Early fault early warning method and system based on dynamic evolution of complex industrial map

Abstract

The invention relates to the technical field of fault early warning, in particular to an early fault early warning method and system based on dynamic evolution of a complex industrial map. The method comprises the following steps: constructing a multi-level industrial system knowledge graph based on acquired industrial equipment parameters; performing adaptive knowledge updating on the multilevel industrial system knowledge graph based on a graph dynamic evolution mechanism; constructing a joint reasoning framework based on the large model and the knowledge graph; optimizing a joint reasoning framework based on adaptive knowledge updating and a continuous learning mechanism; and carrying out early fault early warning on the industrial equipment by utilizing the optimized joint reasoning framework. According to the early fault early warning system based on complex industrial system graph dynamic evolution and large model joint reasoning, a dynamic knowledge graph and a large language model are fused, and early, accurate and interpretable industrial fault early warning is achieved.

Description

Technical Field

[0001] The present invention relates to the field of fault warning technology, and in particular to an early fault warning method and system based on the dynamic evolution of complex industrial graphs. Background Art

[0002] Traditional failure maintenance strategies are mainly divided into three categories: post-failure maintenance, preventive maintenance, and predictive maintenance. Post-failure maintenance involves repairs after a device fails. While this avoids unnecessary maintenance costs, it often results in greater losses. Preventive maintenance involves inspecting and maintaining equipment at fixed intervals, disregarding the equipment's actual condition and easily leading to over- or under-maintenance. Predictive maintenance is based on the equipment's actual operating status, continuously monitoring key equipment parameters to predict potential failures and taking action before they occur. It is currently the most advanced maintenance strategy. Research shows that compared to traditional post-failure maintenance strategies, effective early warning systems can reduce equipment downtime by 50%-70%, lower maintenance costs by 25%-30%, and extend equipment life by 20%-40%.

[0003] In early fault warning technology, rule-based approaches use pre-set expert rules or thresholds to determine whether a system is abnormal. These methods are simple to implement, computationally efficient, and offer highly interpretable results. However, they rely on expert experience, have poor generalization capabilities, and struggle to handle dynamic changes and uncertainty. Statistical methods, such as multivariate statistical process control and principal component analysis, can handle correlations between multivariate data and offer some robustness to data noise. However, they typically assume that data follows a specific distribution, have limited ability to model nonlinear relationships, and struggle to handle data with strong time-series dependencies.

[0004] Large language models (LLMs), with their powerful natural language processing and knowledge representation capabilities, have been widely used in the industrial field for applications such as intelligent document analysis, expert knowledge acquisition, fault diagnosis assistance, and anomaly description generation. However, in terms of fault warning, large language models still have limitations, such as limited numerical computing capabilities, insufficient domain knowledge, real-time challenges, reliability concerns, and a lack of structured representation for industrial systems.

[0005] Overall, current industrial fault warning technology faces core challenges, including system complexity and dynamics, the separation of knowledge and data, the trade-off between explainability and accuracy, the weakness of early warning signs, the formalization and updating of domain knowledge, the integration of multi-source heterogeneous data, computing efficiency and resource consumption, the scarcity of fault samples, the lack of integration between system knowledge structure and data-driven models, and the lack of operability of warning results. Specifically, they include: The complexity of industrial system structures and component relationships makes it difficult for traditional fault warning methods to fully grasp the overall status of the system and internal interaction mechanisms, resulting in inaccurate predictions of fault propagation paths and impact ranges, seriously affecting the timeliness and effectiveness of warnings.

[0006] Industrial knowledge and real-time data have long been disconnected, and existing technologies are unable to organically integrate the two. Knowledge graphs can represent structured knowledge but lack integration with dynamic data. Data-driven models can process real-time data but struggle to incorporate domain knowledge. This disconnect limits fault early warning capabilities.

[0007] The early signals of faults are weak and drowned out by a large amount of normal data and background noise. Existing technologies are unable to identify such weak signals, resulting in the inability to issue early warnings in the early stages of faults and missing the best time for maintenance.

[0008] Large-scale language models face domain adaptability challenges when applied in industrial settings. General-purpose large-scale models lack sufficient industrial expertise and have limited capabilities in numerical calculations and time series analysis, making them difficult to directly apply to highly specialized industrial fault warning tasks.

[0009] In order to overcome the limitations of existing technologies, it is urgent to develop a new type of industrial fault warning system that can organically integrate system structure knowledge, expert experience and data-driven models, dynamically capture system evolution, and achieve early, accurate and explainable fault warning. Summary of the Invention

[0010] In order to solve the above-mentioned problems, the present invention provides an early fault warning method and system based on the dynamic evolution of complex industrial graphs.

[0011] In a first aspect, the present invention provides an early fault warning method based on the dynamic evolution of complex industrial graphs, which adopts the following technical solutions: An early fault warning method based on the dynamic evolution of complex industrial graphs, including: Obtain industrial equipment parameters; Construct a multi-level industrial system knowledge graph based on the acquired industrial equipment parameters; Adaptive knowledge update of multi-level industrial system knowledge graph based on graph dynamic evolution mechanism; Build a joint reasoning framework based on large models and knowledge graphs; Optimize the joint reasoning framework based on adaptive knowledge updating and continuous learning mechanisms; Utilize the optimized joint reasoning framework for early failure warning of industrial equipment.

[0012] Furthermore, the multi-level industrial system knowledge graph is constructed based on the acquired industrial equipment parameters, including defining the overall architecture of the graph, constructing a basic graph representing the physical structure of the industrial system and the relationship between equipment components through physical entity node components and parameter characteristics, forming a physical layer subgraph; constructing an abstract representation describing the system functional modules and their interactive relationships through the functional relationship of functional unit nodes, forming a functional layer subgraph; and establishing a network representation of monitoring parameters and their association relationships through monitoring node data and monitoring devices, forming a monitoring layer subgraph. The multi-level industrial system knowledge graph is represented as follows: , in, Represents the physical layer subgraph, including the physical entities of devices and components and their connection relationships; Represents the functional layer subgraph, describing the functional units of the system and their interactions; Represents the monitoring layer subgraph, which contains sensor and monitoring point information; They represent the cross-layer relationships between the physical layer and the functional layer, the functional layer and the monitoring layer, and the physical layer and the monitoring layer respectively.

[0013] Furthermore, the construction of a multi-level industrial system knowledge graph based on the acquired industrial equipment parameters also includes establishing mapping relationships between the three-layer graphs based on the physical layer, monitoring layer, and functional layer to ensure the integrity and consistency of knowledge representation. Specifically, the importance of physical components to functional realization is quantified through contribution analysis, and mapping weights between physical components and functional units are established; the ability of monitoring parameters to reflect functional status is determined through sensitivity and coverage analysis, and the association weights between functional status and monitoring parameters are established; finally, cross-layer constraints are used to ensure the consistency of overall semantics, which is expressed as: , in, A measure of semantic consistency of the representation graph; Represents a set of paths that satisfy semantic consistency, i.e., a valid path from the physical entity through the functional unit to the monitoring point; Represents the set of all possible paths.

[0014] Furthermore, the adaptive knowledge update of the multi-level industrial system knowledge graph based on the dynamic evolution mechanism of the graph includes establishing a mathematical representation of the graph changing over time, expanding the static graph into a dynamic time series structure, and establishing a mathematical model of the graph evolution to describe the driving factors and change laws of the graph state change; designing a dynamic update mechanism for the node and edge attributes in the graph so that the graph can reflect the system state changes in real time, wherein a dynamic update model for the health status of physical components is established, comprehensively considering monitoring data, historical status and external environmental factors; establishing an integral update model for cumulative attributes such as usage time to accurately record the cumulative operating status of the equipment. For cumulative attributes such as usage time, an integral update model is adopted, which is expressed as: , in, Presentation Component i Deadline t The cumulative running time of Indicates that the component is at time running status.

[0015] Furthermore, the adaptive knowledge update of the multi-level industrial system knowledge graph based on the dynamic evolution mechanism of the graph also includes the design of a dynamic adjustment mechanism for the graph topology structure, including node addition and deletion, edge addition and deletion, and weight update, to adapt to changes in the system structure, wherein the node addition and deletion operation is expressed as: , in, V ( t ) means at time t The set of all nodes in the graph; the addition and deletion of edges is handled by establishing a dynamic management mechanism for relationship edges to handle the establishment and disconnection of connection relationships in the system, which is expressed as: , in, Indicates time t The edge set of represents a new edge set, Indicates the removal of edge sets.

[0016] Furthermore, the joint reasoning framework based on the large model and knowledge graph is constructed, including first designing a three-layer enhancement strategy to improve the industrial domain understanding and reasoning ability of the large model. The first layer is domain knowledge enhancement, which improves the industrial domain understanding ability of the large model by injecting professional knowledge, and adopts efficient parameter fine-tuning technology to avoid the high computational cost of full parameter fine-tuning; the second layer is multimodal information fusion, which enhances the ability of the large model to process multiple data types, and realizes the effective fusion of different modal information through multimodal encoder and attention mechanism; the third layer is graph structure perception, which enhances the large model's ability to understand the knowledge graph structure, encodes the graph structure through graph neural network, and establishes a cross-modal attention mechanism, which is expressed as: , in, For the l The node feature matrix of the layer, is the adjacency matrix with self-loops added, is the corresponding degree matrix, is the learnable weight matrix, is a non-linear activation function.

[0017] Furthermore, the construction of a joint reasoning framework based on a large model and knowledge graph also includes designing a three-level joint reasoning mechanism, combining graph reasoning, statistical reasoning, and large model reasoning to form a multi-level collaborative early warning mechanism. The first level is the graph reasoning layer, which performs fault propagation analysis and impact assessment based on the structure of the knowledge graph and uses graph structure information for logical reasoning; the second level is the statistical reasoning layer, which uses multivariate anomaly detection and degradation trend analysis methods based on monitoring data to perform anomaly detection and trend analysis; the third level reasoning mechanism is the large model reasoning layer, which uses a large language model for deep semantic understanding and complex reasoning, realizes multi-source information integration and chain thinking reasoning, and finally designs a weighted decision fusion mechanism to integrate the results of the three reasoning layers, expressed as: , in, Indicates the final decision, 、 and Represent the decision results of the graph reasoning layer, statistical reasoning layer and large model reasoning layer respectively, is the weight coefficient, satisfying .

[0018] Furthermore, the optimization of the joint reasoning framework based on adaptive knowledge updating and continuous learning mechanism includes establishing a multi-source knowledge acquisition framework, continuously collecting and integrating new knowledge from multiple channels, automatically extracting knowledge patterns from historical monitoring data, and using unsupervised learning methods to mine data regularities and implicit patterns; extracting empirical knowledge from historical failure cases, analyzing recorded failure cases and summarizing failure modes and solutions; using large models to extract structured knowledge from technical documents, and converting unstructured technical information into usable knowledge representations; converting feedback from engineers and maintenance personnel into structured knowledge, integrating manual experience and system automatic learning results, and finally designing a knowledge quality evaluation mechanism, including knowledge consistency evaluation, reliability evaluation and practicality evaluation, through which the accuracy and practicality of knowledge in the system are effectively ensured.

[0019] Furthermore, the optimized joint reasoning framework is used to perform early fault warning for industrial equipment, including the design of a fault warning verification and false alarm optimization mechanism to solve the problems of high false alarm rate and low credibility in traditional warning systems. The accuracy and practicality of warning results are improved through multi-source verification, false alarm control and feedback learning mechanisms. Among them, a multi-source data cross-validation mechanism is used to reduce the misjudgment that may be caused by a single data source, and the credibility of the warning is improved through consistency verification of multiple data sources. The cross-validation score is defined as: , in, Validation_Score Indicates the cross-validation score of the warning (0-1), Indicator i Indicates the i The indication results of the data source, It represents the weight of the i-th data source, reflecting the reliability of the data source.

[0020] The second aspect is an early fault warning system based on the dynamic evolution of complex industrial graphs, including: The data acquisition module is configured to acquire industrial equipment parameters; The knowledge graph module is configured to construct a multi-level industrial system knowledge graph based on the acquired industrial equipment parameters; The updating module is configured to adaptively update the knowledge graph of the multi-level industrial system based on the dynamic evolution mechanism of the graph; The reasoning module is configured to build a joint reasoning framework based on the large model and the knowledge graph; The optimization module is configured to optimize the joint reasoning framework based on adaptive knowledge updating and continuous learning mechanisms; The early warning module is configured to use the optimized joint reasoning framework to provide early warning of industrial equipment failures.

[0021] In a third aspect, the present invention provides a computer-readable storage medium storing a plurality of instructions, wherein the instructions are suitable for being loaded and executed by a processor of a terminal device, for carrying out an early fault warning method based on the dynamic evolution of a complex industrial graph.

[0022] In a fourth aspect, the present invention provides a terminal device comprising a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, and the instructions are suitable for being loaded and executed by the processor to implement the early fault warning method based on the dynamic evolution of complex industrial maps.

[0023] In summary, the present invention has the following beneficial technical effects: This invention, "An Early Fault Warning System Based on Dynamic Evolution of Complex Industrial System Graphs and Joint Reasoning of Large Models," integrates dynamic knowledge graphs and large language models to achieve early, accurate, and explainable industrial fault warnings. The system uses a multi-level knowledge graph to represent the structure of complex industrial systems and, combined with the graph's dynamic evolution mechanism, captures system state changes in real time, significantly improving its modeling capabilities for complex systems. Attention enhancement and sparse pattern learning mechanisms, designed to target subtle early signs of faults, effectively identify fault signals buried in background noise, providing a reliable basis for maintenance decisions.

[0024] This system overcomes many limitations of traditional early warning methods and achieves a deep integration of knowledge and data. The adaptive knowledge update mechanism can continuously and automatically extract new knowledge from multi-source data, enabling the system to adapt to changes in the industrial environment and avoiding the problem of early warning failure caused by outdated knowledge. The large model and graph joint reasoning framework provides highly interpretable early warning results, including fault cause analysis, development trend prediction, and treatment recommendations, greatly enhancing the practical value of the early warning results. Through the organic combination of these innovative technologies, this system provides a comprehensive solution for predictive maintenance of industrial equipment, effectively reducing maintenance costs and extending equipment life, with significant economic benefits and social value. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 Schematic diagram of an early fault warning method based on the dynamic evolution of a complex industrial graph according to Example 1 of the present invention. DETAILED DESCRIPTION

[0026] The present invention will be further described in detail below with reference to the accompanying drawings.

[0027] Example 1 Reference Figure 1 In this embodiment, an early fault warning method based on the dynamic evolution of a complex industrial graph includes: Step S1: Construction of multi-level industrial system knowledge graph: This step aims to construct a multi-layered industrial system knowledge graph, which serves as the foundational knowledge representation framework for the fault warning system. Unlike the traditional single-plane structure, this knowledge graph adopts a three-layer architecture: physical, functional, and monitoring layers. This fully represents the physical structure, functional relationships, and monitoring parameters of the industrial system.

[0028] Step S1.1: Definition of the overall architecture of the graph: The multi-level industrial system knowledge graph can be expressed as: , in, Represents the physical layer subgraph, which includes physical entities such as devices and components and their connection relationships; Represents the functional layer subgraph, describing the functional units of the system and their interactions; Represents the monitoring layer subgraph, which contains sensor and monitoring point information; They represent the cross-layer relationships between the physical layer and the functional layer, the functional layer and the monitoring layer, and the physical layer and the monitoring layer respectively.

[0029] Step S1.1.1: Physical layer subgraph construction: Construct a basic graph that represents the physical structure of the industrial system and the relationship between device components to form a complete representation of the system's physical topology. The physical layer subgraph is defined as: , in, Represents a set of physical entity nodes, such as pumps, valves, bearings and other equipment components; Represents physical relationship edge sets, such as mechanical connections, electrical connections, etc. Represents a set of properties, containing the characteristics of each physical entity.

[0030] Establishment of the hierarchical relationship of physical layer nodes: The hierarchical relationship between physical layer nodes is represented by inclusion edges, forming a multi-level hierarchy of device-component-assembly-element. The probability of fault propagation between physical components can be expressed as: , in, Indicates a faulty slave component i Propagate to components j probability; Indicates the normalized value (0-1) of the physical distance between components. The closer the distance, the larger the value. Indicates the degree of functional dependence (0-1), the stronger the dependence, the larger the value; Indicates the strength of historical fault propagation evidence (0-1); is the weight coefficient, satisfying .

[0031] Step S1.1.2: Functional layer subgraph construction: The functional layer subgraph is an abstract representation of the system functional modules and their interactions, which establishes a function-based system understanding framework. The functional layer subgraph is defined as: , in, Represents a set of functional unit nodes, such as power transmission, signal processing and other functional modules; Represents the functional relationship edge set, describing the interaction between functions; Represents a set of functional attributes, including information such as performance indicators.

[0032] Step S1.1.3: Construction of monitoring layer subgraph: Establish a network representation of monitoring parameters and their relationships to form a complete description of system observability. The monitoring layer subgraph is defined as: , in, Represents a set of monitoring point nodes, including various sensors and monitoring devices; Represents the parameter association edge set, describing the correlation between monitoring parameters; Represents a set of monitoring attributes, including measurement parameters, sampling frequency, and other characteristics. The correlation between parameters can be calculated using the correlation coefficient: , in, Representation parameters i and j Correlation coefficient of Representation parameters i and j covariance of Represents parameters respectively i and j The standard deviation of .

[0033] Step S1.1.3.1, definition of abnormality of monitoring layer subgraph: Establishing a quantitative index for the degree of abnormality of monitoring parameters, comprehensively considering single parameter abnormality and multi-parameter correlation abnormality, provides a basis for early fault sign identification. The abnormality of monitoring parameters can be defined as: , in, Representation parameters i At the moment t degree of abnormality; Display parameters i At the moment t The measured value of and Represents parameters respectively i the historical mean and standard deviation of Representation and parameters i A collection of related parameters; Representation parameters i and j The degree of correlation (0-1); wiw_i and Represents parameters respectively i and related parameters j The weight of .

[0034] Step S1.2: Establish cross-layer relationship mapping: Establish mapping relationships between the three layers of graphs to ensure the integrity and consistency of knowledge representation.

[0035] S1.2.1. Relationship between the physical layer and the functional layer: Establish the mapping weights between physical components and functional units, and quantify the importance of physical components to functional realization through contribution analysis. The relationship between the physical layer and the functional layer can be expressed as: , in, Representing physical components i Functional Unit j The association weight of Representing physical components i Functional Unit j The denominator represents the contribution of all physical components to the functional unit j The sum of the contributions of is used for normalization.

[0036] S1.2.2. The relationship between the functional layer and the monitoring layer: Establish the association weight between functional status and monitoring parameters, and determine the ability of monitoring parameters to reflect functional status through sensitivity and coverage analysis. The association relationship between the functional layer and the monitoring layer can be expressed as: , in, Indicates monitoring parameters j Functional Unit i The association weight of Indicates monitoring parameters j Functional Unit i Sensitivity of state changes (0-1); Indicates monitoring parameters j Functional Unit i coverage (0-1).

[0037] 1.2.3, Semantic consistency constraints: The graph ensures the overall semantic consistency through cross-layer constraints. The formula is as follows: , in, A measure of semantic consistency of the representation graph; Represents a set of paths that satisfy semantic consistency, i.e., a valid path from the physical entity through the functional unit to the monitoring point; Represents the set of all possible paths.

[0038] S2. Design of dynamic graph evolution mechanism: The second key step is to design a dynamic graph evolution mechanism, enabling the knowledge graph to continuously and adaptively update as the state of the industrial system changes. Unlike traditional static knowledge graphs, this system's knowledge graph is a dynamic knowledge base that can capture dynamic changes in the system over time, reflecting the impact of factors such as component aging, changing operating conditions, and system upgrades.

[0039] S2.1. Overall framework of dynamic graph evolution: Design an overall framework for the dynamic evolution of graphs, and expand static graphs into dynamic time-series structures. The dynamic evolution of graphs can be formally represented as a time-series graph sequence: , in, G ( t ) indicates the time t The system graph state, including three layers of subgraphs and their cross-layer relationships, changes continuously with time t; S2.2 Dynamic update mechanism of graph attributes, Design a dynamic update mechanism for node and edge attributes in the graph so that the graph can reflect changes in system status in real time.

[0040] S2.2.1 Node health status update, A dynamic update model for the health status of physical components is established, taking into account monitoring data, historical status, and external environmental factors. The attributes of nodes and edges in the graph are dynamically updated based on real-time data. Taking the health status attributes of physical components as an example, the update model is: , in, Presentation Component i At the moment t Health status index (0-1), is the smoothing factor (0-1), is the health status evaluation function; Representation and Components i relevant monitoring data; Presentation Component i Attributes such as age and maintenance history; Indicates external environmental conditions.

[0041] S2.2.2 Cumulative attribute update model, Establish an integral update model for cumulative attributes such as usage time to accurately record the cumulative operating status of the equipment. For cumulative attributes such as usage time, the integral update model is adopted, and the formula is as follows: , in, Presentation Component i Deadline t The cumulative running time of Indicates that the component is at time The operating status (0 or 1).

[0042] S2.3 Graph topology evolution, Design a dynamic adjustment mechanism for the graph topology, including adding and deleting nodes, adding and deleting edges, and updating weights, to adapt to changes in the system structure.

[0043] S2.3.1 Node addition and deletion operations, Establish a node lifecycle management mechanism to handle the addition and removal of devices and components in the system. The topology of the graph is adjusted as the system changes, mainly including three operations: node addition and deletion, edge addition and deletion, and weight update. The node addition and deletion operation can be expressed as: , in, V ( t ) means at time t The set of all nodes in the graph, which may represent physical devices, functional units or monitoring points in the industrial system. Indicates that at the next moment The node set in the graph of (t+Δt) is the updated node set. Indicates the time interval A collection of new nodes that need to be added to the graph, representing newly added equipment, functions, or monitoring points. Indicates the time interval The set of nodes that need to be removed from the graph represents the equipment, functions or monitoring points that are eliminated or disabled. ∪ is the set union operator, which means that the new node set Add to existing node collection ∖ is the difference operator of the set, which means removing from the merged set Nodes in .

[0044] S2.3.2 Add and delete operations, A dynamic management mechanism for relationship edges is established to handle the establishment and disconnection of connection relationships in the system. The edge addition and deletion operation formula is as follows: , in, Indicates time t The edge set of represents a new edge set, Indicates the removal of edge sets.

[0045] S2.3.3 Dynamic update of relationship weights, An incremental learning model for relationship weights is established to dynamically adjust the relationship strength based on the system operation status. For the dynamic update of relationship weights, an incremental learning model is used, and the formula is as follows: , in, At the moment t The weight of is the learning rate (0-1), is a weight adjustment amount that can be calculated based on correlation changes, fault samples, or expert feedback.

[0046] S2.4 Parameter-related dynamic modeling, A dynamic modeling mechanism for monitoring the correlation between parameters is established to capture the changes in parameter correlation strength with system status.

[0047] S2.4.1 Dynamic Mutual Information Calculation, The correlation strength between parameters in the monitoring layer is not static, but changes dynamically with the system state. Dynamic mutual information can be used to represent time-varying correlations: , in, Representation parameters i and j At the moment t The mutual information of . Indicates at time t Nearby time window and The joint probability distribution of . and Respectively represent the corresponding marginal probability distribution; S2.4.2 Sliding time window mechanism, A sliding time window mechanism is introduced to capture the dynamic changes of parameter associations and adapt to the time-varying characteristics of the system operation state. The formula is as follows: , in, Representation parameters i and j At the moment t correlation. Indicates parameters Indicates parameters i exist[t - W , t ]Data sequence within the time window, W is the window length; S2.5 Fault propagation mode evolution, Establish a dynamic evolution mechanism of fault propagation mode and update the fault propagation probability and propagation time model.

[0048] S2.5.1 Dynamic update of fault propagation probability, A dynamic update model of fault propagation probability is established, and the propagation probability parameters are adjusted according to the latest observation data. The dynamic update model of fault propagation probability is: , in, Indicates a faulty slave component i Propagate to components j The probability of is the update coefficient (0-1); represents the probability of transmission calculated based on the most recent observation data.

[0049] S2.5.2 Dynamic update of fault propagation time, A dynamic update model of fault propagation time is established to reflect the changes in propagation speed with system conditions. The dynamic update model of fault propagation time is: , in, Indicates a faulty slave component i Propagate to components j The estimated time, is the update coefficient (0-1), Represents the propagation time calculated based on the most recent observation data.

[0050] S2.6 System evolution status monitoring, Establish a monitoring mechanism for the overall evolution status of the system and quantify the system evolution trend through multi-dimensional indicators.

[0051] S2.6.1 Definition of graph evolution state vector, In order to capture the overall evolution trend of the system, the graph evolution state vector is introduced. The formula is as follows: , in, The system at time t The evolution state vector of the system contains multiple indicators that reflect changes in different aspects of the system. Each component Si(t) represents a specific evolution indicator. This system designs three key indicators: system stability index, system stability index, and correlation structure change rate.

[0052] S2.6.2 System Stability Index Calculation, A quantitative indicator of system stability is established to measure the stability of the system relationship structure. The formula is as follows: , This metric measures the stability of a system structure. When system relationships (edge ​​weights) change slightly, the stability index is close to 1, indicating a stable system. When system relationships change dramatically, the stability index is close to 0, indicating an unstable system state and possible anomalies or failures.

[0053] S2.6.3 System degradation rate calculation, Calculate the overall degradation rate of the system to reflect the downward trend of system performance. The formula is as follows: , This metric calculates the average rate of decline in the health of all physical components. Positive values ​​indicate system recovery (e.g., after repair), while negative values ​​indicate system degradation. A higher absolute value of the degradation rate indicates faster system performance changes and requires closer monitoring.

[0054] S2.6.4 Calculation of the rate of change of associated structures, The correlation structure change rate is calculated to quantify the speed of the map structure adjustment. The formula is as follows: , This metric measures the rate of change in the graph structure. A high rate of change indicates that system relationships are rapidly adjusting, possibly reflecting a change in operating conditions, a configuration update, or the emergence of an abnormal state.

[0055] Step S3: Construction of a large model and graph joint reasoning framework: This step builds a joint reasoning framework for large models and graphs, combining the powerful reasoning capabilities of large language models with the structured knowledge of industrial knowledge graphs to design an innovative fault warning reasoning mechanism. This framework aims to overcome the limitations of single approaches and achieve a deep fusion of knowledge-driven and data-driven approaches.

[0056] S3.1 Large Model Domain Enhancement Strategy, In order to adapt the general large-scale language model to the industrial fault warning task, this paper designs a three-layer enhancement strategy to improve the industrial domain understanding and reasoning ability of the large model.

[0057] S3.1.1 Domain knowledge enhancement, The first-tier enhancement strategy is domain knowledge enhancement. This improves the large model's ability to understand the industrial domain by injecting specialized knowledge, and employs efficient parameter fine-tuning techniques to avoid the high computational cost of full parameter fine-tuning. The specific formula is as follows: , in, Represents the basic large language model, To represent industrial domain knowledge datasets, knowledge injection uses parameter efficient fine-tuning (PEFT) technology to avoid the high computational cost of full parameter fine-tuning.

[0058] S3.1.2 Enhanced multimodal information fusion, The second-level enhancement strategy is multimodal information fusion, which enhances the ability of large models to handle multiple data types. It achieves effective fusion of different modal information through multimodal encoders and attention mechanisms. The formula is as follows: , in, is a multimodal encoder, encoders for text, time series, numerical, and spectral data, respectively. is a multimodal fusion function. The fusion method uses the attention mechanism: , in, Q is the query vector, and Respectively i The key vector and value vector of the modality.

[0059] S3.1.3 Enhanced graph structure perception, The third-layer enhancement strategy is graph structure perception, which enhances the large model's ability to understand the knowledge graph structure. Graph structure encoding is performed through graph neural networks, and a cross-modal attention mechanism is established. The formula is as follows: , in, For the l The node feature matrix of the layer, is the adjacency matrix with self-loops added, is the corresponding degree matrix, is the learnable weight matrix, is a nonlinear activation function. The graph structure features are fused with the text representation through a cross-modal attention mechanism.

[0060] S3.2 Three-level joint reasoning mechanism, A three-level joint reasoning mechanism is designed, combining the advantages of graph reasoning, statistical reasoning and large model reasoning to form a multi-level collaborative early warning mechanism.

[0061] S3.2.1 Graph Inference Layer, The first-level reasoning mechanism is the graph reasoning layer, which performs fault propagation analysis and impact assessment based on the structure of the knowledge graph and uses graph structure information for logical reasoning. The specific formula is as follows: , in, Indicates a faulty slave node Propagate to nodes The probability of Indicates from arrive The path, represents the propagation probability between adjacent nodes.

[0062] , in, Representation node The impact of the failure, Representation node The importance weight of .

[0063] S3.2.2 Statistical Inference Layer, The second-level reasoning mechanism is the statistical reasoning layer, which performs anomaly detection and trend analysis based on monitoring data. It uses multivariate anomaly detection and degradation trend analysis methods. The specific formula is as follows: , in, Indicates time t Multivariate monitoring data, and represent the mean vector and covariance matrix under normal conditions, respectively. Represents the Mahalanobis distance calculation function.

[0064] , in, Indicates the remaining service life, represents the number of monitoring within the time window [tW,t], Indicates the current health status. Represents the degradation rate.

[0065] S3.2.3 Large Model Inference Layer, The third-level reasoning mechanism is the large-model reasoning layer, which uses a large language model for deep semantic understanding and complex reasoning, and implements multi-source information integration and chain-based reasoning. The specific formula is as follows: , in, Represents the comprehensive context, Indicates the state of the graph. Represents real-time data, Indicates event record, Represents historical cases, Indicates related documents.

[0066] , in, A prompt template that guides chain thinking. Represents the reasoning task description.

[0067] S3.2.4 Weighted Decision Fusion, The system designs a weighted decision fusion mechanism to integrate the results of the three reasoning layers. The formula is as follows: , in, Indicates the final decision, 、 and Represent the decision results of the graph reasoning layer, statistical reasoning layer and large model reasoning layer respectively, is the weight coefficient, satisfying .

[0068] S3.3 Early fault sign capture mechanism, To address the problem of early faults being difficult to detect, this invention has designed a mechanism to capture weak early fault signs. The attention mechanism is used to amplify weak signal features and improve the ability to identify early fault signs. The formula is as follows: , in, Indicates the parameters At the moment t The attention score, For the attention network, Represents the current data context, Indicates the current graph status; S4. Adaptive knowledge updating and continuous learning mechanism, S4.1 Multi-source knowledge acquisition framework, Establish a multi-source knowledge acquisition framework to continuously collect and integrate new knowledge from multiple channels.

[0069] S4.1.1 Monitoring data mining, Automatically extract knowledge patterns from historical monitoring data and use unsupervised learning methods to mine data regularities and implicit patterns. The formula is as follows: , in, Indicates the time period Clustering, Association, and Anomaly represent cluster analysis, association rule mining, and anomaly detection methods, respectively.

[0070] S4.1.2 Case library learning, Case library learning extracts empirical knowledge from historical failure cases. The system analyzes the recorded failure cases and summarizes the failure modes and solutions. The formula is as follows: , in, Experience ( Case ) represents the empirical knowledge extracted from the case, including fault symptoms, root causes, propagation paths, solutions, and effect evaluation.

[0071] S4.1.3 Document Knowledge Extraction, Using a large model to extract structured knowledge from technical documents, unstructured technical information is transformed into usable knowledge representation. The formula is as follows: , in, represents the knowledge extracted from the document, represents the knowledge extraction function of the large model, Represents a predefined knowledge pattern.

[0072] S4.1.4 Integration of expert feedback, Transform feedback from engineers and maintenance personnel into structured knowledge, integrating manual experience and system automatic learning results. The formula is as follows: , in, represents the knowledge extracted from expert feedback, Indicates the original feedback from the experts, Context Indicates relevant context, G ( t ) indicates the current graph status.

[0073] S4.2 Knowledge quality assessment mechanism, Design a knowledge quality assessment mechanism, including knowledge consistency assessment, reliability assessment and practicality assessment, through which the accuracy and practicality of the knowledge in the system can be effectively ensured.

[0074] S4.2.1 Knowledge consistency assessment, Check the consistency between new knowledge and existing knowledge, identify and resolve knowledge conflicts, and ensure the logical consistency of the knowledge system. The formula is as follows: , in, Representing new knowledge With the current map The consistency score, Function identification conflicting knowledge, Represents the subset of the graph that is relevant to the new knowledge.

[0075] S4.2.2 Knowledge reliability assessment, According to the evaluation knowledge reliability of the knowledge source, the source reliability, extraction method reliability and confidence are comprehensively considered. The formula is as follows: , in, represents the knowledge reliability score, Indicates the reliability of the knowledge source, Indicates the reliability of the extraction method, Confidence Indicates confidence.

[0076] S4.2.3 Knowledge practicality assessment, The actual value of existing knowledge for fault warning is evaluated, taking into account potential impact, novelty, and applicability. The formula is as follows:

[0077] in, represents the knowledge practicality score, Indicates potential influence, Indicates novelty, Indicates applicability.

[0078] S4.3 Knowledge Fusion Strategy, Design knowledge fusion strategies to handle heterogeneous knowledge sources and potential conflicts, including incremental knowledge fusion and knowledge conflict detection and resolution.

[0079] S4.3.1 Incremental Knowledge Fusion, Incremental knowledge fusion integrates new knowledge into the existing knowledge graph in an incremental manner to maintain the continuity and stability of the knowledge base. Incremental knowledge fusion integrates new knowledge into the existing knowledge graph in an incremental manner. The formula is as follows: , in, represents the updated graph, Fusion represents the fusion function, Represents a new set of validated knowledge.

[0080] S4.3.2 Knowledge conflict detection and resolution, Identify and resolve knowledge conflicts, and handle contradictions between knowledge through quality assessment, expert arbitration, etc. The formula is as follows: , in, Resolve represents the conflict resolution function, Quality Function to evaluate the quality of knowledge, is the threshold parameter, and Merge represents the merge operation.

[0081] S4.4 Model Adaptive Optimization, Continuously adjust model parameters and structure based on operational feedback, including early warning effect recording, parameter self-optimization, and model structure optimization.

[0082] S4.4.1 Recording and evaluation of early warning effects, The system records and evaluates the effect of each warning, establishes a warning performance evaluation system, and provides a basis for model optimization. The formula is as follows: , in, Indicates time t The early warning performance score, represents the F1 score calculation function, TP 、 FP 、 FN 、 TN They represent the number of true faults, false faults, false faults, and true faults, respectively. A true fault is when the system predicts a fault and it actually occurs. A false fault is when the system predicts a fault but it does not actually occur.

[0083] S4.4.2 Parameter Adaptive Adjustment, The model parameters are automatically adjusted based on the warning effect, and gradient optimization and other methods are used to achieve automatic parameter tuning. The formula is as follows: , in, represents the model parameter set at time t, η is the learning rate, represents the gradient of performance with respect to the parameters.

[0084] S4.4.3 Model structure optimization, Dynamically adjust the model structure based on performance feedback to adapt to different task requirements and changes in system characteristics. The formula is as follows: , in, Indicates time t The model structure, EvolutionRepresents a structural evolution function that adjusts model complexity and focus based on historical performance and task distribution.

[0085] Step S5: Fault warning verification and false alarm optimization: This step designs a fault warning verification and false alarm optimization mechanism to solve the problems of high false alarm rate and low credibility in traditional warning systems. Through mechanisms such as multi-source verification, false alarm control and feedback learning, the accuracy and practicality of warning results are improved.

[0086] S5.1 Multi-source data cross-validation mechanism, A multi-source data cross-validation mechanism is used to reduce the possibility of misjudgment caused by a single data source and improve the credibility of warnings through consistency verification of multiple data sources. The warning trigger threshold is set based on the cross-validation score. When the verification score exceeds the preset threshold, the system triggers a warning. The cross-validation score is defined as: , in, Validation_Score Indicates the cross-validation score of the warning (0-1), Indicator i Indicates the i The result of the data source indication (0 or 1), It represents the weight of the i-th data source, reflecting the reliability of the data source.

[0087] S5.2 Early warning feedback learning mechanism, Establish an early warning feedback learning mechanism to continuously improve the accuracy of early warnings through feedback, achieving self-optimization and improvement of the system. This mechanism can collect multi-dimensional feedback information on early warning effects and automatically adjust model parameters and early warning strategies based on the feedback results, forming a closed-loop optimization system.

[0088] S5.2.1 Feedback information collection, The system has established a comprehensive feedback information collection mechanism, collecting feedback data on early warning effectiveness from multiple dimensions. Feedback information is mainly divided into three categories: early warning accuracy feedback, timeliness feedback, and expert evaluation feedback.

[0089] The early warning accuracy feedback evaluates the correctness of the early warning by comparing the early warning results with the actual fault occurrence, including four situations: true positive (TP, the early warning is correct and the fault occurs), false positive (FP, false alarm), true negative (TN, correct but no warning) and false negative (FN, missed alarm).

[0090] Timeliness feedback evaluates the timeliness of early warnings, analyzes the relationship between the time when early warnings are issued and the time when actual faults occur, and ensures that early warnings are issued within the effective time window.

[0091] Expert evaluation feedback collects subjective evaluations of field experts on warning relevance, severity assessment, root cause analysis, and recommended measures.

[0092] The comprehensive score calculation formula for feedback information is: , in, is the weight coefficient, satisfying ; represents the accuracy score (0-1); Indicates the timeliness score (0-1); Indicates the expert evaluation score (0-1).

[0093] The accuracy score is calculated based on the confusion matrix: , Among them, TP represents the number of true positives; TN represents the number of true negatives; FP represents the number of false positives; and FN represents the number of false negatives.

[0094] S5.2.2 Dynamic optimization of early warning strategies, Optimize early warning strategies based on feedback results, including warning level setting and customized optimization of warning content, to improve the practicality and effectiveness of early warnings.

[0095] Optimize warning level settings. Dynamically adjust warning level classification criteria based on the historical development patterns and actual impact of different fault types. The system analyzes the match between warning levels in historical fault cases and actual fault severity, identifies deviations in level settings, and makes adjustments accordingly.

[0096] Alert content customization and optimization dynamically adjusts the level of detail and presentation of alert information based on the needs of different user roles. The system collects user feedback on alert content, analyzes which information is most helpful for decision-making, and optimizes the structure and content of alert reports accordingly.

[0097] Example 2 This embodiment provides an early fault warning system based on the dynamic evolution of complex industrial graphs, including: The data acquisition module is configured as follows: A computer-readable storage medium stores a plurality of instructions, wherein the instructions are suitable for being loaded and executed by a processor of a terminal device, for example, an early fault warning method based on the dynamic evolution of a complex industrial graph.

[0098] A terminal device includes a processor and a computer-readable storage medium, the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, and the instructions are suitable for being loaded and executed by the processor to implement an early fault warning method based on the dynamic evolution of complex industrial maps.

[0099] The above are all preferred embodiments of the present invention, and are not intended to limit the scope of protection of the present invention. Therefore, any equivalent changes made based on the structure, shape, and principle of the present invention should be included in the scope of protection of the present invention.

Claims

1. An early fault warning method based on the dynamic evolution of complex industrial graphs, characterized by: include: Obtain industrial equipment parameters; Construct a multi-level industrial system knowledge graph based on the acquired industrial equipment parameters; Adaptive knowledge update of multi-level industrial system knowledge graph based on graph dynamic evolution mechanism; Build a joint reasoning framework based on large models and knowledge graphs; Optimize the joint reasoning framework based on adaptive knowledge updating and continuous learning mechanisms; Utilize the optimized joint reasoning framework for early failure warning of industrial equipment.

2. The early fault warning method based on the dynamic evolution of complex industrial graphs according to claim 1 is characterized in that: The multi-level industrial system knowledge graph based on the acquired industrial equipment parameters is constructed, including defining the overall architecture of the graph, constructing a basic graph representing the physical structure of the industrial system and the relationship between equipment components through physical entity node components and parameter characteristics, forming a physical layer subgraph; constructing an abstract representation describing the system functional modules and their interactive relationships through the functional relationship of functional unit nodes, forming a functional layer subgraph; and establishing a network representation of monitoring parameters and their associations through monitoring node data and monitoring devices, forming a monitoring layer subgraph. The multi-level industrial system knowledge graph is represented as follows: ， in, Represents the physical layer subgraph, including the physical entities of devices and components and their connection relationships; Represents the functional layer subgraph, describing the functional units of the system and their interactions; Represents the monitoring layer subgraph, which contains sensor and monitoring point information; They represent the cross-layer relationships between the physical layer and the functional layer, the functional layer and the monitoring layer, and the physical layer and the monitoring layer respectively.

3. The early fault warning method based on the dynamic evolution of complex industrial graphs according to claim 2 is characterized in that: The multi-level industrial system knowledge graph constructed based on the acquired industrial equipment parameters also includes establishing mapping relationships between the three-level graphs based on the physical layer, monitoring layer, and functional layer to ensure the integrity and consistency of knowledge representation. Specifically, the importance of physical components to functional realization is quantified through contribution analysis, and mapping weights between physical components and functional units are established; the ability of monitoring parameters to reflect functional status is determined through sensitivity and coverage analysis, and the association weights between functional status and monitoring parameters are established; finally, cross-layer constraints are used to ensure the consistency of overall semantics, which is expressed as: ， in, A measure of semantic consistency of the representation graph; Represents a set of paths that satisfy semantic consistency, i.e., a valid path from the physical entity through the functional unit to the monitoring point; Represents the set of all possible paths.

4. The early fault warning method based on the dynamic evolution of complex industrial graphs according to claim 3 is characterized in that: The adaptive knowledge update of the multi-level industrial system knowledge graph based on the dynamic evolution mechanism of the graph includes establishing a mathematical representation of the graph's changes over time, expanding the static graph into a dynamic time series structure, and establishing a mathematical model of the graph evolution to describe the driving factors and change laws of the graph state changes; designing a dynamic update mechanism for the node and edge attributes in the graph so that the graph can reflect the system state changes in real time. Among them, a dynamic update model for the health status of physical components is established, which comprehensively considers monitoring data, historical status and external environmental factors; and an integral update model for the cumulative usage time attribute is established to accurately record the cumulative operating status of the equipment. For the cumulative usage time attribute, the integral update model is adopted, which is expressed as: ， in, Presentation Component i Deadline t The cumulative running time of Indicates that the component is at time running status.

5. The early fault warning method based on the dynamic evolution of complex industrial graphs according to claim 4 is characterized in that: The adaptive knowledge update of the multi-level industrial system knowledge graph based on the dynamic evolution mechanism of the graph also includes the design of a dynamic adjustment mechanism for the graph topology structure, including node addition and deletion, edge addition and deletion, and weight update, to adapt to changes in the system structure. The node addition and deletion operation is expressed as: ， in, V ( t ) means at time t The set of all nodes in the graph; the addition and deletion of edges is handled by establishing a dynamic management mechanism for relationship edges to handle the establishment and disconnection of connection relationships in the system, which is expressed as: ， in, Indicates time t The edge set of represents a new edge set, Represents the removal of edge sets.

6. The early fault warning method based on the dynamic evolution of complex industrial graphs according to claim 5 is characterized in that: The joint reasoning framework based on the large model and knowledge graph is constructed. First, a three-layer enhancement strategy is designed to improve the industrial domain understanding and reasoning capabilities of the large model. The first layer is domain knowledge enhancement, which improves the industrial domain understanding capabilities of the large model by injecting professional knowledge. The efficient parameter fine-tuning technology is used to avoid the high computational cost of full parameter fine-tuning. The second layer is multimodal information fusion, which enhances the large model's ability to process multiple data types and achieves effective fusion of different modal information through multimodal encoders and attention mechanisms. The third layer is graph structure perception, which enhances the large model's ability to understand the knowledge graph structure. Graph structure encoding is performed through graph neural networks, and a cross-modal attention mechanism is established, which can be expressed as: ， in, For the l The node feature matrix of the layer, is the adjacency matrix with self-loops added, is the corresponding degree matrix, is the learnable weight matrix, is a non-linear activation function.

7. The early fault warning method based on the dynamic evolution of complex industrial graphs according to claim 6 is characterized in that: The joint reasoning framework based on the large model and knowledge graph is constructed, and further includes designing a three-level joint reasoning mechanism, combining graph reasoning, statistical reasoning, and large model reasoning to form a multi-level collaborative early warning mechanism. The first level is the graph reasoning layer, which performs fault propagation analysis and impact assessment based on the structure of the knowledge graph and uses graph structure information for logical reasoning; the second level is the statistical reasoning layer, which uses multivariate anomaly detection and degradation trend analysis methods based on monitoring data for anomaly detection and trend analysis; the third level reasoning mechanism is the large model reasoning layer, which uses a large language model for deep semantic understanding and complex reasoning, realizing multi-source information integration and chain thinking reasoning. Finally, a weighted decision fusion mechanism is designed to integrate the results of the three reasoning layers, expressed as: ， in, Indicates the final decision, 、 and Represent the decision results of the graph reasoning layer, statistical reasoning layer and large model reasoning layer respectively, is the weight coefficient, satisfying .

8. The early fault warning method based on the dynamic evolution of complex industrial graphs according to claim 7 is characterized in that: The optimization of the joint reasoning framework based on adaptive knowledge updating and continuous learning mechanisms includes establishing a multi-source knowledge acquisition framework, continuously collecting and integrating new knowledge from multiple channels, automatically extracting knowledge patterns from historical monitoring data, and using unsupervised learning methods to mine data regularities and implicit patterns; extracting empirical knowledge from historical failure cases, analyzing recorded failure cases and summarizing failure modes and solutions; and using large models to extract structured knowledge from technical documents and transform unstructured technical information into usable knowledge representations. The feedback from engineers and maintenance personnel is converted into structured knowledge, integrating human experience and system automatic learning results. Finally, a knowledge quality assessment mechanism is designed, including knowledge consistency assessment, reliability assessment and practicality assessment. Through this mechanism, the accuracy and practicality of the knowledge in the system are effectively ensured.

9. The early fault warning method based on the dynamic evolution of complex industrial graphs according to claim 8 is characterized in that: The optimized joint reasoning framework is used to perform early fault warning for industrial equipment, including the design of fault warning verification and false alarm optimization mechanisms to address the high false alarm rate and low credibility issues in traditional warning systems. The accuracy and practicality of warning results are improved through multi-source verification, false alarm control, and feedback learning mechanisms. A multi-source data cross-validation mechanism is used to reduce possible misjudgments caused by a single data source, and consistency checks on multiple data sources are used to improve warning credibility. The cross-validation score is defined as: ， in, Validation_Score Indicates the cross-validation score of the warning (0-1), Indicator i Indicates the i The indication results of the data source, It represents the weight of the i-th data source, reflecting the reliability of the data source.

10. An early fault warning system based on the dynamic evolution of complex industrial graphs, characterized by: include: The data acquisition module is configured to acquire industrial equipment parameters; The knowledge graph module is configured to construct a multi-level industrial system knowledge graph based on the acquired industrial equipment parameters; The updating module is configured to adaptively update the knowledge graph of the multi-level industrial system based on the dynamic evolution mechanism of the graph; The reasoning module is configured to build a joint reasoning framework based on the large model and the knowledge graph; The optimization module is configured to optimize the joint reasoning framework based on adaptive knowledge updating and continuous learning mechanisms; The early warning module is configured to use the optimized joint reasoning framework to provide early warning of industrial equipment failures.

Images