Glass fiber production energy efficiency diagnosis knowledge graph construction method and system fusing industrial mechanism and data driving

Abstract

The invention discloses a glass fiber production energy efficiency diagnosis knowledge graph construction method and system fusing an industrial mechanism and data driving. The method comprises the following steps: 1, constructing a glass fiber production energy efficiency knowledge graph ontology fusing the industrial mechanism; and 2, driving the knowledge graph to realize dual-channel dynamic evolution. A real-time data source is connected, so that the static graph has learning and updating capabilities and comprises two parallel and collaborative evolution channels. And step 3, carrying out interpretable energy efficiency diagnosis reasoning based on the dynamic knowledge graph. And when it is monitored that the energy efficiency index is abnormal, a diagnosis process is started. The glass fiber production energy efficiency diagnosis knowledge graph construction system integrating the industrial mechanism and the data driving is a software and hardware integrated closed-loop intelligent agent and comprises the following three core function modules, and the modules closely cooperate with knowledge flow through data flow to jointly achieve the whole process from knowledge construction, dynamic evolution to intelligent diagnosis. According to the method, the diagnosis precision and robustness are remarkably improved.

Description

Technical Field

[0001] This invention relates to a method and system for constructing a knowledge graph for energy efficiency diagnosis in fiberglass production that integrates industrial mechanisms and data-driven approaches, belonging to the field of industrial artificial intelligence and knowledge engineering technology. Background Technology

[0002] Currently, in energy-intensive process manufacturing industries such as fiberglass, energy efficiency management is the core of cost reduction and efficiency improvement. Current energy efficiency analysis and diagnostic technologies mainly fall into three fragmented paradigms, making it difficult to meet the needs for accurate, real-time, and interpretable diagnostics in complex industrial scenarios:

[0003] Deep learning, neural networks, and other purely data-driven models rely on massive amounts of historical data to predict or classify by mining statistical correlations between variables. Their core flaw lies in their "black box" nature; diagnostic conclusions lack interpretability at the physical or technological level. When new operating conditions not covered by the training data (such as new raw material formulations or severe equipment degradation) occur in the production process, model performance drops sharply or even fails, and the recommendations given may contradict common industrial sense, making it difficult for on-site process personnel to trust and adopt them.

[0004] Pure mechanistic models based on first principles are constructed according to physical and chemical laws such as mass conservation, energy conservation, and chemical reaction kinetics, possessing strict logic and reliability. However, constructing a complete and accurate production line-level mechanistic model is extremely complex and costly. More importantly, key parameters in the model, such as heat transfer coefficients and reaction rate constants, are usually assumed to be ideal or design values, failing to reflect the dynamic performance degradation of equipment caused by scaling, wear, aging, etc., and also failing to capture soft factors such as human operating habits. This results in insufficient practicality for real-time online diagnostics, making the model "rigid."

[0005] Rule-based expert systems summarize the experience of domain experts into an "IF-THEN" rule base. The problem lies in the significant bottleneck in knowledge acquisition; expert experience is difficult to fully and unambiguously transform into rules; and the rule base is difficult to maintain, struggles to adapt to process adjustments and equipment updates, and easily becomes outdated. Therefore, the industry urgently needs a hybrid intelligent diagnostic solution that can integrate data, mechanisms, and experience to form a computable, evolvable, and interpretable approach to address these issues. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings and deficiencies of existing technologies, namely, the invisibility of data-driven models, the rigidity of mechanistic models, and the slow updates of expert systems. It provides a method and system for constructing a knowledge graph for energy efficiency diagnosis in fiberglass production that deeply integrates industrial mechanisms and real-time data and possesses self-evolving capabilities. This invention constructs a dynamic knowledge fusion and reasoning framework driven by both "mechanism and data," specifically including the following three levels:

[0007] Constructing a computable knowledge base with embedded mechanisms: This pioneering approach defines and stores equipment, processes, and energy efficiency indicators in the fiberglass production field through strong associations using first-principles formulas, empirical formulas, and other mechanistic models. This transforms the knowledge graph from a "database" storing static relationships into a "computable model" containing inherent computational logic.

[0008] To achieve a data-driven dynamic lifecycle for knowledge graphs, a dual evolutionary mechanism of online parameter calibration and implicit relationship mining is proposed. On the one hand, using real-time production data, key parameters such as the heat loss coefficient of the mechanistic model in the knowledge graph are dynamically corrected through recursive estimation algorithms such as extended Kalman filtering, synchronizing the model state with the actual health state of the equipment. On the other hand, through temporal correlation analysis, undefined causal relationships and implicit causal relationships in the data are automatically discovered and supplemented, enabling the knowledge graph to autonomously expand and update.

[0009] Provides interpretable diagnostic reasoning services: Based on the aforementioned dynamic knowledge graph, a multi-hop causal reasoning engine based on a probabilistic graphical model was designed. When energy efficiency anomalies occur, the engine can perform reverse tracing and confidence propagation along the semantic relationships in the knowledge graph, outputting a root cause diagnostic report with a complete causal path and quantified confidence, realizing transparent tracing of the diagnostic process from phenomenon to root cause.

[0010] The technical solution adopted by this invention to solve its technical problem is: a method for constructing a knowledge graph for energy efficiency diagnosis of fiberglass production that integrates industrial mechanisms and data-driven approaches, the method comprising the following steps:

[0011] Step 1: Construct an ontology of knowledge graphs on the energy efficiency of fiberglass production that integrates industrial mechanisms.

[0012] Based on the fiberglass production process, equipment principles, and energy balance, a standardized body model is defined as follows:

[0013] Step 1-1: Equipment layer entity construction: Define key equipment such as smelting furnace and wire drawing machine as entities, and encapsulate their static attributes, dynamic operating parameters and energy consumption data.

[0014] Step 1-2: Construction of the process layer: Define the core processes such as melting, drawing, and drying as entities, and define the material / energy flow input and output relationships and timing logic relationships between processes.

[0015] Steps 1-3: Energy Efficiency Layer Ontology Construction: Define indicators such as unit product energy consumption and thermal efficiency as entities. Through predefined mechanisms or empirical formulas, such as: unit product power consumption = total power consumption of process / qualified product output, strongly calculate and associate the energy efficiency indicator entities with the dynamic attributes of the equipment layer and the parameter entities of the process layer. Store these mathematical relationships as calculation relationship edges between entities in the knowledge graph to form a computable knowledge network skeleton.

[0016] Step 2: Drive the knowledge graph to achieve dual-channel dynamic evolution. Connect to real-time data sources to enable the static graph to learn and update, including two parallel and collaborative evolution channels.

[0017] Step 2-1: Implicit Knowledge Mining Channel: Periodically extract high-confidence statistical associations between variables from historical / real-time data using time-series association rule mining or graph neural network methods. If the association is not defined by the mechanistic knowledge in Step 1, it is treated as a new relation edge with confidence weights, submitted for review, and then integrated into the knowledge graph to supplement the blind spots of the mechanistic model.

[0018] Step 2-2: Mechanism Model Parameter Calibration Channel: For the mechanism formula defined in Step 1, which includes adjustable parameters such as heat transfer coefficient and efficiency decay factor, real-time production data is used as the observation sequence. Recursive parameter estimation algorithms such as extended Kalman filtering are employed to perform online, rolling estimation of these key parameters. Subsequently, the parameter values ​​bound to the corresponding entities in the knowledge graph are dynamically updated, ensuring that the mechanism model embedded in the graph always tracks the actual state of the equipment, achieving co-aging of the model and the equipment.

[0019] Step 3: Perform interpretable energy efficiency diagnostic reasoning based on dynamic knowledge graphs. When abnormal energy efficiency indicators are detected, the following diagnostic process is initiated, specifically including:

[0020] Step 3-1: Anomaly localization and reasoning trigger: Set the anomaly indicator node as the starting point for reasoning.

[0021] Step 3-2: Multi-hop Causal Tracing and Probability Calculation: Centered on the starting point, perform a reverse breadth / depth search along the edges of influencing factors and other relationships in the knowledge graph to enumerate all possible causal paths. For each node on the path, calculate a local anomaly probability based on its current real-time data state. Then, based on Bayes' theorem or a confidence propagation algorithm, aggregate the probabilities along the causal paths to calculate the global confidence of each terminal root cause hypothesis, such as "excessive vibration of a certain bearing".

[0022] Step 3-3: Generate an interpretable diagnostic report: Sort the root cause hypotheses by confidence level and output a structured report. The report not only lists the Top-K root causes, but more importantly, clearly shows the complete causal chain from the anomaly to each root cause, and provides real-time data of each node in the chain as evidence, thereby making the diagnostic conclusion verifiable and interpretable.

[0023] This invention also provides a knowledge graph construction system for energy efficiency diagnosis in fiberglass production that integrates industrial mechanisms and data-driven approaches to implement the aforementioned methods. This system is a closed-loop intelligent entity integrating hardware and software, comprising three core functional modules. These modules work closely together through data and knowledge flows to achieve the entire process from knowledge construction and dynamic evolution to intelligent diagnosis. Specific functions are as follows:

[0024] (1) Knowledge graph construction and storage module. This module is the knowledge foundation and static model carrier of the system. Its functions include:

[0025] Ontology Management: Provides graphical or configurable ontology editing tools, with a built-in general ontology library for equipment, processes, and indicators in the fiberglass production field. It supports users in instantiating entities based on specific production lines and defining static relationships between entities, such as "belongs to," "connected to," and "predecessor to."

[0026] Mechanism model embedding: Allows users to attach predefined mathematical formulas or computational logic to specific relationships, such as the relationship between energy efficiency indicators and process parameters. For example, the entity "unit product power consumption" can be bound to the "total power consumption" attribute of the "wire drawing machine" entity, and the "output" attribute of the "production batch" entity, using the formula SEC = total power consumption / output. These formulas are stored in the knowledge graph as "computational relationships," upgrading the graph from a relational network to a semantic model capable of performing computations.

[0027] Graph storage and query: The constructed entities, attributes, relationships, and additional mechanistic logic are persistently stored in the graph database, and standard Cypher and Gremlin query interfaces are provided for other modules to access.

[0028] (2) Data fusion and graph evolution module. This module is the intelligent engine and dynamic core of the system, and its functions include:

[0029] Multi-source data access and processing: Connect and collect heterogeneous data (including time series data, event logs, test results, etc.) from MES, SCADA, and sensor networks in real time, clean, align and format them, and convert them into information that can be digested by the knowledge graph.

[0030] The implicit relationship mining engine runs improved temporal association rule algorithms (such as the temporal extension of FP-Growth) or graph representation learning models to analyze historical and real-time data. Its function is to automatically discover strong statistical associations that exist stably between variables. If this association is not covered by the mechanistic relationships defined in the ontology, the engine generates a "data-driven relationship" suggestion with confidence weights (such as "ambient humidity → drying oven gas consumption"), which is then automatically or semi-automatically added to the knowledge graph after review.

[0031] Mechanism Model Parameter Online Calibration Engine: For mechanism models with adjustable parameters stored in the knowledge graph, such as the kiln thermal efficiency model, online parameter estimation is performed.

[0032] Its built-in recursive estimation algorithm framework (such as extended Kalman filter or recursive least squares method).

[0033] The engine obtains the input power and output temperature observation variables from the real-time data stream and combines them with the mechanism formulas retrieved from the knowledge graph.

[0034] In each calculation cycle, the engine outputs the latest estimates of key parameters such as the heat loss coefficient and actively writes the new values ​​back to the knowledge graph to update the attributes of the corresponding entities or relationships.

[0035] (3) Diagnostic Reasoning and Report Generation Module: This module serves as the system's application interface and value output, providing transparent diagnostic services. Its functions include:

[0036] Real-time monitoring and anomaly detection: Continuously monitor key energy efficiency indicators obtained from knowledge graphs or data streams, and use methods such as statistical process control or threshold judgment to automatically identify anomalies and trigger diagnostic processes.

[0037] Explainable Inference Engine: Upon receiving an anomaly trigger signal, the engine starts from the anomaly indicator node and performs reverse graph traversal and multi-hop inference on the current version of the dynamic knowledge graph. Based on the relationship types in the graph, especially causal relationships, multiple possible root cause tracing paths are constructed. The engine combines the latest real-time state data of each node entity on the path from the data fusion module and the weights of the relationship edges, and uses algorithms based on Bayesian networks or confidence propagation to calculate the global probability or confidence score of each potential root cause at the end of each path.

[0038] Structured Report Generator: Encapsulates the results of the inference engine to generate easy-to-understand diagnostic reports. The reports not only list the most likely root causes and their confidence rankings, but more importantly, they clearly present the complete and traceable causal chain from the anomaly to each root cause, along with the theoretical, actual, and deviation values ​​of each node in the chain, thus achieving complete visualization and verifiability of the diagnostic conclusions.

[0039] Beneficial effects:

[0040] 1. This invention significantly improves diagnostic accuracy and robustness. It ensures the lower limit of physical rationality of diagnosis through mechanistic model and adapts to equipment changes and discovers new knowledge through data-driven approach. It overcomes the inherent defects of pure data or pure mechanistic model and can maintain the expected high diagnostic accuracy of >90% under variable working conditions.

[0041] 2. This invention achieves white-box and interpretability of the diagnostic process. The output is no longer a simple fault code or classification label, but a clear causal chain that conforms to the process logic, enabling field engineers to understand, verify and take action quickly, which greatly improves the credibility and adoption rate of AI diagnostic results.

[0042] 3. This invention constructs an evolvable enterprise digital asset. The knowledge graph integrates scattered mechanistic knowledge, expert experience, and data insights into a unified, computable, queryable, and updatable digital model. This model continuously learns and evolves over time, becoming a core, sustainably value-added intelligent asset for the enterprise. Attached Figure Description

[0043] Figure 1 This is a diagram of the overall system architecture of the present invention.

[0044] Figure 2 This is a flowchart of the method of the present invention. Detailed Implementation

[0045] The invention will now be described in further detail with reference to the accompanying drawings.

[0046] Example 1

[0047] like Figure 1 As shown, this invention provides a knowledge graph construction system for energy efficiency diagnosis in fiberglass production that integrates industrial mechanisms and data-driven approaches. This system is a closed-loop intelligent entity integrating hardware and software, comprising three core functional modules. These modules work closely together through data and knowledge flows to achieve the entire process from knowledge construction and dynamic evolution to intelligent diagnosis. Specific functions are as follows:

[0048] (1) Knowledge graph construction and storage module. This module is the knowledge foundation and static model carrier of the system. Its functions include:

[0049] Ontology Management: Provides graphical or configurable ontology editing tools, with a built-in general ontology library for equipment, processes, and indicators in the fiberglass production field. It supports users in instantiating entities based on specific production lines and defining static relationships between entities, such as "belongs to," "connected to," and "predecessor to."

[0050] Mechanism model embedding: Allows users to attach predefined mathematical formulas or computational logic to specific relationships, such as the relationship between energy efficiency indicators and process parameters. For example, the entity "unit product power consumption" can be bound to the "total power consumption" attribute of the "wire drawing machine" entity, and the "output" attribute of the "production batch" entity, using the formula SEC = total power consumption / output. These formulas are stored in the knowledge graph as "computational relationships," upgrading the graph from a relational network to a semantic model capable of performing computations.

[0051] Graph storage and query: The constructed entities, attributes, relationships, and additional mechanistic logic are persistently stored in the graph database, and standard Cypher and Gremlin query interfaces are provided for other modules to access.

[0052] (2) Data fusion and graph evolution module. This module is the intelligent engine and dynamic core of the system, and its functions include:

[0053] Multi-source data access and processing: Connect and collect heterogeneous data (including time series data, event logs, test results, etc.) from MES, SCADA, and sensor networks in real time, clean, align and format them, and convert them into information that can be digested by the knowledge graph.

[0054] The implicit relationship mining engine runs improved temporal association rule algorithms (such as the temporal extension of FP-Growth) or graph representation learning models to analyze historical and real-time data. Its function is to automatically discover strong statistical associations that exist stably between variables. If this association is not covered by the mechanistic relationships defined in the ontology, the engine generates a "data-driven relationship" suggestion with confidence weights (such as "ambient humidity → drying oven gas consumption"), which is then automatically or semi-automatically added to the knowledge graph after review.

[0055] Mechanism Model Parameter Online Calibration Engine: For mechanism models with adjustable parameters stored in the knowledge graph, such as the kiln thermal efficiency model, online parameter estimation is performed.

[0056] Its built-in recursive estimation algorithm framework (such as extended Kalman filter or recursive least squares method).

[0057] The engine obtains the input power and output temperature observation variables from the real-time data stream and combines them with the mechanism formulas retrieved from the knowledge graph.

[0058] In each calculation cycle, the engine outputs the latest estimates of key parameters such as the heat loss coefficient and actively writes the new values ​​back to the knowledge graph to update the attributes of the corresponding entities or relationships.

[0059] (3) Diagnostic Reasoning and Report Generation Module: This module serves as the system's application interface and value output, providing transparent diagnostic services. Its functions include:

[0060] Real-time monitoring and anomaly detection: Continuously monitor key energy efficiency indicators obtained from knowledge graphs or data streams, and use methods such as statistical process control or threshold judgment to automatically identify anomalies and trigger diagnostic processes.

[0061] Explainable Inference Engine: Upon receiving an anomaly trigger signal, the engine starts from the anomaly indicator node and performs reverse graph traversal and multi-hop inference on the current version of the dynamic knowledge graph. Based on the relationship types in the graph, especially causal relationships, multiple possible root cause tracing paths are constructed. The engine combines the latest real-time state data of each node entity on the path from the data fusion module and the weights of the relationship edges, and uses algorithms based on Bayesian networks or confidence propagation to calculate the global probability or confidence score of each potential root cause at the end of each path.

[0062] Structured Report Generator: Encapsulates the results of the inference engine to generate easy-to-understand diagnostic reports. The reports not only list the most likely root causes and their confidence rankings, but more importantly, they clearly present the complete and traceable causal chain from the anomaly to each root cause, along with the theoretical, actual, and deviation values ​​of each node in the chain, thus achieving complete visualization and verifiability of the diagnostic conclusions.

[0063] Example 2

[0064] like Figure 2 As shown, this invention provides a method for constructing a knowledge graph for energy efficiency diagnosis in fiberglass production that integrates industrial mechanisms and data-driven approaches. The method includes the following steps:

[0065] Step 1: Construct an ontology model and initial knowledge graph for the fiberglass production field.

[0066] Based on the fiberglass production process, equipment principles, and the law of conservation of energy, a structured knowledge graph ontology model is constructed, specifically including:

[0067] (1) Define the equipment layer entities and their attributes: Define key equipment such as smelting furnace, wire drawing machine, and drying furnace as entities. Each entity contains static attributes such as equipment model and rated power, and dynamic attributes such as real-time power and operating status.

[0068] (2) Define the process layer entities and relationships: Define the core processes such as glass melting, wire drawing, drying and curing as entities, and establish the material flow, energy flow and temporal logic relationships between processes;

[0069] (3) Define the relationship between energy efficiency indicators and mechanisms: Define energy efficiency indicators such as unit product energy consumption, thermal efficiency, and overall equipment efficiency as entities, and associate and bind them with underlying equipment and process parameters through predefined mathematical formulas, for example: SEC 拉丝机 = P 实时功率 × t 运行时间 / Q 合格产量 The computational logic is stored in the knowledge graph as "computational relation" edges, forming a computable semantic network.

[0070] Step 2: Establish a multi-source heterogeneous data access and preprocessing channel

[0071] It interfaces with manufacturing execution systems, data acquisition and monitoring systems, and various sensor networks to achieve real-time acquisition and preprocessing of production data, specifically including:

[0072] (1) Data access: Real-time acquisition of equipment operating parameters, energy consumption data, process setpoints and environmental monitoring data through standard interfaces such as OPC UA and REST API;

[0073] (2) Data cleaning and alignment: The original data is processed by filling missing values, correcting outliers, and aligning timestamps to ensure that the data quality meets the requirements of subsequent analysis;

[0074] (3) Data formatting: Convert the processed data into a standardized format that conforms to the entity-attribute model of the knowledge graph, providing a data foundation for the real-time updating of the knowledge graph.

[0075] Step 3: Implement a dual-channel dynamic evolution mechanism for knowledge graphs

[0076] Based on real-time data streams, the knowledge graph is driven to achieve autonomous evolution and adaptive updates, comprising two parallel and collaborative evolutionary pathways:

[0077] (1) Implicit Relationship Mining and Channel Supplementation: Improved temporal association rule mining algorithms, such as temporal FP-Growth or graph neural network methods, are used to analyze the statistical associations between variables in historical and real-time data. When a new association is found where both support and confidence exceed preset thresholds, such as support > 0.3 and confidence > 0.7, and it is not defined by the mechanism relationship in the ontology, a "data-driven relationship" edge with confidence weight (such as "ambient humidity → drying oven gas consumption") is automatically generated and integrated into the knowledge graph after manual review or automatic rule verification.

[0078] (2) Online calibration channel for mechanism model parameters: For mechanism models stored in the knowledge graph that contain adjustable parameters such as the heat loss coefficient in the kiln thermal efficiency model, a recursive parameter estimation algorithm is used for real-time calibration. The algorithm takes real-time observed data such as input power and actual discharge temperature as input, dynamically estimates the optimal values ​​of the model parameters, and writes the updated parameter values ​​back to the knowledge graph in real time, so that the mechanism model always reflects the real state of the equipment.

[0079] Step 4: Real-time energy efficiency monitoring and anomaly detection based on dynamic knowledge graph

[0080] By utilizing the constructed dynamic knowledge graph, continuous monitoring of energy efficiency in the production process is achieved, specifically including:

[0081] (1) Real-time indicator calculation: Based on the calculation logic stored in the knowledge graph and combined with real-time data, the current value of each energy efficiency indicator is automatically calculated;

[0082] (2) Anomaly detection: Statistical process control methods or adaptive threshold algorithms are used to monitor key energy efficiency indicators. When the indicator value exceeds its dynamic control range for multiple consecutive sampling points, the system automatically marks the anomaly and triggers the diagnostic reasoning process.

[0083] Step 5: Perform interpretable multi-hop causal diagnostic reasoning

[0084] When an energy efficiency anomaly is detected, transparent diagnostic reasoning based on a knowledge graph is initiated, specifically including:

[0085] (1) Reasoning starting point location: Starting from the abnormal energy efficiency index node, locate it in the knowledge graph;

[0086] (2) Multi-path causal tracing: Perform reverse breadth / depth traversal along the "influencing factors", "cause" and other relational edges in the knowledge graph to enumerate all possible causal paths;

[0087] (3) Confidence propagation calculation: For each node on the causal path, an initial confidence score is assigned based on the degree of deviation between its current real-time data and the normal state. Then, a confidence propagation algorithm based on Bayesian networks is used to aggregate the confidence scores along the causal chain to obtain the global confidence score of the terminal potential root cause;

[0088] (4) Root cause hypothesis ranking and pruning: The reasoning path is pruned according to the preset depth threshold and confidence threshold, and the remaining root cause hypotheses are sorted in descending order according to the global confidence.

[0089] Step 6: Generate a structured and interpretable diagnostic report and provide feedback for optimization.

[0090] Based on the reasoning results, a diagnostic report that is easy to understand and act upon is generated, forming a knowledge loop, specifically including:

[0091] (1) Automatic report generation: The system automatically generates a structured diagnostic report, which includes: anomaly description, Top-K root cause list, complete causal chain corresponding to each root cause, and real-time value / theoretical value / deviation value of each node in the chain;

[0092] (2) Recommendations: Retrieve maintenance or operation optimization suggestions related to the diagnosed root cause from the "Expert Knowledge Base" subgraph of the knowledge graph, such as "It is recommended to check the bearing lubrication status" and "It is recommended to adjust the wire drawing speed setting".

[0093] (3) Knowledge feedback and iteration: The diagnostic cases, including triggering conditions, reasoning paths, and verification results, are used as new knowledge samples. After evaluation and confirmation, they can be selectively transformed into rules or cases to supplement the knowledge graph, so as to realize the continuous accumulation of diagnostic knowledge and the self-improvement of system performance.

[0094] The above six steps constitute a complete technical closed loop from knowledge construction, data fusion, dynamic evolution, monitoring and early warning, diagnostic reasoning to feedback optimization. Steps 1-3 focus on the construction and evolution of the knowledge graph, while steps 4-6 focus on graph-based intelligent applications. The steps are closely coupled through data flow and knowledge flow, jointly achieving deep integration of mechanisms and data, as well as the interpretability and adaptability of the energy efficiency diagnosis process.

[0095] Example 3

[0096] This invention establishes a standardized semantic framework in the field of energy efficiency in glass fiber production, forming the skeleton of the framework. Specifically, it includes:

[0097] Step S110: Equipment Layer Ontology Construction. Define all critical energy-consuming equipment on the production line as entities, such as Electric_Furnace_01, Drawing_Machine_05. Each entity has attributes, including:

[0098] Static attributes: Device ID, model, rated power, design efficiency curve, manufacturer, etc.

[0099] Dynamic attributes: real-time operating status such as running, stopping, and alarm; key process parameters such as kiln temperature and wire drawing machine speed; and real-time energy consumption values ​​such as active power and cumulative power consumption.

[0100] Relationships: Partly located between equipment and workshop, and material / energy flow connections between equipment.

[0101] Step S120: Construction of the process layer entity. Define the core production processes as entities (e.g., Glass_Melting, Fiber_Drawing, Drying). Each process entity contains:

[0102] Input / Output: Define material flows (e.g., Raw_Batch input, Molten_Glass output) and energy flows (Electricity_Input, Heat_Loss).

[0103] Process parameters: Key control variables for this process, such as melting temperature, wire drawing speed, and drying temperature.

[0104] Timing and logical relationships: Define the relationships between processes, such as predecessor (Glass_Melting precedes Fiber_Drawing) and parallelism.

[0105] Step S130: Energy Efficiency Layer Ontology Construction. Define energy efficiency evaluation index entities, such as Specific_Energy_Consumption, Thermal_Efficiency, and OEE. The key to this invention lies in strongly associating index entities with equipment attributes and process parameter entities through predefined mathematical models containing inherent mechanisms / empirical data. For example:

[0106] The formula for calculating Specific Energy Consumption of Drawing Machine 05 is as follows: This formula links energy efficiency metrics to equipment energy consumption readings (from the equipment layer) and production data (from the MES, linked to the process layer).

[0107] The formula for calculating the thermal efficiency of electric furnace 01 is: η = / Theoretical_Heat_Required can be calculated based on the feeding amount, glass specific heat capacity, and other mechanism formulas, and is related to the material input of the process layer and the energy input of the equipment layer.

[0108] Mathematical relationships are stored as "attributes" or "relationships" in the knowledge graph, so that indicators are no longer isolated values, but derived nodes with clear computational paths to the underlying data.

[0109] This invention brings static knowledge graphs to life, enabling them to dynamically evolve with production data. Specifically, it includes:

[0110] Step S210: Implicit Relationship Mining and Completion. The system periodically extracts multivariate time-series data within a time window from the historical data warehouse.

[0111] Methods: Improved temporal association rule mining algorithms such as FP-Growth or graph neural network link prediction models are used.

[0112] Process: Algorithm analysis of data such as "Environment.humidity" and "Oven fuel consumption per unit area". "Two columns of data. If the support and confidence exceed the preset thresholds, such as support > 0.3 and confidence > 0.7, and the relationship does not exist in the existing knowledge graph."

[0113] Action: The system automatically generates a new candidate relationship edge: Environment.humidity -- significant influence --> Drying_Furnace.gas_SEC, and assigns a confidence weight. After online confirmation by knowledge engineers or automatic review according to preset rules, this candidate relationship is officially added to the graph, enriching the original mechanism-based knowledge network.

[0114] Step S220: Online calibration of mechanistic model parameters. The mechanistic formulas stored in the knowledge graph, such as Heat_Transfer_Coefficient in the thermal efficiency formula, contain adjustable parameters, and the design values ​​will drift as the equipment ages.

[0115] Method: A recursive parameter estimation algorithm, such as extended Kalman filter or recursive least squares, is used.

[0116] Process: Taking the kiln thermal efficiency model as an example. The system uses real-time collected data such as "input electrical energy," "actual output," and "output temperature" as observations. EKF uses the kiln thermal balance equation stored in the knowledge graph as the state equation and estimates the "comprehensive heat loss coefficient" to be calibrated as the state variable.

[0117] Action: The algorithm outputs the latest estimate of the parameters in each sampling period. The system uses this new value to dynamically update the attribute `current_heat_loss_coefficient` of the corresponding entity `lectric_Furnace_01` in the knowledge graph. This allows the mechanistic model in the graph to reflect the true health status of the equipment in real time, keeping the diagnostic foundation up-to-date.

[0118] Step S230: Graph Version Management and Snapshots. All new relationships and parameter updates generated through steps S210 and S220 are recorded, and historical snapshots of the graph can be generated based on timestamps to trace the knowledge evolution process.

[0119] This invention utilizes dynamic knowledge graphs for interpretable diagnosis. Specifically, it includes:

[0120] Step S310: Anomaly Detection and Inference Trigger. The real-time monitoring module detects that the "unit product power consumption of wire drawing process A" exceeds its dynamic threshold for three consecutive points, triggering the inference engine.

[0121] Step S320: Multi-path probabilistic reasoning. The reasoning engine starts with anomaly indicator nodes and performs a directed search in the knowledge graph.

[0122] Causal chain unrolling: The engine retrieves all "influencing factor" relationship edges pointing to the anomalous node. For example, "Drawing_SEC_A is abnormally high" may be affected by "Drawing_Machine_A.efficiency is reduced" and "Feeding_Glass_Temperature is insufficient".

[0123] Evidence collection and probability calculation: For each potential cause node (e.g., Drawing_Machine_A.efficiency), the engine:

[0124] Check its current status, such as whether the real-time efficiency calculation value is too low.

[0125] Continue upstream along this cause node to find deeper underlying reasons, such as low efficiency possibly due to excessive bearing_vibration or insufficient cooling_water_flow.

[0126] Assign a "local confidence score" based on the deviation of real-time data to the node state on each causal path, and perform propagation and aggregation calculations along the path. Use Bayesian network inference or simple weight superposition to finally obtain the "global confidence score" of each possible root cause at the end, such as "bearing vibration exceeds the standard".

[0127] Path pruning and sorting: The inference path is pruned based on preset search depth and confidence thresholds. Finally, all terminal root cause hypotheses are sorted in descending order of their global confidence scores.

[0128] Step S330: Diagnostic Report Generation. The system generates a structured diagnostic report, including:

[0129] Anomaly Description: The triggered anomaly indicator and the degree of deviation.

[0130] Top-K root cause list: Possible root causes sorted by confidence level, with each root cause accompanied by its complete causal path: Drawing_SEC_A too high ← Drawing_Machine_A.efficiency low ← Bearing_vibration exceeding the limit.

[0131] Supporting evidence: List the real-time values ​​of each node in the causal path and compare them with the normal range.

[0132] Optimization suggestion: Retrieve maintenance or operation suggestions related to the diagnosed root cause from the "Expert Knowledge Base" subgraph of the knowledge graph, such as suggesting bearing inspection and lubrication.

[0133] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for constructing a knowledge graph for energy efficiency diagnosis in fiberglass production that integrates industrial mechanisms and data-driven approaches, characterized in that, The method includes the following steps: Step 1: Construct an ontology of energy efficiency knowledge graph for fiberglass production that integrates industrial mechanisms. Based on the fiberglass production process, equipment principles, and energy balance relationships, define a standardized ontology model. Step 1-1: Equipment layer entity construction, defining key equipment such as smelting furnace and wire drawing machine as entities, and encapsulating their static attributes, dynamic operating parameters and energy consumption data; Steps 1-2: Construct the process layer body, define the core processes of melting, drawing, and drying as entities, and define the material / energy flow input and output relationships and timing logic relationships between processes; Steps 1-3: Energy efficiency layer ontology construction. Define indicators such as unit product energy consumption and thermal efficiency as entities. Through predefined mechanisms or empirical formulas, i.e., unit product power consumption = total power consumption of process / qualified product output, strongly calculate and associate the energy efficiency indicator entities with the dynamic attributes of the equipment layer and the parameter entities of the process layer. Store these mathematical relationships as the calculation relationship edges between entities in the knowledge graph to form a computable knowledge network skeleton. Step 2: Drive the knowledge graph to achieve dual-channel dynamic evolution, connect to real-time data sources, and enable the static graph to have learning and updating capabilities, including two parallel and collaborative evolution channels; Step 2-1: Implicit knowledge mining channel. Periodically, high-confidence statistical associations between variables are discovered from historical / real-time data using time-series association rule mining or graph neural network methods. If the association is not defined by the mechanistic knowledge in Step 1 above, it is treated as a new relation edge with confidence weight, submitted for review, and then integrated into the knowledge graph to supplement the blind spots of the mechanistic model. Step 2-2: Mechanism Model Parameter Calibration Channel. For the mechanism formula defined in Step 1, which includes adjustable parameters such as heat transfer coefficient and efficiency decay factor, real-time production data is used as the observation sequence. Recursive parameter estimation algorithms such as extended Kalman filter are used to estimate these key parameters online and in a rolling manner. Subsequently, the parameter values ​​bound to the corresponding entities in the knowledge graph are dynamically updated so that the mechanism model embedded in the graph always tracks the real state of the equipment. Step 3: Perform interpretable energy efficiency diagnosis reasoning based on dynamic knowledge graphs. When abnormal energy efficiency indicators are detected, initiate the diagnosis process. Step 3-1: Anomaly localization and inference triggering: Set the anomaly indicator node as the starting point for inference; Step 3-2: Multi-hop causal tracing and probability calculation: Centered on the starting point, perform a reverse breadth / depth search along the relationship edges of influencing factors in the knowledge graph to enumerate all possible causal paths. For each node on the path, calculate a local anomaly probability based on its current real-time data status. Then, based on Bayes' theorem or confidence propagation algorithm, aggregate the probabilities along the causal path to calculate the global confidence of each terminal root cause hypothesis, such as "a certain bearing vibration exceeds the standard". Step 3-3: Generate an interpretable diagnostic report: Sort the root cause hypotheses by confidence level and output a structured report. The report not only lists the Top-K root causes, but more importantly, it clearly shows the complete causal chain from the anomaly to each root cause, and provides real-time data of each node in the chain as evidence, thereby making the diagnostic conclusion verifiable and interpretable.

2. The method for constructing a knowledge graph for energy efficiency diagnosis of fiberglass production that integrates industrial mechanisms and data-driven approaches, as described in claim 1, is characterized in that... Step 1, based on the fiberglass production process, equipment principles, and the law of conservation of energy, constructs a structured knowledge graph ontology model, specifically including: (1) Define the equipment layer entities and their attributes: Define key equipment such as smelting furnace, wire drawing machine, and drying furnace as entities. Each entity contains static attributes such as equipment model and rated power, and dynamic attributes such as real-time power and operating status. (2) Define the process layer entities and relationships: Define the core processes such as glass melting, wire drawing, drying and curing as entities, and establish the material flow, energy flow and temporal logic relationships between processes; (3) Define the relationship between energy efficiency indicators and mechanisms: Define energy efficiency indicators such as unit product energy consumption, thermal efficiency, and overall equipment efficiency as entities, and associate and bind them with underlying equipment and process parameters through predefined mathematical formulas, for example: SEC 拉丝机 = P 实时功率 × t 运行时间 / Q 合格产量 The computational logic is stored in the knowledge graph as "computational relation" edges, forming a computable semantic network.

3. The method for constructing a knowledge graph for energy efficiency diagnosis of fiberglass production that integrates industrial mechanisms and data-driven approaches, as described in claim 1, is characterized in that... Step 2 involves connecting to the manufacturing execution system, data acquisition and monitoring system, and various sensor networks to perform real-time acquisition and preprocessing of production data, specifically including: (1) Data access: Real-time acquisition of equipment operating parameters, energy consumption data, process setpoints and environmental monitoring data through standard interfaces such as OPC UA and REST API; (2) Data cleaning and alignment: The original data is processed by filling missing values, correcting outliers, and aligning timestamps to ensure that the data quality meets the requirements of subsequent analysis; (3) Data formatting: Convert the processed data into a standardized format that conforms to the entity-attribute model of the knowledge graph, providing a data foundation for the real-time updating of the knowledge graph.

4. The method for constructing a knowledge graph for energy efficiency diagnosis of fiberglass production that integrates industrial mechanisms and data-driven approaches, as described in claim 1, is characterized in that... Step 3, based on real-time data streams, drives the knowledge graph to achieve autonomous evolution and adaptive updates. It includes two parallel and collaborative evolutionary pathways, specifically: (1) Implicit Relationship Mining and Supplementing Channels: Using improved temporal association rule mining algorithms such as temporal FP-Growth or graph neural network methods, the statistical associations between variables in historical and real-time data are analyzed. When the support and confidence both exceed the preset thresholds, such as support > 0.3 and confidence > 0.7, and there is no new association defined by the mechanism relationship in the ontology, a "data-driven relationship" edge with confidence weight is automatically generated, namely "environmental humidity → drying oven gas consumption". After manual review or automatic rule verification, it is integrated into the knowledge graph. (2) Online calibration channel for mechanism model parameters: For mechanism models stored in the knowledge graph that contain adjustable parameters such as heat loss coefficient in the kiln thermal efficiency model, a recursive parameter estimation algorithm is used for real-time calibration. The algorithm takes real-time observation data such as input power and actual discharge temperature as input, dynamically estimates the optimal value of model parameters, and writes the updated parameter values ​​back to the knowledge graph in real time, so that the mechanism model always reflects the real state of the equipment.

5. The method for constructing a knowledge graph for energy efficiency diagnosis of fiberglass production that integrates industrial mechanisms and data-driven approaches, as described in claim 1, is characterized in that... Step 4 utilizes the constructed dynamic knowledge graph to continuously monitor the energy efficiency of the production process, specifically including: (1) Real-time indicator calculation: Based on the calculation logic stored in the knowledge graph and combined with real-time data, the current value of each energy efficiency indicator is automatically calculated; (2) Anomaly detection: The system uses statistical process control or adaptive threshold algorithm to monitor key energy efficiency indicators. When the indicator value exceeds its dynamic control range for multiple consecutive sampling points, the system automatically marks the anomaly and triggers the diagnostic reasoning process.

6. The method for constructing a knowledge graph for energy efficiency diagnosis of fiberglass production that integrates industrial mechanisms and data-driven approaches, as described in claim 1, is characterized in that... When an energy efficiency anomaly is detected in step 5, transparent diagnostic reasoning based on a knowledge graph is initiated, specifically including: (1) Reasoning starting point location: Starting from the abnormal energy efficiency index node, locate it in the knowledge graph; (2) Multi-path causal tracing: Perform reverse breadth / depth traversal along the "influencing factors", "cause" and other relational edges in the knowledge graph to enumerate all possible causal paths; (3) Confidence propagation calculation: For each node on the causal path, an initial confidence score is assigned based on the degree of deviation between its current real-time data and the normal state. Then, a confidence propagation algorithm based on Bayesian network is used to aggregate the confidence scores along the causal chain to obtain the global confidence score of the terminal potential root cause. (4) Root cause hypothesis ranking and pruning: The reasoning path is pruned according to the preset depth threshold and confidence threshold, and the remaining root cause hypotheses are sorted in descending order according to the global confidence.

7. The method for constructing a knowledge graph for energy efficiency diagnosis of fiberglass production that integrates industrial mechanisms and data-driven approaches, as described in claim 1, is characterized in that... Step 6, based on the reasoning results, generates a diagnostic report that is easy to understand and act upon, forming a knowledge loop, specifically including: (1) Automatic report generation: The system automatically generates a structured diagnostic report, which includes: anomaly description, Top-K root cause list, complete causal chain corresponding to each root cause, and real-time value / theoretical value / deviation value of each node in the chain; (2) Recommendations: Retrieve maintenance or operation optimization suggestions related to the diagnosed root cause from the "Expert Knowledge Base" subgraph of the knowledge graph, such as "It is recommended to check the bearing lubrication status" and "It is recommended to adjust the wire drawing speed setting"; (3) Knowledge feedback and iteration: The diagnostic cases, including triggering conditions, reasoning paths, and verification results, are used as new knowledge samples. After evaluation and confirmation, they can be selectively transformed into rules or cases to supplement the knowledge graph, so as to realize the continuous accumulation of diagnostic knowledge and the self-improvement of system performance.

8. A knowledge graph construction system for energy efficiency diagnosis of fiberglass production that integrates industrial mechanisms and data-driven approaches, characterized in that, The system is a closed-loop intelligent agent integrating hardware and software, including the following three core functional modules, which work closely together through data flow and knowledge flow; The knowledge graph construction and storage module serves as the system's knowledge foundation and static model carrier, and its functions include: Ontology Management: Provides graphical or configurable ontology editing tools, with a built-in general ontology library for equipment, process, and indicator categories in the fiberglass production field. It supports users to instantiate entities according to specific production lines and define static relationships between entities, such as "belongs to", "connected to", and "predecessor to". Mechanism Model Embedding: This is the core innovative function of this module. It allows users to attach predefined mathematical formulas or calculation logic to specific relationships such as the relationship between energy efficiency indicators and process parameters. That is, the "total power consumption" attribute of the "unit product power consumption" entity and the "output" attribute of the "wire drawing machine" entity, and the "output" attribute of the "production batch" entity are bound together by the formula SEC = total power consumption / output. These formulas are stored in the knowledge graph as "computational relationships", which upgrades the graph from a relational network to a semantic model that can perform calculations. Graph storage and query: Persistently store the constructed entities, attributes, relationships and additional mechanistic logic in the graph database, and provide standard Cypher and Gremlin query interfaces for other modules to access; The data fusion and graph evolution module is the intelligent engine and dynamic core of the system. Multi-source data access and processing: Connect and collect heterogeneous data from MES, SCADA, and sensor networks in real time, including time series data, event logs, and test results, and clean, align, and format them to convert them into information that can be digested by the knowledge graph; Implicit Relationship Mining Engine: Runs an improved temporal association rule algorithm, namely the temporal extension or graph representation learning model of FP-Growth, to analyze historical and real-time data and automatically discover strong statistical associations that exist stably between variables. If this association is not covered by the mechanism relationship defined in the ontology, the engine will generate a "data-driven relationship" suggestion with confidence weight, namely "ambient humidity → drying oven gas consumption", which will be automatically or semi-automatically completed into the knowledge graph after review. Mechanism model parameter online calibration engine: For mechanism models with adjustable parameters stored in the knowledge graph, such as the kiln thermal efficiency model, online parameter estimation is performed; Its built-in recursive estimation algorithm framework is the extended Kalman filter or recursive least squares method. The engine obtains the input power and output temperature observation variables from the real-time data stream and combines them with the mechanism formulas retrieved from the knowledge graph; In each calculation cycle, the engine outputs the latest estimates of key parameters such as heat loss coefficient and actively writes the new values ​​back to the knowledge graph to update the attributes of the corresponding entities or relationships. (3) Diagnostic Reasoning and Report Generation Module This module is the system's application interface and value output, responsible for providing transparent diagnostic services, including: Real-time monitoring and anomaly detection: Continuously monitor key energy efficiency indicators obtained from knowledge graphs or data streams, and use methods such as statistical process control or threshold judgment to automatically identify anomalies and trigger diagnostic processes; Explainable reasoning engine: When an anomaly trigger signal is received, the engine starts from the anomaly indicator node and performs reverse graph traversal and multi-hop reasoning on the current version of the dynamic knowledge graph. Based on the relationship types in the graph, especially causal relationships, multiple possible root cause tracing paths are constructed. The engine combines the latest real-time status data of each node entity on the path from the data fusion module and the weight of the relationship edges, and uses an algorithm based on Bayesian network or confidence propagation to calculate the global probability or confidence score of each potential root cause at the end of each path. Structured report generator: It encapsulates the results of the inference engine to generate easy-to-understand diagnostic reports. The reports not only list the most likely root causes and their confidence ranking, but more importantly, they clearly present the complete and traceable causal chain from the anomaly to each root cause, and attach the theoretical value, actual value and deviation of each node in the chain, thereby achieving complete visualization and verification of the diagnostic conclusions.

Images