Equipment fault diagnosis method, device and equipment based on graph enhanced industrial large model

Abstract

The application provides an equipment fault diagnosis method and device based on graph-enhanced industrial large model, and equipment, the method comprising: receiving multi-source heterogeneous fault data collected by a sensor; obtaining unified analysis data according to each heterogeneous fault data; constructing an industrial fault diagnosis graph according to a node set, an edge set, a relationship adjacency representation and a node attribute matrix determined by the unified analysis data, a preset node type set, a preset edge type set, a node set, an edge set, historical fault data; determining a graph evidence set according to a key subgraph, a key fault path set and path confidence statistical data determined by the industrial fault diagnosis graph; constructing joint input data according to a preset fault sample, the graph evidence set, preset historical case data and a constraint condition; determining corresponding complete input data according to the joint input data; inputting the complete input data into an industrial large model for inference processing to obtain fault diagnosis data, thereby improving the accuracy of equipment fault diagnosis.

Description

Technical Field

[0001] This application relates to the field of equipment fault diagnosis technology, and in particular to a method, apparatus and equipment for equipment fault diagnosis based on graph-enhanced industrial large model. Background Technology

[0002] Equipment fault diagnosis is a technology that monitors equipment operating status, identifies abnormal symptoms, locates the fault location and cause, and predicts its development trend. It integrates sensor data, signal processing, and intelligent algorithms to achieve early warning and health management for complex equipment.

[0003] In the existing technology, the traditional equipment fault diagnosis method based on graph-enhanced industrial large model obtains multiple fault data of the equipment and performs fault analysis on single type of data in each fault data to obtain the equipment fault diagnosis result.

[0004] However, the existing fault diagnosis method, which involves acquiring multiple fault data points from the equipment and then analyzing each single type of data point separately, is too simplistic and reduces the accuracy of equipment fault diagnosis. Summary of the Invention

[0005] This application provides a method, apparatus, and equipment for equipment fault diagnosis based on a graph-enhanced industrial large model. This addresses the problem that the fault diagnosis method, which involves acquiring multiple fault data points from the equipment and then performing fault analysis on a single type of data in each fault data point, is too simplistic and reduces the accuracy of equipment fault diagnosis.

[0006] Firstly, this application provides a fault diagnosis method for equipment based on a graph-enhanced industrial large model, applied to electronic equipment, including:

[0007] Receive multi-source heterogeneous fault data generated by sensor acquisition equipment during operation;

[0008] Preprocess the heterogeneous fault data to obtain unified analysis data;

[0009] Based on the unified analysis data, the preset node type set, the preset edge type set, and the historical fault data, determine the node set, edge set, relational adjacency representation, and node attribute matrix;

[0010] The industrial fault diagnosis map of the equipment is constructed based on the preset node type set, the preset edge type set, the node set, the edge set, the relational adjacency representation, and the node attribute matrix.

[0011] The corresponding key sub-graphs are determined based on the industrial fault diagnosis diagram;

[0012] The graph evidence set is determined based on the key subgraph, the pre-screened set of key fault paths, and the statistical data corresponding to the pre-generated path confidence.

[0013] The joint input data is constructed based on the preset fault samples, the set of graphical evidence, the preset historical case data, and the constraints.

[0014] The combined input data is subjected to a structured transformation to obtain enhanced prompt data;

[0015] The corresponding complete input data is determined based on the combined input data and the enhanced prompt data;

[0016] The complete input data is input into the graph-enhanced industrial model for inference processing to obtain fault diagnosis data.

[0017] In one possible design, the heterogeneous fault data includes continuous monitoring time-series data, event and alarm data, text log data, equipment structure and process data, and operating condition and environmental context data. Accordingly, the preprocessing of the heterogeneous fault data to obtain unified analysis data includes: preprocessing the continuous monitoring time-series data to obtain standardized time-series segments characterizing changes in equipment status; preprocessing the event and alarm data to obtain standard alarm data for corresponding events; preprocessing the text log data to obtain standard log data for corresponding texts; preprocessing the equipment structure and process data to obtain structural relationship data between any two components; and performing cross-source alignment and correlation mapping based on the standardized time-series segments, standard alarm data for each event, standard log data for each text, structural relationship data, and operating condition and environmental context data to obtain unified analysis data.

[0018] In one possible design, determining the node set, edge set, relational adjacency representation, and node attribute matrix based on the unified analysis data, preset node type set, preset edge type set, and historical fault data includes: determining the corresponding node set based on the preset node type set; determining the correlation degree between any two random parameters determined by the unified analysis data and historical fault data, and the standard deviation corresponding to each random parameter; determining the corresponding weight based on the correlation degree, and determining the corresponding edge type based on the preset edge type set; determining the corresponding candidate edge based on the weight, any two nodes, and the edge type, and determining the edge set based on each candidate edge; determining the state subgraphs formed by the equipment under different time windows, and determining the time sequence diagram sequence based on each state subgraph to represent the relational adjacency representation; and obtaining the initial attributes of each node based on the node set to obtain the node attribute matrix.

[0019] In one possible design, determining the corresponding key subgraph based on the industrial fault diagnosis map includes: performing heterogeneous node feature mapping based on the industrial fault diagnosis map to obtain initial embedding values; updating based on the initial embedding values, the neighborhood set and mapping matrix corresponding to each node, the self-loop transformation matrix, and the relation weight coefficients to obtain enhanced node representations; determining the time-varying enhanced representation of the corresponding node at any given time based on the enhanced node representations; determining the corresponding time-varying enhanced representation based on the time-varying enhanced representations and preset fusion coefficients; and after the time-varying enhanced representations, selecting key subgraphs related to preset fault samples from the industrial fault diagnosis map.

[0020] In one possible design, the path confidence generation process includes: determining the fault concern score of the corresponding node based on the preset monitoring signal anomaly score, the preset event or alarm information score, the preset knowledge rule score, and the preset historical case support score; and determining the path confidence based on the fault concern score, the candidate propagation paths of pre-constructed fault mode nodes, and the edge association strength between adjacent nodes.

[0021] In one possible design, the fault diagnosis data includes a first fault diagnosis result and a second fault diagnosis result; correspondingly, the step of inputting the complete input data into the graph-augmented industrial big data model for inference processing to obtain fault diagnosis data includes: inputting the complete input data into the graph-augmented industrial big data model for inference processing to obtain a first fault diagnosis result; during the inference process of the graph-augmented industrial big data model, generating a second fault diagnosis result based on the fault category, fault location, root cause analysis results, fault propagation or evidence path, and maintenance or inspection suggestions.

[0022] In one possible design, after inputting the complete input data into the graph-augmented industrial big data model for inference processing to obtain fault diagnosis data, the method further includes: calculating a corresponding consistency evaluation value based on the fault diagnosis data and the graph evidence set; calculating the rule satisfaction degree corresponding to the fault diagnosis data based on the rule set; determining whether the consistency evaluation value is greater than a preset threshold and whether the rule satisfaction degree satisfies the corresponding rule; if the consistency evaluation value is greater than or equal to the preset threshold and the rule satisfaction degree satisfies the corresponding rule, then the fault diagnosis data is determined as a reliable diagnosis result; if the consistency evaluation value is less than the preset threshold and / or the rule satisfaction degree does not satisfy the corresponding rule, then the complete input data is corrected to obtain corrected complete input data; and the corrected complete input data is input into the graph-augmented industrial big data model for inference processing to obtain the final fault diagnosis data.

[0023] In one possible design, after the complete input data is input into the graph-enhanced industrial big data model for inference processing to obtain fault diagnosis data, the method further includes: determining, based on the multi-source heterogeneous fault data, original sensor data fragments and event records or text records, key nodes, key relationships, critical paths, and time windows or timestamp sets related to the fault diagnosis data, to calculate the corresponding evidence chain; determining a credibility evaluation value based on the confidence level of the fault diagnosis data, the consistency evaluation value, the rule satisfaction level, historical case support, or multi-source evidence support; and determining the structured diagnostic result of the equipment based on the final fault category, fault location result, root cause description, fault propagation or evidence path, maintenance or testing recommendations, the evidence chain, and the credibility evaluation value.

[0024] Secondly, this application provides an equipment fault diagnosis device based on a graph-enhanced industrial large model, applied to electronic equipment, comprising:

[0025] The receiving module is used to receive multi-source heterogeneous fault data generated by the sensor acquisition equipment during operation;

[0026] The processing module is used to preprocess the heterogeneous fault data to obtain unified analysis data;

[0027] The first determining module is used to determine the node set, edge set, relational adjacency representation, and node attribute matrix based on the unified analysis data, the preset node type set, the preset edge type set, and historical fault data.

[0028] The construction module is used to construct an industrial fault diagnosis map of the equipment based on the preset node type set, the preset edge type set, the node set, the edge set, the relational adjacency representation, and the node attribute matrix.

[0029] The second determining module is used to determine the corresponding key sub-graph based on the industrial fault diagnosis graph;

[0030] The third determining module is used to determine the graph evidence set based on the key subgraph, the pre-screened set of key fault paths, and the statistical data corresponding to the pre-generated path confidence.

[0031] The component module is used to construct joint input data based on the preset fault samples, the set of graphical evidence, the preset historical case data, and constraints.

[0032] A conversion module is used to perform structured conversion on the combined input data to obtain enhanced prompt data;

[0033] The fourth determining module is used to determine the corresponding complete input data based on the combined input data and the enhanced prompt data;

[0034] The first reasoning module is used to input the complete input data into the graph-enhanced industrial model for reasoning processing to obtain fault diagnosis data.

[0035] Thirdly, this application provides an electronic device, including: at least one processor and a memory;

[0036] The memory stores computer-executed instructions;

[0037] The at least one processor executes computer execution instructions stored in the memory, causing the at least one processor to perform the equipment fault diagnosis method based on graph-enhanced industrial large model as described in the first aspect and various possible designs of the first aspect.

[0038] The equipment fault diagnosis method, apparatus, and equipment based on graph-enhanced industrial large-scale model provided in this application acquire multi-source heterogeneous fault data generated by the equipment during operation by receiving sensors; preprocess the heterogeneous fault data to obtain unified analysis data; determine the node set, edge set, relational adjacency representation, and node attribute matrix based on the unified analysis data, a preset set of node types, a preset set of edge types, and historical fault data; construct an industrial fault diagnosis graph for the equipment based on the preset set of node types, preset set of edge types, node set, edge set, relational adjacency representation, and node attribute matrix; and determine the corresponding key subgraph based on the industrial fault diagnosis graph. The graph evidence set is determined based on the key subgraph, the pre-screened set of key fault paths, and the statistical data corresponding to the pre-generated path confidence. Joint input data is constructed based on preset fault samples, the graph evidence set, preset historical case data, and constraints. The joint input data undergoes structured transformation to obtain enhanced prompt data. The corresponding complete input data is determined based on the joint input data and the enhanced prompt data. The complete input data is then input into the graph-enhanced industrial big data model for inference processing to obtain fault diagnosis data. Through the fusion processing of multi-source heterogeneous fault data, fault diagnosis data is generated, thereby improving the accuracy of equipment fault diagnosis. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 A schematic diagram illustrating an application scenario of the equipment fault diagnosis method based on a graph-enhanced industrial large model provided in this application embodiment;

[0041] Figure 2A flowchart illustrating the equipment fault diagnosis method based on a graph-enhanced industrial large model provided in this application embodiment. Figure 1 ;

[0042] Figure 3 A flowchart illustrating the equipment fault diagnosis method based on a graph-enhanced industrial large model provided in this application embodiment. Figure 2 ;

[0043] Figure 4 A schematic diagram of the equipment fault diagnosis device based on a graph-enhanced industrial large model provided in this application embodiment;

[0044] Figure 5 This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0046] Equipment fault diagnosis is a technology that monitors equipment operating status, identifies abnormal symptoms, locates the fault location and cause, and predicts its development trend. It integrates sensor data, signal processing, and intelligent algorithms to achieve early warning and health management for complex equipment. In existing technologies, traditional equipment fault diagnosis methods based on graph-augmented industrial models acquire multiple fault data points from the equipment and perform fault analysis on a single type of data within each fault data point to obtain the fault diagnosis result. However, this existing fault diagnosis approach, which involves acquiring multiple fault data points and performing fault analysis on a single type of data within each fault data point, is overly simplistic and reduces the accuracy of equipment fault diagnosis.

[0047] To address the aforementioned technical problems, this application proposes the following technical concept: Considering heterogeneous fault data, the inventors perform data processing based on the heterogeneous fault data, a preset set of node types, a preset set of edge types, and historical fault data until an industrial fault diagnosis map of the equipment is obtained. The corresponding key subgraphs are determined based on the industrial fault diagnosis map. A graph evidence set is determined using the key subgraphs, the set of key fault paths, and statistical data corresponding to the path confidence. Joint input data is constructed based on preset fault samples, the graph evidence set, preset historical case data, and constraints. The corresponding complete input data is determined using the joint input data. This complete input data is then input into a graph-enhanced industrial large-scale model for inference processing to obtain fault diagnosis data, thereby improving the accuracy of equipment fault diagnosis.

[0048] Figure 1 This is a schematic diagram illustrating an application scenario of the equipment fault diagnosis method based on a graph-enhanced industrial large model provided in this application embodiment.

[0049] like Figure 1 As shown, the scene includes: sensor 101 and electronic device 102.

[0050] Among them, sensor 101 can be a single sensor or a cluster of multiple sensors.

[0051] Electronic device 102 can be a standalone device or a cluster of multiple devices.

[0052] Electronic device 102 receives multi-source heterogeneous fault data generated by sensor 101 during equipment operation; preprocesses the heterogeneous fault data to obtain unified analysis data; determines node sets, edge sets, relational adjacency representations, and node attribute matrices based on the unified analysis data, preset node type sets, preset edge type sets, and historical fault data; constructs an industrial fault diagnosis graph for the equipment based on the preset node type sets, preset edge type sets, node sets, edge sets, relational adjacency representations, and node attribute matrices; determines corresponding key subgraphs based on the industrial fault diagnosis graph; determines graph evidence sets based on the key subgraphs, pre-screened key fault path sets, and statistical data corresponding to pre-generated path confidence; constructs joint input data based on preset fault samples, graph evidence sets, preset historical case data, and constraints; performs structured transformation on the joint input data to obtain enhanced prompt data; determines corresponding complete input data based on the joint input data and enhanced prompt data; and inputs the complete input data into a graph-enhanced industrial large-scale model for inference processing to obtain fault diagnosis data. Detailed embodiments are described below.

[0053] Figure 2 A flowchart illustrating the equipment fault diagnosis method based on a graph-enhanced industrial large model provided in this application embodiment. Figure 1 The execution entity in this embodiment can be Figure 1 The electronic devices shown in the illustrated embodiments are not specifically limited in this embodiment. Figure 2 As shown, the method includes:

[0054] S201: Receives multi-source heterogeneous fault data generated by the sensor acquisition equipment during operation.

[0055] In this embodiment, the equipment can be high-end industrial equipment or other types of equipment.

[0056] In this embodiment, the multi-source heterogeneous fault data includes continuous monitoring time-series data, event and alarm data, text record data, equipment structure and process data, and operating condition and environmental context data.

[0057] The continuous monitoring time-series data includes continuous time-series data such as vibration, temperature, pressure, current, voltage, speed, displacement, flow rate, and acoustic signals; the event and alarm data includes alarm information, fault codes, start-up and shutdown records, process switching events, control command events, and maintenance operation events; the text record data includes maintenance records, fault work orders, repair reports, expert experience, and historical case texts; the equipment structure and process data includes equipment composition hierarchy, component connection relationships, functional dependencies, parameter upper and lower limits, process ranges, and operating constraints; and the operating condition and environmental context data includes contextual information that affects fault diagnosis results, such as load status, working mode, processing conditions, and external ambient temperature and humidity.

[0058] The above data can be obtained through various channels such as industrial field sensors, control system interfaces, edge acquisition terminals, equipment maintenance management systems, manufacturing execution systems, and historical databases.

[0059] For example, the dataset corresponding to multi-source heterogeneous fault data is:

[0060]

[0061] in, This is a dataset corresponding to multi-source heterogeneous fault data; To continuously monitor time-series data; For event and alarm data; Record data for text; For equipment structure and process data; This is for operating condition and environmental context data.

[0062] S202: Preprocess the heterogeneous fault data to obtain unified analysis data.

[0063] In this embodiment, the heterogeneous fault data includes continuous monitoring time-series data, event and alarm data, text record data, equipment structure and process data, and operating condition and environmental context data; correspondingly, step S202 specifically includes steps a~e:

[0064] Step a: Preprocess the continuous monitoring time series data to obtain standardized time series segments used to characterize changes in equipment status.

[0065] In this embodiment, since different fault data have significant differences in sampling frequency, time granularity, representation form and semantic level, they need to be preprocessed in a hierarchical manner.

[0066] Specifically, the continuous monitoring time series data is processed by denoising, missing value compensation, outlier correction, normalization and time window segmentation to obtain standardized time series segments for characterizing equipment status changes.

[0067] For example, the continuously monitored time series data is The standardized time segment is of length [length missing]. sample fragments .

[0068] Furthermore, based on the above, features such as mean, variance, kurtosis, frequency domain energy, envelope spectrum peak value, and time-frequency statistics can be extracted to form a monitoring feature vector. .

[0069] Step b: Preprocess the event and alarm data to obtain the standard alarm data for the corresponding events.

[0070] Specifically, the event and alarm data are sorted by time, deduplicated, standardized in encoding, and standardized in level to obtain standard alarm data for the corresponding events.

[0071] For example, the standard alarm data for an event is a unified set of events mapped from alarm codes, status identifiers, and event descriptions in different systems, where each alarm event can be represented as:

[0072]

[0073] in, For alarm events; The time when the event occurred; Event category; Event level; This refers to the object associated with the event.

[0074] Furthermore, the above representation allows for the establishment of a time-aligned relationship between discrete events and continuous monitoring signals.

[0075] Step c: Preprocess the text record data to obtain the standard record data of the corresponding text.

[0076] Specifically, the text data such as maintenance records, repair reports, and fault work orders in the text record data are processed by terminology normalization, entity recognition, keyword extraction, and semantic segmentation to obtain the standard record data of the corresponding text.

[0077] The standard record data consists of associative fault semantic units converted from unstructured text.

[0078] Among them, any text record of the standard record data is:

[0079]

[0080] in, It is a collection of component entities, fault entities, and parameter entities; It is a set of semantic relations; This is a collection of fault description statements.

[0081] Furthermore, through the above steps, historical experience and knowledge can be transformed into candidate nodes and relationships required for subsequent graph construction.

[0082] Step d: Preprocess the equipment structure and process data to obtain the structural relationship data between any two components.

[0083] Specifically, the equipment structure and process data are processed by unifying object numbering, sorting out hierarchical relationships, and standardizing rule expression to obtain structural relationship data between any two components.

[0084] In this way, equipment, subsystems, components, parameters, and process constraints can be mapped into unified structural knowledge units.

[0085] For example, any two components are components. With components .

[0086]

[0087] in, For structural relational data; It is a set of structural relationships.

[0088] The aforementioned structural relationships include compositional relationships, connectivity relationships, dependency relationships, and constraint relationships. This process lays the foundation for the subsequent construction of structural edges and regular edges in the heterogeneous graph.

[0089] Step e: Perform cross-source alignment and correlation mapping based on standardized time sequence segments, standard alarm data for each event, standard record data for each text, structural relationship data, and operating condition and environmental context data to obtain unified analysis data.

[0090] Specifically, cross-source alignment and association mapping are performed based on standardized time-series segments, standard alarm data for each event, standard record data for each text, structural relationship data, and operating condition and environmental context data. That is, based on equipment identifiers, component identifiers, timestamps, and fault event indexes, multi-source data are merged into a unified analysis unit to obtain unified analysis data.

[0091] For example, the unified analysis data is as follows:

[0092]

[0093] In the formula, To standardize data analysis; This refers to the monitoring signal segment within the corresponding time window in the standardized time sequence segment; This is the set of related events in the standard alarm data for each event; For the standard record data of each text, the related text knowledge, This refers to the structural rule information within each structural relationship data. This refers to the operating condition context in the operating condition and environmental context data.

[0094] The unified representation method described above can aggregate the originally scattered multi-source heterogeneous fault data into a comprehensive input sample oriented towards the same fault process.

[0095] Furthermore, quality assessments and confidence level labels can be set for different types of data. For example, quality scores can be assigned to data collection completeness, time synchronization accuracy, text standardization, and rule reliability, forming a data quality vector. .

[0096] The aforementioned data quality vectors are used to weight and adjust the contributions of data from different sources during subsequent graph augmentation and graph augmentation industrial large model inference stages. This can reduce the interference of noisy data, incomplete data, or low-reliability text records on diagnostic results.

[0097] S203: Determine the node set, edge set, relational adjacency representation, and node attribute matrix based on unified analysis data, preset node type set, preset edge type set, and historical fault data.

[0098] Specifically, step S203 includes steps a~f:

[0099] Step a: Determine the corresponding node set based on the preset node type set.

[0100] For example, a preset set of node types This includes device node types, component node types, parameter node types, signal characteristic node types, fault phenomenon node types, fault mode node types, operating condition node types, and knowledge rule node types.

[0101] For example, the device node is Component nodes are Parameter nodes are Signal feature nodes are Fault Phenomenon Nodes Fault mode nodes are Operating condition nodes are and knowledge rule nodes are .

[0102] For example, the node set is:

[0103]

[0104] Among them, equipment nodes are used to represent the equipment or its subsystems being diagnosed; component nodes are used to represent key components such as spindles, bearings, gears, servo drive units, hydraulic circuits, and electrical control modules; parameter nodes are used to represent operational monitoring quantities such as temperature, current, pressure, vibration amplitude, and frequency band energy; signal feature nodes are used to represent diagnostic features extracted from the time domain, frequency domain, and time-frequency domain; fault phenomenon nodes are used to represent fault symptoms such as abnormal noise, abnormal temperature rise, waveform abrupt changes, and power fluctuations; fault mode nodes are used to represent specific fault types such as wear, cracks, loosening, imbalance, blockage, and insulation degradation; operating condition nodes are used to represent load levels, speed ranges, processing stages, and operating modes; and knowledge rule nodes are used to represent knowledge information such as historical cases, expert rules, maintenance strategies, and safety constraints.

[0105] Step b: Determine the correlation between any two random parameters determined based on unified analysis data and historical fault data, as well as the standard deviation of each random parameter.

[0106] In this embodiment, to enhance the ability of the graph structure to express industrial mechanisms, this solution adopts a combination of prior knowledge-driven and data relationship learning to construct heterogeneous association graphs. On the one hand, based on equipment structure diagrams, process flows, maintenance manuals, and expert experience, equipment composition relationships, connection dependencies, and rule constraint relationships are established. On the other hand, based on unified analysis data and historical fault data, the correlation, similarity, and co-occurrence between parameters, features, and fault phenomena are calculated, thereby supplementing data-driven relationship edges.

[0107] For example, any two random parameters determined based on unified analysis data and historical fault data are... and .

[0108] In this embodiment, the formula for calculating the correlation between two random parameters, determined based on unified analysis data and historical fault data, and the standard deviation of each random parameter, is as follows:

[0109]

[0110] In the formula, for and The corresponding degree of relevance; For random parameters and covariance; Random parameters and The standard deviation.

[0111] Furthermore, similarity association edges can be generated based on feature distance or semantic matching score to determine the similarity relationship between fault samples and historical cases.

[0112] Step c: Determine the corresponding weights based on the degree of relevance, and determine the corresponding edge types based on the preset set of edge types.

[0113] For example, depending on the degree of relevance Determine the corresponding weights as .

[0114] For example, a preset set of edge types The edge types include compositional relationship edges between devices and components, connection relationship edges between components, attribution relationship edges between components and parameters, mapping relationship edges between parameters and features, indication relationship edges between features and fault phenomena, inference relationship edges between fault phenomena and fault modes, support relationship edges between fault modes and knowledge nodes, conditional constraint relationship edges between operating conditions and parameters or fault modes, and evolutionary relationship edges between different time slice states.

[0115] Step d: Determine the corresponding candidate edges based on the weights, any two nodes, and edge types, and determine the edge set based on each candidate edge.

[0116] In this embodiment, the calculation formula for determining the corresponding candidate edge based on the weight, any two nodes, and edge type is as follows:

[0117]

[0118] In the formula, Candidate edges; and This corresponds to any two nodes; This refers to the relationship type between nodes, i.e., the edge type; As weight.

[0119] The aforementioned weights can be determined by the strength of prior rules, relevance, event co-occurrence frequency, or support from historical cases.

[0120] For example, the set of edges is E.

[0121] Step e: Determine the state subgraphs formed by the equipment under different time windows, and determine the timing diagram sequence based on each state subgraph to represent the relational adjacency representation.

[0122] For example, the state subgraph is as follows: The time series diagram sequence is .

[0123] For example, the relational adjacency is represented as A.

[0124] Furthermore, the above process connects corresponding nodes at adjacent time points through time evolution edges to describe the dynamic process of a fault from early signs to overt anomalies and then to the formation of a fault. This design is beneficial for capturing the fault evolution trend in subsequent diagnosis, rather than just relying on static judgment at a single moment.

[0125] Step f: Obtain the initial attributes of each node based on the node set to obtain the node attribute matrix.

[0126] In this embodiment, feature vectors matching the semantics of different types of nodes need to be configured. For parameter nodes, statistical features, trend features, and spectral features can be assigned; for component nodes, structural location, functional category, and historical failure frequency can be assigned; for knowledge nodes, text semantic embedding and rule type labeling can be assigned.

[0127] For example, any node The initial properties are Then the node attribute matrix can be represented as:

[0128]

[0129] Furthermore, by unifying the expression of node attributes, subsequent graph enhancement modules can perform cross-modal and cross-level joint modeling in heterogeneous graph spaces.

[0130] S204: Construct an industrial fault diagnosis diagram for equipment based on a preset set of node types, a preset set of edge types, a set of nodes, a set of edges, a relational adjacency representation, and a node attribute matrix.

[0131] In this embodiment, the calculation formula for constructing the industrial fault diagnosis map of the equipment based on the preset node type set, preset edge type set, node set, edge set, relational adjacency representation, and node attribute matrix is ​​as follows:

[0132]

[0133] in, This is an industrial fault diagnosis diagram; A set of nodes; Let it be the set of edges; A collection of node types; A set of edge types; This is a matrix of node attributes; This represents the relational adjacency representation.

[0134] Among them, nodes and edges have heterogeneous attributes, and different types of nodes and relationships jointly describe the multi-level fault semantic space of the equipment.

[0135] Furthermore, to avoid noise introduced by too many weakly correlated edges in the industrial fault diagnosis graph, this scheme can also set up a relation filtering and constraint mechanism. Specifically, only relation edges that satisfy the following criteria are retained: structural rationality, process constraint, statistical significance, or knowledge support threshold:

[0136]

[0137] in, To satisfy the requirements of structural rationality, process constraints, statistical significance, or knowledge support thresholds for relational edges; Indicates the threshold for retaining relationships.

[0138] By using the above method, candidate edges can be compressed into an industrial fault diagnosis graph that is more suitable for fault diagnosis tasks, reducing the interference of irrelevant edges on subsequent reasoning.

[0139] S205: Determine the corresponding critical sub-graphs based on the industrial fault diagnosis diagram.

[0140] Specifically, step S205 includes steps a~e:

[0141] Step a: Perform heterogeneous node feature mapping based on the industrial fault diagnosis map to obtain initial embedding values.

[0142] Specifically, heterogeneous node feature mapping is performed based on the industrial fault diagnosis map, projecting parameter nodes, component nodes, fault phenomenon nodes, fault mode nodes, and knowledge nodes onto a unified representation space to obtain initial embedding values.

[0143] For example, the initial embedding value is ,in Represents type node type This indicates that it is for this type of node. The feature mapping function is set.

[0144] Furthermore, through the above methods, a comparable and aggregateable representation basis can be established in a unified feature space for industrial fault information from different sources, different modalities, and different levels.

[0145] Step b: Update the node representation based on the initial embedding value, the neighborhood set and mapping matrix corresponding to each node, the self-loop transformation matrix, and the relation weight coefficients to obtain the enhanced node representation. In this embodiment, regarding relation modeling, this scheme considers that the relations in equipment fault diagnosis include not only structural connection relations, but also functional dependency relations, parameter coupling relations, anomaly indication relations, causal inference relations, operating condition constraint relations, and knowledge support relations. Therefore, for nodes... Its neighboring nodes Relationship between edges A relationship-aware message passing mechanism is introduced to assign differentiated relationship weight coefficients to different relationship types.

[0146] In this embodiment, the enhanced node representation is updated based on the initial embedding value, the neighborhood set and mapping matrix corresponding to each node, the self-loop transformation matrix, and the relation weight coefficients, to obtain the calculation formula:

[0147]

[0148] In the formula, To enhance node representation; For nodes In relation types The neighborhood set below; For the first Hierarchical Relationship Type The corresponding mapping matrix, Values ​​start from 0; It is a self-loop transformation matrix; It is a non-linear activation function; This refers to the relational weighting coefficients when aggregating neighborhood information.

[0149] Furthermore, the above expressions enable differentiated modeling of different relationship types, avoiding the simplistic and equivalent treatment of various relationships in industrial fault diagnosis.

[0150] Step c: Determine the time-enhanced representation of the corresponding node at any given time based on the enhanced node representation.

[0151] For example, the time augmentation of the corresponding node at any given time is represented as: node At any moment Moment enhancement is represented as .

[0152] Step d: Determine the corresponding time-varying augmentation representation based on the time-enhanced representation and the preset fusion coefficients.

[0153] In this embodiment, the calculation formula for determining the corresponding time-varying enhancement representation based on the time-enhanced representation and the preset fusion coefficient is as follows:

[0154]

[0155] In the formula, For time-varying augmentation representation; The preset fusion coefficient represents the fusion coefficient between information at the current time and information at historical time. The time is enhanced.

[0156] Furthermore, this design helps to capture the dynamic pattern of faults gradually evolving from early, weak symptoms to significant anomalies, thereby improving the ability to identify progressive faults and faults under complex operating conditions.

[0157] Step e: After time-varying enhanced representation, select key subgraphs from the industrial fault diagnosis graph that are related to preset fault samples.

[0158] Specifically, after time-varying augmented representation, key subgraphs related to preset fault samples are extracted from the industrial fault diagnosis graph based on node importance, relation weight, path confidence and task relevance.

[0159] For example, the preset fault sample is the current fault sample; the key subgraph is:

[0160]

[0161] In the formula, Key subgraph; A subset of nodes; It is a subset of edges.

[0162] The key subgraph is used to centrally express the core evidence required for this fault diagnosis, including key abnormal parameters, key components, key fault phenomena, candidate fault modes, and key relationships and key paths that support the connections between these objects.

[0163] Furthermore, compared to directly inputting the entire image into subsequent modules, this method can reduce interference from irrelevant structures, improve fault diagnosis efficiency and result reliability.

[0164] S206: Determine the graph evidence set based on the key subgraph, the pre-screened set of key failure paths, and the statistical data corresponding to the pre-generated path confidence.

[0165] In this embodiment, the process of generating path confidence includes steps a~b:

[0166] Step a: Determine the fault concern score for the corresponding node based on the preset monitoring signal anomaly score, preset event or alarm information score, preset knowledge rule score, and preset historical case support score.

[0167] In this embodiment, to highlight the components, parameters, and relationships that truly play a key role in the fault diagnosis process, a fault attention weighting mechanism is introduced during graph augmentation. This mechanism dynamically weights the diagnostic contributions of nodes and edges based on the anomaly intensity, event level, knowledge matching degree, and historical case support in the current sample to be diagnosed.

[0168] For example, the corresponding node is determined based on the preset monitoring signal anomaly score, the preset event or alarm information score, the preset knowledge rule score, and the preset historical case support score. The formula for calculating the fault concern score is:

[0169]

[0170] In the formula, Fault attention score; The preset score is used to indicate the degree of anomaly in the monitoring signal. Assign a score to the preset event or alarm information; Preset knowledge rule scores; Preset historical case support scores; These are the corresponding weighting coefficients.

[0171] Furthermore, based on this fault attention score, key nodes and key edges most relevant to the current fault can be prioritized during graph augmentation, thereby improving the relevance of structural representation.

[0172] Step b: Determine the path confidence based on the fault concern score, the candidate propagation paths of the pre-constructed fault mode nodes, and the edge association strength between adjacent nodes.

[0173] In this embodiment, to further characterize the chain-like propagation of equipment failures from local anomalies to system failures, a critical failure path modeling mechanism is introduced. An example of a candidate propagation path from anomaly parameter nodes to failure mode nodes is shown below. .

[0174] In this embodiment, the formula for calculating the path confidence based on the fault concern score, the pre-constructed candidate propagation paths of fault mode nodes, and the edge association strength between adjacent nodes is as follows:

[0175]

[0176] In the formula, For path confidence; The strength of edge association between adjacent nodes; Fault concern score for nodes on the path; This is a consistency evaluation item for the mechanism of path relationships. These are candidate propagation paths for nodes in failure mode; This is the adjustment coefficient.

[0177] In addition, candidate propagation paths can be scored and ranked to select the key fault propagation paths with the most diagnostic and explanatory value, which can be used for subsequent root cause analysis and diagnostic basis generation in graph-enhanced industrial large-scale models.

[0178] Specifically, step S206 involves determining the graph evidence set based on the key subgraph, the pre-screened set of key failure paths, and the statistical data corresponding to the pre-generated path confidence levels. The formula for this determination is as follows:

[0179]

[0180] In the formula, For the collection of evidence; Based on the key subgraph Sure; A set of critical failure paths; This refers to the statistical data corresponding to the path confidence level.

[0181] Furthermore, this graph evidence set not only retains the core diagnostic criteria in the graph structure, but also facilitates semantic reasoning, root cause localization, and diagnostic interpretation generation in subsequent graph-enhanced industrial large-scale models.

[0182] S207: Combined input data is formed based on preset fault samples, graphical evidence sets, preset historical case data, and constraints.

[0183] In this embodiment, the calculation formula for the joint input data, which is based on preset fault samples, a set of graphical evidence, preset historical case data, and constraints, is as follows:

[0184]

[0185] In the formula, For combined input data; Preset fault samples; For the collection of evidence; This is a pre-defined set of historical case data, used to represent the set of historical cases similar to the current sample; These are constraints used to represent equipment mechanism rules, process boundaries, safety constraints, and logical consistency constraints.

[0186] Furthermore, through this multi-source joint input method, the graph-enhanced industrial large model not only receives surface semantics, but also simultaneously receives relational evidence and constraints from the graph structure, thereby avoiding the semantic ambiguity and lack of industrial knowledge caused by relying solely on natural language descriptions.

[0187] S208: Perform a structured transformation on the combined input data to obtain enhanced prompt data.

[0188] In this embodiment, during the input organization stage, the joint input can be further constructed with hierarchical prompts based on the fault diagnosis task objectives. For example, the input content can be organized in the order of equipment object layer—anomaly representation layer—relationship layer—candidate fault layer—constraint verification layer, so that the graph-augmented industrial large model can prioritize understanding the current diagnostic object, anomaly symptoms, and key relationships before performing root cause reasoning and conclusion generation.

[0189] In this embodiment, the combined input data undergoes a structured transformation to obtain the calculation formula for the enhanced prompt data:

[0190]

[0191] In the formula, To enhance the prompt data; This is a function for summarizing fault evidence.

[0192] S209: Determine the corresponding complete input data based on the combined input data and the enhanced prompt data.

[0193] In this embodiment, the calculation formula for determining the corresponding complete input data based on the combined input data and enhanced prompt data is as follows:

[0194]

[0195] In the formula, To input complete data; This indicates splicing or template-based merging operations.

[0196] Furthermore, through the above design, the structural evidence output by the graph augmentation module is explicitly integrated into the diagnostic reasoning context of the graph augmentation industrial big model. When answering fault diagnosis questions, the model can rely on key components, relational paths, and knowledge support in the graph, rather than just performing general semantic completion.

[0197] S210: Input the complete input data into the graph-enhanced industrial model for inference processing to obtain fault diagnosis data.

[0198] For example, the fault diagnosis data is Y.

[0199] In this embodiment, the fault diagnosis data includes a first fault diagnosis result and a second fault diagnosis result; correspondingly, step S210 specifically includes steps a~b:

[0200] Step a: Input the complete input data into the graph-enhanced industrial model for inference processing to obtain the first fault diagnosis result.

[0201] In this embodiment, the complete input data is input into the graph-enhanced industrial large-scale model for inference processing, and the calculation formula for the first fault diagnosis result is as follows:

[0202]

[0203] In the formula, This is the first fault diagnosis result; To enhance the inference mapping of the large industrial model; To input complete data.

[0204] The first fault diagnosis result shall include at least one or more of the following: fault category determination, fault component location, fault root cause description, fault propagation path, fault severity evaluation, maintenance recommendations, and further testing recommendations.

[0205] Step b: During the reasoning process of the graph-enhanced industrial large model, a second fault diagnosis result is generated based on the fault category, fault location, root cause analysis results, fault propagation or evidence path, and maintenance or inspection suggestions.

[0206] In this embodiment, during the graph-enhanced industrial large-scale model inference process, the calculation formula for generating the second fault diagnosis result based on the fault category, fault location, root cause analysis results, fault propagation or evidence path, and maintenance or inspection suggestions is as follows:

[0207]

[0208] In the formula, This is the second fault diagnosis result; Fault category; Location of the fault; This is the result of the root cause analysis; For fault propagation or evidence path; This provides repair or inspection recommendations.

[0209] Furthermore, this output structure facilitates the one-to-one correspondence between the diagnostic results of the graph-enhanced industrial large model and the structural evidence of the graph-enhanced module, providing an interface for subsequent consistency verification and evidence tracing.

[0210] In addition, after step S210, the method further includes: in order to enhance the constraint response capability of the graph-enhanced industrial big data model to the diagnostic task, a candidate fault scoring mechanism based on graph evidence is introduced during the inference process of the graph-enhanced industrial big data model, so that the graph-enhanced industrial big data model generates a set of candidate faults; any candidate fault is selected from the set of candidate faults, and the corresponding comprehensive diagnostic confidence is determined based on the semantic confidence of the candidate fault, the consistency score with the key subgraph evidence, the support of historical cases, and the degree of satisfaction of rule constraints.

[0211] An example of the candidate fault set is as follows:

[0212]

[0213] In the formula, For the set of candidate faults; This is the kth candidate fault.

[0214] The formula for calculating the comprehensive diagnostic confidence score, which involves selecting any candidate fault from the candidate fault set and determining its corresponding semantic confidence score, consistency score with key subgraph evidence, historical case support, and rule constraint satisfaction, is as follows:

[0215]

[0216] in, To assess the overall diagnostic confidence level; To enhance the industrial large model's ability to detect candidate faults The semantic generation confidence; Candidate fault Consistency score with evidence from key subgraphs; To demonstrate the support of historical cases; For the degree of satisfaction of rule constraints; These are the weighting coefficients.

[0217] Furthermore, this comprehensive scoring method can be used to sort and filter the fault diagnosis results generated by the graph-augmented industrial large model, thereby improving the reliability of the final diagnosis results.

[0218] In addition, after step S210, steps c~e are also included:

[0219] Step c: Based on the multi-source heterogeneous fault data, determine the original sensor data fragments, event records or text records, key nodes, key relationships, critical paths, and time windows or timestamp sets related to the fault diagnosis data, so as to calculate the corresponding evidence chain.

[0220] In this embodiment, based on multi-source heterogeneous fault data, the original sensor data fragments, event records or text records, key nodes, key relationships, critical paths, and time windows or timestamp sets related to the fault diagnosis data are determined, and the calculation formula for the corresponding evidence chain is as follows:

[0221]

[0222] In the formula, For the chain of evidence; This includes raw sensor data fragments and event or text records related to fault diagnosis data; As a key node; As a key relationship, For the critical path; A time window or a set of timestamps.

[0223] Furthermore, through the aforementioned backtracking mechanism, a clear correspondence can be established between the fault category, the faulty component, the propagation link, and the original data source, making the output results traceable. This not only makes it easier for maintenance personnel to understand the diagnostic basis, but also facilitates subsequent operation and maintenance review and model optimization.

[0224] Step d: Determine the credibility evaluation value based on the confidence level, consistency evaluation value, rule satisfaction, historical case support, or multi-source evidence support of the fault diagnosis data.

[0225] In this embodiment, the formula for calculating the credibility evaluation value based on the confidence level, consistency evaluation value, rule satisfaction, historical case support, or multi-source evidence support of the fault diagnosis data is as follows:

[0226]

[0227] In the formula, This is a credibility rating value; To enhance the industrial large model's ability to evaluate candidate results The generation confidence level; This is the consistency evaluation value; For rule satisfaction, For historical case support or multi-source evidence support, These are the corresponding weighting coefficients. The system can, according to... The system ranks multiple candidate diagnostic conclusions and selects the result with the highest confidence level as the current output candidate.

[0228] The calculation of the consistency evaluation value and rule satisfaction is detailed in S301 and S302 of the following embodiments.

[0229] Step e: Determine the structured diagnostic results of the equipment based on the final fault category, fault location results, root cause description, fault propagation or evidence path, maintenance or testing recommendations, evidence chain and credibility evaluation value.

[0230] In this embodiment, the formula for calculating the structured diagnostic results of the equipment based on the final fault category, fault location results, root cause description, fault propagation or evidence path, maintenance or testing recommendations, evidence chain, and credibility evaluation value is as follows:

[0231]

[0232] In the formula, The results are structured diagnostics. The final fault category; The results of the fault location; Explain the root cause; For fault propagation or evidence path; For repair or inspection recommendations; For the chain of evidence; This is the credibility rating.

[0233] Furthermore, this output method enables the system not only to answer what fault occurred, but also to answer where the fault occurred, why the conclusion was reached, where the basis for the conclusion came from, and how to proceed next, thus better meeting the engineering requirements of equipment fault diagnosis.

[0234] Furthermore, graph-enhanced industrial large-scale models can employ large language models adapted to industrial corpora or fine-tuned for the task, multimodal industrial models, or industry knowledge-enhanced models. The key is not to limit oneself to a specific large-scale model structure, but to provide structured fault evidence and industrial constraint information through graph enhancement mechanisms, thereby achieving collaborative work between graphs and graph-enhanced industrial large-scale models in fault diagnosis tasks. In other words, the focus of protection is on the graph-enhanced evidence-driven large-scale model diagnostic reasoning mechanism, rather than being limited to a specific form of model parameters.

[0235] In addition, the graph-enhanced industrial big data model also supports a hierarchical output mode. For on-site maintenance personnel, it can output concise diagnostic conclusions, key component locations and handling suggestions. For expert users or maintenance management systems, it can further output key evidence nodes, fault paths, rule verification results and raw data indexes. This hierarchical output mechanism can take into account both on-site readability and technical auditability.

[0236] In summary, the equipment fault diagnosis method based on a graph-enhanced industrial large-scale model provided in this embodiment receives multi-source heterogeneous fault data generated by sensors during equipment operation; preprocesses the heterogeneous fault data to obtain unified analysis data; determines the node set, edge set, relational adjacency representation, and node attribute matrix based on the unified analysis data, a preset set of node types, a preset set of edge types, and historical fault data; constructs an industrial fault diagnosis graph for the equipment based on the preset set of node types, the preset set of edge types, the node set, the edge set, the relational adjacency representation, and the node attribute matrix; and determines the corresponding key subgraph based on the industrial fault diagnosis graph. The graph evidence set is determined based on the key subgraph, the pre-screened set of key fault paths, and the statistical data corresponding to the pre-generated path confidence. Joint input data is constructed based on preset fault samples, the graph evidence set, preset historical case data, and constraints. The joint input data undergoes structured transformation to obtain enhanced prompt data. The corresponding complete input data is determined based on the joint input data and the enhanced prompt data. The complete input data is then input into the graph-enhanced industrial big data model for inference processing to obtain fault diagnosis data. Through the fusion processing of multi-source heterogeneous fault data, fault diagnosis data is generated, thereby improving the accuracy of equipment fault diagnosis.

[0237] Furthermore, the equipment fault diagnosis method based on a graph-augmented industrial large-scale model provided in this embodiment constructs an industrial fault diagnosis graph for equipment based on a preset set of node types, a preset set of edge types, a set of nodes, a set of edges, relational adjacency representations, and a node attribute matrix. Compared to existing methods that rely solely on time-series models or simple feature splicing, this industrial fault diagnosis graph can not only express the complex structural dependencies and fault propagation chains within the equipment, but also incorporate historical knowledge, expert rules, and dynamic operating condition information into a unified modeling framework. This significantly improves the effectiveness and interpretability of subsequent graph-augmented representations and graph-augmented industrial large-scale model fault diagnosis reasoning. Moreover, the industrial fault diagnosis graph enables previously scattered and differently expressed fault information to establish relationships within a unified framework, thereby improving the completeness of fault information expression and the sufficiency of diagnostic basis. Furthermore, by introducing the industrial fault diagnosis graph, the relationships between equipment components, component couplings, parameter correlations, and the causal relationships between fault phenomena and fault modes are uniformly expressed. This allows fault diagnosis to move beyond local feature recognition and reflect the propagation path and evolution process of faults within the equipment, thereby improving the ability to identify complex and chain-like faults.

[0238] In addition, the equipment fault diagnosis method based on graph-enhanced industrial large model provided in this embodiment introduces structured evidence such as key subgraphs, key relationships and key paths into the input of the graph-enhanced industrial large model, enabling the graph-enhanced industrial large model to have structure perception and relationship constraint capabilities during the reasoning process, avoiding relying solely on surface semantics for judgment, thereby reducing diagnostic bias and conclusion drift problems.

[0239] Furthermore, the equipment fault diagnosis method based on graph-enhanced industrial large-scale models provided in this embodiment determines the original sensor data fragments, event records or text records, key nodes, key relationships, critical paths, and time windows or timestamp sets related to the fault diagnosis data based on multi-source heterogeneous fault data, in order to calculate the corresponding evidence chain; determines the credibility evaluation value based on the confidence level, consistency evaluation value, rule satisfaction, historical case support, or multi-source evidence support of the fault diagnosis data; and determines the structured diagnosis result of the equipment based on the final fault category, fault location result, root cause description, fault propagation or evidence path, maintenance or testing suggestions, evidence chain, and credibility evaluation value. This ensures that the output structured diagnosis result not only includes fault conclusions but also clearly provides the diagnostic basis and associated paths, meeting the requirements of industrial field for the credibility and auditability of diagnostic results.

[0240] Furthermore, the equipment fault diagnosis method based on graph-augmented industrial big data model provided in this embodiment preprocesses continuous monitoring time-series data to obtain standardized time-series segments for characterizing equipment state changes; preprocesses event and alarm data to obtain standard alarm data for corresponding events; preprocesses text record data to obtain standard record data for corresponding texts; preprocesses equipment structure and process data to obtain structural relationship data between any two components; and performs cross-source alignment and correlation mapping processing based on standardized time-series segments, standard alarm data for each event, standard record data for each text, structural relationship data, and operating condition and environmental context data to obtain unified analysis data. This ensures that the subsequent construction of fault diagnosis graphs is no longer limited to a single modality but can comprehensively utilize multi-dimensional information such as sensor status, event trajectory, structural relationships, historical knowledge, and operating condition context, thereby providing a complete, accurate, and computable data foundation for intelligent fault diagnosis of graph-augmented industrial big data model.

[0241] In addition, the equipment fault diagnosis method based on graph-augmented industrial big model provided in this embodiment can effectively identify key anomaly sources, characterize fault propagation links, and extract structural evidence most relevant to the diagnosis task, thereby significantly improving the accuracy and reliability of subsequent graph-augmented industrial big model in fault diagnosis, root cause reasoning, and result interpretation.

[0242] Furthermore, the equipment fault diagnosis method based on graph-enhanced industrial large model provided in this embodiment transforms the equipment structure relationships, parameter coupling relationships, fault propagation relationships, and knowledge support relationships contained in the heterogeneous graph into enhanced inputs that can be directly utilized by the graph-enhanced industrial large model. This can more effectively reduce semantic bias, lack of evidence, and conclusion drift in the diagnostic results, thereby significantly improving the accuracy, reliability, and interpretability of equipment fault diagnosis.

[0243] In addition, the equipment fault diagnosis method based on graph-augmented industrial large model provided in this embodiment can be flexibly expanded according to different equipment structures and business needs, and has strong versatility and promotion value, which can provide effective technical support for industrial intelligent operation and maintenance and fault diagnosis systems.

[0244] Figure 3 A flowchart illustrating the equipment fault diagnosis method based on a graph-enhanced industrial large model provided in this application embodiment. Figure 2 In the embodiments of this application, in Figure 2 Based on the provided embodiments, a detailed explanation is given of the specific implementation method for correcting fault diagnosis data after step S210. For example... Figure 3 As shown, the method includes:

[0245] S301: Calculate the corresponding consistency evaluation value based on the fault diagnosis data and the set of graphical evidence.

[0246] In this embodiment, the formula for calculating the corresponding consistency evaluation value based on the fault diagnosis data and the graphical evidence set is as follows:

[0247]

[0248] in, The degree of matching between the faulty components, abnormal parameters, and fault modes involved in the results and the set of key nodes; This represents the degree of matching between the relationships involved in the results and the set of key relationships; This refers to the degree of matching between the root cause link or explanation path in the results and the set of critical failure paths; Y represents the weighting coefficients; Y represents the fault diagnosis data. For the collection of evidence.

[0249] S302: Calculate the rule satisfaction degree corresponding to the fault diagnosis data based on the rule set.

[0250] For example, the rule set is as follows: .

[0251] In this embodiment, the formula for calculating the rule satisfaction degree corresponding to the fault diagnosis data based on the rule set is as follows:

[0252]

[0253] In the formula, The degree of rule satisfaction is represented by M, and the number of rules is M. Does the fault diagnosis data meet the requirements of section 1? Rules.

[0254] In addition, when A lower value indicates a conflict between the diagnostic conclusion and the equipment mechanism or process logic. For example, if the fault propagation path corresponding to a certain fault conclusion violates the equipment connection direction, or if the recommended maintenance measures are inconsistent with safety constraints, then the result can be determined to have a mechanism inconsistency problem.

[0255] S303: Determine whether the consistency evaluation value is greater than the preset threshold, and determine whether the rule satisfaction meets the corresponding rule.

[0256] For example, the rule is the m-th rule.

[0257] S304: If the consistency evaluation value is greater than or equal to the preset threshold and the rule satisfaction meets the corresponding rule, then the fault diagnosis data is determined as a reliable diagnosis result.

[0258] S305: If the consistency evaluation value is less than the preset threshold and / or the rule satisfaction does not meet the corresponding rule, then the complete input data is corrected to obtain the corrected complete input data.

[0259] In this embodiment, if the consistency evaluation value is determined to be less than a preset threshold and / or the rule satisfaction does not meet the corresponding rule, a result optimization mechanism is triggered. This mechanism includes four methods: evidence recombination, local subgraph supplementation, constraint strengthening hints, and re-reasoning. Specifically, key subgraphs can be re-extracted based on conflict items. Supplement the local evidence related to the disputed faulty component, failure mode, or operating condition, and then revise the complete input data based on the newly added graphical evidence and rule constraints to obtain the revised complete input data. .

[0260] S306: Input the corrected complete input data into the graph-enhanced industrial large model for inference processing to obtain the final fault diagnosis data.

[0261] In this embodiment, the corrected complete input data is input into the graph-enhanced industrial large-scale model for inference processing to obtain the calculation formula for the final fault diagnosis data:

[0262]

[0263] In the formula, This is the final fault diagnosis data.

[0264] In addition, if It outperforms the initial results in consistency score, rule satisfaction score, and reliable output score. Then Alternative This serves as the final fault diagnosis data. Thus, a closed-loop optimization mechanism of generation—verification—correction—re-output is formed.

[0265] In summary, the equipment fault diagnosis method based on a graph-augmented industrial big data model provided in this embodiment calculates the corresponding consistency evaluation value based on the fault diagnosis data and the graph evidence set; calculates the rule satisfaction degree corresponding to the fault diagnosis data based on the rule set; determines whether the consistency evaluation value is greater than a preset threshold and whether the rule satisfaction degree satisfies the corresponding rule; if the consistency evaluation value is greater than or equal to the preset threshold and the rule satisfaction degree satisfies the corresponding rule, the fault diagnosis data is determined as a reliable diagnosis result; if the consistency evaluation value is less than the preset threshold and / or the rule satisfaction degree does not satisfy the corresponding rule, the complete input data is corrected to obtain corrected complete input data; the corrected complete input data is input into the graph-augmented industrial big data model for inference processing to obtain the final fault diagnosis data. This allows the fault diagnosis data to be optimized through evidence recombination and re-inference, thereby avoiding the direct output of erroneous fault diagnosis data and improving the reliability of equipment fault diagnosis.

[0266] Figure 4 This is a schematic diagram of the equipment fault diagnosis device based on a graph-enhanced industrial large model provided in an embodiment of this application. Figure 4 As shown, the equipment fault diagnosis device based on graph-enhanced industrial large model includes: a receiving module 401, a processing module 402, a first determining module 403, a construction module 404, a second determining module 405, a third determining module 406, a composition module 407, a conversion module 408, a fourth determining module 409, and a first reasoning module 410.

[0267] Receiver module 401 is used to receive multi-source heterogeneous fault data generated by the sensor acquisition equipment during operation;

[0268] The processing module 402 is used to preprocess the heterogeneous fault data to obtain unified analysis data;

[0269] The first determining module 403 is used to determine the node set, edge set, relational adjacency representation and node attribute matrix based on unified analysis data, preset node type set, preset edge type set and historical fault data;

[0270] Module 404 is used to construct an industrial fault diagnosis diagram of equipment based on a preset set of node types, a preset set of edge types, a set of nodes, a set of edges, a relational adjacency representation, and a node attribute matrix.

[0271] The second determining module 405 is used to determine the corresponding key sub-graph based on the industrial fault diagnosis diagram;

[0272] The third determination module 406 is used to determine the graph evidence set based on the key subgraph, the pre-screened set of key fault paths and the statistical data corresponding to the pre-generated path confidence.

[0273] The component module 407 is used to construct joint input data based on preset fault samples, graphical evidence sets, preset historical case data, and constraints.

[0274] The conversion module 408 is used to perform structured conversion on the combined input data to obtain enhanced prompt data;

[0275] The fourth determining module 409 is used to determine the corresponding complete input data based on the combined input data and the enhanced prompt data;

[0276] The first reasoning module 410 is used to input the complete input data into the graph-enhanced industrial model for reasoning processing to obtain fault diagnosis data.

[0277] In one possible implementation, the heterogeneous fault data includes continuous monitoring time-series data, event and alarm data, text log data, equipment structure and process data, and operating condition and environmental context data; correspondingly, the processing module 402 specifically includes:

[0278] The first processing unit is used to preprocess the continuous monitoring time series data to obtain standardized time series segments that characterize changes in equipment status.

[0279] The second processing unit is used to preprocess the event and alarm data to obtain the standard alarm data for the corresponding event.

[0280] The third processing unit is used to preprocess the text record data to obtain the standard record data of the corresponding text.

[0281] The fourth processing unit is used to preprocess the equipment structure and process data to obtain structural relationship data between any two components;

[0282] The fifth processing unit is used to perform cross-source alignment and correlation mapping based on standardized time sequence segments, standard alarm data of each event, standard record data of each text, structural relationship data, and operating condition and environmental context data to obtain unified analysis data.

[0283] In one possible implementation, the first determining module 403 specifically includes:

[0284] The first determining unit is used to determine the corresponding node set based on a preset node type set;

[0285] The second determining unit is used to determine the degree of correlation between two random parameters based on any two random parameters determined by unified analysis data and historical fault data, as well as the standard deviation of each random parameter.

[0286] The third determining unit is used to determine the corresponding weight based on the degree of relevance and to determine the corresponding edge type based on the preset edge type set;

[0287] The fourth determining unit is used to determine the corresponding candidate edges based on the weights, any two nodes, and edge types, and to determine the edge set based on each candidate edge.

[0288] The fifth determining unit is used to determine the state subgraphs formed by the equipment under different time windows, and to determine the sequence of time diagrams based on each state subgraph to represent the relational adjacency representation;

[0289] The acquisition unit is used to obtain the initial attributes of each node based on the node set in order to obtain the node attribute matrix.

[0290] In one possible implementation, the second determining module 405 specifically includes:

[0291] The mapping unit is used to perform heterogeneous node feature mapping based on the industrial fault diagnosis map to obtain the initial embedding value;

[0292] The update unit is used to update the node representation based on the initial embedding value, the neighborhood set and mapping matrix corresponding to each node, the self-loop transformation matrix and the relation weight coefficients.

[0293] The first determining unit is used to determine the time-enhanced representation of the corresponding node at any given time based on the enhanced node representation;

[0294] The second determining unit is used to determine the corresponding time-varying enhancement representation based on the time-enhanced representation and the preset fusion coefficients;

[0295] The filtering unit is used to filter key subgraphs related to preset fault samples from the industrial fault diagnosis graph after time-varying enhanced representation.

[0296] In one possible implementation, the device further includes:

[0297] The fifth determination module is used to determine the fault concern score of the corresponding node based on the preset monitoring signal anomaly score, preset event or alarm information score, preset knowledge rule score, and preset historical case support score.

[0298] The sixth determination module is used to determine the path confidence based on the fault concern score, the candidate propagation paths of pre-constructed fault mode nodes, and the edge association strength between adjacent nodes.

[0299] In one possible implementation, the fault diagnosis data includes a first fault diagnosis result and a second fault diagnosis result; correspondingly, the first inference module 410 specifically includes:

[0300] The reasoning unit is used to input the complete input data into the graph-enhanced industrial model for reasoning processing to obtain the first fault diagnosis result;

[0301] The generation unit is used to generate a second fault diagnosis result during the reasoning process of graph-enhanced industrial large model, based on the fault category, fault location, root cause analysis results, fault propagation or evidence path, and maintenance or inspection suggestions.

[0302] In one possible implementation, the device further includes:

[0303] The first calculation module is used to calculate the corresponding consistency evaluation value based on the fault diagnosis data and the set of graphical evidence.

[0304] The second calculation module is used to calculate the rule satisfaction degree corresponding to the fault diagnosis data based on the rule set.

[0305] The judgment module is used to determine whether the consistency evaluation value is greater than the preset threshold and whether the rule satisfaction meets the corresponding rule.

[0306] The seventh determination module is used to determine the fault diagnosis data as a reliable diagnosis result if the consistency evaluation value is greater than or equal to the preset threshold and the rule satisfaction meets the corresponding rule.

[0307] The correction module is used to correct the complete input data if the consistency evaluation value is less than a preset threshold and / or the rule satisfaction does not meet the corresponding rule, so as to obtain the corrected complete input data.

[0308] The second inference module is used to input the corrected complete input data into the graph-enhanced industrial model for inference processing to obtain the final fault diagnosis data.

[0309] In one possible implementation, the device further includes:

[0310] The eighth determination module is used to determine the original sensor data fragments, event records or text records, key nodes, key relationships, critical paths and time windows or timestamp sets related to the fault diagnosis data based on the multi-source heterogeneous fault data, so as to calculate the corresponding evidence chain.

[0311] The ninth determination module is used to determine the credibility evaluation value based on the confidence level, consistency evaluation value, rule satisfaction, historical case support, or multi-source evidence support of the fault diagnosis data.

[0312] The tenth determination module is used to determine the structured diagnostic results of the equipment based on the final fault category, fault location results, root cause description, fault propagation or evidence path, maintenance or testing recommendations, evidence chain and credibility evaluation value.

[0313] The apparatus provided in this embodiment can be used to execute the technical solutions of the above method embodiments. Its implementation principle and technical effects are similar, and will not be described again here.

[0314] Figure 5 This is a schematic diagram of the hardware structure of the electronic device provided in an embodiment of this application. Figure 5 As shown, the electronic device of this embodiment includes: a processor 501 and a memory 502; the memory stores computer-executable instructions; at least one processor executes the computer-executable instructions stored in the memory, causing at least one processor to execute the above-described equipment fault diagnosis method based on a graph-enhanced industrial large model.

[0315] Alternatively, the memory 502 can be either standalone or integrated with the processor 501.

[0316] When the memory 502 is set up independently, the electronic device also includes a bus 503 for connecting the memory 502 and the processor 501.

[0317] This application also provides a computer storage medium storing computer execution instructions. When the processor executes the computer execution instructions, it implements the above-mentioned equipment fault diagnosis method based on graph-enhanced industrial large model.

[0318] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described equipment fault diagnosis method based on a graph-enhanced industrial large model.

[0319] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or modules, and may be electrical, mechanical, or other forms.

[0320] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to implement the solution of this embodiment according to actual needs.

[0321] Furthermore, the functional modules in the various embodiments of this application can be integrated into one processing unit, or each module can exist physically separately, or two or more modules can be integrated into one unit. The unit composed of the above modules can be implemented in hardware or in the form of hardware plus software functional units.

[0322] The integrated modules described above, implemented as software functional modules, can be stored in a computer-readable storage medium. These software functional modules, stored in a storage medium, include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute some steps of the methods of the various embodiments of this application.

[0323] It should be understood that the aforementioned processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. A general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly manifested as execution by a hardware processor, or execution by a combination of hardware and software modules within the processor.

[0324] The memory may include high-speed RAM, and may also include non-volatile storage (NVM), such as at least one disk storage device, and may also be a USB flash drive, external hard drive, read-only memory, disk or optical disc, etc.

[0325] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0326] The aforementioned storage media can be implemented from any type of volatile or non-volatile storage device or a combination thereof, such as Static Random-Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The storage media can be any available medium accessible to general-purpose or special-purpose computers.

[0327] An exemplary storage medium is coupled to a processor, enabling the processor to read information from and write information to the storage medium. Alternatively, the storage medium can be an integral part of the processor. Both the processor and the storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and storage medium can exist as discrete components in an electronic device or host device.

[0328] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0329] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A method for equipment fault diagnosis based on a graph-enhanced industrial large-scale model, characterized in that, Applied to electronic devices, including: Receive multi-source heterogeneous fault data generated by sensor acquisition equipment during operation; Preprocess the heterogeneous fault data to obtain unified analysis data; Based on the unified analysis data, the preset node type set, the preset edge type set, and the historical fault data, determine the node set, edge set, relational adjacency representation, and node attribute matrix; The industrial fault diagnosis map of the equipment is constructed based on the preset node type set, the preset edge type set, the node set, the edge set, the relational adjacency representation, and the node attribute matrix. The corresponding key sub-graphs are determined based on the industrial fault diagnosis diagram; The graph evidence set is determined based on the key subgraph, the pre-screened set of key fault paths, and the statistical data corresponding to the pre-generated path confidence. The joint input data is constructed based on preset fault samples, the set of graphical evidence, preset historical case data, and constraints. The combined input data is subjected to a structured transformation to obtain enhanced prompt data; The corresponding complete input data is determined based on the combined input data and the enhanced prompt data; The complete input data is input into the graph-enhanced industrial model for inference processing to obtain fault diagnosis data.

2. The method according to claim 1, characterized in that, The heterogeneous fault data includes continuous monitoring time series data, event and alarm data, text record data, equipment structure and process data, and operating condition and environmental context data; Accordingly, the preprocessing of the heterogeneous fault data to obtain unified analysis data includes: The continuous monitoring time series data is preprocessed to obtain standardized time series segments used to characterize the changes in the equipment state; The event and alarm data are preprocessed to obtain standard alarm data for the corresponding events; The text record data is preprocessed to obtain the standard record data of the corresponding text; The equipment structure and process data are preprocessed to obtain structural relationship data between any two components; Cross-source alignment and correlation mapping are performed based on the standardized time sequence segments, standard alarm data for each event, standard record data for each text, structural relationship data, and operating condition and environmental context data to obtain unified analysis data.

3. The method according to claim 1, characterized in that, The step of determining the node set, edge set, relational adjacency representation, and node attribute matrix based on the unified analysis data, preset node type set, preset edge type set, and historical fault data includes: The corresponding node set is determined based on the preset node type set; The correlation between the two random parameters is determined based on any two random parameters determined from the unified analysis data and historical fault data, and the standard deviation of each random parameter. The corresponding weight is determined based on the degree of relevance, and the corresponding edge type is determined based on the preset edge type set; Based on the weights, any two nodes, and the edge type, determine the corresponding candidate edges, and determine the edge set based on each candidate edge; The state subgraphs formed by the equipment under different time windows are determined, and the timing diagram sequence is determined according to each state subgraph to represent the relational adjacency representation; The initial attributes of each node are obtained from the set of nodes to obtain the node attribute matrix.

4. The method according to claim 1, characterized in that, The step of determining the corresponding key sub-graph based on the industrial fault diagnosis graph includes: Heterogeneous node feature mapping is performed based on the industrial fault diagnosis map to obtain initial embedding values; The node representation is updated based on the initial embedding value, the neighborhood set and mapping matrix corresponding to each node, the self-loop transformation matrix and the relation weight coefficients to obtain the enhanced node representation. Determine the time-enhanced representation of the corresponding node at any given moment based on the enhanced node representation; The corresponding time-varying enhancement representation is determined based on the time-enhanced representation and the preset fusion coefficient; Following the time-varying enhanced representation, key subgraphs related to preset fault samples are selected from the industrial fault diagnosis graph.

5. The method according to claim 1, characterized in that, The process of generating path confidence includes: The fault concern score for the corresponding node is determined based on the preset abnormality score of the monitoring signal, the preset event or alarm information score, the preset knowledge rule score, and the preset historical case support score. The path confidence is determined based on the fault concern score, the candidate propagation paths of the pre-constructed fault mode nodes, and the edge association strength between adjacent nodes.

6. The method according to claim 1, characterized in that, The fault diagnosis data includes a first fault diagnosis result and a second fault diagnosis result; Accordingly, the step of inputting the complete input data into the graph-enhanced industrial model for inference processing to obtain fault diagnosis data includes: The complete input data is input into the graph-enhanced industrial model for inference processing to obtain the first fault diagnosis result; During the reasoning process of the graph-enhanced industrial large model, a second fault diagnosis result is generated based on the fault category, fault location, root cause analysis results, fault propagation or evidence path, and maintenance or inspection suggestions.

7. The method according to any one of claims 1 to 6, characterized in that, After the complete input data is input into the graph-enhanced industrial large model for inference processing to obtain fault diagnosis data, the method further includes: Calculate the corresponding consistency evaluation value based on the fault diagnosis data and the graphical evidence set; Calculate the rule satisfaction degree corresponding to the fault diagnosis data based on the rule set; Determine whether the consistency evaluation value is greater than a preset threshold, and determine whether the rule satisfaction meets the corresponding rule; If the consistency evaluation value is determined to be greater than or equal to the preset threshold and the rule satisfaction degree satisfies the corresponding rule, then the fault diagnosis data is determined as a reliable diagnosis result. If the consistency evaluation value is determined to be less than the preset threshold and / or the rule satisfaction does not meet the corresponding rule, then the complete input data is corrected to obtain the corrected complete input data; The corrected complete input data is then input into the graph-enhanced industrial model for inference processing to obtain the final fault diagnosis data.

8. The method according to claim 7, characterized in that, After the complete input data is input into the graph-enhanced industrial large model for inference processing to obtain fault diagnosis data, the method further includes: Based on the multi-source heterogeneous fault data, determine the original sensor data fragments, event records or text records, key nodes, key relationships, key paths, and time windows or timestamp sets related to the fault diagnosis data, so as to calculate the corresponding evidence chain. The credibility evaluation value is determined based on the confidence level of the fault diagnosis data, the consistency evaluation value, the rule satisfaction level, the historical case support level, or the multi-source evidence support level. The structured diagnostic results of the equipment are determined based on the final fault category, fault location results, root cause description, fault propagation or evidence path, maintenance or testing recommendations, the evidence chain, and the credibility evaluation value.

9. A fault diagnosis device for equipment based on a graph-enhanced industrial large-scale model, characterized in that, Applied to electronic devices, including: The receiving module is used to receive multi-source heterogeneous fault data generated by the sensor acquisition equipment during operation; The processing module is used to preprocess the heterogeneous fault data to obtain unified analysis data; The first determining module is used to determine the node set, edge set, relational adjacency representation, and node attribute matrix based on the unified analysis data, the preset node type set, the preset edge type set, and historical fault data. The construction module is used to construct an industrial fault diagnosis map of the equipment based on the preset node type set, the preset edge type set, the node set, the edge set, the relational adjacency representation, and the node attribute matrix. The second determining module is used to determine the corresponding key sub-graph based on the industrial fault diagnosis graph; The third determining module is used to determine the graph evidence set based on the key subgraph, the pre-screened set of key fault paths, and the statistical data corresponding to the pre-generated path confidence. The component module is used to construct joint input data based on preset fault samples, the set of graphical evidence, preset historical case data, and constraints. A conversion module is used to perform structured conversion on the combined input data to obtain enhanced prompt data; The fourth determining module is used to determine the corresponding complete input data based on the combined input data and the enhanced prompt data; The first reasoning module is used to input the complete input data into the graph-enhanced industrial model for reasoning processing to obtain fault diagnosis data.

10. An electronic device, characterized in that, include: At least one processor and memory; The memory stores computer-executed instructions; The at least one processor executes computer execution instructions stored in the memory, causing the at least one processor to perform the equipment fault diagnosis method based on graph-enhanced industrial large model as described in any one of claims 1 to 8.

Images