An industrial metamodel autonomous intelligent construction method and base architecture

Abstract

The application relates to the technical field of industrial meta-model construction, and discloses an industrial meta-model autonomous intelligent construction method and a base architecture, which comprises completing meta extraction, completing meta division, completing meta reasoning and completing meta tracing. According to the application, the standardization of industrial knowledge and meta granularity expression are realized through meta extraction, meta knowledge clusters with continuous mechanism and clear boundaries are formed through meta division, the meta knowledge clusters are further encapsulated into industrial meta models with clear structure, clear interface and identifiable function through meta reasoning, and a combinable, reasonable and reusable industrial meta model knowledge graph is constructed; meanwhile, the construction process and results are subjected to credible verification, abnormality identification and root cause tracing through meta tracing, so that an all-process industrial meta model intelligent construction framework covering the standardized expression, automatic organization, model construction, system expansion and credible verification of industrial knowledge is formed.

Description

Technical Field

[0001] This application relates to the field of industrial metamodel construction technology, specifically an autonomous intelligent construction method and base architecture for industrial metamodels. Background Technology

[0002] With the rapid development of next-generation artificial intelligence technologies, represented by large-scale models, the fundamental role of data in enhancing model capabilities and implementing applications is becoming increasingly prominent. Compared to simply relying on computing power expansion or model structure optimization, building high-quality datasets has become a key support for driving the continuous evolution of artificial intelligence systems. Especially in the industrial sector, a large amount of knowledge exists in the form of technical documents, rules of experience, etc. Without unified organization and high-quality expression, it is difficult to effectively serve model training and intelligent reasoning. Therefore, how to construct industrial knowledge resources that are clearly structured, semantically accurate, and computable has become an important issue for the development of intelligent technologies in the industrial field. Summary of the Invention

[0003] The purpose of this application is to provide an autonomous intelligent construction method and base architecture for industrial meta-models, so as to solve the technical problems of multi-source heterogeneity of industrial knowledge, unclear organizational boundaries, difficulty in model encapsulation, and difficulty in verifying and tracing the construction results in the prior art.

[0004] To achieve the above objectives, this application provides an autonomous intelligent construction method for an industrial meta-model, comprising: The system performs information hot structure based on semantic chain and meta-granular reconstruction based on knowledge fiber bundle to transform multi-source heterogeneous industrial information into industrial meta-knowledge with unified form and granularity, thus completing meta-extraction; wherein, the multi-source heterogeneous industrial information includes text, data, images and expert experience. Perform knowledge distillation-based construction of industrial meta-knowledge module space and homotopy-based partitioning of industrial meta-knowledge clusters to obtain industrial meta-knowledge clusters and complete meta-partitioning; Perform knowledge distribution-driven generative meta-model construction to obtain industrial meta-models, perform potential energy constraint-based meta-model knowledge graph construction, associate the obtained industrial meta-models to obtain industrial meta-model knowledge graphs, and complete meta-reasoning. The system performs hierarchical Markov-driven generative chain construction and structural causal model-driven conflict root cause tracing to determine the root cause of the conflict and complete meta-tracing. Based on the completion status of meta-tracing, it provides feedback and optimizes meta-extraction, meta-partitioning, and meta-reasoning.

[0005] To achieve the above objectives, this application also provides an industrial meta-model foundation architecture, which applies the autonomous intelligent construction method of the industrial meta-model described above, including: The meta-extraction module is used to perform information hot structure based on semantic chain and meta-granular reconstruction based on knowledge fiber bundle, transforming multi-source heterogeneous industrial information into industrial meta-knowledge with unified form and granularity, and completing meta-extraction; wherein, the multi-source heterogeneous industrial information includes text, data, images and expert experience. The meta-partitioning module is used to perform the construction of the industrial meta-knowledge module space based on knowledge distillation and the partitioning of industrial meta-knowledge clusters based on homotopy relations, thereby obtaining industrial meta-knowledge clusters and completing the meta-partitioning. The meta-reasoning module is used to perform knowledge distribution-driven generative meta-model construction to obtain industrial meta-models, perform potential energy constraint-based meta-model knowledge graph construction, associate the obtained industrial meta-models to obtain industrial meta-model knowledge graphs, and complete meta-reasoning. The meta-source tracing module is used to perform hierarchical Markov-driven generative chain construction and structural causal model-driven conflict root cause tracing, determine the root cause of the conflict to complete the meta-source tracing, and based on the completion status of the meta-source tracing, provide feedback and optimize meta-extraction, meta-partitioning and meta-reasoning.

[0006] The autonomous intelligent construction method and base architecture for the industrial meta-model of this application achieve at least the following beneficial effects: To achieve standardized expression of industrial knowledge; to unify industrial knowledge scattered from different sources, in different forms and at different granularities into industrial meta-knowledge with clear structure, consistent semantics, and computability, thereby improving the standardization and reusability of industrial knowledge resources.

[0007] Achieve orderly organization and boundary definition of industrial knowledge; automatically aggregate massive industrial meta-knowledge into meta-knowledge clusters with stable internal relationships, continuous mechanisms, and clear boundaries, avoiding the impact of knowledge mixing, duplication, and incorrect classification on subsequent modeling; To achieve automated construction and systematic organization of industrial meta-models; to encapsulate scattered local mechanistic relationships, variable constraints and engineering rules into industrial meta-models with stable structure, clear interfaces and clear functions; and to further establish compatibility and combination relationships between different industrial meta-models, forming an industrial meta-model knowledge graph that is reasonable, searchable and sustainably expandable, thereby supporting the automatic construction, combination and invocation and continuous expansion of complex industrial models. To achieve reliable verification and conflict tracing of the results of industrial meta-model construction; to verify the correctness, rationality and reliability of the meta-model and its relationships, and to locate the source of the problem when anomalies or conflicts occur, thereby improving the transparency, reliability and traceability of the automated construction process of industrial knowledge. Attached Figure Description

[0008] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0009] Figure 1 A flowchart illustrating the autonomous intelligent construction method for industrial meta-models provided in this application embodiment; Figure 2 A flowchart illustrating the autonomous intelligent construction method for industrial meta-models provided in this application embodiment; Figure 3 A diagram illustrating the meta-extraction process provided in this application embodiment; Figure 4 A diagram illustrating the meta-partitioning process provided in this application embodiment; Figure 5 Meta-inference process diagram provided for embodiments of this application; Figure 6 A diagram illustrating the meta-tracing process provided in this application's embodiments; Figure 7 This is a structural block diagram of the industrial meta-model base architecture provided in the embodiments of this application.

[0010] The implementation, functional features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0011] The technical solutions in the embodiments of this application will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0012] In this document, the term "comprising" is intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0013] In response to the technical challenges of multi-source heterogeneity, unclear organizational boundaries, difficulties in model encapsulation, and the inability to verify and trace the construction results of industrial knowledge in existing technologies, current technologies have explored automated applications of industrial knowledge, such as industrial knowledge graph construction and domain mechanism knowledge base organization. However, significant limitations remain in scenarios involving the automated and intelligent construction of industrial meta-models. Specifically, the main shortcomings are as follows: In a major aspect, the shortcomings are as follows: the focus is on the graph fusion of multi-source industrial knowledge, without forming a reusable modular knowledge representation; the graph relationships mainly reflect the associations between entities, making it difficult to express the high-order coupling relationships and combinatorial dependencies between multiple industrial knowledge; and there is a lack of a combinatorial judgment mechanism oriented towards model interfaces, functional semantics, and mechanistic compatibility, making it difficult to guarantee the compatibility and composability between knowledge and models.

[0014] In another major aspect, the shortcomings are as follows: it mainly relies on preset ontology levels and triplet filling methods to organize knowledge, making it difficult to reconstruct multi-source heterogeneous industrial knowledge in a unified granularity; the knowledge boundary is mainly determined by the ontology structure and artificial rules, lacking a mechanism for automatic division based on knowledge distribution, mechanism continuity and internal relationships; the organization of knowledge content serves more for retrieval and expression, making it difficult to further complete the automatic encapsulation from knowledge units to computable industrial models.

[0015] Against this backdrop, establishing a seamless connection across the entire chain from data to knowledge, and then to models and reasoning, has become a crucial direction for achieving intelligent knowledge automation. By structuring and systematically organizing massive amounts of industrial information, gradually transforming it into knowledge units capable of expressing mechanistic relationships, and further forming a computable and composable model-based knowledge system, not only can artificial intelligence systems be better understood and reasoned about complex industrial problems, but high-quality corpus support can also be provided for training large models, thereby promoting the in-depth application of artificial intelligence in industrial scenarios.

[0016] In response to the detailed analysis above, this embodiment discloses an autonomous intelligent construction method and base architecture for industrial meta-models. The purpose is to form an intelligent construction framework for industrial meta-models that covers standardized expression of industrial knowledge, automated organization, meta-model construction, model system expansion, and credible verification and traceability. This framework provides a high-quality, computable, and verifiable model-based knowledge foundation for industrial large-scale model training, industrial knowledge reasoning, and intelligent modeling of complex industrial systems.

[0017] The autonomous intelligent construction method of the industrial meta-model in this embodiment will now be described.

[0018] Reference Figure 1 , Figure 1 A flowchart illustrating the autonomous intelligent construction method for industrial meta-models provided in this application embodiment.

[0019] Firstly, such as Figure 1 As shown in the figure, this embodiment discloses an autonomous intelligent construction method for an industrial meta-model, including: S10: Execute information hot structure based on semantic chain and meta-granular reconstruction based on knowledge fiber bundle to transform multi-source heterogeneous industrial information into industrial meta-knowledge with unified form and granularity, and complete meta-extraction; among which, multi-source heterogeneous industrial information includes text, data, images and expert experience.

[0020] S20: Perform the construction of the industrial meta-knowledge module space based on knowledge distillation and the partitioning of the industrial meta-knowledge clusters based on homotopy relations to obtain the industrial meta-knowledge clusters and complete the meta-partitioning.

[0021] S30: Perform knowledge distribution-driven generative meta-model construction to obtain industrial meta-models, perform potential energy constraint-based meta-model knowledge graph construction, associate the obtained industrial meta-models to obtain industrial meta-model knowledge graphs, and complete meta-reasoning.

[0022] S40: Execute hierarchical Markov-driven generative chain construction and structural causal model-driven conflict root cause tracing, determine the root cause of the conflict to complete meta-tracing, and based on the completion status of meta-tracing, provide feedback and optimize meta-extraction, meta-partitioning and meta-reasoning.

[0023] Reference Figure 2 , Figure 2 A flowchart illustrating the autonomous intelligent construction method for industrial meta-models provided in this application embodiment.

[0024] Combination Figure 2 The autonomous intelligent construction method for the industrial meta-model proposed in this embodiment includes meta-extraction, meta-partitioning, meta-reasoning, and meta-originating. S10 to S40 of the autonomous intelligent construction method for the industrial meta-model in this embodiment will now be described in detail.

[0025] Reference Figure 3 , Figure 3 This is a diagram illustrating the meta-extraction process provided in an embodiment of this application.

[0026] Regarding the meta-extraction of S10, combined with Figure 3 As shown, meta-extraction is the process of identifying and extracting industrial knowledge from multi-source heterogeneous industrial information and outputting it as industrial meta-knowledge with unified form and granularity, providing standardized input for meta-partitioning and meta-reasoning. Meta-extraction mainly includes a semantic chain-based information deconstruction stage and a knowledge fiber bundle-based meta-granularity reconstruction stage.

[0027] Specifically, information one-hot structures based on semantic chains include: By mapping multi-source heterogeneous industrial information to a unified semantic space, an industrial semantic representation matrix is ​​obtained to characterize industrial knowledge factors.

[0028] In this embodiment, the multi-source industrial information set is: , Representing text modality, Representing time-series structured data modes, The image modality is represented. Through multimodal semantic encoding and fusion, it is mapped to a unified semantic space to obtain the industrial semantic representation matrix: (1) in, Indicates the first A unified semantic representation of industrial knowledge factors.

[0029] Based on the causal dependence, process constraints, and interaction relationships of industrial knowledge factors, an industrial information semantic chain is constructed.

[0030] In this embodiment, the constructed industrial information semantic chain is as follows: (2) in: (3) in: Represents the semantic chain of industrial information; This represents a set of semantic chain nodes, with each node corresponding to an industrial knowledge factor. Represents a set of semantic chain edges; Represents a node and The types of relationships between them include causal relationships, constraint relationships, sequence relationships, and action relationships; Indicates in relation type Next node point to The strength of semantic association; Indicates the edge generation threshold; This indicates vector concatenation; Indicates element-wise product; and These are learnable parameters; This represents a non-linear activation function.

[0031] Guided by different prior knowledge, orthogonal projection is performed on the semantic chain of industrial information to generate multiple industrial knowledge views. The generation of industrial knowledge views satisfies preset optimization constraints, which are used to control semantic preservation loss, semantic coupling, and consistency distance between the industrial knowledge view and the corresponding prior knowledge.

[0032] In this embodiment, the generated industrial knowledge views are as follows: (4) The corresponding preset optimization constraints can be those that satisfy the following optimization constraints: (5) in: Indicates the first Orthogonal industrial knowledge view; This represents a semantic propagation function based on a semantic chain structure, used to fuse node semantics and link relationships; Represents the orthogonal projection function guided by knowledge; This represents semantic preservation loss, used to ensure that the original industrial semantics are not lost after the knowledge view is projected. Orthogonal constraints are used to represent different knowledge views, which are used to reduce semantic coupling between knowledge views; Represents a knowledge view Corresponding industrial a priori Consistent distance between them; This represents the weighting coefficient.

[0033] One-hot masking is performed on the industrial knowledge view to obtain the semantic attribution mask of each industrial knowledge factor, so as to deconstruct the semantic chain of industrial information.

[0034] In this embodiment, a one-hot mask is applied to the generated orthogonal industrial knowledge view to obtain the semantic attribution mask for each knowledge factor. And based on this, the original semantic chain is deconstructed: (6) in: Indicates the first Does the knowledge factor belong to the first...? Industry-like knowledge view; Indicates the first The industrial semantic chain after class deconstruction; and These represent the set of nodes and the set of edges in the deconstructed semantic chain, respectively.

[0035] The deconstructed industrial information semantic chain is parametrically mapped, and the industrial knowledge factors, attributes, relationships and constraints are uniformly transformed into a parametric industrial knowledge embryo, thus completing the information unique structure based on the semantic chain.

[0036] In this embodiment, the uniformly transformed parameterized industrial knowledge roughness is: (7) in: , Indicates the first A fine-grained industrial knowledge rough embryo; This represents a knowledge parameterization mapping function; Represents a node In the The set of neighboring nodes in the semantic chain is used to provide the contextual relationships of the knowledge factor; Indicates the source of knowledge; Represents industrial knowledge objects; Represents a collection of object property parameters; This represents the set of relationships between the knowledge object and other objects; This represents the set of rules or constraints corresponding to the knowledge object. This represents the final set of fine-grained industrial knowledge prototypes.

[0037] Specifically, meta-granular reconstruction based on knowledge fiber bundles includes: Driven by a large model, with the local triviality and fibrillation of industrial knowledge as pre-defined constraints, and with the constructed parameterized industrial knowledge embryo as input and industrial meta-knowledge as the reconstruction target, the discrete parameterized industrial knowledge embryo is subjected to generative reconstruction and evolutionary reasoning, and the output is candidate industrial meta-knowledge that meets the pre-defined formal requirements.

[0038] The technical principle of the meta-granular reconstruction based on knowledge fiber bundles in this embodiment will now be explained.

[0039] In the process of meta-granular reconstruction, a parameterized industrial knowledge rough embryo is used as input. Through knowledge reconstruction, it is organized into industrial meta-knowledge with a unified structure and consistent granularity. To this end, this embodiment first clarifies the basic properties that industrial knowledge should satisfy, and then further gives the basic form of industrial meta-knowledge based on this.

[0040] Industrial meta-knowledge is not a simple aggregation of primitive knowledge, but rather a standardized knowledge unit with stable structure and uniform granularity, meaning that industrial knowledge satisfies local triviality and fibrousness.

[0041] Local triviality is used to ensure that industrial knowledge has a consistent structural representation within the local parameter domain. Let... For industrial knowledge parameter source space, Representing examples of industrial knowledge, This represents the source mapping from industrial knowledge instances to knowledge parameter sources. For any local parameter domain... There exists an isomorphic mapping: (8) in: Represents the local parameter domain The corresponding knowledge set; Represents a unified knowledge space; Represents a local triviality mapping; Indicates from arrive The first projection mapping. This property shows that knowledge can be computed and accumulated within any local parameter domain.

[0042] Fiberity refers to the correspondence between knowledge and its parameter sources, meaning that every piece of knowledge evolves from a knowledge source. It is defined as follows: (9) in: It represents a specific source of knowledge at the underlying level; express The collection of all knowledge generated from the source (i.e., the fibers on that source). To represent a specific instance of industrial knowledge; Indicates and All knowledge from the same origin constitutes an equivalent class. This property constrains the knowledge generated by large models to have a clear evolutionary origin path.

[0043] Based on the local triviality and fibrousness of knowledge, the basic form of industrial meta-knowledge is defined as follows: (10) in: express The inverse mapping; Indicates the first A knowledge parameter source. Driven by a large model, with the local triviality and fibrillation of knowledge as mathematical constraints, and with the knowledge parameters in the constructed industrial knowledge parameter base space as input, and with industrial meta-knowledge as the reconstruction target, the discrete parameters are generatively reconstructed and evolutionarily reasoned, and finally output candidate industrial meta-knowledge that meets the formal requirements.

[0044] The candidate industrial meta-knowledge undergoes meta-granularity testing and correction. The testing process includes a mechanism completeness test to determine whether the knowledge can be derived from a single parameter source, an information gain test to determine whether the knowledge is indivisible, and a knowledge independence test to determine whether the knowledge depends solely on its own parameter source without causing cross-source drift. If the test fails, the feedback information is returned to the reconstruction process to adjust the knowledge parameter mapping or combination relationship and regenerate the meta-granularity until industrial meta-knowledge that meets the preset constraints and formal requirements is obtained.

[0045] In this embodiment, after obtaining candidate meta-knowledge, it undergoes meta-granularity testing and correction. The testing process includes mechanism completeness testing, information gain testing, and knowledge independence testing. Mechanism completeness testing determines whether knowledge can be derived from a single parameter source; information gain testing determines whether knowledge is indivisible; and knowledge independence testing determines whether knowledge depends solely on its own parameter source without cross-source drift. If the tests fail, the feedback information is returned to the reconstruction process to adjust the knowledge parameter mapping or combination relationship, and meta-granularity generation is performed again until industrial meta-knowledge that satisfies the constraints is obtained.

[0046] Reference Figure 4 , Figure 4 A diagram illustrating the meta-partitioning process provided in this application embodiment.

[0047] Regarding the meta-partition of S20, combined with Figure 4 As shown, the meta-partition takes the massive industrial meta-knowledge obtained by meta-extraction as input, quantifies the similarity between meta-knowledge and maps it to a unified modulus space, mines the knowledge evolution path and defines the cluster boundary in this space, and finally generates a knowledge homotopy cluster with internal regularity and order and clear external boundary.

[0048] Specifically, the construction of an industrial meta-knowledge module space based on knowledge distillation aims, in part, to construct a unified spatial representation that transforms the industrial meta-knowledge obtained in the meta-extraction stage into a comparable and organizeable representation. This embodiment constructs the industrial meta-knowledge module space from two aspects: knowledge similarity measurement and meta-knowledge basis distillation, including: Each industrial meta-knowledge is represented as a distribution on a shared fiber pattern through dictionary learning.

[0049] In the application of this embodiment, the first The activation coefficient of individual industrial meta-knowledge in the shared fiber model is: Then its normalized distribution is expressed as: (11) in: Indicates the number of shared fiber patterns. To prevent extremely small positive numbers with a denominator of zero, this distribution representation transforms structural differences between meta-knowledge into distributional differences.

[0050] For any two industrial meta-knowledges, calculate the knowledge difference entropy to obtain the similarity.

[0051] For any two industrial meta-knowledge and The corresponding activation distribution is and The Jensen-Shannon divergence was used to calculate the knowledge difference entropy. And further obtain similarity : (12) in: This is the similarity adjustment coefficient. The larger the value, the closer the distribution of the two meta-knowledge items on the shared fiber pattern, and the higher the structural similarity.

[0052] After obtaining similarity, a meta-knowledge base is extracted from all industrial meta-knowledge using knowledge distillation driven by a large model.

[0053] In this embodiment, the meta-knowledge base set is: ,in Indicates the first Individual knowledge base Let be the total number of meta-knowledge bases. The distillation process is constrained by representation sparsity, basis orthogonality, and knowledge reproducibility, ensuring that the meta-knowledge bases can both represent the original knowledge structure and reduce redundancy between bases. This process can be represented as: (13) In this embodiment, a large-model-driven knowledge distillation method is used to extract a representative meta-knowledge base from all industrial meta-knowledge. The meta-knowledge base serves as a reference direction for a unified space, ensuring that each piece of meta-knowledge no longer exists merely as pairwise similarities, but can be mapped to a coordinate representation in the unified space. Let the set of meta-knowledge bases be: where, Indicates the first The weight vector of each industrial meta-knowledge on each meta-knowledge basis. , and represents knowledge reproduction, and represents the weight coefficients between sparsity and basis orthogonality, respectively.

[0054] Based on the extracted meta-knowledge base, and combined with the weights corresponding to the preset meta-knowledge base, an industrial meta-knowledge module space is constructed.

[0055] In this embodiment, the meta-knowledge base set and weight representation Constructing the industrial meta-knowledge module space: (14) in, This represents the modular space mapping operator. Through this mapping, all industrial meta-knowledge is transformed into coordinate points in a unified modular space. Similar meta-knowledge is close in space, while meta-knowledge with large differences is separated in different basis directions, thus providing a computable spatial basis for subsequent homotopy cluster generation and boundary determination.

[0056] Specifically, the industrial meta-knowledge cluster partitioning based on homotopy relations aims to identify sets of meta-knowledge with continuous evolutionary relationships from the industrial meta-knowledge module space. This process takes meta-knowledge coordinate points as input and, through steps such as manifold projection, evolutionary path search, and boundary segmentation, obtains knowledge clusters with homotopy continuity; including: By projecting the industrial meta-knowledge module space into a low-dimensional manifold, the meta-knowledge that was originally distributed in the high-dimensional module space is divided into several locally continuous low-dimensional manifolds.

[0057] In this embodiment, the low-dimensional manifold projection is performed as follows: (15) in, Represents the industrial meta-knowledge model space. Represents the manifold projection operator. Represents the low-dimensional coordinates after projection. Indicates the first A low-dimensional manifold, This represents the total number of low-dimensional manifolds. Through this projection, similar meta-knowledge with local continuity is concentrated into the same low-dimensional manifold, providing a foundation for subsequent path search.

[0058] Within a low-dimensional manifold, connection costs are constructed based on the magnitude of adjacent changes between meta-knowledge, and knowledge evolution paths are sought based on the principles of minimizing the total path increment and minimizing adjacent jumps; the set of path nodes that meet the continuity requirement is determined as candidate homotopy clusters.

[0059] In this embodiment, in low-dimensional manifolds Internally, connection costs are constructed based on the magnitude of adjacent changes between meta-knowledge, and knowledge evolution paths are sought based on the principles of minimizing the total path increment and minimizing adjacent jumps. (16) in, Representing a low-dimensional manifold The knowledge evolution path in China This represents the connection cost between two adjacent pieces of knowledge in the path. Let be the number of element-wise knowledge in this manifold. The weighting parameter represents the path. This path reflects the evolutionary order of meta-knowledge from near to far and from continuous to variable. The set of path nodes that meet the continuity requirement is determined as the candidate homotopy cluster.

[0060] The mechanism change region is extracted based on the sorting jump point and the cluster outer edge point, and a mechanism constant section is constructed to characterize the location where the continuity of mechanism is broken during the knowledge evolution process; a regular constraint surface is constructed by combining expert experience preferences and domain constraints; a regular constant section for boundary determination is formed based on the mechanism constant section and the regular constraint surface.

[0061] In principle, since candidate homotopy clusters may still contain nodes with mechanistic changes or structural mutations, further boundary determination is required. In the knowledge evolution path, when there is a significant jump in the connection cost between adjacent meta-knowledge points, it usually means that the corresponding process mechanism, equipment operating state, or parameter interaction relationship has changed; simultaneously, meta-knowledge points located at the cluster edge are often at transitional positions between different evolutionary branches. Specifically, mechanistic change regions are extracted based on sorting jump points and cluster outer edge points, and mechanistic constant cross-sections are constructed in these regions. This is used to characterize the locations where the continuity of mechanisms is disrupted during knowledge evolution. Furthermore, a regularized constraint surface is constructed by combining expert experience preferences and domain constraints. Together, they form a regular constant cross section used for boundary determination. : (17) If the connecting segment between adjacent meta-knowledge passes through a regular constant section, then the connecting segment is considered to have a boundary, and the candidate homotopy cluster needs to be segmented; if it does not pass through the regular constant section, then the connecting segment is considered to maintain a homotopy continuity.

[0062] In this embodiment, the determination of the boundary split point is as follows: (18) in, This indicates that the location is a boundary dividing point. This indicates that the position remains continuous.

[0063] Based on all the split points, the candidate homotopy clusters are branched and cut, and the connectivity of the sub-clusters is judged until the continuity requirement is met within each sub-cluster, resulting in a set of knowledge homotopy clusters that are smooth and ordered within the clusters and have clear boundaries between the clusters.

[0064] Thus, the set of knowledge homotopy clusters in this embodiment provides a structured foundation for subsequent meta-model reasoning, knowledge graph construction, and attribution of new meta-knowledge.

[0065] Reference Figure 5 , Figure 5 A diagram illustrating the meta-inference process provided in this application embodiment.

[0066] Regarding the meta-reasoning of S30, combined with Figure 5 As shown, the meta-model reasoning part takes the meta-knowledge clusters obtained in the meta-partitioning stage as input, and constructs an industrial meta-model by organizing, reasoning and constraining the meta-knowledge within the clusters and their relationships, and further forms an industrial meta-model knowledge graph.

[0067] Specifically, knowledge distribution-driven generative meta-model construction includes: The backbone center of the industrial meta-knowledge cluster is estimated based on the knowledge distribution within the cluster.

[0068] In this embodiment, a unified representation mapping is performed on each industrial meta-knowledge within the meta-knowledge cluster to obtain its vector representation in the knowledge module space. The backbone center of the meta-knowledge cluster, i.e., the cluster center in the knowledge module space, is estimated based on the knowledge distribution within the cluster. For the first... Individual knowledge is used to calculate the degree of deviation from the trunk center and the degree of local support within the cluster: (19) in, Indicates the first Vector representation of individual knowledge Indicates the core center within the cluster. This is used to measure the degree to which the meta-knowledge deviates from the cluster backbone. It is used to measure the degree to which this meta-knowledge is supported by other meta-knowledge in the same cluster.

[0069] For an industrial meta-knowledge, determine the degree of global deviation and local support of the industrial meta-knowledge relative to the backbone center; combine the degree of global deviation and local support to obtain the backbone confidence of the industrial meta-knowledge; based on the backbone confidence, combined with the preset backbone confidence threshold, knowledge is proposed to obtain a meta-knowledge cluster backbone with consistent structure, continuous mechanism and clear boundary.

[0070] Based on the global deviation and local support levels, and by combining both, a meta-knowledge backbone confidence level is constructed: (20) in, The distance sensitivity coefficient is used. If the confidence of the backbone of a certain meta-knowledge is lower than a preset threshold, it is marked as a candidate anomalous knowledge. Subsequently, the candidate anomalous knowledge is further subjected to semantic support verification, that is, to determine whether it can be reasonably supported by other meta-knowledge in the same cluster in terms of mechanism relationships, variable constraints, or engineering rules. If the meta-knowledge is significantly deviated from the backbone distribution within the cluster and lacks semantic support, it is judged as anomalous knowledge and is removed; if it is on the distribution boundary but still has a clear mechanism support relationship with the knowledge within the cluster, it is retained as special working condition knowledge or boundary knowledge. In this way, a meta-knowledge cluster backbone with consistent structure, continuous mechanism, and clear boundaries is obtained, providing reliable input for subsequent meta-model generation.

[0071] Based on the distribution and mechanistic relationships of various industrial meta-knowledge in the industrial meta-knowledge model space, the phylogenetic dependencies between meta-knowledge are identified, and the backbone of the meta-knowledge cluster is organized into a hypergraph structure. The hypergraph structure is input into the hypergraph reasoning model, and the meta-knowledge nodes and their corresponding hyperedge relationships are jointly encoded and combined inference to generate candidate meta-model structures. Among them, the combined inference is constrained by the dynamic changes of state variables, the conservation of mass, the conservation of energy or the material balance relationship, and the constraints of parameter range, boundary conditions and engineering feasible region.

[0072] After obtaining the backbone of the meta-knowledge cluster, the resulting hypergraph structure is as follows: (twenty one) Among them, the node set The superedge set represents the industrial meta-knowledge within the backbone of the meta-knowledge cluster. This represents the high-order coupling or common constraint relationships formed between multiple industrial meta-knowledges. Through this hypergraph structure, multi-variable, multi-constraint, and multi-mechanism coupling relationships that are difficult to cover with a single meta-knowledge can be uniformly expressed. Subsequently, the hypergraph structure is input into the hypergraph inference model, where meta-knowledge nodes and their hyperedge relationships are jointly encoded and combined inferenced to generate candidate meta-model structures. During the inference process, a meta-model inference engine is introduced to constrain and guide candidate combined paths. The meta-model inference engine includes a dynamic evolution engine, a physical conservation engine, and a mechanism constraint engine; the dynamic evolution engine is used to constrain the dynamic changes of state variables, the physical conservation engine is used to constrain mass conservation, energy conservation, or material balance relationships, and the mechanism constraint engine is used to constrain parameter ranges, boundary conditions, and engineering feasible regions. For candidate paths that do not satisfy the above mechanism laws or engineering constraints, the inference model reduces their generation weight or excludes them directly.

[0073] After constraint-guided hypergraph reasoning, an industrial meta-model prototype structure is obtained, which represents the variable relationships, parameter relationships, mechanistic equations, computational logic, and constraint expressions in the industrial system. The prototype structure of the industrial meta-model is then interface-identified to determine the input variables, output variables, parameter terms, and boundary conditions required for model operation, thus forming the industrial meta-model interface. Based on the contribution of key industrial meta-knowledge during the reasoning process, the functional semantics represented by the model are sampled and summarized to form the functional attributes of the industrial meta-model. Finally, the prototype structure of the industrial meta-model, the industrial meta-model interface, and the functional attributes of the industrial meta-model are encapsulated to obtain the industrial meta-model.

[0074] In this embodiment, the prototype structure of the industrial metamodel is obtained after constraint-guided hypergraph reasoning. The prototype structure is used to represent the variable relationships, parameter relationships, mechanistic equations, computational logic, and constraint expressions in an industrial system. It can be represented in the form of a mathematical model, a rule model, or a code model. Furthermore, interface identification is performed on the metamodel prototype structure to determine the input variables, output variables, parameter terms, and boundary conditions required for model operation, thus forming the metamodel interface. Simultaneously, based on the contribution of key meta-knowledge in the reasoning process, the functional semantics represented by the model are sampled and summarized to form meta-model functional attributes. Finally, the metamodel prototype structure, metamodel interface, and metamodel functional attributes are uniformly encapsulated to obtain the industrial metamodel: (twenty two) Specifically, the construction of a meta-model knowledge graph based on potential energy constraints aims to characterize the composable relationships between different meta-models, including: For any two industrial meta-models to be associated, a compatibility potential function is constructed. This compatibility potential function outputs compatibility potential energy based on mechanistic consistency judgment, interface closure judgment, and functional synergy judgment. Among them: mechanistic consistency is used to judge whether the variable relationships, constraint relationships, and physical laws of the two industrial meta-models are contradictory; interface closure is used to judge whether the output of one industrial meta-model can meet the input requirements of the other industrial meta-model; functional synergy is used to judge whether the two industrial meta-models have synergistic effects on industrial tasks, process links, or control objectives.

[0075] In this embodiment, for any two meta-models to be correlated, their compatibility is evaluated from three aspects: mechanistic consistency, interface closure, and functional synergy. Mechanistic consistency is used to determine whether the variable relationships, constraint relationships, and physical laws of the two meta-models contradict each other; interface closure is used to determine whether the output of one meta-model can meet the input requirements of the other; and functional synergy is used to determine whether the two have synergistic effects on industrial tasks, process links, or control objectives. A compatibility potential function is constructed based on these factors: (twenty three) in, These represent the potential energy for mechanistic consistency, the potential energy for interface closure, and the potential energy for functional synergy, respectively. These correspond to the weighting coefficients. The lower the potential energy value, the stronger the compatibility between the two meta-models; the higher the potential energy value, the stronger the inconsistency or conflict between them.

[0076] Using compatibility potential as a supervision signal, the vector representation space of the industrial meta-model is shaped through comparative learning. After the representation space is shaped, candidate associations are selected based on the vector distance of the industrial meta-model, and interface closure verification is performed to form an industrial meta-model knowledge graph with industrial meta-models as nodes and compatibility combination relationships as edges.

[0077] In this embodiment, compatibility potential energy is used as a monitoring signal to perform comparative learning and shaping of the metamodel vector representation space. This ensures that highly compatible metamodels move closer to each other in the representation space, while metamodels with conflicting mechanisms, mismatched interfaces, or inconsistent functional goals move further apart. After shaping the representation space, candidate associations are selected based on the metamodel vector distance, and interface closure verification is further performed. If two metamodels meet a preset potential energy threshold and their input / output interfaces can form an effective connection, an association edge is established between them. If they have conflicting mechanisms, incomplete interface closures, or inconsistent functional goals, no combination edge is established, or they are marked as conflicting relationships. This forms a metamodel knowledge graph with industrial metamodels as nodes and compatible combination relationships as edges.

[0078] Reference Figure 6 , Figure 6 A diagram illustrating the meta-tracing process provided in this application embodiment.

[0079] Regarding the origins of S40, combined with Figure 6 As shown, meta-tracing takes the monitoring data of the entire process of meta-model construction as input, and combines hierarchical Markov chains and structural causal models to complete state link modeling, multi-dimensional cross-validation, anomaly identification, and chain-based root cause tracing, thereby realizing reliable verification and problem localization in the industrial meta-model construction process. Meta-tracing can be divided into two parts: state decision coding and hierarchical Markov chain construction, and multi-dimensional verification and chain-based causal tracing.

[0080] Specifically, hierarchical Markov-driven generative chain construction includes: Collect state data, decision-making behavior, and operational feedback information throughout the entire process of building the industrial meta-model, perform data denoising and standardization, and construct a continuous time series. Perform vectorization transformation on the operational state and various decision-making actions of the industrial meta-model, and then fuse the two types of vector features through a joint mapping function to generate a unified state-decision joint code.

[0081] In this embodiment, the computational formula for generating a unified state-decision joint code can be: (twenty four) in, Indicates the first Joint encoding of state and decision at time step. , These are the state vector and decision vector after encoding, respectively. For joint mapping functions, These are learnable parameters for the joint encoding process. This encoding transforms heterogeneous multi-source data into a standardized chain-like representation, providing a foundation for subsequent dynamic link modeling.

[0082] Using a temporally arranged joint encoding sequence as input, a hierarchical Markov chain is constructed by combining the multi-level architecture characteristics of the industrial meta-model. Pre-set interference terms are introduced to establish the state transition relationship between adjacent time steps.

[0083] In this embodiment, the preset interference term can be the introduction of environmental latent features, model latent factors, etc., thereby establishing the state transition relationship between adjacent time steps can be: (25) in, , These represent the meta-model running states at two adjacent time points. To drive decision-making actions that change state, For potential influencing factors such as environment and model latent features, This is the state transition function. These are the learnable parameters corresponding to the function. Hierarchical Markov chains can completely reproduce the state transitions, decision-driven processes, and temporal evolution patterns of multi-level meta-models.

[0084] The maximum likelihood estimation method is used to iteratively train the hierarchical Markov chain. The fitting effect of the model with the actual running link is evaluated by the log-likelihood function, and the state transition rules between levels are continuously optimized.

[0085] In this embodiment, the continuous optimization of the inter-level state transition rules can be expressed as: (26) in, Let be the log-likelihood function. This represents the total number of training samples. This represents the conditional probability of the next state occurring under the combined influence of the current state and the decision. The trained hierarchical Markov chain can serve as a benchmark model for end-to-end tracing.

[0086] Specifically, the structured causal model-driven conflict root cause tracing aims to achieve multi-dimensional verification and chain-like causal tracing, including: Based on the multi-level state topology and decision evolution link output by the hierarchical Markov chain, and combined with the structural causal model, multi-dimensional cross-validation is carried out to verify whether the relationship between the industrial meta-model and knowledge entities conforms to the industrial mechanism, process flow and physical constraints.

[0087] In this embodiment, the causal reasoning calculation formula is: (27) in, , The feature vectors of the entity to be verified and the meta-model are: Output the result for causal reasoning. For causal mapping functions, For model parameters, This represents the system noise term. By combining isomorphic constraints, logical boundary checks, and standard graph comparisons, the self-consistency of the meta-model system is comprehensively determined, and abnormal links with logical conflicts and mechanistic violations are marked.

[0088] For the abnormal links identified by cross-validation, the degree of abnormality of a single causal chain is quantified by an anomaly metric function to distinguish problematic sub-chains from normal links.

[0089] In this embodiment, the anomaly measurement function can be: (28) in, The causal link to be detected. For anomaly measurement functions, For measuring process parameters, This is the anomaly evaluation value for the causal chain. The higher the evaluation value, the more significant the errors and deviations in the chain.

[0090] Based on the results of anomaly quantification, and combined with counterfactual reasoning techniques, we trace back along the time-series path of the hierarchical Markov chain to locate the abnormal state nodes and erroneous decision-making links layer by layer, sort out the complete error propagation path and the root cause of the problem, and determine the root cause of the conflict.

[0091] In this embodiment, the traceability results are simultaneously sent back to the meta-extraction, meta-partition, and meta-inference front-end modules, forming a closed-loop mechanism of construction-verification-traceability-optimization. Finally, a traceability analysis report and optimization suggestions are output, comprehensively improving the reliability and traceability of the industrial meta-model system.

[0092] Therefore, it is clear that this embodiment incorporates at least the following key technical aspects, including: Industrial knowledge meta-extraction mechanism: This embodiment targets multi-source heterogeneous industrial information such as text, data, images and expert experience. First, it constructs a unified semantic representation and causal semantic chain. Then, it obtains a fine-grained knowledge embryo through one-hot deconstruction and parameterized mapping. Furthermore, it introduces knowledge fiber bundle constraints to reconstruct the knowledge at the meta-granularity and perform consistency checks, forming industrial meta-knowledge that is uniform in form, traceable in parameters, and indivisible.

[0093] Industrial knowledge element partitioning mechanism: This embodiment learns the meta-knowledge base through knowledge distillation, constructs a unified industrial meta-knowledge module space, and enables different meta-knowledge to be measured and compared in the same space; on this basis, the local manifold characteristics and homotopy continuity of the meta-knowledge distribution are used, combined with mechanistic constraints, to cut the knowledge boundary and form a meta-knowledge cluster with continuous internal mechanism and clear boundary.

[0094] Industrial knowledge meta-reasoning mechanism: This embodiment takes meta-knowledge clusters as input, identifies the core knowledge within the clusters and removes abnormal knowledge, represents the meta-knowledge and its higher-order coupling relationships as a spectral hypergraph, and generates industrial meta-models under the guidance of reasoning engines such as dynamics, conservation relations and engineering constraints; further, it evaluates the combinatorial relationships between meta-models through compatibility potential constraints, and constructs a reasonable, searchable and scalable industrial meta-model knowledge graph.

[0095] Industrial knowledge meta-source tracing mechanism: This embodiment organizes the state, relationship and decision information in the meta-model generation process into a hierarchical generation chain to characterize the meta-model formation and relationship propagation process; on this basis, structural causal analysis is introduced to verify the meta-model and its relationships, and when conflicts or anomalies occur, key states, key relationships and key decision links are located in reverse, so as to realize the reliable judgment and root cause tracing of the meta-model construction results.

[0096] Reference Figure 7 , Figure 7 This is a structural block diagram of the industrial meta-model base architecture provided in the embodiments of this application.

[0097] Secondly, such as Figure 7 As shown, this embodiment also discloses an industrial meta-model base architecture, which applies the autonomous intelligent construction method of the industrial meta-model described above, including: The meta-extraction module is used to perform information hot structure based on semantic chain and meta-granular reconstruction based on knowledge fiber bundle, transforming multi-source heterogeneous industrial information into industrial meta-knowledge with unified form and granularity, thus completing meta-extraction; among which, multi-source heterogeneous industrial information includes text, data, images and expert experience.

[0098] The meta-partitioning module is used to perform the construction of the industrial meta-knowledge module space based on knowledge distillation and the partitioning of industrial meta-knowledge clusters based on homotopy relations, thereby obtaining industrial meta-knowledge clusters and completing the meta-partitioning.

[0099] The meta-reasoning module is used to perform knowledge distribution-driven generative meta-model construction to obtain industrial meta-models, perform potential energy constraint-based meta-model knowledge graph construction, associate the obtained industrial meta-models to obtain industrial meta-model knowledge graphs, and complete meta-reasoning.

[0100] The meta-source tracing module is used to perform hierarchical Markov-driven generative chain construction and structural causal model-driven conflict root cause tracing, determine the root cause of the conflict to complete the meta-source tracing, and based on the completion status of the meta-source tracing, provide feedback and optimize meta-extraction, meta-partitioning and meta-reasoning.

[0101] It should be noted that the industrial meta-model base architecture of this embodiment corresponds to the aforementioned autonomous intelligent construction method of the industrial meta-model. Therefore, the contents not specifically described in the industrial meta-model base architecture of this embodiment, including but not limited to functional definitions, working principles, and technical effects, can all be referred to the description in the aforementioned autonomous intelligent construction method of the industrial meta-model, and will not be repeated here.

[0102] In summary, this embodiment aims to overcome the problems of inconsistent multi-source heterogeneous knowledge representation, unclear knowledge boundary delineation, insufficient modeling and encapsulation capabilities, and lack of reliable verification and traceability mechanisms in existing industrial knowledge organization methods. It provides an autonomous intelligent construction method and foundation architecture for industrial meta-models. Specifically, it achieves standardization and meta-granular expression of industrial knowledge through meta-extraction, forms mechanistically continuous and clearly defined meta-knowledge clusters through meta-division, and further encapsulates these meta-knowledge clusters into well-structured, clearly defined, and functionally identifiable industrial meta-models through meta-reasoning. It also constructs a composable, reasonable, and reusable industrial meta-model knowledge graph. Simultaneously, it performs reliable verification, anomaly identification, and root cause tracing of the construction process and results through meta-traceability, thereby forming a full-process intelligent construction framework for industrial meta-models covering standardized expression, automatic organization, model construction, system expansion, and reliable verification of industrial knowledge.

[0103] Taking the alumina evaporation process as an example, this embodiment illustrates the application of the autonomous intelligent construction method and base architecture for the industrial meta-model.

[0104] In the meta-extraction stage, multi-source industrial information includes process text, runtime sequence data, and process flow diagrams: (29) in, It represents textual information such as process specifications and operation records; It represents time-series data such as temperature, pressure, liquid level, steam flow rate, mother liquor concentration, and heat transfer temperature difference; This represents image information such as evaporation flow diagrams and equipment structure diagrams. Semantic encoding is performed on the three types of information to obtain standardized semantic descriptions: (30) in, Includes textual semantics such as evaporator, mother liquor, live steam, and scaling; Indicates the temporal variation characteristics of operating parameters; The visual semantics of the evaporator, heating chamber, circulating pump, and pipeline connections are represented. Finally, the three types of semantic descriptions are projected onto a unified semantic space through an isomorphic mapping function to obtain a unified knowledge vector for the alumina evaporation process: (31) in, It is used to achieve a unified semantic representation of text, time-series data and image information in equipment objects, operating parameters, process relationships and abnormal states.

[0105] In practical applications, Using this as input, construct an industrial information semantic chain for the alumina evaporation process. and in mechanistic knowledge, empirical knowledge and causal knowledge Guided by this, orthogonal decoupling projection is performed to generate an orthogonal industrial knowledge view. Subsequently, the orthogonal industrial knowledge view is transformed into a one-heat mask, and semantic chain deconstruction is guided. The deconstructed semantic chain is then parameterized and mapped to obtain a fine-grained industrial knowledge roughness of the alumina evaporation process: (32) in, , , and These represent the knowledge object, attribute parameters, relationships, and constraint rules in the evaporation process, respectively.

[0106] The knowledge parameter source space for the alumina evaporation process is: (33) in, , , , , and These represent knowledge parameter sources such as heat transfer temperature difference, steam flow rate, mother liquor concentration, evaporation pressure, liquid level, and scaling state. The heat transfer temperature difference parameter source... For example, by using a large model to generatively reconstruct the coarse-grained knowledge embryo, the scattered parameters, states, mechanistic relationships, and control rules are organized into unified industrial meta-knowledge: (34) in, This represents the set of meta-knowledge instances generated by associating heat transfer temperature difference parameter sources. This indicates the source mapping relationship. This represents the evolutionary relationships between knowledge instances. For example, the reconstructed meta-knowledge might include: (35) Therefore, the role of large-scale model reconstruction is not simply to extract individual entities or relationships, but to organize the mechanistic relationships, operating states and control rules in the alumina evaporation process into industrial meta-knowledge with unified form and stable granularity, using the parameter source as the anchor point.

[0107] In the meta-partitioning stage, following the aforementioned alumina evaporation process, we take the topological clustering of evaporation-related industrial meta-knowledge as an example. Using the meta-extraction method of industrial knowledge fiberization reconstruction, we obtain uniformly formatted industrial meta-knowledge from the processes of alumina multi-effect evaporation, seawater multi-effect distillation, sugar refining multi-effect evaporation, and mechanical vapor recompression evaporation, forming the set of industrial meta-knowledge to be partitioned. (36) in, This indicates the multi-effect evaporation process of alumina. This indicates the multi-effect distillation process of seawater. This indicates the multi-effect evaporation process in sugar production. This indicates a mechanical vapor recompression evaporation process; , , and These represent subsets of industrial meta-knowledge obtained from the different industrial processes mentioned above. This represents the set of all industrial meta-knowledge to be partitioned. Each industrial meta-knowledge, anchored by a knowledge parameter source, contains knowledge instances with a source-tracing relationship to that parameter source and their evolutionary relationships, and can be represented as follows: (37) in, Indicates the source of knowledge parameters. This represents a set of industrial knowledge instances generated by associating parameters from a source. This represents a specific instance of industrial knowledge. This represents the source mapping between knowledge instance domain parameter sources. It represents the mechanistic connections and evolutionary relationships between knowledge instances within meta-knowledge.

[0108] In practical applications, some of the industrial meta-knowledge obtained from various industrial processes can be represented as: (38) The superscript is used to identify the industrial process to which the industrial meta-knowledge belongs, and the subscript is used to identify the core knowledge content described by the industrial meta-knowledge. This indicates phase change heat transfer. This indicates the cascade utilization of latent heat of steam. This indicates mechanical steam compression and reuse.

[0109] Alumina multi-effect evaporation phase change heat transfer element knowledge Knowledge of phase change heat transfer elements in multi-effect distillation of seawater Knowledge of phase change heat transfer elements in sugar production multi-effect evaporation Although they are designed for different materials, each unit includes pressure. Affecting saturation temperature The saturation temperature determines the effective heat transfer temperature difference. Heat transfer temperature difference driven heat exchange Examples of knowledge such as change can be summarized by the following common mechanistic evolutionary relationships: (39) The phase change heat transfer element knowledge of mechanical steam during compression mainly describes the process where the secondary steam after compression returns to the evaporator, releases latent heat, and drives the material to continue evaporating through an effective heat transfer temperature difference. Its mechanistic evolution can be summarized as follows: (40) in, and These represent the pressure and temperature of the compressed secondary steam, respectively. This refers to the secondary steam generated after heat exchange. Although its source of added steam is different from that of other evaporation processes, its heat exchange process after returning to the evaporator still follows the temperature difference-driven and gas-liquid phase change mechanism.

[0110] The steam volume utilization meta-knowledge in multi-effect evaporation of alumina, multi-effect distillation of seawater, and multi-effect evaporation of sugar production all include examples of secondary steam from the previous stage serving as a supplementary heat source for the next stage. Their common evolutionary relationship can be summarized as follows: (41) That is, the first The secondary steam generated in the first stage is directed to the second stage. The latent heat of steam is transferred in stages, driving the subsequent material to continue evaporating, thus realizing the step-by-step utilization of the latent heat of steam.

[0111] Mechanical vapor recompression meta-knowledge This mainly describes the process where secondary steam enters the compressor, where its pressure and temperature are increased through mechanical work, and then it is reused as heating steam. The evolution of this mechanism can be summarized as follows: (42) in, This indicates the work done by the compressor. This indicates the energy provided after the compressed steam returns to the evaporator. Although and Both involve the utilization of secondary steam, but the former relies on the pressure gradient between effects to achieve the gradual transfer of latent heat, while the latter relies on mechanical work to increase the pressure and temperature of the secondary steam. The core energy utilization mechanisms of the two are different.

[0112] Set of industrial meta-knowledge By inputting the aforementioned meta-knowledge similarity measurement method, fiber structure encoding is performed on the parameter nodes, associations, and evolution directions between knowledge instances of each industrial meta-knowledge, and the similarity relationships between different meta-knowledge can be calculated, resulting in: , , and All of them include shared fiber modes such as temperature difference driven heat exchange, and they are relatively similar. , and All three systems contain a topology in which secondary steam is transferred stage by stage along a multi-effect system, thus exhibiting a high degree of similarity. In contrast, Since the compressor performs work, its structural differences from the meta-knowledge of multi-effect steam latent heat utilization are relatively large, therefore we obtain: (43) Subsequently, the similarity measurement results are input into the aforementioned knowledge distillation network to extract representative meta-knowledge bases from all industrial meta-knowledge. In this embodiment, the set of partial meta-knowledge bases obtained by knowledge distillation can be represented as: (44) in, This represents the knowledge base of phase change heat transfer elements, used to characterize common knowledge structures such as temperature difference-driven heat transfer and vapor-liquid phase change. The latent heat gradient utilization meta-knowledge base is used to characterize the knowledge structure of secondary steam being transferred step by step along the multi-effect system; This represents the knowledge base indicating that external work increases energy, used to characterize the knowledge structure that causes secondary steam to increase in pressure and temperature and then reheat due to external work. After mapping each industrial meta-knowledge to the industrial meta-knowledge module space constituted by the above meta-knowledge base, , , and Mainly along They are directionally distributed and therefore adjacent to each other in the modulus space. , and Mainly along Directional distribution; Then mainly along the direction Distribution. As a result, similar industrial meta-knowledge approaches each other in the module space, while industrial meta-knowledge with different core mechanisms separates along different meta-knowledge base directions.

[0113] The resulting industrial meta-knowledge model space is input into the aforementioned manifold projection and path sorting method to obtain several low-dimensional manifolds and their corresponding candidate homotopy paths. Among them, industrial meta-knowledge related to phase change heat transfer forms candidate homotopy paths: (45) This path indicates that although the composition and production objectives of alumina mother liquor, sugar solution, seawater, and MVR evaporation materials are different, the corresponding meta-knowledge is driven by temperature difference, latent heat transfer, and vapor-liquid phase change, and can be interconnected along a continuous and smooth path in a low-dimensional manifold.

[0114] Steam utilization-related industrial meta-knowledge is used to form candidate homotopy paths: (46) Among them, the first three meta-knowledges are all based on the inter-effect pressure gradient and the stepwise transmission of secondary steam, and the connection cost between adjacent meta-knowledges in the path changes relatively gently; when the path extends to At that time, the energy utilization method changed from the inter-effect latent heat cascade transfer to the steam pressurization and reuse driven by mechanical work, and the connection cost showed a significant leap.

[0115] The candidate homotopy paths are further input into the aforementioned regular constant cross-section boundary partitioning method. Based on the change in the knowledge connection cost of adjacent elements in the candidate paths, ranking jump points and candidate cluster outer edge points are extracted, and a mechanistic constant cross-section is fitted accordingly. When the path extends to... The connection cost shows a significant leap, indicating that the core structure of meta-knowledge has shifted from the cascade transfer of latent heat between effects to steam reuse driven by mechanical compression. Further regularization constraints, including process hierarchy, knowledge type, and variable relationships, are then used to form a regularized constant cross section. Since the connection segment between the two passes through the regularized constant cross section, the candidate path is segmented at this location.

[0116] In contrast, candidate paths Although the various industrial meta-knowledge items belong to different industrial processes, they all maintain the core mechanism of temperature difference-driven heat transfer and vapor-liquid phase change. No significant mechanism jumps occur in the paths, and adjacent connecting segments do not cross regular constant cross sections, thus maintaining a continuous homotopy cluster of knowledge. After completing candidate path segmentation and subspace connectivity judgment, the final output is the set of industrial meta-knowledge homotopy clusters: (47) in, Represents a homotopic group for phase transition heat transfer, representing Multi-effect latent heat cascade utilization of homotopy clusters, This indicates a homotopic cluster that uses external work to drive heat recovery.

[0117] When adding external electrical power to the evaporation process of ohmic heating vacuum concentration of fruit juice, the knowledge of evaporation element is needed. At that time, among them, This represents the ohmic heating vacuum concentration process. This meta-knowledge describes the evolutionary relationship between electrical energy input, Joule heating, juice temperature rise, and water evaporation: (48) in, This represents the electrical work input from the outside. This indicates the Joule heating generated inside the juice. Indicates the temperature of the juice. This indicates newly generated water vapor.

[0118] First of all Mapped to the existing meta-knowledge module space and projected onto the corresponding low-dimensional manifold. This is because it relates to the meta-knowledge of mechanical vapor recompression. Both exhibit a continuous structure of "external work input - energy conversion - evaporation for energy supply," and the connection path between them does not cross a canonical constant cross section. Therefore, they are classified into the homotopy group of external work-driven evaporation for energy supply. And insert the existing path according to the principle of minimizing adjacent jumps: (49) In the meta-reasoning stage, following the aforementioned alumina evaporation process, meta-reasoning sequentially completes the elimination of abnormal meta-knowledge, the construction of the industrial meta-model, and the construction of the industrial meta-model knowledge graph.

[0119] In practical applications, the input is the heat transfer and evaporation meta-knowledge cluster obtained from meta-partitioning, which contains 4 meta-knowledge elements. , , , In this embodiment, each piece of industrial knowledge can be represented by an independent formula. Wherein: (50) in: Describe the effects of mother liquor concentration and evaporation pressure on the boiling point of the mother liquor; Describe the effect of fouling thermal resistance on the overall heat transfer coefficient; Describe the energy conversion between heat transfer temperature difference, overall heat transfer coefficient, and evaporation rate; The abnormal metadata to be identified incorrectly subtracts the scale thermal resistance from the total thermal resistance. In the formula, The concentration of the mother liquor. For evaporation pressure, For pressure The solvent saturation temperature below The increase in boiling point due to concentration. The boiling point of the mother liquor; and These are the scale thickness and the scale thermal conductivity, respectively. The reference heat transfer coefficient under scale-free conditions. This is the actual overall heat transfer coefficient; For the temperature of the live steam, For heat exchange area, The latent heat of vaporization of water, This represents the amount of water evaporated per unit time.

[0120] The above meta-knowledge is encoded and mapped to a knowledge module space: (51) Based on the degree of deviation of each meta-knowledge from the cluster core structure and the supporting evidence within the cluster, the corresponding meta-knowledge confidence level is obtained. In this embodiment, , and The continuous mechanism of "mother liquor boiling point - overall heat transfer coefficient - evaporation water volume" can be derived between them; and from The overall heat transfer coefficient increases with increasing scale thickness, which is inconsistent with knowledge of intracranial structures and heat transfer mechanisms. Therefore, its confidence level satisfies: (52) Therefore, The anomalous meta-knowledge was identified and removed, resulting in the heat transfer and evaporation backbone meta-knowledge cluster: (53) Subsequently, the variable dependencies among the three meta-knowledge formulas were analyzed: Boiling point of the mother liquor , quilt Quotation; Generate overall heat transfer coefficient , quilt Quotation; Simultaneously dependent and And generate evaporated water Construct a hierarchy hypergraph based on the above dependencies: (54) Based on the spectral hypergraph, inference is performed using a hypergraph neural network, and constraints are applied by the inference engine to obtain the main body of the heat transfer evaporation model: (55) Further analysis of the computational inference chain in the model to obtain input / output interface information: (56) Based on the attention score of the hypergraph neural network reasoning process, functional semantic sampling is performed on key meta-knowledge to obtain its functional attributes. The evaporation water volume is calculated based on the mother liquor concentration, evaporation pressure, live steam temperature, and scale thickness. This results in the final industrial meta-model for heat transfer evaporation. (57) For the construction of meta-model knowledge graphs, suppose the following four industrial meta-models have been obtained using the same method: (58) in: This is a steam heating element model, with the output interface being... Its function is to calculate the temperature of live steam for heating. The input interface for the heat transfer and evaporation element model constructed above is: The output interface is ; For material concentration meta-models, the input interface includes The output interface includes the mother liquor concentration. Its function is to calculate the concentration of the mother liquor and the change in liquid level; This is a vacuum pressure element model; the input interface includes the amount of water evaporated. The output interface is the evaporation pressure. Its function is to calculate the pressure state inside the evaporator.

[0121] Calculate the compatibility potential between any two metamodels: (59) in, This indicates the degree of mechanistic compatibility after combining meta-models. Indicates the degree of interface matching. Indicates the degree of functional complementarity. , and For weights.

[0122] Based on compatibility potential energy calculations, the following meta-model combination satisfies the potential energy threshold requirement: (60) Meanwhile, the interface matching result is: (61) Therefore, by establishing directed connections between the corresponding meta-models, we obtain the industrial meta-model knowledge graph: (62) in: (63) (64) In the meta-tracing phase, following the aforementioned alumina evaporation process, this paper takes the multi-effect evaporation system in alumina production as an example to conduct a source tracing analysis of anomaly knowledge in the industrial meta-model. This process involves multiple stages, including steam pressure regulation, material concentration, heat transfer, and equipment condition monitoring, and contains a large amount of industrial knowledge from process documents, operational data, and expert experience. This example selects the evaporation heat transfer meta-model as the verification object to illustrate the process of anomaly knowledge discovery and source localization.

[0123] In this example, an industrial meta-model set is first constructed based on the alumina evaporation process: (65) in, Indicates the first The industrial meta-models include an evaporator heat transfer model, a material balance model, a steam control model, and an energy consumption optimization model. Each industrial meta-model contains model structure information, operating status information, control decision information, and historical verification information, which are used to describe the dynamic evolution of the industrial system under different operating conditions.

[0124] During the operation of the industrial metamodel, the model state is affected by control decisions and generates feedback information through actual operating results. Therefore, the continuous change process within the metamodel's lifecycle can be represented as a state-decision-feedback sequence: (66) in, Indicates the first The state of the industrial meta-model at any given time includes equipment operating parameters, process variables, and the internal state of the model. This refers to control decisions that affect the current state, such as steam pressure regulation, flow control, and equipment maintenance strategies. This indicates the verification results after implementing the corresponding decision, including prediction error, degree of process constraint satisfaction, and expert evaluation information.

[0125] Evaporator heat transfer element model For example, its core state variables include steam temperature, mother liquor concentration, heat exchange capacity, and equipment scaling status, which can be represented as: (67) in, Indicates steam temperature. Indicates the concentration of the mother liquor. Indicates the overall heat transfer coefficient. This indicates the thermal resistance due to scaling. The above description of these conditions provides a complete picture of the changes in heat transfer performance of the evaporator at different operating stages.

[0126] Subsequently, the heterogeneous information in the industrial meta-model is uniformly transformed through the state decision encoding module, mapping the continuously changing state information, control behavior, and feedback results to a unified feature space: (68) in, Represents the encoded state-decision joint features. This represents the state-decision encoding function. After encoding, industrial data from different sources can be compared in a unified form, providing a data foundation for subsequent state evolution analysis and anomaly tracing. Furthermore, the state-decision joint encoding sequence from historical operation processes is input into a hierarchical Markov chain model to learn the state transition patterns during alumina evaporation: (69) Here, the probability represents the likelihood of a change in state at the next moment under the combined influence of the current industrial state and control decisions. By learning the state transition patterns during normal operation, a dynamic evolution model that conforms to actual industrial mechanisms can be established, serving as an important basis for subsequent anomaly detection.

[0127] In the actual process of extracting industrial knowledge, if, due to text comprehension bias or model reasoning errors, the statement "increased fouling reduces heat transfer capacity" is incorrectly extracted as "increased fouling enhances heat transfer capacity," then the knowledge relationships within the industrial metamodel will be erroneous. Although this erroneous knowledge has a complete semantic expression, it does not conform to the actual heat transfer process.

[0128] According to the industrial heat transfer mechanism, the overall heat transfer coefficient and thermal resistance of evaporation equipment satisfy the following relationship: (70) in, Indicates the thermal resistance of the equipment foundation. This represents the additional thermal resistance caused by scaling. From this relationship, it can be seen that as the thermal resistance from scaling increases, the overall heat transfer coefficient should decrease; therefore, the aforementioned erroneous knowledge will be identified by the multidimensional trusted verification module.

[0129] The industrial knowledge to be tested can be represented as a knowledge triple: (71) in, , , These represent knowledge entities, entity relationships, and target entities, respectively. For example, erroneous knowledge can be represented as: G1 = (Fouling thermal resistance, enhanced heat transfer, overall heat transfer coefficient) Through the credibility evaluation function: (72) The system comprehensively assesses the rationality of the knowledge structure, its logical consistency, and its alignment with industrial mechanisms. When the evaluation result falls below a set threshold, the knowledge is marked as an anomalous node, and a chain-like causal tracing process is initiated.

[0130] During the source tracing process, based on the state evolution path stored in the hierarchical Markov chain, the state changes before and after the occurrence of the abnormal node are traced in reverse to form an anomaly propagation chain: (73) in, Indicates an abnormal causal propagation chain. This represents the state node corresponding to the location where the anomaly occurred. By analyzing the state changes, decision-making behaviors, and knowledge sources in the anomaly chain, the link that caused the error can be further located.

[0131] Ultimately, this embodiment can output abnormal knowledge nodes, error propagation paths, and root cause analysis results. For example, it can determine that the error originates from a mismatch of mechanistic relationships in the knowledge extraction stage, rather than from errors in model structure design or operating parameter settings. The source tracing results are then fed back to the knowledge extraction and model optimization stages, realizing a complete closed loop of knowledge generation, model construction, reliable verification, anomaly localization, and feedback optimization, thereby improving the reliability of the industrial metamodel in complex industrial scenarios.

[0132] In the embodiments provided in this application, it should be understood that the embodiments described herein can be implemented in hardware, software, firmware, middleware, code, or any suitable combination thereof. For hardware implementation, the processor may be implemented in one or more of the following: application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), processors, controllers, microcontrollers, microprocessors, other electronic units designed to implement the functions described herein, or combinations thereof. For software implementation, some or all of the processes of the embodiments may be performed by a computer program instructing the associated hardware. During implementation, the program may be stored in a computer-readable storage medium or transmitted as one or more instructions or code on a computer-readable storage medium. Computer-readable storage media include computer storage media and communication media, wherein communication media include any medium that facilitates the transmission of a computer program from one place to another. Storage media may be any available medium accessible to a computer. Computer-readable storage media may include, but are not limited to, RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code having the form of instructions or data structures and accessible to a computer.

[0133] Finally, it should be noted that the above description is only a preferred embodiment of this application and is not intended to limit this application. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for autonomous intelligent construction of an industrial meta-model, characterized in that, include: The system performs information hot structure based on semantic chain and meta-granular reconstruction based on knowledge fiber bundle to transform multi-source heterogeneous industrial information into industrial meta-knowledge with unified form and granularity, thus completing meta-extraction; wherein, the multi-source heterogeneous industrial information includes text, data, images and expert experience. Perform knowledge distillation-based construction of industrial meta-knowledge module space and homotopy-based partitioning of industrial meta-knowledge clusters to obtain industrial meta-knowledge clusters and complete meta-partitioning; Perform knowledge distribution-driven generative meta-model construction to obtain industrial meta-models, perform potential energy constraint-based meta-model knowledge graph construction, associate the obtained industrial meta-models to obtain industrial meta-model knowledge graphs, and complete meta-reasoning. The system performs hierarchical Markov-driven generative chain construction and structural causal model-driven conflict root cause tracing to determine the root cause of the conflict and complete meta-tracing. Based on the completion status of meta-tracing, it provides feedback and optimizes meta-extraction, meta-partitioning, and meta-reasoning.

2. The autonomous intelligent construction method for industrial meta-models according to claim 1, characterized in that, The semantic chain-based one-hot information structure includes: Multi-source heterogeneous industrial information is mapped to a unified semantic space to obtain an industrial semantic representation matrix for representing industrial knowledge factors; Based on the causal dependence, process constraints, and interaction relationships of industrial knowledge factors, an industrial information semantic chain is constructed. Based on different prior knowledge, orthogonal projection is performed on the semantic chain of industrial information to generate multiple industrial knowledge views. Among them, the generation of industrial knowledge views meets the preset optimization constraints, which are used to control semantic preservation loss, semantic coupling and consistency distance between industrial knowledge views and corresponding prior knowledge. One-hot masking is performed on the industrial knowledge view to obtain the semantic attribution mask of each industrial knowledge factor, so as to deconstruct the semantic chain of industrial information. The deconstructed industrial information semantic chain is parametrically mapped, and the industrial knowledge factors, attributes, relationships and constraints are uniformly transformed into a parametric industrial knowledge embryo, thus completing the information unique structure based on the semantic chain.

3. The autonomous intelligent construction method for industrial meta-models according to claim 2, characterized in that, The meta-granular reconstruction based on knowledge fiber bundles includes: Driven by a large model, with the local triviality and fibrillation of industrial knowledge as pre-defined constraints, and with the constructed parameterized industrial knowledge embryo as input and industrial meta-knowledge as the reconstruction target, the discrete parameterized industrial knowledge embryo is subjected to generative reconstruction and evolutionary reasoning, and the output is candidate industrial meta-knowledge that meets the pre-defined formal requirements. The candidate industrial meta-knowledge undergoes meta-granularity testing and correction. The testing process includes a mechanism completeness test to determine whether the knowledge can be derived from a single parameter source, an information gain test to determine whether the knowledge is indivisible, and a knowledge independence test to determine whether the knowledge depends solely on its own parameter source without causing cross-source drift. If the test fails, the feedback information is returned to the reconstruction process to adjust the knowledge parameter mapping or combination relationship and regenerate the meta-granularity until industrial meta-knowledge that meets the preset constraints and formal requirements is obtained.

4. The autonomous intelligent construction method for industrial meta-models according to claim 1, characterized in that, The construction of the industrial meta-knowledge module space based on knowledge distillation includes: Each industrial meta-knowledge is represented as a distribution on a shared fiber pattern through dictionary learning; For any two industrial meta-knowledges, calculate the knowledge difference entropy to obtain the similarity; After obtaining similarity, knowledge distillation driven by a large model is used to extract meta-knowledge base from all industrial meta-knowledge. Based on the extracted meta-knowledge base, and combined with the weights corresponding to the preset meta-knowledge base, an industrial meta-knowledge module space is constructed.

5. The autonomous intelligent construction method for industrial meta-models according to claim 4, characterized in that, The industrial meta-knowledge cluster partitioning based on homotopy relations includes: By projecting the industrial meta-knowledge module space into a low-dimensional manifold, the meta-knowledge that was originally distributed in the high-dimensional module space is divided into several locally continuous low-dimensional manifolds. Within a low-dimensional manifold, the connection cost is constructed based on the magnitude of adjacent changes between meta-knowledge, and the knowledge evolution path is sought based on the principle of minimizing the total path increment and the adjacent jump; the set of path nodes that meet the continuity requirement is determined as the candidate homotopy cluster. The mechanism change region is extracted based on the sorting jump point and the cluster outer edge point, and a mechanism constant section is constructed to characterize the location where the continuity of mechanism is broken during the knowledge evolution process; a regular constraint surface is constructed by combining expert experience preferences and domain constraints; a regular constant section for boundary determination is formed based on the mechanism constant section and the regular constraint surface. If the connection segment between adjacent meta-knowledge passes through a regular constant section, then the connection segment is considered to have a boundary, and the candidate homotopy cluster needs to be segmented; if it does not pass through the regular constant section, then the connection segment is considered to maintain a homotopy continuity. Based on all the split points, the candidate homotopy clusters are branched and cut, and the connectivity of the sub-clusters is judged until the continuity requirement is met within each sub-cluster, resulting in a set of knowledge homotopy clusters that are smooth and ordered within the clusters and have clear boundaries between the clusters.

6. The autonomous intelligent construction method for industrial meta-models according to claim 1, characterized in that, The knowledge distribution-driven generative meta-model construction includes: Estimate the backbone center of the industrial meta-knowledge cluster based on the knowledge distribution within the industrial meta-knowledge cluster; For an industrial meta-knowledge, determine the degree of global deviation and local support of the industrial meta-knowledge relative to the backbone center; combine the degree of global deviation and local support to obtain the backbone confidence of the industrial meta-knowledge; based on the backbone confidence, combined with the preset backbone confidence threshold, knowledge is proposed to obtain a meta-knowledge cluster backbone with consistent structure, continuous mechanism and clear boundary. Based on the distribution and mechanistic relationships of various industrial meta-knowledge in the industrial meta-knowledge model space, the phylogenetic dependencies between meta-knowledge are identified, and the backbone of the meta-knowledge cluster is organized into a hypergraph structure. The hypergraph structure is input into the hypergraph reasoning model, and the meta-knowledge nodes and their corresponding hyperedge relationships are jointly encoded and combined inference to generate candidate meta-model structures. Among them, the combined inference is constrained by the dynamic changes of state variables, the conservation of mass, the conservation of energy or the material balance relationship, and the constraints of parameter range, boundary conditions and engineering feasible region. After constraint-guided hypergraph reasoning, an industrial meta-model prototype structure is obtained, which represents the variable relationships, parameter relationships, mechanistic equations, computational logic, and constraint expressions in the industrial system. The prototype structure of the industrial meta-model is then interface-identified to determine the input variables, output variables, parameter terms, and boundary conditions required for model operation, thus forming the industrial meta-model interface. Based on the contribution of key industrial meta-knowledge during the reasoning process, the functional semantics represented by the model are sampled and summarized to form the functional attributes of the industrial meta-model. Finally, the prototype structure of the industrial meta-model, the industrial meta-model interface, and the functional attributes of the industrial meta-model are encapsulated to obtain the industrial meta-model.

7. The autonomous intelligent construction method for industrial meta-models according to claim 6, characterized in that, The construction of the meta-model knowledge graph based on potential energy constraints includes: For any two industrial meta-models to be associated, a compatibility potential function is constructed. This compatibility potential function outputs a compatibility potential energy based on mechanistic consistency judgment, interface closure judgment, and functional synergy judgment. Among them: mechanistic consistency is used to judge whether the variable relationships, constraint relationships, and physical laws of the two industrial meta-models are contradictory; interface closure is used to judge whether the output of one industrial meta-model can meet the input requirements of the other industrial meta-model; functional synergy is used to judge whether the two industrial meta-models have synergistic effects on industrial tasks, process links, or control objectives. Using compatibility potential as a supervision signal, the vector representation space of the industrial meta-model is shaped through comparative learning. After the representation space is shaped, candidate associations are selected based on the vector distance of the industrial meta-model, and interface closure verification is performed to form an industrial meta-model knowledge graph with industrial meta-models as nodes and compatibility combination relationships as edges.

8. The autonomous intelligent construction method for industrial meta-models according to claim 1, characterized in that, The hierarchical Markov-driven generative chain construction includes: Collect state data, decision-making behavior and operational feedback information throughout the entire process of building the industrial meta-model, complete data denoising and standardization, and construct a continuous time series; perform vectorization transformation on the operational state and various decision-making actions of the industrial meta-model respectively, and then fuse the two types of vector features through a joint mapping function to generate a unified state-decision joint code; Using a temporally arranged joint coding sequence as input, a hierarchical Markov chain is constructed by combining the multi-level architecture characteristics of the industrial meta-model, and a preset interference term is introduced to establish the state transition relationship between adjacent time steps. The maximum likelihood estimation method is used to iteratively train the hierarchical Markov chain. The fitting effect of the model with the actual running link is evaluated by the log-likelihood function, and the state transition rules between levels are continuously optimized.

9. The autonomous intelligent construction method for industrial meta-models according to claim 8, characterized in that, The conflict root cause tracing driven by the structural causal model includes: Based on the multi-level state topology and decision evolution link output by hierarchical Markov chain, multi-dimensional cross-validation is carried out in combination with structural causal model to verify whether the relationship between industrial meta-model and knowledge entities conforms to industrial mechanism, process flow and physical constraints. For the abnormal links identified by cross-validation, the degree of abnormality of a single causal chain is quantified by an anomaly metric function to distinguish problematic sub-chains from normal links; Based on the results of anomaly quantification, and combined with counterfactual reasoning techniques, we trace back along the time-series path of the hierarchical Markov chain to locate the abnormal state nodes and erroneous decision-making links layer by layer, sort out the complete error propagation path and the root cause of the problem, and determine the root cause of the conflict.

10. An industrial meta-model base architecture, employing the autonomous intelligent construction method for industrial meta-models as described in any one of claims 1 to 9, characterized in that, include: The meta-extraction module is used to perform information hot structure based on semantic chain and meta-granular reconstruction based on knowledge fiber bundle, transforming multi-source heterogeneous industrial information into industrial meta-knowledge with unified form and granularity, and completing meta-extraction; wherein, the multi-source heterogeneous industrial information includes text, data, images and expert experience. The meta-partitioning module is used to perform the construction of the industrial meta-knowledge module space based on knowledge distillation and the partitioning of industrial meta-knowledge clusters based on homotopy relations, thereby obtaining industrial meta-knowledge clusters and completing the meta-partitioning. The meta-reasoning module is used to perform knowledge distribution-driven generative meta-model construction to obtain industrial meta-models, perform potential energy constraint-based meta-model knowledge graph construction, associate the obtained industrial meta-models to obtain industrial meta-model knowledge graphs, and complete meta-reasoning. The meta-source tracing module is used to perform hierarchical Markov-driven generative chain construction and structural causal model-driven conflict root cause tracing, determine the root cause of the conflict to complete the meta-source tracing, and based on the completion status of the meta-source tracing, provide feedback and optimize meta-extraction, meta-partitioning and meta-reasoning.

Images