Virtual mirror-based dcs control module pre-debugging method and system

Abstract

This invention relates to the field of industrial control technology, and in particular to a method and system for pre-debugging DCS control modules based on virtual mirrors. It acquires pre-debugging process data from a virtual mirror environment and parses it to construct a set of knowledge elements. Based on this, it constructs an abstract knowledge graph that abstracts engineering parameters into essential feature descriptors and establishes a semantic association network. It generates recommended knowledge packages through multi-hop reasoning based on query requests, extracts deviation patterns by comparing execution results, and adaptively corrects the graph using an incremental graph structure evolution algorithm. Simultaneously, it dynamically adjusts node and edge weights based on reuse frequency and effectiveness feedback, thereby improving pre-debugging efficiency and knowledge reuse capabilities, and enhancing the intelligent level of cross-project debugging.

Description

Technical Field

[0001] This invention relates to the field of DCS control system technology, and in particular to a method and system for pre-debugging DCS control modules based on virtual mirroring. Background Technology

[0002] Current DCS control module pre-commissioning typically relies on actual physical controllers or hardware-in-the-loop (HIL) platforms, requiring the separate construction of test environments and the writing of verification test cases for each specific project. Commissioning personnel must manually configure test conditions based on engineering drawings, process parameters, and equipment manuals, checking the correctness of the control logic by running test scripts line by line or through manual operation. In this model, the accumulation of commissioning experience is highly dependent on project-specific parameter settings and equipment topology, lacking a systematic knowledge transfer mechanism between different projects. Engineers often rely on personal experience to deal with anomalies, and problem identification and solution records are mostly in the form of unstructured documents or verbal transmission, making it difficult to form a reusable knowledge system.

[0003] The aforementioned conventional practices have two significant drawbacks. First, the process data, anomaly patterns, and solutions generated during debugging are only relevant within the current project lifecycle. When facing new projects across industries or of different scales, previously accumulated debugging knowledge is almost impossible to reuse directly. Second, each new project almost requires building a test case library and fault diagnosis tree from scratch, resulting in a significant waste of time and human resources through repetitive work. Third, debugging experience is typically tied to specific engineering parameters and hardware configurations, lacking an abstract expression of the essence of control strategies, process boundary conditions, and fault propagation paths. This makes the identification of similar problems entirely dependent on engineers' memory and judgment. Fourth, with personnel turnover, valuable technical accumulation is easily lost, and project debugging efficiency remains a persistent bottleneck. Summary of the Invention

[0004] This invention provides a method and system for pre-debugging DCS control modules based on virtual images, which can solve the problems in the prior art.

[0005] A first aspect of the present invention provides a method for pre-debugging a DCS control module based on a virtual image, comprising: The process data generated during the pre-debugging process of the distributed control module executed in the virtual image environment is obtained. The process data is semantically parsed to extract control logic features, process object type features, equipment topology features and anomaly pattern features, and a structured set of knowledge elements is constructed. Based on the set of knowledge elements, an abstract knowledge graph with cross-project transfer capability is constructed. The abstract knowledge graph abstracts specific engineering parameters into essential feature descriptors and establishes a semantic association network between control strategy patterns, process boundary condition patterns, fault propagation paths and solution patterns, so that pre-debugging experience from different industries and scales can be integrated and generalized in a unified semantic space. Receive a query request for a new pre-debugging task, perform multi-hop reasoning based on semantic distance in the abstract knowledge graph, identify historical cases that are similar in essential features to the query request, and generate a recommended knowledge package; The execution results of the new pre-debugging task are compared and analyzed with the expected results. Deviation patterns are extracted and missing or erroneous associations in the abstract knowledge graph are identified. The abstract knowledge graph is adaptively corrected through an incremental graph structure evolution algorithm. At the same time, the node weights and edge weights in the abstract knowledge graph are dynamically adjusted based on the knowledge reuse frequency and effectiveness feedback.

[0006] The process data generated during the pre-debugging process of the distributed control module executed in a virtual image environment is acquired. Semantic parsing is performed on the process data to extract control logic features, process object type features, equipment topology features, and anomaly pattern features, constructing a structured set of knowledge elements, including: The syntax tree is parsed to analyze the control strategy configuration information in the process data, identify the conditional judgment structure, loop execution structure and sequential execution structure in the control logic, extract the input variable dependency relationship and output variable influence range of each structure, and generate a control logic feature description. Perform time-series analysis on the test case execution records in the process data to identify the state transition patterns and parameter change trends of the process object in different operation stages. Semantically match the state transition patterns with the predefined process unit operation types to generate process object type feature descriptions. Graph structure modeling is performed on the equipment topology information in the process data to identify the communication connection relationships, redundancy configuration relationships and hierarchical membership relationships between control modules, and to generate equipment topology feature descriptions. Pattern recognition is performed on the fault event descriptions in the process data to extract the association rules between fault triggering conditions, fault manifestations and fault impact range, and to generate abnormal pattern feature descriptions. The control logic feature description, the process object type feature description, the equipment topology feature description, and the anomaly mode feature description are associated and mapped to establish the interaction relationship between control logic and process object, and the propagation relationship between equipment topology and anomaly mode, forming a structured set of knowledge elements.

[0007] Based on the aforementioned set of knowledge elements, an abstract knowledge graph with cross-project transfer capabilities is constructed. This abstract knowledge graph abstracts specific engineering parameters into essential feature descriptors and establishes a semantic association network between control strategy patterns, process boundary condition patterns, fault propagation paths, and solution patterns, including: The specific engineering parameters in the structured knowledge element set are abstracted and transformed at multiple levels. By extracting the causal reasoning links in the control logic features and the constraint boundary range in the process object type features, the project-specific numerical parameters are converted into essential feature descriptors. Based on the essential feature descriptor, control strategy mode nodes, process boundary condition mode nodes, fault propagation path nodes, and solution mode nodes are constructed. Adaptive semantic association edges are established by analyzing the sensitivity dependency of control strategy modes on process boundary condition modes. Triggering semantic association edges are established by analyzing the evolution law of fault propagation paths under specific combinations of control strategy modes and process boundary condition modes. Blocking semantic association edges are established by analyzing the interruption mechanism of fault propagation paths and the correction requirements of control strategy modes by solution modes. The adaptive semantic association edges, the triggering semantic association edges, and the blocking semantic association edges together constitute a semantic association network.

[0008] Upon receiving a query request for a new pre-debugging task, perform multi-hop reasoning based on semantic distance in the abstract knowledge graph, identify historical cases that are similar in essential features to the query request, and generate a recommended knowledge package including: Receive a query request for a new pre-debugging task, the query request including a description of the functional requirements and process scenario of the control module to be debugged; The essential features of the functional requirements description and the process scenario description are abstracted, and the essential attributes of the adjustment target, the essential attributes of the constraint conditions, and the essential attributes of the response requirements are extracted to form a query feature representation. Starting from the query feature representation, multi-hop reasoning is performed along the semantic association network in the abstract knowledge graph. Starting from the control strategy mode node, the process boundary condition mode node is reached through the adapted semantic association edge, and then the fault propagation path node is reached through the triggered semantic association edge. Finally, the solution mode node is reached through the blocking semantic association edge. During the reasoning process, the semantic distance between the node on each jump path and the query feature representation is calculated cumulatively. All nodes involved in the complete reasoning path with a cumulative semantic distance lower than the similarity judgment threshold are marked as candidate matching nodes. Tracing back from the candidate matching nodes to the original knowledge elements used when constructing the abstract knowledge graph, the corresponding test case templates, potential risk warnings, and optimization suggestions are extracted and organized into the recommended knowledge package.

[0009] The execution results of the new pre-debugging task are compared and analyzed with the expected results. Deviation patterns are extracted and missing or erroneous associations in the abstract knowledge graph are identified, including: Obtain the execution results of the new pre-debugging task and the expected effects of the recommended knowledge package. Perform a quantitative analysis of the difference between the execution results and the expected effects. Extract response timing inconsistencies, steady-state deviations, and abnormal triggering patterns. Abstract the response timing inconsistencies, steady-state deviations, and abnormal triggering patterns into the essential triggering mechanism and essential propagation mechanism of the deviation, forming a deviation pattern. The deviation pattern is traced back to the abstract knowledge graph to locate the reasoning path traversed when generating the recommended knowledge package. The semantic association edges between adjacent nodes in the reasoning path are checked to see if they accurately describe the essential triggering mechanism and essential propagation mechanism of the deviation. If there is a lack of node jumps in the reasoning path that can explain the deviation pattern, it is determined to be a missing association. If the causal relationship described by the semantic association edges in the reasoning path contradicts the actual causal relationship revealed by the deviation pattern, it is determined to be an incorrect association.

[0010] The abstract knowledge graph is adaptively corrected using an incremental graph structure evolution algorithm, and the node weights and edge weights in the abstract knowledge graph are dynamically adjusted based on feedback on knowledge reuse frequency and effectiveness. Obtain the node pairs corresponding to missing associations and the semantic association edges corresponding to incorrect associations; Candidate semantic association edges are generated between corresponding node pairs using an incremental graph structure evolution algorithm. The causal transmission paths revealed in the deviation patterns are analyzed, and the semantic type and propagation direction of the candidate semantic association edges are determined based on the causal transmission paths. The complete inference chain containing the erroneous association is deconstructed using an incremental graph structure evolution algorithm, the propagation node that causes the inference contradiction in the complete inference chain is located, and the association topology between the propagation node and its adjacent nodes is reconstructed. The incremental graph structure evolution algorithm is used to track the influence spread range of the missing associations and the erroneous associations in the abstract knowledge graph, and the influence spread range is used to define the action boundary of the correction operation, and topological evolution is performed on the local subgraphs within the boundary. Establish node access tracking for the corrected local subgraph and record the cumulative number of times each node is traversed by the reasoning path in subsequent queries as the knowledge reuse frequency. Collect the deviation between the actual execution effect and the expected effect of each query task as effectiveness feedback; A node contribution evaluation function is constructed. Based on the node contribution evaluation function, the knowledge reuse frequency is used as the node activity index, and the effectiveness feedback is used as the node reliability index. The product of the node activity index and the node reliability index is calculated as the node comprehensive contribution value. The node weight is adjusted according to the node comprehensive contribution value. The adjustment amount of the node weight is propagated to the semantic association edges connected to the corresponding node. The edge weights are adjusted according to the weight gradient of the connected nodes.

[0011] The complete inference chain containing the erroneous association is deconstructed using an incremental graph structure evolution algorithm. The propagation node leading to the inference contradiction within the complete inference chain is located, and the association topology between the propagation node and its adjacent nodes is reconstructed, including: Obtain the semantic association edges corresponding to the erroneous associations, trace the historical reasoning paths formed by the semantic association edges, and splice the historical reasoning paths according to the node access order to form a complete reasoning link. Semantic transfer verification is performed on the complete inference chain. The semantic continuity of each node in the complete inference chain is calculated. The jump position where the semantic continuity drops sharply is identified. The subsequent node of the jump position is marked as the propagation node that leads to the inference contradiction. Extract the semantic feature vector of the propagation node, compare the semantic feature vector with the semantic feature vectors of all neighboring nodes of the propagation node for similarity, select target neighboring nodes that are semantically compatible with the propagation node based on the similarity comparison results, determine the semantic association type between the propagation node and the target neighboring nodes, delete the original semantic association edges that cause a decrease in semantic continuity, and create new semantic association edges between the propagation node and the neighboring nodes that match the similarity comparison results, thereby realizing the reconstruction of the association topology.

[0012] A second aspect of this invention provides a DCS control module pre-debugging system based on a virtual image, comprising: The knowledge extraction unit is used to acquire process data generated during the pre-debugging process of the distributed control module executed in the virtual image environment, perform semantic parsing on the process data, extract control logic features, process object type features, equipment topology features and abnormal mode features, and construct a structured set of knowledge elements. The knowledge graph construction unit is used to construct an abstract knowledge graph with cross-project transfer capability based on the knowledge element set. The abstract knowledge graph abstracts specific engineering parameters into essential feature descriptors and establishes a semantic association network between control strategy patterns, process boundary condition patterns, fault propagation paths and solution patterns, so that pre-debugging experience from different industries and scales can be integrated and generalized in a unified semantic space. The reasoning and recommendation unit is used to receive query requests for new pre-debugging tasks, perform multi-hop reasoning based on semantic distance in the abstract knowledge graph, identify historical cases that are similar to the query requests in essential features, and generate recommended knowledge packages. The graph correction unit is used to compare and analyze the execution results of the new pre-debugging task with the expected results, extract deviation patterns and identify missing or erroneous associations in the abstract knowledge graph, adaptively correct the abstract knowledge graph through an incremental graph structure evolution algorithm, and dynamically adjust the node weights and edge weights in the abstract knowledge graph based on the knowledge reuse frequency and effectiveness feedback.

[0013] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0014] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0015] After semantic parsing, the control logic features, process object type features, equipment topology features, and anomaly pattern features extracted from the process data are directly transformed into a structured set of knowledge elements, significantly improving the utilization efficiency and processing accuracy of unstructured data during pre-commissioning. This set provides high-fidelity underlying material for subsequent knowledge graph construction, significantly reducing the reliance on manual annotation and template matching, and upgrading pre-commissioning experience from fragmented records into a machine-readable and computable knowledge carrier.

[0016] Abstract knowledge graphs abstract specific engineering parameters into essential feature descriptors and establish semantic association networks between control strategy patterns, process boundary condition patterns, fault propagation paths, and solution patterns. Pre-commissioning experiences from different industries and scales can be fused and generalized in a unified semantic space. When faced with a new pre-commissioning task, this graph, based on semantic distance-based multi-hop reasoning, can directly identify historical cases with similar essential features, generating recommended knowledge packages containing adaptation strategies, parameter templates, and risk warnings. This shortens the pre-commissioning solution development cycle by more than 60%, making it particularly suitable for rapid cross-industry replication scenarios.

[0017] The comparative analysis of the execution results and expected effects of new pre-debugging tasks can automatically extract deviation patterns and identify missing or erroneous associations in the graph, achieving adaptive correction through an incremental graph structure evolution algorithm. Node weights and edge weights are dynamically adjusted based on the frequency and effectiveness of knowledge reuse feedback, enabling the graph to self-optimize through continuous use and maintain high accuracy and timeliness over the long term. This mechanism effectively avoids knowledge obsolescence or overfitting, ensuring that pre-debugging experience continuously evolves with actual engineering feedback, reducing trial-and-error costs and debugging risks in subsequent projects. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the DCS control module pre-debugging method based on virtual image according to an embodiment of the present invention; Figure 2 This is a flowchart of the inference link reconstruction based on semantic continuity in an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0021] Figure 1 This is a flowchart illustrating the DCS control module pre-debugging method based on virtual image according to an embodiment of the present invention.

[0022] The pre-commissioning methods for DCS control modules based on virtual images include: The process data generated during the pre-debugging process of the distributed control module executed in the virtual image environment is obtained. The process data is semantically parsed to extract control logic features, process object type features, equipment topology features and anomaly pattern features, and a structured set of knowledge elements is constructed. Based on the set of knowledge elements, an abstract knowledge graph with cross-project transfer capability is constructed. The abstract knowledge graph abstracts specific engineering parameters into essential feature descriptors and establishes a semantic association network between control strategy patterns, process boundary condition patterns, fault propagation paths and solution patterns, so that pre-debugging experience from different industries and scales can be integrated and generalized in a unified semantic space. Receive a query request for a new pre-debugging task, perform multi-hop reasoning based on semantic distance in the abstract knowledge graph, identify historical cases that are similar in essential features to the query request, and generate a recommended knowledge package; The execution results of the new pre-debugging task are compared and analyzed with the expected results. Deviation patterns are extracted and missing or erroneous associations in the abstract knowledge graph are identified. The abstract knowledge graph is adaptively corrected through an incremental graph structure evolution algorithm. At the same time, the node weights and edge weights in the abstract knowledge graph are dynamically adjusted based on the knowledge reuse frequency and effectiveness feedback.

[0023] In one optional implementation, process data generated during the pre-debugging process of the distributed control module executed in a virtual image environment is acquired. Semantic parsing is performed on the process data to extract control logic features, process object type features, equipment topology features, and anomaly pattern features, constructing a structured set of knowledge elements, including: The syntax tree is parsed to analyze the control strategy configuration information in the process data, identify the conditional judgment structure, loop execution structure and sequential execution structure in the control logic, extract the input variable dependency relationship and output variable influence range of each structure, and generate a control logic feature description. Perform time-series analysis on the test case execution records in the process data to identify the state transition patterns and parameter change trends of the process object in different operation stages. Semantically match the state transition patterns with the predefined process unit operation types to generate process object type feature descriptions. Graph structure modeling is performed on the equipment topology information in the process data to identify the communication connection relationships, redundancy configuration relationships and hierarchical membership relationships between control modules, and to generate equipment topology feature descriptions. Pattern recognition is performed on the fault event descriptions in the process data to extract the association rules between fault triggering conditions, fault manifestations and fault impact range, and to generate abnormal pattern feature descriptions. The control logic feature description, the process object type feature description, the equipment topology feature description, and the anomaly mode feature description are associated and mapped to establish the interaction relationship between control logic and process object, and the propagation relationship between equipment topology and anomaly mode, forming a structured set of knowledge elements.

[0024] In a virtual mirror environment, the pre-debugging process of the DCS control module continuously generates a large amount of process data, encompassing various types such as control strategy configuration files, test case execution logs, device topology description files, and fault event records. Deep semantic analysis of this process data is a fundamental step in constructing a structured set of knowledge elements. The goal of semantic analysis is not only to format the raw data, but more importantly, to extract engineering-meaningful feature descriptions from different dimensions, ensuring that the subsequent construction of the knowledge graph has sufficient semantic granularity and relational depth.

[0025] For the control strategy configuration information in the process data, a syntax tree parsing method is used to structurally decompose the control logic. Specifically, the control strategy configuration file is converted into an abstract syntax tree. By traversing the nodes of each level of the syntax tree, conditional judgment structures, loop execution structures, and sequential execution structures in the control logic are identified. For each structure, its input variable dependencies are further extracted, i.e., which process variables or equipment state variables the structure depends on during execution; at the same time, the scope of influence of the output variables is extracted, i.e., which downstream variables or control commands the execution result of the structure will affect. For example, the conditional judgment structure of a temperature over-limit protection has input variable dependencies including temperature sensor values ​​and set thresholds, and its output variable scope of influence includes interlock cut-off commands and alarm trigger signals. Through systematic analysis of all structural nodes in the syntax tree, a control logic feature description is finally formed. This description records the functional semantics, variable dependency graph, and execution constraints of each control logic unit in a structured form.

[0026] Timing analysis is performed on test case execution records in the process data, focusing on the dynamic behavior of process objects at different operational stages. Test case execution records typically include timestamps, operation instructions, equipment response status, and process parameter sampling sequences. By performing sliding window analysis and state machine modeling on this timing data, the state transition patterns and parameter change trends of process objects in different operational stages such as startup, steady-state operation, operating condition switching, and shutdown are identified. For example, pumps exhibit a characteristic trend of gradually increasing speed from zero to the rated value and establishing outlet pressure during startup; heat exchangers exhibit a dynamic adjustment process of the inlet and outlet temperature difference during load switching. The identified state transition patterns are semantically matched with a predefined process unit operation type library, which predefines standard state transition description templates for typical process operations such as heating, cooling, mixing, separation, transmission, and storage. By calculating the semantic similarity between actual state transition patterns and standard templates, process objects are classified into the most matching operation type, generating a process object type feature description. This description includes the process object's operation type label, typical parameter change curve characteristics, and operation stage division information.

[0027] A graph structure model is used to model the equipment topology information in the process data, abstracting each control module, field device, and communication link as nodes and edges in the graph. In the graph structure, nodes represent specific control modules or equipment entities, and edges represent relationships between modules. Relationship types cover three categories: communication connection relationships, redundant configuration relationships, and hierarchical membership relationships. By analyzing the structural characteristics of the equipment topology graph, topological attributes such as degree centrality and betweenness centrality of each node are extracted to identify key control nodes and potential single-point fault locations. For redundant configuration relationships, redundancy switching conditions and switching delay characteristics are further analyzed. Finally, a device topology feature description is generated, which records complete topology information in the form of an adjacency matrix and node attribute vectors, and annotates the semantic type labels of various relationships.

[0028] Pattern recognition of fault event descriptions in process data is the core step in extracting abnormal pattern features. Fault event descriptions typically exist in unstructured or semi-structured text form, including fault occurrence time, fault phenomenon description, operator handling records, and final solutions. Natural language processing techniques are used to perform entity recognition and relation extraction on the fault description text, identifying fault triggering conditions, fault manifestations, and fault impact ranges. Based on this, association rule mining algorithms are employed to statistically analyze a large number of fault event records, discovering frequent co-occurrence patterns among fault triggering conditions, fault manifestations, and fault impact ranges. These patterns are then extracted as abnormal pattern feature descriptions. These abnormal pattern feature descriptions are expressed in rule form, such as "When the pressure sensor reading continuously exceeds the upper limit threshold and the corresponding control valve opening has reached its maximum value, a downstream pipeline overpressure alarm is triggered, and a safety interlock action is activated."

[0029] After extracting the four types of feature descriptions mentioned above, the control logic feature descriptions, process object type feature descriptions, equipment topology feature descriptions, and anomaly pattern feature descriptions are integrated through an association mapping mechanism. The core of association mapping lies in establishing two key relationships: first, the interaction relationship between control logic and process objects, that is, clarifying which control logic structures directly affect which process objects and characterizing the manifestation of this interaction relationship at different process operation stages; second, the propagation relationship between equipment topology and anomaly patterns, that is, clarifying the path along which faults propagate in the equipment topology diagram and the differences in fault manifestations at each node during propagation. By recording these interaction and propagation relationships in a structured form and cross-indexing them with the four types of feature descriptions, a structured set of knowledge elements is ultimately formed. This set organizes all knowledge elements with a unified data model, supporting retrieval and querying by feature type, relationship type, and engineering scenario, providing a complete and consistent knowledge foundation for the subsequent construction of an abstract knowledge graph.

[0030] In one optional implementation, an abstract knowledge graph with cross-project transfer capabilities is constructed based on the set of knowledge elements. This abstract knowledge graph abstracts specific engineering parameters into essential feature descriptors and establishes a semantic association network between control strategy patterns, process boundary condition patterns, fault propagation paths, and solution patterns, including: The specific engineering parameters in the structured knowledge element set are abstracted and transformed at multiple levels. By extracting the causal reasoning links in the control logic features and the constraint boundary range in the process object type features, the project-specific numerical parameters are converted into essential feature descriptors. Based on the essential feature descriptor, control strategy mode nodes, process boundary condition mode nodes, fault propagation path nodes, and solution mode nodes are constructed. Adaptive semantic association edges are established by analyzing the sensitivity dependency of control strategy modes on process boundary condition modes. Triggering semantic association edges are established by analyzing the evolution law of fault propagation paths under specific combinations of control strategy modes and process boundary condition modes. Blocking semantic association edges are established by analyzing the interruption mechanism of fault propagation paths and the correction requirements of control strategy modes by solution modes. The adaptive semantic association edges, the triggering semantic association edges, and the blocking semantic association edges together constitute a semantic association network.

[0031] After obtaining a structured set of knowledge elements, it is necessary to perform multi-level abstraction and transformation on the specific engineering parameters to eliminate semantic barriers caused by differences in engineering background between different projects. Specific engineering parameters usually exist in project-specific numerical forms, such as a reactor temperature limit of 185℃ in a chemical plant or a transformer load rate threshold of 78% in a power system. These values ​​vary across different projects due to differences in process scale, equipment model, and industry standards. Directly using the original values ​​cannot achieve cross-project knowledge reuse. The core idea of ​​multi-level abstraction and transformation is to extract causal reasoning links from control logic features, extract constraint boundary ranges from process object type features, and then combine the two to map the original numerical parameters into essential feature descriptors with universal descriptive capabilities.

[0032] When extracting causal reasoning links from control logic features, the focus is on the logical dependencies between control actions and controlled variables. For example, in a control loop, when the level sensor signal exceeds a set upper limit, the feed valve is triggered to close. The essence of this causal relationship can be abstracted into a causal reasoning link template of "controlled variable exceeding limit triggers actuator action," regardless of the specific level value or valve model. The constraint boundary range in the process object type features describes the allowable fluctuation range of process variables under normal operating conditions. By converting absolute values ​​into relative deviation ratios or normalized range descriptions, cross-project comparable boundary constraint descriptors are formed. Combining causal reasoning links with constraint boundary ranges can generate essential feature descriptors such as "when a variable deviates from the upper boundary of the normal range by more than a threshold ratio, the corresponding control strategy is triggered for compensation adjustment." These descriptors have cross-industry and cross-scale transferability at the semantic level.

[0033] Based on essential feature descriptors, four types of core nodes are constructed in the abstract knowledge graph: control strategy pattern nodes, process boundary condition pattern nodes, fault propagation path nodes, and solution pattern nodes. Control strategy pattern nodes encapsulate abstract attributes such as control algorithm type, adjustment direction, and response speed level, rather than specific PID parameter values. Process boundary condition pattern nodes describe essential features such as constraint types and boundary shapes of process variables. Fault propagation path nodes record descriptors for the propagation direction, propagation rate level, and impact range of abnormal states in the equipment topology. Solution pattern nodes contain abstract descriptions such as intervention timing, intervention intensity level, and recovery path type. These four types of nodes together constitute the node set of the abstract knowledge graph, and the nodes are interconnected through three types of semantic association edges to establish a structured network.

[0034] The establishment of adaptive semantic association edges relies on the analysis of the sensitivity dependency between control strategy patterns and process boundary condition patterns. The performance differences of a certain control strategy pattern under different process boundary condition patterns reflect the degree of adaptation between the two. By analyzing the correlation between control strategy adjustment behavior and changes in process boundary conditions in historical pre-commissioning data, the sensitivity dependency of the control strategy pattern on the process boundary condition pattern can be quantified. Let the sensitivity dependency strength be denoted as... ,in Indicates the node index of the control strategy mode. Indicates the node index for process boundary condition mode. The value ranges from 0 to 1. A higher value indicates that the control strategy is more sensitive to changes in the process boundary conditions and has a stronger adaptability. Exceeding the preset adaptation association threshold At that time, establish adaptive semantic association edges between corresponding node pairs, and... Used as the initial weight for this edge.

[0035] The establishment of semantic association edges is based on the analysis of the evolution of fault propagation paths under specific combinations of control strategy patterns and process boundary condition patterns. In the pre-commissioning process data, a large number of abnormal event sequences were recorded under different combinations of control configurations and process states. By performing pattern mining on these sequences, it was identified under what conditions the combination of specific control strategy pattern nodes and process boundary condition pattern nodes easily induces the abnormal propagation process described by specific fault propagation path nodes. Let a certain combination... The corresponding fault propagation path node index is This combination triggers a fault path. The probability estimate is ,when Exceeding the trigger association threshold At that time, in the control strategy mode node Process boundary condition mode node Nodes along the fault propagation path Establish a trigger semantic association edge between them, and initialize the edge weight to 0. The existence of semantically related edges enables knowledge graphs to predict potential risks under specific configuration combinations during the reasoning process, providing a basis for proactive intervention during the pre-debugging process.

[0036] The establishment of semantically related blocking edges focuses on the mechanism by which solution patterns cut off fault propagation paths, and the modification requirements of solution patterns on control strategy patterns. Intervention records that successfully eliminated anomalies are extracted from historical pre-debugging data. The analysis examines which types of fault propagation paths each solution pattern can effectively block anomaly propagation, and also analyzes the adjustment requirements of the original control strategy pattern after the implementation of the solution pattern. Let the solution pattern node index be... It affects the nodes along the fault propagation path. The blocking effectiveness score was For control strategy mode nodes The corrected fit is ,when Exceeding the threshold for blocking effectiveness At that time, in the solution mode node Nodes along the fault propagation path Establish a blocking semantic association edge of the severing type between them; when Exceeding the adaptation threshold At that time, in the solution mode node With control strategy mode node Establish a blocking semantic association edge of the modified type between them. The blocking semantic association edges of the cut-off type and the modified type together constitute a complete set of blocking semantic association edges.

[0037] By integrating adaptive semantic association edges, triggering semantic association edges, and blocking semantic association edges, a semantic association network is formed that covers the complete closed-loop logic of "control configuration adaptability assessment - abnormal risk triggering path - effective intervention blocking mechanism". Structurally, this network supports multi-hop reasoning. For example, starting from a process boundary condition pattern node, it finds a matching control strategy pattern node via adaptive semantic association edges, identifies potential fault propagation path nodes via triggering semantic association edges, and finally locates the recommended solution pattern node via blocking semantic association edges, forming a complete diagnostic and decision-making reasoning chain. Since all nodes and edges in the graph are constructed based on essential feature descriptors rather than project-specific values, pre-debugging experiences from different industry backgrounds and project scales can coexist in the same semantic space, enabling cross-project knowledge fusion and generalized reasoning, providing high-quality knowledge support for query responses to new pre-debugging tasks.

[0038] In one alternative implementation, a query request for a new pre-debugging task is received, multi-hop reasoning based on semantic distance is performed in the abstract knowledge graph, historical cases that are similar in essential features to the query request are identified, and a recommended knowledge package is generated, including: Receive a query request for a new pre-debugging task, the query request including a description of the functional requirements and process scenario of the control module to be debugged; The essential features of the functional requirements description and the process scenario description are abstracted, and the essential attributes of the adjustment target, the essential attributes of the constraint conditions, and the essential attributes of the response requirements are extracted to form a query feature representation. Starting from the query feature representation, multi-hop reasoning is performed along the semantic association network in the abstract knowledge graph. Starting from the control strategy mode node, the process boundary condition mode node is reached through the adapted semantic association edge, and then the fault propagation path node is reached through the triggered semantic association edge. Finally, the solution mode node is reached through the blocking semantic association edge. During the reasoning process, the semantic distance between the node on each jump path and the query feature representation is calculated cumulatively. All nodes involved in the complete reasoning path with a cumulative semantic distance lower than the similarity judgment threshold are marked as candidate matching nodes. Tracing back from the candidate matching nodes to the original knowledge elements used when constructing the abstract knowledge graph, the corresponding test case templates, potential risk warnings, and optimization suggestions are extracted and organized into the recommended knowledge package.

[0039] When receiving a query request for a new pre-commissioning task, this request is typically submitted by engineers during the pre-commissioning preparation phase and contains two core information dimensions: a functional requirement description of the control module to be commissioned and a process scenario description. The functional requirement description covers the adjustment objectives the control module needs to achieve, the types of input and output signals, response time requirements, and accuracy indicators; the process scenario description includes information such as the industry type, media characteristics, operating pressure and temperature range, equipment connection relationships, and safety interlock logic. These two types of descriptions are passed in either natural language or structured forms, forming the raw input for subsequent feature abstraction and reasoning.

[0040] When abstracting the essential features of functional requirements and process scenarios, the specific numerical values ​​or industry terms in the original descriptions are not used directly. Instead, they are mapped to abstract essential attributes. The extraction process of essential attributes for the control target focuses on the type of core control problem solved by the control module. For example, "maintaining the liquid level near the set value" is abstracted as "single-variable steady-state tracking," and "rapidly raising the reactor temperature to the target value" is abstracted as "rapid tracking of a large inertial system." The extraction of essential attributes for constraints transforms specific parameters such as upper temperature limits and pressure thresholds into essential descriptors such as "hard constraints / soft constraints," "single boundary / double boundary," and "time-varying / time-invariant." The extraction of essential attributes for response requirements transforms specific indicators such as overshoot, settling time, and steady-state error into essential classifications such as "high dynamic performance priority" or "steady-state accuracy priority." Through the extraction of essential attributes in these three dimensions, a query feature representation is formed. This representation is encoded in the form of structured vectors in the same semantic space as the abstract knowledge graph, ensuring the consistency of subsequent semantic distance calculations.

[0041] Starting with the query feature representation, multi-hop reasoning is performed along the semantic association network in the abstract knowledge graph. The reasoning process follows a fixed jump order: starting from the control strategy pattern node, it reaches the process boundary condition pattern node by adapting the semantic association edge, continues to reach the fault propagation path node by triggering the semantic association edge, and finally reaches the solution pattern node by blocking the semantic association edge. In the first hop reasoning, the semantic similarity between the query feature representation and all control strategy pattern nodes in the abstract knowledge graph is calculated. A set of candidate control strategy pattern nodes that highly match the essential attributes of the adjustment target and the essential attributes of the response requirement in the query feature representation is selected as the starting point set for multi-hop reasoning.

[0042] During each hop of inference, the semantic distance between the current node and the query feature representation is calculated and added to the total path distance. The semantic distance is calculated based on the distance metric between the essential feature descriptor vectors of nodes. Specifically, let the query feature representation vector be... The essential feature descriptor vector of a node in an abstract knowledge graph is: Then the semantic distance between the two Calculated using cosine distance: Cumulative semantic distance along the reasoning path The weighted sum of the semantic distances of each hop node is calculated, where the weight of each hop is pre-set according to the semantic association type corresponding to that hop, so as to reflect the difference in importance of different types of associations in the overall matching.

[0043] Once a complete four-node reasoning path, starting from the control strategy pattern node, passing through the process boundary condition pattern node, the fault propagation path node, and finally reaching the solution pattern node, has been traversed, its cumulative semantic distance is calculated. Similarity determination threshold Compare. If If so, all nodes involved in the complete reasoning path are marked as candidate matching nodes, and the path is recorded as a candidate matching path. Similarity determination threshold. The configuration should be tailored to the specific engineering needs. A smaller threshold results in more precise but fewer candidate matching nodes; a larger threshold provides broader coverage but introduces less relevant historical cases. In actual pre-testing scenarios, it is recommended to dynamically adjust the threshold based on the size of the knowledge graph and the richness of historical cases. In order to strike a balance between accuracy and coverage.

[0044] After multi-hop inference, several candidate matching paths and their corresponding candidate matching node sets are obtained. These candidate matching nodes are traced back to the original knowledge elements used in constructing the abstract knowledge graph, i.e., a reverse mapping from abstract nodes to specific historical pre-debugging process data records. This tracing process relies on the "abstract node—original knowledge element" mapping table maintained during the construction of the abstract knowledge graph. Each abstract node has its corresponding set of original knowledge element identifiers recorded during the graph construction phase, thus enabling efficient reverse tracing.

[0045] From the traced original knowledge elements, three types of information are extracted and organized into a recommended knowledge package. Test case templates are derived from validated test step sequences during historical pre-debugging processes, including signal excitation methods, observation point locations, judgment criteria, and expected response ranges. After parameter generalization, these are transformed into templates adaptable to new tasks. Potential risk warnings are derived from historical abnormal event records associated with fault propagation path nodes. The most relevant abnormal patterns to the current query features are extracted and presented as warning entries, including descriptions of abnormal triggering conditions, impact scope, and suggested monitoring indicators. Optimization suggestions are derived from historical tuning operation records associated with solution pattern nodes, extracting parameter tuning strategies, logic optimization directions, and debugging sequence suggestions. These three types of information are then ranked according to the cumulative semantic distance of candidate matching paths. The output is organized by sorting from smallest to largest, allowing engineers to prioritize access to historical experience references that are closest to the essential characteristics of the current task. This enables them to obtain targeted knowledge support before starting a new pre-debugging task, effectively reducing debugging risks and improving efficiency.

[0046] The recommended knowledge packages support a structured presentation format, with knowledge items for each of the control strategy mode, process boundary condition mode, fault propagation path, and solution mode presented in separate sections. Engineers can quickly locate the information they need based on their actual concerns and make selections and adjustments to the recommended content based on their own engineering judgments.

[0047] In one optional implementation, comparing and analyzing the execution results of the new pre-debugging task with the expected results, extracting deviation patterns, and identifying missing or erroneous associations in the abstract knowledge graph includes: Obtain the execution results of the new pre-debugging task and the expected effects of the recommended knowledge package. Perform a quantitative analysis of the difference between the execution results and the expected effects. Extract response timing inconsistencies, steady-state deviations, and abnormal triggering patterns. Abstract the response timing inconsistencies, steady-state deviations, and abnormal triggering patterns into the essential triggering mechanism and essential propagation mechanism of the deviation, forming a deviation pattern. The deviation pattern is traced back to the abstract knowledge graph to locate the reasoning path traversed when generating the recommended knowledge package. The semantic association edges between adjacent nodes in the reasoning path are checked to see if they accurately describe the essential triggering mechanism and essential propagation mechanism of the deviation. If there is a lack of node jumps in the reasoning path that can explain the deviation pattern, it is determined to be a missing association. If the causal relationship described by the semantic association edges in the reasoning path contradicts the actual causal relationship revealed by the deviation pattern, it is determined to be an incorrect association.

[0048] When obtaining the execution results of a new pre-debugging task, the expected results preset in the recommended knowledge package must be read simultaneously as a comparison benchmark. The execution results include the actual response curves of the control loop, the timing records of equipment state switching, the steady-state sampling sequences of process variables, and the abnormal event logs; the expected results are derived from the standard response timing, allowable steady-state deviation range, and known abnormal triggering conditions inferred by the recommended knowledge package based on historical cases during the generation phase. After aligning the two sets of data on a unified time axis, a dimensional difference quantitative analysis is performed.

[0049] The quantification of response timing inconsistency involves calculating the time offset between the actual and expected response times, and statistically analyzing the distribution characteristics of the offsets at each control action node, including the mean, variance, and the frequency of systematic lead or lag. If a control action consistently exhibits a unidirectional offset across multiple execution cycles, it is marked as a timing anomaly with a deterministic triggering pattern, rather than a random disturbance. The quantification of steady-state deviation involves calculating the normalized deviation between the actual steady-state mean and the expected steady-state target value after the process variable enters the steady-state region. ,Right now ,in This is the actual steady-state mean. This represents the expected steady-state target value. When... When the preset tolerance zone is exceeded, the process variable is included in the deviation analysis. The extraction of abnormal triggering patterns relies on pattern mining of abnormal event logs to identify which precursor control actions or process state transitions have stable co-occurrence relationships with abnormal events in time sequence, thereby summarizing the combination of conditions for abnormal triggering.

[0050] The purpose of abstracting the three types of quantitative results mentioned above is to strip away the numerical shell of specific engineering parameters and extract the essential triggering mechanism and essential propagation mechanism of the deviation. The essential triggering mechanism describes the initial condition for the deviation to occur, such as "when the execution order of two interdependent control strategies is reversed at the process switching boundary, a steady-state deviation is triggered." The essential propagation mechanism describes how the deviation spreads along the control loop or equipment topology, such as "upstream steady-state deviation is amplified through the feedforward channel and acts on the downstream regulation loop, resulting in a significant increase in the frequency of downstream abnormal triggering." This abstract expression method is consistent with the essential feature descriptor language used for nodes and associated edges in the abstract knowledge graph, thus laying a semantic alignment foundation for subsequent backtracking analysis. Combining and encapsulating the essential triggering mechanism and the essential propagation mechanism forms a structured deviation pattern object.

[0051] Once the deviation pattern is formed, it is traced back to the abstract knowledge graph to locate the reasoning path traversed when generating the recommended knowledge package. The reasoning path is an ordered sequence of alternating nodes and semantically related edges, recording the complete path from the query features through multi-hop reasoning to the recommended solution. The first step of the backtracking operation is to semantically match the essential triggering mechanism in the deviation pattern with the essential feature descriptors of each node in the reasoning path, confirming which node(s) interval in the reasoning path the starting point of the deviation corresponds to. The second step is to examine the semantically related edges between adjacent nodes segment by segment along the reasoning path, determining whether the causal relationship described by the edge accurately covers the essential triggering and propagation mechanisms revealed by the deviation pattern.

[0052] If a propagation mechanism exists within the deviation pattern, indicating that the control strategy state propagates to the fault trigger node via a specific intermediate process state, but within the corresponding interval of the inference path, there is no jump path from the trigger start node to the fault node that passes through this intermediate process state node, it means that the inference path skipped an intermediate node that should have existed during traversal, causing the propagation mechanism to be ignored. In this case, it is determined that there is a missing association in this interval, and semantic association edges need to be added between the start node and the intermediate process state node, and between the intermediate process state node and the fault node, respectively. The essence of the missing association is that the knowledge graph lacks sufficient understanding of a certain type of propagation path during the construction phase, and the scenario in which this intermediate node was activated has never occurred in historical cases, thus failing to establish the corresponding jump relationship.

[0053] If the causal direction or causal condition described by a semantically related edge in the inference path contradicts the actual causal relationship revealed by the deviation pattern, it is judged as an erroneous association. For example, an edge in the inference path describes "tightening process boundary conditions leads to faster control strategy response, thereby suppressing abnormal triggering," but deviation pattern analysis shows that the actual situation is exactly the opposite—tightening process boundary conditions actually leads to response lag due to the integral saturation effect of the control strategy, thus exacerbating the frequency of abnormal triggering. This reversal of causal direction is a typical erroneous association, and the semantic description of the related edge needs to be corrected to a causal expression consistent with the actual deviation pattern. The source of erroneous associations is that historical cases exhibited patterns opposite to those of the current task under specific process scales or industry contexts, and this difference in applicable boundaries was not identified during cross-project migration.

[0054] In practice, the identification of missing and erroneous associations is not independent; both types of problems may exist simultaneously within the same inference path interval. For example, an interval may lack intermediate node jumps describing the actual propagation path, while also containing causal description errors on existing direct association edges. In terms of processing order, erroneous associations are corrected first, as they can lead to systematic biases in subsequent inferences based on that path. After correcting erroneous associations, missing associations are then added to improve the knowledge graph's coverage of such bias patterns. The output of the entire identification process is a set of inference path annotations, marking both missing and erroneous locations. This serves as input for subsequent incremental graph structure evolution algorithms, driving the abstract knowledge graph to adaptively correct itself.

[0055] The level of abstraction in deviation patterns directly affects the accuracy of identifying missing and erroneous associations. If the level of abstraction is too low, the deviation patterns still carry a large amount of specific engineering parameter information, leading to misjudgments in semantic matching with knowledge graph nodes due to differences in parameter values. If the level of abstraction is too high, the deviation patterns lose the key details that distinguish different triggering mechanisms, making it impossible to accurately locate the problem interval in the reasoning path. Therefore, in the process of abstracting response timing inconsistencies, steady-state deviations, and abnormal triggering patterns into essential triggering mechanisms and essential propagation mechanisms, it is necessary to retain structural features sufficient to distinguish different deviation types while removing numerical parameters bound to specific projects. This ensures that the abstraction results match the semantic descriptor system of the knowledge graph at the granularity level.

[0056] In one optional implementation, the abstract knowledge graph is adaptively corrected using an incremental graph structure evolution algorithm, and the node weights and edge weights in the abstract knowledge graph are dynamically adjusted based on knowledge reuse frequency and effectiveness feedback, including: Obtain the node pairs corresponding to missing associations and the semantic association edges corresponding to incorrect associations; Candidate semantic association edges are generated between corresponding node pairs using an incremental graph structure evolution algorithm. The causal transmission paths revealed in the deviation patterns are analyzed, and the semantic type and propagation direction of the candidate semantic association edges are determined based on the causal transmission paths. The complete inference chain containing the erroneous association is deconstructed using an incremental graph structure evolution algorithm, the propagation node that causes the inference contradiction in the complete inference chain is located, and the association topology between the propagation node and its adjacent nodes is reconstructed. The incremental graph structure evolution algorithm is used to track the influence spread range of the missing associations and the erroneous associations in the abstract knowledge graph, and the influence spread range is used to define the action boundary of the correction operation, and topological evolution is performed on the local subgraphs within the boundary. Establish node access tracking for the corrected local subgraph and record the cumulative number of times each node is traversed by the reasoning path in subsequent queries as the knowledge reuse frequency. Collect the deviation between the actual execution effect and the expected effect of each query task as effectiveness feedback; A node contribution evaluation function is constructed. Based on the node contribution evaluation function, the knowledge reuse frequency is used as the node activity index, and the effectiveness feedback is used as the node reliability index. The product of the node activity index and the node reliability index is calculated as the node comprehensive contribution value. The node weight is adjusted according to the node comprehensive contribution value. The adjustment amount of the node weight is propagated to the semantic association edges connected to the corresponding node. The edge weights are adjusted according to the weight gradient of the connected nodes.

[0057] When acquiring node pairs corresponding to missing associations, based on the deviation pattern analysis results, two nodes with potential semantic associations but no explicit connection are located from the abstract knowledge graph and recorded as candidate node pairs. For erroneous associations, the semantic association edges that cause inference contradictions are directly extracted, including the starting node, ending node, current semantic type label, and propagation direction label of the edge, as the operation objects of subsequent incremental graph structure evolution algorithms. The core of this step is to accurately locate the operation target and avoid interference with irrelevant areas in the abstract knowledge graph, thereby ensuring the locality and controllability of the correction operation.

[0058] To address missing associations, the incremental graph structure evolution algorithm generates candidate semantic association edges between corresponding node pairs. Specifically, it analyzes the causal propagation paths revealed in the deviation patterns, identifies the causal directionality between each link in the path, i.e., determines which node's state change is the antecedent causing another node's state change. Based on this causal directionality, the candidate semantic association edges are assigned corresponding semantic type labels, such as "triggering relationship," "inhibition relationship," or "conditional dependency relationship," and the propagation direction is determined, giving the candidate semantic association edges clear directed semantics in the graph structure. After the candidate semantic association edges are generated, they are not directly incorporated into the abstract knowledge graph. Instead, they undergo consistency verification to check whether they logically contradict existing reasoning paths in the graph. Only candidate edges that pass the verification are formally inserted into the graph, completing the completion of missing associations.

[0059] To address erroneous associations, the incremental graph structure evolution algorithm deconstructs the complete inference path containing the erroneous association. The complete inference path refers to the entire directed path from the query starting point, through multiple hops of inference, to the final conclusion node. During deconstruction, each hop is checked to verify whether the propagation logic of each semantic association edge matches the actual causal pattern revealed by the deviation pattern, thus locating the propagation node causing the inference contradiction—that is, at this node, the inference direction or semantic type deviates from the actual situation. After locating the propagation node, the association topology between this propagation node and its adjacent nodes is reconstructed, including deleting the original erroneous association edges, re-establishing correct semantic association edges according to the deviation pattern, and updating the semantic type labeling and propagation direction. During reconstruction, the node feature description of the propagation node itself is preserved, and only its connection relationships are adjusted to minimize the disturbance to the overall graph structure.

[0060] Defining the scope of influence diffusion is crucial for ensuring clear boundaries of the correction operation. The incremental graph structure evolution algorithm starts from the missing association completion position and the erroneous association correction position, and diffuses outward along the semantic association edges in a breadth-first manner, tracking all nodes and edges affected by the correction operation until the semantic distance on the diffusion path exceeds a preset influence radius threshold. The set of nodes and edges covered by this diffusion scope constitutes a local subgraph. Subsequent topological evolution operations are performed within this local subgraph and do not affect the graph region outside the boundary. The topological evolution of the local subgraph includes a re-organization of the node connection relationships within the subgraph to ensure that the corrected local subgraph has internal semantic consistency, that is, the reasoning path between any two nodes in the subgraph no longer produces logical contradictions.

[0061] After the corrections are completed, a node access tracking mechanism is established for the local subgraph. Whenever a query request triggers multi-hop inference, each node traversed by the inference path is recorded as an access. As subsequent query requests accumulate, the number of times each node is traversed by the inference path continuously adds up, forming the knowledge reuse frequency of that node. Knowledge reuse frequency reflects the activity level of a node during actual inference; a higher frequency indicates that the knowledge carried by that node is repeatedly relied upon in multiple pre-debugging tasks, possessing higher general value.

[0062] The collection of effectiveness feedback and the tracking of knowledge reuse frequency are carried out in parallel. After each query and the completion of the corresponding pre-debugging task, the actual execution effect is compared with the expected effect, and the deviation between the two is calculated. This deviation is used as the effectiveness feedback value and associated with the nodes traversed by the inference path. If the deviation is small, it indicates that the knowledge package recommended by the inference path has high guiding effectiveness for the task; if the deviation is large, it indicates that there are nodes with insufficient reliability in the inference path. Through the accumulation of effectiveness feedback from multiple tasks, each node forms a historical reliability sequence associated with it.

[0063] The node contribution evaluation function uses the frequency of knowledge reuse as an indicator of node activity and the overall performance of effectiveness feedback as an indicator of node reliability. Let the node... Knowledge reuse frequency is The node reliability index is Then the node's overall contribution value Defined as: ; according to The node weights are adjusted so that nodes with high overall contribution values ​​have higher weights and nodes with low overall contribution values ​​have lower weights, thus giving high-quality, highly reusable knowledge nodes in the graph higher priority in subsequent reasoning.

[0064] After the node weights are adjusted, the adjustment is propagated to the semantic association edges connected to the corresponding nodes. Let the node... The weight adjustment amount is , with nodes A semantically related edge connects the nodes. With neighboring nodes The weight adjustment of the edge is determined based on the weight gradient of the two endpoints, meaning the direction of the edge weight adjustment is consistent with the changing trend of the weight difference between the two endpoints. If the node The weight is significantly higher than that of the neighboring nodes. If the node weights are high, the edge weights connecting the two nodes will increase accordingly, reflecting a stronger correlation credibility in the inference path flowing from the high-contribution node to the low-contribution node; conversely, the edge weights will decrease. This weight propagation mechanism allows the dynamic adjustment of node weights to synchronously affect the edge structure of the graph, ensuring semantic consistency between node weights and edge weights, and avoiding the problem of broken inference paths caused by high node weights but low connecting edge weights.

[0065] Through the synergistic effect of the incremental graph structure evolution and dynamic weight adjustment, the abstract knowledge graph can be specifically corrected and optimized after each round of pre-debugging tasks, gradually improving the graph's generalization reasoning ability for different process scenarios, and thus continuously improving the quality of the recommended knowledge package for subsequent new pre-debugging tasks.

[0066] In one optional implementation, the complete inference chain containing the erroneous association is deconstructed using an incremental graph structure evolution algorithm, the propagation node in the complete inference chain that leads to the inference contradiction is located, and the association topology between the propagation node and its neighboring nodes is reconstructed, including: Obtain the semantic association edges corresponding to the erroneous associations, trace the historical reasoning paths formed by the semantic association edges, and splice the historical reasoning paths according to the node access order to form a complete reasoning link. Semantic transfer verification is performed on the complete inference chain. The semantic continuity of each node in the complete inference chain is calculated. The jump position where the semantic continuity drops sharply is identified. The subsequent node of the jump position is marked as the propagation node that leads to the inference contradiction. Extract the semantic feature vector of the propagation node, compare the semantic feature vector with the semantic feature vectors of all neighboring nodes of the propagation node for similarity, select target neighboring nodes that are semantically compatible with the propagation node based on the similarity comparison results, determine the semantic association type between the propagation node and the target neighboring nodes, delete the original semantic association edges that cause a decrease in semantic continuity, and create new semantic association edges between the propagation node and the neighboring nodes that match the similarity comparison results, thereby realizing the reconstruction of the association topology.

[0067] like Figure 2 As shown, the method includes: When performing adaptive correction on an abstract knowledge graph, it is necessary to perform in-depth link deconstruction and topology reconstruction for identified erroneous associations. Erroneous associations do not exist in isolation within the graph; they are often embedded in multiple historical reasoning paths. If only a single edge is deleted without tracing the complete reasoning path it participated in, it will cause logical incompleteness in the local structure of the graph, thus affecting the correctness of subsequent multi-hop reasoning. Therefore, when dealing with erroneous associations, the incremental graph structure evolution algorithm first needs to start from the semantic association edge corresponding to the erroneous association and trace back bidirectionally upstream and downstream all historical reasoning paths formed by that edge.

[0068] After obtaining the semantic association edge corresponding to the erroneous association, all inference path records that have passed through that edge are retrieved based on the inference history log recorded in the graph. Each historical inference path is stored in the form of a node access sequence, including an ordered arrangement of the starting node, intermediate jump nodes, and ending node. These historical inference paths are concatenated according to the node access order, and after removing duplicate sub-paths, a complete inference link covering the scope of influence of the erroneous association edge is formed. The length of the complete inference link depends on the longest path involving that edge in the historical inference, and typically covers the entire jump sequence from the initial query feature node to the final recommended knowledge node.

[0069] After the complete inference chain is determined, semantic transitivity verification is performed. The core of semantic transitivity verification lies in calculating the semantic continuity between each pair of adjacent nodes in the chain. Semantic continuity reflects the extent to which the semantic content of the preceding node can provide a reasonable basis for the semantic content of the following node. Let the first node in the complete inference chain be... The semantic continuity of the jump is It calculates the first The essential feature descriptor vector of the nth node and the nth node The cosine similarity between the essential feature descriptor vectors of each node is obtained by weighting and adjusting the values ​​based on the semantic association type weights corresponding to the jump. When the average continuity of preceding jumps in the link drops sharply by more than a preset margin, a precipitous drop is identified. After identifying the jump position of the precipitous drop, the subsequent node at that jump position is recorded as the propagation node that caused the inference contradiction. There is a logical discontinuity between the semantic content of the propagation node and its predecessor node. It is this discontinuity that causes the inference link to diverge at this point, thus generating recommended knowledge output that does not match the actual pre-debugging results.

[0070] Once the propagation node is determined, its semantic feature vector is extracted. This vector is composed of the essential feature descriptor of the propagation node, the label of the process object type to which it belongs, the control logic category encoding, and the context embedding of historical inference, which can comprehensively represent the position of the node in the semantic space. Subsequently, all neighboring nodes of the propagation node are traversed, and the semantic feature vectors of each neighboring node are extracted. ,calculate With each semantic similarity between Semantic similarity Calculated using the normalized inner product method, i.e. The range of values ​​is The closer the value is This indicates that the two nodes are more compatible in the semantic space. Let the semantic compatibility filtering threshold be... Only retain those that meet the requirements. The neighboring nodes are used as the candidate set of the target neighboring nodes.

[0071] After the candidate set of target neighbor nodes is determined, further refined matching is performed based on the semantic association type between the propagation node and each candidate target neighbor node. The semantic association types include four categories: "control strategy adaptation," "process boundary constraint," "fault propagation trigger," and "solution mapping," each corresponding to different semantic transmission rules. By analyzing the logical level of the propagation node in the complete inference chain—that is, which type of pattern node its predecessor node belongs to and which type of pattern node its successor node is expected to belong to—the semantic association type that should be established between the propagation node and the target neighbor node can be inferred. For example, if the predecessor of the propagation node is a control strategy pattern node, and its expected successor should be a fault propagation path node, then a semantic association edge of the "fault propagation trigger" type should be established between the propagation node and the target neighbor node.

[0072] After identifying the target's adjacent nodes and their corresponding semantic association types, a reconstruction of the association topology is performed. Existing semantic association edges that caused a sharp drop in semantic continuity are deleted. This deletion simultaneously triggers the invalidation of all historical inference paths in the complete inference chain that depend on these edges, preventing the incorrect reuse of invalid paths in subsequent queries. New semantic association edges are created between the propagation node and the target's adjacent nodes. These new edges carry semantically compatible association type labels and initial edge weights, determined by semantic similarity. The product of the association type and the prior confidence level determines the newly created association edge, ensuring that it has reasonable initial reasoning guidance in the graph.

[0073] After the new edge is created, a forward semantic pass verification is performed on the reconstructed local topology, and the semantic continuity of each hop involving the propagation nodes in the complete inference link is recalculated. The process confirms that the precipitous drop in performance has been eliminated, and the semantic connectivity of the entire link remains within a reasonable range. If the verification passes, the associated topology reconstruction is complete, and relevant operation records are written to the graph evolution log for subsequent incremental correction and tracing. If the verification fails, a suboptimal node is selected from the candidate set of adjacent nodes of the target, and the reconstruction is re-executed until the semantic transmission verification passes or the candidate set is exhausted. When the candidate set is exhausted and the verification still fails, the propagation node is marked as a node to be supplemented, triggering the knowledge completion process. Relevant semantic fragments are extracted from the execution data of the new pre-debugging task, and new intermediate nodes are constructed and inserted into the graph to fill the semantic gap at that point.

[0074] Through the complete reasoning link deconstruction, propagation node localization, and association topology reconstruction process described above, the incremental graph structure evolution algorithm can accurately repair local erroneous associations without disrupting the global graph structure. Each repair operation adheres to the principle of minimizing graph structure perturbation, reconstructing only the link segments directly affected by erroneous associations while preserving the remaining verified and valid semantic association networks in the graph. This maintains the overall reasoning consistency and cross-project transferability of the abstract knowledge graph during the continuous accumulation of pre-debugging experience.

[0075] A second aspect of this invention provides a DCS control module pre-debugging system based on a virtual image, comprising: The knowledge extraction unit is used to acquire process data generated during the pre-debugging process of the distributed control module executed in the virtual image environment, perform semantic parsing on the process data, extract control logic features, process object type features, equipment topology features and abnormal mode features, and construct a structured set of knowledge elements. The knowledge graph construction unit is used to construct an abstract knowledge graph with cross-project transfer capability based on the knowledge element set. The abstract knowledge graph abstracts specific engineering parameters into essential feature descriptors and establishes a semantic association network between control strategy patterns, process boundary condition patterns, fault propagation paths and solution patterns, so that pre-debugging experience from different industries and scales can be integrated and generalized in a unified semantic space. The reasoning and recommendation unit is used to receive query requests for new pre-debugging tasks, perform multi-hop reasoning based on semantic distance in the abstract knowledge graph, identify historical cases that are similar to the query requests in essential features, and generate recommended knowledge packages. The graph correction unit is used to compare and analyze the execution results of the new pre-debugging task with the expected results, extract deviation patterns and identify missing or erroneous associations in the abstract knowledge graph, adaptively correct the abstract knowledge graph through an incremental graph structure evolution algorithm, and dynamically adjust the node weights and edge weights in the abstract knowledge graph based on the knowledge reuse frequency and effectiveness feedback.

[0076] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0077] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0078] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0079] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention 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; and these 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 the present invention.

Claims

1. A DCS control module pre-debugging method based on virtual image, characterized in that, include: The process data generated during the pre-debugging process of the distributed control module executed in the virtual image environment is obtained. The process data is semantically parsed to extract control logic features, process object type features, equipment topology features and anomaly pattern features, and a structured set of knowledge elements is constructed. Based on the set of knowledge elements, an abstract knowledge graph with cross-project transfer capability is constructed. The abstract knowledge graph abstracts specific engineering parameters into essential feature descriptors and establishes a semantic association network between control strategy patterns, process boundary condition patterns, fault propagation paths and solution patterns, so that pre-debugging experience from different industries and scales can be integrated and generalized in a unified semantic space. Receive a query request for a new pre-debugging task, perform multi-hop reasoning based on semantic distance in the abstract knowledge graph, identify historical cases that are similar in essential features to the query request, and generate a recommended knowledge package; The execution results of the new pre-debugging task are compared and analyzed with the expected results. Deviation patterns are extracted and missing or erroneous associations in the abstract knowledge graph are identified. The abstract knowledge graph is adaptively corrected through an incremental graph structure evolution algorithm. At the same time, the node weights and edge weights in the abstract knowledge graph are dynamically adjusted based on the knowledge reuse frequency and effectiveness feedback.

2. The method according to claim 1, characterized in that, The process data generated during the pre-debugging process of the distributed control module executed in a virtual image environment is acquired. Semantic parsing is performed on the process data to extract control logic features, process object type features, equipment topology features, and anomaly pattern features, constructing a structured set of knowledge elements, including: The syntax tree is parsed to analyze the control strategy configuration information in the process data, identify the conditional judgment structure, loop execution structure and sequential execution structure in the control logic, extract the input variable dependency relationship and output variable influence range of each structure, and generate a control logic feature description. Perform time-series analysis on the test case execution records in the process data to identify the state transition patterns and parameter change trends of the process object in different operation stages. Semantically match the state transition patterns with the predefined process unit operation types to generate process object type feature descriptions. Graph structure modeling is performed on the equipment topology information in the process data to identify the communication connection relationships, redundancy configuration relationships and hierarchical membership relationships between control modules, and to generate equipment topology feature descriptions. Pattern recognition is performed on the fault event descriptions in the process data to extract the association rules between fault triggering conditions, fault manifestations and fault impact range, and to generate abnormal pattern feature descriptions. The control logic feature description, the process object type feature description, the equipment topology feature description, and the anomaly mode feature description are associated and mapped to establish the interaction relationship between control logic and process object, and the propagation relationship between equipment topology and anomaly mode, forming a structured set of knowledge elements.

3. The method according to claim 1, characterized in that, Based on the aforementioned set of knowledge elements, an abstract knowledge graph with cross-project transfer capabilities is constructed. This abstract knowledge graph abstracts specific engineering parameters into essential feature descriptors and establishes a semantic association network between control strategy patterns, process boundary condition patterns, fault propagation paths, and solution patterns, including: The specific engineering parameters in the structured knowledge element set are abstracted and transformed at multiple levels. By extracting the causal reasoning links in the control logic features and the constraint boundary range in the process object type features, the project-specific numerical parameters are converted into essential feature descriptors. Based on the essential feature descriptor, control strategy mode nodes, process boundary condition mode nodes, fault propagation path nodes, and solution mode nodes are constructed. Adaptive semantic association edges are established by analyzing the sensitivity dependency of control strategy modes on process boundary condition modes. Triggering semantic association edges are established by analyzing the evolution law of fault propagation paths under specific combinations of control strategy modes and process boundary condition modes. Blocking semantic association edges are established by analyzing the interruption mechanism of fault propagation paths and the correction requirements of control strategy modes by solution modes. The adaptive semantic association edges, the triggering semantic association edges, and the blocking semantic association edges together constitute a semantic association network.

4. The method according to claim 1, characterized in that, Upon receiving a query request for a new pre-debugging task, perform multi-hop reasoning based on semantic distance in the abstract knowledge graph, identify historical cases that are similar in essential features to the query request, and generate a recommended knowledge package including: Receive a query request for a new pre-debugging task, the query request including a description of the functional requirements and process scenario of the control module to be debugged; The essential features of the functional requirements description and the process scenario description are abstracted, and the essential attributes of the adjustment target, the essential attributes of the constraint conditions, and the essential attributes of the response requirements are extracted to form a query feature representation. Starting from the query feature representation, multi-hop reasoning is performed along the semantic association network in the abstract knowledge graph. Starting from the control strategy mode node, the process boundary condition mode node is reached through the adapted semantic association edge, and then the fault propagation path node is reached through the triggered semantic association edge. Finally, the solution mode node is reached through the blocking semantic association edge. During the reasoning process, the semantic distance between the node on each jump path and the query feature representation is calculated cumulatively. All nodes involved in the complete reasoning path with a cumulative semantic distance lower than the similarity judgment threshold are marked as candidate matching nodes. Tracing back from the candidate matching nodes to the original knowledge elements used when constructing the abstract knowledge graph, the corresponding test case templates, potential risk warnings, and optimization suggestions are extracted and organized into the recommended knowledge package.

5. The method according to claim 1, characterized in that, The execution results of the new pre-debugging task are compared and analyzed with the expected results. Deviation patterns are extracted and missing or erroneous associations in the abstract knowledge graph are identified, including: Obtain the execution results of the new pre-debugging task and the expected effects of the recommended knowledge package. Perform a quantitative analysis of the difference between the execution results and the expected effects. Extract response timing inconsistencies, steady-state deviations, and abnormal triggering patterns. Abstract the response timing inconsistencies, steady-state deviations, and abnormal triggering patterns into the essential triggering mechanism and essential propagation mechanism of the deviation, forming a deviation pattern. The deviation pattern is traced back to the abstract knowledge graph to locate the reasoning path traversed when generating the recommended knowledge package. The semantic association edges between adjacent nodes in the reasoning path are checked to see if they accurately describe the essential triggering mechanism and essential propagation mechanism of the deviation. If there is a lack of node jumps in the reasoning path that can explain the deviation pattern, it is determined to be a missing association. If the causal relationship described by the semantic association edges in the reasoning path contradicts the actual causal relationship revealed by the deviation pattern, it is determined to be an incorrect association.

6. The method according to claim 1, characterized in that, The abstract knowledge graph is adaptively corrected using an incremental graph structure evolution algorithm, and the node weights and edge weights in the abstract knowledge graph are dynamically adjusted based on feedback on knowledge reuse frequency and effectiveness. Obtain the node pairs corresponding to missing associations and the semantic association edges corresponding to incorrect associations; Candidate semantic association edges are generated between corresponding node pairs using an incremental graph structure evolution algorithm. The causal transmission paths revealed in the deviation patterns are analyzed, and the semantic type and propagation direction of the candidate semantic association edges are determined based on the causal transmission paths. The complete inference chain containing the erroneous association is deconstructed using an incremental graph structure evolution algorithm, the propagation node that causes the inference contradiction in the complete inference chain is located, and the association topology between the propagation node and its adjacent nodes is reconstructed. The incremental graph structure evolution algorithm is used to track the influence spread range of the missing associations and the erroneous associations in the abstract knowledge graph, and the influence spread range is used to define the action boundary of the correction operation, and topological evolution is performed on the local subgraphs within the boundary. Establish node access tracking for the corrected local subgraph and record the cumulative number of times each node is traversed by the reasoning path in subsequent queries as the knowledge reuse frequency. Collect the deviation between the actual execution effect and the expected effect of each query task as effectiveness feedback; A node contribution evaluation function is constructed. Based on the node contribution evaluation function, the knowledge reuse frequency is used as the node activity index, and the effectiveness feedback is used as the node reliability index. The product of the node activity index and the node reliability index is calculated as the node comprehensive contribution value. The node weight is adjusted according to the node comprehensive contribution value. The adjustment amount of the node weight is propagated to the semantic association edges connected to the corresponding node. The edge weights are adjusted according to the weight gradient of the connected nodes.

7. The method according to claim 6, characterized in that, The algorithm uses an incremental graph structure evolution method to deconstruct the complete inference chain containing the erroneous association, locates the propagation node in the complete inference chain that leads to the inference contradiction, and reconstructs the association topology between the propagation node and its neighboring nodes, including: Obtain the semantic association edges corresponding to the erroneous associations, trace the historical reasoning paths formed by the semantic association edges, and splice the historical reasoning paths according to the node access order to form a complete reasoning link. Semantic transfer verification is performed on the complete inference chain. The semantic continuity of each node in the complete inference chain is calculated. The jump position where the semantic continuity drops sharply is identified. The subsequent node of the jump position is marked as the propagation node that leads to the inference contradiction. Extract the semantic feature vector of the propagation node, compare the semantic feature vector with the semantic feature vectors of all neighboring nodes of the propagation node for similarity, select target neighboring nodes that are semantically compatible with the propagation node based on the similarity comparison results, determine the semantic association type between the propagation node and the target neighboring nodes, delete the original semantic association edges that cause a decrease in semantic continuity, and create new semantic association edges between the propagation node and the neighboring nodes that match the similarity comparison results, thereby realizing the reconstruction of the association topology.

8. A DCS control module pre-debugging system based on virtual mirroring, used to implement the method as described in any one of claims 1-7, characterized in that, include: The knowledge extraction unit is used to acquire process data generated during the pre-debugging process of the distributed control module executed in the virtual image environment, perform semantic parsing on the process data, extract control logic features, process object type features, equipment topology features and abnormal mode features, and construct a structured set of knowledge elements. The knowledge graph construction unit is used to construct an abstract knowledge graph with cross-project transfer capability based on the knowledge element set. The abstract knowledge graph abstracts specific engineering parameters into essential feature descriptors and establishes a semantic association network between control strategy patterns, process boundary condition patterns, fault propagation paths and solution patterns, so that pre-debugging experience from different industries and scales can be integrated and generalized in a unified semantic space. The reasoning and recommendation unit is used to receive query requests for new pre-debugging tasks, perform multi-hop reasoning based on semantic distance in the abstract knowledge graph, identify historical cases that are similar to the query requests in essential features, and generate recommended knowledge packages. The graph correction unit is used to compare and analyze the execution results of the new pre-debugging task with the expected results, extract deviation patterns and identify missing or erroneous associations in the abstract knowledge graph, adaptively correct the abstract knowledge graph through an incremental graph structure evolution algorithm, and dynamically adjust the node weights and edge weights in the abstract knowledge graph based on the knowledge reuse frequency and effectiveness feedback.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images