Method for constructing heavy die forging press fault knowledge graph

Abstract

The present application belongs to the technical field of large-scale die forging press, and particularly relates to a heavy die forging press fault knowledge graph construction method, comprising: S1, obtaining multi-source heterogeneous text data and processing to obtain preprocessed text data; S2, constructing a fault knowledge graph annotation dataset; S3, training an enhanced cascade binary tagging model; S4, using the enhanced cascade binary tagging model to perform knowledge extraction on the preprocessed text data of S1 to obtain fault knowledge triples; S5, based on a pre-set mode layer, performing knowledge fusion on the fault knowledge triples of S4 to obtain fused fault knowledge; S6, using a graph neural network to perform missing association link prediction on the fused fault knowledge of S5 to complete the missing association, and storing the completed fault knowledge into a graph database. The method can realize high-precision structured analysis of multi-source heterogeneous fault data of a heavy die forging press and dynamic complete construction of a knowledge graph.

Description

Technical Field

[0001] This invention belongs to the field of large-scale die forging press technology, and in particular relates to a method for constructing a fault knowledge graph of a heavy-duty die forging press. Background Technology

[0002] Heavy-duty forging presses are core foundational equipment in the heavy equipment manufacturing field, mainly composed of key subsystems such as the main frame, main hydraulic cylinder, and servo control system. This equipment operates under extreme conditions of high load, high temperature, and high impact for extended periods, resulting in faults exhibiting significant characteristics such as strong coupling of multiple parameters, complex causal chains, and concealed evolutionary patterns. As the service life of the equipment increases, ensuring its safe and stable operation becomes increasingly crucial.

[0003] However, existing technologies still face many serious challenges in fault knowledge management and intelligent operation and maintenance of heavy-duty forging presses, as detailed below:

[0004] First, processing and utilizing equipment fault data for knowledge is extremely challenging. On the one hand, heavy-duty forging presses accumulate massive amounts of multi-source heterogeneous text data, including alarm records, maintenance logs, and technical documents, during long-term operation. This data generally suffers from problems such as non-standard semantic descriptions, mixed technical terms, and high temporal noise. On the other hand, due to the multi-parameter coupling characteristics of faults, a single data source cannot reflect the full picture of the fault. Existing natural language processing technologies, when faced with such highly specialized and unstructured industrial texts, often lack targeted domain dictionaries and terminology discovery mechanisms, resulting in low word segmentation accuracy and blurred entity recognition boundaries, making it difficult to accurately extract structured fault knowledge triples from noisy data. Furthermore, general-purpose models are prone to overfitting or missed detections when processing such imbalanced and semantically complex industrial data, making it difficult to meet the accuracy requirements of knowledge extraction for engineering applications.

[0005] Secondly, existing knowledge graph construction methods are ill-suited to the dynamic completion and fusion requirements of fault knowledge for large-scale forging presses. After extracting knowledge from multi-source heterogeneous data, inconsistencies in entity referencing and relationship conflicts often arise. Existing knowledge fusion methods lack schema-level constraints for four specific knowledge domains: equipment functional structure, condition monitoring, faults, and maintenance, resulting in poor consistency of the fused knowledge base. More critically, due to limitations in data collection completeness and the sparsity of fault occurrences, the initial knowledge graphs constructed generally suffer from missing associations. Traditional link prediction methods fail to fully exploit high-order semantic information within the graph structure, hindering the effective completion of potential fault-cause-maintenance associations and limiting the application value of knowledge graphs in practical fault diagnosis and decision support. Furthermore, traditional manual inspections and experience-based troubleshooting methods are inefficient and cannot meet the stringent requirements of high reliability and rapid response for such large-scale equipment.

[0006] Therefore, how to achieve high-precision structured analysis of multi-source heterogeneous fault data of heavy-duty forging presses and dynamic and complete construction of knowledge graphs, so as to effectively improve the intelligent diagnosis capability of equipment faults and the reliability of operation and maintenance decisions under extreme working conditions, has become an urgent problem to be solved. Summary of the Invention

[0007] To address the aforementioned shortcomings of existing technologies, the present invention aims to provide a method for constructing a fault knowledge graph for heavy-duty forging presses. This method enables high-precision structured analysis of multi-source heterogeneous fault data from heavy-duty forging presses and dynamic and complete construction of the knowledge graph, thereby improving the intelligent diagnostic capability of equipment faults and the reliability of operation and maintenance decisions under extreme operating conditions.

[0008] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0009] The method for constructing a fault knowledge graph for heavy-duty forging presses includes the following steps:

[0010] S1. Acquire multi-source heterogeneous text data of a heavy-duty forging press, including alarm records, maintenance logs, and technical documents; clean and extract values ​​from the alarm records and maintenance logs, and clean and perform hierarchical block processing with structure awareness on the technical documents to obtain formatted text data; perform term discovery on the formatted text data based on inter-point mutual information and left and right entropy algorithms to update the domain dictionary, perform synonym normalization, and perform forced word segmentation based on the updated domain dictionary to obtain preprocessed text data;

[0011] S2. Extract X% from the preprocessed text data in S1 as training samples, and perform triplet annotation to construct a fault knowledge graph annotation dataset.

[0012] S3, based on the fault knowledge graph annotation dataset of S2, train an enhanced cascaded binary labeling model;

[0013] The enhanced cascaded binary labeling model is derived from an improvement of the cascaded binary labeling model. The improvement includes introducing a global pointer mechanism into the model structure for entity boundary modeling and using a multi-head attention mechanism to enhance the contextual semantic representation. During training, adversarial training and the Focal Loss function are combined to optimize the model parameters.

[0014] S4. Using the enhanced cascaded binary labeling model trained in S3, knowledge extraction is performed on the text data preprocessed in S1 to obtain fault knowledge triples.

[0015] S5. Based on the preset pattern layer, the fault knowledge triples of S4 are fused. The knowledge fusion includes entity alignment, relation merging and conflict resolution to obtain the fused fault knowledge. The pattern layer predefines entity type, relation type and attribute type. The entity type covers four knowledge domains: equipment functional structure, status monitoring, fault and maintenance.

[0016] S6. A graph neural network is used to predict the missing associations of the fault knowledge fused in S5 in order to complete the missing associations. The completed fault knowledge is then stored in the graph database to obtain the fault knowledge graph of the heavy-duty forging press.

[0017] Compared with the prior art, the present invention has the following advantages:

[0018] 1. Improved preprocessing accuracy and domain adaptability of multi-source heterogeneous industrial text. Existing technologies typically use general-purpose word segmentation tools to process industrial text, which struggles to handle the issues of mixed technical terminology and semantic non-standardization in heavy-duty forging presses. This solution achieves forced word segmentation for the equipment-specific corpus by extracting numerical values ​​from alarm records and maintenance logs, performing structure-aware hierarchical segmentation of technical documents, and dynamically updating the domain dictionary and synonym normalization using inter-point mutual information and left / right entropy algorithms. Compared to traditional methods, this approach effectively overcomes the interference of temporal noise and non-standard descriptions, significantly improving the quality of input data for subsequent knowledge extraction.

[0019] 2. Enhanced accuracy and robustness of fault knowledge triple extraction in complex contexts. Addressing the performance degradation of existing models when handling overlapping entities and imbalanced samples, this approach proposes an enhanced cascaded binary labeling model. By introducing a global pointer mechanism to accurately model entity boundaries, utilizing a multi-head attention mechanism to strengthen contextual semantic understanding, and combining adversarial training and Focal Loss function to optimize parameters, this effectively solves the challenges of entity nesting, ambiguous relationships, and positive / negative sample imbalance in heavy-duty forging press fault data. Compared to conventional information extraction models, this method demonstrates superior generalization ability and extraction accuracy in low-resource, high-noise industrial scenarios.

[0020] 3. Deep semantic fusion across the entire equipment lifecycle knowledge domain is achieved. Unlike general knowledge graphs that lack domain constraints leading to knowledge conflicts, this solution pre-defines a schema layer covering four knowledge domains: equipment functional structure, status monitoring, fault and maintenance, and performs entity alignment, relationship merging, and conflict resolution accordingly. This fusion strategy based on top-level ontology design ensures the logical consistency and standardization of knowledge extracted from multi-source heterogeneous data, laying the foundation for building a high-quality knowledge base that conforms to the operation and maintenance logic of heavy-duty forging presses.

[0021] 4. It completes the potential fault associations under sparse data, improving the completeness and reasoning ability of the knowledge graph. Existing technologies often stop at static storage to address the problem of missing knowledge graph associations caused by sparse fault occurrence and incomplete data collection. This solution uses graph neural networks to predict links in the fused knowledge, fully mining high-order semantic features in the graph structure and automatically inferring and completing implicit fault-cause-maintenance associations. This not only solves the graph sparsity problem but also endows the knowledge graph with dynamic evolution and intelligent reasoning capabilities, making it more adaptable to the analysis needs of complex faults under extreme operating conditions.

[0022] In summary, this method can achieve high-precision structured analysis of multi-source heterogeneous fault data of heavy-duty forging presses and dynamic and complete construction of knowledge graphs, thereby improving the intelligent diagnosis capability of equipment faults and the reliability of operation and maintenance decisions under extreme working conditions.

[0023] Preferably, in S1, the step of performing term discovery on formatted text data based on inter-point mutual information and left / right entropy algorithms to update the domain dictionary and perform synonym normalization includes:

[0024] Obtain candidate words from the formatted text data ,in, For the i-th character, The length of the character string;

[0025] Calculate candidate words Inter-point information:

[0026] ;

[0027] in, This represents the probability of a candidate word appearing. For a single character The probability of its occurrence;

[0028] Calculate the left entropy of candidate words and right entropy :

[0029] ;

[0030] ;

[0031] in, and Candidate words The set of characters to the left and right of the string; count(⋅) is the frequency of the string;

[0032] when and When the candidate words are selected, they are considered as new term candidates, resulting in candidate word w; where, , This is a preset empirical threshold.

[0033] A standard terminology seed set is constructed using words from the domain dictionary, and words in the standard terminology seed set are used as candidate standard terms; the comprehensive similarity between candidate term w and each candidate standard term s in the standard terminology seed set is calculated respectively.

[0034] ;

[0035] in, To be become Minimum number of single-character edits required; , Each word and The length of the characters; These are pre-trained Chinese word vectors; As a balance factor;

[0036] Find the candidate standard word s with the highest comprehensive similarity; if the highest comprehensive similarity is greater than a preset threshold, then map the candidate word w to the candidate standard word s; if the highest comprehensive similarity is not greater than the preset threshold, then add the candidate word w as an independent new term to the domain dictionary, and use it as a candidate standard word for subsequent independent new term discovery and synonym normalization processes of the corpus.

[0037] This setup, by introducing adversarial perturbations during training to enhance the model's robustness to input noise, and by using the Focal Loss function to dynamically adjust the weights of easy and difficult samples, effectively alleviates the problem of extreme imbalance between positive and negative samples that is common in heavy forging press fault data. Compared with existing techniques that only use conventional loss functions, this combined optimization strategy enables the model to focus more on mining key fault features when faced with semantically ambiguous and sparsely labeled industrial text, thereby improving the accuracy and generalization ability of triple extraction.

[0038] Preferably, in S1, the hierarchical block processing for cleaning and structure-aware analysis of technical documents includes:

[0039] The technical documents were converted into structured text with Markdown tags using a parsing engine;

[0040] The title tags in the structured text are parsed, and the text is recursively segmented using chapters and sections as basic units, treating each title and its subordinate paragraphs, tables, and lists as an independent logical block.

[0041] The tables in the structured text are identified and converted into JSON format for independent storage, without being involved in subsequent text segmentation;

[0042] When the number of characters in the independent logic block exceeds a preset length threshold, a sliding window is used for overlapping segmentation. The window size is 256 characters, the step size is 128 characters, and the overlap rate is 50%.

[0043] This approach addresses the shortcomings of traditional document segmentation methods, which often disrupt the semantic coherence and structural integrity of technical documents. By preserving the title-content hierarchy, isolating non-textual structures, and introducing overlapping sliding windows within very long logical blocks, it achieves a balance between semantic integrity and processability in technical documents. It avoids gaps in fault mechanism descriptions caused by improper segmentation while ensuring sufficient coverage of contextual information during subsequent terminology discovery and triple extraction. Compared to general NLP preprocessing workflows, this structure-aware segmentation strategy significantly improves the mineability and extraction consistency of implicit fault knowledge in technical documents.

[0044] Preferably, in S1, the cleaning and value extraction of alarm records and maintenance logs includes:

[0045] Regular expressions are used to match numeric strings in alarm records and maintenance logs, where the numeric strings include pressure values, temperature values, and displacement amounts.

[0046] The matched numerical strings are normalized to convert similar physical quantities with different units into a unified standard unit.

[0047] The normalized numerical values ​​are bound to their corresponding physical quantity semantic tags to form structured numerical fields, which are then embedded in the formatted text data.

[0048] This setup effectively overcomes the structural limitations of traditional information extraction models in processing fault texts of specialized equipment, providing crucial model support for constructing high-confidence fault knowledge triples.

[0049] Preferably, in S1, the forced word segmentation based on the updated domain dictionary includes:

[0050] The updated domain dictionary is organized into a Trie tree structure. The longest forward match is performed on the segmented text from left to right, and the matched domain terms are marked as indivisible units.

[0051] Use a general word segmenter to perform statistical segmentation on the remaining character sequences that were not matched;

[0052] The two results are merged in their original order to obtain the preprocessed text data.

[0053] This approach addresses common issues in the extraction of heterogeneous data from multiple sources for heavy-duty forging presses, such as inconsistent referentiality, redundant relationships, and semantic contradictions. It integrates scattered triples into a unified knowledge framework through top-level ontology constraints. On one hand, it uses four core knowledge domains as anchors to achieve semantic unification of entities across texts and modalities, avoiding knowledge fragmentation caused by differences in expression. On the other hand, it automatically identifies and corrects logical conflicts based on predefined relation types and attribute constraints, ensuring the inherent consistency of the fused knowledge. Compared to schema-less open knowledge fusion methods, this strategy significantly reduces manual verification costs while improving the reliability and usability of the knowledge graph in professional operation and maintenance scenarios.

[0054] Preferably, in S2, the triplet annotation adopts a strategy combining large model-assisted pre-annotation and manual verification and correction, the process of which includes:

[0055] The preprocessed text data is pre-labeled using a large language model to generate an initial candidate set of triples.

[0056] Domain experts validate and correct the initial triplet candidate set to generate structured labeled data containing cleaned text, entity lists, relation lists, and attribute lists, which serves as the fault knowledge graph labeled dataset.

[0057] This setup enables the effective conversion of fault knowledge from symbolic logic to geometric semantics, enhancing the intelligent analysis and generalization capabilities of knowledge graphs in complex industrial diagnostic scenarios.

[0058] Preferably, in S3, the enhanced concatenated binary tagging model includes a semantic encoding layer, a head entity decoding layer, and a relation-aware tail entity concatenated decoding layer:

[0059] The semantic encoding layer uses a pre-trained language model to process the input text sequence. Encode to obtain the context semantic representation The semantic encoding layer also introduces an adversarial training strategy, which modifies the input representation at the word embedding layer. Add perturbation The perturbated adversarial samples The input is encoded into the pre-trained language model, wherein, The gradient of the loss function with respect to the input embedding. It is the L2 norm. The disturbance amplitude;

[0060] The head entity decoding layer calculates the head entity score matrix based on the context semantic representation using a global pointer mechanism. Fragments with scores exceeding a preset threshold are identified as head entities; among them, , Let S and S be the learnable weight matrices for the start and end positions of the head entity, respectively. and These are the learnable bias terms for the start and end positions of the head entity, respectively. and These are the context-enhanced representation vectors for the i-th and j-th positions in the sequence, respectively.

[0061] The relation-aware tail entity concatenated decoding layer introduces a multi-head self-attention mechanism to perform deep interaction and reconstruction on the context semantic representation, and outputs a context-enhanced representation after residual connection and layer normalization. ,in Indicates cross-attention mechanism, For context-enhanced representation, This provides the semantic representation of the head entity; and employs a global pointer mechanism with independent parameters to calculate the tail entity score matrix for each relation r. ,in, and Let be the vector representations of the i-th and j-th positions, respectively. and These are the weight matrices for the starting and ending positions of the tail entity corresponding to relation r, respectively. and These are the tail entity start position bias term and tail entity end position bias term corresponding to relation r, respectively; segments with scores exceeding a preset threshold are determined as tail entities associated with the head entity, and form candidate triples with the head entity and relation.

[0062] This approach addresses the challenges commonly found in fault texts from large forging presses, such as long-distance dependencies, nested entities, and relationships dependent on contextual semantics. It enhances model robustness by introducing adversarial perturbations at the semantic encoding layer, employs global pointers at the head entity recognition layer to avoid boundary offsets, and utilizes a cross-attention mechanism to dynamically model fine-grained interactions between head entities and candidate tail entities during the tail entity generation stage. Crucially, it injects relation types as explicit conditions into the tail entity decoding process, achieving precise relation-guided localization and significantly reducing mismatches of tail entities caused by relation ambiguity.

[0063] Preferably, in S3, the enhanced cascaded binary labeling model employs a hybrid loss function. :

[0064] ;

[0065] in, Loss due to head entity recognition; For tail entity recognition loss; The standard cross-entropy loss function; These are the weighting coefficients;

[0066] ;

[0067] in, The set of positive sample fragments; A set of negative sample fragments;

[0068] ;

[0069] in, For a set of relations; and These are the sets of positive and negative sample segments corresponding to relation r, respectively.

[0070] ;

[0071] in, A collection of attribute categories; Let be the true attribute category label at position i. When there is no attribute label at position i, the value is -1, indicating that position does not participate in the attribute loss calculation. The i-th position represents the actual label belonging to attribute category c, with a value of 0 or 1; Let be the predicted probability that the i-th position belongs to attribute category c; This is an indicator function that takes the value 1 when the condition inside the parentheses is true, and 0 otherwise.

[0072] This setup effectively balances the depth of semantic understanding with the rigor of the annotation structure, and effectively solves the bottlenecks in accuracy and robustness in industrial fault text entity recognition.

[0073] Preferably, in S5, the entity alignment includes: aligning candidate entity pairs. Using semantic similarity With structural similarity The weighted sum is used to calculate the overall similarity. :

[0074] ;

[0075] in, Preset weighting coefficients;

[0076] When the overall similarity > At that time, they are merged into the same entity;

[0077] The relationship merging includes: when there are multiple candidate relationships between the same head entity and tail entity, retaining the relationship with the highest confidence.

[0078] The conflict resolution includes: when there is contradictory information about the attribute values ​​of the same entity from different sources, assigning preset weights according to the data source type and performing weighted voting, and obtaining the attribute value with the highest score that exceeds the preset confidence threshold as the final attribute value; the data source type includes structured data and model-extracted data.

[0079] This setup, through structured ontology constraints, enables the leap of fault knowledge from fragmented information to a systematic semantic network, laying a solid foundation for the automated construction of highly reliable industrial knowledge graphs.

[0080] Preferably, in S6, the graph neural network is a relational graph convolutional network; the relational graph convolutional network updates the entity embedding layer by layer within a preset number of layers, wherein the first... The embedding update formula for entity i in layer i is:

[0081] ;

[0082] in, Let i be the embedding vector of entity i at layer l; Let r be the set of neighboring entities that point to entity i through relation r; Let r be the weight matrix of relation r at layer l; σ is the self-loop transformation matrix; σ is the activation function; For a set of relations;

[0083] After the preset number of embedding updates, the score of the triple (h,r,t) is calculated using the DistMult bilinear function. :

[0084] ;

[0085] in, and These are the final embedding vectors for the head and tail entities, respectively. The embedding vector of the relation;

[0086] Based on the scores, link prediction is performed to fill in missing associations;

[0087] The relational graph convolutional network is trained using the following loss function:

[0088] ;

[0089] in, This is the set of training triples for the current batch; This refers to the spacing hyperparameter. As a positive sample, The negative sample is constructed by randomly replacing the tail entity.

[0090] This setup effectively balances the depth of semantic understanding with the rigor of the annotation structure, and effectively solves the bottlenecks in accuracy and robustness in industrial fault text entity recognition. Attached Figure Description

[0091] To make the objectives, technical solutions, and advantages of the invention clearer, the invention will now be described in further detail with reference to the accompanying drawings, wherein:

[0092] Figure 1 This is a flowchart of the method;

[0093] Figure 2 This is a schematic diagram illustrating the impact of each hyperparameter on model performance in Example 2;

[0094] Figure 3 This is a schematic diagram of the comparative experimental results in Example 2;

[0095] Figure 4 This is a schematic diagram illustrating the impact of hyperparameters on model performance in Example 2;

[0096] Figure 5 This is a schematic diagram comparing the performance of different models in Example 2;

[0097] Figure 6 This is a partial illustration of the fault knowledge graph of the large forging press in Example 2;

[0098] Figure 7 This is a visualization diagram of the fault knowledge graph of the heavy-duty forging press in Example 2. Detailed Implementation

[0099] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0100] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of them. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to represent selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0101] Example 1

[0102] like Figure 1 As shown, this invention provides a method for constructing a fault knowledge graph of a heavy-duty forging press, including the following steps:

[0103] S1. Acquire multi-source heterogeneous text data of a heavy-duty forging press, including alarm records, maintenance logs, and technical documents; clean and extract values ​​from the alarm records and maintenance logs, and clean and perform hierarchical block processing with structure awareness on the technical documents to obtain formatted text data; perform term discovery on the formatted text data based on inter-point mutual information and left and right entropy algorithms to update the domain dictionary, perform synonym normalization, and perform forced word segmentation based on the updated domain dictionary to obtain preprocessed text data.

[0104] In specific implementation, the alarm records and maintenance logs are cleaned and data extracted, wherein the data extraction includes:

[0105] Regular expressions are used to match numeric strings in alarm records and maintenance logs, where the numeric strings include pressure values, temperature values, and displacement amounts.

[0106] The matched numerical strings are normalized to convert similar physical quantities with different units into a unified standard unit.

[0107] The normalized numerical values ​​are bound to their corresponding physical quantity semantic tags to form structured numerical fields, which are then embedded in the formatted text data.

[0108] In specific implementation, the cleaning and structure-aware hierarchical block processing of technical documents (PDF format) includes:

[0109] Use the MinerU parsing engine to convert PDFs into structured text with Markdown tags;

[0110] Hierarchical segmentation: The Markdown heading tags (#, ##, ###) are parsed and recursively segmented into chapters and sections. Each heading and its subordinate paragraphs, tables, lists, etc., are treated as an independent logical block. This approach maintains the semantic integrity within chapters and avoids artificially separating strongly related content such as headings from body text, table headers from table bodies, etc.

[0111] Table Structure Extraction: Recognize Markdown tables and convert them to JSON format. The text content of each cell is preserved, and the correspondence between table headers and rows is recorded. The structured table is stored independently and is not involved in subsequent text segmentation, allowing for direct referencing of the table's internal relationships during subsequent knowledge extraction.

[0112] Long text protection: When the number of characters in a text block (e.g., a section) exceeds a preset threshold (in practice, Lmax=512 is used as an alternative strategy), a sliding window is used for overlapping segmentation: window size W=256 characters, step size s=128 characters, and overlap rate 50%. This strategy is only for very long logical blocks to avoid truncation of subsequent model input due to excessively long single blocks, while the overlapping window retains boundary information.

[0113] The sliding window size W = 256 characters, step size s = 128 characters, and overlap rate of 50% are mainly based on two points: First, according to statistics, the length of each semantic unit (such as a complete paragraph or 4 to 6 sentences) in technical literature is mostly between 200 and 300 characters, and 256 characters can accommodate enough context without exceeding the input length limit of subsequent models; Second, 50% overlap ensures that causal relationships are not cut off by the window boundary, while also taking into account computational efficiency, which is a commonly used empirical configuration in sliding window tasks.

[0114] The above method prioritizes document structure segmentation, preserving the integrity of inherent logical units such as chapters and tables in technical documents to the greatest extent, and providing semantically coherent input fragments for subsequent research.

[0115] In specific implementation, the term discovery based on inter-point mutual information and left / right entropy algorithms for formatted text data to update the domain dictionary, and the synonym normalization, includes:

[0116] Obtain candidate words from the formatted text data ,in, For the i-th character, The length of the character string;

[0117] Calculate candidate words Inter-point information:

[0118] ;

[0119] in, This represents the probability of a candidate word appearing. For a single character The probability of its occurrence;

[0120] For candidate words, the PMI value of their sub-word combinations indicates that the larger the PMI, the tighter the internal binding of the words.

[0121] Calculate the left entropy of candidate words and right entropy :

[0122] ;

[0123] ;

[0124] wherein, and are respectively the left neighbor character set and the right neighbor character set of the candidate word ; count(·) is the occurrence frequency of the character string;

[0125] when and , the candidate word is determined as a new term candidate to obtain the candidate word w; wherein, , are preset empirical thresholds.

[0126] Generally, the pointwise mutual information takes a value between 0 and 10 after being logarithmized. For term discovery, 3 is a commonly used empirical value, which can ensure that the internal characters of the candidate word are closely associated. If it is lower than 3, noise is easily introduced (such as the combination of "de" and subsequent characters); if it is higher than 5, terms with shorter length may be missed. After testing on 50 annotated terms, a recall rate of about 90% can be achieved. The left and right entropies reflect the richness of the context. It is generally considered that an entropy greater than 1.5 indicates high boundary freedom and good word stability. This value is obtained through statistical analysis of the distribution of left and right adjacencies of words in the corpus, which is consistent with the recommended values in most Chinese new word discovery literatures. In specific implementation, the empirical thresholds are set , .

[0127] constructing a standard term seed set by using the words in the domain dictionary, and taking the words in the standard term seed set as candidate standard words; respectively calculating the comprehensive similarity between the candidate word w and each candidate standard word s in the standard term seed set:

[0128] ;

[0129] wherein, is the minimum number of single-character editing operations required to convert into ; , are respectively the character lengths of the word and ; is a pre-trained Chinese word vector (Word2Vec, with a dimension of 300); is a balance factor;

[0130] finding the candidate standard word s with the highest comprehensive similarity; if the highest comprehensive similarity is greater than a preset threshold, mapping the candidate word w to the candidate standard word s; if the highest comprehensive similarity is not greater than the preset threshold, adding the candidate word w to the domain dictionary as an independent new term, and using it as a candidate standard word for subsequent independent new term discovery and synonym normalization processes of the corpus.

[0131] In practice, =0.6; the preset threshold is 0.85.

[0132] By introducing adversarial perturbations during training to enhance the model's robustness to input noise, and by dynamically adjusting the weights of easy and difficult samples using the Focal Loss function, the problem of extreme imbalance between positive and negative samples, which is common in heavy forging press fault data, is effectively alleviated. Compared with existing techniques that only use conventional loss functions, this combined optimization strategy enables the model to focus more on mining key fault features when faced with semantically ambiguous and sparsely labeled industrial text, thereby improving the accuracy and generalization ability of triple extraction.

[0133] In specific implementation, the forced word segmentation based on the updated domain dictionary includes:

[0134] The updated domain dictionary is organized into a Trie tree structure. The longest forward match is performed on the segmented text from left to right, and the matched domain terms are marked as indivisible units. For example, if the dictionary contains "main cylinder pressure" for "master cylinder pressure drop", the whole text is locked to avoid splitting it into "main cylinder" and "pressure".

[0135] Then, a general word segmenter is used to perform statistical segmentation on the remaining unmatched character sequences; in practice, the remaining unmatched character sequences are fed into the jieba word segmenter (which uses bidirectional maximum matching + hidden Markov model) for statistical segmentation.

[0136] Finally, the two results are merged in their original order to obtain the preprocessed text data.

[0137] S2. Extract X% from the preprocessed text data in S1 as training samples, and perform triplet annotation to construct a fault knowledge graph annotation dataset.

[0138] In practice, the value of X can be set according to the amount of preprocessed text data, which will not be elaborated here.

[0139] In practice, the triplet annotation adopts a strategy combining large model-assisted pre-annotation and manual verification and correction. The process includes:

[0140] The preprocessed text data is pre-labeled using a large language model to generate an initial candidate set of triples.

[0141] Domain experts validate and correct the initial triplet candidate set to generate structured labeled data containing cleaned text, entity lists, relation lists, and attribute lists, which serves as the fault knowledge graph labeled dataset.

[0142] S3, based on the fault knowledge graph annotation dataset of S2, train an enhanced cascaded binary labeling model;

[0143] The enhanced cascaded binary labeling model is derived from an improvement of the cascaded binary labeling model. The improvement includes introducing a global pointer mechanism into the model structure for entity boundary modeling and using a multi-head attention mechanism to enhance the contextual semantic representation. During training, adversarial training and the Focal Loss function are combined to optimize the model parameters.

[0144] In specific implementation, the enhanced concatenated binary tagging model includes a semantic encoding layer, a context enhancement layer, a head entity decoding layer, and a relation-aware tail entity concatenated decoding layer:

[0145] Semantic coding layer

[0146] The semantic encoding layer uses a pre-trained language model to process the input text sequence. Encode to obtain the context semantic representation In practice, to convert unstructured fault repair text into structured vector representations, the RoBERTa-wwm-ext model, pre-trained on Chinese corpora, was selected as the semantic encoder. This model consists of 12 stacked Transformer encoder layers with a hidden layer dimension of 768 and 12 heads for the multi-head attention mechanism. The input encoding vector of RoBERTa is composed of three parts: word embedding, segment embedding, and position embedding, which are added element-wise to represent the semantic information of the token, the information of its sentence category, and its position information in the sequence, respectively.

[0147] The semantic encoding layer also introduces an adversarial training strategy, which modifies the input representation at the word embedding layer. Add perturbation:

[0148] ;

[0149] in, The gradient of the loss function with respect to the input embedding. It is the L2 norm. The amplitude of the disturbance;

[0150] Perturbated adversarial examples The pre-trained language model is encoded and then fed into the model for forward propagation and parameter updates, thereby enhancing the model's robustness to noisy data and minor perturbations. It should be noted that adversarial perturbations are introduced only during the training phase and not during the model inference phase. That is, adversarial inputs are introduced during the training phase. The input is fed into a pre-trained language model for encoding; at the inference node, the original input is used. .

[0151] For a unified representation, the input is fed into the RoBERTa-wwm-ext model, where context modeling is performed through a multi-layer Transformer encoder. The encoding process can be represented as follows:

[0152] ;

[0153] ;

[0154] in, For the first The hidden layer output vector corresponding to each token This is the hidden dimension of RoBERTa. This vector integrates the semantics of the current token itself and its contextual information.

[0155] Context enhancement layer

[0156] To further enhance the model's ability to model long-distance dependencies and improve the fine-grained characterization of entity boundaries, a multi-head self-attention mechanism is introduced based on the semantic encoding layer output to perform deep interaction and reconstruction of the context representation. Let the encoding layer output be... ,in Indicates the sequence length. Represents the hidden layer dimension.

[0157] By performing a linear mapping on the input representation, we obtain a query, key, and value matrix:

[0158] ;

[0159] in, For learnable parameter matrix, The subspace dimension for each attention head (usually taken as...) , (Number of heads).

[0160] The calculation process for single-head self-attention is as follows:

[0161] ;

[0162] To enhance the model's expressive power, a multi-head mechanism is introduced, mapping the input to multiple subspaces and computing attention in parallel:

[0163] ;

[0164] ;

[0165] in, This is the number of attention heads (in practice, it is set to 8). This is the output mapping matrix. Multi-head structures can capture multi-granularity dependencies from different semantic subspaces, which helps in modeling entity boundary information in complex contexts.

[0166] Subsequently, residual connections and layer normalization are introduced to stabilize the training process and alleviate the gradient vanishing problem, resulting in the final context-enhanced representation:

[0167] ;

[0168] ;

[0169] Through the above process, the model can establish dependencies between tokens globally, ensuring that the representation of each position not only contains its own semantic information but also incorporates related contextual information. Compared to encoding results that rely solely on BERT, this enhancement layer further strengthens the ability to characterize long-distance semantic associations and entity boundary features, thereby providing a more discriminative representation for subsequent entity recognition and relation extraction.

[0170] Header Entity Decoding Layer

[0171] The head entity decoding layer calculates the head entity score matrix based on the context semantic representation using a global pointer mechanism:

[0172] ;

[0173] in, , Let S and S be the learnable weight matrices for the start and end positions of the head entity, respectively. and These are the learnable bias terms for the start and end positions of the head entity, respectively. and These are the context-enhanced representation vectors for the i-th and j-th positions in the sequence, respectively.

[0174] Fragments with scores exceeding a preset threshold are identified as head entities; in specific implementation, to obtain probabilistic prediction results, a Sigmoid mapping is applied to the scores:

[0175] ;

[0176] in, The Sigmoid activation function is defined as follows:

[0177] ;

[0178] This function maps the score to the interval (0,1) to represent position pairs. The probability of forming the head entity. The reason for using the Sigmoid function instead of Softmax is that this task is a multi-label discrimination problem, and multiple entity segments may exist simultaneously in the same sequence.

[0179] During the model training phase, multi-label cross-entropy loss is used for optimization. During the inference phase, all... The score exceeds the preset threshold fragments It was identified as a candidate header entity. The preset threshold is set to 0.5 in practice. Global pointers can naturally handle nested and overlapping entities without requiring subsequent pairing and processing.

[0180] The relation-aware tail entity concatenated decoding layer introduces a multi-head self-attention mechanism to perform deep interaction and reconstruction on the context semantic representation, and outputs a context-enhanced representation after residual connection and layer normalization. ,in Indicates cross-attention mechanism, For context-enhanced representation, This provides the semantic representation of the head entity; and employs a global pointer mechanism with independent parameters to calculate the tail entity score matrix for each relation r:

[0181] ;

[0182] in, and Let be the vector representations of the i-th and j-th positions, respectively. and These are the weight matrices for the starting and ending positions of the tail entity corresponding to relation r, respectively. and These are the tail entity start position offset term and tail entity end position offset term corresponding to relation r, respectively;

[0183] Fragments with scores exceeding a preset threshold are identified as tail entities associated with the head entity, and are then combined with the head entity and the relation to form candidate triples.

[0184] The enhanced cascaded binary labeling model employs a hybrid loss function. :

[0185] ;

[0186] in, Loss due to head entity recognition; For tail entity recognition loss; The standard cross-entropy loss function; These are the weighting coefficients;

[0187] ;

[0188] in, The set of positive sample fragments; A set of negative sample fragments;

[0189] ;

[0190] in, For a set of relations; and These are the sets of positive and negative sample segments corresponding to relation r, respectively.

[0191] ;

[0192] in, A collection of attribute categories; Let be the true attribute category label at position i. When there is no attribute label at position i, the value is -1, indicating that position does not participate in the attribute loss calculation. The i-th position represents the actual label belonging to attribute category c, with a value of 0 or 1; Let be the predicted probability that the i-th position belongs to attribute category c; This is an indicator function that takes the value 1 when the condition inside the parentheses is true, and 0 otherwise.

[0193] To address the challenges commonly found in fault texts of large forging presses, such as long-distance dependencies, nested entities, and relationships dependent on contextual semantics, this paper improves model robustness by introducing adversarial perturbations at the semantic encoding layer, employs global pointers at the head entity recognition layer to avoid boundary offsets, and utilizes a cross-attention mechanism to dynamically model fine-grained interaction relationships between head entities and candidate tail entities during the tail entity generation stage. Crucially, it injects relation type as an explicit condition into the tail entity decoding process, thereby achieving precise localization guided by relations and significantly reducing mismatches of tail entities caused by relation ambiguity.

[0194] The hybrid loss function effectively balances the depth of semantic understanding with the rigor of the annotation structure, and effectively solves the bottleneck of accuracy and robustness in industrial fault text entity recognition.

[0195] S4. Using the enhanced cascaded binary labeling model trained in S3, knowledge is extracted from the preprocessed text data in S1 to obtain fault knowledge triples.

[0196] S5. Based on the preset pattern layer, perform knowledge fusion on the fault knowledge triples of S4. The knowledge fusion includes entity alignment, relation merging and conflict resolution to obtain the fused fault knowledge. The pattern layer predefines entity types, relation types and attribute types. Among them, the entity types cover four knowledge domains: equipment functional structure, status monitoring, fault and maintenance.

[0197] In specific implementation, the entity alignment includes: aligning candidate entity pairs Using semantic similarity With structural similarity The weighted sum is used to calculate the overall similarity. :

[0198] ;

[0199] in, Preset weighting coefficients;

[0200] When the overall similarity > At that time, they are merged into the same entity;

[0201] The relationship merging includes: when there are multiple candidate relationships between the same head entity and tail entity, retaining the relationship with the highest confidence.

[0202] The conflict resolution includes: when there is contradictory information about the attribute values ​​of the same entity from different sources, assigning preset weights according to the data source type and performing weighted voting, and obtaining the attribute value with the highest score that exceeds the preset confidence threshold as the final attribute value; the data source type includes structured data and model-extracted data.

[0203] S6. A graph neural network is used to predict the missing associations of the fault knowledge fused in S5 in order to complete the missing associations. The completed fault knowledge is then stored in the graph database to obtain the fault knowledge graph of the heavy-duty forging press.

[0204] In specific implementation, the graph neural network is a relational graph convolutional network; the relational graph convolutional network updates the entity embedding layer by layer within a preset number of layers, wherein the first... The embedding update formula for entity i in layer i is:

[0205] ;

[0206] in, Let i be the embedding vector of entity i at layer l; Let r be the set of neighboring entities that point to entity i through relation r; Let r be the weight matrix of relation r at layer l; σ is the self-loop transformation matrix; σ is the activation function; For a set of relations;

[0207] To facilitate better understanding, the following explanation is provided.

[0208] The above formula can be broken down into two steps: message construction and message aggregation.

[0209] Message construction phase: For each relation, each neighbor generates a message through a linear transformation.

[0210] ;

[0211] Message aggregation phase: for the same relation The average of all neighbor messages is taken to obtain the aggregate information under this relation type; then the summation is applied to all relation types to obtain the overall contribution of all neighbors; finally, the self-loop transformed features are added, and a new embedding is obtained through the activation function.

[0212] .

[0213] After the preset number of embedding updates, the score of the triple (h,r,t) is calculated using the DistMult bilinear function. :

[0214] ;

[0215] in, and These are the final embedding vectors for the head and tail entities, respectively. The embedding vector of the relation;

[0216] The specific number of preset layers can be set according to specific needs. In this embodiment, the preset number of layers is 2.

[0217] Based on the scores, link prediction is performed to fill in missing associations;

[0218] The relational graph convolutional network is trained using the following loss function:

[0219] ;

[0220] in, This is the set of training triples for the current batch; This refers to the spacing hyperparameter. As a positive sample, The negative sample is constructed by randomly replacing the tail entity.

[0221] This method can achieve high-precision structured analysis of multi-source heterogeneous fault data of heavy-duty forging presses and dynamic and complete construction of knowledge graphs, thereby improving the intelligent diagnosis capability of equipment faults and the reliability of operation and maintenance decisions under extreme working conditions.

[0222] Example 2

[0223] To better illustrate the effectiveness of this method, the following experimental verification was conducted.

[0224] 1. Fault knowledge extraction experiment verification

[0225] (1) Experimental evaluation indicators and environmental configuration

[0226] Precision, recall, and F1 score—common metrics in triplet extraction tasks—are used as evaluation metrics. Let the number of correctly predicted triples by the model be... (True example), the total number of predicted triples is ( (As a false positive example), the total number of real triples is ( (This is a false negative example).

[0227] A triple is considered correct only when both the entity and relation in the triple match the ground truth. Specifically, accuracy was not used as an evaluation metric in this experiment because the triple extraction task suffers from a severe imbalance between positive and negative samples: the number of non-entity tokens and invalid triples far exceeds the number of true triples. In this situation, the model is prone to predicting all samples as negative, resulting in potentially high accuracy but zero recall, failing to effectively reflect true performance. Therefore, precision, recall, and F1 score more objectively measure the model's extraction capability.

[0228] The preprocessed corpus dataset was used for model training, parameter tuning, and performance evaluation. The training, validation, and test sets were divided in an 8:1:1 ratio.

[0229] To verify the effectiveness of the proposed enhanced cascaded binary labeling model in extracting textual knowledge about forging equipment faults, a detailed explanation is provided regarding the experimental setup, model hyperparameters, comparative experiments, and ablation experiments. The experimental environment is shown in Table 1.

[0230] Table 1

[0231]

[0232] (2) Model hyperparameter configuration

[0233] To determine the optimal hyperparameter configuration for the model on the dataset and to explore the impact of key hyperparameters on model performance, a series of hyperparameter tuning experiments were designed and conducted. The experiments focused on four hyperparameters that significantly affect model performance: learning rate, batch size, maximum entity length, and decoding threshold.

[0234] The search range and baseline values ​​for each hyperparameter are shown in Table 2.

[0235] Table 2

[0236]

[0237] The macroscopic F1 score on the validation set was used as the evaluation metric for hyperparameter selection. To ensure statistical significance, each experiment was repeated three times, and the average value was taken as the final result to reduce random error. Finally, the parameter combination with the best performance on the validation set was selected as the final hyperparameter configuration for the model.

[0238] Because RoBERTa-wwm-ext is used as the semantic encoding layer, this pre-trained model already possesses general semantic representation capabilities. To avoid catastrophic forgetting under large gradient updates and to protect the structural stability of the pre-trained weights, the learning rate search interval is set at... .

[0239] To determine the optimal hyperparameter configuration, experiments were conducted on learning rate, batch size, decoding threshold, and entity length limit. The results are as follows: Figure 2 As shown, the learning rate is 3×10. -5 The model achieved an F1 score of 84.0%. A lower score resulted in slow convergence, while a higher score caused training oscillations. A batch size of 16 resulted in the fastest convergence and the best F1 score on the test set. Smaller batches converged quickly but had higher gradient noise, while larger batches were more stable but converged more slowly. A decoding threshold of 0.5 achieved the best balance between precision and recall. A threshold that was too low could lead to overprediction, while a threshold that was too high would be sensitive to atypical features. A limit of 20 for entity length resulted in stable training, moderate computational cost, and a good balance between entity boundary recognition and inference efficiency. Subsequent experiments all used the above configuration to ensure high accuracy and stable convergence of the model.

[0240] Based on the above experimental analysis, the optimal hyperparameter configuration of the model is shown in Table 3.

[0241] Table 3

[0242]

[0243] This configuration achieved the best F1 score on the test set, improving it by 3.9 percentage points compared to the baseline parameters. Applying this configuration to the test set for final evaluation yielded an F1 score of 88.9%, validating the effectiveness and generalization ability of the proposed hyperparameter configuration. All subsequent comparative experiments and ablation studies were conducted based on this configuration.

[0244] (3) Ablation test

[0245] To verify the effectiveness of the improved modules in the proposed model (AECasRel model), multiple ablation experiments were designed to observe the impact on knowledge extraction performance by progressively removing or replacing key components. All ablation experiments used the same training strategy, random seed, and evaluation metrics to ensure the comparability of results.

[0246] Using the original CasRel model as a baseline, global pointer, multi-head attention, adversarial training, attribute prediction, and hybrid loss are progressively integrated. All variants use the same dataset and hyperparameters (learning rate 3×10⁻⁻). 5 (Batch size 16, maximum sequence length 128). The experimental results are shown in Table 4.

[0247] Table 4

[0248]

[0249] As shown in Table 4, the introduction of each module positively contributes to the model performance. Replacing the traditional pointer network with a global pointer improves the F1 score by 2.4 percentage points, validating its advantages in handling overlapping and nested entities. The introduction of the multi-head attention mechanism brings a 1.0 percentage point improvement, indicating that enhanced contextual representation and head entity fusion effectively improve relation perception. Adversarial training further improves the score by 1.5 percentage points, confirming its robustness to input noise. The attribute prediction task brings a slight improvement of 0.6 percentage points through multi-task learning. The hybrid loss effectively alleviates sample imbalance, further improving the F1 score by 1.4 percentage points. The final complete model improves by 7.7 percentage points compared to the baseline, fully validating the effectiveness and synergistic effect of each module.

[0250] (4) Comparative experiment

[0251] To verify the superiority of our proposed AECasRel model in the task of extracting fault knowledge from forging equipment, the following mainstream models were selected as baselines for comparison: BERT-BiLSTM-CRF (sequence labeling method), BERT+Multi-HeadSelection (multi-head selection mechanism), Tagging Joint Tagging (joint annotation model), GraphRel graph neural network relation extraction, CasRel (cascaded binary tagging framework), and TPLinker (pointer network joint extraction). Among them, BERT-BiLSTM-CRF, as a representative of sequence labeling methods, was used to compare basic information extraction capabilities; BERT+Multi-HeadSelection embodies the characteristics of a multi-relation extraction mechanism; Joint Tagging achieves joint extraction of entities and relations through a unified annotation method; GraphRel can utilize structural information to improve the recognition effect of complex semantic relations; CasRel, as a typical method of cascaded binary tagging framework, is directly comparable to our proposed model; and TPLinker, as a currently superior joint extraction model, was used to evaluate the model's performance in complex overlapping triplet scenarios. The above methods reflect the development level of existing knowledge extraction technologies from different perspectives, and can provide a relatively comprehensive comparative analysis of this method.

[0252] All models were trained and tested on the same dataset, with hyperparameters tuned to optimal performance, and the same evaluation metrics used. Results are shown in Table 5. Figure 3 As shown.

[0253] Table 5

[0254]

[0255] As shown in the table, the traditional sequence labeling model BERT-BiLSTM-CRF is limited by its inability to handle entity overlap, with an F1 score of only 76.9%. The joint labeling model Tagging has a high recall (81.4%) but a low precision (65.8%), indicating a serious over-prediction problem. Although GraphRel and BERT+Multi-Head Selection introduce relation graphs and multi-head attention, they still do not completely solve the problem of extracting overlapping and nested entities, with F1 scores of 80.7% and 80.2%, respectively. CasRel significantly improves performance through cascaded decoding, achieving an F1 score of 84.4%. TPLinker, using a pointer network, further improves the F1 score to 85.9%. Our proposed AECasRel model, based on CasRel, introduces global pointers, multi-head attention, adversarial training, and auxiliary attribute prediction, effectively handling entity overlap, imbalanced samples, and long-distance dependencies, achieving an F1 score of 92.1%, which is 6.2 percentage points higher than the best baseline TPLinker, validating its state-of-the-art performance in industrial fault knowledge extraction tasks.

[0256] 2. Fault Knowledge Graph Link Prediction Experiment Verification

[0257] The constructed fault knowledge graph was used as the dataset for the experiment. This dataset covers multiple types of faults, including 16 entity types such as equipment, components, fault manifestations, and maintenance measures, and 18 relation types such as "corresponding maintenance plan," "included location," and "fault occurred." The original set of triples totaled 7989 entries, which were randomly divided into training, validation, and test sets in an 8:1:1 ratio.

[0258] (1) Evaluation indicators

[0259] To quantify model performance, the mean reciprocal ranking (MRR) and hit rate (Hits@n) were used as evaluation metrics for the fault knowledge graph link prediction experiment.

[0260] (2) Parameter settings

[0261] To determine the optimal configuration of the RGCN model for the fault knowledge graph completion task, a grid search was used to fine-tune the key hyperparameters. The search range and final values ​​are shown in Table 6. The evaluation metrics were the mean reciprocal rank (MRR) and Hits@10 on the validation set.

[0262] Table 6

[0263]

[0264] During the hyperparameter search process, the mean reciprocal rank (MRR) on the validation set is used as the evaluation metric to observe the influence trend of different parameters on model performance, such as... Figure 4As shown in the figure. Experiments show that the model performs best with a learning rate of 5×10−4, an embedding dimension of 128, 3 RGCN layers, 100 negative samples, the AdamW optimizer, a batch size of 512, and an edge dropout rate of 0.2. Compared to the baseline parameter configuration, this configuration improves the MRR on the test set by 3.9 percentage points, achieving a final MRR of 0.39 on the test set and 0.52 for Hits@10.

[0265] (3) Algorithm performance comparison

[0266] To verify the effectiveness and superiority of the graph neural network (RGCN model) used in this method for predicting fault knowledge graph links in forging equipment, several representative knowledge graph embedding and graph neural network models were selected as comparative baselines, including: TransE, DistMult, CompGCN, RotatE, TuckER, CompTuck, and RGCN. Among them, TransE, as a classic translation distance model, is used to measure the modeling ability of basic embedding methods; DistMult uses a bilinear function to characterize entity relationship interactions, but is limited by the symmetry assumption; CompGCN enhances node representation ability by introducing graph structure information; RotatE models relational semantics through a complex space rotation mechanism; TuckER achieves multi-relation modeling based on tensor decomposition; CompTuck introduces a combinatorial enhancement mechanism based on TuckER; and RGCN achieves the aggregation and propagation of multi-relational neighborhood information through relation-aware graph convolution.

[0267] All models were trained and tested on the same dataset, using the same data partitioning method and evaluation metrics. Hyperparameters of each model were tuned to their optimal performance state to ensure the fairness and reliability of the experimental comparison. Evaluation metrics included MRR, Hits@1, Hits@3, and Hits@10. Experimental results are shown in Table 7. Figure 5 As shown.

[0268] Table 7

[0269]

[0270] In all evaluation metrics, the RGCN model proposed in this method achieved the best performance, with an MRR of 0.3991, Hits@1 of 0.3279, Hits@3 of 0.4399, and Hits@10 of 0.5225. Compared with other baseline models, RGCN showed a stable advantage in all evaluation metrics. This indicates that RGCN can effectively utilize the graph structure information and relational diversity in knowledge graphs, and has a significant advantage in fault link prediction tasks. RotatE performed extremely poorly on this dataset (MRR of only 0.0019), indicating that its complex embedding assumption is not applicable to the complex relational patterns of this fault knowledge graph. Although models such as CompGCN and TuckER introduce combinatorial operations or tensor decomposition, their performance is still inferior to RGCN, further validating the applicability of relational graph convolutional networks in industrial fault knowledge graph completion tasks.

[0271] 3. Examples of Fault Knowledge Graph Construction and Application

[0272] After verifying the accuracy of the above model, to visually demonstrate the overall effectiveness of this method in fault knowledge extraction and graph completion, a typical alarm record is selected for case analysis: "Insufficient pilot pressure of hydraulic cylinder C balance control valve V0202C". This record involves entity types including specific location and fault manifestation, and relationship types including attribution and fault occurrence.

[0273] Entity relations were extracted using the AECasRel model, and the extraction results are shown in Table 8.

[0274] Table 8

[0275]

[0276] By linking entities, "insufficient pilot pressure" is mapped to existing fault manifestation entities in the knowledge graph. The extracted triples accurately reflect the compositional relationships between components, fault phenomena, and corresponding maintenance solutions.

[0277] In the knowledge graph, the fault symptom "insufficient pilot pressure" is associated with multiple candidate repair solutions through the relationship "corresponding repair solution." To assist field engineers in selecting the most appropriate solution, a trained RGCN link prediction model is used to analyze each candidate repair solution. Calculate the score This score reflects the degree of fit between the solution and the failure performance. The Top-3 results of the model output are shown in Table 9.

[0278] Table 9

[0279]

[0280] The model successfully ranked and recommended multiple candidate repair solutions associated with the knowledge graph using the fault symptom "insufficient pilot pressure" extracted from the alarm text as the query. Among them, "cleaning / grinding the valve core" ranked first with a score of 98.5%, and its reasoning was based on "physical topology matching," indicating that the model accurately located the fault to the valve core level through the contained component relationships in the graph structure, thus prioritizing the recommendation of low-cost, replacement-free solutions for the valve core. This result verifies that the proposed method can effectively integrate graph structure information and industrial economic constraints, providing a reasonable ranking of repair solutions for fault symptoms and significantly improving the practicality of diagnostic decisions.

[0281] 4. Fault Knowledge Graph Storage and Display

[0282] The constructed fault knowledge graph integrates three types of knowledge sources: structured data mapping, AECasRel model extraction, and RGCN link prediction completion. In practical applications, it needs to support multi-hop relationship queries. For example, starting from the fault symptom "insufficient pilot pressure," it can be linked to different solutions such as "cleaning the valve core" and "replacing the balance valve" through "corresponding maintenance plan," and further find the required spare parts through the "replace spare parts" relationship, thus forming a complete diagnostic decision path.

[0283] Neo4j graph database was chosen as the storage solution. Neo4j's native graph storage structure and efficient transaction processing capabilities support complex path queries and real-time inference, meeting the data retrieval needs in fault diagnosis scenarios. Entity-relationship triples were exported in CSV format, and nodes, relationships, and attributes were created in batches using the LOAD CSV command. A selection of entities and relationships were selected for visualization, such as... Figure 6 As shown.

[0284] like Figure 7 The image shows the overall visualization result of the fault knowledge graph constructed using this method. This knowledge graph contains 5031 entities and 7989 relationships. Entity types cover 16 categories, including equipment, specific locations, fault manifestations, repair methods, and spare parts; relationship types cover 18 categories, including included locations, fault occurrences, corresponding repair plans, and replacement spare parts. The knowledge graph constructed using this method has considerable scale and coverage, providing sufficient knowledge support for subsequent fault diagnosis.

[0285] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit the technical solutions. Those skilled in the art should understand that any modifications or equivalent substitutions to the technical solutions of the present invention without departing from the spirit and scope of the present invention should be covered within the scope of the claims of the present invention.

Claims

1. A method for constructing a fault knowledge graph for heavy-duty forging presses, characterized in that, Includes the following steps: S1. Obtain multi-source heterogeneous text data of the heavy-duty forging press, the multi-source heterogeneous text data including alarm records, maintenance logs and technical documents; clean and extract values ​​from the alarm records and maintenance logs, and clean and perform hierarchical block processing with structure awareness on the technical documents to obtain formatted text data. The term discovery algorithm based on inter-point mutual information and left and right entropy is used to update the domain dictionary of formatted text data, and synonym normalization is performed. Forced word segmentation is then performed based on the updated domain dictionary to obtain preprocessed text data. S2. Extract X% from the preprocessed text data in S1 as training samples, and perform triplet annotation to construct a fault knowledge graph annotation dataset. S3, based on the fault knowledge graph annotation dataset of S2, train an enhanced cascaded binary labeling model; The enhanced cascaded binary labeling model is obtained by improving the cascaded binary labeling model. The improvements include introducing a global pointer mechanism into the model structure for entity boundary modeling and using a multi-head attention mechanism to enhance the contextual semantic representation. During training, adversarial training and the Focal Loss function are combined to optimize the model parameters; S4. Using the enhanced cascaded binary labeling model trained in S3, knowledge extraction is performed on the text data preprocessed in S1 to obtain fault knowledge triples. S5. Based on the preset pattern layer, the fault knowledge triples of S4 are fused. The knowledge fusion includes entity alignment, relation merging and conflict resolution to obtain the fused fault knowledge. The pattern layer predefines entity type, relation type and attribute type. The entity type covers four knowledge domains: equipment functional structure, status monitoring, fault and maintenance. S6. A graph neural network is used to predict the missing associations of the fault knowledge fused in S5 in order to complete the missing associations. The completed fault knowledge is then stored in the graph database to obtain the fault knowledge graph of the heavy-duty forging press.

2. The method for constructing a fault knowledge graph for heavy-duty forging presses according to claim 1, characterized in that, In S1, the step of performing term discovery on formatted text data based on inter-point mutual information and left / right entropy algorithms to update the domain dictionary and perform synonym normalization includes: Obtain candidate words from the formatted text data ,in, For the i-th character, The length of the character string; Calculate candidate words Inter-point information: ； in, This represents the probability of a candidate word appearing. For a single character The probability of its occurrence; Calculate the left entropy of candidate words and right entropy : ； ； in, and Candidate words The set of characters to the left and right of the string; count(⋅) is the frequency of the string; when and When the candidate words are selected, they are considered as new term candidates, resulting in candidate word w; where, , This is a preset empirical threshold. A standard terminology seed set is constructed using words from the domain dictionary, and words in the standard terminology seed set are used as candidate standard terms; the comprehensive similarity between candidate term w and each candidate standard term s in the standard terminology seed set is calculated respectively. ； in, To be become Minimum number of single-character edits required; , Each word and The length of the characters; These are pre-trained Chinese word vectors; As a balance factor; Find the candidate standard word s with the highest comprehensive similarity; if the highest comprehensive similarity is greater than a preset threshold, then map the candidate word w to the candidate standard word s; if the highest comprehensive similarity is not greater than the preset threshold, then add the candidate word w as an independent new term to the domain dictionary, and use it as a candidate standard word for subsequent independent new term discovery and synonym normalization processes of the corpus.

3. The method for constructing a fault knowledge graph for heavy-duty forging presses according to claim 1, characterized in that, In S1, the hierarchical block processing for cleaning and structure-aware analysis of technical documents includes: The technical documents were converted into structured text with Markdown tags using a parsing engine; The title tags in the structured text are parsed, and the text is recursively segmented using chapters and sections as basic units, treating each title and its subordinate paragraphs, tables, and lists as an independent logical block. The tables in the structured text are identified and converted into JSON format for independent storage, without being involved in subsequent text segmentation; When the number of characters in the independent logic block exceeds a preset length threshold, a sliding window is used for overlapping segmentation. The window size is 256 characters, the step size is 128 characters, and the overlap rate is 50%.

4. The method for constructing a fault knowledge graph for heavy-duty forging presses according to claim 1, characterized in that, In S1, the alarm records and maintenance logs are cleaned and data extracted, wherein the data extraction includes: Regular expressions are used to match numeric strings in alarm records and maintenance logs, where the numeric strings include pressure values, temperature values, and displacement amounts. The matched numerical strings are normalized to convert similar physical quantities with different units into a unified standard unit. The normalized numerical values ​​are bound to their corresponding physical quantity semantic tags to form structured numerical fields, which are then embedded in the formatted text data.

5. The method for constructing a fault knowledge graph for heavy-duty forging presses according to claim 1, characterized in that, In S1, the forced word segmentation based on the updated domain dictionary includes: The updated domain dictionary is organized into a Trie tree structure. The longest forward match is performed on the segmented text from left to right, and the matched domain terms are marked as indivisible units. Use a general word segmenter to perform statistical segmentation on the remaining character sequences that were not matched; The two results are merged in their original order to obtain the preprocessed text data.

6. The method for constructing a fault knowledge graph for a heavy-duty forging press according to claim 1, characterized in that, In S2, the triplet annotation adopts a strategy combining large model-assisted pre-annotation and manual verification and correction. The process includes: The preprocessed text data is pre-labeled using a large language model to generate an initial candidate set of triples. Domain experts validate and correct the initial triplet candidate set to generate structured labeled data containing cleaned text, entity lists, relation lists, and attribute lists, which serves as the fault knowledge graph labeled dataset.

7. The method for constructing a fault knowledge graph for a heavy-duty forging press according to claim 1, characterized in that, In S3, the enhanced concatenated binary tagging model includes a semantic encoding layer, a head entity decoding layer, and a relation-aware tail entity concatenated decoding layer: The semantic encoding layer uses a pre-trained language model to process the input text sequence. Encode to obtain contextual semantic representation The semantic encoding layer also introduces an adversarial training strategy, which modifies the input representation at the word embedding layer. Add perturbation The perturbated adversarial samples The input is encoded into the pre-trained language model, wherein, The gradient of the loss function with respect to the input embedding. It is the L2 norm. The disturbance amplitude; The head entity decoding layer calculates the head entity score matrix based on the context semantic representation using a global pointer mechanism. Fragments with scores exceeding a preset threshold are identified as head entities; among them, , Let S and S be the learnable weight matrices for the start and end positions of the head entity, respectively. and These are the learnable bias terms for the start and end positions of the head entity, respectively. and These are the context-enhanced representation vectors for the i-th and j-th positions in the sequence, respectively. The relation-aware tail entity concatenated decoding layer introduces a multi-head self-attention mechanism to perform deep interaction and reconstruction on the context semantic representation, and outputs a context-enhanced representation after residual connection and layer normalization. ,in Indicates cross-attention mechanism, For context-enhanced representation, This provides the semantic representation of the head entity; and employs a global pointer mechanism with independent parameters to calculate the tail entity score matrix for each relation r. ,in, and Let be the vector representations of the i-th and j-th positions, respectively. and These are the weight matrices for the starting and ending positions of the tail entity corresponding to relation r, respectively. and These are the tail entity start position bias term and tail entity end position bias term corresponding to relation r, respectively; segments with scores exceeding a preset threshold are determined as tail entities associated with the head entity, and form candidate triples with the head entity and relation.

8. The method for constructing a fault knowledge graph for a heavy-duty forging press according to claim 7, characterized in that, In S3, the enhanced cascaded binary labeling model employs a hybrid loss function. : ； in, Loss due to head entity recognition; For tail entity recognition loss; The standard cross-entropy loss function; These are the weighting coefficients; ； in, The set of positive sample fragments; A set of negative sample fragments; ； in, For a set of relations; and These are the sets of positive and negative sample segments corresponding to relation r, respectively. ； in, A collection of attribute categories; Let be the true attribute category label at position i. When there is no attribute label at position i, the value is -1, indicating that position does not participate in the attribute loss calculation. The i-th position represents the actual label belonging to attribute category c, with a value of 0 or 1; Let be the predicted probability that the i-th position belongs to attribute category c; This is an indicator function that takes the value 1 when the condition inside the parentheses is true, and 0 otherwise.

9. The method for constructing a fault knowledge graph for a heavy-duty forging press according to claim 1, characterized in that, In S5, the entity alignment includes: aligning candidate entity pairs. Using semantic similarity With structural similarity The weighted sum is used to calculate the overall similarity. : ； in, Preset weighting coefficients; When the overall similarity > At that time, they are merged into the same entity; The relationship merging includes: when there are multiple candidate relationships between the same head entity and tail entity, retaining the relationship with the highest confidence. The conflict resolution includes: when there is contradictory information about the attribute values ​​of the same entity from different sources, assigning preset weights according to the data source type and performing weighted voting, and obtaining the attribute value with the highest score that exceeds the preset confidence threshold as the final attribute value; the data source type includes structured data and model-extracted data.

10. The method for constructing a fault knowledge graph for a heavy-duty forging press according to claim 1, characterized in that, In S6, the graph neural network is a relational graph convolutional network; the relational graph convolutional network updates the entity embedding layer by layer within a preset number of layers, wherein the first... The embedding update formula for entity i in layer i is: ； in, Let i be the embedding vector of entity i at layer l; Let r be the set of neighboring entities that point to entity i through relation r; Let r be the weight matrix of relation r at layer l; σ is the self-loop transformation matrix; σ is the activation function; For a set of relations; After the preset number of embedding updates, the score of the triple (h,r,t) is calculated using the DistMult bilinear function. : ； in, and These are the final embedding vectors for the head and tail entities, respectively. The embedding vector of the relation; Based on the scores, link prediction is performed to fill in missing associations; The relational graph convolutional network is trained using the following loss function: ； in, This is the set of training triples for the current batch; This refers to the spacing hyperparameter. As a positive sample, The negative sample is constructed by randomly replacing the tail entity.

Images