A method for monitoring equipment lifecycle faults based on large models and multi-agent systems

Abstract

This invention discloses a method for monitoring equipment lifecycle faults based on a large model and multiple agents, comprising the following steps: acquiring data and performing standardization processing to obtain equipment lifecycle monitoring data; generating a sequence of equipment lifecycle stage identifiers; fine-tuning a language model based on instructions for equipment fault semantic parsing to obtain a set of equipment semantic constraints; obtaining a set of multi-agent state evidence based on a multi-agent analysis system; determining fault candidates to obtain current fault candidate results; constructing an equipment lifecycle fault backtracking tensor; decomposing the equipment lifecycle fault backtracking tensor using a fault inversion tensor train decomposition algorithm to obtain a core tensor sequence; performing reverse core tensor propagation to obtain a fault inversion contribution sequence; and performing continuous path filtering to generate equipment lifecycle fault monitoring results. This invention utilizes tensor inversion technology to achieve equipment fault tracing, offering advantages such as accurate location and clear path identification.

Description

Technical Field

[0001] This invention relates to the field of intelligent fault diagnosis technology, and in particular to a method for monitoring equipment faults throughout its entire lifecycle based on a large model and multiple agents. Background Technology

[0002] With the long-term operation of industrial equipment, energy equipment, and production line equipment, equipment operation data, sensor data, alarm record data, inspection text data, and maintenance text data accumulate continuously. Utilizing artificial intelligence technology for equipment fault monitoring has become an important direction for equipment operation and maintenance. Existing methods typically detect anomalies based on equipment operation data or extract maintenance, alarm, and inspection information through text parsing, and then combine multi-source data to identify faults or predict fault risks in the current equipment status.

[0003] However, the aforementioned methods primarily rely on positive judgments from historical data to the current state or future risks. This makes it difficult to further reverse-engineer the starting component, initial equipment lifecycle stage, and initial abnormal signal after a candidate fault has already appeared on the target device. Furthermore, existing methods do not adequately utilize the correlation between device semantic constraints, multi-agent state evidence, and device lifecycle stages, making it difficult to form a reversible fault formation path and resulting in unclear fault tracing results.

[0004] Therefore, how to provide a method for monitoring equipment failures throughout its entire lifecycle based on large models and multiple agents is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a method for monitoring equipment faults throughout their entire lifecycle based on a large model and multiple agents. This invention utilizes tensor inversion technology to trace the source of equipment faults, and has the advantages of accurate location and clear path.

[0006] A device lifecycle fault monitoring method based on a large model and multiple agents according to an embodiment of the present invention includes the following steps: Obtain raw data of the target device throughout its entire lifecycle and perform standardized processing to obtain monitoring data of the entire lifecycle of the device. Based on the equipment's full lifecycle monitoring data, a sequence of equipment lifecycle stage identifiers is generated. Input the equipment operation and maintenance text data from the equipment lifecycle monitoring data into the instruction fine-tuning language model for equipment fault semantic parsing to obtain the set of equipment semantic constraints. By inputting the equipment's full lifecycle monitoring data, the equipment's lifecycle stage identifier sequence, and the equipment's semantic constraint set into the multi-agent analysis system, a multi-agent state evidence set is obtained. Based on the multi-agent state evidence set and the device semantic constraint set, the current fault candidate result is obtained by determining the fault candidate. Based on the multi-agent state evidence set and the current fault candidate results, construct a fault backtracking tensor for the entire life cycle of the device. The fault inversion tensor train decomposition algorithm is used to decompose the fault backtracking tensor of the entire equipment life cycle to obtain the core tensor sequence; Write the current fault candidate results into the fault category core tensor as the terminal constraint, use the device semantic constraint set as the inverse verification constraint, perform inverse core tensor propagation, and obtain the fault inverse contribution sequence. By filtering continuous paths based on the fault inversion contribution sequence, fault source tracing results are obtained and equipment lifecycle fault monitoring results are generated.

[0007] Optionally, the raw data of the entire life cycle of the equipment includes equipment history data, equipment operation data, and equipment maintenance text data.

[0008] Optionally, the generation of the device lifecycle stage identifier sequence includes: Extract equipment history data and equipment operation data from the equipment lifecycle monitoring data, and divide the equipment lifecycle monitoring data into multiple historical time windows according to the collection time of the equipment operation data; Based on the equipment history data corresponding to each historical time window, determine the history status characteristics of each historical time window; Based on the equipment operation data corresponding to each historical time window, determine the operation status characteristics of each historical time window; Based on the historical status characteristics and operational status characteristics of each historical time window, the stage of each historical time window is determined to obtain the equipment life cycle stage corresponding to each historical time window. Arrange the lifecycle stages of each device in chronological order according to the historical time window, and generate a sequence of device lifecycle stage identifiers.

[0009] Optionally, the generation of the device semantic constraint set includes: Equipment maintenance text data is extracted from the equipment lifecycle monitoring data, and sorted according to the text generation time, associated equipment object, and associated component object to obtain the equipment maintenance text sequence. The equipment maintenance text sequence is cleaned and its fields are standardized to obtain a standardized equipment maintenance text sequence. The standardized equipment operation and maintenance text sequence is input into the instruction fine-tuning language model for equipment fault semantic parsing. The fault object, fault phenomenon, fault category, fault occurrence time window and abnormal duration are identified from the standardized inspection text and standardized historical fault text to obtain equipment fault semantic constraints. Standardized equipment operation and maintenance text sequences are input into a command fine-tuning language model for equipment fault semantic parsing. The maintenance object, maintenance action, maintenance occurrence time, post-maintenance status, and fault recurrence status are identified from the standardized maintenance text to obtain equipment maintenance semantic constraints. Based on the associated equipment objects, associated component objects, and text generation time, the semantic constraints of equipment failure and equipment maintenance are organized to obtain a set of equipment semantic constraints.

[0010] Optionally, the generation of the multi-agent state evidence set includes: Equipment operation data is extracted from the equipment lifecycle monitoring data, and the equipment operation data is organized according to equipment object, component object and historical time window to obtain equipment operation evidence input data; The operational status data from the equipment operation evidence input data is input into the equipment status intelligent agent to generate equipment status evidence; Based on the set of device semantic constraints, determine the associated component objects, and input the device operation evidence input data corresponding to the associated component objects into the component anomaly agent to generate component anomaly evidence; The sensor-collected data from the equipment operation evidence input data is input into the sensor response agent to generate sensor response evidence; Input the operating condition status data from the equipment operation evidence input data into the operating condition disturbance intelligent agent to generate operating condition disturbance evidence. The maintenance and recovery agent is generated by inputting equipment operation evidence data, equipment lifecycle stage identifier sequence, and equipment semantic constraint set into the maintenance and recovery agent. Input the device lifecycle stage identifier sequence into the lifecycle assessment agent to generate lifecycle stage evidence; Based on the equipment object, associated component object, and historical time window, the evidence of equipment status, component anomaly, sensor response, operating condition disturbance, maintenance and recovery, and life cycle stage is aligned to obtain a multi-agent status evidence set.

[0011] Optionally, the generation of the current fault candidate result includes: Extract the state evidence corresponding to the current time window from the multi-agent state evidence set, and align the evidence according to the device object, component object and the current time window to obtain the current time window state evidence group; Extract the semantic constraints corresponding to the current time window from the set of device semantic constraints, and perform semantic alignment according to the device object, component object and the current time window to obtain the semantic constraint group of the current time window; Determine the abnormal pointing results of each component object within the current time window based on the evidence group of the current time window status; Based on the semantic constraint group of the current time window, perform semantic matching on the abnormal pointing results to obtain the semantic matching abnormal pointing results; The current faulty object is determined based on the semantic matching anomaly pointing to the result, and the current fault category and the current fault occurrence time window are determined based on the semantic constraint group of the current time window; The current fault candidate results are composed of the current fault object, the current fault category, and the current fault occurrence time window.

[0012] Optionally, the construction of the device lifecycle fault backtracking tensor includes: Extract equipment objects, component objects, sensor responses, operating conditions, and historical time windows from the equipment lifecycle monitoring data, and determine the equipment lifecycle stage corresponding to each historical time window based on the equipment lifecycle stage identifier sequence; Extract fault categories and maintenance events from the set of device semantic constraints, and extract the current fault object, current fault category, and current fault occurrence time window from the current fault candidate results; A set of fault backtracking dimensions is constructed using equipment objects, component objects, sensor responses, operating conditions, equipment lifecycle stages, historical time windows, maintenance events, and fault categories as fault backtracking dimensions. Based on the fault backtracking dimension set, the evidence dimension of the multi-agent state evidence set is processed to obtain state evidence mapping data; the semantic dimension of the device semantic constraint set is processed to obtain semantic constraint mapping data; and the fault location of the current fault candidate result is processed to obtain the current fault mapping data. Write the state evidence mapping data, semantic constraint mapping data and current fault mapping data into the tensor positions corresponding to the fault backtracking dimension set to form the initial fault backtracking tensor. Based on the equipment object, component object, historical time window, and equipment lifecycle stage, the state evidence mapping data, semantic constraint mapping data, and current fault mapping data in the initial fault backtracking tensor are organized according to their positions to obtain the equipment full lifecycle fault backtracking tensor.

[0013] Optionally, the generation of the core tensor sequence includes: The fault backtracking tensor of the entire equipment life cycle is input into the fault inversion tensor train decomposition algorithm. The fault inversion tensor train decomposition algorithm includes a dimension arrangement layer, a dimension-by-dimensional expansion layer, a connection rank determination layer, a core tensor generation layer, and a core tensor concatenation layer. In the dimensional arrangement layer, the fault backtracking tensor of the entire equipment life cycle is arranged dimensionally according to the order of equipment object, component object, sensor response, operating condition, equipment life cycle stage, historical time window, maintenance event and fault category, to obtain the inverse dimensional arrangement tensor. In the dimension-by-dimensional expansion layer, the dimension-by-dimensional matrix expansion process is performed on the tensors arranged according to the dimensional order of the inverse dimension to obtain a set of dimension-by-dimensional expansion matrices. In the connection rank determination layer, the tensor train connection rank between adjacent dimensions is determined based on the non-zero position correspondence between adjacent expansion matrices in the set of dimension-by-dimensional expansion matrices, the evidence distribution density, and the fault category correlation strength, thus obtaining the connection order column; In the core tensor generation layer, based on the set of dimension-wise expansion matrices and the connection order column, the equipment object expansion matrix, component object expansion matrix, sensor response expansion matrix, operating condition expansion matrix, equipment life cycle stage expansion matrix, historical time window expansion matrix, maintenance event expansion matrix, and fault category expansion matrix are sequentially subjected to low-rank decomposition processing to obtain the equipment object core tensor, component object core tensor, sensor response core tensor, operating condition core tensor, equipment life cycle stage core tensor, historical time window core tensor, maintenance event core tensor, and fault category core tensor. In the core tensor concatenation layer, the core tensors of the equipment object, component object, sensor response, operating condition, equipment life cycle stage, historical time window, maintenance event, and fault category are sequentially connected according to the connection order to obtain the core tensor sequence.

[0014] Optionally, the generation of the fault inversion contribution sequence includes: Write the current fault candidate results into the fault category core tensor in the core tensor sequence to obtain the terminal constraint core tensor. The set of semantic constraints of the equipment is organized into inverse verification constraints. The core tensor of the end constraint is used as the starting point for the inverse core tensor propagation. According to the reverse connection order from the core tensor of the fault category to the core tensor of the equipment object in the core tensor sequence, the core tensor of the maintenance event, the core tensor of the historical time window, the core tensor of the equipment life cycle stage, the core tensor of the operating condition, the core tensor of the sensor response, and the core tensor of the component object are passed in sequence to obtain the initial inverse contribution connection set. The initial inversion contribution connection set is verified according to the inversion verification constraints. The initial inversion contribution connections that meet the inversion verification constraints are retained to obtain the verified inversion contribution connection set. The fault inversion contribution sequence is obtained by arranging the check inversion contribution set according to the reverse connection order of the core tensor sequence.

[0015] Optionally, the generation of the equipment's full lifecycle fault monitoring results includes: According to the contribution order in the fault inversion contribution sequence, the contribution connection relationship between adjacent contributing nodes is extracted. Based on the continuity of the contribution connection relationship, the continuous path of the fault inversion contribution sequence is filtered to obtain the fault formation path. Identify the initiating component, the initial equipment lifecycle stage, and the initial abnormal signal based on the fault formation path; The fault tracing result consists of the starting component, the initial equipment life cycle stage, the initial abnormal signal, and the fault formation path; Based on the fault tracing results, generate equipment lifecycle fault monitoring results.

[0016] The beneficial effects of this invention are: This invention constructs a fault backtracking tensor for the entire equipment lifecycle by combining equipment lifecycle monitoring data, equipment lifecycle stage identifier sequences, equipment semantic constraint sets, multi-agent state evidence sets, and current fault candidate results. This allows equipment history, operating status, sensor responses, operating condition changes, maintenance records, and fault semantics to form a corresponding relationship within a unified data structure. Therefore, this invention no longer relies solely on a single operating parameter or alarm information for fault diagnosis, but rather integrates the state changes of the equipment across different historical time windows, different component objects, and different equipment lifecycle stages, improving the data support completeness and reliability of fault monitoring results.

[0017] This invention employs a fault inversion tensor train decomposition algorithm to decompose the fault backtracking tensor of the entire equipment lifecycle. This yields a core tensor sequence comprising core tensors for the equipment object, component object, sensor response, operating condition, equipment lifecycle stage, historical time window, maintenance event, and fault category. The current fault candidate result is written into the fault category core tensor as an end constraint, and reverse core tensor propagation is performed from the fault category core tensor towards the equipment object core tensor. Through this processing method, this invention can trace the fault origin backward from the current fault candidate result, avoiding the problem of traditional methods that rely solely on forward fault prediction from historical data and struggle to explain the fault initiation point.

[0018] This invention also verifies the reverse propagation direction and contribution connection relationship through a set of device semantic constraints, and performs continuous path filtering by combining the fault inversion contribution sequence, thereby obtaining the starting component, the starting equipment life cycle stage, the initial abnormal signal, and the fault formation path. Therefore, this invention can identify weak abnormal signals that have participated in fault evolution but did not directly form obvious alarms in the early stages, expanding equipment fault monitoring results from simple fault indications to fault tracing, path reconstruction, and maintenance assistance, resulting in more accurate fault location, clearer tracing paths, and more targeted maintenance checks. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of a device lifecycle fault monitoring method based on a large model and multiple agents proposed in this invention; Figure 2 This is a schematic diagram illustrating the construction of the equipment lifecycle fault backtracking tensor in the equipment lifecycle fault monitoring method based on large models and multiple agents proposed in this invention. Figure 3 This is a schematic diagram illustrating the reverse core tensor propagation using the current fault candidate result as the end constraint in a device lifecycle fault monitoring method based on a large model and multiple agents proposed in this invention. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0021] refer to Figures 1-3 A method for monitoring equipment lifecycle faults based on large models and multi-agent systems includes the following steps: Acquire raw data of the target device throughout its entire lifecycle, standardize the raw data to obtain monitoring data of the device throughout its entire lifecycle. The raw data of the device throughout its entire lifecycle includes device history data, device operation data and device maintenance text data. Based on the equipment lifecycle monitoring data, determine the equipment lifecycle stage corresponding to each historical time window of the target equipment, and generate an equipment lifecycle stage identifier sequence; Input the equipment operation and maintenance text data from the equipment lifecycle monitoring data into the instruction fine-tuning language model for equipment fault semantic parsing to obtain the set of equipment semantic constraints. The equipment's full lifecycle monitoring data, equipment lifecycle stage identifier sequence, and equipment semantic constraint set are input into the multi-agent analysis system to obtain a multi-agent state evidence set, which includes equipment state evidence, component anomaly evidence, sensor response evidence, operating condition disturbance evidence, maintenance and recovery evidence, and lifecycle stage evidence. Based on the multi-agent state evidence set and the device semantic constraint set, the operating state of the target device in the current time window is determined to identify fault candidates, and the current fault candidate results are obtained. Based on the equipment lifecycle monitoring data, the equipment lifecycle stage identifier sequence, the equipment semantic constraint set, the multi-agent state evidence set, and the current fault candidate results, a fault backtracking tensor for the equipment lifecycle is constructed. The fault inversion tensor train decomposition algorithm is used to decompose the fault backtracking tensor of the entire equipment life cycle to obtain the core tensor sequence. The core tensor sequence includes the core tensor of equipment object, core tensor of component object, core tensor of sensor response, core tensor of operating condition, core tensor of equipment life cycle stage, core tensor of historical time window, core tensor of maintenance event, and core tensor of fault category. Write the current fault candidate result into the fault category core tensor as the terminal constraint, use the device semantic constraint set as the inverse verification constraint, and perform inverse core tensor propagation from the fault category core tensor to the device object core tensor in the core tensor sequence to obtain the fault inverse contribution sequence. Continuous path filtering is performed based on the fault inversion contribution sequence to obtain fault source tracing results. Based on the fault source tracing results, equipment full life cycle fault monitoring results are generated. The fault source tracing results include the starting component, the starting equipment life cycle stage, the initial abnormal signal, and the fault formation path.

[0022] In this embodiment, the original data of the entire life cycle of the equipment includes equipment history data, equipment operation data, and equipment maintenance text data.

[0023] In this embodiment, the generation of the device lifecycle stage identifier sequence includes: Extract equipment history data and equipment operation data from the equipment lifecycle monitoring data, and divide the equipment lifecycle monitoring data into multiple historical time windows according to the collection time of the equipment operation data; Based on the equipment history data corresponding to each historical time window, the history status characteristics of each historical time window are determined. The history status characteristics include commissioning time, cumulative running time, and component replacement status. Based on the equipment operation data corresponding to each historical time window, the operation status characteristics of each historical time window are determined. The operation status characteristics include operation fluctuation status, alarm occurrence status, and operation stability status. Based on the historical status characteristics and operational status characteristics of each historical time window, the stage of each historical time window is determined to obtain the equipment life cycle stage corresponding to each historical time window. Specifically, during the phase determination, the commissioning duration, cumulative operating time, and component replacement status are used as history determination factors, while the operating fluctuation status, alarm occurrence status, and stable operating status are used as operating determination factors. When the commissioning duration is within the initial operating range and the operating fluctuation status shows a continuous convergence trend, the corresponding historical time window is determined as the commissioning period. When the cumulative operating time increases continuously and the stable operating status meets the stable operating conditions, the corresponding historical time window is determined as the stable operating period. When the operating fluctuation status, alarm occurrence status, and component replacement status all show signs of degradation within a continuous historical time window, the corresponding historical time window is determined as the performance degradation period or maintenance recovery period. Arrange the lifecycle stages of each device in chronological order according to the historical time window, and generate a sequence of device lifecycle stage identifiers.

[0024] In this embodiment, the generation of the device semantic constraint set includes: Equipment maintenance text data is extracted from the equipment lifecycle monitoring data, and sorted according to the text generation time, associated equipment object, and associated component object to obtain the equipment maintenance text sequence. Text cleaning and field unification processing are performed on the equipment operation and maintenance text sequence to obtain a standardized equipment operation and maintenance text sequence, which includes standardized inspection text, standardized maintenance text, and standardized historical fault text. The standardized equipment operation and maintenance text sequence is input into the instruction fine-tuning language model for equipment fault semantic parsing. The fault object, fault phenomenon, fault category, fault occurrence time window and abnormal duration are identified from the standardized inspection text and standardized historical fault text to obtain equipment fault semantic constraints. Among them, the instruction fine-tuning language model for equipment fault semantic parsing reads standardized inspection text and standardized historical fault text, first extracts equipment object words, component object words, abnormal status words and alarm description words from the text, then maps equipment object words and component object words to fault objects respectively, and merges abnormal status words and alarm description words into fault phenomena; fault phenomena with the same or similar semantics under the same fault object are grouped into the same fault category, and the fault occurrence time window and abnormal duration are determined based on the text generation time, duration description and recurrence description; Standardized equipment operation and maintenance text sequences are input into a command fine-tuning language model for equipment fault semantic parsing. The maintenance object, maintenance action, maintenance occurrence time, post-maintenance status, and fault recurrence status are identified from the standardized maintenance text to obtain equipment maintenance semantic constraints. Among them, the instruction fine-tuning language model extracts maintenance object words, maintenance action words, and maintenance time words from standardized maintenance text to form maintenance object, maintenance action, and maintenance occurrence time; reads retest description, recovery description, and re-alarm description from maintenance text to obtain the post-maintenance status; when the same maintenance object reappears in the text after the maintenance occurrence time with the same fault category or the same fault phenomenon, it is identified as a fault recurrence status. Based on the associated equipment objects, associated component objects, and text generation time, the semantic constraints of equipment failure and equipment maintenance are organized to obtain a set of equipment semantic constraints.

[0025] In this embodiment, the generation of the multi-agent state evidence set includes: Equipment operation data is extracted from the equipment lifecycle monitoring data, and the equipment operation data is organized according to equipment object, component object and historical time window to obtain equipment operation evidence input data. The equipment operation evidence input data includes operation status data, sensor data, operating condition data and alarm record data corresponding to equipment object, component object and historical time window. The operating status data in the equipment operation evidence input data is input into the equipment status intelligent agent. The equipment status intelligent agent generates equipment status evidence based on the direction, magnitude, duration of change of equipment operation data in each historical time window and the continuity of change between adjacent historical time windows. The device status agent encodes the direction, magnitude, and duration of changes in operating status data between adjacent historical time windows to obtain operating status change features. When the operating status change features maintain the same direction of change in consecutive historical time windows, corresponding device status evidence is generated. The device status evidence includes at least the device object, historical time window, direction of operating status change, and operating status persistence identifier. The associated component objects are determined based on the set of semantic constraints of the equipment. The equipment operation evidence input data corresponding to the associated component objects is input into the component anomaly agent. The component anomaly agent generates component anomaly evidence based on the changes in operation deviation, duration of deviation and recovery status of the associated component objects in adjacent historical time windows. The component anomaly agent calculates the difference between the running status data of the associated component object in the current historical time window and the running status data in the adjacent historical time windows to obtain the component running deviation value; when the component running deviation value changes in the same direction in consecutive historical time windows, the deviation duration is recorded; when the component running deviation value falls back to the running status range of the corresponding component object in subsequent historical time windows, the deviation recovery status is recorded, and component anomaly evidence is generated accordingly. The sensor data collected from the equipment operation evidence input data is input into the sensor response agent. The sensor response agent generates sensor response evidence based on the sensor response changes, response delays, response duration, and response direction consistency corresponding to the same associated component object. The sensor response agent reads sensor data collected by the same associated component object according to the historical time window, calculates the sensor response change between adjacent historical time windows, compares the historical time window in which the sensor response change first appears with the historical time window corresponding to the component anomaly evidence, and obtains the response delay. When multiple sensor data collections show the same rising or falling direction or the same fluctuation direction in a continuous historical time window, the consistency of the response direction is determined and sensor response evidence is generated. The operating condition status data in the equipment operation evidence input data is input into the operating condition disturbance intelligent agent, which generates operating condition disturbance evidence based on load changes, start-stop changes and environmental changes in each historical time window. The equipment operation evidence input data, the equipment life cycle stage identifier sequence, and the equipment semantic constraint set are input into the maintenance and recovery intelligent agent. The maintenance and recovery intelligent agent generates maintenance and recovery evidence based on the maintenance semantic content in the equipment semantic constraint set, the maintenance occurrence time, and the changes in equipment operation data before and after the maintenance occurrence time. Among them, the maintenance and recovery agent takes the maintenance occurrence time in the semantic constraints of equipment maintenance as the dividing point, extracts the equipment operation data before and after the maintenance occurrence time, compares the changes in the operation status data, sensor acquisition data and alarm record data before and after maintenance, and obtains the state difference before and after maintenance; when there are still operation deviations, sensor abnormalities or alarm records corresponding to the maintenance object after maintenance, maintenance and recovery evidence that characterizes insufficient maintenance and recovery is generated. The device lifecycle stage identifier sequence is input into the lifecycle assessment agent, which then generates lifecycle stage evidence according to the correspondence between historical time windows and device lifecycle stages. Based on the equipment object, associated component object, and historical time window, the evidence of equipment status, component anomaly, sensor response, operating condition disturbance, maintenance and recovery, and life cycle stage is aligned to obtain a multi-agent status evidence set.

[0026] In this embodiment, the generation of the current fault candidate result includes: Extract the state evidence corresponding to the current time window from the multi-agent state evidence set, and align the evidence according to the device object, component object and the current time window to obtain the current time window state evidence group; Extract the semantic constraints corresponding to the current time window from the set of device semantic constraints, and perform semantic alignment according to the device object, component object and the current time window to obtain the semantic constraint group of the current time window; Based on the current time window status evidence group, determine the abnormal pointing results of each component object within the current time window. The abnormal pointing results include the abnormal pointing component object, the abnormal pointing sensor response, and the abnormal pointing operation status. When determining the anomaly pointing result, first extract the equipment status evidence, component anomaly evidence, and sensor response evidence corresponding to the same component object from the current time window status evidence group; when at least two of the three types of evidence point to the same component object, the component object is determined as the anomaly pointing component object; the sensor response evidence corresponding to the anomaly pointing component object is taken as the anomaly pointing sensor response, and the equipment status evidence corresponding to the anomaly pointing component object is taken as the anomaly pointing operating status. Based on the semantic constraint group of the current time window, semantic matching is performed on the abnormal pointing results to obtain the semantic matching abnormal pointing results. The semantic matching abnormal pointing results include the abnormal pointing component object, abnormal pointing sensor response, and abnormal pointing operation status that match the semantic constraint group of the current time window. During semantic matching, the abnormal pointing component object is matched with the fault object in the semantic constraint group of the current time window, and the abnormal pointing sensor response and abnormal pointing operating status are matched with the fault phenomenon in the semantic constraint group of the current time window. When both object matching and semantic matching are successful, the corresponding abnormal pointing result is retained as the semantic matching abnormal pointing result. When neither object matching nor semantic matching is successful, the abnormal pointing result is not used as the current fault candidate. The current faulty object is determined based on the semantic matching anomaly pointing to the result, and the current fault category and the current fault occurrence time window are determined based on the semantic constraint group of the current time window; The current fault candidate results are composed of the current fault object, the current fault category, and the current fault occurrence time window.

[0027] In this embodiment, the construction of the equipment lifecycle fault backtracking tensor includes: Extract equipment objects, component objects, sensor responses, operating conditions, and historical time windows from the equipment lifecycle monitoring data, and determine the equipment lifecycle stage corresponding to each historical time window based on the equipment lifecycle stage identifier sequence; Extract fault categories and maintenance events from the set of device semantic constraints, and extract the current fault object, current fault category, and current fault occurrence time window from the current fault candidate results; A set of fault backtracking dimensions is constructed using equipment objects, component objects, sensor responses, operating conditions, equipment lifecycle stages, historical time windows, maintenance events, and fault categories as fault backtracking dimensions. According to the fault backtracking dimension set, the evidence dimension of the multi-agent state evidence set is processed to obtain state evidence mapping data. The state evidence mapping data includes equipment state mapping data formed by equipment state evidence, component anomaly mapping data formed by component anomaly evidence, sensor response mapping data formed by sensor response evidence, operating condition disturbance mapping data formed by operating condition disturbance evidence, maintenance and recovery mapping data formed by maintenance and recovery evidence, and life cycle stage mapping data formed by life cycle stage evidence. In the evidence dimension positioning process, each piece of status evidence is written to a tensor based on its corresponding device object, component object, and historical time window. Device status evidence is written at the intersection of the device object dimension and the historical time window dimension; component anomaly evidence is written at the intersection of the component object dimension and the historical time window dimension; sensor response evidence is written at the intersection of the sensor response dimension and the historical time window dimension; and operating condition disturbance evidence, maintenance recovery evidence, and life cycle stage evidence are written to the corresponding positions in the operating condition status dimension, maintenance event dimension, and device life cycle stage dimension, respectively. Based on the fault backtracking dimension set, the semantic dimension alignment processing of the equipment semantic constraint set is performed to obtain semantic constraint mapping data. The semantic constraint mapping data includes fault category mapping data formed by fault category alignment and maintenance event mapping data formed by maintenance event alignment. In the semantic dimension alignment process, the fault objects in the device semantic constraint set are mapped to the component object dimension, the fault category is written into the fault category dimension, and the maintenance objects and maintenance actions in the device semantic constraint set are written into the maintenance event dimension. When the fault category or maintenance event has a text generation time, the text generation time is mapped to the corresponding historical time window, and a correspondence is established between the fault category dimension, the maintenance event dimension, and the historical time window dimension. Based on the fault backtracking dimension set, the current fault candidate results are processed to return the fault location to obtain the current fault mapping data. The current fault mapping data includes the current fault object mapping data formed by returning the current fault object to its location, the current fault category mapping data formed by returning the current fault category to its location, and the current fault time mapping data formed by returning the current fault occurrence time window to its location. During the fault location relocation process, the current fault object is mapped to the device object dimension or component object dimension, the current fault category is mapped to the fault category dimension, and the current fault occurrence time window is mapped to the historical time window dimension. The fault end identifier is written to the tensor position that corresponds to the current fault object, the current fault category, and the current fault occurrence time window to obtain the current fault mapping data. Write the state evidence mapping data, semantic constraint mapping data and current fault mapping data into the tensor positions corresponding to the fault backtracking dimension set to form the initial fault backtracking tensor. Based on the equipment object, component object, historical time window, and equipment life cycle stage, the state evidence mapping data, semantic constraint mapping data, and current fault mapping data in the initial fault backtracking tensor are organized according to their positions to obtain the equipment full life cycle fault backtracking tensor. When organizing the location, the same equipment object, the same component object, and the same historical time window are used as indices to align the state evidence mapping data, semantic constraint mapping data, and current fault mapping data. When state evidence mapping data and semantic constraint mapping data exist under the same index, the evidence semantic association is established at the corresponding tensor position. When current fault mapping data exists under the same index, the position is marked as the fault inversion endpoint position, thus obtaining the fault backtracking tensor of the entire equipment life cycle.

[0028] In this embodiment, the generation of the core tensor sequence includes: The fault backtracking tensor of the entire equipment life cycle is input into the fault inversion tensor train decomposition algorithm. The fault inversion tensor train decomposition algorithm includes a dimension arrangement layer, a dimension-by-dimensional expansion layer, a connection rank determination layer, a core tensor generation layer, and a core tensor concatenation layer. The internal connection relationship of the fault inversion tensor train decomposition algorithm consists of a dimensional arrangement layer, a dimensional expansion layer, a connection rank determination layer, a core tensor generation layer, and a core tensor concatenation layer connected in sequence. Among them, the fault backtracking tensor of the entire equipment life cycle first enters the dimensional arrangement layer, forming an inversion dimensional arrangement tensor according to the order of equipment object, component object, sensor response, operating condition, equipment life cycle stage, historical time window, maintenance event, and fault category. Then it enters the dimensional expansion layer to obtain a set of dimensional expansion matrices. The connection rank determination layer generates a connection order column based on the non-zero position correspondence between adjacent expansion matrices, evidence distribution density, and fault category correlation strength. The core tensor generation layer generates core tensors of equipment object, component object, sensor response, operating condition, equipment life cycle stage, historical time window, maintenance event, and fault category based on the dimensional expansion matrix set and the connection order column. The core tensor concatenation layer then connects the above core tensors in sequence to form a core tensor sequence. During training, the input data is a historical equipment full life cycle fault backtracking tensor. The data format is a multi-dimensional tensor composed of equipment object dimension, component object dimension, sensor response dimension, operating condition dimension, equipment life cycle stage dimension, historical time window dimension, maintenance event dimension, and fault category dimension. Each tensor position is written with state evidence mapping data, semantic constraint mapping data, and current fault mapping data. The training data comes from the equipment history data, equipment operation data, equipment maintenance text data, multi-agent state evidence set, historical fault records and post-maintenance verification records of the target equipment and similar equipment. The annotation method is to annotate the historical fault samples with the current fault candidate result, starting component, starting equipment life cycle stage, initial abnormal signal and fault formation path based on the maintenance confirmation result, manual review result and fault recurrence record. The loss function is composed of the reconstruction error of the fault backtracking tensor throughout the equipment's life cycle, the expression error of the core tensor sequence to the fault formation path, the end constraint error after the current fault candidate result is written into the fault category core tensor, the semantic consistency error between the equipment semantic constraint set and the inverse contribution connection, the starting component location error, and the starting equipment life cycle stage location error. The training parameters include the upper limit of the core tensor connection rank, the number of input tensors in each batch, the number of training rounds, the initial learning step size, the connection rank adjustment interval, and the number of early stop rounds. During training, the arrangement order of the core tensors is fixed first, and then the internal parameters of each core tensor and the connection rank between adjacent core tensors are updated. When the total loss decreases below the convergence threshold in multiple consecutive training rounds, and the fault formation path integrity rate, the starting component location accuracy, and the starting equipment life cycle stage location accuracy in the verification samples no longer improve, the fault inverse tensor train decomposition algorithm is considered to have completed training. The fault inversion tensor train decomposition algorithm writes the current fault candidate result into the fault category core tensor as the terminal constraint, and then uses the equipment semantic constraint set as the inversion verification constraint. It performs inverse core tensor propagation from the fault category core tensor to the equipment object core tensor to obtain the fault inversion contribution sequence. Through this improvement, tensor train decomposition no longer only expresses the low-rank relationship between high-dimensional data, but can trace back from the current fault category to the starting component, the starting equipment life cycle stage and the initial abnormal signal in the historical time window, and form a fault formation path. This solves the problem that traditional equipment fault monitoring can only identify the current abnormality and it is difficult to reverse the fault starting point and the source of early weak abnormality. In the dimensional arrangement layer, the fault backtracking tensor of the entire equipment life cycle is arranged dimensionally according to the order of equipment object, component object, sensor response, operating condition, equipment life cycle stage, historical time window, maintenance event and fault category, to obtain the inverse dimensional arrangement tensor. In the dimension-by-dimensional expansion layer, the dimension-by-dimensional matrix expansion process is performed on the tensors arranged according to the dimensional order of the inverse dimension arrangement to obtain a set of dimension-by-dimensional expansion matrices. The set of dimension-by-dimensional expansion matrices includes the equipment object expansion matrix, component object expansion matrix, sensor response expansion matrix, operating condition expansion matrix, equipment life cycle stage expansion matrix, historical time window expansion matrix, maintenance event expansion matrix, and fault category expansion matrix. In the stepwise matrix expansion process, the inverse dimension arrangement tensor is expanded into a matrix by using the current dimension as the row index and combining the remaining dimensions after the current dimension as the column index. This matrix expansion process is repeated in the order of equipment object, component object, sensor response, operating condition, equipment life cycle stage, historical time window, maintenance event and fault category to obtain a set of stepwise expanded matrices. In the connection rank determination layer, the tensor train connection rank between adjacent dimensions is determined based on the non-zero position correspondence between adjacent expansion matrices in the set of dimension-by-dimensional expansion matrices, the evidence distribution density, and the fault category correlation strength, thus obtaining the connection order column; The generation of the connection order column specifically includes: for each pair of adjacent expanded matrices in the set of dimension-wise expanded matrices, firstly, reading the non-zero positions of the written data in the adjacent expanded matrices, and indexing and matching the non-zero positions according to device objects, component objects, historical time windows, and fault categories, counting the number of non-zero positions with the same index in the two pairs of adjacent expanded matrices, and obtaining the corresponding values ​​of the non-zero positions; then, counting the number of written positions of state evidence mapping data, semantic constraint mapping data, and current fault mapping data in the adjacent expanded matrices, and calculating the proportion of the number of written positions in all writable positions in the adjacent expanded matrices, obtaining the evidence distribution density value; then, based on the fault category mapping data, counting its relationship with the device... The number of times that state mapping data, component anomaly mapping data, sensor response mapping data, and maintenance event mapping data form a corresponding relationship within the same historical time window is calculated, and the proportion of this number in the number of occurrences of fault category mapping data is calculated to obtain the fault category association strength value. Then, the non-zero position correspondence value, evidence distribution density value, and fault category association strength value are normalized to the same value interval, and weighted summation is performed according to the set weights to obtain the adjacent dimension connection evaluation value. Finally, based on the connection rank interval where the adjacent dimension connection evaluation value is located, the tensor train connection rank corresponding to this set of adjacent expansion matrices is determined, and the connection order column is calculated sequentially according to the arrangement order of the dimension-by-dimensional expansion matrix set. In the core tensor generation layer, based on the set of dimension-wise expansion matrices and the connection order column, the equipment object expansion matrix, component object expansion matrix, sensor response expansion matrix, operating condition expansion matrix, equipment life cycle stage expansion matrix, historical time window expansion matrix, maintenance event expansion matrix, and fault category expansion matrix are sequentially subjected to low-rank decomposition processing to obtain the equipment object core tensor, component object core tensor, sensor response core tensor, operating condition core tensor, equipment life cycle stage core tensor, historical time window core tensor, maintenance event core tensor, and fault category core tensor. In the low-rank decomposition process, singular value decomposition or low-rank matrix decomposition is performed on each expansion matrix in the set of dimension-wise expansion matrices, and the corresponding number of decomposition components are truncated according to the connection order. The truncated decomposition components are rearranged according to the current dimension size, the previous connection rank and the next connection rank to generate the corresponding core tensor. Each expansion matrix is ​​processed in sequence to obtain the device object core tensor to the fault category core tensor. In the core tensor concatenation layer, the core tensors of the equipment object, component object, sensor response, operating condition, equipment life cycle stage, historical time window, maintenance event, and fault category are sequentially connected according to the connection order to obtain the core tensor sequence.

[0029] In this embodiment, the generation of the fault inversion contribution sequence includes: Write the current fault candidate results into the fault category core tensor in the core tensor sequence to obtain the terminal constraint core tensor. Specifically, when writing the current fault candidate result into the fault category core tensor, the fault category index corresponding to the current fault category is used as the writing position, and the current fault object and the current fault occurrence time window are used as additional constraints and written to the tensor connection position connected to the fault category index. After writing, the fault category connections in the fault category core tensor that do not correspond to the current fault category are suppressed to obtain the terminal constraint core tensor. The set of device semantic constraints is organized into inverse verification constraints, which are used to verify the backward propagation direction and contribution connection relationship in the core tensor sequence. Among them, the inverse verification constraints include fault object verification constraints, fault category verification constraints, maintenance event verification constraints, and time window verification constraints; the fault object verification constraints are derived from the fault objects in the equipment semantic constraint set, the fault category verification constraints are derived from the fault categories in the equipment semantic constraint set, the maintenance event verification constraints are derived from the maintenance objects and maintenance actions in the equipment semantic constraint set, and the time window verification constraints are derived from the text generation time and the fault occurrence time window in the equipment semantic constraint set. Using the end-constraint core tensor as the starting point for reverse core tensor propagation, and following the reverse connection order from the fault category core tensor to the equipment object core tensor in the core tensor sequence, the core tensor passes through the maintenance event core tensor, the historical time window core tensor, the equipment life cycle stage core tensor, the operating condition core tensor, the sensor response core tensor, and the component object core tensor in sequence to obtain the initial inverse contribution connection set. During the reverse propagation of the core tensor, the current fault category index in the end-constraint core tensor is used as the starting point, and the tensor connection contribution between adjacent core tensors is calculated sequentially along the reverse connection order of the core tensor sequence. For each core tensor, the contribution connection that has a connection relationship with the current fault category index, the current fault occurrence time window, or the current fault object is retained. The contribution connections obtained from the fault category core tensor to the device object core tensor are collected in the propagation order to form the initial reverse contribution connection set. The initial inversion contribution connection set is verified according to the inversion verification constraints. The initial inversion contribution connections that meet the inversion verification constraints are retained to obtain the verified inversion contribution connection set. Specifically, when validating the initial inversion contribution connection set, the device object, component object, maintenance event, historical time window, and fault category corresponding to each initial inversion contribution connection are matched with the inversion validation constraints. When the initial inversion contribution connection simultaneously satisfies the requirements of consistent fault category, corresponding historical time window, and consistent device object or component object, it is retained as a valid inversion contribution connection. When the initial inversion contribution connection does not match the inversion validation constraints, it is removed from the initial inversion contribution connection set. The fault inversion contribution sequence is obtained by arranging the check inversion contribution set according to the reverse connection order of the core tensor sequence.

[0030] In this embodiment, the generation of equipment lifecycle fault monitoring results includes: Extract the contribution connection relationship between adjacent contributing nodes according to the contribution order in the fault inversion contribution sequence; Among them, the contributing nodes are the nodes in the fault inversion contributing sequence that correspond to the equipment object, component object, sensor response, operating condition, equipment life cycle stage, historical time window, maintenance event or fault category. The contributing connection relationship between adjacent contributing nodes includes the connection direction, connection order and connection strength of adjacent contributing nodes. The above contributing connection relationship is extracted according to the contribution order in the fault inversion contributing sequence. Based on the continuity of the contribution connection relationship, the fault inversion contribution sequence is filtered for continuous paths to obtain the fault formation path; During continuous path filtering, the adjacent contributing nodes are sequentially judged according to the contribution order of the fault inversion contribution sequence to determine whether they have a continuous historical time window, the same equipment object, or the same component object correspondence. When adjacent contributing nodes are continuous in the historical time window and maintain a correspondence in the equipment object or component object, the adjacent contributing nodes are merged into the same fault formation path. When adjacent contributing nodes do not meet the continuity requirement, the path is terminated and path filtering is restarted. Identify the initiating component, the initial equipment lifecycle stage, and the initial abnormal signal based on the fault formation path; In the fault formation path, the earliest contributing node with abnormal contributing connection is found in the order from the earliest to the latest historical time window. The component object corresponding to the contributing node is determined as the starting component, the equipment life cycle stage corresponding to the contributing node is determined as the starting equipment life cycle stage, and the sensor response or operating status change corresponding to the contributing node is determined as the initial abnormal signal. When multiple contributing nodes are in the same earliest historical time window, the contributing node with the strongest connection strength is selected as the basis for determination. The fault tracing result consists of the starting component, the initial equipment life cycle stage, the initial abnormal signal, and the fault formation path; Based on the fault tracing results, generate equipment lifecycle fault monitoring results.

[0031] Example 1: To verify the feasibility of this invention in practice, it was applied to a group of continuous production equipment in a large manufacturing park. Multiple electromechanical interconnected devices in this park operate continuously, generating equipment history data, equipment operation data, and equipment maintenance text data. Equipment history data includes commissioning records, cumulative operation records, and component replacement records. Equipment operation data includes operating status data, sensor-collected data, operating condition data, and alarm record data. Equipment maintenance text data includes inspection texts, maintenance texts, and historical fault texts. In this scenario, some devices exhibit potential fault phenomena within the current time window, but based solely on current alarm records or changes in a single sensor, it is difficult to determine the earliest corresponding starting component, the initial equipment lifecycle stage, and the initial abnormal signal.

[0032] In the application process, the raw data of the target device's entire lifecycle is first acquired and standardized to obtain the device's entire lifecycle monitoring data. Standardization includes time alignment of data from different sources, unification of device and component objects, anomaly data cleanup, and organization of equipment maintenance texts, ensuring that operational data, inspection texts, maintenance texts, and historical fault texts can be mapped to the same historical time window. Subsequently, based on the device's entire lifecycle monitoring data, the corresponding device lifecycle stage for each historical time window is determined, generating a sequence of device lifecycle stage identifiers. This allows the device status to be distinguished according to the commissioning period, stable operation period, performance degradation period, maintenance and recovery period, and end-of-life period.

[0033] Next, the equipment operation and maintenance text data is input into a fine-tuning language model for equipment fault semantic parsing, resulting in a set of equipment semantic constraints. This model identifies fault objects, fault phenomena, fault categories, fault occurrence time windows, and abnormal durations from standardized inspection texts and standardized historical fault texts, and identifies maintenance objects, maintenance actions, maintenance occurrence times, post-maintenance status, and fault recurrence status from standardized maintenance texts. Through this processing, fault descriptions and maintenance descriptions, originally scattered in manual records, are transformed into a set of equipment semantic constraints that can participate in subsequent analysis.

[0034] Subsequently, the equipment's full lifecycle monitoring data, equipment lifecycle stage identifier sequences, and equipment semantic constraint sets are input into the multi-agent analysis system to obtain a multi-agent state evidence set. The equipment state agent generates equipment state evidence, the component anomaly agent generates component anomaly evidence, the sensor response agent generates sensor response evidence, the operating condition disturbance agent generates operating condition disturbance evidence, the maintenance and recovery agent generates maintenance and recovery evidence, and the lifecycle assessment agent generates lifecycle stage evidence. All types of evidence are aligned according to equipment objects, component objects, and historical time windows for subsequent determination of current fault candidate results.

[0035] Within the current time window, fault candidates are determined based on the multi-agent state evidence set and the device semantic constraint set, resulting in the current fault candidate result. This current fault candidate result is not directly used as the final conclusion, but rather as the final constraint in the fault inversion process. Subsequently, the device's full lifecycle monitoring data, the device lifecycle stage identifier sequence, the device semantic constraint set, the multi-agent state evidence set, and the current fault candidate result are jointly used to construct the device's full lifecycle fault backtracking tensor, forming a unified backtracking data structure among device objects, component objects, sensor responses, operating conditions, device lifecycle stages, historical time windows, maintenance events, and fault categories.

[0036] Then, the fault inversion tensor train decomposition algorithm is used to decompose the fault backtracking tensor of the entire equipment lifecycle, obtaining a core tensor sequence. The current fault candidate result is written into the fault category core tensor as the terminal constraint, and the equipment semantic constraint set is used as the inversion verification constraint. Inverse core tensor propagation is performed from the fault category core tensor to the equipment object core tensor to obtain the fault inversion contribution sequence. This process can trace back the abnormal contributions in the historical time window from the current fault candidate result, rather than judging the fault only based on the current moment.

[0037] Finally, continuous path filtering is performed based on the fault inversion contribution sequence to obtain the fault tracing results. These results include the starting component, the initial equipment lifecycle stage, the initial abnormal signal, and the fault formation path. Based on these results, maintenance personnel can clearly identify which component the current fault likely originated from, which equipment lifecycle stage it first occurred in, and which type of sensor response corresponds to the initial abnormal signal. This allows for the generation of full lifecycle fault monitoring results for the equipment. Compared to methods relying solely on current alarms or single threshold judgments, this invention transforms scattered operational data, maintenance records, and status evidence into clear fault formation paths, resulting in more accurate fault location, a more coherent tracing process, and clearer maintenance and inspection targets.

[0038] Table 1 Comparison of Fault Source Tracing Performance Throughout the Equipment Lifecycle

[0039] As shown in Table 1, the method of this invention achieves an accuracy of 90.8% in identifying current fault candidates, which is an improvement over the 78.6% of the threshold alarm method, the 84.3% of the single time-series prediction method, and the 88.7% of the multi-source fusion diagnostic method. This is because this invention does not rely solely on a single operating parameter or a single sensor threshold to determine the current fault. Instead, it simultaneously utilizes equipment lifecycle monitoring data, equipment semantic constraint sets, and multi-agent state evidence sets to determine fault candidates. This allows equipment state evidence, component anomaly evidence, sensor response evidence, operating condition disturbance evidence, maintenance and recovery evidence, and lifecycle stage evidence to participate in the judgment, thus reducing misjudgments caused by a single alarm trigger.

[0040] In terms of the accuracy of initial component location, this invention achieves 87.9%, significantly higher than the threshold alarm method's 61.4%, the single time-series prediction method's 68.9%, and the multi-source fusion diagnostic method's 75.6%. This result demonstrates that this invention can not only identify the current faulty object but also unify the relationship between component objects, sensor responses, maintenance events, and historical time windows through the equipment's full lifecycle fault backtracking tensor. Furthermore, by employing the fault inversion tensor train decomposition algorithm, it performs reverse core tensor propagation from the fault category core tensor to the equipment object core tensor, thereby reversibly locating the earliest corresponding component source of the fault.

[0041] In terms of the accuracy of locating faults at the beginning of the equipment lifecycle, this invention achieves 85.6%, which is an improvement over the 54.8% of the threshold alarm method, and also higher than the 63.5% of the single time-series prediction method and the 70.8% of the multi-source fusion diagnostic method. This indicates that by writing the equipment lifecycle stage identifier sequence into the equipment's full lifecycle fault backtracking tensor, this invention can determine in the fault tracing process which stage the anomaly first occurred in: the commissioning period, the stable operation period, the performance degradation period, the maintenance and recovery period, or the end of the lifecycle. Compared to methods that only focus on the current operating status, this invention can trace the fault within the entire equipment lifecycle, thus more accurately distinguishing anomalies caused by operating condition disturbances, natural degradation, and insufficient maintenance and recovery.

[0042] In terms of initial abnormal signal identification accuracy, this invention achieves 86.7%, higher than the 57.2% of threshold alarm methods, 66.1% of single time-series prediction methods, and 73.4% of multi-source fusion diagnostic methods. The main reason for this is that this invention uses continuous path filtering based on the fault inversion contribution sequence, enabling it to find the earliest location of abnormal contribution connections in the fault formation path and identify the corresponding sensor response or operating state change as the initial abnormal signal. This allows for the identification of weak abnormal signals that have not yet formed a clear alarm, rather than only making a judgment after an alarm occurs.

[0043] In terms of fault path integrity rate, this invention achieves 83.2%, higher than the threshold alarm method's 49.5%, the single time-series prediction method's 58.4%, and the multi-source fusion diagnostic method's 66.9%. This significant improvement is mainly due to the invention's inclusion of the current fault candidate result in the fault category core tensor as an end constraint, and the use of the equipment semantic constraint set as an inverse verification constraint. This allows the inverse core tensor propagation to form a continuous contribution chain along the fault category, maintenance event, historical time window, equipment lifecycle stage, operating condition, sensor response, component object, and equipment object. The resulting fault tracing result not only includes the starting component and initial abnormal signal but also outputs the fault formation path, facilitating subsequent maintenance and inspection.

[0044] Regarding the average time taken for a single fault tracing attempt, this invention achieves 26.8 seconds, which is higher than the 18.7 seconds of the threshold alarm method but lower than the 31.5 seconds of the multi-source fusion diagnostic method, and close to the 24.2 seconds of the single time-series prediction method. This result indicates that although this invention introduces the equipment lifecycle fault backtracking tensor and the fault inversion tensor train decomposition algorithm, the tensor train decomposition can decompose the high-dimensional fault backtracking tensor into a core tensor sequence, reducing the computational pressure caused by directly calculating the full high-dimensional data. Therefore, this invention improves the ability to locate the starting component, locate the lifecycle stage, identify the initial abnormal signal, and reconstruct the fault formation path while maintaining an acceptable fault tracing time.

[0045] In summary, the advantages of this invention are not only reflected in the improvement of the current fault candidate identification accuracy from 88.7% to 90.8%, but more importantly, in the improvement of the starting component positioning accuracy to 87.9%, the starting equipment life cycle stage positioning accuracy to 85.6%, the initial abnormal signal identification accuracy to 86.7%, and the fault formation path integrity rate to 83.2%. These indicators collectively demonstrate that this invention can trace the fault source backward from the current fault candidate results, solving the problem that it is difficult to locate the fault starting point by relying solely on the current alarm or forward prediction. It has the beneficial effects of accurate fault location, clear source tracing path, and clear maintenance and inspection objects.

[0046] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for monitoring equipment lifecycle faults based on large models and multi-agent systems, characterized in that, Includes the following steps: Obtain raw data of the target device throughout its entire lifecycle and perform standardized processing to obtain monitoring data of the entire lifecycle of the device. Based on the equipment's full lifecycle monitoring data, a sequence of equipment lifecycle stage identifiers is generated. Input the equipment operation and maintenance text data from the equipment lifecycle monitoring data into the instruction fine-tuning language model for equipment fault semantic parsing to obtain the set of equipment semantic constraints. By inputting the equipment's full lifecycle monitoring data, the equipment's lifecycle stage identifier sequence, and the equipment's semantic constraint set into the multi-agent analysis system, a multi-agent state evidence set is obtained. Based on the multi-agent state evidence set and the device semantic constraint set, the current fault candidate result is obtained by determining the fault candidate. Based on the multi-agent state evidence set and the current fault candidate results, construct a fault backtracking tensor for the entire life cycle of the device. The fault inversion tensor train decomposition algorithm is used to decompose the fault backtracking tensor of the entire equipment life cycle to obtain the core tensor sequence; Write the current fault candidate results into the fault category core tensor as the terminal constraint, use the device semantic constraint set as the inverse verification constraint, perform inverse core tensor propagation, and obtain the fault inverse contribution sequence. By filtering continuous paths based on the fault inversion contribution sequence, fault source tracing results are obtained and equipment lifecycle fault monitoring results are generated.

2. The method for monitoring equipment lifecycle faults based on a large model and multiple agents according to claim 1, characterized in that, The raw data for the entire lifecycle of the equipment includes equipment history data, equipment operation data, and equipment maintenance text data.

3. The method for monitoring equipment lifecycle faults based on a large model and multiple agents according to claim 1, characterized in that, The generation of the device lifecycle stage identifier sequence includes: Extract equipment history data and equipment operation data from the equipment lifecycle monitoring data, and divide the equipment lifecycle monitoring data into multiple historical time windows according to the collection time of the equipment operation data; Based on the equipment history data corresponding to each historical time window, determine the history status characteristics of each historical time window; Based on the equipment operation data corresponding to each historical time window, determine the operation status characteristics of each historical time window; Based on the historical status characteristics and operational status characteristics of each historical time window, the stage of each historical time window is determined to obtain the equipment life cycle stage corresponding to each historical time window. Arrange the lifecycle stages of each device in chronological order according to the historical time window, and generate a sequence of device lifecycle stage identifiers.

4. The method for monitoring equipment lifecycle faults based on a large model and multiple agents according to claim 1, characterized in that, The generation of the device semantic constraint set includes: Equipment maintenance text data is extracted from the equipment lifecycle monitoring data, and sorted according to the text generation time, associated equipment object, and associated component object to obtain the equipment maintenance text sequence. The equipment maintenance text sequence is cleaned and its fields are standardized to obtain a standardized equipment maintenance text sequence. The standardized equipment operation and maintenance text sequence is input into the instruction fine-tuning language model for equipment fault semantic parsing. The fault object, fault phenomenon, fault category, fault occurrence time window and abnormal duration are identified from the standardized inspection text and standardized historical fault text to obtain equipment fault semantic constraints. Standardized equipment operation and maintenance text sequences are input into a command fine-tuning language model for equipment fault semantic parsing. The maintenance object, maintenance action, maintenance occurrence time, post-maintenance status, and fault recurrence status are identified from the standardized maintenance text to obtain equipment maintenance semantic constraints. Based on the associated equipment objects, associated component objects, and text generation time, the semantic constraints of equipment failure and equipment maintenance are organized to obtain a set of equipment semantic constraints.

5. The method for monitoring equipment lifecycle faults based on a large model and multiple agents according to claim 1, characterized in that, The generation of the multi-agent state evidence set includes: Equipment operation data is extracted from the equipment lifecycle monitoring data, and the equipment operation data is organized according to equipment object, component object and historical time window to obtain equipment operation evidence input data; The operational status data from the equipment operation evidence input data is input into the equipment status intelligent agent to generate equipment status evidence; Based on the set of device semantic constraints, determine the associated component objects, and input the device operation evidence input data corresponding to the associated component objects into the component anomaly agent to generate component anomaly evidence; The sensor-collected data from the equipment operation evidence input data is input into the sensor response agent to generate sensor response evidence; Input the operating condition status data from the equipment operation evidence input data into the operating condition disturbance intelligent agent to generate operating condition disturbance evidence. The maintenance and recovery agent is generated by inputting equipment operation evidence data, equipment lifecycle stage identifier sequence, and equipment semantic constraint set into the maintenance and recovery agent. Input the device lifecycle stage identifier sequence into the lifecycle assessment agent to generate lifecycle stage evidence; Based on the equipment object, associated component object, and historical time window, the evidence of equipment status, component anomaly, sensor response, operating condition disturbance, maintenance and recovery, and life cycle stage is aligned to obtain a multi-agent status evidence set.

6. The method for monitoring equipment lifecycle faults based on a large model and multiple agents according to claim 1, characterized in that, The generation of the current fault candidate result includes: Extract the state evidence corresponding to the current time window from the multi-agent state evidence set, and align the evidence according to the device object, component object and the current time window to obtain the current time window state evidence group; Extract the semantic constraints corresponding to the current time window from the set of device semantic constraints, and perform semantic alignment according to the device object, component object and the current time window to obtain the semantic constraint group of the current time window; Determine the abnormal pointing results of each component object within the current time window based on the evidence group of the current time window status; Based on the semantic constraint group of the current time window, perform semantic matching on the abnormal pointing results to obtain the semantic matching abnormal pointing results; The current faulty object is determined based on the semantic matching anomaly pointing to the result, and the current fault category and the current fault occurrence time window are determined based on the semantic constraint group of the current time window; The current fault candidate results are composed of the current fault object, the current fault category, and the current fault occurrence time window.

7. The method for monitoring equipment lifecycle faults based on a large model and multiple agents according to claim 1, characterized in that, The construction of the device lifecycle fault backtracking tensor includes: Extract equipment objects, component objects, sensor responses, operating conditions, and historical time windows from the equipment lifecycle monitoring data, and determine the equipment lifecycle stage corresponding to each historical time window based on the equipment lifecycle stage identifier sequence; Extract fault categories and maintenance events from the set of device semantic constraints, and extract the current fault object, current fault category, and current fault occurrence time window from the current fault candidate results; A set of fault backtracking dimensions is constructed using equipment objects, component objects, sensor responses, operating conditions, equipment lifecycle stages, historical time windows, maintenance events, and fault categories as fault backtracking dimensions. Based on the fault backtracking dimension set, the evidence dimension of the multi-agent state evidence set is processed to obtain state evidence mapping data; the semantic dimension of the device semantic constraint set is processed to obtain semantic constraint mapping data; and the fault location of the current fault candidate result is processed to obtain the current fault mapping data. Write the state evidence mapping data, semantic constraint mapping data and current fault mapping data into the tensor positions corresponding to the fault backtracking dimension set to form the initial fault backtracking tensor. Based on the equipment object, component object, historical time window, and equipment lifecycle stage, the state evidence mapping data, semantic constraint mapping data, and current fault mapping data in the initial fault backtracking tensor are organized according to their positions to obtain the equipment full lifecycle fault backtracking tensor.

8. The method for monitoring equipment lifecycle faults based on a large model and multiple agents according to claim 1, characterized in that, The generation of the core tensor sequence includes: The fault backtracking tensor of the entire equipment life cycle is input into the fault inversion tensor train decomposition algorithm. The fault inversion tensor train decomposition algorithm includes a dimension arrangement layer, a dimension-by-dimensional expansion layer, a connection rank determination layer, a core tensor generation layer, and a core tensor concatenation layer. In the dimensional arrangement layer, the fault backtracking tensor of the entire equipment life cycle is arranged dimensionally according to the order of equipment object, component object, sensor response, operating condition, equipment life cycle stage, historical time window, maintenance event and fault category, to obtain the inverse dimensional arrangement tensor. In the dimension-by-dimensional expansion layer, the dimension-by-dimensional matrix expansion process is performed on the tensors arranged according to the dimensional order of the inverse dimension to obtain a set of dimension-by-dimensional expansion matrices. In the connection rank determination layer, the tensor train connection rank between adjacent dimensions is determined based on the non-zero position correspondence between adjacent expansion matrices in the set of dimension-by-dimensional expansion matrices, the evidence distribution density, and the fault category correlation strength, thus obtaining the connection order column; In the core tensor generation layer, based on the set of dimension-wise expansion matrices and the connection order column, the equipment object expansion matrix, component object expansion matrix, sensor response expansion matrix, operating condition expansion matrix, equipment life cycle stage expansion matrix, historical time window expansion matrix, maintenance event expansion matrix, and fault category expansion matrix are sequentially subjected to low-rank decomposition processing to obtain the equipment object core tensor, component object core tensor, sensor response core tensor, operating condition core tensor, equipment life cycle stage core tensor, historical time window core tensor, maintenance event core tensor, and fault category core tensor. In the core tensor concatenation layer, the core tensors of the equipment object, component object, sensor response, operating condition, equipment life cycle stage, historical time window, maintenance event, and fault category are sequentially connected according to the connection order to obtain the core tensor sequence.

9. The method for monitoring equipment lifecycle faults based on a large model and multiple agents according to claim 1, characterized in that, The generation of the fault inversion contribution sequence includes: Write the current fault candidate results into the fault category core tensor in the core tensor sequence to obtain the terminal constraint core tensor. The set of semantic constraints of the equipment is organized into inverse verification constraints. The core tensor of the end constraint is used as the starting point for the inverse core tensor propagation. According to the reverse connection order from the core tensor of the fault category to the core tensor of the equipment object in the core tensor sequence, the core tensor of the maintenance event, the core tensor of the historical time window, the core tensor of the equipment life cycle stage, the core tensor of the operating condition, the core tensor of the sensor response, and the core tensor of the component object are passed in sequence to obtain the initial inverse contribution connection set. The initial inversion contribution connection set is verified according to the inversion verification constraints. The initial inversion contribution connections that meet the inversion verification constraints are retained to obtain the verified inversion contribution connection set. The fault inversion contribution sequence is obtained by arranging the check inversion contribution set according to the reverse connection order of the core tensor sequence.

10. The method for monitoring equipment lifecycle faults based on a large model and multiple agents according to claim 1, characterized in that, The generation of the equipment's full lifecycle fault monitoring results includes: According to the contribution order in the fault inversion contribution sequence, the contribution connection relationship between adjacent contributing nodes is extracted. Based on the continuity of the contribution connection relationship, the continuous path of the fault inversion contribution sequence is filtered to obtain the fault formation path. Identify the initiating component, the initial equipment lifecycle stage, and the initial abnormal signal based on the fault formation path; The fault tracing result consists of the starting component, the initial equipment life cycle stage, the initial abnormal signal, and the fault formation path; Based on the fault tracing results, generate equipment lifecycle fault monitoring results.

Images