes at step t+1.

[0187] For example, the interval of credible evidence information can be expressed as:

[0188]

[0189] in, This represents the candidate root cause with the highest support at step t+1. Fault confidence, C t This represents the set of first candidate root causes.

[0190] For example, the normalized evidence credibility information entropy can be expressed as:

[0191]

[0192] Where log represents the logarithmic operation. Let represent the base entropy of the first credible evidence information at step t+1.

[0193] For example, the reliability-weighted strength of evidence conflict can be expressed as:

[0194]

[0195] Among them, D JS Indicates JS divergence, e represents the credibility of the second piece of evidence. t This represents the sum of the failure support probabilities corresponding to each candidate root cause.

[0196] For example, the first evidence closure score can be expressed as:

[0197]

[0198] in, and Both are weighting coefficients, and their values ​​are in the range [0,1]. For example, It is 0.5. It is 0.3.

[0199] In one possible implementation, determining the first closure state based on the first evidence closure score and the evidence conflict strength weighted by reliability may include:

[0200] When the first evidence closure score is greater than or equal to a preset closure threshold and the evidence conflict intensity is less than or equal to a preset conflict threshold, the first closure state is determined to indicate that the calibrated evidence is sufficient to pinpoint the root cause of the fault and there is no evidence conflict (evidence is closed); when the first evidence closure score is less than a preset closure threshold and the evidence conflict intensity is less than or equal to a preset conflict threshold, the first closure state is determined to indicate that the calibrated evidence is insufficient to pinpoint the root cause of the fault (evidence is not closed); when the evidence conflict intensity is greater than a preset conflict threshold, the first closure state is determined to indicate that the calibrated evidence has evidence conflict (evidence conflict).

[0201] For example, if the preset closure threshold is 0.6 and the preset conflict threshold is 0.4, then when the first evidence closure score is between [0.6, 1.0] and the conflict intensity is less than or equal to 0.4, the first closure state is determined to indicate that the calibrated evidence is sufficient to lock the root cause of the fault and there is no evidence conflict.

[0202] When the first evidence closure score is in the range of [-1.0, 0.6) and the conflict strength is less than or equal to 0.4, the first closure state is determined to indicate that the calibrated evidence is insufficient to lock down the evidence state of the root cause of the fault.

[0203] If the conflict intensity is greater than 0.4, the first closure state is determined to indicate that there is a conflict of evidence after calibration.

[0204] In this embodiment, based on the calibrated evidence, a first candidate root cause set and a first candidate root cause evidence distribution are determined. The first candidate root cause evidence distribution is used to indicate the degree of fault support of the calibrated evidence for each candidate root cause in the first candidate root cause set. The first candidate root cause set is aligned, and the aligned first evidence credibility information is determined. The first evidence credibility information is used to indicate the accumulated fault confidence of each candidate root cause in the first candidate root cause set. Based on the first evidence credibility information and the first candidate root cause evidence distribution, second evidence credibility information is determined. The second evidence credibility information is used to indicate the fault confidence of each candidate root cause updated after fusing the calibrated evidence. Based on the first evidence credibility information and the second evidence credibility information, a first closure state is determined.

[0205] In the above method, candidate root causes and fault support levels are determined by combining calibrated evidence, historical fault confidence is obtained by aligning the candidate root cause set, historical confidence and fault support levels are merged to update fault confidence, and the first closed state is determined by combining the new and old fault confidence. This method can make full use of the calibration evidence in each round and the fault judgment accumulated in history, objectively distinguish between three types of working conditions: sufficient evidence, insufficient evidence, and conflicting evidence, and improve the accuracy of fault diagnosis in industrial systems.

[0206] Next, combined Figure 4 The training method of the first detection model is described in detail. For example, Figure 4 A flowchart illustrating a training method for a first detection model provided in an embodiment of this application is shown below. Figure 4 As shown, the method includes:

[0207] S401. Obtain multiple fault samples, multiple first expert diagnostic trajectories corresponding to each fault sample, and the confidence scores of multiple nodes corresponding to each fault sample. The first expert diagnostic trajectories sequentially include the nodes to be investigated and the first tool to be called when investigating each node.

[0208] In one possible implementation, obtaining multiple fault samples, multiple first-expert diagnostic trajectories corresponding to each fault sample, and the confidence levels of multiple nodes corresponding to each fault sample can include: pulling complete historical fault records from the industrial system database as fault samples; operation and maintenance experts independently complete the entire process of manual investigation for a single fault sample, recording each step of the investigation node and the tools to be called by the current node in time sequence, generating multiple first-expert diagnostic trajectories; experts, combining fault phenomena, equipment media flow direction and control coupling relationship, label the fault probability of each topology node within the single-step investigation range with a value range of [0,1], obtaining the confidence level of each node, and establishing a unique binding index for the fault samples, corresponding multiple expert trajectories, and node confidence levels and storing them uniformly in the training dataset.

[0209] S402. Based on multiple fault samples, multiple first expert diagnostic trajectories, and the confidence levels of multiple nodes corresponding to each fault sample, determine the second detection model.

[0210] In some implementations, a second detection model is determined based on multiple fault samples, multiple first-expert diagnostic trajectories, and the confidence levels of multiple nodes corresponding to each fault sample, including:

[0211] Based on multiple fault samples, multiple first-expert diagnostic trajectories, and the confidence scores of multiple nodes corresponding to each fault sample, structured samples are determined. The structured samples are used to indicate the local topology, single-step diagnostic action, single-step tool observation data, and subsequent detection status updated after the action is executed for a single detection. Supervised fine-tuning training is performed based on the structured samples to obtain the second detection model.

[0212] In one possible implementation, the first expert diagnostic trajectory includes diagnostic action primitives and specific tool calls in several steps. The diagnostic action primitives refer to the basic action identifiers used to standardize and abstractly classify fault investigation behaviors, uniformly distinguishing different diagnostic intentions. These primitives include: symptom grounding, topology expansion, propagation tracing, candidate verification, and diagnostic termination. Symptom grounding refers to locking the initial investigation starting node based on the fault description; topology expansion refers to querying newly added related topologies upstream and downstream of the current node to expand the observation range; propagation tracing refers to tracing the source of fault propagation backward along the medium and control coupling link; candidate verification refers to calling tools to verify the status of the current high-confidence potential fault node; diagnostic termination refers to stopping further tool calls to output diagnostic conclusions when the evidence meets the closure condition; and specific tool calls refer to matching the diagnostic action primitives to executable query, read, and verification tool instructions in the integrated tool environment, including tool identifiers, query nodes, and parameter information.

[0213] In one possible implementation, structured samples refer to single-step standardized training units decomposed from multiple first-expert diagnostic trajectories, including: input information, current local topology, single-step diagnostic action, tool observation data, and the propagation front path after the action is completed, used for supervised fine-tuning of the input. The current input information includes: fault samples, currently observed local topology, tools invoked in steps prior to the current step, and historical observation evidence. For example, the input information can be represented as:

[0214]

[0215] Where d0 represents the fault sample, This indicates the tools and historical observational evidence invoked in steps preceding the current step. This represents the local topology that has been observed so far.

[0216] In one possible implementation, supervised fine-tuning training refers to using expert-generated structured samples as standard answers to adjust the internal parameters of the second detection model in reverse, so that the second detection model learns the node selection and tool calling logic of the human standard.

[0217] In one possible implementation, the second detection model is an intermediate base model obtained only through supervised fine-tuning, which merely replicates the expert screening steps without optimizing tool resource consumption or strengthening optimization by incorporating evidence closure states.

[0218] In one possible implementation, based on multiple fault samples, multiple first-expert diagnostic trajectories, and the confidence levels of multiple nodes corresponding to each fault sample, a structured sample is determined. This can include: disassembling each first-expert diagnostic trajectory line by line and replaying the trajectory, that is, reproducing each step of the operation in the integrated tool environment according to the tool call order of the experts, and reproducing the current local topology, symptom grounding node, candidate root cause set, and propagation front path according to the return results of each tool to construct the visible diagnostic state. In addition, the diagnostic action primitives and specific tool calls in the tool return results are extracted to form the structured diagnostic actions. The visible diagnostic states and structured diagnostic actions of multiple steps of multiple first-expert diagnostic trajectories are integrated to obtain the structured sample.

[0219] Here, the current local topology refers to the local system structure of each node and its connection edges that has been obtained through tool queries at step t of the diagnosis; the symptom grounding node refers to the device, component, and sensor node that matches the fault description and serves as the initial starting point for this diagnosis; the candidate root cause set is a summary of all potential fault nodes with fault probability and the possibility of fault propagation at the current step; the propagation front path refers to the topology link where the fault can continue to be traced outwards and has not yet been verified; the diagnostic action primitive is a standardized abstract diagnostic intent classification identifier; and the specific tool call is the executable tool query instruction corresponding to the primitive.

[0220] For example, the visible diagnostic status can be represented as:

[0221]

[0222] Where t represents the sequence number of the step during reproduction, S represents the currently observed local topology. t Indicates the symptom grounding node, C1 t Let P represent the set of candidate root causes. t This indicates the frontier path of dissemination.

[0223] For example, a structured diagnostic action can be represented as:

[0224]

[0225] Among them, ut Indicates a diagnostic action primitive, a t This indicates a specific tool call.

[0226] For example, a structured sample can be represented as:

[0227]

[0228] in, This represents the input information for the t-th step in the first expert diagnostic trajectory corresponding to the i-th fault sample. This represents the visible diagnostic status at the t-th step in the first expert diagnostic trajectory corresponding to the i-th fault sample. This represents the structured diagnostic action at step t in the first expert diagnostic trajectory corresponding to the i-th fault sample. Let represent the set of symptom grounding nodes and propagation front paths for the t-th step in the first expert diagnosis trajectory corresponding to the i-th fault sample. This represents the propagation frontier path of the t-th step in the first expert diagnosis trajectory corresponding to the i-th fault sample.

[0229] In one possible implementation, supervised fine-tuning training is performed based on structured samples to obtain a second detection model. This may include: initializing a large model with tool-calling capabilities as initial weights, taking multiple fault samples as inputs and structured diagnostic actions as outputs, and using the structured samples as ground truths and iteratively training with a target loss function; after each training iteration, the action matching rate and topology reconstruction accuracy can be verified using a validation set; if there is no improvement in the metrics for the first consecutive rounds (e.g., 10 times), the training is terminated, and the converged topology internalization agent, i.e., the second detection model, is output.

[0230] The method for determining the validation set includes: extracting it from fault samples and expert diagnostic trajectory data in a stratified manner according to a fixed ratio (e.g., dividing the training set and validation set into a 9:1 ratio). The stratification rule ensures that the distribution of fault types, equipment subsystems, and fault severity in the validation set is consistent with that in the training set. After extraction, the validation set is isolated and does not participate in model parameter updates. It is only used to verify the model action matching rate and topology restoration accuracy after each round of training, and to objectively measure the model's generalization ability.

[0231] In one possible implementation, the training objective is to enable the agent (the second detection model) to recover the visible diagnostic state given input information, generate structured diagnostic actions, and update the propagation front path based on the set of symptom grounding nodes and propagation front paths. For example, the objective loss function can be expressed as:

[0232]

[0233] Where t represents the step number, Indicates the diagnostic status monitoring loss. This indicates the structured action supervision loss. Indicates the subsequent state monitoring loss. The weights correspond to the pre-defined diagnostic state supervision loss. The weights correspond to the pre-defined structured action supervision loss. The weights are the corresponding weights for the pre-defined subsequent state supervision loss.

[0234] The diagnostic state supervision loss is used to constrain and train the model to accurately recover the current visible diagnostic state. For example, for the t-th step, the diagnostic state supervision loss can be expressed as:

[0235]

[0236] in, Indicates including parameters The output probability of the second detection model, The current visible diagnostic status of the expert trajectory offline replay (including observed topology, symptom grounding nodes, candidate root causes and exploration frontiers, context and diagnostic actions, and diagnostic result information); This represents the input information for the t-th step.

[0237] The structured action supervision loss is used to generate tool invocation actions with explicit industrial diagnostic semantics based on the current state. For example, for the t-th step, the structured action supervision loss can be expressed as:

[0238]

[0239] in, This refers to the structured diagnostic action labels for experts, composed of diagnostic intent primitives (such as tracking and verification) and specific tool calls and parameter combinations. This indicates that the expert's trajectory was replayed offline in the first... Step-by-step discrete diagnostic action selection This indicates the specific tool selected and invoked at step t for offline replay of the expert trajectory.

[0240] The successor state supervision loss is used to train the model's ability to correctly update the next diagnostic state after performing an action and receiving environmental feedback. For example, for the t-th step, the successor state supervision loss can be expressed as:

[0241]

[0242] in, This refers to the local observation data (such as new topology or prediction results) returned by the environment after the current tool is executed. The visible diagnostic state is evolved to the next detection step after incorporating new observational evidence.

[0243] It should be noted that before determining the subsequent state supervision loss of the current detection step, a structured diagnostic action can be performed on the candidate root cause set of the current detection step to obtain the visible diagnostic state of the next detection step.

[0244] S403. Based on multiple fault samples and the second detection model, determine the first detection model.

[0245] In some implementations, a first detection model is determined based on multiple fault samples and a second detection model, including:

[0246] For each fault sample, a fault diagnosis trajectory is determined based on the second detection model. For each fault diagnosis trajectory, tool usage information is determined, which indicates the resource consumption cost incurred by calling tools during the determination of the fault diagnosis trajectory. For each fault diagnosis trajectory, credibility calibration is performed on each observation evidence obtained using multiple tools to be called within the trajectory, and the second candidate root cause set and the second candidate root cause evidence distribution of the calibrated evidence are determined. For each fault diagnosis trajectory, a second closure state is determined based on the second candidate root cause set and the second candidate root cause evidence distribution. Based on the second closure state of each fault diagnosis trajectory and the tool usage information, the second detection model is iteratively optimized to obtain the first detection model.

[0247] In one possible implementation, the fault diagnosis trajectory refers to the fully automated troubleshooting process that is automatically generated after each fault sample is input into the second detection model, without any human intervention.

[0248] In one possible implementation, tool usage information is a quantitative record of the total cost of resources such as computing power and database bandwidth consumed by each tool to be invoked within a fault diagnosis trajectory, which is used to optimize model constraints.

[0249] In one possible implementation, performing credibility calibration on each observation evidence obtained using multiple tools to be invoked within the trajectory refers to correcting the deviation of the original observation data returned by the tools, eliminating sensor drift and model-inherent false alarms, and outputting the node failure probability and evidence reliability after calibration.

[0250] In one possible implementation, the second candidate root cause set is a set of all devices, components, and sensor nodes that are likely to fail, selected from the calibrated evidence.

[0251] In one possible implementation, the second candidate root cause evidence distribution is used to quantify the fault support strength of the current round of calibration evidence for each node in the set, and the sum of the values ​​is 1.

[0252] In one possible implementation, the second closure state refers to the diagnostic termination marker determined based on the fusion of the candidate root cause set and evidence distribution to establish confidence in the new and old faults. It is categorized into three types: evidence closed, evidence not closed, and evidence conflict. When the second closure state is evidence closed, it indicates that the calibrated evidence is sufficient to pinpoint the root cause of the fault and there is no evidence conflict. When the second closure state is evidence not closed, it indicates that the calibrated evidence is insufficient to pinpoint the root cause of the fault. When the second closure state is evidence conflict, it indicates that the calibrated evidence contains evidence conflicts.

[0253] In one possible implementation, iterative optimization refers to updating the parameters of the second detection model cyclically with the dual optimization objectives of reducing tool resource consumption and accelerating the attainment of a closed evidence state, until the model converges and the first detection model is output.

[0254] In one possible implementation, for each fault sample, the fault diagnosis trajectory is determined based on the second detection model. This may include: for each fault sample, inputting the fault sample into the second detection model to obtain the fault diagnosis trajectory.

[0255] In one possible implementation, determining tool usage information for each fault diagnosis trajectory may include: pre-setting computational power and bandwidth consumption coefficients for each tool call; for multiple fault diagnosis trajectories, traversing all tool call steps for each trajectory, summing the bandwidth consumption coefficients of the tools used in each step to obtain the total cost, and then normalizing the result to obtain the tool usage information for this fault diagnosis trajectory. For example, the tool usage information can take values ​​in the range [0.1, 0.5] based on the complexity of the tool to be called; for instance, 0.1 for low topology lookup costs and 0.5 for high fault prediction costs. This embodiment does not limit this.

[0256] In one possible implementation, the specific implementation method for performing credibility calibration on each observation evidence obtained using multiple tools to be invoked within the trajectory can refer to the specific implementation method for performing credibility calibration on the first local detection evidence in the aforementioned embodiments. The specific implementation method for determining the second candidate root cause set can refer to the specific implementation method for determining the first candidate root cause set in the aforementioned embodiments. The specific implementation method for determining the distribution of the second candidate root cause evidence can refer to the specific implementation method for determining the distribution of the first candidate root cause evidence in the aforementioned embodiments. The specific implementation method for determining the second closure state can refer to the specific implementation method for determining the first closure state in the aforementioned embodiments. These details will not be elaborated here.

[0257] In one possible implementation, based on the second closed state of each fault diagnosis trajectory and tool usage information, the second detection model is iteratively optimized to obtain the first detection model. This may include: for each fault diagnosis trajectory, inputting the second closed state into a preset potential energy function to obtain the evidence potential energy return; then constructing an optimized return for a single trajectory based on the evidence potential energy return, tool usage information, and the correctness of the trajectory's final root cause determination; using the maximization of the average optimized return of each fault diagnosis trajectory as the optimization objective, and updating the internal parameters of the second detection model in reverse; after each round of parameter updates, re-sampling all fault samples to generate fault diagnosis trajectories, and repeating the calculation of evidence potential energy return, the calculation of optimized return, and the update process of model parameters; in continuous multiple iterations, when the average evidence potential energy and average tool usage consumption of all trajectories are stable within a preset fluctuation range, the model is determined to have converged, and the converged second detection model is determined as the first detection model.

[0258] Among them, the evidence potential return refers to the quantitative value calculated based on the second closed state, the second evidence closure score, and the evidence conflict intensity. The evidence potential return is used to quantify the uncertainty of evidence within the current diagnostic trajectory. The more insufficient the evidence and the more serious the evidence conflict, the higher the evidence potential return value, which means that the current diagnostic path is less effective. The model will impose a penalty on such paths during optimization.

[0259] Among them, the correctness of the final root cause determination of the trajectory is used to characterize whether the root cause output by the diagnostic trajectory matches the actual root cause of the fault. It is the basis for reward and punishment for model optimization. The preset fluctuation range is a fixed numerical range used to determine whether the model has converged. The preset fluctuation range can be a very small numerical range that can be set in advance. For example, the average evidence potential energy fluctuation threshold is set to 0.02 and the average tool usage consumption fluctuation threshold is set to 0.01. When the absolute value of the change in the average evidence potential energy of the full trajectory in each round is less than 0.02 and the absolute value of the change in the average tool usage consumption is less than 0.01 within a specified number of rounds (such as 10 rounds), the index fluctuation is considered to meet the convergence requirements.

[0260] For example, the evidence potential reward corresponding to the (t+1)th step of fault diagnosis can be expressed as:

[0261]

[0262] in, This indicates taking the non-negative part. This represents the preset evidence closure threshold. This indicates the preset penalty weight for conflict states. Indicates an indicator function, The second evidence closure score represents the t-th step in fault diagnosis. This score is used to determine the second closure state. The method for determining the second evidence closure score can refer to the method for determining the first evidence closure score in the above embodiment, and will not be elaborated here. The evidence state indicated by the first evidence closure score at step t+1 is evidence conflict. For example, the preset evidence closure threshold is 0.6, and the preset conflict state penalty weight is 1.0.

[0263] Specifically, the termination decision can be satisfied when the second evidence closure score reaches the preset evidence closure threshold. However, once a conflict of evidence occurs, the evidence potential reward will increase the penalty weight of the conflict state to force the agent to continue exploring to resolve the conflict.

[0264] In one possible implementation, constructing the optimized reward for a single trajectory based on evidence potential energy reward, tool usage information, and the correctness of the final root cause determination of the trajectory can include: determining the immediate reward for each step based on the evidence potential energy reward, tool usage information, and the correctness of the final root cause determination of the trajectory; and then determining the optimized reward for the trajectory based on the immediate rewards for each step. The immediate reward is used to indicate the benefit or loss brought by a single diagnostic action. The immediate reward refers to the single-step reward or penalty value calculated after a single round of tool invocation and evidence verification, combined with the increment of evidence potential energy, the single-step tool resource cost, and the current root cause identification accuracy. If the single-step action reduces evidence potential energy, consumes no unnecessary tools, and improves the root cause identification accuracy, the immediate reward is a positive reward value; if the single-step action causes evidence conflict, generates invalid tool cost, or fails to advance the root cause determination, the immediate reward is a negative penalty value. The total optimized reward for the entire trajectory is obtained by weighted summing of the immediate rewards for all steps of a fault diagnosis trajectory. It should be noted that before determining the immediate reward corresponding to the t-th step of fault diagnosis, the evidence closure score and evidence potential reward of the next detection step can be determined based on the candidate root cause set and candidate root cause evidence distribution of the next detection step, and the immediate reward corresponding to the t-th step can be determined based on this.

[0265] For example, the immediate reward corresponding to the t-th step of fault diagnosis can be represented as:

[0266]

[0267] Where, d t This represents the tool usage information for the t-th step. This indicates the preset tool cost weight. Indicates whether the current step is a termination action (1 if yes, 0 if no). Indicate the true root cause or the root cause identified by experts. This indicates that the reward has been correctly terminated. Indicates the punishment for the root cause of the error. This indicates that the punishment was terminated prematurely before the evidence was conclusive. The evidence closure score at step t+1 indicates that the evidence is closed. The first evidence closure score at step t+1 indicates that the evidence state is not evidence closure.

[0268] The preset tool cost weight refers to a fixed coefficient used to scale the penalty for tool resource consumption. The larger the weight, the more the model tends to reduce tool calls to save computing power and bandwidth resources. The preset tool cost weight is, for example, 0.1. The correct termination reward is a positive reward value given to a single step when the model outputs a diagnostic conclusion that matches the true root cause when the evidence is closed. It is used to incentivize the model to accurately terminate the diagnosis when the evidence is sufficient. The correct termination reward is, for example, 10.0. The incorrect root cause penalty is a negative loss value applied when the model terminates the diagnosis but the output root cause does not match the true root cause of the fault. It is used to constrain the model to avoid incorrectly locking the source of the fault. The incorrect root cause penalty is, for example, 10.0. The early termination penalty when the evidence is not closed is a penalty value applied when the model actively ends the diagnosis process before the evidence has reached the closure condition. It is used to force the model to continue to carry out evidence collection and review when the evidence is insufficient. The early termination penalty when the evidence is not closed is, for example, 5.0.

[0269] For example, the optimization reward for a single trajectory can be expressed as:

[0270]

[0271] Where t represents the sequence number of each step in the trajectory, and T represents the total number of steps in the trajectory. denoted as the decay factor in reinforcement learning, used to exponentially decay the immediate reward as the diagnostic steps increase, thereby penalizing the sluggishness and high-potential states of intermediate steps, forcing the agent (second detection model) to converge the diagnosis more efficiently. For example, the decay factor for the t-th step is, for example, 0.95.

[0272] In this way, by jointly constructing an optimized return based on evidence potential energy, tool resource consumption, and root cause matching results, and continuously iterating and updating the second detection model, the probability of the model selecting diagnostic actions with low evidence potential energy, capable of forming a normal evidence loop, and accurate root cause determination can be increased. This reduces premature termination when there is insufficient evidence or conflict, as well as the behavior of ineffective tool calls. The optimized first detection model can automatically perform review and evidence collection when there is low confidence or evidence conflict, and accurately output diagnostic conclusions when there is sufficient evidence, so that the root cause of the fault, the fault propagation path, and the diagnostic evidence are matched and unified.

[0273] In this embodiment, multiple fault samples, multiple first expert diagnostic trajectories corresponding to each fault sample, and the confidence levels of multiple nodes corresponding to each fault sample are obtained. The first expert diagnostic trajectories sequentially include the nodes to be investigated and the first tool to be called when investigating each node. Based on the multiple fault samples, the multiple first expert diagnostic trajectories, and the confidence levels of multiple nodes corresponding to each fault sample, a second detection model is determined. Based on the multiple fault samples and the second detection model, a first detection model is determined.

[0274] In the above method, a second detection model with topology investigation capability is obtained by first combining expert diagnostic trajectory supervision training. Then, based on the automatic diagnostic trajectory of fault sample sampling, the first detection model is obtained by combining evidence potential energy, tool resource consumption, and root cause matching results for iterative optimization. This method can replicate mature manual diagnostic logic, reduce invalid tool calls, actively review evidence when there is conflict or insufficient confidence, and achieve accurate closed-loop output when the evidence is sufficient. This improves the consistency and overall diagnostic efficiency of industrial fault root cause location, propagation path deduction and evidence chain generation.

[0275] Figure 5 This is a schematic diagram of a fault diagnosis device based on an intelligent agent, provided as an embodiment of this application. Please refer to... Figure 5 The agent-based fault diagnosis device 500 includes: an acquisition module 501, a first determination module 502, a first processing module 503, a second processing module 504, and a second determination module 505, wherein...

[0276] Module 501 is used to acquire first text describing the device fault;

[0277] The first determining module 502 is used to determine the starting node of fault diagnosis based on the first text;

[0278] The first processing module 503 is used to obtain a first tool to be called based on a pre-trained first detection model, starting from the starting node, and to obtain first local detection evidence based on the first detection model and the first tool to be called. The first detection model is trained based on multiple fault samples, expert diagnosis trajectories corresponding to each fault sample, and the confidence of each of the multiple nodes corresponding to each fault sample.

[0279] The second processing module 504 is used to perform credibility calibration on the first local detection evidence based on the first detection model, and determine the first closure state of the calibrated evidence. The first closure state is used to indicate whether the calibrated evidence is sufficient to lock the root cause of the fault and whether there is evidence conflict.

[0280] The second determining module 505 is used to determine the fault diagnosis result based on the first closed state.

[0281] The fault diagnosis device based on intelligent agents provided in this application embodiment can execute the technical solution shown in the above method embodiment. Its implementation principle and beneficial effects are similar, and will not be repeated here.

[0282] In one possible implementation, the first tool to be invoked includes one or more of the following: a topology query tool, a node fault diagnosis tool, and a sensor data reading tool;

[0283] The topology query tool is used to query the topological connection relationships between devices, components, and sensors within the system. The node fault diagnosis tool is used to output the fault probability and local faults of nodes. The sensor data reading tool is used to collect the sensing information and time-series operation data of the corresponding measurement points.

[0284] In one possible implementation, the first processing module 503 is specifically used for:

[0285] When the first tool to be invoked includes a topology query tool, the local observation topology and upstream and downstream related node information corresponding to the node are determined based on the first tool to be invoked, and the first local detection evidence is determined based on the local observation topology and upstream and downstream related node information.

[0286] When the first tool to be invoked includes a node fault diagnosis tool, the fault judgment result of the node is determined based on the first tool to be invoked, and the first local detection evidence is determined based on the fault judgment result;

[0287] When the first tool to be invoked includes a sensor data reading tool, the sensing time series operation data of the target measurement point corresponding to the time period is determined based on the first tool to be invoked, and the first local detection evidence is determined based on the sensing time series operation data.

[0288] In one possible implementation, the second determining module 505 is specifically used for:

[0289] When the first closed state indicates that the calibrated evidence is sufficient to pinpoint the root cause of the fault and there is no conflict of evidence, the fault diagnosis results, including the root cause node, the fault propagation path across devices, and the tracing evidence chain, are determined based on the first local detection evidence.

[0290] When the first closed state indicates that the evidence after calibration is insufficient to pinpoint the root cause of the fault or that there is a conflict of evidence, a second tool to be invoked is obtained, the second tool to be invoked is invoked to obtain second local detection evidence, and the fault diagnosis result is determined based on the first local detection evidence and the second local detection evidence.

[0291] In one possible implementation, the second processing module 504 is specifically used for:

[0292] Based on the calibrated evidence, a first candidate root cause set and a first candidate root cause evidence distribution are determined. The first candidate root cause evidence distribution is used to indicate the degree of fault support of the calibrated evidence for each candidate root cause in the first candidate root cause set.

[0293] Align the first candidate root cause set and determine the first evidence credibility information after alignment. The first evidence credibility information is used to indicate the failure confidence of each candidate root cause in the first candidate root cause set.

[0294] Based on the first evidence credibility information and the first candidate root cause evidence distribution, the second evidence credibility information is determined. The second evidence credibility information is used to indicate the failure confidence of each candidate root cause obtained after the evidence is fused and calibrated.

[0295] Based on the credible information of the second piece of evidence, the first closed state is determined.

[0296] In one possible implementation, the agent-based fault diagnosis device 500 further includes a third processing module 506, which is used for:

[0297] The system acquires multiple fault samples, multiple first expert diagnostic trajectories corresponding to each fault sample, and the confidence scores of multiple nodes corresponding to each fault sample. The first expert diagnostic trajectories sequentially include the nodes to be investigated and the first tool to be called when investigating each node.

[0298] Based on multiple fault samples, multiple first-expert diagnostic trajectories, and the confidence levels of multiple nodes corresponding to each fault sample, a second detection model is determined.

[0299] Based on multiple fault samples and the second detection model, the first detection model is determined.

[0300] In one possible implementation, the third processing module 506 is specifically used for:

[0301] Based on multiple fault samples, multiple first-expert diagnostic trajectories, and the confidence levels of multiple nodes corresponding to each fault sample, structured samples are determined. The structured samples are used to indicate the local topology, single-step diagnostic action, single-step tool observation data, and subsequent detection status updated after the action is executed for a single detection.

[0302] The second detection model is obtained by performing supervised fine-tuning training based on structured samples.

[0303] In one possible implementation, the third processing module 506 is specifically used for:

[0304] For each fault sample, a fault diagnosis trajectory is determined based on the second detection model;

[0305] For each fault diagnosis trajectory, determine the tool usage information. The tool usage information is used to indicate the resource consumption cost incurred by calling the tool during the process of determining the fault diagnosis trajectory.

[0306] For each fault diagnosis trajectory, the credibility calibration is performed on each observation evidence obtained using multiple tools to be invoked within the trajectory, and the second candidate root cause set and the distribution of the second candidate root cause evidence are determined after calibration.

[0307] For each fault diagnosis trajectory, a second closure state is determined based on the second candidate root cause set and the distribution of evidence for the second candidate root cause.

[0308] Based on the second closed state of each fault diagnosis trajectory and the tool usage information, the second detection model is iteratively optimized to obtain the first detection model.

[0309] The fault diagnosis device based on intelligent agents provided in this application embodiment can execute the technical solution shown in the above method embodiment. Its implementation principle and beneficial effects are similar, and will not be repeated here.

[0310] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 6 As shown, the electronic device 600 may include: a transceiver 601, a processor 602, and a memory 603.

[0311] The processor 602 executes computer execution instructions stored in the memory, causing the processor 602 to perform the scheme in the above embodiments. The processor 602 can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0312] The memory 603 is connected to the processor 602 via the system bus and completes communication between them. The memory 603 is used to store computer program instructions.

[0313] Transceiver 601 can be used to obtain the task to be run and its configuration information.

[0314] The system bus can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The system bus can be divided into address bus, data bus, control bus, etc. For ease of representation, only one thick line is used in the diagram, but this does not indicate that there is only one bus or one type of bus. Transceivers are used to enable communication between database access devices and other computers (e.g., clients, read-write libraries, and read-only libraries). Memory may include random access memory (RAM) and may also include non-volatile memory.

[0315] This application also provides a chip for executing instructions, which is used to execute the technical solution of the agent-based fault diagnosis method in the above embodiments.

[0316] This application also provides a computer-readable storage medium storing computer instructions. When the computer instructions are executed on a computer, the computer performs the technical solution of the fault diagnosis method based on intelligent agents described in the above embodiments.

[0317] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or modules, and may be electrical, mechanical, or other forms.

[0318] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to implement the solution of this embodiment according to actual needs.

[0319] Furthermore, the functional modules in the various embodiments of this application can be integrated into one processing unit, or each module can exist physically separately, or two or more modules can be integrated into one unit. The unit composed of the above modules can be implemented in hardware or in the form of hardware plus software functional units.

[0320] The integrated modules described above, implemented as software functional modules, can be stored in a computer-readable storage medium. These software functional modules, stored in a storage medium, include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute some steps of the methods of the various embodiments of this application.

[0321] It should be understood that the aforementioned processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. A general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly manifested as execution by a hardware processor, or execution by a combination of hardware and software modules within the processor.

[0322] The memory may include high-speed RAM, and may also include non-volatile memory (NVM), such as at least one disk storage device, and may also be a USB flash drive, external hard drive, read-only memory, disk or optical disc, etc.

[0323] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0324] The aforementioned storage media can be implemented from any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The storage media can be any available medium accessible to general-purpose or special-purpose computers.

[0325] An exemplary storage medium is coupled to a processor, enabling the processor to read information from and write information to the storage medium. Alternatively, the storage medium can be an integral part of the processor. The processor and storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and storage medium can exist as discrete components in an electronic control unit or main control device.

[0326] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0327] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A fault diagnosis method based on intelligent agents, characterized in that, include: Retrieve the first text used to describe the device malfunction; Based on the first text, determine the starting node for fault diagnosis; Starting from the starting node, a first tool to be invoked is obtained based on a pre-trained first detection model, and first local detection evidence is obtained based on the first detection model and the first tool to be invoked. The first detection model is trained based on multiple fault samples, expert diagnosis trajectories corresponding to each fault sample, and the confidence of each of the multiple nodes corresponding to each fault sample. Based on the first detection model, the credibility of the first local detection evidence is calibrated, and the first closure state of the calibrated evidence is determined. The first closure state is used to indicate whether the calibrated evidence is sufficient to lock the root cause of the fault and whether there is evidence conflict. Based on the first closed state, the fault diagnosis result is determined.

2. The method according to claim 1, characterized in that, The first tool to be invoked includes one or more of the following: topology query tool, node fault diagnosis tool, and sensor data reading tool; The topology query tool is used to query the topological connection relationships between devices, components, and sensors within the system. The node fault diagnosis tool is used to output the fault probability and local faults of nodes. The sensor data reading tool is used to collect the sensing information and time-series operation data of the corresponding measurement points.

3. The method according to claim 1, characterized in that, First local detection evidence is obtained based on the first detection model and the first tool to be invoked, including: When the first tool to be invoked includes a topology query tool, based on the first tool to be invoked, the local observation topology and upstream and downstream related node information corresponding to the node are determined, and based on the local observation topology and the upstream and downstream related node information, the first local detection evidence is determined; When the first tool to be invoked includes a node fault diagnosis tool, the fault judgment result of the node is determined based on the first tool to be invoked, and the first local detection evidence is determined based on the fault judgment result; When the first tool to be invoked includes a sensor data reading tool, the sensing time series operation data of the target measurement point corresponding to the time period is determined based on the first tool to be invoked, and the first local detection evidence is determined based on the sensing time series operation data.

4. The method according to claim 1, characterized in that, Based on the first closed state, the fault diagnosis result is determined, including: When the first closed state indicates that the calibrated evidence is sufficient to pinpoint the root cause of the fault and there is no evidence conflict, the fault diagnosis result, including the root cause node, the fault propagation path across devices, and the traceable evidence chain, is determined based on the first local detection evidence. When the first closed state indicates that the evidence after calibration is insufficient to pinpoint the root cause of the fault or that there is a conflict of evidence, a second tool to be invoked is obtained, the second tool to be invoked is invoked to obtain second local detection evidence, and the fault diagnosis result is determined based on the first local detection evidence and the second local detection evidence.

5. The method according to claim 1, characterized in that, Determine the first closure state of the calibrated evidence, including: Based on the calibrated evidence, a first candidate root cause set and a first candidate root cause evidence distribution are determined, wherein the first candidate root cause evidence distribution is used to indicate the degree of fault support of the calibrated evidence for each candidate root cause in the first candidate root cause set. Align the first candidate root cause set and determine the first evidence credibility information after alignment. The first evidence credibility information is used to indicate the failure confidence level accumulated by each candidate root cause in the first candidate root cause set. Based on the first evidence credibility information and the first candidate root cause evidence distribution, a second evidence credibility information is determined. The second evidence credibility information is used to indicate the fault confidence of each candidate root cause obtained after the evidence is fused and calibrated. Based on the credible information of the second piece of evidence, the first closed state is determined.

6. The method according to claim 1, characterized in that, The training method for the first detection model includes: Multiple fault samples are obtained, as well as multiple first expert diagnostic trajectories corresponding to each fault sample and the confidence scores of multiple nodes corresponding to each fault sample. The first expert diagnostic trajectory includes, in sequence, the nodes to be investigated and the first tool to be called when investigating each node. Based on the multiple fault samples, the multiple first expert diagnostic trajectories, and the confidence levels of the multiple nodes corresponding to each fault sample, a second detection model is determined. Based on the multiple fault samples and the second detection model, a first detection model is determined.

7. The method according to claim 6, characterized in that, Based on the multiple fault samples, the multiple first expert diagnostic trajectories, and the confidence levels of the multiple nodes corresponding to each fault sample, a second detection model is determined, including: Based on the multiple fault samples, the multiple first expert diagnostic trajectories, and the confidence scores of the multiple nodes corresponding to each fault sample, a structured sample is determined. The structured sample is used to indicate the local topology, single-step diagnostic action, single-step tool observation data, and subsequent detection status updated after the action is executed for a single detection. Supervised fine-tuning training is performed based on the structured samples to obtain the second detection model.

8. The method according to claim 7, characterized in that, Based on the multiple fault samples and the second detection model, a first detection model is determined, including: For each fault sample, a fault diagnosis trajectory is determined based on the second detection model; For each fault diagnosis trajectory, tool usage information is determined, which indicates the resource consumption cost incurred by calling the tool during the process of determining the fault diagnosis trajectory. For each fault diagnosis trajectory, the credibility calibration is performed on each observation evidence obtained using multiple tools to be invoked within the trajectory, and the second candidate root cause set and the distribution of the second candidate root cause evidence are determined after calibration. For each fault diagnosis trajectory, a second closure state is determined based on the second candidate root cause set and the second candidate root cause evidence distribution; Based on the second closed state of each fault diagnosis trajectory and the tool usage information, the second detection model is iteratively optimized to obtain the first detection model.

9. An electronic device, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory to implement the method as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1 to 8.

Images