Industrial software enhancement system with rational number fusion

Abstract

The invention discloses an industrial software enhancement system based on rational number fusion. The system comprises a knowledge graph construction module, a model training module, a context sensing module, a model activation module and an instruction generation module. Constructing a domain knowledge graph in an off-line manner; pre-training a group of artificial intelligence compensation model clusters; during online operation, the system senses working condition characteristics of an industrial site in real time, deep understanding and risk pre-judgment are performed on working conditions by utilizing a knowledge graph, and a most matched compensation model is activated in a targeted manner from a model cluster; and generating and executing a compensation adjustment instruction of physical consistency. According to the method, the advantages of a mechanism model and a data model are effectively fused, the defect that a traditional data driving model possibly outputs a result violating a physical law is overcome, and the safety, the adaptability and the intelligent level of a control decision in a complex industrial scene are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial software and artificial intelligence, and particularly relates to an industrial software enhancement system based on numerical and intelligent fusion. BACKGROUND

[0002] Industrial software is the core driving force of modern high-end manufacturing, energy and chemical industries, and other key industrial fields. At present, industrial processes are developing towards the direction of extreme, complex, and personalized, which brings great difficulties to industrial software used for process control and simulation.

[0003] There are two main technical routes in the prior art. The first is a modeling method based on physical mechanism. This method is based on first principles and describes industrial processes by establishing partial differential equations. The advantage of this method is that it has a solid theoretical foundation and good extrapolation ability, and the decision-making process is interpretable. However, this method also has significant disadvantages: on the one hand, the real industrial environment is full of disturbance factors such as material batch differences, sensor noise, and environmental fluctuations that are difficult to model accurately. Mechanism models often need to make idealized assumptions, which leads to a significant decrease in prediction accuracy when dealing with these uncertainties, resulting in model distortion problems. On the other hand, high-fidelity mechanism models are computationally intensive and difficult to meet the needs of real-time online control and optimization in industrial sites.

[0004] The second is a data-driven modeling method based on artificial intelligence. This method uses machine learning or deep learning algorithms to learn the nonlinear relationship between process parameters and process results from a large amount of historical data. The advantage of this method is that it is fast and can fit highly complex nonlinear relationships, making it suitable for real-time applications. However, it also has inherent limitations: on the one hand, the reliability of data-driven models depends entirely on the quality and coverage of the training data. When encountering new conditions that have not appeared in the training set, the model may output prediction results that seriously violate basic physical laws such as conservation of energy and conservation of mass, introducing risks in industrial production systems. On the other hand, data-driven models are not transparent in their internal working mechanism, making their decision-making process difficult to understand and trust, and their generalization ability is poor. When production processes change, expensive data collection and model training are usually required.

[0005] This situation is particularly prominent in precision manufacturing processes in the aerospace, high-end materials, and other fields. For example, the high-temperature alloy smelting process required for aircraft engine turbine blades needs to precisely control the uniformity of alloy composition and trace elements. However, small fluctuations in impurity composition in different batches of raw materials, especially recycled materials, can cause unpredictable disturbances to the smelting process, such as key element loss or segregation.

[0006] This poses a dilemma for existing industrial control software: if the underlying model on which the software relies is a mechanism model based on ideal assumptions, it is difficult to accurately describe such complex disturbances, resulting in insufficient control accuracy; if the software uses a purely data-driven model, its decision-making results may violate basic physical principles such as the laws of thermodynamics when faced with new batches of raw materials, outputting incorrect feeding or heating instructions, introducing significant risks in terms of safety and quality. SUMMARY

[0007] To solve the above problems, the present application discloses an industrial software enhancement system based on the fusion of physical and numerical models, wherein the "fusion of physical and numerical models" refers to a technical framework that deeply combines models based on physical mechanisms (referred to as "physical") and data-driven artificial intelligence models (referred to as "numerical"). This method builds a collaborative intelligent closed loop of physical guiding numerical, physical integrating numerical, and numerical feeding back physical, interacts with existing industrial control software in terms of data and instructions, and realizes enhanced control of industrial processes.

[0008] The present application provides an industrial software enhancement system based on the fusion of physical and numerical models, which is used to interact with an industrial control system in terms of data and instructions. The industrial software enhancement system comprises: a knowledge graph construction module, which is used to formalize and structure the knowledge in the industrial field in the offline preparation stage, to construct a domain knowledge graph containing a reasoning engine; a model training module, which is used to train a group of special artificial intelligence compensation models for a variety of typical working conditions in the industrial field in the offline preparation stage, to form a cluster of artificial intelligence compensation models, and to embed physical mechanism constraints into the training process of each special artificial intelligence compensation model; a context perception module, which is used to collect real-time data from the industrial site, extract structured working condition features therefrom, and convert them into semantic queries understandable by the domain knowledge graph, when the system is running online; a model activation module, which is used to input the semantic queries into the domain knowledge graph, use the reasoning engine contained in the domain knowledge graph to reason, and select and target activate one or a group of artificial intelligence compensation models from the cluster of artificial intelligence compensation models according to the reasoning results; an instruction generation module, which is used to receive real-time sensor data as input from the activated artificial intelligence compensation model, calculate a group of compensation adjustment instructions for the industrial control system, and send them to the actuator in the industrial control system through the output interface of the industrial software enhancement system, to regulate the industrial process.

[0009] The application also provides a method applied to the above-mentioned industrial software enhancement system for data and instruction interaction with an industrial control system, and the method specifically comprises the following steps: step one, formalization of domain knowledge and construction of a graph. In the offline preparation stage of the method, first principles, empirical formulas, expert rules, equipment manuals, process flows and other unstructured and semi-structured knowledge in a specific industrial field are formalized and structured to construct a domain knowledge graph. The domain knowledge graph at least contains core entities such as materials, equipment and process parameters; entity attributes such as the thermal conductivity and viscosity of materials; and relationships between entities, which specifically include causal relationships defined based on physical laws, correlation relationships defined based on statistics or experience, classification relationships of entity attribution, constraint relationships of parameter value ranges, and mapping relationships between working conditions and strategies for subsequent decision-making.

[0010] Step two, pre-training of an artificial intelligence model cluster with physical constraints. Also in the offline preparation stage of the method, a group of special artificial intelligence compensation models are designed and trained for a variety of typical working conditions that may occur in the field, such as "start-up phase", "steady operation", "emergency shutdown", etc., thereby forming an artificial intelligence compensation model cluster. In the training process, the physical mechanism constraint is deeply embedded into the training process of the artificial intelligence model. Specifically, the total loss function L_total of the model is designed to be composed of a data-driven loss term L_data and a physical mechanism residual loss term L_phys, which can be expressed as: L_total = λ_data L_data + λ_phys L_phys. Wherein, L_data is used to measure the gap between the model prediction value and the real data label, L_phys is to take the residual form of the core control equation (such as the energy conservation equation, the mass conservation equation) in this field as a penalty term, if the model output does not satisfy the physical equation, then L_phys increases, thereby correcting the model weight in back propagation and forcing the model to learn the physically valid solution. λ_data and λ_phys are preset weight coefficients.

[0011] Step three, context perception and understanding of real-time working conditions. In the online operation of the system, sensor data from the industrial field is collected at a high frequency through a data interface, and combined with the current production task information being executed. The system uses signal processing and pattern recognition technologies to extract structured working condition features from the raw data, and converts them into semantic queries understandable by the domain knowledge graph.

[0012] Step four, model targeted activation based on knowledge graph reasoning. The working condition query generated in the previous step is input into the domain knowledge graph module. Using the reasoning engine of the knowledge graph, first, the most matching entities and relationships in the current working condition characteristics are retrieved in the graph, then causal or association reasoning is performed to predict the subsequent state or potential risks of the current working condition, and finally, according to the pre-set "working condition-strategy mapping relationship" in step one, one or a group of artificial intelligence compensation models most suitable for processing the current specific working condition are accurately selected and targeted activated from the artificial intelligence compensation model cluster constructed in step two.

[0013] Compared with the model switching method based on simple condition judgment in the prior art, the model targeted activation mechanism based on knowledge graph reasoning proposed by the present application utilizes the knowledge graph to perform deep semantic understanding on the real-time working condition, and directly maps the reasoning result to the most suitable artificial intelligence model, so that the system can accurately call the optimal solution in the model cluster according to the subtle changes of the industrial process, and realize higher level adaptive control, and improve the pertinence and effectiveness of control.

[0014] Step five, compensation instruction generation and execution of physical consistency. The activated artificial intelligence compensation model receives real-time sensor data as input and calculates one or a group of compensation adjustment instructions for the existing control system. Since the model has been injected with physical constraints during training, the output instructions naturally ensure that the core physical laws are not violated, thereby ensuring the reliability and safety of the decision. These compensation instructions are sent to the lower execution mechanism through the output interface for real-time and fine regulation and control of the industrial process.

[0015] Step six, solidification of experience chain and closed-loop evolution of knowledge. The system continuously records each intervention process to form experience chain data containing time stamp, working condition characteristics, graph reasoning path, activated artificial intelligence compensation model, output compensation instruction, process state feedback and final result. The knowledge evolution module periodically or after a specific event, uses algorithms such as causal inference and association rule mining to analyze the accumulated experience chain offline to refine new knowledge.

[0016] The present application also proposes a knowledge graph closed-loop self-evolution method driven by experience chain. This method realizes the transformation from data to information and then to knowledge through structured experience chain data. When the algorithm analysis finds a new knowledge with high confidence, for example, a new and better "working condition-strategy mapping relationship", the new knowledge will be verified and automatically and structuredly updated back to the domain knowledge graph in step one. This self-learning mechanism enables the knowledge base and decision-making ability of the system to be continuously and automatically iterated and optimized as the running time elapses and data accumulates.

[0017] Further, to ensure the robustness and reliability of the knowledge evolution process, a new knowledge confidence assessment and fusion module can be added in step six. When the knowledge evolution module refines a candidate knowledge from the experience chain, the system first assigns an initial confidence score to the candidate knowledge based on factors such as the number of experience chains, the consistency of results, etc., and stores it in a temporary observation knowledge base. In future runs, the system can use an exploratory strategy to try out the candidate knowledge with a certain probability and continuously monitor its application effect. If multiple verifications produce good results, the confidence score of the candidate knowledge is raised accordingly. When the score exceeds a pre-set knowledge solidification threshold, the candidate knowledge is officially and automatically incorporated into the main knowledge graph. Alternatively, for candidate knowledge that conflicts with existing knowledge, the system can also initiate a knowledge arbitration process assisted by experts, thereby avoiding the risk of knowledge base pollution caused by a small amount of abnormal data.

[0018] Another aspect of the present application also provides an industrial software enhancement system for the integration of reasoning and numerical methods, which comprises: a knowledge graph construction module for performing step one described above; a model training module for performing step two described above; a context awareness module for performing step three described above; a model activation module for performing step four described above; an instruction generation module for performing step five described above; and a knowledge evolution module for performing step six described above.

[0019] Compared with the prior art, the present application has the following beneficial effects: by introducing a physical mechanism residual loss term in the training of artificial intelligence models, it is ensured that any control instructions output by the system comply with the basic physical laws, effectively improving the reliability and safety of industrial software decision-making, and overcoming the defect that traditional data-driven models may produce serious accident error outputs when facing new working conditions; by constructing a domain knowledge graph and using its reasoning ability to deeply understand real-time working conditions, targeted activation of special artificial intelligence models in the model cluster is realized, so that the system can call the most suitable solution for the rapidly changing working conditions, thereby realizing adaptive and fine control of complex industrial processes and improving control precision and efficiency; and by recording and analyzing experience chains, a closed-loop self-evolution ability of system knowledge is constructed, which can extract and refine new high-value domain knowledge from its own running experience and automatically update it to the knowledge graph, realizing continuous self-learning and optimization of the knowledge base and decision-making ability, and converting implicit expert experience into explicit and structured knowledge. BRIEF DESCRIPTION OF DRAWINGS

[0020] The embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0021] Figure 1 is a schematic diagram of the overall architecture and interaction of an industrial software enhancement system for the integration of reasoning and numerical methods according to an embodiment of the present application.

[0022] Figure 2 is a structural schematic diagram of a domain knowledge graph according to an embodiment of the present application.

[0023] Figure 3 is a principle schematic diagram of a physical constraint artificial intelligence model training process according to an embodiment of the present application.

[0024] Figure 4 is an empirical chain-based knowledge closed loop evolution flowchart according to an embodiment of the present application.

[0025] Figure 5 is a schematic diagram of the influence of physical constraints on model parameter optimization paths according to an embodiment of the present application. DETAILED DESCRIPTION

[0026] The present application will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0027] Embodiment 1 The present embodiment provides a method for enhancing industrial software by fusing logic and numbers, or a method applied to an industrial software enhancement system for fusing logic and numbers, with reference to Figure 1 The method can be applied to an industrial software enhancement system that interacts with existing industrial control software (for example, a distributed control system DCS or a programmable logic controller PLC). The method specifically includes the following steps: Step S1, formalization of domain knowledge and construction of a graph, in the offline preparation stage, formalize and structure the industrial domain knowledge to construct a domain knowledge graph containing a reasoning engine.

[0028] In the offline preparation stage of the method, formalize and structure the unstructured and semi-structured knowledge such as first principles, empirical formulas, expert rules, equipment operation manuals, process flow files, etc. of a specific industrial domain (for example, the industrial domain of polymer material polymerization reaction) to construct a domain knowledge graph.

[0029] Specifically, this step can be decomposed into the following sub-steps: Step S101, knowledge extraction. Using natural language processing (NLP) techniques, such as named entity recognition (NER) and relation extraction (RE) models, automatically extract core entities, entity attributes, and their relationships from textual device manuals, process procedures, and other documents. For example, from the operation manual of a polymerization reactor, we can extract “reactor R-101” (device entity), “stirring rate” (process parameter entity), “catalyst A” (material entity), and their “influence” (association relationship) and “set range 50-100 RPM” (constraint relationship).

[0030] Step S102, knowledge representation and storage. The extracted knowledge is represented in the form of triples “entity-relation-entity” or “entity-attribute-attribute value” and stored in a graph database (such as Neo4j or JanusGraph database software). Referring to Figure 2 , the domain knowledge graph 200 at least contains: Core entities 201, such as materials (e.g., monomers, initiators), devices (e.g., reactors, pumps), and process parameters (e.g., temperature, pressure); Entity attributes 202, such as the thermal conductivity and viscosity of materials; And the relationship 203 between entities, which includes: causal relationship 203a based on physical laws, such as temperature rise leading to accelerated reaction rate; association relationship 203b based on statistics or experience, such as “catalyst A batch number is batch X” associated with “product viscosity is high”; classification relationship 203c of entity attribution, such as “R-101” belongs to batch reactor; constraint relationship 203d of parameter value range, such as “reactor jacket temperature” is constrained to “80-120 degrees Celsius”; and mapping relationship 203e between working conditions and strategies for subsequent decision-making, such as “detecting a decrease in monomer conversion” mapping to “increasing the initiator injection rate” compensation strategy.

[0031] In this way, scattered domain knowledge is converted into a structured network that can be understood and reasoned by machines, providing a knowledge base for subsequent intelligent decision-making.

[0032] Step S2, pre-training of physical constraint artificial intelligence model cluster. In the offline preparation stage, a set of dedicated artificial intelligence compensation models are trained for various typical working conditions in the industry, forming an artificial intelligence compensation model cluster, and embedding physical mechanism constraints into the training process of each dedicated artificial intelligence compensation model.

[0033] Likewise in the offline preparation stage of the method, a group of special artificial intelligence compensation models are designed and trained for various typical working conditions that may occur in the field, such as the "warm-up start-up phase", "stable reaction phase", "viscosity sudden increase abnormal working condition", "product grade switching phase" and the like in the polymerization reaction process, thereby forming a cluster of artificial intelligence compensation models.

[0034] With reference to Figure 3 In the training process, the physical mechanism constraint 301 is deeply embedded into the training process of the artificial intelligence model. Specifically, the total loss function L_total of the model is designed to be composed of a data-driven loss term L_data and a physical mechanism residual loss term L_phys weighted. Its form can be expressed as: L_total = λ_data L_data + λ_phys L_phys.

[0035] L_data is used to measure the gap between the model prediction value and the real data label, and can use conventional loss functions such as Mean Squared Error (MSE) or cross-entropy. For example, the difference between the model-predicted reaction endpoint viscosity and the real viscosity measured in the laboratory.

[0036] L_phys is to transform the core control equation in this field, such as the energy conservation equation 302 or the mass conservation equation 303 based on polymerization reaction kinetics, into its residual form as a penalty term. For example, for energy conservation, the residual can be expressed as |(generated heat + input heat) - (output heat + accumulated heat)|. If the model output (such as the predicted temperature, reaction rate) is substituted into the physical equation, the equation will be unbalanced, and the absolute value of L_phys will increase, thereby correcting the model weight during back propagation 304, forcing the model to learn parameters that not only fit historical data, but also meet basic physical laws.

[0037] λ_data and λ_phys are preset weight coefficients for balancing data fitting and physical consistency, and their value range can be between 0.1 and 10, and the optimal value is determined by cross-validation and other methods.

[0038] Optionally, the specific implementation of the artificial intelligence compensation model can be a long short-term memory network (LSTM) for processing time-series related process data, or a graph neural network (GNN) for fusing topological structure information of a knowledge graph for prediction.

[0039] The principle of this step is to impose strong constraints from the physical world on the huge solution space of the neural network by introducing L_phys, avoiding it from producing solutions that violate physical laws in areas not covered by the training data, thereby significantly improving the model's generalization ability and reliability in safety-critical scenarios.

[0040] To more intuitively illustrate the technical principle of this step S2, please refer to Figure 5 . Figure 5 In the figure, the contour lines of the background represent the distribution of the data-driven loss L_data, where there are multiple low valley regions of the loss function, i.e. local optimal solutions. The area enclosed by the gray contour in the figure represents the physically feasible region, i.e. when the model parameters are within this region, its output satisfies the physical mechanism constraints (L_phys ≈ 0). Traditional data-driven models, when training, their parameter optimization path (as shown by the black dot path in the figure) only aims to minimize L_data, so it will converge to the local optimal solution with the lowest data loss, but this solution is outside the physically feasible region, resulting in the model output violating the physical laws (i.e. physical violation). The method of the present application, when training, since the total loss function contains the physical mechanism residual loss term L_phys, its parameter optimization path (as shown by the white triangle path in the figure) will be guided by both L_data and L_phys, and the path is constrained within the physically feasible region for optimization, and finally converges to a physically consistent global optimal solution. In this way, it fundamentally ensures that the model fits the data while its decision-making behavior does not conflict with the basic physical laws of the field.

[0041] Step S3, real-time context perception and understanding, when the system is online, collect real-time data from the industrial site, extract structured operating condition features from it, and convert the operating condition features into semantic queries understandable by the domain knowledge graph.

[0042] When the system is online, high-frequency (e.g. 1Hz to 10Hz) sensor data streams such as temperature, pressure, flow, and component concentration are collected from the industrial site through data interfaces such as OPC-UA or Modbus TCP, and combined with the current production task information (such as product model, production batch) obtained from the manufacturing execution system (MES). The system uses signal processing and pattern recognition techniques to extract structured operating condition features from the raw data.

[0043] Specifically, this step can include: filtering high-frequency noise data, such as using Kalman filtering; trend analysis of key variables, such as calculating their first and second derivatives to judge their change rate and trend; using clustering algorithms (such as K-Means) or classifiers (such as support vector machine SVM) to identify the operating condition category to which the current data pattern belongs.

[0044] For example, the system monitors that the temperature sensor readings rise by 10 degrees Celsius within 5 minutes, and the monomer concentration sensor readings drop rapidly, while the MES information shows that the current is in the "reaction initiation" phase. The system combines these features and converts them into a semantic query that the domain knowledge graph can understand, such as a structured query statement: query the working conditions and their coping strategies related to "fast temperature rise rate", "fast monomer consumption", and "reaction initiation phase".

[0045] Step S4, model targeted activation based on knowledge graph reasoning, input the semantic query generated in step S3 into the domain knowledge graph, use the reasoning engine contained in the domain knowledge graph to reason, and select and target activate one or a group of artificial intelligence compensation models from the artificial intelligence compensation model cluster according to the reasoning result.

[0046] Specifically, input the working condition query generated in the previous step S3 into the domain knowledge graph module, use the reasoning engine of the knowledge graph to perform the following operations: First, retrieve the most matching entities and relationships in the graph according to the current working condition characteristics. For example, match to the working condition node representing "fast temperature rise".

[0047] Then, multi-step reasoning is performed. The reasoning engine traverses along the causal or associated relationship chain in the graph to predict the subsequent state or potential risks of the current working condition. For example, starting from the fast temperature rise node, it may infer the potential risk of explosive polymerization (a common chain polymerization reaction in polymer chemical industry) along the causal relationship chain.

[0048] Finally, according to the pre-set working condition-strategy mapping relationship in step S1, accurately select and target activate one or a group of artificial intelligence compensation models that are most suitable for processing the current specific working condition from the artificial intelligence compensation model cluster constructed in step S2. For example, the reasoning result of potential explosive polymerization risk activates the compensation model in the model cluster that is specifically used for emergency cooling and inhibition of polymerization reaction through the mapping relationship.

[0049] Compared with the model switching method based on simple threshold or isolated rules, the present application uses the knowledge graph to perform deep and global semantic understanding and risk prediction on the real-time working condition, and directly maps the reasoning result to the optimal artificial intelligence compensation model, so that the system can accurately call the special model in the model cluster according to the subtle changes and potential evolution trend of the industrial process, realize more accurate adaptive control of the working condition changes, and improve the pertinence and effectiveness of the control.

[0050] As an alternative, model activation can also be based on similarity calculation. The real-time operating condition feature vector is calculated with the cosine similarity of each model in the model cluster corresponding to the typical operating condition feature vector, and the model with the highest similarity is activated. However, the knowledge graph reasoning method can better handle complex causal relationships, so it has wider applicability.

[0051] Step S5, compensation instruction generation and execution of physical consistency. The artificial intelligence compensation model activated in step S4 receives real-time sensor data as input, calculates a set of compensation adjustment instructions for the industrial control system, and sends them to the actuator in the industrial control system through the output interface of the industrial software enhancement system to regulate the industrial process.

[0052] Specifically, the activated artificial intelligence compensation model (such as the "aggregation reaction emergency cooling and inhibition model") receives real-time sensor data as input and calculates one or a set of compensation adjustment instructions for the existing control system. For example, the output instructions are: "cooling water valve opening degree increases by 20%", "inhibitor injection pump starts, flow rate is set to 5L / min".

[0053] Since the model has been injected with physical constraints during training, the output instructions naturally ensure that they do not violate core physical laws (such as the conservation of energy), avoiding dangerous instructions that are physically contradictory, such as "cooling water valve opening degree increases while heating jacket power also increases", thereby ensuring the reliability and safety of the decision. These compensation instructions are sent to the lower-level actuators (such as PLC) through the output interface to regulate the industrial process in real time and in detail.

[0054] Step S6, solidification of experience chain and closed-loop evolution of knowledge.

[0055] Referring to Figure 4 , the system continuously records each intervention process to form structured experience chain data 407 containing timestamps, operating condition features 401, graph reasoning paths 402, activated artificial intelligence compensation models 403, output compensation instructions 404, process state feedback 405, and final results (such as quality improvement, energy consumption reduction) 406.

[0056] The knowledge evolution module periodically (such as once a day) or after a specific event (such as the end of a batch production), uses algorithms such as causal inference and association rule mining to analyze the accumulated experience chain offline to extract new knowledge 408.

[0057] For example, by analyzing a large amount of experience chain data through association rule mining algorithms (such as the Apriori algorithm), the system can discover a new, pre-defined high-confidence rule by experts: “When the raw material batch is Y and the environmental humidity is greater than 80%, the compensation control is performed using model B, and the product excellent rate is 15% higher than using the default model A”.

[0058] This newly discovered rule, after confidence evaluation and fusion as described below, will be converted into a new “working condition-strategy mapping relationship” and automatically and structurally updated back to the domain knowledge graph 409 in step S1. This constitutes a high-order cognitive learning closed loop, enabling the system’s knowledge base and decision-making ability to be continuously, automatically iterated and optimized over time, and the rules hidden in the operator’s experience and running data are converted into explicit, structured knowledge.

[0059] Further, to ensure the robustness and reliability of the knowledge evolution process, a new knowledge confidence evaluation and fusion module can be added in step S6. When the knowledge evolution module refines a candidate knowledge from the experience chain, the system first assigns it an initial confidence score based on statistical support, result consistency, and other factors, and stores it in a temporary observation knowledge base. In future operation, the system can use an exploratory strategy (such as the ε-greedy strategy) to try out the candidate knowledge with a certain probability, and continuously monitor its application effect. If multiple verifications produce good results, the confidence score of the candidate knowledge is increased. When the score exceeds a preset knowledge solidification threshold (for example, 0.9), the candidate knowledge is formally and automatically incorporated into the main knowledge graph. For candidate knowledge that conflicts with existing knowledge (especially knowledge derived from first principles), the system initiates a knowledge arbitration process assisted by domain experts, thereby avoiding the risk of knowledge base pollution caused by a small amount of abnormal data.

[0060] Embodiment 2 The present embodiment provides a numerical and knowledge fusion enhanced industrial software system for data and instruction interaction with an industrial control system, which can perform the method described in embodiment 1. Referring to Figure 1 , the industrial software enhanced system can be deployed on an edge server or a cloud platform and interact with the industrial control system 101 on site. The system includes: A knowledge graph construction module 102 for performing step S1, responsible for extracting domain knowledge from multiple sources and constructing and maintaining a structured domain knowledge graph. This module can be implemented by a software program running on a server.

[0061] The model training module 103 is configured to perform step S2, and is responsible for designing, training and managing a model cluster containing multiple dedicated artificial intelligence compensation models. In particular, when training the model, the module uses a special loss function containing a physical mechanism residual loss term to ensure the physical consistency of the model output.

[0062] The context perception module 104 is configured to perform step S3, and is responsible for collecting sensor data and production task information in real time from the industrial site through a data interface, and extracting structured working condition features therefrom, and converting the features into a knowledge graph processable query.

[0063] The model activation module 105 is configured to perform step S4, and is responsible for receiving the working condition query from the context perception module, using the reasoning capability of the knowledge graph to perform working condition understanding and risk prediction, and activating the most suitable artificial intelligence compensation model from the model cluster according to the reasoning result.

[0064] The instruction generation module 106 is configured to perform step S5, and is responsible for inputting real-time data into the activated artificial intelligence compensation model, generating a physically consistent compensation control instruction, and sending the instruction to the lower industrial control system 101 through an output interface.

[0065] The knowledge evolution module 107 is configured to perform step S6, and is responsible for recording the complete intervention process to form an experience chain, and performing offline analysis and mining on the experience chain data to extract new knowledge, and realizing closed loop, automatic updating and iteration of the domain knowledge graph.

[0066] Those skilled in the art can understand that all or part of the steps of the above embodiments can be completed by programs instructing relevant hardware, and the programs can be stored in a computer readable storage medium. When the program is executed, it can include the steps of the above method embodiments. Among them, the storage medium can be ROM, RAM, magnetic disk or optical disk, etc. Similarly, the system modules in the above embodiments can also be implemented by hardware circuits (such as FPGA, ASIC), or in the form of software function units.

[0067] Embodiment 3 This embodiment further illustrates the application of the method of the present application in a specific complex industrial scenario, taking the vacuum induction melting process of a high-temperature alloy required for an aero-engine turbine blade as an example. This process has strict requirements for the uniformity of alloy composition and the precise control of trace elements, which directly affects the mechanical properties and service life of the final product. In the prior art, the main method is to program control by using solidified process parameter cards, supplemented by manual experience for fine tuning. However, the limitation of this method is that it cannot effectively respond to the slight fluctuations in impurity composition in different batches of raw materials (including new materials and recycled materials), which may lead to unpredictable element burning or segregation during the melting process. It is difficult to accurately model by relying solely on mechanism models, and unconstrained pure data models may propose heating or feeding instructions that violate the laws of thermodynamics, posing a huge safety and quality risk.

[0068] In this embodiment, a domain knowledge graph for high-temperature alloy melting is first constructed offline. The graph contains the chemical composition standards of various alloy grades, the thermophysical properties of each element, phase diagram data, the interaction between elements (such as eutectic and segregation tendency), the performance parameters of melting equipment (such as vacuum furnaces), and the expert rules of "process phenomenon-failure cause-adjustment measure" summarized in history. At the same time, special artificial intelligence compensation models are pre-trained for several key stages of the melting process, such as "feeding and melting", "composition refining", "trace element control" and "pouring and cooling", forming a model cluster. When training these models, in addition to using historical production data, key physical constraints, such as the furnace temperature change equation based on energy conservation (input power = alloy heat absorption + furnace heat dissipation) and the element concentration change equation based on mass conservation, are added as physical residual loss terms to the total loss function to ensure the reasonableness of the model output.

[0069] In an online operation, the system enters the "trace element regulation" phase, aiming to accurately adjust the final content of several key trace elements (such as rhenium, hafnium). At this time, the context perception module discovers that the actual melting rate of the rhenium element is lower than the theoretical model prediction value by collecting real-time spectral analyzer data, and at the same time, the vacuum gauge reading shows a slight abnormal fluctuation. The system converts the two real-time working condition characteristics of "low rhenium melting rate" and "abnormal vacuum fluctuation" into a query to the knowledge graph. The reasoning engine of the model activation module reasons in the knowledge graph and finds that the two phenomena are highly associated with a historical expert rule, which states: "When the content of a certain specific impurity in the recycled material is slightly high, a trace of oxidation film will be formed on the surface of the melt, hindering the melting of high-melting-point elements, and accompanied by a slight local outgassing phenomenon". Based on this reasoning, the system determines that the current working condition is not a simple melting dynamics problem, but a complex problem caused by raw material uncertainty. Therefore, the system does not activate the standard "trace element regulation" model, but activates a special working condition model that specifically deals with "surface oxidation film inhibition and composition compensation".

[0070] The activated special model receives real-time data and generates a set of physically consistent collaborative compensation instructions: on the one hand, the instructions appropriately increase the stirring power of the induction heating coil to physically destroy the oxidation film on the surface of the melt and accelerate the melting of elements; on the other hand, based on the mass conservation constraint, the dose of additional trace rhenium elements needed to be added is accurately calculated to compensate for the melting loss in the early stage. This set of instructions is sent to the PLC for execution. Subsequent spectral data shows that the melting rate of rhenium elements returns to normal and the final product composition inspection is qualified.

[0071] The entire intervention process, including the initial working condition characteristics, the reasoning path of the knowledge graph (i.e. identifying the oxidation film problem), the activated special model, the output collaborative instructions, and the final good results, is recorded as a structured experience chain. When similar melting anomalies caused by batches of raw materials occur multiple times in subsequent production and are successfully handled, the knowledge evolution module analyzes these accumulated experience chains and automatically refines a new high-confidence knowledge: "When detecting that the batch of recycled materials is a specific batch, and the spectrum and vacuum degree show a certain combination of characteristics, the initial stirring power should be adjusted in advance and the rhenium element should be pre-compensated". This new knowledge is automatically updated to the knowledge graph after passing the confidence evaluation, so that the system can upgrade from passive compensation to active prediction and prevention when encountering similar situations in the future, thereby realizing the closed-loop self-evolution of knowledge.

[0072] As can be seen from the above embodiment of the high-temperature alloy smelting, by constructing a field knowledge graph, the present application can identify deep-rooted problems (for example, surface oxidation film problems caused by raw material fluctuations) that are difficult to find by traditional methods through deep semantic understanding and reasoning of complex industrial field information, and accordingly target the activation of a special artificial intelligence model to generate a physically consistent collaborative control instruction. The process records the experience chain formed, and also provides a basis for the self-learning and knowledge evolution of the system, enabling the system to upgrade from passive response to active prevention. The numerical and intelligent fusion framework disclosed by the present application has universal applicability, and is not only suitable for the smelting process, but also can be widely applied to other complex industrial software enhancement scenarios, such as chemical process optimization, semiconductor manufacturing process control, intelligent power grid scheduling, etc., providing an effective technical approach for improving the adaptability, robustness and intelligent level of the industrial software in these fields.

[0073] The above merely describes preferred embodiments of the present application, but is not intended to limit the protection scope of the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A theoretical and mathematical integration industrial software enhancement system, wherein the industrial software enhancement system is used for data and instruction interaction with an industrial control system, characterized in that, The industrial software enhancement system includes: The knowledge graph construction module is used to formalize and structure industrial domain knowledge during the offline preparation phase in order to build a domain knowledge graph that includes an inference engine. The model training module is used to train a set of dedicated artificial intelligence compensation models for various typical working conditions in the industrial field during the offline preparation stage, thereby forming an artificial intelligence compensation model cluster, and embedding physical mechanism constraints into the training process of each dedicated artificial intelligence compensation model. The context-aware module is used to collect real-time data from the industrial site when the system is running online, extract structured working condition features from it, and convert them into semantic queries that can be understood by the domain knowledge graph. The model activation module is used to input the semantic query into the domain knowledge graph, perform reasoning using the reasoning engine contained in the domain knowledge graph, and select and target the activation of one or a group of artificial intelligence compensation models from the artificial intelligence compensation model cluster based on the reasoning results. The instruction generation module receives real-time sensor data as input from the activated artificial intelligence compensation model, calculates a set of compensation and adjustment instructions for the industrial control system, and sends them to the actuators in the industrial control system through the output interface of the industrial software enhancement system to regulate the industrial process.

2. The industrial software enhancement system integrating rational and mathematical methods according to claim 1, characterized in that, The domain knowledge graph includes core entities, entity attributes, and relationships between entities. The relationships between entities include at least one of the following: causal relationships defined based on physical laws, correlation relationships defined based on statistics or experience, and mapping relationships between preset working conditions and strategies.

3. The industrial software enhancement system integrating rational and mathematical methods according to claim 1, characterized in that, The model training module is specifically used to: design a total loss function L_total composed of a weighted sum of a data-driven loss term L_data and a physical mechanism residual loss term L_phys; Wherein, L_data is used to measure the difference between the predicted value of the artificial intelligence compensation model and the real data label; L_phys is a physical mechanism penalty term, the value of which is used to characterize the degree to which the output of the artificial intelligence compensation model conforms to a core control equation of the industrial field. The less the output conforms to the core control equation, the larger the absolute value of L_phys.

4. The industrial software enhancement system integrating rational and mathematical methods according to claim 1, characterized in that, The context awareness module is specifically used to filter high-frequency noise data, perform trend analysis on variables to determine their rate of change and trend, and use pattern recognition algorithms to identify the operating condition category to which the current data pattern belongs.

5. The industrial software enhancement system integrating rational and mathematical methods according to claim 1, characterized in that, The model activation module is specifically used to: retrieve entities and relationships in the graph that best match the characteristics of the current working condition; reason along the causal or correlational relationships in the graph to predict the subsequent state or potential risks of the current working condition; and determine one or a set of artificial intelligence compensation models to be activated based on the mapping relationship between the working condition and the strategy.

6. The industrial software enhancement system integrating rational and mathematical methods according to claim 1, characterized in that, It also includes a knowledge evolution module for the solidification of experience chains and the closed-loop evolution of knowledge. Specifically, it includes: recording each intervention process to form structured experience chain data containing timestamps, working condition characteristics, graph reasoning paths, activated artificial intelligence compensation models, output compensation adjustment instructions, process status feedback, and final results; and periodically or after the triggering of set events, performing offline analysis on the accumulated experience chain data to extract new knowledge, and automatically and structurally updating the new knowledge into the domain knowledge graph.

7. The industrial software enhancement system integrating rational and mathematical methods according to claim 6, characterized in that, The knowledge evolution module is also used for: When a candidate knowledge is extracted from the experience chain data, an initial confidence score is assigned to the candidate knowledge based on statistical support and result consistency, and it is stored in a temporary observation knowledge base. In subsequent operation, the candidate knowledge is tested with a preset probability and its application effect is monitored, and its confidence score is updated according to the application effect. Once the confidence score exceeds the preset knowledge solidification threshold, the candidate knowledge is formally incorporated into the domain knowledge graph.

8. The industrial software enhancement system integrating rational and mathematical methods according to claim 7, characterized in that, Also includes: When the candidate knowledge conflicts with existing knowledge derived from first principles in the domain knowledge graph, a knowledge arbitration process assisted by domain experts is initiated to determine whether the candidate knowledge should be incorporated into the domain knowledge graph.

9. The industrial software enhancement system integrating rational and mathematical methods according to claim 1, characterized in that, When the real-time data from the industrial site is time-related process data, the artificial intelligence compensation model uses a long short-term memory network; when it is necessary to integrate the topological structure information of the domain knowledge graph for prediction, the artificial intelligence compensation model uses a graph neural network.

10. A method for applying to an industrial software enhancement system that integrates rational and mathematical data as described in any one of claims 1-9, wherein the industrial software enhancement system is used for data and instruction interaction with an industrial control system, characterized in that, The method includes the following steps: Step S1: In the offline preparation stage, the industrial domain knowledge is formalized and structured to construct a domain knowledge graph that includes an inference engine. Step S2: In the offline preparation stage, a set of dedicated artificial intelligence compensation models are trained for various typical working conditions in the industrial field, thereby forming an artificial intelligence compensation model cluster, and physical mechanism constraints are embedded into the training process of each dedicated artificial intelligence compensation model. Step S3: When the system is running online, collect real-time data from the industrial site, extract structured working condition features from it, and convert the working condition features into semantic queries that the domain knowledge graph can understand. Step S4: Input the semantic query into the domain knowledge graph, use the reasoning engine contained in the domain knowledge graph to perform reasoning, and select and target one or a group of artificial intelligence compensation models from the artificial intelligence compensation model cluster based on the reasoning results. In step S5, the artificial intelligence compensation model activated in step S4 receives real-time sensor data as input, calculates a set of compensation and adjustment instructions for the industrial control system, and sends them to the actuators in the industrial control system through the output interface of the industrial software enhancement system to regulate the industrial process.

Images