An autonomous control method based on layered knowledge architecture and object history inheritance

Abstract

The application discloses a kind of based on layered knowledge architecture and object history inheritance autonomous control method, belong to complex technical system control technical field.The method includes: the seven-layer progressive layered semantic knowledge architecture consisting of original data layer, state variable layer, natural science model layer, field knowledge layer, artificial intelligence layer, decision control layer and object history layer is constructed;Each technical object is assigned a unique identifier and initialized object history;Data is collected in each process stage and converted into state variables, and personalized control parameters are autonomously generated through layered reasoning;The updated history is passed to the downstream, and the downstream makes decisions based on the inherited history and feeds back to the upstream to achieve self-optimization;Based on object history, digital twin is built to realize long-cycle performance prediction.The application realizes decision-making full-link traceability, single-product personalized control and full-life-cycle self-learning iteration, and is suitable for manufacturing and operation control of complex systems such as fuel cells, lithium batteries and semiconductors.

Description

Technical Field

[0001] This invention relates to the technical field of industrial artificial intelligence and cyber-physical systems, and in particular to a method for autonomous control and regulation of a complex technical system throughout its entire lifecycle based on a hierarchical semantic knowledge architecture and object history inheritance. Background Technology

[0002] As industrial production systems become increasingly digitized, the volume of process, sensor, and operational data continues to grow. Currently, various methods are used to evaluate this data, including statistical methods, data-driven models, machine learning methods, neural networks, digital twins, and knowledge-based expert systems. Simultaneously, digital object traceability is being built along the value chain to store quality data, test information, and operational status of materials, process parameters, individual components, or products. Furthermore, systems are known to combine individual physical models with artificial intelligence methods to optimize production processes or predict the future state of technological systems.

[0003] However, existing technologies have the following significant drawbacks:

[0004] First, data, models, expert knowledge, product information, and decision-making algorithms are either isolated or coupled only at specific points, lacking a consistently layered knowledge architecture. This prevents the systematic connection of all knowledge levels, from physical laws to autonomous decision-making, within a unified semantic structure. Existing full-lifecycle cyber-physical fusion solutions employ a flat four-dimensional architecture, where data, physical models, domain knowledge, AI methods, decision-making algorithms, and object history lack clear hierarchical divisions. Various types of information are loosely coupled and spliced ​​together, lacking a unified semantic knowledge graph and making it impossible to trace causal relationships in decision-making layer by layer.

[0005] Second, digital object history primarily serves as a document record or traceability for manufacturing, testing, and operational data. The information contained within is typically not utilized as part of a hierarchically constructed knowledge structure, thus hindering the independent derivation of new insights, technical conclusions, or optimized control strategies. Existing object history is only used for static quality evidence preservation and traceability, lacking a cross-process history inheritance and transfer mechanism. Upstream information is not fully transferred downstream, and downstream control relies solely on real-time data from its own process, making it impossible to achieve personalized parameter adjustment for individual products.

[0006] Third, known machine learning or neural network systems primarily derive decisions from the statistical correlation between input and output data, with limited interpretability of the underlying natural scientific causal relationships. The integration of physical laws, chemical reaction mechanisms, materials science models, or formalized expert knowledge is often only supplementary or application-specific. Purely data-driven AI relies on statistical correlations, lacks support from physical and chemical causal mechanisms, and exhibits poor predictive robustness.

[0007] Fourth, digital twins primarily map the current or predicted state of a technical system, typically lacking higher-level knowledge organization. They cannot interconnect measurement data, models, domain knowledge, decision rules, and product information within a unified semantic architecture. The use of digital twins is often limited to a single component, device, or operational state, merely reproducing instantaneous operational conditions and failing to predict aging and remaining service life based on a complete lifecycle timeline. Equipment fault diagnosis solutions focus only on a single aspect of maintenance fault diagnosis; knowledge graphs are only used for causal fault retrieval, not covering the entire process from front-end manufacturing, assembly, off-line testing, to recycling, and lack the ability to autonomously adjust manufacturing processes.

[0008] Fifth, most existing systems lack a unified descriptive level for different types of technical information. Raw data, measured values, derived state variables, physical models, expert knowledge, statistical relationships, and decision rules are often treated equally, despite belonging to different levels of abstraction and knowledge. This hinders both the traceability of autonomous decision-making and the migration of the system to new products, processes, or application areas. The lack of a standardized, universal data structure and the architecture's reliance on a single production line make reuse across fields such as electrochemistry, chemical engineering, and aerospace extremely difficult.

[0009] Sixth, existing technologies only support one-way data flow from manufacturing to operation and maintenance. There is no closed-loop logic for downstream processes to provide feedback to optimize upstream processes. The knowledge base and simulation model cannot be continuously updated as the product is put into service, and there is a lack of a full life cycle self-learning and iteration mechanism.

[0010] The information disclosed in this background section is intended only to enhance the understanding of the overall background of the invention and should not be construed as an admission or in any way implying that the information constitutes prior art known to those skilled in the art. Summary of the Invention

[0011] The purpose of this invention is to provide an autonomous control method based on a hierarchical knowledge architecture and object history inheritance, in order to solve the technical problems in the prior art such as data and knowledge separation, untraceable decision-making, object history only being stored as evidence without participating in control, lack of cross-process inheritance mechanism, lack of bidirectional closed-loop self-learning, and digital twins only being instantaneous simulations.

[0012] To achieve the above objectives, this invention provides a self-control method based on a hierarchical knowledge architecture and object history inheritance, comprising the following steps:

[0013] S1: Construct a multi-level progressive semantic knowledge architecture, which consists of a seven-layer progressive structure from bottom to top: raw data layer, state variable layer, natural science model layer, domain knowledge layer, artificial intelligence layer, decision control layer and object history layer. Each knowledge object is configured with a unique identifier, attribute definition, semantic relationship and state transition rules. The layers are connected by semantic association to realize progressive reasoning from raw data to autonomous decision-making.

[0014] S2: Assign a unique and unchangeable object identifier to each technical object, create a corresponding object history and complete initialization. Initialization data includes raw material parameters, supplier data, incoming characterization test data and initial state variables. The object identifier remains unchanged throughout the entire life cycle, achieving complete correspondence of all state information.

[0015] S3: At each process stage, raw measurement data is collected and converted into state variables with physical meaning. The state variables and their semantic classifications are written into the object history. Consistency checks, model mapping and knowledge reasoning are completed through a hierarchical knowledge architecture. The personalized control parameters and optimization strategies for the single product are generated autonomously, and the decision results are fed back to the object history as new knowledge elements.

[0016] S4: The updated complete object history is standardized and passed to the downstream process. The downstream process carries out its control, adjustment and decision-making based on the inherited complete upstream history. At the same time, the newly added status information, process deviation and performance data of the downstream process are fed back to the upstream process, driving the upstream process parameters to self-optimize and adjust and the knowledge base to expand iteratively. Among them, the updated object life cycle is passed to at least one subsequent method step in the form of state and knowledge information inheritance.

[0017] In one embodiment of the present invention, the unique identifier of each knowledge object in step S1 is generated by combining product ID, process step, timestamp and version number, and each knowledge element can be explicitly referenced through this unique identifier; the attributes of the knowledge object include physical characteristics, process parameters, measured values, uncertainties, quality characteristics and calculated state variables; the semantic relationships between knowledge objects include time predecessor / successor relationships, cause-effect relationships, whole-part relationships, state transition relationships and deductive reasoning relationships.

[0018] In one embodiment of the present invention, the object history layer in step S1 is the seventh layer. It is not an independent data source, but an integration of all information from the first to the sixth layers in the time dimension. It stores the complete time-series document of the temporal state change process, process conversion records, measurement values, decision results and knowledge elements derived therefrom throughout the entire life cycle. The object history not only stores the original data, but also stores the semantic classification of the data and the knowledge elements that are valid at that time.

[0019] In one embodiment of the present invention, the reasoning process of the decision control layer in step S3 is executed according to the following logical steps:

[0020] Read the semantically described state variables;

[0021] Perform consistency and reasonableness checks;

[0022] Map state variables to semantic models and object histories;

[0023] Reasoning is performed using domain knowledge, physical models, and rules;

[0024] Evaluate multiple possible course of action;

[0025] Choose the optimal strategy;

[0026] Output target values, control variables, and action recommendations;

[0027] The decision results are fed back into the object's resume as new knowledge elements.

[0028] In one embodiment of the present invention, the iterative expansion of the knowledge base in step S4 follows the following four rules:

[0029] Rule 1: When a previously unknown relationship is identified between process state and product characteristics, a new knowledge element is generated;

[0030] Rule 2: Expand an existing knowledge element when new measurement data confirms or quantitatively refines its statement.

[0031] Rule 3: When new knowledge indicates that a previous description is incomplete or incorrect, replace one knowledge element;

[0032] Rule 4: When a component first demonstrates a provable functional association within a component group, a new whole-part relationship is generated.

[0033] In one embodiment of the present invention, the sources of initial data for the object history in step S2 include: for objects manufactured within the enterprise's internal value chain, the initial data begins with their respective raw materials and manufacturing process data; for purchased objects, the initial data uses specifications, quality data, and batch information provided by the supplier; all objects are recorded through their own characterization and testing procedures before further processing, including mechanical, physical, chemical, and electrochemical studies and microscopic structural analysis methods, and the measurement data obtained therefrom, together with the supplier data, constitute the initial state space of the object history.

[0034] In one embodiment of the present invention, the object identifier is generated by the supplier batch number or serial number, internal allocation within the enterprise, or by a combination of multiple features such as supplier information, batch, timestamp, material category and serial number; the carrier form of the identifier includes QR code, RFID, DataMatrix code or universally unique identifier UUID, and each physical object clearly corresponds to at least one independent digital object lifecycle.

[0035] In one embodiment of the present invention, each decision made and its effect in step S3 is written as a new knowledge element into the object's history, becoming part of the object's history, and expanding the semantic knowledge model for subsequent decisions, forming a recursive knowledge growth mechanism that combines object history, semantic ontology, and autonomous decision-making; the knowledge object is not only stored, but is continuously re-evaluated semantically, so that each state change can generate new knowledge, and by identifying new causal relationships, performing semantic classification, and persistently integrating them into the knowledge base, the structured knowledge is continuously developed.

[0036] In one embodiment of the present invention, the following steps are also included: S5: Constructing a digital twin based on the object history, wherein the digital twin not only describes the current state of the technical object, but also includes all known states, the state formation process and the underlying causal relationships. By loading the complete object history, long-term prediction of the product's future performance, aging process, quality development and remaining service life is realized. The prediction results are written back to the object history and participate in subsequent knowledge reasoning.

[0037] In one embodiment of the invention, the method is used for the manufacturing, assembly, testing, operation management, condition monitoring, maintenance and / or recycling of complex technical products, as well as for technical systems used in the multi-stage manufacturing of fuel cells, electrolyzers, batteries, semiconductor products, chemical process systems, machines, vehicle components or similar products.

[0038] Compared with existing technologies, the autonomous control method based on a hierarchical knowledge architecture and object history inheritance according to the present invention has the following beneficial effects: It achieves end-to-end traceability from raw data to autonomous decision-making through a seven-layer progressive hierarchical semantic knowledge architecture, solving the problem of unexplainable pure data-driven AI; it enables complete cross-process transmission and personalized adaptive control of single-product lifecycle status information through an object history inheritance mechanism; it achieves full lifecycle self-learning iteration through upstream and downstream bidirectional closed-loop feedback; it achieves long-term accurate prediction of aging evolution and performance degradation through a full-time history-driven digital twin; it achieves universal migration across devices, software, and industries through a standardized general data structure; it achieves dynamic knowledge generation and structured knowledge development through continuous semantic evaluation of knowledge objects; it connects all knowledge elements through a unified semantic ontology, forming a consistent knowledge graph and avoiding information fragmentation and redundancy; and it significantly improves the robustness and transferability of autonomous decision-making by applying artificial intelligence methods to structured knowledge objects rather than raw data, combined with natural science models and domain knowledge constraints. Attached Figure Description

[0039] Figure 1 This is a flowchart of an autonomous control method based on a hierarchical knowledge architecture and object history inheritance according to an embodiment of the present invention. Detailed Implementation

[0040] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings, but it should be understood that the scope of protection of the present invention is not limited to the specific embodiments.

[0041] Unless otherwise expressly stated, throughout the specification and claims, the term "comprising" or its variations such as "including" or "comprises" shall be understood to include the stated elements or components without excluding other elements or other components.

[0042] like Figure 1 As shown, according to a preferred embodiment of the present invention, an autonomous control method based on a hierarchical knowledge architecture and object history inheritance has an overall architecture comprising a seven-layer progressive hierarchical semantic knowledge system. The knowledge objects in each layer are associated through unified semantic relationships, realizing progressive reasoning from raw data to autonomous decision-making.

[0043] Specifically, the method includes the following steps:

[0044] S1: Construct a multi-level progressive semantic knowledge architecture, which consists of a seven-layer progressive structure from bottom to top: raw data layer, state variable layer, natural science model layer, domain knowledge layer, artificial intelligence layer, decision control layer and object history layer. Each knowledge object is configured with a unique identifier, attribute definition, semantic relationship and state transition rules. The layers are connected by semantic association to realize progressive reasoning from raw data to autonomous decision-making.

[0045] In this step, the first layer of raw data acquisition technology system acquires the raw sensor data of the entire process, including sensor values, image data, measurement signals, process parameters, control variables, operating data, test values ​​and environmental conditions, which constitute the objective data foundation for all subsequent knowledge processes.

[0046] The second layer of state variables transforms the raw measurement data into technically interpretable state variables after verification, synchronization, and rationalization. These include descriptive variables such as temperature, pressure, concentration, material properties, electrical characteristics, mechanical stress, deformation, aging state, mass characteristics, efficiency, effective area, and reaction kinetics, thus achieving the separation of direct observation and physical description of the system.

[0047] The third layer, the natural science model layer, organizes physical models, chemical models, electrochemical models, materials science models, mathematical models, and empirical verification relationships in a common knowledge base, describing the causal relationships between state variables, thus making it possible to provide technical interpretations of the collected state variables.

[0048] The fourth layer, the domain knowledge layer, integrates expert knowledge, design guidelines, manufacturing knowledge, quality rules, safety rules, standards and specifications, experiential knowledge, optimization strategies, and technical boundary conditions. This knowledge is formally described and associated with natural science models.

[0049] The fifth layer of artificial intelligence includes machine learning, neural networks, probabilistic models, fuzzy logic, reasoning methods, optimization algorithms, and other adaptive methods. These methods do not operate in isolation on raw data, but on knowledge objects that have been structured by the aforementioned layers, thereby significantly improving the robustness, interpretability, and transferability of autonomous decision-making.

[0050] The sixth-level decision-making and control layer autonomously generates setpoints, manipulated variables, control strategies, optimization measures, maintenance decisions, predictions, and action recommendations based on all the aforementioned knowledge levels.

[0051] The seventh layer, the object history layer, constitutes a continuously growing object biography, describing the complete history of all relevant information about the technology object throughout its entire lifecycle. This seventh layer is not an independent data source, but rather an integration of all information from layers one through six along a timeline. It stores a complete time-series document of temporal state changes, process transition records, measurements, decision results, and the knowledge elements derived from them throughout the entire lifecycle. The object history not only stores raw data but also its semantic classification and the knowledge elements that were valid at that time. New insights are continuously fed back into the knowledge architecture, expanding its knowledge base and thus forming a self-learning overall technology system.

[0052] Each knowledge object at each layer is uniquely identified by a combination of product ID, process step, timestamp, and version number. Each knowledge element can be explicitly referenced using this unique identifier. Knowledge object attributes include physical properties, process parameters, measured values, uncertainties, quality characteristics, and calculated state variables. Semantic relationships between knowledge objects include time predecessor / successor relationships, causal-effect relationships, whole-part relationships, state transition relationships, and deductive reasoning relationships. All knowledge elements are not merely interconnected through databases or file structures, but organized through a formalized semantic structure, forming a consistent knowledge graph that maps deterministic and probabilistic relationships within the same architecture.

[0053] The hierarchical semantic knowledge architecture in this step is built upon an ontology based on natural sciences. This ontology includes the following basic knowledge categories: Natural Laws, describing the fundamental physical, chemical, mechanical, electrical, or mathematical laws that determine the behavior of technological systems, constituting the boundary conditions for all model construction; Models, representing formal descriptions of natural science relationships, mapping the relationships between state variables of technological systems, which can be analytical, numerical, empirical, or hybrid; Objects, including materials, components, subassemblies, products, processes, equipment, machines, or complete technological systems, each object possessing a unique identifier and object history; States and State Variables, describing the physical, chemical, mechanical, electrical, geometric, or mathematical states of objects through a set of state variables, achieving separation between direct observation and physical description of the system; Processes and State Transitions, describing the processes that change the state of objects, each process connecting at least one initial state and one final state, with continuous state transition sequences describing the entire lifecycle development of objects; Knowledge Elements, including natural science models, domain knowledge, expert rules, mathematical relationships, statistical models, and artificial intelligence methods, used for interpreting state variables; and Decisions, technological decisions derived from knowledge elements, including setting variables, manipulating variables, controlling parameters, maintenance measures, quality decisions, and optimization strategies.

[0054] The core technical effect of this step is that, by organizing all knowledge levels in a hierarchical manner, it creates for the first time a technical architecture that can transform data into information, information into technically interpretable state variables, state variables into knowledge, knowledge into autonomous decision-making, and decision-making into adaptive control and regulation processes. This ensures that each decision can logically find its basis in its respective lower-level knowledge, thereby achieving continuous improvement throughout the entire lifecycle of the technical system.

[0055] S2: Assign a unique and unchangeable object identifier to each technical object, create a corresponding object history and complete initialization. Initialization data includes raw material parameters, supplier data, incoming characterization test data and initial state variables. The object identifier remains unchanged throughout the entire lifecycle, achieving complete correspondence of all state information.

[0056] In this step, the initial data sources for the object history are divided into two categories: for objects manufactured within the enterprise's internal value chain, such as catalyst materials, catalyst coatings, or bipolar plates, their object history begins with their respective raw materials and manufacturing process data; for externally purchased objects, such as membranes or gas diffusion layers, the initial data uses specifications, quality data, and batch information provided by the supplier. Furthermore, all objects are characterized and tested through their own procedures before further processing, including mechanical, physical, chemical, and electrochemical studies, as well as microscopic structural analysis methods such as transmission electron microscopy and scanning electron microscopy. The resulting measurement data, along with supplier data, constitute the initial state space of the corresponding object history.

[0057] Each object acquires a unique, immutable object identifier upon creation or entry into the value chain. This identifier perfectly maps all states, process steps, measurement variables, and state transitions throughout its lifecycle. The object identifier remains unchanged across all levels of the knowledge architecture and can explicitly inherit all state information from raw materials through components to the complete final product and its operation, establishing a correspondence between an object, a unique identifier, and an object history.

[0058] The object identifier is derived from the supplier batch number or serial number, internally assigned by the enterprise, or generated by a combination of multiple features such as supplier information, batch, timestamp, material category, and serial number; the carrier form of the identifier includes QR code, RFID, DataMatrix code, or universally unique identifier UUID, and each physical object clearly corresponds to at least one independent digital object lifecycle.

[0059] The core technical effect of this step is that by establishing a one-to-one correspondence between a unique identifier and the object's history, a complete anchor point for the entire life cycle information of the technical object is established, ensuring that all state changes, process conversions and decision results can be traced back to a unique physical object, laying the foundation for subsequent cross-process history inheritance and autonomous decision-making.

[0060] S3: At each process stage, raw measurement data is collected and converted into state variables with physical meaning. The state variables and their semantic classifications are written into the object history. Consistency checks, model mapping and knowledge reasoning are completed through a hierarchical knowledge architecture. The personalized control parameters and optimization strategies for the single product are generated autonomously, and the decision results are fed back to the object history as new knowledge elements.

[0061] In this step, the standardized processing loop within a single process includes: First, continuously collecting measurement variables for this process, including temperature, pressure, roll speed, material elongation, flatness, coating thickness, coating weight, humidity, conductivity, image data, geometric deviation, surface structure, etc.; then, in the second layer of state variables, the original measurement data are converted into physically meaningful state variables based on natural science models. For example, platinum loading is calculated based on coating thickness and coating area density, Young's modulus of the material is calculated based on roll tension and elongation, membrane surface waviness is calculated based on image data, and membrane proton conductivity is calculated based on conductivity measurements. The uncertainty of the state variables is quantified and calculated using the measurement error propagation formula and stored along with the state variables; the converted state variables and their semantic classifications are written into the object's history, recording the state transitions brought about by this process.

[0062] The reasoning process of the decision control layer is executed according to the following logical steps: read the state variables of the semantic description; perform consistency and rationality checks, such as verifying whether the coating thickness matches the coating weight; map the state variables to the semantic model and object history, and associate them with the initial parameters in the upstream object history; use domain knowledge, physical models and rules to reason, and combine process experience knowledge and relevant models to evaluate the current quality; evaluate multiple possible action plans, such as adjusting tension, adjusting speed or adjusting slurry viscosity; select the optimal strategy; output the target value, control variables and action suggestions, such as parameter correction values ​​for the current workpiece; and feed the decision results back to the object history as new knowledge elements.

[0063] Each decision made and its effects are written as new knowledge elements into the object's history, becoming part of the object's history and expanding the semantic knowledge model for subsequent decisions. This forms a recursive knowledge growth mechanism that combines object history, semantic ontology, and autonomous decision-making. Knowledge objects are not merely stored but are continuously re-evaluated semantically, ensuring that each state change generates new knowledge. By identifying new causal relationships, performing semantic classification, and persistently integrating them into the knowledge base, structured knowledge is continuously developed. It is this recursive knowledge growth—the combination of object history, semantic ontology, and autonomous decision-making—that constitutes the core inventive idea of ​​this invention.

[0064] The core technical effect of this step is that it enables personalized control of individual products through hierarchical reasoning. The decisions of each process stage are based on the product's complete history and underlying physical model, rather than uniform process benchmark parameters, thereby significantly improving the manufacturing consistency and product quality of complex systems.

[0065] S4: The updated complete object history is standardized and passed to the downstream process. The downstream process carries out its control, adjustment and decision-making based on the inherited complete upstream history. At the same time, the newly added status information, process deviation and performance data of the downstream process are fed back to the upstream process, driving the upstream process parameters to self-optimize and adjust and the knowledge base to expand iteratively. Among them, the updated object life cycle is passed to at least one subsequent method step in the form of state and knowledge information inheritance.

[0066] In this step, the object history inheritance and transfer mechanism is as follows: After each process is completed, the fully updated object history is transferred to the downstream process through a standardized data structure. When the downstream process starts, all control, detection, optimization, and decision-making functions are based on the inherited upstream complete history, rather than relying solely on the real-time data of the current process. During the history transfer process, not only the values ​​of state variables are transferred, but also their semantic classification, uncertainty, relevant model parameters, and historical decision records, ensuring that the downstream process can fully understand the upstream state and its causes. For example, the downstream hot pressing process can adjust the hot pressing pressure and temperature in different areas based on the coating uniformity distribution recorded in the upstream coating history, thereby achieving personalized processing parameter settings for individual products.

[0067] The two-way closed-loop feedback mechanism works as follows: new status information, process deviations, and performance data from downstream processes are fed back to upstream processes, driving the upstream process parameters to self-optimize and adjust. For example, during the fuel cell stack offline testing, it was found that the performance deviation of a certain batch of components was large. By tracing the object's history, the main cause was located in a certain process parameter fluctuation in the upstream process. Subsequently, this deviation was fed back to the upstream process, automatically correcting the control baseline parameters and early warning thresholds.

[0068] The iterative expansion of the knowledge base follows four rules: Rule 1, when a previously unknown relationship between process state and product characteristics is identified, a new knowledge element is generated; Rule 2, when new measurement data confirms or quantitatively refines the statement of an existing knowledge element, that existing knowledge element is expanded; Rule 3, when new knowledge indicates that a previous description is incomplete or incorrect, a knowledge element is replaced; Rule 4, when a component first exhibits a provable functional relationship within a component group, a new whole-part relationship is generated. The expansion of the knowledge base is not achieved by arbitrarily adding more data, but rather by identifying new semantic relationships between state variables, process transitions, whole-part relationships, and the resulting changes in product characteristics, based on these rules.

[0069] The core technical effect of this step is that it enables product-level adaptive control through cross-process history inheritance, achieves full life-cycle self-learning iteration through two-way closed-loop feedback, and forms a permanent self-learning system, which can significantly shorten the production line debugging cycle.

[0070] The method also includes the following steps:

[0071] S5: Based on object history, a digital twin is built. The digital twin not only describes the current state of the technical object, but also includes all known states, the state formation process and the underlying causal relationships. By loading the complete object history, long-term predictions of the product's future performance, aging process, quality development and remaining service life can be achieved. The prediction results are written back to the object history and participate in subsequent knowledge reasoning.

[0072] In this step, the construction of the digital twin uses a complete inherited object history as the underlying data source, loads full lifecycle time-series state data, and combines a third-layer natural science model (such as a physical aging model, an electrochemical decay model, and a material creep model) with a fifth-layer artificial intelligence prediction algorithm to simulate the evolution process of performance degradation, seal aging, and catalyst degradation under long-term operation. Unlike traditional digital twins that only map instantaneous operating states, the digital twin of this invention includes all known states of the technical object, its formation process, and the underlying causal relationships, enabling prediction of the product's future performance, aging process, quality development, and remaining service life during the manufacturing process.

[0073] The prediction results are written back to the object's history and participate in subsequent knowledge reasoning and control decisions, forming a complete closed loop of history construction, digital twin simulation, prediction output, history writing back, and optimization decisions. During the product operation phase, real-time collected operational data continuously updates the object's history. Based on this, the digital twin dynamically adjusts the prediction model, predicting product performance, efficiency, failure probability, and remaining service life in real time, and automatically outputting maintenance and correction strategies. Repair and replacement data are written into the history, further updating the aging model and life prediction. During the scrapping and recycling phase, disassembly, testing, and material recycling data are entered into the history, creating a complete closed-loop record of the product's entire life cycle.

[0074] The core technical effect of this step is that, through a digital twin driven by a full-time history, it can achieve high-precision prediction of the long-term performance evolution, aging degradation and extreme working condition performance of products. The life prediction error is far better than that of existing twin solutions based only on real-time data. The prediction results are written back into the history and participate in subsequent reasoning, making the digital twin an organic component of the continuously evolving hierarchical knowledge architecture, rather than an independent simulation tool.

[0075] The method of the present invention can be used for the manufacturing, assembly, testing, operation management, condition monitoring, maintenance and / or recycling of complex technical products, as well as for technical systems for the multi-stage manufacturing of fuel cells, electrolyzers, batteries, semiconductor products, chemical process systems, machines, vehicle parts or similar products.

[0076] Example: Autonomous control of the entire life cycle of a PEM fuel cell stack

[0077] This embodiment uses the manufacturing and lifecycle management of a proton exchange membrane (PEM) fuel cell stack as an example to illustrate the specific implementation process of the method of the present invention. The embodiment is for illustrative purposes only and does not constitute a limitation on the scope of protection.

[0078] S1: Constructing a seven-layer progressive hierarchical semantic knowledge architecture

[0079] This embodiment first constructs a seven-layer progressive hierarchical semantic knowledge architecture, from bottom to top as follows:

[0080] The first layer is the raw data layer, which collects raw sensor data from the entire fuel cell manufacturing and operation process, including temperature, pressure, roll speed, coating thickness, humidity, conductivity, image data, geometric deviation, surface structure, positioning accuracy, clamping force, leakage rate, and electrochemical characteristic values. This layer forms the objective data foundation for all subsequent knowledge processes.

[0081] The second layer is the state variable layer, which converts the raw measurement data into technically interpretable state variables after verification, synchronization, and rationalization. These include descriptive variables with clear physical meaning, such as material elongation, smoothness / wrinkle, platinum loading, membrane resistance, catalyst activity, contact pressure, sealing compression, single cell voltage uniformity, aging state, quality characteristics, efficiency, effective area, and reaction kinetics.

[0082] The third layer is the natural science model layer, which organizes physical models, chemical models, electrochemical models, materials science models, mathematical models, and empirically verified relationships in a common knowledge base, describing the causal relationships between state variables. For example, the Ohmic polarization model describes the relationship between membrane resistance and voltage drop, the activation polarization model describes the relationship between catalyst activity and reaction kinetics, the mass transfer model describes the relationship between gas diffusion and concentration distribution, and the material creep model describes the long-term deformation of seals under clamping force.

[0083] The fourth layer is the domain knowledge layer, which integrates expert knowledge, design guidelines, manufacturing knowledge, quality rules, safety rules, standards and specifications, experiential knowledge, optimization strategies, and technical boundary conditions. This knowledge is formally described and associated with natural science models.

[0084] The fifth layer is the artificial intelligence layer, which includes machine learning, neural networks, probabilistic models, fuzzy logic, reasoning methods, optimization algorithms, and other adaptive methods. These methods do not operate in isolation on raw data, but rather on knowledge objects that have been structured by the aforementioned layers, thereby significantly improving the robustness, interpretability, and transferability of autonomous decision-making.

[0085] The sixth layer is the decision control layer, which autonomously generates setpoints, manipulated variables, control strategies, optimization measures, maintenance decisions, predictions, and action suggestions based on all the aforementioned knowledge layers.

[0086] The seventh layer is the object history layer, constituting a continuously growing object biography. It contains a complete time-series document of all state changes, technological transformations, measurements, decisions, and the knowledge derived from them, generated from layers one through six. The seventh layer is not a new data source, but rather an integration of all layers along the time dimension. The object history not only stores the raw data but also its semantic classification and the knowledge elements that were valid at that time.

[0087] Each knowledge object is configured with a unique identifier, attribute definitions, semantic relationships, and state transition rules. The unique identifier is generated by combining the product ID, process step, timestamp, and version number; each knowledge element can be explicitly referenced through this unique identifier. The attributes of a knowledge object include physical characteristics, process parameters, measured values, uncertainties, quality characteristics, and calculated state variables. The semantic relationships between knowledge objects include time predecessor / successor relationships, causal cause-effect relationships, whole-part relationships, state transition relationships, and deductive reasoning relationships.

[0088] S2: Object Identifier Allocation and History Initialization

[0089] Each component of the fuel cell system is given an independent object history, including catalyst materials, membranes, catalyst-coated membranes (CCM), membrane electrode assemblies (MEA), gas diffusion layers (GDL), bipolar plates (BPP), seals, cooling plates, single cells, stacks, and end products. Each object is assigned a unique, immutable object identifier when it is generated or enters the value chain. This identifier can completely map all states, process steps, measurement variables, and state transitions throughout the entire lifecycle.

[0090] The initial data sources for the object's history include: for catalyst materials, CCMs, or bipolar plates manufactured in-house, the object's history begins with its respective raw materials and manufacturing process data; for purchased membranes or gas diffusion layers, the specifications, quality data, and batch information provided by the supplier are used as initial data. Furthermore, all objects are characterized and tested according to their own procedures before further processing, including mechanical, physical, chemical, and electrochemical studies, as well as microscopic structural analysis methods such as transmission electron microscopy (TEM) and scanning electron microscopy (REM). The resulting measurement data, along with supplier data, constitute the initial state space of the corresponding object's history.

[0091] The object identifier remains unchanged across all levels of the knowledge architecture and can explicitly inherit all state information from raw materials through components to the complete fuel cell stack and its operation, realizing the basic concept of "one object – one unique identifier – one object history". The identifier can be inherited from the supplier batch number or serial number, assigned internally by the enterprise, or generated by a combination of multiple features such as supplier information, batch, timestamp, material category, and serial number; the carrier form includes QR code, RFID, DataMatrix code, or universally unique identifier (UUID).

[0092] S3: Single-process layered processing and autonomous decision-making

[0093] Taking the CCM roll-to-roll coating process as an example, this paper illustrates the standardized processing cycle within a single process. During the coating process, measurement variables such as temperature, pressure, roll speed, material elongation, flatness / wrinkle, coating thickness, coating weight, humidity, electrical conductivity, image data, geometric deviation, and surface structure are continuously collected.

[0094] In the second layer of state variables, the raw measurement data are converted into physically meaningful state variables based on natural science models. For example, platinum loading is calculated based on coating thickness and coating area density, Young's modulus of the material is calculated based on coil tension and elongation, surface waviness of the membrane is calculated based on image data, and membrane proton conductivity is calculated based on conductivity measurements. The uncertainty of the state variables is quantified and calculated using measurement error propagation formulas and stored along with the state variables.

[0095] The converted state variables and their semantic classifications are written into the object history of the CCM object to record the state transitions brought about by this coating process.

[0096] The reasoning process then proceeds to the sixth-level decision-control layer, executing the following logical steps: Reading the state variables of the semantic description; performing consistency and rationality checks, such as verifying whether the coating thickness matches the coating weight; mapping the state variables to the semantic model and object history, and associating them with the initial parameters in the upstream catalyst material history; using domain knowledge, physical models, and rules for reasoning, and combining coating process experience knowledge and mass transfer models to evaluate the current coating quality; evaluating multiple possible action plans, such as adjusting tension, adjusting coating speed, or adjusting slurry viscosity; selecting the optimal strategy; outputting target values, control variables, and action suggestions, such as tension correction values ​​for the current roll material; and feeding the decision results back to the object history as new knowledge elements.

[0097] Each decision made and its effects become part of the object's history, expanding the semantic knowledge model for subsequent decisions. It is this recursive knowledge growth—the combination of object history, semantic ontology, and autonomous decision-making—that constitutes the core inventive idea of ​​this invention.

[0098] S4: Object History Inheritance and Two-Way Closed-Loop Feedback

[0099] After the CCM coating process is completed, the fully updated CCM object history is transferred to the downstream MEA manufacturing process through a standardized data structure. When the MEA process starts, all control, inspection, optimization, and decision-making functions are carried out based on the inherited upstream complete history, rather than relying solely on the real-time data of this process.

[0100] For example, the MEA hot pressing process can adjust the hot pressing pressure and temperature of different areas according to the coating uniformity distribution recorded in the CCM history, so as to realize personalized processing parameter settings for individual products, thereby improving the consistency and performance of MEA.

[0101] During the process of transmitting the history, not only are the numerical values ​​of the state variables transmitted, but also their semantic classification, uncertainty, relevant model parameters and historical decision records are transmitted to ensure that downstream processes can fully understand the upstream state and its causes.

[0102] New status information, process deviations, and performance data from downstream processes are fed back to upstream processes, driving self-optimization adjustments of upstream process parameters. For example, during fuel cell stack offline testing, it was found that the voltage uniformity of single cells corresponding to a certain batch of CCMs had a large deviation. By tracing the CCM history, the main cause was identified as tension fluctuations in a certain section of the roll material during the coating process. Subsequently, this deviation was fed back to the coating process, automatically correcting the tension control baseline parameters and warning thresholds.

[0103] The iterative expansion of the knowledge base follows these four rules:

[0104] Rule 1: When a previously unknown correlation between process state and product characteristics is identified, a new knowledge element is generated. For example, when a quantitative relationship between coating rate and catalyst layer porosity is discovered for the first time, a new knowledge element is generated.

[0105] Rule 2: Expand an existing knowledge element when new measurement data confirms or quantitatively refines its statement. For example, update the parameter accuracy of the knowledge element when new batch data further validates the relationship between tension and coating thickness deviation.

[0106] Rule 3: When new knowledge indicates that a previous description was incomplete or incorrect, replace a knowledge element. For example, if it is found that the original aging model has a large prediction bias under high temperature and high humidity conditions, replace it with a revised aging model.

[0107] Rule 4: A new whole-part relationship is generated when a component first exhibits a provable functional association within a component group. For example, a seal, gas diffusion layer, membrane electrode assembly, and bipolar plate initially exist as independent objects, only forming a stack as a new functional whole system during the pressing process; and during operation, characteristics such as membrane swelling, changes in contact pressure, changes in catalyst layer, and changes in performance parameters interact with each other, thus continuously changing the object history and generating new whole-part semantic relationships.

[0108] The expansion of the knowledge base is not achieved by arbitrarily adding more data, but rather by rules-based identification of new semantic relationships between state variables, process transitions, whole-part relationships, and the resulting changes in product characteristics. Only these new relationships are incorporated into the knowledge architecture as new or extended knowledge elements.

[0109] S5: Digital Twins and Long-Term Forecasting

[0110] A digital twin is constructed from the object's history, describing not only the current state of the fuel cell stack but also all known states, their formation processes, and the underlying causal relationships. This allows for the prediction of the stack's future performance, aging process, quality development, and remaining service life during the manufacturing process.

[0111] The digital twin loads full lifecycle time-series data from the complete object's history, and combines it with physical aging models and AI prediction algorithms to simulate the evolution processes such as performance degradation, seal aging, and catalyst degradation under long-term operation. The prediction results are written back to the object's history and participate in subsequent knowledge reasoning and control decisions.

[0112] After the fuel cell stack is put into operation, operational data is collected in real time, and the object history is continuously updated. Based on a hierarchical knowledge architecture, product performance, efficiency, failure probability, and remaining service life are predicted in real time, and maintenance and correction strategies are automatically generated. Repair and replacement data are written into the history, updating the aging model and life prediction. During the end-of-life recycling stage, dismantling, testing, and material recycling data are entered into the history, creating a complete closed-loop record of the product's entire life cycle.

[0113] All new data, new patterns, and optimization experiences throughout the entire process are continuously fed back into the hierarchical knowledge architecture, expanding the knowledge base, iterating the AI ​​model, and improving the material and aging models, forming a permanent self-learning system.

[0114] The foregoing description of specific exemplary embodiments of the invention is for illustrative and explanatory purposes. These descriptions are not intended to limit the invention to the precise forms disclosed, and it will be apparent that many changes and variations can be made in accordance with the foregoing teachings. The exemplary embodiments were chosen and described in order to explain the specific principles of the invention and its practical application, thereby enabling those skilled in the art to implement and utilize various different exemplary embodiments of the invention, as well as various different choices and variations. The scope of the invention is intended to be defined by the claims and their equivalents.

Claims

1. A self-control method based on hierarchical knowledge architecture and object history inheritance, characterized in that, Includes the following steps: S1: Construct a multi-level progressive semantic knowledge architecture, which consists of a seven-layer progressive structure from bottom to top: raw data layer, state variable layer, natural science model layer, domain knowledge layer, artificial intelligence layer, decision control layer and object history layer. Each knowledge object is configured with a unique identifier, attribute definition, semantic relationship and state transition rules. The layers are connected by semantic association to realize progressive reasoning from raw data to autonomous decision-making. S2: Assign a unique and unchangeable object identifier to each technical object, create a corresponding object history and complete initialization. Initialization data includes raw material parameters, supplier data, incoming characterization test data and initial state variables. The object identifier remains unchanged throughout the entire life cycle, achieving complete correspondence of all state information. S3: At each process stage, raw measurement data is collected and converted into state variables with physical meaning. The state variables and their semantic classifications are written into the object history. Consistency checks, model mapping and knowledge reasoning are completed through a hierarchical knowledge architecture. The personalized control parameters and optimization strategies for the single product are generated autonomously, and the decision results are fed back to the object history as new knowledge elements. S4: The updated complete object history is standardized and passed to the downstream process. The downstream process carries out its control, adjustment and decision-making based on the inherited complete upstream history. At the same time, the newly added status information, process deviation and performance data of the downstream process are fed back to the upstream process, driving the upstream process parameters to self-optimize and adjust and the knowledge base to expand iteratively. Among them, the updated object life cycle is passed to at least one subsequent method step in the form of state and knowledge information inheritance.

2. The autonomous control method based on hierarchical knowledge architecture and object history inheritance as described in claim 1, characterized in that, In step S1, the unique identifier of each knowledge object is generated by combining the product ID, process step, timestamp, and version number. Each knowledge element can be explicitly referenced through this unique identifier. The attributes of the knowledge object include physical characteristics, process parameters, measured values, uncertainties, quality characteristics, and calculated state variables. The semantic relationships between knowledge objects include time predecessor / successor relationships, cause-effect relationships, whole-part relationships, state transition relationships, and deductive reasoning relationships.

3. The autonomous control method based on hierarchical knowledge architecture and object history inheritance as described in claim 1, characterized in that, In step S1, the object history layer is the seventh layer. It is not an independent data source, but an integration of all information from the first to the sixth layers in the time dimension. It stores the complete time-series document of the temporal state change process, process conversion records, measurement values, decision results and knowledge elements derived from them throughout the entire life cycle. The object history not only stores the original data, but also stores the semantic classification of the data and the knowledge elements that are valid at that time.

4. The autonomous control method based on hierarchical knowledge architecture and object history inheritance as described in claim 1, characterized in that, The reasoning process of the decision control layer in step S3 is executed according to the following logical steps: Read the semantically described state variables; Perform consistency and reasonableness checks; Map state variables to semantic models and object histories; Reasoning is performed using domain knowledge, physical models, and rules; Evaluate multiple possible course of action; Choose the optimal strategy; Output target values, control variables, and action recommendations; The decision results are fed back into the object's resume as new knowledge elements.

5. The autonomous control method based on hierarchical knowledge architecture and object history inheritance according to claim 1, characterized in that, The iterative expansion of the knowledge base in step S4 follows these four rules: Rule 1: When a previously unknown relationship is identified between process state and product characteristics, a new knowledge element is generated; Rule 2: Expand an existing knowledge element when new measurement data confirms or quantitatively refines its statement. Rule 3: When new knowledge indicates that a previous description is incomplete or incorrect, replace one knowledge element; Rule 4: When a component first demonstrates a provable functional association within a component group, a new whole-part relationship is generated.

6. The autonomous control method based on hierarchical knowledge architecture and object history inheritance according to claim 1, characterized in that, The sources of initial data for the object history in step S2 include: for objects manufactured within the enterprise's internal value chain, the initial data begins with their respective raw materials and manufacturing process data; for purchased objects, the initial data uses the specifications, quality data, and batch information provided by the supplier; all objects are recorded through their own characterization and testing procedures before further processing, including mechanical, physical, chemical, and electrochemical studies and microscopic structural analysis methods. The measurement data obtained thereby, together with the supplier data, constitute the initial state space of the object history.

7. The autonomous control method based on hierarchical knowledge architecture and object history inheritance according to claim 1, characterized in that, The object identifier is derived from the supplier batch number or serial number, internally assigned by the enterprise, or generated by a combination of multiple features such as supplier information, batch, timestamp, material category, and serial number; the carrier form of the identifier includes QR code, RFID, DataMatrix code, or universally unique identifier UUID, and each physical object clearly corresponds to at least one independent digital object lifecycle.

8. The autonomous control method based on hierarchical knowledge architecture and object history inheritance according to claim 1, characterized in that, In step S3, each decision made and its effect are written into the object's history as a new knowledge element, becoming part of the object's history and expanding the semantic knowledge model for subsequent decisions, forming a recursive knowledge growth mechanism that combines object history, semantic ontology, and autonomous decision-making. Knowledge objects are not merely stored, but are continuously re-evaluated semantically, so that each state change can generate new knowledge. By identifying new causal relationships, performing semantic classification, and persistently integrating them into the knowledge base, structured knowledge is continuously developed.

9. The autonomous control method based on hierarchical knowledge architecture and object history inheritance according to claim 1, characterized in that, It also includes the following steps: S5: Construct a digital twin based on the object's history. The digital twin not only describes the current state of the technical object, but also includes all known states, the state formation process, and the underlying causal relationships. By loading the complete object history, long-term predictions of the product's future performance, aging process, quality development, and remaining service life are achieved. The prediction results are written back to the object history and participate in subsequent knowledge reasoning.

10. The autonomous control method based on hierarchical knowledge architecture and object history inheritance according to any one of the preceding claims, characterized in that, The method is used for the manufacturing, assembly, testing, operation management, condition monitoring, maintenance and / or recycling of complex technology products, as well as for technology systems used in the multi-stage manufacturing of fuel cells, electrolyzers, batteries, semiconductor products, chemical process systems, machines, vehicle parts or similar products.

Images