A large model multi-agent industrial equipment abnormal root cause positioning method and system

Abstract

The application discloses a large model multi-agent industrial equipment abnormal root cause positioning method and system, relates to the technical field of industrial equipment intelligent operation and maintenance and fault diagnosis, and comprises the following steps: constructing a candidate root cause propagation relationship from a candidate root cause object to an abnormal performance object and containing a state transmission relationship between equipment objects based on an abnormal state of industrial equipment and equipment correlation knowledge; verifying actual running states of each equipment object and propagation expectations, combining root cause support information to form a root cause evidence state; correcting the propagation relationship according to an evidence gap of an associated unexplained abnormal parameter or a propagation mismatch equipment object, and determining a target root cause propagation relationship according to the same, and outputting a root cause positioning result. The method improves the accuracy, stability and interpretability of complex industrial fault root cause positioning by constructing, verifying and correcting the root cause propagation relationship in cooperation with multiple agents.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance and fault diagnosis technology for industrial equipment, specifically a method and system for locating the root cause of anomalies in large-scale multi-agent industrial equipment. Background Technology

[0002] Modern industrial systems consist of numerous devices, sensors, and control nodes, which are continuously interconnected through process media, energy transfer, and control signals. After a localized fault occurs, the abnormal state often spreads along the connection direction of the devices, generating parameter deviations, alarm records, and operational logs at different times. Existing industrial equipment diagnostic technologies mainly rely on threshold detection, time series analysis, machine learning, expert rules, and knowledge graphs to identify anomalies. Threshold detection and time series analysis are suitable for detecting parameter exceeding limits or trend changes, but they typically only confirm the existence of an anomaly and struggle to distinguish the fault source, propagation stage, and end-point manifestation. Machine learning can output fault categories based on historical samples, but the number of industrial fault samples is limited, making it difficult for model conclusions to correspond to the actual propagation process between devices. Expert rules and knowledge graphs can incorporate equipment knowledge, but preset rules do not adequately cover complex faults, and static correlations are difficult to verify as field conditions change. Large language models can parse alarm texts and maintenance records, but when a single model simultaneously handles anomaly confirmation, knowledge retrieval, and root cause reasoning, it is prone to unclear task dependencies and mismatched reasoning bases. Multi-agent diagnostics is increasingly being used to share different tasks, but some solutions still remain at the level of summarizing results and have failed to establish a stable correspondence between multi-source information and root cause reasoning under the same abnormal event.

[0003] Existing root cause localization methods suffer from insufficient verification of propagation relationships. While device topology can indicate connection locations, it cannot independently prove whether anomalies propagate according to expected mechanisms and chronological order. Historical cases or diagnostic rules can provide empirical support, but they cannot replace object-by-object verification of on-site device status. When multiple candidate root causes appear simultaneously, existing methods often select conclusions based on overall similarity or confidence levels, failing to identify which anomalous parameters remain unexplained or mismatched devices in the propagation path. When evidence is insufficient, common approaches include regenerating candidate conclusions or expanding the search scope, making it difficult to establish a correspondence between unexplained anomalies and specific propagation locations, and even more difficult to perform targeted correction based on device connection relationships and fault propagation mechanisms. On-site verification and user feedback are often separated from the original reasoning process; erroneous conclusions cannot be traced back to specific evidence gaps, and new cases are difficult to form reusable diagnostic criteria. Therefore, in scenarios where multi-source anomaly information and device-related knowledge jointly participate in diagnosis, existing root cause localization methods still struggle to construct root cause propagation relationships that can be verified by on-site operating conditions, and also struggle to determine the locations to be corrected in the propagation path and perform targeted correction based on unexplained anomalous parameters or device objects that do not conform to propagation expectations. Summary of the Invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by this invention is that existing methods for locating the root causes of industrial equipment anomalies suffer from insufficient coordination between multi-source anomaly information and equipment-related knowledge, lack of object-by-object verification of root cause propagation paths based on the actual operating status of the equipment, and difficulty in identifying and correcting gaps in propagation paths when evidence is insufficient, resulting in low accuracy and interpretability of root cause location. The invention also addresses the problem of how to construct, verify, locate gaps, and directionally correct candidate root cause propagation relationships under the constraints of equipment connection relationships and fault propagation mechanisms.

[0006] To address the aforementioned technical problems, this invention provides the following technical solution: a method for locating the root cause of anomalies in large-scale multi-agent industrial equipment, comprising: forming a candidate root cause propagation relationship based on the abnormal state of industrial equipment and equipment association knowledge, which points from candidate root cause objects to abnormal manifestation objects and includes the state transmission relationship between equipment objects; verifying the correspondence between the actual operating state of each equipment object and the propagation expectation of the state transmission relationship, and determining the root cause evidence state of the candidate root cause propagation relationship in conjunction with the root cause support information; associating unexplained abnormal parameters or equipment objects that do not conform to the propagation expectation with evidence gaps in the root cause evidence state, and limiting the correction of the candidate root cause propagation relationship; determining the target root cause propagation relationship based on the corrected candidate root cause propagation relationship and the corresponding root cause evidence state, and outputting the root cause location result.

[0007] As a preferred embodiment of the large-scale multi-agent industrial equipment anomaly root cause localization method described in this invention, the method further includes: mapping the equipment operation data and equipment text information corresponding to the same abnormal event according to the equipment identifier and the occurrence time to form an abnormal information association relationship; defining the abnormal state of the industrial equipment by the parameter change state and alarm parsing state in the abnormal information association relationship, and directionally associating the abnormal state of the industrial equipment with the abnormal equipment, abnormal parameters, and candidate faults.

[0008] As a preferred embodiment of the large-scale multi-agent industrial equipment anomaly root cause localization method described in this invention, it further includes: a large-scale multi-agent collaborative reasoning team composed of a task scheduling agent, an anomaly monitoring agent, a context retrieval agent, a root cause reasoning agent, and a result generation agent; the anomaly monitoring agent outputs the abnormal state of the industrial equipment based on the anomaly information correlation; the context retrieval agent organizes equipment correlation knowledge and fault diagnosis knowledge around the abnormal equipment into the root cause support information; the abnormal state of the industrial equipment and the root cause support information are jointly associated with the root cause reasoning agent after being formed in parallel, and the task scheduling agent maintains the collaborative reasoning state according to the dependency relationship between anomaly confirmation, context retrieval, root cause reasoning, and result generation.

[0009] As a preferred embodiment of the large-scale multi-agent industrial equipment anomaly root cause localization method described in this invention, wherein: the equipment connection direction represented by the equipment association knowledge limits the connection order of the candidate root cause object, associated equipment object, and anomaly manifestation object; the fault propagation mechanism and propagation time characteristics limit the state transmission relationship between adjacent equipment objects; the abnormal parameters in the abnormal state of the industrial equipment are associated with the corresponding equipment object or state transmission relationship according to the equipment affiliation and occurrence sequence, so that the candidate root cause propagation relationship satisfies the abnormal parameter coverage and fault propagation time sequence constraints.

[0010] As a preferred embodiment of the large-scale multi-agent industrial equipment anomaly root cause localization method described in this invention, the root cause support information includes on-site state support information and diagnostic knowledge support information; along the candidate root cause propagation relationship, the conformity between the actual operating state of each equipment object and the corresponding propagation expectation is determined as the node propagation conformity state, and the support relationship between the candidate root cause propagation relationship and historical fault knowledge and inference verification information is determined as the path knowledge support state; the inference verification information represents the diagnostic rule matching relationship or counterfactual consistency relationship, and the diagnostic rules limit the correspondence between the abnormal state combination and the candidate root cause object or fault propagation direction; the root cause evidence state is jointly limited by the node propagation conformity state and the path knowledge support state.

[0011] As a preferred embodiment of the large-scale multi-agent industrial equipment anomaly root cause localization method described in this invention, the evidence gaps include anomaly coverage gaps associated with unexplained anomaly parameters and propagation mismatch gaps associated with equipment objects that do not conform to propagation expectations; the anomaly coverage gaps and propagation mismatch gaps are used to mark the propagation positions to be corrected in the corresponding candidate root cause propagation relationships; the intermediate equipment objects or state transmission relationships corresponding to the propagation positions to be corrected are defined by the equipment association knowledge and fault propagation mechanism; the corrected candidate root cause propagation relationships are re-associated with the actual operating state of each equipment object and the root cause support information to form an updated root cause evidence state.

[0012] As a preferred embodiment of the large-scale multi-agent industrial equipment anomaly root cause localization method described in this invention, the method further includes: correlating the root cause localization results with on-site verification results and user feedback to form a diagnostic verification state; when the diagnostic verification state is a confirmed state, associating the corresponding industrial equipment anomaly state, target root cause propagation relationship, and root cause support information as new fault cases, and forming diagnostic rules based on the fault symptom recurrence relationship between different new fault cases; when the diagnostic verification state is a mismatch state, associating the user feedback with the corresponding evidence gap and propagation position to be corrected to form diagnostic correction information for adjusting the inference strategy or the large-scale model domain adaptation state.

[0013] Another objective of this invention is to provide a large-scale multi-agent industrial equipment anomaly root cause localization system. This system can verify the equipment operating status of candidate root cause propagation relationships through the coordinated operation of a propagation relationship generation module, an evidence state correction module, and a root cause localization update module. It can also locate and correct propagation relationships based on evidence gaps. This addresses the problems in existing industrial equipment anomaly root cause localization technologies, such as insufficient coordination between multi-source anomaly information and equipment-related knowledge, lack of equipment-by-equipment verification of propagation paths, and difficulty in targeted correction when evidence is insufficient.

[0014] As a preferred embodiment of the large-scale multi-agent industrial equipment anomaly root cause localization system described in this invention, the system includes: a propagation relationship generation module, an evidence state correction module, and a root cause localization update module. The propagation relationship generation module is used to associate multi-source equipment information of the same abnormal event, collaboratively form the abnormal state and root cause support information of industrial equipment, and construct candidate root cause propagation relationships limited by equipment connection direction, fault propagation mechanism, and occurrence sequence. The evidence state correction module is used to verify the equipment state and diagnostic knowledge of the candidate root cause propagation relationships, locate the propagation position to be corrected based on evidence gaps, and update the candidate root cause propagation relationships and root cause evidence state. The root cause localization update module is used to form root cause localization results based on the updated candidate root cause propagation relationships and root cause evidence state, and form new fault cases or diagnostic correction information based on on-site verification and user feedback.

[0015] Another object of the present invention is to provide a large-scale multi-agent industrial equipment anomaly root cause localization device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the large-scale multi-agent industrial equipment anomaly root cause localization method.

[0016] Another object of the present invention is to provide a large-scale multi-agent industrial equipment anomaly root cause localization storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the large-scale multi-agent industrial equipment anomaly root cause localization method are implemented.

[0017] The beneficial effects of this invention are as follows: The large-scale multi-agent industrial equipment anomaly root cause localization method provided by this invention constructs candidate root cause propagation relationships constrained by equipment association knowledge, verifies the correspondence between the actual operating state of the equipment and the propagation expectation, and uses evidence gaps to perform directional correction on mismatch positions in the propagation relationship. This transforms the root cause localization process from a single anomaly judgment to a closed-loop analysis of propagation relationship construction, evidence verification, and gap correction, thereby reducing the diagnostic bias caused by intermediate abnormal equipment being misjudged as root cause objects and the omission of propagation chains. This invention achieves better results in terms of root cause localization accuracy, propagation path interpretability, and diagnostic stability under complex associated faults. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 The overall flowchart of a large-scale multi-agent industrial equipment anomaly root cause localization method provided by the present invention is shown.

[0020] Figure 2 This is the overall timing diagram of a large-scale multi-agent industrial equipment anomaly root cause localization method provided by the present invention.

[0021] Figure 3 This invention provides a schematic diagram of the topology and abnormal state of a raw material pump-regulating valve system, which is used in a large-scale multi-agent industrial equipment anomaly root cause localization method.

[0022] Figure 4 This invention provides a schematic diagram of the root cause propagation relationship of a raw material pump-regulating valve system in a large-scale multi-agent industrial equipment anomaly root cause localization method.

[0023] Figure 5 The flowchart of the task scheduling agent for a large-scale multi-agent industrial equipment anomaly root cause localization method provided by the present invention is shown.

[0024] Figure 6 The present invention provides a causal reasoning flowchart for a multi-agent industrial equipment anomaly root cause localization method based on the present invention. Detailed Implementation

[0025] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0026] During operation, the specific equipment objects refer to the raw material pump P-101, inlet pipeline L-101, inlet filter S-101, regulating valve FCV-201, heat exchanger E-101, and their corresponding detection locations involved in the generation, propagation, or manifestation of abnormalities.

[0027] Equipment operating data specifically refers to raw sampled values ​​such as pressure, valve position, temperature, vibration, and equipment efficiency. The actual operating status is determined based on the direction, magnitude, and timing of the changes in the raw sampled values ​​relative to the normal baseline.

[0028] An abnormal state of industrial equipment is a correlated state formed by multiple actual operating states and alarm analysis results under the same abnormal event.

[0029] Candidate root cause objects refer to the initial fault objects in candidate propagation relationships.

[0030] Abnormal behavior objects refer to equipment objects that directly exhibit abnormalities at the transmission end.

[0031] Status transfer relationship refers to the abnormal transfer relationship between adjacent equipment objects formed by material flow, mechanical action, control response or heat exchange.

[0032] Root cause support information includes on-site condition support information and diagnostic knowledge support information. The former is formed by comparing the actual operating status with the propagation expectations, while the latter is formed by historical fault knowledge, diagnostic rule matching results, and counterfactual consistency relationships.

[0033] An evidence gap refers to a specific location in a candidate propagation relationship where anomalous parameters or propagation expectations are not explained.

[0034] Example 1, referring to Figures 1-4 This embodiment provides a large-scale multi-agent industrial equipment anomaly root cause localization method, which is used to determine the target root cause object that can completely explain the current abnormal event from multiple possible fault objects when there are material transfer, control adjustment and fault chain effects between industrial equipment.

[0035] The large-scale multi-agent model referred to in this embodiment refers to multiple agents with independent inputs, outputs, and task constraints formed by loading different task prompts onto the same basic large language model. Each agent performs task scheduling, anomaly confirmation, context retrieval, root cause reasoning, and result generation around the same abnormal event. The multiple agents do not output multiple diagnostic conclusions independently, but rather, according to a preset task dependency relationship, the structured data generated by the previous agent is used as the input of the next agent, so that equipment operation data, equipment text information, equipment association knowledge, fault propagation mechanism, and historical diagnostic knowledge are continuously transmitted under the same abnormal event.

[0036] In this embodiment, the Tongyi Qianwen 2.5-7B instruction model is selected as the basic large language model. The basic large language model adopts a decoding Transformer network structure with a parameter scale of approximately 7 billion and a context length of 32,768 text tags during inference.

[0037] The basic large language model is deployed on a computing server connected to the industrial control network. It receives device text information in natural language form and fielded device operation data, and generates structured diagnostic data according to a preset output format. In other implementations, other instruction-based large language models with semantic parsing, knowledge association, and structured text generation capabilities can also be used. The parameter scale can be set to 3 billion, 7 billion, 13 billion, or larger scales depending on the computing resources available in the industrial field.

[0038] Based on the training results of a general corpus, the basic large language model is adapted to the industrial equipment domain using data. This industrial equipment domain data includes equipment manuals, equipment ledgers, process flow descriptions, equipment connection relationships, operational alarms, maintenance records, historical fault cases, and diagnostic rules. The domain adaptation samples include input data and target output data. The input data includes at least anomaly event identifiers, equipment identifiers, anomaly parameters, parameter change states, alarm parsing states, and equipment association knowledge. The target output data includes at least candidate root cause objects, equipment object connection order, state propagation relationships, propagation expectations, the corresponding locations of anomaly parameters, and root cause objects confirmed on-site.

[0039] In this embodiment, a low-rank adaptation method is used to train the basic large language model in the domain. The rank of the low-rank matrix is ​​set to 16, the scaling factor is set to 32, the dropout ratio is set to 0.05, the learning rate is set to 2×10^-5, and the number of training rounds is set to 3.

[0040] During training, the low-rank adaptation parameters are adjusted based on the difference between the model's output field and the target output field. After training is complete, the basic parameters of the base language model are frozen, and the industrial equipment domain adaptation parameters are loaded during inference. Other implementations may also employ full parameter fine-tuning, cue-based fine-tuning, prefix fine-tuning, or continuous training based on confirmed fault cases.

[0041] To reduce the occurrence of different results from the same input, this embodiment sets the generation temperature of the basic large language model to 0.1, the cumulative probability threshold to 0.9, and limits the content generated by the model through output field constraints. The device identifier output by the basic large language model must exist in the device ledger, the output device connection relationship must exist in the device association knowledge, and the output fault propagation relationship must be associated with the corresponding fault mechanism knowledge or diagnostic rules.

[0042] When the model output lacks device identifiers, state transit relationships, propagation expectations, or knowledge sources, the current output is marked as an invalid result and will not proceed to the subsequent root cause localization process.

[0043] Task scheduling agent 201, anomaly monitoring agent 202, context retrieval agent 203, root cause reasoning agent 204, and result generation agent 205 each invoke independent session instances of the basic large language model. Each session instance loads different task prompts and maintains consistency of input objects through anomaly event identifiers, device identifiers, and data time ranges.

[0044] Data transmitted between agents is recorded using structured fields, which can be saved as relational data tables, graph data records, or JavaScript object representation data. Unvalidated free text is not directly used as input for the next agent.

[0045] The task prompts loaded by the task scheduling agent 201 include: reading the diagnostic request, identifying the target device and the abnormal event; decomposing the diagnostic task into anomaly confirmation, context retrieval, root cause reasoning, and result generation; executing anomaly confirmation and context retrieval in parallel; initiating root cause reasoning only when both return the same abnormal event identifier and target device identifier; and not generating candidate root causes independently. The task scheduling agent 201 outputs the task identifier, task type, abnormal event identifier, target device identifier, input data location, expected output fields, and task dependency status.

[0046] The task prompts loaded by the anomaly monitoring agent 202 include: determining the abnormal state of industrial equipment based on the equipment operation data and alarm analysis status in the same abnormal event; outputting the abnormal equipment, abnormal parameters, normal baseline, actual value, direction of change, magnitude of change, and occurrence time respectively; and not directly outputting the final root cause based on a single abnormal parameter.

[0047] The task prompts loaded by the context retrieval agent 203 include: retrieving device connection relationships, device mechanism knowledge, historical fault cases, and diagnostic rules related to abnormal devices; each search result must include a knowledge source identifier; distinguishing upstream devices, current devices, and downstream devices according to the device connection direction; and not generating device objects or fault propagation relationships that do not exist in the knowledge base.

[0048] The task prompts loaded by the root cause reasoning agent 204 include: generating candidate root cause objects only based on the abnormal state of industrial equipment and root cause support information; candidate propagation paths must conform to the equipment connection direction and fault propagation mechanism; each abnormal parameter must be associated with the corresponding equipment object or state transmission relationship; comparing the actual operating state of the equipment object with the propagation expectation item by item; unexplained abnormal parameters are marked as abnormal coverage gaps, and equipment objects that do not conform to the propagation expectation are marked as propagation mismatch gaps; equipment objects or state transmission relationships must not be supplemented without equipment association knowledge support.

[0049] The task prompts loaded by the resulting intelligent agent 205 include: determining the target root cause propagation relationship from the candidate root cause propagation relationships that meet the root cause determination conditions; outputting the root cause device, root cause type, upstream inducing cause, abnormal propagation path, comprehensive confidence level, key on-site evidence, and key knowledge evidence; and not changing the device objects and state transfer relationships that have been verified in the root cause reasoning process.

[0050] Thus, the five types of intelligent agents respectively undertake task management, abnormal data transformation, diagnostic knowledge acquisition, constrained causal reasoning, and structured result output, forming a large model multi-agent collaborative processing process.

[0051] This embodiment focuses on the raw material conveying system in an ethylene production plant as the diagnostic object. The raw material conveying system includes a centrifugal raw material pump P-101 with a rated power of 75kW, a rated head of 60m, and a rated flow rate of 150m³ / h; a pneumatic diaphragm regulating valve FCV-201 with a diameter of DN100 and a stroke of 50mm; a pressure sensor PT-101 with a range of 0~1.6MPa and an accuracy of 0.5%FS; a shell-and-tube heat exchanger E-101 with a heat exchange area of ​​120m²; an inlet storage tank T-101; an inlet pipeline L-101; and an inlet filter S-101 installed in the inlet pipeline L-101.

[0052] The medium flows in the following sequence: inlet storage tank T-101, inlet filter S-101, raw material pump P-101, regulating valve FCV-201, and heat exchanger E-101. Pressure sensor PT-101 is installed on the outlet pipeline of raw material pump P-101 to obtain the outlet pressure of raw material pump P-101. When pressure sensor PT-101 itself does not malfunction, it only serves as the location for obtaining operational data and is not considered the root cause of abnormal propagation.

[0053] In this embodiment, candidate root cause objects are determined jointly by the equipment object and its candidate fault state. For example, cavitation in raw material pump P-101 constitutes one candidate root cause object, and impeller wear in raw material pump P-101 constitutes another candidate root cause object. Abnormal manifestation objects are equipment objects and their abnormal states that directly exhibit parameter deviations or changes in operating status after the abnormality propagates, such as increased opening of control valve FCV-201 and fluctuations in outlet temperature of heat exchanger E-101.

[0054] Reference Figure 3At 14:32:17 on a certain day, the anomaly monitoring agent 202 read 600 pressure sampling values ​​generated by pressure sensor PT-101 in the previous 60 seconds. The outlet pressure of raw material pump P-101 gradually decreased from 0.80MPa to 0.78MPa, and further decreased from 0.78MPa to 0.45MPa between 14:32:05 and 14:32:10, subsequently remaining between 0.44MPa and 0.46MPa. During the same period, the valve position feedback value of regulating valve FCV-201 increased from 45% to 78%, the inlet temperature of heat exchanger E-101 remained around 32℃, and the outlet temperature fluctuated repeatedly between 68℃ and 82℃.

[0055] The anomaly monitoring agent 202 converts equipment operation data into anomaly status records. Each anomaly status record includes at least an anomaly event identifier, equipment identifier, parameter identifier, normal baseline, actual value, direction of change, magnitude of change, and occurrence time. In this embodiment, the states of decreased raw material pump outlet pressure, increased regulating valve opening compensation, and fluctuating heat exchanger outlet temperature are identified and grouped into the same anomaly event according to the equipment identifier and occurrence time, thus forming anomaly states of industrial equipment.

[0056] The context retrieval agent 203 reads equipment association knowledge around the raw material pump P-101. Equipment association knowledge can be stored using a directed equipment graph, equipment connection table, or process flow relationship data. This embodiment uses a directed equipment graph, setting the inlet storage tank T-101, inlet pipeline L-101, inlet filter S-101, raw material pump P-101, regulating valve FCV-201, and heat exchanger E-101 as equipment nodes, and defining the medium flow direction, inlet resistance influence, pump head change, valve control compensation, and heat exchange state change as directed relationships between the equipment nodes.

[0057] The contextual retrieval agent 203 also retrieves the relationships from the fault knowledge base: the relationship between increased inlet resistance and decreased effective net positive suction head (NPSH); the relationship between decreased NPSH and feed pump cavitation; the relationship between feed pump cavitation and pump vibration and decreased head; and the relationship between decreased outlet pressure and increased control valve opening compensation. Each root cause support information record at least the knowledge source, applicable equipment, preceding state, subsequent state, propagation direction, propagation time range, and support type.

[0058] The root cause reasoning agent 204 receives abnormal states and root cause support information of industrial equipment, converts the abnormal states into abnormal state nodes, converts equipment association knowledge into equipment relationship edges, and generates three candidate root cause objects from the abnormal equipment and its associated equipment: raw material pump cavitation, impeller wear, and control valve malfunction. Blockage of the inlet pipeline or inlet filter forms an inlet blockage state. This inlet blockage state, as an upstream inducing cause associated with raw material pump cavitation, is used to explain abnormal suction resistance and decreased effective net positive suction head (NPSH) of the raw material pump, but is not considered a direct root cause state ultimately determined in this embodiment.

[0059] Each candidate root cause object establishes a candidate root cause propagation relationship 100 to the abnormal behavior object. The candidate root cause propagation relationship 100 can be represented by a directed graph, a causal chain, or an event sequence with propagation time. This embodiment uses a causal graph composed of device nodes, state nodes, and directed propagation edges. Each directed propagation edge records the previous state, subsequent state, propagation mechanism, propagation expectation, and knowledge source.

[0060] When the root cause reasoning agent 204 generates candidate root cause propagation relationship 100, it first obtains the upstream and downstream devices associated with the candidate root cause object according to the device connection direction, and then determines whether there is a fault propagation mechanism between adjacent device objects.

[0061] When there is no device connection relationship or fault propagation mechanism between adjacent device objects, the expansion along that direction stops; when a corresponding relationship exists, the adjacent device objects and their state propagation relationships are written into the candidate root cause propagation relationship 100. Therefore, the large model cannot arbitrarily combine device objects, but rather forms candidate propagation paths under the constraints of device connection relationships and fault propagation mechanisms.

[0062] Cavitation of the raw material pump P-101 was selected as the first candidate root cause propagation relationship. The main propagation path of the first candidate root cause propagation relationship is as follows: cavitation of the raw material pump causes bubbles to be generated and collapse in the pump, resulting in a decrease in pump head and outlet pressure; after the actual outlet flow rate is lower than the control target, the adjustment process increases the opening of the control valve FCV-201; after the unstable flow rate of the medium enters the heat exchanger E-101, it causes fluctuations in the outlet temperature of the heat exchanger.

[0063] Clogging of the inlet filter S-101, increased inlet resistance, and decreased effective net positive suction head (NPSH) are used as upstream inducing factors associated with feed pump cavitation to verify the formation conditions of feed pump cavitation.

[0064] Reference Figure 4The candidate root cause propagation relationship 100 is constructed according to the diagnostic interpretation chain of "root cause - intermediate cause - abnormal manifestation". Among them, raw material pump cavitation is the candidate root cause state, inlet blockage and its corresponding increase in suction resistance are the intermediate causes, and the pump outlet pressure drops from 0.8MPa to 0.45MPa, the regulating valve opening increases from 45% to 78%, and the heat exchanger outlet temperature fluctuates between 68℃ and 82℃ as abnormal manifestations formed along the equipment correlation relationship. Figure 4 The solid arrows in the text are used to represent positive diagnostic associations in the candidate root cause propagation relationship. That is, based on the equipment connection relationship, fault propagation mechanism and actual operating status, the root cause status, intermediate causes and various abnormal manifestations are associated in sequence to verify whether the current candidate root cause can explain the downstream anomaly. Figure 4 The dashed arrows do not indicate reverse physical propagation of the fault between devices, but rather represent the feedback verification relationship between the abnormal behavior and upstream candidate root causes and intermediate causes. That is, after obtaining terminal anomalies such as heat exchanger outlet temperature fluctuations, this abnormal state is back-linked along the candidate root cause propagation relationship to feed pump cavitation and inlet blockage. It is checked whether the root cause states and intermediate causes can jointly explain the pump outlet pressure drop, the increase in regulating valve opening, and the heat exchanger temperature fluctuation. When a consistent explanation can be formed, the corresponding verification result is written into root cause evidence state 200. When an abnormal behavior cannot be explained by the corresponding root cause state or intermediate cause, an anomaly coverage gap or propagation mismatch gap is formed, and the position to be corrected in the candidate root cause propagation relationship 100 is defined accordingly, thereby... Figure 4 It also demonstrates the forward construction process of candidate root cause propagation relationships and the reverse evidence verification process based on abnormal results.

[0065] The second candidate root cause propagation relationship uses impeller wear of raw material pump P-101 as the candidate root cause, and is sequentially associated with decreased pump efficiency, decreased pump head, decreased outlet pressure, and increased control valve opening compensation. The third candidate root cause propagation relationship uses abnormality of control valve FCV-201 as the candidate root cause, and is sequentially associated with valve position deviation, changes in medium flow rate, and fluctuations in heat exchanger outlet temperature.

[0066] Each propagation edge is associated with a state transmission relationship that is configured with a propagation expectation. The propagation expectation can be the direction of parameter increase or decrease, the range of change that the parameter should enter, or the order in which different abnormal states should be satisfied.

[0067] In this embodiment, the expected propagation between inlet filter blockage and increased inlet resistance includes an increase in the pressure difference across the filter; the expected propagation between raw material pump cavitation and decreased head includes increased pump vibration; the expected propagation between decreased outlet pressure and increased regulating valve opening includes an increase in the valve position feedback value after the pressure drops; and the expected propagation between unstable flow rate and heat exchanger temperature fluctuation includes fluctuations in outlet temperature after the valve position changes.

[0068] The root cause reasoning agent 204 associates each abnormal parameter with the device node or state propagation relationship in the candidate root cause propagation relationship 100 according to the device affiliation and occurrence time. The raw material pump outlet pressure is associated with the raw material pump P-101, the regulating valve opening is associated with the state propagation relationship between the outlet pressure drop and the valve control compensation, and the heat exchanger outlet temperature is associated with the state propagation relationship of the medium flow rate change acting on the heat exchanger E-101.

[0069] Further data was collected on the pressure difference across the inlet filter S-101, the vibration of the raw material pump P-101, and the equipment efficiency. The pressure difference across the inlet filter S-101 was 0.12 MPa, which is less than 0.05 MPa under normal conditions; the vibration of the raw material pump P-101 was 4.5 mm / s, which is less than 2.8 mm / s under normal conditions; and the pump efficiency decreased from 78% to 72%. The anomaly monitoring agent 202 converted the above raw data into states of increased filter pressure difference, increased pump vibration, and decreased pump efficiency, respectively, and associated them with the same anomaly event.

[0070] The root cause reasoning agent 204 reads the actual operating state of the device objects one by one along the candidate root cause propagation relationship 100 and compares it with the corresponding propagation expectation. When the actual operating state meets the propagation expectation, the corresponding device node or propagation edge is recorded as a propagation compliance state; when the actual operating state does not meet the propagation expectation, it is recorded as a propagation mismatch state; when the confirmed abnormal parameters are not associated with the candidate root cause propagation relationship 100, they are recorded as an unexplained state.

[0071] The on-site condition support information is formed by comparing the results of increased filter pressure differential, increased pump vibration, decreased outlet pressure, increased control valve opening, and heat exchanger outlet temperature fluctuations with corresponding propagation expectations. The diagnostic knowledge support information is formed by historical failure cases, diagnostic rules, and counterfactual consistency relationships. Together, the on-site condition support information and the diagnostic knowledge support information constitute the root cause support information.

[0072] The contextual retrieval agent 203 retrieved historical fault cases that matched the current abnormal state of the industrial equipment. These historical fault cases included records of raw material pump cavitation, inlet filter blockage, pump outlet pressure drop, and increased control valve opening. The contextual retrieval agent 203 also retrieved diagnostic rule R3-1, which specifies that when the outlet pressure drops and the control valve opening increases, the fault direction points to the raw material pump or the raw material pump inlet side.

[0073] The root cause reasoning agent 204 performs counterfactual verification on the inlet blockage. After removing the inlet blockage state from the current diagnostic scenario, it infers, based on the equipment mechanism, that the suction resistance decreases, the feed pump inlet pressure recovers, the effective net positive suction head (NPSH) increases, and the pump outlet pressure recovers. This inference result is consistent with the equipment operating mechanism, thus forming a counterfactual consistency relationship.

[0074] The root cause evidence state 200 is jointly defined by the node propagation conformity state and the path knowledge support state. The root cause evidence state 200 records at least the candidate path identifier, the candidate root cause object, the actual operating status of each device object, the node verification result, the historical case support result, the diagnostic rule matching result, the counterfactual consistency relationship, the evidence gap, the propagation position to be corrected, and the overall confidence level.

[0075] In this embodiment, the overall confidence level C is determined according to the following formula:

[0076]

[0077] in, This indicates the degree of conformity in node propagation, determined by the ratio of the number of device nodes and state transmission relationships that conform to propagation expectations to the total number of nodes to be verified. The degree of path knowledge support is determined by the number of state transit relationships supported by historical fault cases or diagnostic rules. It indicates the degree of counterfactual consistency, determined by whether the corresponding abnormal behavior changes according to the device mechanism after the candidate cause is removed.

[0078] The overall confidence level is 0.85, which is calculated by weighting the node propagation consistency (1.00), path knowledge support (0.80), and counterfactual consistency (0.70), i.e.: 0.4×1.00+0.3×0.80+0.3×0.70=0.85.

[0079] The overall confidence level is only used as a quantitative field in the root cause evidence state 200 and does not replace the verification of device connectivity, fault propagation mechanism and abnormal parameter coverage.

[0080] The initial first candidate root cause propagation relationship can explain the increased pump vibration, decreased outlet pressure, increased control valve opening, and heat exchanger outlet temperature fluctuations, but it does not include inlet blockage, increased suction resistance, and decreased effective net positive suction head (NPSH), and therefore cannot explain the increased pressure difference across the inlet filter S-101. The root cause reasoning agent 204 marks this pressure difference state as an unexplained anomalous parameter and creates an anomalous coverage gap on the inlet side of the feed pump P-101.

[0081] The root cause reasoning agent 204 sends the gap type, gap association parameters, adjacent equipment objects, and the propagation position to be corrected to the context retrieval agent 203. The context retrieval agent 203 obtains equipment mechanism knowledge regarding the relationship between inlet filter blockage and increased inlet resistance, and between increased inlet resistance and decreased effective net positive suction head (NPSH), focusing on the inlet side of the raw material pump P-101. After the knowledge is returned, the root cause reasoning agent 204 supplements the first candidate root cause propagation relationship with information on inlet filter S-101, inlet blockage state, increased suction resistance state, and decreased effective NPSH state, forming a corrected candidate root cause propagation relationship.

[0082] The second candidate root cause propagation relationship can explain the decrease in pump efficiency and the decrease in outlet pressure, but impeller wear usually corresponds to gradual performance degradation and cannot explain the decrease in outlet pressure from 0.78MPa to 0.45MPa in a short period of time, thus forming a propagation mismatch gap.

[0083] The third candidate root cause propagation relationship can explain the changes in control valve opening and heat exchanger temperature fluctuations, but the increase in control valve opening actually occurs after the outlet pressure drops, which does not conform to the propagation sequence of control valve abnormality as the initial root cause, thus forming a propagation mismatch gap.

[0084] The corrected first candidate root cause propagation relationship was re-linked to the actual operating status and root cause support information of each device. The increases in inlet filter differential pressure, increased pump vibration, decreased outlet pressure, increased regulating valve opening, and heat exchanger outlet temperature fluctuations were all explained accordingly.

[0085] The overall confidence level for the first candidate root cause propagation relationship is 0.85, for the second candidate root cause propagation relationship it is 0.45, and for the third candidate root cause propagation relationship it is 0.35.

[0086] The root cause determination threshold was determined by statistically analyzing the overall confidence levels of historically confirmed fault cases, with 0.6 being the critical value that maximized the distinction between correct and incorrect root cause propagation relationships. The overall confidence level of the first candidate root cause propagation relationship was higher than the root cause determination threshold, and no unexplained outliers were found. Therefore, the corrected first candidate root cause propagation relationship was determined as the target root cause propagation relationship, and the root cause localization result 300 was output.

[0087] The resulting intelligent agent 205 generates a structured root cause localization result 300 based on the target root cause propagation relationship. Root cause localization result 300 records the direct root cause device as raw material pump P-101, the direct root cause state as raw material pump cavitation, the upstream inducing device as inlet filter S-101, the upstream inducing cause as inlet filter blockage, and the target root cause propagation relationship is represented according to the actual state of the equipment as "inlet filter blockage—increased suction resistance—decreased effective net positive suction head—raw material pump cavitation—decreased outlet pressure—increased control valve opening compensation—heat exchanger outlet temperature fluctuation", with a comprehensive confidence level of 0.85, and is associated with historical fault cases, diagnostic rule R3-1, counterfactual consistency relationship, and corresponding field operation data.

[0088] Table 1. Comparison of the present invention with conventional industrial equipment root cause localization methods.

[0089]

[0090] As shown in Table 1, in a production scenario where raw material pumps, regulating valves, and heat exchangers are interconnected, when the outlet pressure of the raw material pump drops from the normal value to 0.45 MPa, the opening of the regulating valve increases from 45% to 78%, and the outlet temperature of the heat exchanger fluctuates between 68℃ and 82℃, a single threshold detection can only identify low pressure, high valve position, and temperature fluctuation separately. The diagnostic method based on the fault classification model may also only output raw material pump fault or regulating valve fault, making it difficult to determine whether there is a continuous propagation relationship between different anomalies.

[0091] This invention correlates inlet-side flow obstruction, feed pump cavitation, outlet pressure drop, control valve compensation, and heat exchanger temperature fluctuations according to equipment connection direction and fault propagation mechanism through candidate root cause propagation relationships. Then, based on the actual operating status of each equipment object, historical fault knowledge, diagnostic rules, and counterfactual consistency relationships, a root cause evidence state is formed. When the initial propagation relationship cannot explain the inlet-side state, valve opening, or heat exchanger temperature, the specific propagation location to be corrected is determined through evidence gaps, and corresponding equipment objects or state transmission relationships are supplemented, ultimately forming the root cause localization result.

[0092] Therefore, the technical effects of this invention are particularly prominent in scenarios where multiple devices malfunction simultaneously, direct root causes and upstream triggering causes coexist, malfunctions propagate continuously along the process chain or control chain, and initial evidence cannot fully cover all malfunction parameters.

[0093] This invention can avoid misjudging control compensation states such as increased valve opening as the root cause of the fault, can distinguish between the direct root cause state of raw material pump cavitation and the upstream inducing cause of inlet blockage, and can clarify the corresponding position of each abnormal parameter in the propagation relationship, thereby improving the accuracy of root cause localization, the interpretability of propagation path and the diagnostic stability in the case of insufficient evidence under complex associated faults.

[0094] Example 2, refer to Figure 5 Based on Example 1, this example further defines the formation method of abnormal information association and the collaborative reasoning process among multiple agents. It is used to solve the technical problems that it is difficult to classify device operation data and device text information into the same abnormal event, and that object mismatch and time mismatch are easy to occur between different diagnostic tasks. It also achieves the technical effect of transmitting multi-source abnormal information under the same device object, the same event time range and the same task dependency relationship.

[0095] Equipment operation data is generated by on-site testing equipment.

[0096] In this embodiment, pressure sensor PT-101 and regulating valve FCV-201 collect outlet pressure and valve position feedback values ​​at a frequency of 10Hz, respectively; heat exchanger E-101 collects inlet and outlet temperatures at a frequency of 1Hz; and raw material pump P-101 collects equipment efficiency and motor current at a frequency of 1Hz. The equipment identifier, parameter identifier, sampled values, and sampling time are written into the time-series data record. Equipment text information is generated from operation alarms and equipment logs, and at least the original text, source, equipment identifier, and timestamp are saved.

[0097] At 14:32:12, an alarm was generated for "P-101 pump outlet pressure low" and at 14:32:15, an alarm was generated for "FCV-201 valve opening high".

[0098] The equipment text information can be parsed using keyword rules, entity recognition models, or large language models. In this embodiment, the original alarm text, equipment ledger, and parameter dictionary are input into the large language model. The first alarm is parsed as equipment P-101, the parameter is outlet pressure, the actual value is 0.45MPa, the lower limit value is 0.50MPa, and the alarm level is high. The second alarm is parsed as equipment FCV-201, the parameter is valve position, the actual value is 78%, the normal range is 40% to 50%, and the alarm level is medium. These are used to form alarm parsing statuses.

[0099] Parameter change status is used to characterize the deviation of equipment operating parameters from the normal operating baseline.

[0100] This embodiment establishes a dynamic pressure baseline using data from the same operating conditions over the 7 days prior to the anomaly. The historical average outlet pressure is 0.82 MPa with a standard deviation of 0.05 MPa. The standardized deviation corresponding to the current pressure of 0.45 MPa is -7.4, exceeding the threshold of 3 times the standard deviation. The cumulative sum detection value is 2.35, higher than the threshold of 1.2; the exponentially weighted moving average statistic is 0.53, lower than the control lower limit of 0.72, thus determining that the outlet pressure is in a downward abrupt change state. The opening of the regulating valve FCV-201 increases from 45% to 78%, which is determined to be a valve position compensation increase state; the outlet temperature of heat exchanger E-101 varies between 68℃ and 82℃, which is determined to be a temperature fluctuation state. The above parameter change states can also be formed through dynamic thresholds, time-series abrupt change detection, or fault mode matching. This embodiment fuses the statistical detection and time-series detection results in a way that triggers the process if any result meets the abnormal condition.

[0101] The period from 14:32:05 to 14:32:10, where the pressure abrupt change occurred, was designated as the event center, with the correlation time range being 60 seconds before and 30 seconds after the anomaly. The low pressure alarm and high valve position alarm correspond to raw material pump P-101 and its downstream regulating valve FCV-201, respectively, and occur 3 seconds apart. Therefore, the resolution status of these two alarms, along with the parameter changes in pressure, valve position, and temperature, are recorded under the same anomaly event, forming an anomaly information correlation. This anomaly information correlation specifically records the anomaly event identifier, equipment identifier, anomaly parameters, parameter change status, alarm resolution status, data time range, and data source, ensuring that subsequent reasoning uses data generated by the same equipment anomaly.

[0102] A large-scale model multi-agent collaborative reasoning team is composed of a task scheduling agent 201, an anomaly detection agent 202, a context retrieval agent 203, a root cause reasoning agent 204, and a result generation agent 205. Each agent can use the same basic large language model and load different task instructions, or it can use models adapted for anomaly detection, knowledge retrieval, or causal reasoning. This embodiment uses the same basic large language model, configuring independent task prompts, input fields, and output fields for each of the five types of agents.

[0103] Anomaly monitoring agent 202 reads the correlation of anomaly information, accurately quantifies the anomaly event, determines that the outlet pressure drop is 42.7%, the rate of drop during the sharp drop phase is 0.035 MPa / s, and the control valve opening compensation increment is 33%, and writes the above quantification results into the abnormal status of industrial equipment. Context retrieval agent 203 reads the equipment connection relationships around raw material pump P-101, obtaining the upstream inlet storage tank T-101 and inlet pipeline L-101 and its inlet filter S-101, as well as the downstream control valve FCV-201 and heat exchanger E-101; it retrieves cases with a similarity greater than 0.75 from 128 historical cases and obtains diagnostic rules R3-1 and R3-2, forming root cause support information.

[0104] Reference Figure 5 After receiving a diagnostic request, the task scheduling agent 201 first extracts the diagnostic object and diagnostic constraints from the request content. The agent then decomposes the diagnostic task corresponding to the abnormal event into an anomaly confirmation task, a context retrieval task, a root cause reasoning task, and a result generation task. The anomaly confirmation task and the context retrieval task have no data dependency on each other and are set to execute in parallel. The root cause reasoning task uses the abnormal state of the industrial equipment generated by the anomaly confirmation task and the root cause support information generated by the context retrieval task as common inputs. The result generation task uses the root cause localization result 300 generated by the root cause reasoning task as input. Each task instruction carries a task identifier, anomaly event identifier, target device identifier, input data location, expected output field, and deadline.

[0105] Reference Figure 5 The collaborative reasoning state is used to record whether each diagnostic task managed by the task scheduling agent 201 has obtained data that meets the conditions for subsequent task execution.

[0106] After receiving a diagnostic request, the task scheduling agent 201 first parses the request content, extracts the diagnostic target, abnormal event identifier, target device identifier, and data time range, and decomposes the diagnostic task into an anomaly confirmation task, context retrieval task, root cause reasoning task, and result generation task, while identifying the data dependencies between each subtask. When there is no dependency between the anomaly confirmation task and the context retrieval task, the task scheduling agent 201 issues tasks to the anomaly monitoring agent 202 and the context retrieval agent 203 in parallel, and marks their collaborative reasoning state as parallel processing. When the anomaly monitoring agent 202 returns the abnormal status of the industrial equipment and the context retrieval agent 203 returns root cause support information, and the abnormal event identifier, target device identifier, and data time range of the two are consistent, the anomaly confirmation task and the context retrieval task are changed to the completed state, the root cause reasoning task changes from the waiting state to the executable state, and the task scheduling agent 201 issues the corresponding task to the root cause reasoning agent 204. When there is a clear data dependency between the subtasks, they are executed sequentially according to the dependency order. After the root cause reasoning agent 204 forms the target root cause propagation relationship based on the abnormal state of industrial equipment and root cause support information and outputs the root cause localization result 300, the result generation task is converted to an executable state, and the result generation agent 205 forms a diagnostic result accordingly. The task scheduling agent 201 further collects the results returned by each agent, verifies the abnormal event identifier, equipment identifier, data time range and preset output fields, and forms a complete diagnostic report after the verification is passed and returns it to the diagnostic requester, so that each agent can complete collaborative reasoning under the same abnormal event in parallel or sequential order.

[0107] Example 3, based on Example 2, further limits the device connection direction, fault propagation mechanism, propagation time characteristics and abnormal parameter coverage method in the candidate root cause propagation relationship 100, in order to solve the technical problem that the candidate root cause propagation relationship may not conform to the field device connection structure, control logic or fault evolution sequence, and achieve the technical effect that the candidate propagation path is verifiable in terms of device object, state transmission relationship and abnormal occurrence sequence.

[0108] The equipment connection direction is used to define the connection order of candidate root cause objects, associated equipment objects, and anomalous behavior objects in the candidate root cause propagation relationship 100. In this embodiment, the equipment sequence of the raw material conveying system is determined according to the actual flow direction of the medium as inlet tank T-101, inlet pipeline L-101, inlet filter S-101, raw material pump P-101, regulating valve FCV-201, and heat exchanger E-101. The equipment connection direction can also be formed according to the energy transfer direction or the control signal direction; when the outlet pressure decreases and causes the regulating valve opening to increase, the state transfer relationship between raw material pump P-101 and regulating valve FCV-201 is determined by the control signal direction.

[0109] The fault propagation mechanism defines the types of state transmission relationships that can be established between adjacent equipment. Clogged inlet filter S-101 reduces the effective flow area of ​​the filter channel, leading to increased inlet pipeline resistance. Increased inlet resistance lowers the inlet pressure of raw material pump P-101 and causes a decrease in effective net positive suction head (NPSH). The decrease in NPSH causes liquid vaporization within the pump, forming bubbles. The generation and collapse of these bubbles reduce the pump head and outlet pressure. When the outlet flow rate falls below the control target, the opening of regulating valve FCV-201 is increased during the adjustment process. Unstable flow entering heat exchanger E-101 causes changes in the heat exchange per unit time and results in outlet temperature fluctuations. These propagation relationships respectively belong to material resistance transmission, fluid mechanical action, control response, and heat exchange state transmission.

[0110] Fault propagation mechanisms can be defined by knowledge of equipment mechanics, historical failure modes, or validated diagnostic rules.

[0111] This embodiment uses the relationship between the feed pump head and flow characteristics, the inlet resistance and effective net positive suction head (NPSH), and the historical fault mode "inlet pipeline blockage - feed pump cavitation - outlet pressure drop" to limit the formation of propagation edges. When two abnormal parameters occur simultaneously only in time, but the equipment connection relationship and fault propagation mechanism cannot explain the existence of state transmission between them, a corresponding propagation edge is not established in the candidate root cause propagation relationship 100.

[0112] The propagation time characteristic is used to define the sequence in which fault states should be passed along the equipment. In this embodiment, the raw material pump outlet pressure slowly decreases from 0.80 MPa to 0.78 MPa between 14:30 and 14:32; the regulating valve FCV-201 begins to increase from 45% at 14:31:50; the outlet pressure drops sharply to 0.45 MPa between 14:32:05 and 14:32:10; a low pressure alarm is generated at 14:32:12, and a high valve position alarm is generated at 14:32:15; the heat exchanger outlet temperature fluctuates after 14:32:20. The above timing sequence satisfies the propagation order of inlet-side flow obstruction, raw material pump performance degradation, valve compensation, and downstream temperature fluctuation.

[0113] Abnormal parameters are associated with corresponding equipment objects or state propagation relationships according to equipment affiliation and occurrence sequence. For example, a pressure difference of 0.12 MPa across the inlet filter is associated with inlet filter S-101; pump vibration of 4.5 mm / s is associated with the cavitation state of feed pump P-101; outlet pressure of 0.45 MPa is associated with feed pump P-101; valve opening of 78% is associated with the state propagation relationship between outlet pressure drop and valve control compensation; and heat exchanger outlet temperature fluctuation is associated with the state propagation relationship of flow rate change acting on heat exchanger E-101. This association ensures that each abnormal parameter can find a corresponding equipment node or propagation edge in the candidate root cause propagation relationship 100.

[0114] When the candidate root cause propagation relationship uses inlet filter S-101 blockage as the candidate root cause, feed pump P-101 is used to transmit the abnormal inlet resistance to abnormal outlet pressure, and is thus an associated device. Control valve FCV-201 is used to transmit the pressure change to flow compensation, and is also an associated device. Heat exchanger E-101 is the abnormal manifestation at the end of the propagation. When only the drop in outlet pressure of feed pump P-101 is used as the terminating abnormality, feed pump P-101 can also be considered an abnormal manifestation in the corresponding candidate relationship. This means that the role of the device is determined based on the start and end positions of the current candidate propagation relationship, rather than being fixed.

[0115] The candidate root cause propagation relationships corresponding to inlet filter blockage cover five categories of abnormal parameters: filter pressure differential, pump vibration, outlet pressure, valve position, and heat exchanger temperature, and satisfy the aforementioned propagation time characteristics. Feed pump cavitation, as an independent root cause, cannot explain the increase in filter pressure differential; impeller wear, as a root cause, cannot explain the sharp drop in outlet pressure within 10 seconds. Therefore, the former is marked as insufficient coverage of abnormal parameters, and the latter is marked as not meeting the propagation time characteristics, for further processing under root cause evidence state 200.

[0116] Example 4, refer to Figure 6 Based on Example 3, this example further defines the correction process for root cause support information, root cause evidence status 200, evidence gaps, and candidate root cause propagation relationships 100. This is used to solve the technical problem that it is difficult to determine the specific location of insufficient evidence in the propagation relationship based solely on overall similarity or comprehensive confidence. It also achieves the technical effect that candidate propagation relationships can be verified, supplemented in a targeted manner, and re-evaluated according to the device object and propagation location.

[0117] Root cause support information includes field condition support information and diagnostic knowledge support information. Field condition support information is used to verify whether the actual operating status of each equipment object conforms to the corresponding propagation expectation, and can be formed by pressure, differential pressure, valve position, temperature, vibration, and equipment efficiency. Diagnostic knowledge support information is used to verify whether the candidate root cause propagation relationship 100 is supported by existing diagnostic knowledge, and can be formed by historical fault knowledge and inference verification information. Historical fault knowledge includes historical fault cases and historical propagation paths, while inference verification information represents diagnostic rule matching relationships or counterfactual consistency relationships.

[0118] Matching the first candidate root cause propagation relationship with historical cases reveals anomalies including decreased pump outlet pressure and increased valve opening, thus forming historical case support. Matching the current abnormal state of the industrial equipment with rule R3-1 shows that the combination of decreased pressure and increased valve opening satisfies the pump inlet-side fault direction defined by this rule, thus forming a diagnostic rule matching relationship. Assuming that clearing the inlet filter blockage leads to decreased suction resistance, restored pump inlet pressure, restored outlet pressure, and normalized valve opening, this reasoning result is consistent with the equipment mechanism, thus forming a counterfactual consistency relationship. These supporting results collectively form the path knowledge support state.

[0119] Along the intermediate cause of the first candidate root cause propagation, the actual pressure difference of the inlet filter S-101 (0.12 MPa) is compared with the propagation expectation that the pressure difference should increase under the blockage state, forming a support state; the vibration value of the raw material pump P-101 (4.5 mm / s) is compared with the propagation expectation that the vibration should increase under the cavitation state, forming a support state; the actual state of the regulating valve FCV-201 increasing from 45% to 78% is compared with the propagation expectation that the valve position should increase after the outlet pressure drops, forming a support state; the temperature fluctuation at the outlet of the heat exchanger E-101 is compared with the propagation expectation that the temperature should change after the flow rate changes, forming a support state.

[0120] The root cause evidence state 200 is jointly defined by the node propagation compliance state and the path knowledge support state. In this embodiment, an evidence record is generated for each candidate root cause propagation relationship. The evidence record includes at least the path identifier, candidate root cause object, node propagation compliance state, path knowledge support state, unexplained abnormal parameters, device objects that do not conform to propagation expectations, propagation location to be corrected, and overall confidence level.

[0121] Evidence gaps include anomaly coverage gaps and propagation mismatch gaps. Anomaly coverage gaps are used to identify anomalous parameters that have been confirmed but not explained in candidate root cause propagation relationships 100. Propagation mismatch gaps are used to identify equipment objects whose actual operating state is inconsistent with propagation expectations. Gap types can be represented by field status or missing nodes marked by path markers. In this embodiment, gap types and their adjacent nodes are recorded on the candidate propagation graph.

[0122] The first candidate root cause propagation relationship, with feed pump cavitation as the initial root cause, can explain pump vibration, outlet pressure drop, valve position increase, and heat exchanger temperature fluctuations, but cannot explain the 0.12 MPa pressure difference across inlet filter S-101. This pressure difference parameter forms an anomalous coverage gap, and the propagation location to be corrected is marked at the upstream inlet of feed pump P-101 in the first candidate root cause propagation relationship. The propagation location to be corrected refers to the specific path location of the equipment object or state transit relationship that needs to be supplemented, replaced, or deleted, not to the lack of evidence for the entire path.

[0123] Based on equipment correlation knowledge, the direct upstream of feed pump P-101 includes inlet pipeline L-101 and inlet filter S-101. According to the fault propagation mechanism, inlet filter blockage can increase inlet resistance and cause a decrease in effective net positive suction head (NPSH). Therefore, inlet filter S-101 is added as an intermediate equipment object at the location to be corrected for propagation, and the state transmission relationship of "inlet filter blockage - increased inlet resistance - decreased effective NPSH - feed pump cavitation" is supplemented to form a corrected candidate root cause propagation relationship.

[0124] The second candidate root cause propagation relationship predicted that impeller wear would cause a gradual decrease in pump efficiency and outlet pressure. However, the actual outlet pressure dropped from 0.78 MPa to 0.45 MPa within 10 seconds, and the actual rate of change was inconsistent with the propagation prediction. Therefore, a propagation mismatch gap was formed between the impeller wear state and the outlet pressure drop state. After consulting equipment-related knowledge and fault propagation mechanisms, no intermediate equipment objects or propagation relationships were found that could directly cause the aforementioned pressure change due to impeller wear. Therefore, no unfounded supplementation was made, and the corresponding mismatch state was retained.

[0125] The corrected candidate root cause propagation relationships are re-associated with the actual operating states of inlet filter S-101, feed pump P-101, control valve FCV-201, and heat exchanger E-101, and historical cases, rule R3-1, and counterfactual consistency relationships are re-matched to form an updated root cause evidence state 200. The inlet filter pressure difference anomaly is explained, and the corrected overall confidence score is updated to 0.85, which is higher than the root cause determination threshold of 0.6, therefore the correction is stopped. In other embodiments, when the highest overall confidence score is lower than the root cause determination threshold, correction can continue to be performed based on the updated evidence gap, with the maximum number of iterations set to 3.

[0126] Reference Figure 6The root cause reasoning agent 204 first performs fault hypothesis generation based on the abnormal state of industrial equipment, extracts abnormal equipment, abnormal parameters and their change characteristics, and generates candidate root cause objects including raw material pump cavitation, inlet blockage, etc.; then enters the causal path construction stage, and associates the candidate root cause objects with the corresponding abnormal manifestation objects based on the equipment topology, connection direction and fault propagation mechanism defined by equipment association knowledge, forming a candidate root cause propagation relationship 100 that points from the candidate root cause object to the observed abnormality and includes the state transmission relationship between equipment objects.

[0127] After forming candidate root cause propagation relationship 100, root cause reasoning agent 204 obtains equipment association knowledge, historical fault knowledge, diagnostic rules and relevant field status information through context retrieval agent 203, collects evidence for each equipment object in the candidate propagation path, compares the actual operating status with the corresponding propagation expectation, and forms root cause evidence state 200 by combining historical case support, diagnostic rule matching and fault mechanism support.

[0128] In the hypothesis testing phase, counterfactual analysis is further performed on candidate root cause objects. That is, assuming that the corresponding candidate cause does not exist or has been eliminated, it is determined whether the corresponding abnormal state recovers or disappears according to the device mechanism, and the verification results are used to update the evidence support level of the candidate root cause propagation relationship. When the overall confidence level of the root cause evidence state 200 does not reach the root cause determination threshold, or there are still unexplained abnormal parameters and device objects that do not meet the propagation expectations, it is identified as an anomaly coverage gap or propagation mismatch gap, and the process returns to the causal path construction phase. Based on the position corresponding to the evidence gap, the device objects or state transmission relationships are retrieved and supplemented again, and evidence collection and hypothesis testing are performed again. The number of iterations does not exceed the preset N. When the corrected root cause evidence state 200 reaches the root cause determination threshold and all key anomalies are explained, the iteration stops, and the candidate root cause propagation relationship with the highest evidence support level and meeting the root cause determination conditions is determined as the target root cause propagation relationship, forming the root cause localization result 300. The result generating agent 205 further outputs the root cause object, the target root cause propagation relationship, the overall confidence level, and the corresponding supporting evidence to generate a root cause report.

[0129] Example 5, based on Example 4, further adds on-site verification of root cause localization result 300, fault case formation, diagnostic rule updating and mismatch feedback correction process, to solve the technical problems of separation between diagnostic conclusion and on-site inspection results, difficulty in tracing back erroneous reasoning to specific evidence gaps, and difficulty in reusing verified fault knowledge, and achieves the technical effect that on-site verification results can continuously supplement diagnostic knowledge and correct subsequent root cause reasoning processes.

[0130] After the root cause localization result 300 is generated, the following information is output: occurrence time of the abnormal event, abnormal equipment, abnormal parameters, target root cause object, target root cause propagation relationship, overall confidence level, and key root cause supporting information. In this embodiment, the direct root cause equipment is raw material pump P-101, the direct root cause state is raw material pump cavitation, the upstream inducing equipment is inlet filter S-101, the upstream inducing cause is inlet filter blockage, the overall confidence level is 0.85, and it is associated with the pressure difference across the filter (0.12 MPa), raw material pump vibration (4.5 mm / s), historical cases, diagnostic rule R3-1, and counterfactual consistency relationship.

[0131] On-site maintenance personnel, based on the root cause analysis results, inspected the vibration status of raw material pump P-101, the pump outlet pressure, and the condition of the inlet pipeline and inlet filter. The inspection revealed a blockage in the inlet filter, and raw material pump P-101 also exhibited significant cavitation vibration. After cleaning the inlet filter and resuming operation, the cavitation vibration of the raw material pump decreased, the pump outlet pressure recovered from 0.45 MPa to 0.80 MPa, the opening of regulating valve FCV-201 decreased, and the outlet temperature of heat exchanger E-101 stabilized.

[0132] The filter inspection results, cleaning and treatment results, post-treatment equipment operation data, and maintenance personnel feedback are correlated with the abnormal events and root cause localization results 300 to form a diagnostic verification status. The diagnostic verification status can be a confirmed status or a mismatch status; when it is necessary to retain some consistent results, a partial confirmed status can also be added. In this embodiment, the cavitation status of the raw material pump is verified by on-site operation data and post-treatment status changes, and inlet blockage, as an intermediate cause associated with the root cause status, is supported by on-site inspection results; therefore, the diagnostic verification status is determined to be a confirmed status.

[0133] When the diagnostic verification status is confirmed, the abnormal state of the industrial equipment, the target root cause propagation relationship, the root cause supporting information, and the on-site verification results are associated as a new fault case. The new fault case records the abnormal event identifier, equipment identifiers P-101 and S-101, the outlet pressure drop process, valve opening changes, heat exchanger temperature changes, candidate path correction process, target root cause object, on-site handling process, and post-handling status, avoiding the need to save only the isolated conclusion of "inlet filter blockage". New fault cases can be saved using structured case records, graph-based cases, or semantic vector formats. This embodiment simultaneously saves the structured fields and the corresponding text semantic index.

[0134] By comparing the newly added fault cases with historical cases, both cases exhibit a combination of states: decreased outlet pressure, increased control valve opening, increased inlet filter differential pressure, and restored outlet pressure after filter cleaning. This establishes a fault symptom recurrence relationship. The fault symptom recurrence relationship characterizes the combination of abnormal states in different abnormal events and the recurrence of confirmed root causes. When the preset number of recurrences and consistency are met, diagnostic rules are formed or updated based on the fault symptom recurrence relationship. In this embodiment, the newly added fault case is written into the case support record of rule R3-1, providing new on-site evidence for the correspondence that "when the outlet pressure decreases and the control valve opening increases, the fault location is on the pump inlet side."

[0135] In another abnormal event, if cavitation of the raw material pump and a drop in outlet pressure are also detected, but on-site verification reveals that the inlet valve is mistakenly closed and the inlet filter S-101 is not clogged, then the diagnostic verification status is determined to be a mismatch state. User feedback specifically records that on-site confirmation indicates the root cause is the inlet valve, the error root cause is the inlet filter S-101, and the actual opening degree of the inlet valve was not obtained in the original reasoning process.

[0136] The aforementioned user feedback is linked to the evidence gaps and propagation locations upstream of the raw material pump P-101 to form diagnostic correction information. This information records unexplained inlet valve states, inlet valve objects requiring supplementation, and the state propagation relationship of "inlet valve malfunction—reduced flow area—reduced inlet pressure—raw material pump cavitation." When subsequently handling pump inlet-side anomalies, the actual inlet valve opening is added to the contextual search, and inlet filter blockage and inlet valve malfunction are verified separately as peer-level candidate root causes.

[0137] Diagnostic correction information is also used to adjust inference strategies or the domain adaptation status of the large model. Adjusting the inference strategy can change the generation range of candidate root cause objects, the order of evidence retrieval, or the reliability of evidence; adjusting the domain adaptation status of the large model can use confirmed cases as correct inference samples, and use evidence gaps, propagation locations to be corrected, and on-site confirmed root causes in mismatched cases as correction samples. In this embodiment, abnormal event data, original candidate root cause propagation relationships, on-site verification results, and corrected state transmission relationships are organized into domain adaptation data, which, after review, is used for subsequent domain adaptation, so that new root cause inference continues to be based on on-site verified equipment association knowledge, fault propagation mechanisms, and diagnostic rules.

[0138] Example 6, an embodiment of the present invention, provides a large-scale multi-agent industrial equipment anomaly root cause localization system, including a propagation relationship generation module, an evidence state correction module, and a root cause localization update module.

[0139] The propagation relationship generation module is used to associate multi-source device information of the same abnormal event, collaboratively form abnormal state and root cause support information of industrial equipment, and construct candidate root cause propagation relationships 100 that are limited by device connection direction, fault propagation mechanism and occurrence sequence.

[0140] The evidence status correction module is used to verify the device status and diagnostic knowledge of the candidate root cause propagation relationship 100, locate the propagation position to be corrected based on the evidence gap, and update the candidate root cause propagation relationship and root cause evidence status 200.

[0141] The root cause localization update module is used to generate root cause localization result 300 based on the updated candidate root cause propagation relationship and root cause evidence status 200, and to generate new fault cases or diagnostic correction information based on on-site verification and user feedback.

[0142] This embodiment also provides a computer device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements a large-model multi-agent industrial equipment anomaly root cause localization method as proposed in the above embodiment.

[0143] This embodiment also provides a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it implements a method for locating the root cause of anomalies in large-scale multi-agent industrial equipment as proposed in the above embodiment.

[0144] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0145] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0146] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0147] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0148] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for locating the root cause of anomalies in large-scale multi-agent industrial equipment, characterized in that, include: Based on the knowledge of abnormal state of industrial equipment and equipment association, a candidate root cause propagation relationship is formed, which points from the candidate root cause object to the abnormal manifestation object and includes the state transmission relationship between the equipment objects (100). The actual operating state of each device object is verified in correspondence with the expected propagation of the state transmission relationship, and the root cause evidence state (200) of the candidate root cause propagation relationship (100) is determined by combining the root cause support information. The evidence gap in the root cause evidence state (200) is associated with unexplained anomalous parameters or device objects that do not conform to the propagation expectations, and limits the correction of the candidate root cause propagation relationship (100); Based on the corrected candidate root cause propagation relationship and the corresponding root cause evidence status (200), the target root cause propagation relationship is determined, and the root cause localization result is output (300).

2. The method for locating the root cause of anomalies in large-scale multi-agent industrial equipment as described in claim 1, characterized in that, Also includes: The device operation data and device text information corresponding to the same abnormal event are matched according to the device identifier and the time of occurrence to form an abnormal information association relationship; The abnormal state of the industrial equipment is defined by the parameter change status and alarm parsing status in the abnormal information association relationship, and the abnormal state of the industrial equipment is directionally associated with abnormal equipment, abnormal parameters and candidate faults.

3. The method for locating the root cause of anomalies in large-scale multi-agent industrial equipment as described in claim 2, characterized in that, Also includes: The large model multi-agent collaborative reasoning team consists of a task scheduling agent (201), an anomaly detection agent (202), a context retrieval agent (203), a root cause reasoning agent (204), and a result generation agent (205); The anomaly monitoring agent (202) outputs the abnormal state of the industrial equipment based on the anomaly information association relationship, and the context retrieval agent (203) organizes the equipment association knowledge and fault diagnosis knowledge around the abnormal equipment into the root cause support information. The abnormal state of the industrial equipment and the root cause support information are jointly associated with the root cause reasoning agent (204) after being formed in parallel, and the task scheduling agent (201) maintains the collaborative reasoning state according to the dependency relationship between abnormal confirmation, context retrieval, root cause reasoning and result generation.

4. The method for locating the root cause of anomalies in large-scale multi-agent industrial equipment as described in claim 1 or 3, characterized in that: The device connection direction represented by the device association knowledge limits the connection order of the candidate root cause object, associated device object, and abnormal behavior object. Fault propagation mechanisms and propagation time characteristics define the state transfer relationships between adjacent equipment objects; The abnormal parameters in the abnormal state of the industrial equipment are associated with the corresponding equipment objects or state transmission relationships according to the equipment affiliation and occurrence sequence, so that the candidate root cause propagation relationship (100) satisfies the abnormal parameter coverage and fault propagation sequence constraints.

5. The method for locating the root cause of anomalies in large-scale multi-agent industrial equipment as described in claim 4, characterized in that: The root cause support information includes on-site condition support information and diagnostic knowledge support information; Along the candidate root cause propagation relationship (100), the conformity relationship between the actual operating state of each device object and the corresponding propagation expectation is determined as the node propagation conformity state, and the support relationship between the candidate root cause propagation relationship (100) and historical fault knowledge and reasoning verification information is determined as the path knowledge support state; The reasoning verification information characterizes the matching relationship or counterfactual consistency relationship of the diagnostic rules, and the diagnostic rules define the correspondence between the abnormal state combination and the candidate root cause object or the direction of fault propagation. The root cause evidence state (200) is jointly defined by the node propagation conformity state and the path knowledge support state.

6. The method for locating the root cause of anomalies in large-scale multi-agent industrial equipment as described in claim 5, characterized in that: The evidence gaps include anomalous coverage gaps associated with unexplained anomalous parameters, and propagation mismatch gaps associated with device objects that do not conform to propagation expectations. The abnormal coverage gaps and propagation mismatch gaps are marked in the corresponding candidate root cause propagation relationship (100) to indicate the propagation positions to be corrected; The intermediate device objects or state transmission relationships corresponding to the propagation position to be corrected are defined by the device association knowledge and fault propagation mechanism. The corrected candidate root cause propagation relationship is re-associated with the actual operating state of each device object and the root cause support information to form an updated root cause evidence state (200).

7. The method for locating the root cause of anomalies in large-scale multi-agent industrial equipment as described in claim 6, characterized in that, Also includes: The root cause localization results (300) are correlated with the on-site verification results and user feedback to form a diagnostic verification status; When the diagnostic verification status is confirmed, the corresponding abnormal state of industrial equipment, the target root cause propagation relationship and the root cause support information are associated as new fault cases, and diagnostic rules are formed based on the fault symptom reproduction relationship between different new fault cases. When the diagnostic verification status is a mismatch state, the user feedback is associated with the corresponding evidence gap and the propagation position to be corrected, forming diagnostic correction information for adjusting the inference strategy or the domain adaptation status of the large model.

8. A large-scale multi-agent industrial equipment anomaly root cause localization system, employing the large-scale multi-agent industrial equipment anomaly root cause localization method as described in any one of claims 1 to 7, characterized in that: This includes a propagation relationship generation module, an evidence status correction module, and a root cause location update module; The propagation relationship generation module is used to associate multi-source device information of the same abnormal event, collaboratively form abnormal state and root cause support information of industrial equipment, and construct candidate root cause propagation relationships (100) that are limited by the device connection direction, fault propagation mechanism and occurrence sequence. The evidence status correction module is used to verify the device status and diagnostic knowledge of the candidate root cause propagation relationship (100), locate the propagation position to be corrected based on the evidence gap, and update the candidate root cause propagation relationship and root cause evidence status (200). The root cause localization update module is used to generate root cause localization results (300) based on the updated candidate root cause propagation relationship and root cause evidence status (200), and generate new fault cases or diagnostic correction information based on on-site verification and user feedback.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the large-model multi-agent industrial equipment anomaly root cause localization method as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the method for locating the root cause of anomalies in large-scale multi-agent industrial equipment as described in any one of claims 1 to 7.

Images