Document-level intelligent manufacturing process flow relation extraction method

Abstract

A document-level intelligent manufacturing process flow relation extraction method comprises the following steps: S1, acquiring a process document, and labeling process entities in the process document and a relation between the process entities; s2, performing deep coding on the process document by using a pre-training language model, extracting core semantic information of the process document, and generating process entity representation with consistent semantics based on an entity aggregation strategy of context sensing; s3, nodes and edges of a heterogeneous graph are constructed according to process entity representation, and a relation graph convolutional network R-GCN and a hierarchical attention mechanism are introduced to fully capture the relation between process entities; according to the method, the modeling capability of technological process sequence constraint and the fusion effect of document semantics and graph structure features are respectively enhanced by a sequential construction strategy and a process perception double-graph distillation mechanism oriented to a manufacturing process, the integrity and the utilization rate of information transmission are improved, the performance loss caused by error layer-by-layer propagation is relieved, and the method is suitable for popularization and application. Accurate and efficient extraction of the entity relationship in the process document is realized.

Description

Technical Field

[0001] The present invention relates to the field of natural language processing technology, and specifically to a document-level intelligent manufacturing process flow relationship extraction method. Background Art

[0002] In smart manufacturing plants, process documentation details every step of the production process, including equipment requirements, operating procedures, and related process conditions. Due to the often complex content and processes involved, manual extraction alone is not only inefficient but also difficult to ensure accurate information. With the continuous development of smart manufacturing, the demand for efficient and accurate data extraction is increasing, but traditional methods struggle to meet this requirement, especially when faced with massive amounts of complex data, which further increases the difficulty of information processing. Therefore, developing intelligent process document information extraction systems has become a key technology to support production decision-making and accelerate new plant construction.

[0003] Current research can be divided into two main categories. One is pipeline approaches, which treat entity recognition, mention extraction, and relation extraction as independent tasks, processing them sequentially through their own models. For example, the DWIE model combines four information extraction subtasks, proposes a new entity-driven metric, and leverages graph neural networks to enable cross-document and cross-task information transfer, thereby improving accuracy and efficiency. However, pipeline approaches often suffer from error accumulation due to insufficient information sharing when dealing with complex cross-sentence relationships in process documents, and they also struggle to effectively model the dynamic interactions between tasks. The other category is end-to-end joint learning approaches, which simultaneously extract entities and relations by constructing a document-level graph structure. For example, the SGR model combines graph models and path models to integrate multiple paths between entity pairs into a concise subgraph structure for relational reasoning. The DocRE-Rec model proposes a novel encoder-classifier-reconstructor model for document-level relation extraction. This model effectively reconstructs the true path dependencies through a reconstructor, thereby focusing the training process on entity pairs with relationships. Although this method can capture global context information, most methods still use one-way interactive graph encoders, ignoring the key information contained in non-entity words, and have high computational complexity, making it difficult to efficiently process large-scale process documents.

[0004] To this end, the present invention proposes a graph-structure-based joint relationship extraction method and a document-level intelligent manufacturing process relationship extraction method. Summary of the Invention

[0005] The purpose of the present invention is to propose a document-level intelligent manufacturing process flow relationship extraction method, design a sequential construction strategy for the manufacturing process and a process-aware dual-graph distillation mechanism, which respectively enhance the modeling ability of process flow sequence constraints and the fusion effect of document semantics and graph structure features. This method effectively improves the integrity and utilization efficiency of information transmission, alleviates the performance loss caused by the layer-by-layer propagation of errors, and realizes the accurate and efficient extraction of entity relationships in process documents.

[0006] The technical solution adopted by the present invention is: a document-level intelligent manufacturing process flow relationship extraction method, comprising the following steps:

[0007] S1. Obtain process documents and mark the process entities and the relationships between them in the process documents;

[0008] S2. Use a pre-trained language model to deeply encode the process document, extract the core semantic information of the process document, and generate semantically consistent process entity representations based on a context-aware entity aggregation strategy;

[0009] S3. Construct nodes and edges of heterogeneous graphs based on process entity representations, and introduce the relational graph convolutional network (R-GCN) and hierarchical attention mechanism to fully capture the relationships between process entities.

[0010] S4. Construct a sequence dependency graph according to the natural sequence of each process entity in the process flow document, and embed the timing information to obtain a timing dependency graph;

[0011] Obtain production specifications from relevant industries, divide them into hard process constraints and soft process constraints, and explicitly encode edge weights based on the process rules of the production specifications. The hard process constraints are mapped into binary edge weights in a timing dependency graph, and the soft process constraints are mapped into flexible edge weights in the timing dependency graph, resulting in a process rule-enhanced timing dependency graph.

[0012] The improved temporal graph convolutional network HiWave-TGCN is used to capture the long-term and short-term dependencies of process entities in the process rule enhanced timing dependency graph;

[0013] S5. Use the heterogeneous graph and the process rule enhanced temporal dependency graph as teacher models, and the pre-trained language model as the student model. The two teacher models are integrated to guide the student model.

[0014] S6. Under the joint modeling of process rule enhanced timing dependency graph and temporal graph convolutional network T-GCN, the relationship scores between process entities are calculated based on the dual affine attention mechanism, and the hierarchical agglomerative clustering algorithm HAC is combined to optimize the relationship prediction results.

[0015] As a preferred solution, the hierarchical attention mechanism includes constructing local attention heads for capturing local dependencies between nodes and global attention heads for capturing long-distance dependencies and global information between nodes. The outputs of the local attention heads and the global attention heads are spliced to obtain the final feature representation.

[0016] As the preferred solution, the edge weight E vu Positioning:

[0017]

[0018]

[0019] Among them, s v is the process step v, s u is the process step u, w vu is the flexible edge weight, w vu ∈[0,1], σ(·) represents the Sigmoid activation function, which is used to normalize the edge weight to the [0,1] interval; φ k (v,u) is the characteristic function of the k-th process rule, which describes the structural relationship between the steps in terms of equipment compatibility, parameter compatibility, spatial layout, etc.; α k is the importance weight of the corresponding rule; Attn(v,u) represents the vector h represented by the step v 、h u The calculated attention score is used to model semantic relevance; β is the fusion coefficient, which is used to adjust the influence ratio of structural rules and semantic information.

[0020] As a preferred solution, the improved temporal graph convolutional network HiWave-TGCN is a temporal graph convolutional network T-GCN that introduces continuous wavelet transform CWT under a layered fusion mechanism.

[0021] As a preferred solution, capturing the long-term and short-term dependencies of process entities in the process rule enhanced timing dependency graph using the improved temporal graph convolutional network HiWave-TGCN includes the following steps:

[0022] For a given node v, its feature sequence After wavelet transform, the frequency domain enhanced representation is obtained:

[0023]

[0024] in, is the feature of node v after wavelet transform enhancement at time t; x v (τ) is the original time series signal of node v at continuous time τ; τ is the time point on the continuous time axis; ψ is the mother wavelet function, s is the scale parameter; CWT is continuous wavelet transform;

[0025] In the hierarchical fusion mechanism, the static attribute S of the process entity is v With dynamic parameters Introduced into the representation of nodes at different levels respectively:

[0026]

[0027] in, is the initial representation vector of node v; W is the first layer state of the node output by GRU; s is the weight matrix for static feature transformation; S v is the static attribute of node v; b s is the bias vector; ReLU(·) is the nonlinear activation function; GRU(·) is the gated recurrent unit;

[0028] The temporal graph convolutional network (T-GCN) combines structural adjacency information for updates:

[0029]

[0030] in, is the hidden state representation of node v in the first layer of the temporal graph convolutional network T-GCN, A vu is the adjacency matrix; W (l) is the weight matrix of the lth layer; is the representation of the neighbor node u in the l-1th layer; N(v) is the neighbor set of node v; b (l) is the bias term; σ is the nonlinear activation function;

[0031] Finally, the information of each layer is fused to form a multi-source time series representation of the node:

[0032]

[0033] in, is the final temporal representation of node v; MLP(·) is (multi-layer perceptron); is the output of the last layer of graph convolution.

[0034] As a preferred solution, in step S5, two teacher models are fused through gated attention to generate a global knowledge representation, and the student model is optimized using contrastive learning, adversarial training and rule-constrained loss.

[0035] As a preferred solution, in step S5, the KL divergence between the node representation distributions of the teacher model and the student model is minimized to guide the student model to learn more discriminative feature representations:

[0036]

[0037] Among them, L feat is the characteristic distillation loss, is the representation of node v in the teacher model, is the corresponding node representation in the student model;

[0038] Construct a discriminator D to compare the structures generated by the teacher and student models and enforce consistency. The loss function is as follows:

[0039]

[0040] in, is the expectation of all possible teacher graph structures; D(·) is the adversarial network that discriminates the distribution of graph structures; A teacher A graph structure representation constructed for the teacher model; student A graph structure representation constructed for the student model; is the expectation of all possible student graph structures;

[0041] In order to ensure that the predicted relationship conforms to the prior rules, a rule compliance loss is introduced. The loss function is adjusted based on the predicted relationship frequency:

[0042]

[0043] Among them, L rule Losses due to rule compliance; is the predicted distribution.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] (1) A sequential construction strategy for manufacturing processes is designed, and a process rule-enhanced timing dependency graph is proposed. Traditional modeling methods mostly rely on text order or contextual information to construct graph structures, but these methods are difficult to accurately reflect the explicit constraints in production specifications and the flexibility of processes. To this end, the present invention maps the hard constraints in the production specifications into edge weights in the graph, and structurally cuts off illegal paths; for adjustable process steps, learnable flexible edge weight parameters are introduced to support multi-path and parallel modeling. In addition, periodic time coding is integrated to construct a graph structure with time sensitivity and irreversible dependencies. On this basis, combined with the improved temporal graph convolutional network HiWave-TGCN, continuous wavelet transform is introduced to enhance the long-term dependency modeling capability, and a hierarchical fusion mechanism is designed to integrate static structure and dynamic parameter information to achieve multi-source timing modeling of complex process flows.

[0046] (2) A process-aware dual-graph distillation mechanism was designed to alleviate the one-way semantic transfer problem between document encoding and graph encoding in traditional document-level relationship extraction methods. This mechanism uses a heterogeneous graph encoder and a temporal dependency graph encoder as teacher models, and combines the gated attention mechanism to fuse the dual-graph representation to construct a unified global knowledge graph, thereby providing multi-level supervision signals for the student model. At the feature level, contrastive learning is used to align entity embeddings to ensure that the student model obtains semantic representations consistent with the teacher model; at the structural level, adversarial training is used to optimize the relationship prediction matrix of the student model to make it close to the graph adjacency structure of the teacher model; at the rule level, a process constraint verification loss function is introduced to enhance the student model's ability to comply with production specifications. This mechanism provides technical support for the structured representation and intelligent management of process knowledge.

[0047] (3) Compared with previous studies that mainly focused on static relationships such as “location-affiliation” and “city-country” based on general datasets such as DocRED, this paper focuses on the intelligent manufacturing scenario and the relationship between “step 1→step 2” in process flow documents.

[0048] →…” chain dependencies and complex industrial semantics such as “equipment-parameter constraints”, a scenario-adaptive relationship extraction method is proposed. This method proposes a sequential construction strategy for manufacturing processes, constructs a process-rule-enhanced temporal dependency graph, and improves the temporal graph convolutional network to effectively model sequential constraints and multi-level dependencies in the process. It also designs a process-aware dual-graph distillation mechanism that integrates multi-source structural information from heterogeneous graphs and temporal graphs, guiding student models to learn richer graph semantic representations. The overall solution customizes modeling based on the complexity and hierarchical dependency characteristics of process document structures, enhancing the model's understanding of operational process logic and constraint relationships, and improving its adaptability and practicality in industrial document processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0050] Figure 1 It is a schematic diagram of the framework flow of the present invention. DETAILED DESCRIPTION

[0051] The present invention is described in detail below by way of exemplary embodiments. However, it should be understood that elements, structures, and features in one embodiment may also be beneficially combined in other embodiments without further description.

[0052] It should be noted that: unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning understood by persons having ordinary skills in the field to which the invention belongs. The words "one", "an" or "the" and the like used in the patent application specification and claims of the present invention do not express a quantity limitation, but indicate the existence of at least one; the words "first", "second" and "third" used herein shall not be regarded as a limitation on the order of components, but are only used to distinguish different components; the words "include" or "comprise" and the like indicate that the elements or objects appearing before "include" or "comprise" include the elements or objects listed after "include" or "comprise" and their equivalents, but do not exclude other elements or objects with the same function.

[0053] In order to more clearly describe the document-level intelligent manufacturing process relationship extraction method, combined with the attached Figure 1 Describe this embodiment:

[0054] like Figure 1 As shown, a document-level intelligent manufacturing process flow relationship extraction method includes the following steps:

[0055] S1. Obtain process documents and mark the process entities and the relationships between them in the process documents;

[0056] S101. Collect process-related text data from multiple sources, including industrial documents, technical reports, production manuals, and process documentaries. To ensure data diversity and coverage, the collected content includes key information such as product name, process steps, required raw materials, equipment used, and process parameters (such as temperature and time).

[0057] S102. Use the doccano annotation tool to annotate these texts. The annotation contents include:

[0058] Process entity: product name, process steps, equipment name, raw materials, equipment specifications, temperature, duration, etc.

[0059] Relationships between entities: sequential dependencies of process steps, required equipment, required raw materials, equipment processing time, temperature requirements, raw material consumption, etc.

[0060] S103. After labeling is completed, entities and their relationship information are extracted through regular matching and parsing of JSON data, and the data structure is reorganized to ensure its integrity and consistency so that it meets the input requirements of the model.

[0061] S104. Divide the converted dataset into a training set, a validation set, and a test set in a ratio of 8:1:1 to ensure data balance and high-quality input.

[0062] Suppose there are n process document samples D={D1,D2,…,D n}, each process document D i Contains multiple process entities and the relationships between them. Each entity E ij It is document D i The jth process entity in the process, and each entity E ij Can be classified into a certain process entity type, such as equipment, operation, raw materials, process parameters, etc. The relationship between entities R ij Describes the semantic relationship between two process entities, such as the association between equipment and operations, the sequence dependency of process steps, process constraints, etc. For each relationship R ij , we can use the triple (E ij1 ,R ij ,E ij2 ) to express it, where E ij1 and E ij2 is a process entity, R ij is the relationship between them. The goal of this method is to extract all process entities and their corresponding relationships from process documents.

[0063] Triples={(E ij1 ,R ij ,E ij2 )||E ij1 ,E ij2 ∈Entities,R ij ∈Relations}

[0064] S2. Use a pre-trained language model to deeply encode the process document, extract the core semantic information of the process document, and generate semantically consistent process entity representations based on a context-aware entity aggregation strategy;

[0065] S201, use the built-in word segmenter of the RoBERTa pre-trained language model to segment the process document and divide the text into subword sequences. In order to capture the global semantics, insert <s> Special marker that indicates the beginning of a document. Add between each entity or sentence< / s> Markers are used to clearly define the structural boundaries of the text.

[0066] S202: The segmented subword sequence is fed into the multi-layer encoder of the RoBERTa pre-trained language model to generate a context-aware vector representation for each text unit (token). This process not only provides lexical information for each token but also incorporates its context within the sentence into the vector representation, enabling more accurate semantic understanding.

[0067] S203. Use sequence tagging technology to process the vectors generated by RoBERTa, identify key entities in the document (such as product names, process steps, equipment names, etc.), and accurately determine their boundaries. For entities composed of multiple text units (tokens), a weighted pooling method is used to integrate the vectors of each token into a fixed-length entity vector as a comprehensive semantic representation of the entity. Specifically, suppose the entity consists of n tokens, namely x1, x2, ..., x n , their vectors are represented by z1,z2,…,z n , the weighted coefficients are w1,w2,…,w n The comprehensive vector of the entity is calculated by weighted pooling:

[0068]

[0069] S3. Construct nodes and edges of heterogeneous graphs based on process entity representations, and introduce a relational graph convolutional network (R-GCN) and a hierarchical attention mechanism to fully capture the relationships between process entities.

[0070] S301. Construct nodes and edges in the process flow diagram based on the entities and mention information extracted in the second stage. In this diagram, the types of nodes include mention nodes, entity nodes, and document nodes. Specifically, the mention node represents the mention of each process step in the process document; the entity node represents the specific entity in the process flow, such as process steps, equipment, materials, etc.; the document node represents the entire process document and is used to connect all mention nodes and entity nodes. In order to effectively model the relationship between entities, three types of edges are introduced in the diagram: entity aggregation edges, local co-occurrence edges, and global structure edges. Entity aggregation edges connect mention nodes and entity nodes in the same sentence, local co-occurrence edges connect different mention nodes and entity nodes of the same entity, and global structure edges connect all mention nodes and entity nodes to document nodes.

[0071] S302, the main task of the graph convolution layer during the processing process is to update the feature representation of the node and fully capture the semantic information and structural relationships of neighboring nodes. For heterogeneous graphs, different types of edges carry different semantic information, so the graph convolutional network (R-GCN) assigns different weight matrices to each edge type. In each layer of graph convolution, for the kth edge type, the update formula is:

[0072]

[0073] in, is the adjacency matrix of the k-th edge type, and are the weight matrix and bias term of the edge type, N k(v) represents the set of neighbor nodes connected to node v through edges of type k, and σ(·) is a nonlinear activation function.

[0074] Through the propagation of multiple layers of graph convolution, each layer gradually integrates the information of neighboring nodes, enhancing the contextual semantics of the node. Ultimately, through hierarchical propagation, the relational graph convolutional network can introduce multi-dimensional information from different types of edges, making the node representation richer.

[0075] S303. In order to further improve the performance of graph convolution, a hierarchical attention mechanism is introduced to dynamically adjust the importance of each node in the graph convolution by calculating the attention score between nodes. Specifically, the attention score α between node v and node u is calculated. vu , usually using the following formula:

[0076]

[0077] Among them, Q V and K u are the query and key vectors of node v and node u, respectively, and d is the feature dimension. The model uses the attention score to weight the sum of neighbor features to update the node representation. The feature update formula for node v at layer l is:

[0078]

[0079] Where N(v) is the set of neighbors of node v.

[0080] The attention mechanism consists of three local attention heads and one global attention head. Specifically, the local attention heads are used to capture the local dependencies between nodes. Each local head has an independent query, key, and value matrix. The calculation process can be uniformly expressed as:

[0081]

[0082] Among them, F (l) represents the node features of the lth layer, and are the query and key matrices of the i-th attention head, A i is the attention weight matrix of the head, is the value matrix, d′ is the dimension of the attention head, and E i is the adjacency matrix (if there is an edge between node u and node v, then E uv =1, otherwise 0), (1-E i )·(―∞) is used to ignore unconnected nodes; is the output vector obtained by the i-th attention head of node v in layer l. The global attention head is used to capture long-distance dependencies and global information between nodes. Its calculation formula is:

[0083]

[0084] in, and are the query, key, and value matrices of the global attention head, respectively.

[0085] Finally, the outputs of the three local attention heads and one global attention head are concatenated, and the final feature representation of the node in the lth layer is obtained through linear transformation. The formula is:

[0086]

[0087] in, is the output linear transformation matrix, is the bias term. In this way, the model can effectively fuse the results of different attention heads, capture the relationship between nodes from multiple perspectives, and achieve a more comprehensive and in-depth node representation.

[0088] S4. Construct a sequence dependency graph according to the natural sequence of each process entity in the process flow document, and embed the timing information to obtain a timing dependency graph;

[0089] Obtain production specifications from relevant industries and divide them into hard process constraints and soft process constraints. Explicitly encode edge weights based on the process rules of the production specifications: map the hard process constraints to binary edge weights in the timing dependency graph, and map the soft process constraints to flexible edge weights in the timing dependency graph, resulting in a process rule-enhanced timing dependency graph.

[0090] The improved temporal graph convolutional network HiWave-TGCN is used to capture the long-term and short-term dependencies of process entities in the process rule enhanced timing dependency graph; among them, the improved temporal graph convolutional network HiWave-TGCN is a temporal graph convolutional network T-GCN that introduces continuous wavelet transform CWT under a layered fusion mechanism.

[0091] The specific steps include:

[0092] S401, first construct a sequence dependency graph based on the natural sequence of steps in the process flow document. To construct a process dependency graph with temporal semantics, first construct a sequence dependency graph based on the natural sequence of steps in the process flow document. Each node in the graph represents a process step. If step s v Clearly precede step s u , then a directed edge is established between nodes v and u to indicate their sequential dependency. On this basis, time information is introduced to enhance the model's ability to model temporal features. Suppose that each step s i The corresponding timestamp is t i, periodic time embedding is used to map time information into high-dimensional space. The specific encoding method is as follows:

[0093]

[0094] Among them, PE(t i ) is step s i The corresponding timestamp t i A high-dimensional vector representation of a periodic time embedding; d is the embedding dimension, and i is the dimension index. This time embedding results in additional features for the node, improving the model's ability to model the time series structure between process steps.

[0095] S402, based on the obtained time code, further construct a process rule enhanced timing dependency graph. In the structure of the sequence dependency graph, the graph integrates the strong constraints on the execution order of process steps in the production specification and the requirements of the operating procedures; specifically, using the edge weight E vu Explicitly coded process rules: For pairs of steps with a strict sequence (s v ,s u ), set the binary edge weight E vu = 1, structurally cutting off illegal paths; for step pairs with sequential flexibility or parallel possibility, a learnable flexible edge weight w is introduced vu ∈[0,1], used to model its dependency strength and path selection probability:

[0096]

[0097] Where σ(·) represents the Sigmoid activation function, which is used to normalize the edge weight to the interval [0,1]; φ k (v,u) is the characteristic function of the k-th process rule, which describes the structural relationship between the steps in terms of equipment compatibility, parameter compatibility, spatial layout, etc.; α k is the importance weight of the corresponding rule; Attn(v,u) represents the vector h represented by the step v 、h u The calculated attention score is used to model semantic relevance; β is the fusion coefficient, which adjusts the influence ratio of structural rules and semantic information. vu The definition is as follows:

[0098]

[0099] This enhanced graph structure not only retains timing information, but also embeds hard and flexible constraints in process rules, supports dynamic modeling of multi-path and parallel relationships in complex process flows, and improves the model's structural perception of real production processes.

[0100] S403. In order to further enhance the model’s ability to model the temporal dependencies between process steps, this paper proposes an improved temporal graph convolutional network HiWave-TGCN. This model introduces continuous wavelet transform (CWT) based on T-GCN to improve the multi-scale perception and modeling capabilities of long-term dependency features in the process flow. For a given node v, its feature sequence After wavelet transform, the frequency domain enhanced representation is obtained:

[0101]

[0102] in, is the feature of node v after wavelet transform enhancement at time t; x v (τ) is the original time series signal of node v in continuous time τ; τ is the time point on the continuous time axis; ψ is the mother wavelet function, s is the scale parameter; CWT extracts multi-scale time-frequency features to enhance long-term dependency modeling. In terms of structural fusion, HiWave-TGCN designs a hierarchical fusion mechanism to combine static attributes S v (such as equipment model, process rules) and dynamic parameters (such as real-time operational data and environmental factors) are introduced into different levels of representation:

[0103]

[0104] in, is the initial representation vector of node v; W is the first layer state of the node output by GRU; s is the weight matrix for static feature transformation; S v is the static attribute of node v; b s is the bias vector; ReLU(·) is the nonlinear activation function; GRU(·) is the gated recurrent unit;

[0105] The temporal graph convolutional network (T-GCN) is updated with structural adjacency information:

[0106]

[0107] in, is the hidden state representation of node v in the first layer of the temporal graph convolutional network T-GCN, A vu is the adjacency matrix; W (l) is the weight matrix of the lth layer; is (the representation of neighbor node u in the l-1 layer); N(v) is the neighbor set of node v; b (l) is the bias term; σ is the nonlinear activation function;

[0108] Finally, the information at each layer is fused to form a multi-source time series representation of the node:

[0109]

[0110] in, is the final temporal representation of node v; MLP(·) is (multi-layer perceptron); is the output of the last layer of graph convolution.

[0111] HiWave-TGCN can effectively combine graph structure, temporal dependencies, multi-scale dynamic changes and static prior knowledge to achieve more comprehensive and fine-grained step relationship modeling and improve reasoning capabilities in complex process flow scenarios.

[0112] S5. Use the heterogeneous graph and the process rule enhanced timing dependency graph as teacher models, and the pre-trained language model as the student model. The two teacher models are integrated to guide the student model.

[0113] This paper proposes a joint training framework that integrates graph structure modeling and process-aware dual-graph distillation. This framework enhances the model's structural awareness through negative sampling and node reconstruction tasks. Combined with weighted cross-entropy and an adaptive threshold mechanism, it effectively mitigates learning biases introduced by long-tail relationship distributions. Furthermore, by introducing multiple loss terms, such as feature distillation, structural alignment, and rule constraints, the student model is guided to learn more discriminative relationship representations, thereby improving its ability to understand and extract implicit relationships and structural dependencies in process documents.

[0114] The specific steps include:

[0115] S501. In each training batch, for the positive sample triple (v, u, r), type-aware negative sampling and relation-confused negative sampling are first used to construct negative examples. Type-aware negative sampling replaces entities while maintaining semantic type consistency, while relation-confused negative sampling randomly replaces fixed entity pairs with semantically similar but logically conflicting relation labels. During the training process, a node reconstruction task is introduced to represent h by randomly masking some nodes. v , reconstruct the representation using graph neural network to improve the structural modeling ability of the model. Reconstruction loss L recon The calculation method is as follows:

[0116]

[0117] in, Reconstruct the representation for prediction, A collection of occluded nodes.

[0118] S502, the entity pair is represented as [h v ;h u ] Obtain the predicted distribution through bijective mapping And use weighted cross entropy to calculate the main task loss L rel :

[0119]

[0120] Among them, α r is the weight parameter of each type of relationship. In order to enhance the model's ability to identify low-frequency relationships, an adaptive threshold mechanism is introduced. An independent threshold τ is defined for each relationship category. r , based on the average predicted probability of the validation set Dynamic adjustment to obtain the threshold optimization loss L adapt :

[0121]

[0122] in, Represents a long-tail relationship set.

[0123] S503 uses process-aware dual-graph distillation to align the structural representations of the teacher and student models. The inputs are shared, obtained through the heterogeneous graph encoder and the temporal graph encoder, respectively. During training, three types of losses are introduced. The feature distillation loss guides the student model to learn more discriminative feature representations by minimizing the KL divergence between the node representation distributions of the teacher and student models:

[0124]

[0125] Among them, L feat is the characteristic distillation loss, is the representation of node v in the teacher model, is the corresponding node representation in the student model;

[0126] The structural consistency loss uses the discriminator D to compare the structures generated by the teacher and student models to force them to maintain consistency. The loss function is as follows:

[0127] L struct =E Ateacher [logD(A teacher )]+EA student [log(1―D(A student ))]

[0128] Among them, E Ateacher is the expectation of all possible teacher graph structures; D(·) is the adversarial network that discriminates the distribution of graph structures; A teacher A graph structure representation constructed for the teacher model; student Graph structure representation constructed for the student model; EA student is the expectation of all possible student graph structures;

[0129] In order to ensure that the predicted relationship conforms to the prior rules, a rule compliance loss is introduced. This loss function is adjusted based on the predicted relationship frequency:

[0130]

[0131] The total loss function combines the above-mentioned losses, and the final total loss function L is:

[0132] L=λ1L recon +λ2L adapt +λ3L rel +λ4(L feat +L struct +L rule )

[0133] Among them, L rule is the rule compliance loss, and λ1, λ2, λ3, and λ4 are the weight coefficients for each loss. During training, the loss weight coefficients are dynamically adjusted. Model performance is evaluated on the validation set at regular intervals, and the best model parameters based on the performance on the validation set are saved.

[0134] S6. Under the joint modeling of process rule enhanced temporal dependency graph and temporal graph convolutional network (T-GCN), the relationship scores between process entities are calculated based on the dual affine attention mechanism, and the hierarchical agglomerative clustering algorithm (HAC) is combined to optimize the relationship prediction results.

[0135] The specific steps include:

[0136] S601. Under the joint modeling of process rule enhanced timing dependency graph and HiWave-TGCN, the node representation has fully integrated the time series characteristics and multi-dimensional structure information. Let h v represents the feature vector of node v, h u Represents the feature vector of node u, and the goal is to calculate the relationship score S(v,u) between them. In order to calculate the possible relationship between two nodes, a dual affine attention mechanism is used to calculate the relationship score S(v,u) of the node pair:

[0137]

[0138] Where W1 is the biaffine weight matrix used to learn the interaction information between nodes, W2 is the affine transformation weight matrix used to model the independent influence of nodes, and b is the bias term used to adjust the baseline value of the score. The final calculated relationship score S(v,u) represents the possible relationship strength between nodes v and u. A higher score indicates that a relationship is more likely to exist between them.

[0139] S602, HAC uses a bottom-up approach to clustering. First, the similarity between all relationship pairs is calculated, and the cosine similarity is used to calculate the relationship pairs. where r i and r j After calculating the similarity of all relationship pairs, HAC finds the most similar relationship pair (r i ,r j ), which are then merged into a new relation class r new , which is updated as This process is repeated continuously, that is, continuously searching for the most similar relationship pairs to merge and update their representations until the set number of categories is reached. Algorithm 1 is shown in the table.

[0140]

[0141] Parts not described in detail in the above embodiments are prior art.

[0142] It should be noted that although the present invention has been described with reference to the above embodiments, the present invention may also have other various embodiments. Without departing from the spirit and scope of the present invention, it is obvious that those skilled in the art may make various corresponding changes and modifications to the present invention, and such changes and modifications shall fall within the scope of protection of the appended claims and their equivalents.

Claims

1. A document-level intelligent manufacturing process flow relationship extraction method, characterized by: The following steps are involved: S1. Obtain process documents and mark the process entities and the relationships between them in the process documents; S2. Use a pre-trained language model to deeply encode the process document, extract the core semantic information of the process document, and generate semantically consistent process entity representations based on a context-aware entity aggregation strategy; S3. Construct nodes and edges of heterogeneous graphs based on process entity representations, and introduce the relational graph convolutional network (R-GCN) and hierarchical attention mechanism to fully capture the relationships between process entities. S4. Construct a sequence dependency graph according to the natural sequence of each process entity in the process flow document, and embed the timing information to obtain a timing dependency graph; Obtain production specifications from relevant industries, divide them into hard process constraints and soft process constraints, and explicitly encode edge weights based on the process rules of the production specifications. The hard process constraints are mapped into binary edge weights in a timing dependency graph, and the soft process constraints are mapped into flexible edge weights in the timing dependency graph, resulting in a process rule-enhanced timing dependency graph. The improved temporal graph convolutional network HiWave-TGCN is used to capture the long-term and short-term dependencies of process entities in the process rule enhanced timing dependency graph; S5. Use the heterogeneous graph and the process rule enhanced temporal dependency graph as teacher models, and the pre-trained language model as the student model. The two teacher models are integrated to guide the student model. S6. Under the joint modeling of process rule enhanced timing dependency graph and temporal graph convolutional network T-GCN, the relationship scores between process entities are calculated based on the dual affine attention mechanism, and the hierarchical agglomerative clustering algorithm HAC is combined to optimize the relationship prediction results.

2. A document-level intelligent manufacturing process flow relationship extraction method according to claim 1, characterized in that: The hierarchical attention mechanism includes constructing local attention heads for capturing local dependencies between nodes and global attention heads for capturing long-distance dependencies and global information between nodes. The outputs of the local attention heads and the global attention heads are spliced to obtain the final feature representation.

3. The method for extracting document-level intelligent manufacturing process relationships according to claim 1, characterized in that: Edge weight E vu Positioning: Among them, s v is the process step v, s u is the process step u, w vu is the flexible edge weight, w vu ∈[0,1], σ(·) represents the Sigmoid activation function, which is used to normalize the edge weight to the [0,1] interval; φ k (v,u) is the characteristic function of the k-th process rule, which describes the structural relationship between the step pairs in terms of equipment compatibility, parameter compatibility, and spatial layout dimensions; α k is the importance weight of the corresponding rule; Attn(v,u) represents the vector h represented by the step v 、h u The calculated attention score is used to model semantic relevance; β is the fusion coefficient, which is used to adjust the influence ratio of structural rules and semantic information.

4. The method for extracting document-level intelligent manufacturing process relationships according to claim 1, characterized in that: The improved temporal graph convolutional network HiWave-TGCN is a temporal graph convolutional network T-GCN that introduces continuous wavelet transform CWT under a layered fusion mechanism.

5. The method for extracting document-level intelligent manufacturing process relationships according to claim 4 is characterized by: Capturing the long-term and short-term dependencies of process entities in the process rule-enhanced timing dependency graph using the improved temporal graph convolutional network HiWave-TGCN includes the following steps: For a given node v, its feature sequence After wavelet transform, the frequency domain enhanced representation is obtained: in, is the feature of node v after wavelet transform enhancement at time t; x v (τ) is the original time series signal of node v at continuous time τ; τ is the time point on the continuous time axis; ψ is the mother wavelet function, s is the scale parameter; CWT is continuous wavelet transform; In the hierarchical fusion mechanism, the static attribute S of the process entity is v With dynamic parameters Introduced into the representation of nodes at different levels respectively: in, is the initial representation vector of node v; W is the first layer state of the node output by GRU; s is the weight matrix for static feature transformation; S v is the static attribute of node v; b s is the bias vector; ReLU(·) is the nonlinear activation function; GRU(·) is the gated recurrent unit; The temporal graph convolutional network T-GCN combines structural adjacency information for update: in, is the hidden state representation of node v in the first layer of the temporal graph convolutional network T-GCN, A vu is the adjacency matrix; W (l) is the weight matrix of the lth layer; is the representation of the neighbor node u in the l-1 layer; N(v) is the neighbor set of node v; b (l) is the bias term; σ is the nonlinear activation function; Finally, the information at each layer is fused to form a multi-source time series representation of the node: in, is the final temporal representation of node v; MLP(·) is a multi-layer perceptron; is the output of the last layer of graph convolution.

6. The document-level intelligent manufacturing process flow relationship extraction method according to claim 1 is characterized by: In step S5, the two teacher models are fused through gated attention to generate a global knowledge representation, and the student model is optimized using contrastive learning, adversarial training and rule-constrained loss.

7. The method for extracting document-level intelligent manufacturing process relationships according to claim 6, characterized in that: In step S5, the KL divergence between the node representation distributions of the teacher model and the student model is minimized to guide the student model to learn more discriminative feature representations: Among them, L feat is the characteristic distillation loss, is the representation of node v in the teacher model, is the corresponding node representation in the student model; Construct a discriminator D to compare the structures generated by the teacher and student models and enforce consistency. The loss function is as follows: in, is the expectation of all possible teacher graph structures; D(·) is the adversarial network that discriminates the distribution of graph structures; A teacher A graph structure representation constructed for the teacher model; student A graph structure representation constructed for the student model; is the expectation of all possible student graph structures; A rule compliance loss function is introduced to ensure that the predicted relationship conforms to the prior rules. The loss function is adjusted based on the predicted relationship frequency: Among them, L rule Losses due to rule compliance; is the predicted distribution.

Images