Multi-source heterogeneous data fusion and analysis method in engineering field

Abstract

The application discloses a multi-source heterogeneous data fusion and analysis method in the engineering field, which comprises the following steps: performing entity matching on the data after preprocessing based on attributes, and constructing an entity decision tree according to the matching success result; after performing semantic analysis on the entity of the entity decision tree, fusing the entity and the corresponding semantic information to obtain fusion data; converting each piece of fusion data into a sequence composed of marks, calculating the relationship weight between each mark in the sequence and all other marks by using a self-attention mechanism, processing the mark sequence with context information by a deep learning model, and generating a prediction result and a corresponding explanatory output for providing understanding of the decision-making process; the application can effectively process multi-source heterogeneous data in the engineering field, provide high-quality prediction results, and provide explanations for the prediction process, and finally help to make effective decisions.

Description

Technical Field

[0001] This invention relates to the field of power system technology, and more specifically to a method for dynamic capacity expansion of cable groups based on environmental parameters. Background Technology

[0002] In modern engineering, a major challenge is handling and analyzing massive amounts of multi-source, heterogeneous data. This data may originate from various sensors, devices, and other diverse sources, including structured tabular data, unstructured text data, and even knowledge graphs. The processing and analysis of this data are of paramount importance for engineering design, optimization, and decision-making. However, due to the diversity, scale, and complexity of the data, effectively fusing and analyzing it remains a highly challenging problem.

[0003] In existing technologies, a common solution is to use various machine learning and deep learning models for data processing and analysis. However, these models typically can only handle one type of data, or require manual data fusion and preprocessing, which significantly increases workload and may lead to information loss. Furthermore, these models often lack domain knowledge, making them inefficient and inaccurate when processing engineering data.

[0004] Meanwhile, many existing models, while potentially performing well in prediction, lack interpretability. This makes it difficult to understand the model's decision-making process and the contribution of different input features to the prediction results. This poses challenges to decision analysis and improving decision-making efficiency.

[0005] Furthermore, the outputs of many existing models often require further processing and analysis before they can be directly applied to specific engineering tasks, which presents certain challenges in practical use. Summary of the Invention

[0006] To address the problems of low efficiency, information loss, poor accuracy, and lack of interpretability in the processing of multi-source heterogeneous data in the prior art, this invention provides a method for multi-source heterogeneous data fusion and analysis in the engineering field. This invention improves the accuracy and efficiency of data processing, can capture deep-seated features and complex patterns of data, can generate interpretable outputs, and the prediction results can directly serve specific engineering tasks and help with decision analysis.

[0007] Methods for multi-source heterogeneous data fusion and analysis in the engineering field include the following steps:

[0008] Entity matching is performed on the preprocessed data based on attributes, and an entity decision tree is constructed based on the successful matching results.

[0009] The attribute value of each entity is transformed into a feature vector, and similarity is calculated with the feature vector corresponding to the node in the knowledge graph to initially determine the semantic information of the entity. When there are at least two feature vectors, the weight of each feature vector is calculated based on the information of the knowledge graph, and the best semantic information is selected as the semantic parsing result of the entity attribute. After completing the semantic parsing of the entity decision tree entity, the entity and its corresponding semantic information are fused to obtain fused data.

[0010] Each piece of fused data is transformed into a sequence of labels. A self-attention mechanism is used to calculate the relationship weight between each label in the sequence and all other labels. The label sequence with contextual information is processed by a deep learning model to generate prediction results and corresponding interpretive outputs to provide an understanding of the decision-making process.

[0011] Preferably, the preprocessing is used to convert data from various formats and sources into a uniform form, and the preprocessing includes data cleaning, data standardization, and data transformation.

[0012] Preferably, the entity decision tree is used to organize and display entities and their attributes, wherein each node is an entity and each edge represents a certain relationship between entities;

[0013] The construction of an entity decision tree includes: grouping data labeled as the same entity into the same group, using each group of data rows to create a node in the decision tree, and each node containing all relevant data for the corresponding entity, and using the relationships between nodes to construct an edge between two nodes.

[0014] Preferably, the feature vector is used to obtain semantic information of the attribute value, including: literal meaning, context information, and the relationship between words.

[0015] Preferably, after initially determining the semantic information of an entity, if there is only one feature vector and the similarity is greater than a preset threshold, the semantic information of the entity is directly determined; if the similarity is less than the preset threshold, the preset threshold, data annotations, and knowledge graph are comprehensively adjusted, and the semantic information of the entity is determined.

[0016] Preferably, weights are assigned to each feature vector by calculating the global importance and recent occurrence frequency of the feature vector corresponding to the entity attribute value in the corresponding node of the knowledge graph. The weights are used to reflect the importance of the semantic information represented by the feature vector in the entire knowledge graph and the specific context.

[0017] The best semantic information is selected as the semantic parsing result of the entity attribute based on the weight. When there is a semantic conflict, the conceptual relationship in the knowledge graph and the semantic information of other entities or concepts are used to resolve the conflict, ensuring that the most appropriate semantic information is selected.

[0018] Preferably, the expression for calculating the weights is:

[0019]

[0020] Where C is a concept of a knowledge graph node, W(C) is the weight of the concept; α and β are hyperparameters used to adjust the relative importance of importance values ​​and time decay factors in weight calculation; PR(C) is the importance value of concept C, MaxPR is the maximum importance value of all concepts; T(C) is the time of the most recent occurrence of concept C, T now λ is the current time, and λ is the decay factor.

[0021] Preferably, the step of calculating the relationship weights between each label and all other labels in the sequence using a self-attention mechanism includes the following steps:

[0022] Query, key, and value: For each token in the sequence, generate: query vector, key vector, and value vector;

[0023] Weight score calculation: Calculate the dot product between the query vector of each tag and the key vectors of all tags to obtain a raw weight score;

[0024] Relationship weight calculation: Calculate the similarity of each tag to all other tags, and convert the similarity into relationship weights using the softmax function;

[0025] Weighted sum of values: Multiply the value vector of each tag by its corresponding weight score, and then sum these weighted value vectors to obtain the context representation of the current tag;

[0026] The marker is an encoding composed of any combination of words, subwords, and characters;

[0027] The context refers to the information used to capture long-distance dependencies in a sequence.

[0028] Preferably, the step of processing the labeled sequence with contextual information using a deep learning model to generate prediction results and corresponding interpretive outputs includes:

[0029] S1. Take the tag sequence containing its own context information as input;

[0030] S2. The labeled sequence is processed through at least two layers of a deep learning model;

[0031] S3. Generate a new label sequence, which contains the prediction results for each label;

[0032] S4. Generate explanatory output corresponding to the prediction results.

[0033] The advantages of this invention are:

[0034] (1) Integration of multi-source heterogeneous data: This invention can process and integrate data of various types and sources, including structured data, unstructured data, and knowledge graphs, which makes it possible to make the most of existing information resources and obtain more comprehensive and in-depth insights.

[0035] (2) Introducing engineering domain knowledge graphs: By utilizing engineering domain knowledge graphs, models can better understand and process engineering data, improving the accuracy and efficiency of data processing; at the same time, the introduction of knowledge graphs also enables models to handle complex engineering problems, such as engineering design and optimization.

[0036] (3) Transformer-based model design: The self-attention mechanism of the Transformer model can capture long-distance dependencies and complex contextual information in the data; in addition, the layered structure of the Transformer model can learn the deep features and complex patterns of the data, making the model more expressive.

[0037] (4) Self-interpretive design: The model in this invention can not only provide prediction results, but also generate interpretive outputs to help understand the model's decision-making process and the contribution of different input features to the prediction results; this helps to improve the credibility and transparency of the model, and at the same time can provide more information and basis for engineering decisions.

[0038] (5) Directly serving engineering decisions: Through post-processing, the prediction results of this invention can directly serve specific engineering tasks, help with decision analysis, and improve decision efficiency;

[0039] In summary, the aforementioned advantages enable this invention to effectively process multi-source heterogeneous data in the engineering field, provide high-quality prediction results, and offer explanations of the prediction process, ultimately helping to make effective decisions.

[0040] Other features and advantages of the embodiments of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0041] The accompanying drawings are provided to further illustrate embodiments of the present invention and form part of the specification. They are used together with the following detailed description to explain the embodiments of the present invention, but do not constitute a limitation thereof. In the drawings:

[0042] Figure 1 This is a flowchart illustrating the present invention;

[0043] Figure 2 This is a schematic diagram of semantic parsing of entities in the entity decision tree in this invention;

[0044] Figure 3 This is a schematic diagram of the self-attention mechanism calculation process in this invention;

[0045] Figure 4 This is a schematic diagram of the process for generating prediction results and corresponding explanatory outputs in this invention. Detailed Implementation

[0046] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the scope of the present invention.

[0047] Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. Numerous specific details are provided in the following description to give a thorough understanding of embodiments of this application. However, those skilled in the art will recognize that the technical solutions of this application can be practiced without one or more of the specific details, or other methods, components, apparatuses, steps, etc., can be employed. In other instances, well-known methods, apparatuses, implementations, or operations are not shown or described in detail to avoid obscuring various aspects of this application.

[0048] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0049] The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily need to be performed in the described order. For example, some operations / steps can be broken down, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0050] like Figure 1 As shown, the method for multi-source heterogeneous data fusion and analysis in the engineering field includes the following steps:

[0051] Entity matching is performed on the preprocessed data based on attributes, and an entity decision tree is constructed based on the successful matching results.

[0052] The attribute value of each entity is transformed into a feature vector, and similarity is calculated with the feature vector corresponding to the node in the knowledge graph to initially determine the semantic information of the entity. When there are at least two feature vectors, the weight of each feature vector is calculated based on the information of the knowledge graph, and the best semantic information is selected as the semantic parsing result of the entity attribute. After completing the semantic parsing of the entity decision tree entity, the entity and its corresponding semantic information are fused to obtain fused data.

[0053] Each piece of fused data is transformed into a sequence of labels. A self-attention mechanism is used to calculate the relationship weight between each label in the sequence and all other labels. The label sequence with contextual information is processed by a deep learning model to generate prediction results and corresponding interpretive outputs to provide an understanding of the decision-making process.

[0054] This invention first transforms multi-source heterogeneous data into the form of entities and their attributes through entity recognition and attribute extraction, unifying the data format of different data sources and providing a foundation for subsequent data fusion and analysis.

[0055] By matching knowledge graphs in the engineering field, this invention can map entities and attributes to nodes and attributes in the knowledge graph, thereby obtaining a structured representation of the entities.

[0056] The construction of entity decision trees combines knowledge graphs and entity attribute information, providing entities with richer semantic information and contextual environment, and improving the accuracy of parsing and fusing entity attribute values.

[0057] By using the Transformer model with a self-attention mechanism, this invention can effectively process serialized fused data, capture long-distance dependencies between labels, and extract deep-level features.

[0058] The model's generated predictions and interpretive outputs not only provide predictions for the task but also explanations of the decision-making process, thus improving the model's interpretability.

[0059] This invention is used for the fusion and analysis of multi-source heterogeneous data in the engineering field, improving the accuracy and completeness of data fusion. By utilizing structured representations of knowledge graphs and entity decision trees, this invention effectively resolves semantic ambiguity issues of entities, thus improving the accuracy of data fusion. Simultaneously, by processing multi-source heterogeneous data, this invention can obtain a more complete entity view, further enhancing the completeness of data fusion.

[0060] This invention improves the quality and interpretability of prediction results: By using a Transformer model with a self-attention mechanism, it can extract deep features from fused data and generate high-quality prediction results. Simultaneously, the interpretable output generated by the model provides an explanation of the decision-making process, thus enhancing the model's interpretability.

[0061] Accelerates the decision-making process: This invention can automatically process the fusion and analysis of multi-source heterogeneous data, greatly shortening the time for data processing and decision-making, and improving decision-making efficiency.

[0062] In summary, this invention comprehensively utilizes deep learning and knowledge graph technologies to achieve effective fusion and analysis of multi-source heterogeneous data in the engineering field, thereby improving the efficiency and effectiveness of data processing.

[0063] Preferably, the preprocessing is used to convert data from various formats and sources into a uniform form, and the preprocessing includes data cleaning, data standardization, and data transformation.

[0064] In the data preprocessing stage, the primary task is to convert data from various formats and sources into a unified form to facilitate subsequent processing and analysis. This step typically includes the following parts:

[0065] Data cleaning involves checking and handling issues such as missing values, duplicate values, and erroneous values. For text data, grammar and spell checks are performed. For numerical data, outlier detection is required.

[0066] Data standardization: The scale and distribution of data can affect the results of subsequent analysis. Therefore, in this step, the model needs to transform the data to a common scale, such as using normalization or standardization methods. In addition, for text data, the model may need to perform some preprocessing operations, such as converting to lowercase, removing stop words, stemming, and restoring part-of-speech tags, to reduce data complexity and improve processing efficiency.

[0067] Data Transformation: In this step, the model needs to transform unstructured data, such as text, images, and audio, into a structured form. For example, text data can be converted into numerical vectors using a bag-of-words model or word embedding techniques; image data can be converted into feature vectors using feature extraction techniques.

[0068] Introduce advanced preprocessing techniques: For example, for text data, techniques such as stemming, lemmatization, and stop word removal can be introduced.

[0069] The above steps are all aimed at improving data quality, making subsequent data processing and analysis steps more effective.

[0070] After preprocessing the data, the entity matching stage begins. This first requires standardizing attribute names, such as converting uppercase to lowercase and removing spaces. This process aims to reduce variations in textual expression, thereby increasing the number of potentially matching entity pairs.

[0071] Next, attribute names are compared to identify data rows that may represent the same entity. For example, if two data rows both have the attributes "name" and "address," and the values ​​of these two attributes are similar to some extent, then the two data rows can be considered to represent the same entity. This similarity determination may rely on some metric, such as cosine similarity, Jaccard similarity, or edit distance.

[0072] Then, for the identified candidate matches, the model needs to further determine whether they truly represent the same entity. This process employs more complex methods, such as calculating the semantic similarity of attribute values, or using machine learning algorithms like decision trees, random forests, support vector machines, or neural networks. These methods can consider the combined effects of multiple attributes, thereby improving the accuracy of entity matching.

[0073] To handle complex situations, such as inconsistent units or formats of attribute values, appropriate conversions or standardizations are necessary during processing. For example, for date type attributes, it may be necessary to convert all dates to the same format; for numeric type attributes, unit conversions may be required.

[0074] With the introduction of deep learning techniques, such as BERT or GPT, these models are pre-trained on large-scale corpora and learn rich semantic knowledge, which further improves the accuracy of entity matching. These models can extract deep features of attribute values ​​and capture subtle differences in word meaning and semantic relationships.

[0075] Matching results can be categorized into three states: match, non-match, and uncertain. Matching entities are marked as the same entity; non-matching entities are marked as different entities; and uncertain entities may require further review or be determined using heuristic rules.

[0076] The entity matching stage aims to identify data rows that represent the same entity by comparing attributes, so as to facilitate subsequent data fusion.

[0077] Preferably, the entity decision tree is used to organize and display entities and their attributes, wherein each node is an entity and each edge represents a certain relationship between entities;

[0078] The construction of an entity decision tree includes: grouping data labeled as the same entity into the same group, using each group of data rows to create a node in the decision tree, and each node containing all relevant data for the corresponding entity, and using the relationships between nodes to construct an edge between two nodes.

[0079] Based on the entity matching results, data rows that are marked as the same entity are grouped together. Each group of data rows will be used to create a node in the decision tree. Each node will contain all relevant data for that entity, such as the entity's attributes and attribute values.

[0080] Edges are constructed between these nodes to form a decision tree. The basis for constructing edges is usually the relationship between the nodes, such as shared attributes or similar attribute values. For example, if two nodes represent entities that both have an attribute named "Company Name" and the value of this attribute is the same, then an edge can be constructed between these two nodes.

[0081] To construct a decision tree, heuristic rules are used, such as maximizing the similarity between nodes and minimizing the differences between nodes. These rules help the model identify the most likely entity relationships, thereby constructing the most probable decision tree.

[0082] The result of constructing an entity decision tree is a visual representation of entity relationships. This representation helps users understand the structure of the data, facilitating subsequent data analysis and mining. Furthermore, entity decision trees can serve as input for subsequent steps, providing useful information such as which entities are related and which attributes are important.

[0083] By introducing deep learning techniques, such as Graph Neural Networks (GNNs), into this stage, more complex relationships and attributes can be processed, thereby further improving the quality of decision trees. GNNs can effectively process graph-structured data, learning complex patterns on nodes and edges, thus better understanding and capturing the relationships between entities.

[0084] After constructing the entity decision tree, it is necessary to perform semantic parsing of entity attributes, identification and resolution of semantic conflicts, and data fusion.

[0085] Entity attribute semantic parsing: First, semantic parsing is performed on the constructed entity decision tree. Specifically, the attribute values ​​of each entity in the decision tree are analyzed and transformed into a computational form, namely, a feature vector. Furthermore, the decision tree also helps to understand the meaning of each attribute value in its context, such as its relationship with other entities or attributes, and its position within the entire decision tree. This contributes to a more accurate understanding of the semantic information of entity attribute values.

[0086] Semantic Conflict Identification and Resolution: After semantic parsing is complete, the next step is to identify and resolve potential semantic conflicts. In this step, the role of knowledge graphs becomes crucial. Background knowledge within the knowledge graph, such as relationships between concepts and concept definitions, can help identify ambiguous or potentially conflicting entity attribute values. Then, the causes of these conflicts need to be identified, and reasonable solutions proposed. This typically requires utilizing domain knowledge and experience, as well as information from the knowledge graph.

[0087] Data fusion: After resolving semantic conflicts, data fusion can then proceed. This process involves matching entities and their attribute values ​​from the entity decision tree with concepts in the knowledge graph to form a new data representation. This representation includes not only the original entity attribute values ​​but also the semantic information of those attributes within the knowledge graph. This facilitates a better understanding and utilization of the data, providing more information and convenience for subsequent data processing and decision-making.

[0088] Entity decision trees and knowledge graphs play crucial roles in entity attribute semantic parsing, semantic conflict identification and resolution, and data fusion. Entity decision trees provide structured and contextual information about entities and their attributes, aiding in understanding and parsing the meaning of entity attribute values; while knowledge graphs offer broader background knowledge, facilitating the identification and resolution of semantic conflicts and data fusion. Combining these two technologies allows for more effective processing and understanding of multi-source, heterogeneous data in the engineering field.

[0089] Preferably, the feature vector is used to obtain semantic information of the attribute value, including: literal meaning, context information, and the relationship between words.

[0090] To understand the attribute values ​​of entities, they need to be transformed into a form that machines can understand. In this process, the model converts each entity's attribute value into a feature vector. This feature vector captures the semantic information of the attribute value, including its literal meaning, contextual information, and relationships between words.

[0091] Suppose an entity has the attribute value "coating". This word can have different meanings in different contexts. For example, in construction engineering, it might refer to interior and exterior decoration materials, while in artistic creation, it might refer to the paint used by painters. Therefore, machines need to understand the different meanings of this word in different contexts in order to perform correct semantic parsing.

[0092] In the process of generating feature vectors, pre-trained word vector models, such as Word2Vec, GloVe, or BERT, are typically used. These models are trained on large amounts of text data and can capture the semantic information of words and the relationships between them. For example, the Word2Vec model learns the vector representation of words by predicting the context of a word or the words in the context. These vector representations are in a high-dimensional space, where semantically similar words are mapped to nearby positions.

[0093] After being transformed into feature vectors, these feature vectors can be used to calculate the similarity between attribute values ​​and concepts in the knowledge graph. This provides a basis for subsequent semantic parsing and conflict identification. Furthermore, feature vectors allow for a better understanding of the different meanings of attribute values ​​in different contexts, further improving the accuracy of semantic parsing.

[0094] This feature vector is compared with the feature vectors of each concept in the knowledge graph, and the similarity between them is calculated. If the similarity between an attribute value and a concept exceeds a preset threshold, the model considers that the attribute value may have meaning for that concept.

[0095] After the feature vector is generated, the next step is to compare it with the feature vector of each concept in the knowledge graph. Each concept has a corresponding feature vector in the knowledge graph, which represents the semantic attributes and relationships of the concept. This process mainly aims to identify which concepts the attribute value is closest to, thus requiring a comparison between the feature vector of the attribute value and the feature vector of each concept in the knowledge graph.

[0096] This comparison process is typically performed by calculating the cosine similarity between two feature vectors. Cosine similarity measures the angle between two vectors in a high-dimensional space, thus reflecting their similarity. If the cosine similarity between an attribute value and a concept exceeds a preset threshold, the model considers that the attribute value to potentially represent the concept. This threshold needs to be determined experimentally and is generally set to a relatively high value to ensure the identified concepts have high credibility.

[0097] Preferred, such as Figure 2 As shown, after initially determining the semantic information of an entity, if there is only one feature vector and the similarity is greater than a preset threshold, the semantic information of the entity is directly determined; if the similarity is less than the preset threshold, the preset threshold, data annotations, and knowledge graph are comprehensively adjusted, and the semantic information of the entity is determined.

[0098] At this stage, similarity is calculated between the feature vector of each entity attribute value and the feature vector of the corresponding node in the knowledge graph. If an entity attribute value has only one best-matching (highest similarity) feature vector, then this feature vector (and its corresponding knowledge graph node) can be directly used to determine the semantic information of the entity. However, if there are at least two feature vectors, it means that the entity attribute value has multiple possible semantic interpretations, and further processing is required to resolve this semantic conflict.

[0099] When the calculated similarity is below a preset threshold, it means that the matching degree between the feature vector of the entity's attribute value and the feature vector in the knowledge graph is not high, i.e., the semantic information of the entity cannot be accurately obtained from the knowledge graph. In this case, the following processing methods can be adopted:

[0100] Model adjustment: If many entity attribute values ​​cannot match the feature vectors in the knowledge graph, then the feature extraction and comparison methods may need to be reconsidered, and the model may need to be modified or other feature vector representation methods may need to be adopted.

[0101] Threshold adjustment: If only a few entity attribute values ​​cannot match the feature vectors in the knowledge graph, threshold adjustment can be considered. This requires a trade-off between matching accuracy and coverage, and may require finding a suitable balance between the two.

[0102] Manual annotation and feedback: Entity attribute values ​​that cannot be matched can also be handled manually. For example, experts can manually add annotations to these entities or provide feedback to improve model performance.

[0103] Expanding the knowledge graph: If the knowledge graph does not cover a comprehensive range of domains and entity types, many entities may not find corresponding matches. In this case, expanding the knowledge graph to include more domains and entity types can improve the likelihood of matching.

[0104] When the similarity is lower than the preset threshold, multiple aspects such as the model, threshold, data annotation, and knowledge graph need to be comprehensively considered and adjusted to improve the model's matching performance and processing capabilities.

[0105] Preferably, weights are assigned to each feature vector by calculating the global importance and recent occurrence frequency of the feature vector corresponding to the entity attribute value in the corresponding node of the knowledge graph. The weights are used to reflect the importance of the semantic information represented by the feature vector in the entire knowledge graph and the specific context.

[0106] The best semantic information is selected as the semantic parsing result of the entity attribute based on the weight. When there is a semantic conflict, the conceptual relationship in the knowledge graph and the semantic information of other entities or concepts are used to resolve the conflict, ensuring that the most appropriate semantic information is selected.

[0107] An attribute value may have meanings from multiple concepts, leading to semantic conflicts. To address this, information from a knowledge graph, such as relationships between concepts and the importance of concepts, is used to select the most suitable concept for each attribute value. This process is accomplished by calculating the weight of concepts and the similarity between concepts and attribute values, and then selecting the concept with the highest weight and the highest similarity.

[0108] When dealing with semantic conflicts, you may encounter situations where an attribute value may correspond to multiple concepts. In such cases, you need to select the most appropriate one from these concepts.

[0109] First, consider the relationships between concepts in the knowledge graph. For example, if two concepts are directly related in the knowledge graph, they are more likely to be selected simultaneously. The model also considers the importance of concepts, which is reflected by their weights. Weights are calculated based on a concept's position in the knowledge graph and its relationships with other concepts. For example, a concept with many related concepts is likely to have a higher weight.

[0110] After considering these factors, the most suitable concept is selected for each attribute value. This process is accomplished by calculating the concept's weight and the similarity between the concept and the attribute value, and then selecting the concept with the highest weight and the highest similarity.

[0111] This approach not only resolves semantic conflicts but also ensures that the chosen concept is more aligned with reality. For example, if an attribute value might correspond to two concepts, one very common and the other rare, the model is more likely to choose the common concept because it will have a higher weight.

[0112] This process yields the most likely concept for each attribute value, which will serve as the model's output for subsequent data fusion and parsing steps.

[0113] Weight is a quantitative metric that measures the importance of a concept. In knowledge graphs, it is typically used to reflect the degree of importance of a concept within the entire knowledge graph. There are many methods for calculating weights, and the specific method may vary depending on the characteristics and requirements of the knowledge graph.

[0114] One common method for calculating weights is based on a concept's position and connections within the knowledge graph. Specifically, a concept has more connections within the knowledge graph (e.g., it has more relationships with other concepts), resulting in a higher weight. This is because, within a knowledge graph, the connections a concept has with other concepts often reflect its importance.

[0115] In addition, weights can also be calculated based on the frequency of a concept's use in real-world applications. For example, if a concept appears more frequently in actual data processing, then that concept can be considered to have a higher weight. This approach can help the model better understand and process real-world data.

[0116] The method of weight calculation may vary depending on the application scenario and specific needs. C represents the concept of a knowledge graph node, and the weight W of C in the knowledge graph can be calculated as follows:

[0117]

[0118] Wherein, Degree(C) is the degree of concept C in the knowledge graph, that is, the number of other concepts directly connected to it, which reflects the position and relevance of the concept in the knowledge graph;

[0119] MaxDegree is the degree of the concept with the highest degree in the knowledge graph, used to normalize Degree(C);

[0120] Frequency(C) is the frequency with which concept C is used in actual data processing;

[0121] MaxFrequency is the most frequently used concept of all, used to normalize Frequency(C);

[0122] α and β are hyperparameters used to adjust the influence of Degree(C) and Frequency(C) in the weight calculation. Their values ​​can be determined through methods such as cross-validation to obtain optimal model performance.

[0123] Typically, both degree and frequency of use are considered when calculating weights; that is, both metrics are used simultaneously. This is because degree reflects the importance of a concept in the knowledge graph, while frequency of use reflects its importance in actual data processing.

[0124] However, in some cases, only one metric might be considered. For example, if only the structure of the knowledge graph is of interest, and the actual data processing is irrelevant, then only the degree might be considered. Conversely, if only the actual data processing is of interest, and the structure of the knowledge graph is irrelevant, then only the frequency of use might be considered.

[0125] This formula is just one possible way to calculate weights; the specific calculation method may need to be adjusted based on the actual knowledge graph and task requirements.

[0126] PageRank Algorithm: If a greater focus is placed on the global importance of concepts within the knowledge graph, the PageRank algorithm might be used. PageRank is a link analysis algorithm that assigns a global importance score to each concept in the knowledge graph. Assuming a PageRank value of PR(C), the weight of concept C might be defined as follows:

[0127]

[0128] Among them, MaxPR is the maximum PageRank value of all concepts, used to normalize PR(C); γ is a hyperparameter used to adjust the influence of PR(C) in the weight calculation.

[0129] Weight calculation based on time decay: If the data contains time information, a time decay method may be used to calculate the weights. Let T(C) be the time of the most recent occurrence of concept C, T... now For the current time, the weight of concept C can be defined as:

[0130] W(C)=δ*exp(-λ*(T now -T(C)))

[0131] Where λ is the decay factor, used to control the decay rate of the weights over time; δ is a hyperparameter used to adjust the influence of time decay in the weight calculation; this formula assumes that recently emerging concepts are more important.

[0132] In reality, the importance of a concept can be influenced by multiple factors simultaneously. For example, a concept's global importance in a knowledge graph (PageRank value) and its frequency of occurrence in recent data (time decay factor) can both affect its weight. Therefore, it is necessary to combine these two factors to calculate a comprehensive weight, resulting in the weight calculation expression in this invention.

[0133] Preferably, the expression for calculating the weights is:

[0134]

[0135] Where C is a concept of a knowledge graph node, W(C) is the weight of the concept; α and β are hyperparameters used to adjust the relative importance of importance values ​​and time decay factors in weight calculation; PR(C) is the importance value of concept C, MaxPR is the maximum importance value of all concepts; T(C) is the time of the most recent occurrence of concept C, Tnow λ is the current time, and λ is the decay factor.

[0136] By adjusting the values ​​of α and β, the influence of global importance and recent occurrence frequency can be balanced across different tasks and datasets. For example, if global importance is considered more important than recent occurrence frequency, α can be set larger than β, and vice versa.

[0137] The weight of each feature vector (i.e., the semantic information of an entity attribute) is obtained by calculating the global importance (i.e., the centrality of the node in the entire graph, which may be measured by PageRank or other graph algorithms) and recent occurrence frequency (reflecting the importance of the concept in a specific context) of the corresponding node in the knowledge graph.

[0138] In practice, the purpose of calculating weights is to score the possible semantics (represented by feature vectors) of each entity attribute. Selecting the feature vector with the highest weight means selecting the most likely semantics of that entity attribute. This semantics will be used in subsequent steps, such as building entity decision trees and data fusion.

[0139] In this invention, when semantic conflicts arise, information from a knowledge graph is used to resolve these conflicts. The specific steps include the following:

[0140] Conflict identification: First, it is necessary to identify entity attribute values ​​that have semantic conflicts. Semantic conflicts may arise from multiple possible semantic parsing results, or from inconsistencies between the semantic parsing results corresponding to different entity attribute values.

[0141] Collect relevant information: After identifying conflicting entity attribute values, gather all information related to these attribute values. This includes their own semantic parsing results, as well as all information in the knowledge graph related to these results. This information may include information about other entities or concepts directly related to them, as well as global information in the knowledge graph, such as the hierarchical structure between concepts and the relationships between concepts.

[0142] Conflict Resolution: With this information, you can attempt to resolve the conflicts. The specific resolution method may depend on the specific conflict situation and domain knowledge. Possible strategies include selecting the semantic parsing result with the highest weight, judging which result is more reasonable based on the conceptual relationships in the knowledge graph, or combining semantic information from other entities or concepts to determine which result is more consistent.

[0143] Updated results: Finally, based on the results of conflict resolution, the semantic parsing results of entity attribute values ​​and the related fusion data are updated.

[0144] It is important to note that this process may need to be repeated until all semantic conflicts are resolved. Furthermore, the conflict resolution process may affect other entity attribute values ​​and fused data, thus requiring corresponding updates to this data.

[0145] At the same time, it is also necessary to consider that entity attributes may have multiple semantics, and these semantics may conflict. In this case, it is necessary to use information in the knowledge graph, such as the relationships between concepts, as well as the semantic information of other entities and concepts, to resolve these conflicts and ensure that the most appropriate and meaningful semantic is selected.

[0146] The steps outlined above enable effective data parsing and understanding, as well as conflict identification and resolution. Furthermore, utilizing engineering domain knowledge graphs can help models understand specific knowledge and rules within the engineering domain, thereby improving their ability to process engineering-related data.

[0147] In this invention, a self-interpretive Transformer model is designed, which can explain its decision-making process, thereby increasing the model's interpretability. The working process of the model will be described in detail below:

[0148] Input and Encoding: The Transformer model receives the fused data obtained from the above processing steps. Each data point is transformed into a sequence of tokens. Each token is an encoding composed of any combination of words, subwords, and characters. These tokens are then transformed into a set of high-dimensional vectors called "embedded vectors." Embedded vectors can capture and represent the semantic information of the tokens, such as word meaning, word order, and contextual relationships.

[0149] When the Transformer model receives the fused data obtained from the previous processing steps, its main task is to encode the data. The encoding process can be divided into two main steps: tokenization and embedding.

[0150] Tokenization: First, the model tokenizes the input data, that is, it divides the text data into tokens. The specific token unit is chosen based on the application requirements and the characteristics of the data. For example, if the data is English text, a token is usually a word or a subword. If the data is Chinese text, since Chinese does not have explicit word separators, a token is usually a character or a word. The purpose of tokenization is to divide the text data into units that can be processed by the model.

[0151] Embedding: The model then transforms these labels into high-dimensional vectors called embedding vectors. This process is typically achieved using a parametric matrix called the embedding matrix. Each label corresponds to a row in the embedding matrix, and the vector in that row is the embedding vector for that label. The dimension of the embedding vectors is a configurable parameter, typically ranging from several hundred to several thousand dimensions. This embedding matrix is ​​one of the parameters that the model needs to learn during training.

[0152] Embedding vectors are the model's internal representation of a token, capturing and representing its semantic information. For example, similar tokens (semantically or syntactically similar) should have similar embedding vectors. Furthermore, embedding vectors can also capture and represent the token's contextual information, such as word order and contextual relationships, through the model's self-attention mechanism.

[0153] The core of the Transformer model is the "self-attention" mechanism, which captures the relationships between various labels in a sequence. Specifically, for each label in the sequence, the model calculates its relational weights with all other labels in the sequence. These weights reflect the degree of influence or relevance of other labels on the current label. In this way, the model can understand the contextual information in the sequence.

[0154] The self-attention mechanism is a core component of the Transformer model, and its main function is to understand and capture contextual information in a sequence. Specifically, the self-attention mechanism models the context of each tag in the sequence, which is obtained by calculating the relational weights between the current tag and all other tags in the sequence.

[0155] Preferred, such as Figure 3 As shown, the method for calculating the relationship weights between each label and all other labels in a sequence using a self-attention mechanism includes the following steps:

[0156] Query, key, and value: For each token in the sequence, a query vector, a key vector, and a value vector are generated.

[0157] Weight score calculation: A raw weight score is obtained by calculating the dot product between the query vector of each tag and the key vectors of all tags. This score can be understood as the degree of association between the current tag and other tags.

[0158] Relationship weight calculation: Calculate the similarity of each tag with all other tags, and use the softmax function to convert the similarity into relationship weights. This ensures the relativity of relationship weights, meaning that tags with higher weights are more important and need more attention.

[0159] Weighted sum of values: Multiply the value vector of each tag by its corresponding weight score, and then sum these weighted value vectors to obtain the context representation of the current tag;

[0160] The marker is an encoding composed of any combination of words, subwords, and characters;

[0161] The context refers to the information used to capture long-distance dependencies in a sequence.

[0162] The purpose of the above steps is to enable the model to represent each label while incorporating its contextual information, that is, to consider the influence of other labels in the sequence on the current label. In this way, the generated contextual representation can capture long-range dependencies in the sequence, allowing the model to better understand and process sequence data.

[0163] Through the steps described above, the self-attention mechanism can capture contextual information within a sequence. This contextual information reflects the relationships between various tokens in the sequence, such as syntactic structure and semantic connections. This contextual information is crucial for understanding and generating text.

[0164] The query, key, and value vector for each tag are generated independently, meaning the model can process all tags in the sequence in parallel. This is a significant advantage of the Transformer model compared to other sequence processing models (such as RNNs).

[0165] The Transformer model is a deep learning model composed of multiple stacked "encoders" or "decoders." Each encoder / decoder contains a self-attention mechanism and a fully connected neural network. Data is processed by the self-attention mechanism and the fully connected neural network at each layer, and the result of each layer's processing serves as the input to the next layer. Through this stacked structure, the model can learn deep features and complex patterns in the data.

[0166] The Transformer model employs a layered structure of deep learning, primarily composed of multiple encoders or decoders. Each encoder or decoder includes two main parts: a self-attention mechanism and a fully connected neural network.

[0167] Self-attention mechanism: As mentioned above, the self-attention mechanism can capture the relationships between various tokens in a sequence, which is crucial for understanding and generating text. Specifically, the self-attention mechanism compares the embedding vector of each token with the embedding vectors of all other tokens in the sequence and calculates the weight of the relationship between them. This weight can be understood as the importance or relevance of other tokens to the current token.

[0168] Fully connected neural networks (also known as feedforward neural networks) are a very basic neural network structure consisting of multiple fully connected layers. In the Transformer model, the fully connected neural network receives the output of the self-attention mechanism as input, and then extracts deep features from the input data through a series of linear transformations and non-linear activation functions.

[0169] In the Transformer model, data is processed sequentially through the self-attention mechanism and fully connected neural network at each layer. The output of each layer serves as the input to the next. This layered structure allows the model to extract and abstract features from the data layer by layer, thereby learning deep patterns and complex relationships within the data. This is the main advantage of deep learning models compared to shallow learning models.

[0170] It is particularly important to note that although each layer contains a self-attention mechanism and a fully connected neural network, the parameters of each layer are independent. This means that each layer learns different features and patterns. This is one of the key reasons why deep learning models can learn complex patterns.

[0171] Preferred, such as Figure 4 As shown, the process of using a deep learning model (such as Transformer) to process the labeled sequence with contextual information and generate prediction results and corresponding interpretive outputs includes:

[0172] S1. Take the sequence of tags containing their own contextual information as input. These sequences have been processed by the self-attention mechanism so that each tag contains its own contextual information.

[0173] S2. The labeled sequence is processed through at least two layers of a deep learning model. In each layer, the sequence goes through two main sub-modules: a self-attention mechanism and a fully connected neural network. The role of these two sub-modules is to perform feature extraction and feature transformation, so that the model can learn deep features and complex patterns from the input sequence. The processing result of each layer will be used as the input of the next layer.

[0174] S3. Generate a new label sequence, which contains the prediction results for each label. After processing through the last layer of the model, a new label sequence is obtained, which contains the model's prediction results for each label. The prediction results are usually transformed through a special output layer (such as a softmax layer or a linear layer) so that they can be interpreted as the required output type, such as a probability distribution or continuous values.

[0175] S4. Generate explanatory outputs corresponding to the prediction results. Explanatory outputs can help understand the model's decision-making process, such as which input labels have an important impact on the prediction results and the relationships between them. This information is usually extracted through some internal states in the model (such as attention weights).

[0176] In the steps described above, deep learning models, through their layered structure and self-attention mechanism, are able to capture deep-seated features and long-range dependencies in sequence data, thereby generating high-quality prediction results and interpretive outputs.

[0177] After processing by stacked encoders / decoders, an output sequence is generated, which can be represented as a set of predictions or decision results. In the self-interpretive Transformer model, in addition to these predictions, the model also generates a set of interpretive outputs. These outputs help to understand the model's decision-making process, such as which input labels have a significant impact on the decision results, and the relationships between them.

[0178] After being processed by multiple layers of encoders and decoders, the Transformer model generates an output sequence. This sequence can be interpreted as a set of predictions or decision results. These results can be used for various tasks, such as text translation (the output sequence represents the translated sentence), sentiment analysis (the output sequence represents the probability of various sentiments), classification tasks (the output sequence corresponds to the probability of each category), and so on.

[0179] In self-interpretive Transformer models, in addition to the predictions, a portion of the output is specifically used to generate explanations of the model. These explanations help to understand the model's decision-making process; that is, they help to understand how the model derives its predictions from the input data.

[0180] For self-interpretive outputs, the most important component is likely the attention weights. In a self-attention mechanism, the model calculates the attention weight for each input label to each output label. These weights can be viewed as the degree of influence of the input labels on the output labels. By analyzing these weights, we can understand which input labels have a significant impact on the decision outcome and the relationships between these labels.

[0181] Of course, self-interpretive Transformer models can also be used for problems in the engineering field. For example, in architectural design projects, the model may need to predict whether a particular structural design will meet the required performance standards. In this case, the output sequence might be a probability value representing the likelihood that the design will meet the performance standards.

[0182] Furthermore, the self-explanatory output helps in understanding the reasons behind the prediction. By analyzing the attention weights, we can see how the model determines the prediction from the input design parameters and environmental conditions. We can see which design parameters and environmental conditions have a significant impact on the prediction, and the relationships between these factors.

[0183] For example, if a particular design parameter (such as the building's height) has a particularly high attention weight, it can be inferred that this parameter plays a crucial role in determining whether the building meets performance standards. If two design parameters (such as the building's height and width) both have relatively high attention weights and they have a strong interrelationship in the self-attention mechanism, it can be inferred that these two parameters may need to be considered together to correctly evaluate the building's performance.

[0184] Such explanatory outputs can help engineers better understand and trust the model's decision-making process, thereby increasing their confidence and efficiency in using the model. At the same time, this can also help engineers identify potential design problems, improving the quality and efficiency of their designs.

[0185] The self-explanatory Transformer model enables users to understand and trust its decision-making process through self-explanation, which is crucial for engineering applications involving the fusion of multi-source heterogeneous data.

[0186] The self-interpretive Transformer model in this invention, after receiving and processing multi-source heterogeneous data (such as CAD design data, material properties, mechanical performance data, production parameters, etc.), will output prediction results and interpretive output.

[0187] Regarding the prediction results, the output format depends on the specific engineering task. For example, in a complex structural design task, the model might need to predict the strength of a structure under specific conditions. In this case, the prediction result might be a predicted value of the structural strength. If the task is a classification task, such as predicting whether a material is suitable for a certain application, then the output result might be a probability distribution for each category.

[0188] In terms of interpretable output, it helps in understanding the model's decision-making process, i.e., which information in multi-source heterogeneous data influenced the final prediction results, and which features played a decisive role in the prediction. For engineering tasks, this can aid in understanding, for example, whether design parameters, material properties, or other production parameters had a dominant influence on the prediction results when designing a complex structure.

[0189] In the post-processing stage, interpretable outputs need to be fed back to the engineering team. This helps them understand how the model extracts and utilizes information from multi-source heterogeneous data to improve model performance, or optimize engineering designs based on the interpretable outputs. For example, if the interpretable outputs show that a certain design parameter has a significant impact on the prediction results, engineers may need to consider this parameter more in subsequent designs.

[0190] This data output and post-processing approach can maximize the fusion and analysis of multi-source heterogeneous data, providing more comprehensive and accurate solutions to engineering problems. It also makes the model's decision-making process more transparent, which is conducive to improving the efficiency and accuracy of engineering decisions.

[0191] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0192] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0193] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0194] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0195] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0196] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0197] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0198] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0199] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for multi-source heterogeneous data fusion and analysis in the field of engineering, characterized in that, Includes the following steps: Multi-source heterogeneous data in the engineering field are preprocessed, and entity matching is performed on the preprocessed data based on attributes. Data marked as the same entity are grouped together. Each group of data rows is used to create an entity node in the entity decision tree. Each entity node contains the attributes and attribute values ​​of the corresponding entity. Edges between entity nodes are constructed using shared attributes or similar attribute values ​​between entity nodes to obtain an entity decision tree that provides structured and contextual information about entities and their attributes. The attribute value of each entity is transformed into a feature vector, and similarity is calculated with the feature vector corresponding to the node in the knowledge graph to initially determine the semantic information of the entity. When there are at least two feature vectors, the weight of each feature vector is calculated based on the information of the knowledge graph, and the best semantic information is selected as the semantic parsing result of the entity attribute. After completing the semantic parsing of the entity decision tree entity, the entity and its corresponding semantic information are fused to obtain fused data. The feature vector is used to obtain semantic information of the attribute value, including: literal meaning, context information, and the relationship between words; By calculating the global importance and recent occurrence frequency of the feature vector corresponding to the entity attribute value in the corresponding node of the knowledge graph, a weight is assigned to each feature vector. The weight is used to reflect the importance of the semantic information represented by the feature vector in the entire knowledge graph and the specific context. The best semantic information is selected as the semantic parsing result of the entity attribute based on the weight. When there is a semantic conflict, the conceptual relationship in the knowledge graph and the semantic information of other entities or concepts are used to resolve the conflict and ensure that the most appropriate semantic information is selected. The expression for calculating the weight is: in, It refers to the concept of knowledge graph nodes. It is the weight of the concept; and These are hyperparameters used to adjust the relative importance of importance values ​​and time decay factors in weight calculation; It is a concept Importance value, It is the highest importance value among all concepts; It is a concept The most recent time of occurrence, It is the current time. It is the attenuation factor; Each fused data point is transformed into a sequence of labels. A self-attention mechanism is used to calculate the relationship weight between each label in the sequence and all other labels. The label sequence with contextual information is processed by a deep learning model to generate prediction results and corresponding interpretive outputs. The interpretive outputs are used to represent the design parameters, material properties, or production parameters in multi-source heterogeneous data in the engineering field that affect the engineering prediction results.

2. The method for multi-source heterogeneous data fusion and analysis in the engineering field according to claim 1, characterized in that, The preprocessing is used to convert data from various formats and sources into a unified form. Preprocessing includes data cleaning, data standardization, and data transformation.

3. The method for multi-source heterogeneous data fusion and analysis in the engineering field according to claim 1, characterized in that, The entity decision tree is used to organize and display entities and their attributes, where each node is an entity and each edge represents a relationship between entities; The construction of an entity decision tree includes: grouping data labeled as the same entity into the same group, using each group of data rows to create a node in the decision tree, and each node containing all relevant data for the corresponding entity, and using the relationships between nodes to construct an edge between two nodes.

4. The method for multi-source heterogeneous data fusion and analysis in the engineering field according to claim 1, characterized in that, After initially determining the semantic information of an entity, if there is only one feature vector and the similarity is greater than a preset threshold, the semantic information of the entity is directly determined; if the similarity is less than the preset threshold, the preset threshold, data annotations, and knowledge graph are comprehensively adjusted, and the semantic information of the entity is determined.

5. The method for multi-source heterogeneous data fusion and analysis in the engineering field according to claim 1, characterized in that, The method of calculating the relationship weights between each label in the sequence and all other labels using a self-attention mechanism includes the following steps: Query, key, and value: For each token in the sequence, generate: query vector, key vector, and value vector; Weight score calculation: Calculate the dot product between the query vector of each tag and the key vectors of all tags to obtain a raw weight score; Relationship weight calculation: Calculate the similarity of each tag to all other tags, and convert the similarity into relationship weights using the softmax function; Weighted sum of values: Multiply the value vector of each tag by its corresponding weight score, and then sum these weighted value vectors to obtain the context representation of the current tag; The marker is an encoding composed of any combination of words, subwords, and characters; The context refers to the information used to capture long-distance dependencies in a sequence.

6. The method for multi-source heterogeneous data fusion and analysis in the engineering field according to claim 1, characterized in that, The process of using a deep learning model to process the labeled sequence with contextual information to generate prediction results and corresponding interpretive outputs includes: S1. Take the tag sequence containing its own context information as input; S2. The labeled sequence is processed through at least two layers of a deep learning model; S3. Generate a new label sequence, which contains the prediction results for each label; S4. Generate explanatory output corresponding to the prediction results.

Images