An artificial intelligence-based power distribution network construction standard text matching method and system

Abstract

This invention provides an artificial intelligence-based method and system for text matching of power distribution network construction standards. The method includes: extracting semantic feature vectors from the text data of power distribution network construction standards using a gated attention mechanism; extracting image feature vectors from the image data of power distribution network construction standards using a dynamic feature pyramid network; concatenating the image feature vectors and semantic feature vectors to obtain initial joint features; using an uncertainty measurement mechanism and differentiable logic rules, performing probabilistic knowledge query and logical reasoning enhancement on the initial joint features based on a power distribution network primitive knowledge graph to obtain target joint features; calculating the semantic and procedural relevance between the target joint features and standard clauses in the construction process graph to generate construction standard matching results. The matching results of this invention not only conform to text semantics and image features but also strictly fit the construction stage, effectively filtering out stage-irrelevant interference items and significantly improving the accuracy and practicality of standard matching.

Description

Technical Field

[0001] This invention relates to the field of power digitalization technology, specifically to an artificial intelligence-based method and system for matching standard texts in power distribution network construction. Background Technology

[0002] The standardization and safety of power distribution network construction highly depend on the strict adherence to various standards, regulations, and design drawings. With the increasing scale and complexity of power grid construction, relying solely on manual memorization and consulting massive amounts of paper or discrete electronic construction standard documents is insufficient to meet the demands of efficient and precise power distribution network construction. Therefore, the digital matching and intelligent application of construction standards have become an industry trend, aiming to solidify standard knowledge through technological means, providing on-site personnel with immediate and accurate decision support, thereby improving construction efficiency and quality.

[0003] Existing technologies typically employ general natural language processing models and image recognition techniques to process text and drawings. This enables the rapid localization of basic visual elements in drawings (such as text boxes, lines, and basic shapes) and the batch reading of text content, providing a data foundation for subsequent automated processing and reducing the workload that relies entirely on manual reading.

[0004] However, these models lack targeted optimization for power distribution network terminology, symbols, and primitives, resulting in inaccurate feature extraction and consequently inaccurate matching results with the text. Furthermore, existing matching methods often rely on simple semantic vector similarity calculations, failing to incorporate the process logic of construction activities. This often results in semantically related but irrelevant interference items in the matching results, significantly reducing their practicality.

[0005] Therefore, we urgently need a standard text matching scheme that can accurately extract image and text features and deeply integrate construction processes to solve the problems of inaccurate feature extraction and disconnect between matching results and business processes in existing technologies. Summary of the Invention

[0006] To overcome the shortcomings of the existing technology, this invention proposes an artificial intelligence-based method for matching standard texts in power distribution network construction, comprising: A gating attention mechanism is used to extract semantic features from the text data of power distribution network construction standards to obtain semantic feature vectors. A dynamic feature pyramid network is used to perform adaptive weighted fusion processing on the image data of power distribution network construction standards to obtain image feature vectors. The image feature vector and semantic feature vector are concatenated to obtain the initial joint feature; using the uncertainty measurement mechanism and differentiable logic rules, based on the pre-constructed distribution network primitive knowledge graph, the initial joint feature is enhanced by probabilistic knowledge query and logical reasoning to obtain the target joint feature with confidence. Based on the target joint features, the semantic and procedural relevance between the target joint features and the standard clauses in the construction process diagram is calculated in parallel to generate construction standard matching results.

[0007] Optionally, the step of utilizing uncertainty measurement mechanisms and differentiable logic rules, based on a pre-constructed distribution network primitive knowledge graph, to perform probabilistic knowledge querying and logical reasoning enhancement on the initial joint features, yielding target joint features with confidence levels, including: The initial joint features are semantically similar to the pre-constructed distribution network primitive knowledge graph to obtain the association probability between the initial joint features and each node in the distribution network primitive knowledge graph. The association probabilities of all the nodes are aggregated to obtain the probability association vector of the distribution network primitive knowledge graph. The probability association vector is propagated and its features are aggregated through multi-hop message propagation using a graph propagation mechanism to obtain the node feature vector of the distribution network primitive knowledge graph. The node feature vector is then semantically matched and logically reasoned using differentiable logic rules to generate an enhancement signal for correcting the initial joint features. The activation intensity vector of each activated logic rule in the differentiable logic rules is also recorded. By using a lightweight regression network combined with an uncertainty measurement mechanism, the entropy value of the probabilistic correlation vector and the activation intensity vector of the logical rule are nonlinearly fused to obtain the confidence level of the enhanced signal. The enhanced signal is adaptively weighted using the confidence level, and the initial joint features are semantically corrected and relationally supplemented based on the weighted enhanced signal to obtain the target joint features with confidence level.

[0008] Optionally, the step of using a lightweight regression network combined with an uncertainty measurement mechanism to nonlinearly fuse the entropy value of the probabilistic association vector and the activation intensity vector of the logical rule to obtain the confidence of the enhanced signal includes: Based on the uncertainty measurement mechanism, the statistical characteristics of the activation intensity vector of the logical rule are calculated, and the statistical characteristics are the mean or variance of the activation intensity vector. The entropy value of the probabilistic association vector is concatenated with the statistical features of the activation intensity vector to obtain the uncertainty description vector of the initial joint features; The uncertainty description vector is transformed and fused using the nonlinear activation function of a lightweight regression network to obtain an implicit feature vector. The implicit feature vector is then mapped through a normalized probability transformation layer to obtain the confidence level of the enhanced signal.

[0009] Optionally, the step of using a graph propagation mechanism to perform multi-hop message propagation and feature aggregation on the probabilistic association vector to obtain the node feature vector of the distribution network primitive knowledge graph, and using differentiable logic rules to perform semantic matching and logical reasoning on the node feature vector to generate an enhancement signal for correcting the initial joint features, includes: Based on the probabilistic association vector, the graph propagation mechanism of the multi-head graph attention network is used to perform multi-hop message propagation and feature aggregation on the topology of the power distribution network primitive knowledge graph to obtain the node representation vector of the power distribution network primitive knowledge graph. The semantic similarity is obtained by semantically matching the node representation vector with the preconditions of each rule in the differentiable logic rules. From the differentiable logic rules, select multiple rules with semantic similarity greater than the similarity threshold as applicable logic rules; By using logical operators to perform forward reasoning on multiple applicable logical rules, a unified enhancement signal is generated for correcting the initial joint features.

[0010] Optionally, the pre-construction process of the differentiable logic rules is as follows: Using dependency parsing and a domain keyword database, we performed syntactic structure analysis and core extraction on the standard and specification texts for power distribution network construction to obtain preliminary corpus. Using a pre-trained language model, multiple sets of construction conditions, operational behaviors, and constraint parameters in the preliminary corpus are identified, and multiple structured triples based on construction conditions, operational behaviors, and constraint parameters are constructed. Using a production-based rule framework, each structured triple is converted into a predicate logic expression, and the predicate logic expressions are aggregated to obtain a logic rule library; The discrete operators in the predicate logic expression are replaced with continuously differentiable functions using fuzzy logic functions, and the continuously differentiable functions are encapsulated into callable computing units. The callable computing units are then fused with the vector space of the distribution network element knowledge graph to obtain a rule computing unit library. By using a lightweight pre-training strategy combined with historical power distribution network construction standard data, the link weights in the rule calculation unit library are subjected to supervised fine-tuning to obtain differentiable logic rules.

[0011] Optionally, the step of calculating the semantic and procedural relevance between the target joint features and standard clauses in the construction process diagram in parallel, based on the target joint features, to generate construction standard matching results, includes: Input the joint features of the target into the construction stage classification model to generate the current construction stage and the construction stage context representation; Based on the construction stage context representation, the process nodes of all standard clauses in the construction process graph are traversed and queried to obtain the process relevance of each standard clause to the current construction stage; and the semantic similarity between the target joint feature and the feature of each standard clause in the construction process graph is calculated. After adjusting the process relevance and semantic similarity of each standard clause based on the confidence level, a weighted fusion is performed to obtain the comprehensive matching degree of each standard clause. Based on the comprehensive matching degree of each standard clause, the construction standard matching results are selected from the construction process diagram.

[0012] Optionally, the step of using a dynamic feature pyramid network to adaptively weight and fuse image data of power distribution network construction standards to obtain image feature vectors includes: Using a deep convolutional neural network, the image data of power distribution network construction standards are processed by convolution and pooling layer by layer to obtain multi-level feature maps with decreasing resolution, forming the initial feature pyramid. The overall semantic information of the multi-level feature maps is obtained by global average pooling. A dynamic weight generation network is used to generate a set of dynamic weight coefficients for fusing features at each level based on the overall semantic information. Using a weighted fusion algorithm, based on the dynamic weight coefficients, adaptive weighted fusion processing is performed on the multi-level feature maps to obtain the image feature vector of the image data.

[0013] Based on the same inventive concept, this invention proposes an artificial intelligence-based standard text matching system for power distribution network construction, comprising: The feature extraction module is used to extract semantic features from the text data of power distribution network construction standards using a gated attention mechanism to obtain semantic feature vectors, and to perform adaptive weighted fusion processing on the image data of power distribution network construction standards using a dynamic feature pyramid network to obtain image feature vectors. The feature analysis module is used to concatenate the image feature vector and the semantic feature vector to obtain the initial joint feature; using the uncertainty measurement mechanism and differentiable logic rules, based on the pre-constructed distribution network primitive knowledge graph, the initial joint feature is subjected to probabilistic knowledge query and logical reasoning enhancement to obtain the target joint feature with confidence. The standard matching module is used to calculate the semantic and procedural relevance between the target joint features and the standard clauses in the construction process diagram in parallel based on the target joint features, and generate construction standard matching results.

[0014] Optionally, the feature analysis module includes: The probability association unit is used to perform semantic similarity matching between the initial joint features and the pre-constructed distribution network primitive knowledge graph to obtain the association probability between the initial joint features and each node in the distribution network primitive knowledge graph, and to collect the association probabilities of all the nodes to obtain the probability association vector of the distribution network primitive knowledge graph. The logic reasoning unit is used to perform multi-hop message propagation and feature aggregation on the probability association vector using the graph propagation mechanism to obtain the node feature vector of the distribution network graph element knowledge graph, perform semantic matching and logical reasoning on the node feature vector using differentiable logic rules, generate an enhancement signal for correcting the initial joint features, and record the activation intensity vector of each activated logic rule in the differentiable logic rules. The fusion unit is used to perform nonlinear fusion of the entropy value of the probabilistic correlation vector and the activation intensity vector of the logical rule using a lightweight regression network combined with an uncertainty measurement mechanism to obtain the confidence level of the enhanced signal. An adaptive correction unit is used to adaptively weight the enhanced signal using the confidence level, and to perform semantic correction and relational supplementation on the initial joint features based on the weighted enhanced signal to obtain target joint features with confidence level.

[0015] Optionally, the fusion unit is specifically used for: Based on the uncertainty measurement mechanism, the statistical characteristics of the activation intensity vector of the logical rule are calculated, and the statistical characteristics are the mean or variance of the activation intensity vector. The entropy value of the probabilistic association vector is concatenated with the statistical features of the activation intensity vector to obtain the uncertainty description vector of the initial joint features; The uncertainty description vector is transformed and fused using the nonlinear activation function of a lightweight regression network to obtain an implicit feature vector. The implicit feature vector is then mapped through a normalized probability transformation layer to obtain the confidence level of the enhanced signal.

[0016] Optionally, the logical reasoning unit is specifically used for: Based on the probabilistic association vector, the graph propagation mechanism of the multi-head graph attention network is used to perform multi-hop message propagation and feature aggregation on the topology of the power distribution network primitive knowledge graph to obtain the node representation vector of the power distribution network primitive knowledge graph. The semantic similarity is obtained by semantically matching the node representation vector with the preconditions of each rule in the differentiable logic rules. From the differentiable logic rules, select multiple rules with semantic similarity greater than the similarity threshold as applicable logic rules; By using logical operators to perform forward reasoning on multiple applicable logical rules, a unified enhancement signal is generated for correcting the initial joint features.

[0017] Optionally, the system further includes a rule building module for: Using dependency parsing and a domain keyword database, we performed syntactic structure analysis and core extraction on the standard and specification texts for power distribution network construction to obtain preliminary corpus. Using a pre-trained language model, multiple sets of construction conditions, operational behaviors, and constraint parameters in the preliminary corpus are identified, and multiple structured triples based on construction conditions, operational behaviors, and constraint parameters are constructed. Using a production-based rule framework, each structured triple is converted into a predicate logic expression, and the predicate logic expressions are aggregated to obtain a logic rule library; The discrete operators in the predicate logic expression are replaced with continuously differentiable functions using fuzzy logic functions, and the continuously differentiable functions are encapsulated into callable computing units. The callable computing units are then fused with the vector space of the distribution network element knowledge graph to obtain a rule computing unit library. By using a lightweight pre-training strategy combined with historical power distribution network construction standard data, the link weights in the rule calculation unit library are subjected to supervised fine-tuning to obtain differentiable logic rules.

[0018] Optionally, the standard matching module is specifically used for: Input the joint features of the target into the construction stage classification model to generate the current construction stage and the construction stage context representation; Based on the construction stage context representation, the process nodes of all standard clauses in the construction process graph are traversed and queried to obtain the process relevance of each standard clause to the current construction stage; and the semantic similarity between the target joint feature and the feature of each standard clause in the construction process graph is calculated. After adjusting the process relevance and semantic similarity of each standard clause based on the confidence level, a weighted fusion is performed to obtain the comprehensive matching degree of each standard clause. Based on the comprehensive matching degree of each standard clause, the construction standard matching results are selected from the construction process diagram.

[0019] Optionally, the feature extraction module is specifically used for: Using a deep convolutional neural network, the image data of power distribution network construction standards are processed by convolution and pooling layer by layer to obtain multi-level feature maps with decreasing resolution, forming the initial feature pyramid. The overall semantic information of the multi-level feature maps is obtained by global average pooling. A dynamic weight generation network is used to generate a set of dynamic weight coefficients for fusing features at each level based on the overall semantic information. Using a weighted fusion algorithm, based on the dynamic weight coefficients, adaptive weighted fusion processing is performed on the multi-level feature maps to obtain the image feature vector of the image data.

[0020] In another aspect, this application also provides an electronic device, comprising: at least one processor and a memory; the memory and the processor are connected via a bus; The memory is used to store one or more programs; When the one or more programs are executed by the at least one processor, an artificial intelligence-based standard text matching method for power distribution network construction is implemented as described above.

[0021] Furthermore, this application also provides a computer-readable storage medium having an executable program stored thereon, which, when executed, implements the above-described method for matching standard texts for power distribution network construction based on artificial intelligence.

[0022] Compared with the closest existing technology, the present invention has the following beneficial effects: This invention provides an artificial intelligence-based method and system for matching textual data of power distribution network construction standards. The method includes: extracting semantic features from the textual data of power distribution network construction standards using a gated attention mechanism to obtain a semantic feature vector; and using a dynamic feature pyramid network to adaptively weight and fuse the image data of the power distribution network construction standards to obtain an image feature vector; concatenating the image feature vector and the semantic feature vector to obtain an initial joint feature; using an uncertainty measurement mechanism and differentiable logic rules, and based on a pre-constructed power distribution network primitive knowledge graph, performing probabilistic knowledge query and logical reasoning enhancement on the initial joint feature to obtain a target joint feature with confidence; and, based on the target joint feature, parallelly calculating the target joint feature and the standard clauses in the construction event graph. This invention generates construction standard matching results by analyzing the semantic and process relevance of the data. It achieves accurate extraction of image and text features through a dynamic feature pyramid network and gated attention mechanism, solving the fundamental problem of inaccurate feature extraction in general models. Furthermore, by fusing neural symbolic reasoning with primitive knowledge graphs, image and text features are enhanced into engineering semantic representations with confidence, ensuring accurate understanding of professional terms and symbols. Based on this, a construction process graph is introduced for parallel semantic and process relevance calculation, ensuring that the matching results conform to both text semantics and image features while strictly matching the current construction stage, effectively filtering out stage-irrelevant interference. Ultimately, a complete technical closed loop is formed, from accurate perception and semantic understanding to process-oriented decision-making, significantly improving the accuracy and practicality of standard matching. Attached Figure Description

[0023] Figure 1 A flowchart illustrating an artificial intelligence-based method for matching standard texts in power distribution network construction, provided by this invention. Figure 2 A schematic diagram of the structure of a power distribution network construction standard text matching system based on artificial intelligence provided by the present invention; Figure 3 This is a schematic diagram of the structure of an electronic device provided by the present invention. Detailed Implementation

[0024] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0025] Example 1 This invention provides an artificial intelligence-based method for matching standard texts in power distribution network construction, such as... Figure 1 As shown, it includes: S1. Semantic feature vectors are obtained by extracting semantic features from the text data of power distribution network construction standards using a gated attention mechanism, and image feature vectors are obtained by adaptive weighted fusion processing of the image data of power distribution network construction standards using a dynamic feature pyramid network. S2. The image feature vector and semantic feature vector are concatenated to obtain the initial joint feature; using the uncertainty measurement mechanism and differentiable logic rules, based on the pre-constructed distribution network primitive knowledge graph, the initial joint feature is subjected to probabilistic knowledge query and logical reasoning enhancement to obtain the target joint feature with confidence. S3. Based on the target joint features, calculate in parallel the semantic and procedural relevance between the target joint features and the standard clauses in the construction process diagram, and generate construction standard matching results.

[0026] In step S1, S1-1, semantic feature vectors are obtained by extracting semantic features from the text data of power distribution network construction standards using a gating attention mechanism.

[0027] Specifically, using a pre-trained word embedding model, the text data of the segmented power distribution network construction standards is vectorized and mapped to obtain a distributed word vector representation sequence for each word in the text data; Using the Transformer encoder, based on the distributed word vector representation sequence, context semantic encoding is performed through a self-attention mechanism to obtain a hidden state vector sequence containing global context information; By utilizing a gated attention layer, based on the hidden state vector sequence, a trainable gated network is used to fuse with attention scores to perform core semantic focus and redundant information suppression processing, and finally aggregated to generate a denoised semantic feature vector with enhanced key parameters.

[0028] S1-2. Using a dynamic feature pyramid network, adaptive weighted fusion of multi-scale visual features is performed on the image data of power distribution network construction standards to obtain image feature vectors. The image data refers to construction images or drawings related to the textual data of power distribution network construction standards. Specifically: Using a deep convolutional neural network, the image data of power distribution network construction standards are processed by convolution and pooling layer by layer to obtain multi-level feature maps with decreasing resolution, forming the initial feature pyramid. The overall semantic information of the multi-level feature maps is obtained by global average pooling. A dynamic weight generation network is used to generate a set of dynamic weight coefficients for fusing features at each level based on the overall semantic information. Using a weighted fusion algorithm, based on the dynamic weight coefficients, adaptive weighted fusion processing is performed on the multi-level feature maps to obtain the image feature vector of the image data.

[0029] Feature maps at different levels (such as low-level detail textures and high-level semantic concepts) have different values ​​for understanding different types of drawings. The dynamic weight generation network can determine the importance of features at each level under the current semantic theme by analyzing the feature vectors after global average pooling (for example, more attention should be paid to the detail layer for detailed drawings and more attention should be paid to the semantic layer for layout drawings), thereby predicting the optimal dynamic weight coefficients to achieve adaptive fusion, and finally extracting more accurate features.

[0030] In step S2, using an uncertainty measurement mechanism and differentiable logic rules, based on a pre-constructed distribution network element knowledge graph, probabilistic knowledge querying and logical reasoning enhancement are performed on the initial joint features to obtain target joint features with confidence levels. Specifically: 2-1. Perform semantic similarity matching between the initial joint features and the pre-constructed distribution network primitive knowledge graph to obtain the association probability between the initial joint features and each node in the distribution network primitive knowledge graph. Collect all the association probabilities of the nodes to obtain the probability association vector of the distribution network primitive knowledge graph. The initial joint features are hybrid vectors formed by concatenating image features and text features, which integrate the original visual and semantic information, but are not aligned with explicit engineering concepts. Therefore, the representation of each initial joint feature is fuzzy. Here, fuzzy representation means that although the features contain information from the input data, they do not form a clear corresponding mapping relationship with engineering concepts in the domain (such as circuit breakers and cable joints), and belong to an intermediate state.

[0031] Therefore, the semantic similarity matching performed above is to initially map the ambiguous initial joint features to clear engineering concepts by quantifying the probability of each node in the graph (i.e., the probability association vector).

[0032] The probability association vector obtained here is a multidimensional array, where each dimension represents the association probability between the initial joint feature and a node.

[0033] 2-2. Using the graph propagation mechanism, multi-hop message propagation and feature aggregation are performed on the probabilistic association vector to obtain the node feature vector of the distribution network primitive knowledge graph. Semantic matching and logical reasoning are performed on the node feature vector using differentiable logic rules to generate an enhancement signal for correcting the initial joint features, and the activation intensity vector of each activated logic rule in the differentiable logic rules is recorded.

[0034] Specifically, based on the probabilistic association vector, a multi-head graph attention network is used to perform multi-hop message propagation and feature aggregation on the topology of the power distribution network primitive knowledge graph to obtain the node representation vector of the power distribution network primitive knowledge graph; the probabilistic association vector represents the initial state of the nodes in the power distribution network primitive knowledge graph, and the node representation vector integrates global graph structure semantic information; The semantic similarity is calculated by matching the node representation vector with the preconditions of each rule in the differentiable logic rules; multiple rules with semantic similarity greater than the similarity threshold are selected from the differentiable logic rules as applicable logic rules; By using logical operators to perform forward reasoning on multiple applicable logical rules, a unified enhancement signal is generated for correcting the initial joint features.

[0035] The forward reasoning process includes: obtaining conclusions from multiple applicable logical rules from the so-called logical rules; treating these conclusions as new knowledge fragments; and using differentiable logical operators to aggregate, resolve conflicts, and synthesize the conclusions. For example, if multiple rules point to the same conclusion, the strength of that conclusion is enhanced; if potential conflicts exist between conclusions, weighted reconciliation is performed. Finally, a unified, vector-form logical enhancement signal is synthesized, which encodes all engineering constraints and suggestions inferred from the current scenario. Each rule in the differentiable logical rules is an encapsulated knowledge unit, in the form of premises and conclusions. Forward reasoning integrates and deduces scattered, successfully matched logical rules, thereby deriving a new, globally consistent, deterministic conclusion from known facts. Here, known facts refer to the matching of node features with rule premises, and the deterministic conclusion refers to the enhancement signal. This conclusion provides structured knowledge constraints not explicitly included in the original data features, used to directionally correct initial features and improve their logical rationality and accuracy.

[0036] In this context, the node representation vector is a holistic vector, while the rule premise is typically a composite structure formed by connecting multiple conditions (sub-premises) through logical operators. The matching calculation described above does not directly calculate the overall similarity; instead, it usually decomposes the premise into atomic conditions (e.g., device A is a circuit breaker, device B is a busbar, etc.), calculates the semantic similarity between the node representation vector and each atomic condition, and uses a differentiable logical aggregation function to aggregate the matching degrees of these atomic conditions into a holistic semantic similarity for the rule premise. The probability association vector recorded above represents the activation intensity vector of each triggered logical rule in the differentiable logical rules during logical reasoning, and is used for the next step of uncertainty analysis.

[0037] The pre-construction process of the above differentiable logic rules is as follows: Using dependency parsing and a domain keyword database, syntactic structure analysis and core extraction were performed on the standard and specification texts for power distribution network construction to obtain a preliminary corpus. The preliminary corpus includes coarsely labeled construction actions and equipment entities. The standard and specification texts for power distribution network construction were obtained from publicly available mandatory power distribution network construction design regulations, construction standards, construction acceptance specifications, and standardized construction plans from historical projects. Using a pre-trained language model, multiple sets of construction conditions, operational behaviors, and constraint parameters in the preliminary corpus are identified, and multiple structured triples based on construction conditions, operational behaviors, and constraint parameters are constructed. Using a production-based rule framework, each structured triple is converted into a predicate logic expression, and the predicate logic expressions are aggregated to obtain a logic rule library; The discrete operators in the predicate logic expression are replaced with continuously differentiable functions using fuzzy logic functions, and the continuously differentiable functions are encapsulated into callable computing units. The callable computing units are then fused with the vector space of the distribution network element knowledge graph to obtain a rule computing unit library. By using a lightweight pre-training strategy combined with historical power distribution network construction standard data, the link weights in the rule calculation unit library are subjected to supervised fine-tuning to obtain differentiable logic rules.

[0038] The construction of differentiable logic rules transforms discrete engineering specification texts into knowledge components that can be directly computed with neural networks through structuring and differentiability techniques. This achieves precise and structured encapsulation of domain knowledge, and more importantly, endows it with gradient-optimizable and data-driven self-evolution capabilities. This breaks down the training barriers between symbolic rules and deep learning models, not only deeply embedding domain knowledge into the system in a computable form but also enabling it to continuously learn and self-calibrate from real-world data. Through the aforementioned pre-training, we obtain what can be called logic rule components that have completed domain-adaptive initialization and are directly embedded into the neural inference layer.

[0039] By constructing a two-layer knowledge enhancement mechanism of "graph structure propagation - logical rule reasoning," the global topological semantics of the graph are deeply integrated with the local logical constraints of expert rules, achieving a leap from data association to causal inference. This enables the generated enhanced signal to possess both structured relationship awareness and interpretable logical derivation, allowing for precise correction of initial features that conforms to both engineering context and professional standards, thereby improving the accuracy and reliability of subsequent matching.

[0040] 2-3. Using a lightweight regression network combined with an uncertainty measurement mechanism, the entropy value of the probabilistic correlation vector and the activation intensity vector of the logical rule are nonlinearly fused to obtain the confidence level of the enhanced signal.

[0041] Specifically, based on the uncertainty measurement mechanism, the statistical characteristics of the activation intensity vector of the logical rule are calculated, and the statistical characteristics are the mean or variance of the activation intensity vector; The entropy value of the probabilistic association vector is concatenated with the statistical features of the activation intensity vector to obtain the uncertainty description vector of the initial joint features; The uncertainty description vector is transformed and fused using a nonlinear activation function of a lightweight regression network to obtain an implicit feature vector. This implicit feature vector is then mapped through a normalized probability transformation layer to obtain the confidence level of the enhanced signal. The normalized probability transformation layer can be a Sigmoid output layer (an output layer that uses the Sigmoid function as the activation function).

[0042] Among them, the statistical characteristics of the activation intensity vector reflect the internal consistency and determinism of the reasoning logic, the entropy value of the probabilistic association vector reflects the fuzziness of the original query, and the fuzziness (high entropy value) of the input features (probabilistic association vector) directly reflects the degree of ambiguity of the feature extraction model's initial perception of drawings and text, which is the key to identifying the source of error. This fuzziness index provides an objective calibration benchmark for confidence calculation and works together with the consistency index of logical rules: even if the reasoning logic itself seems consistent, if the input features themselves are highly fuzzy, the system should output a lower overall confidence, thereby effectively avoiding the decision risk of 'logical precision but weak foundation' caused by fuzzy underlying perception. That is, systematically suppressing the risk caused by ambiguity of the original features, even if the subsequent logical reasoning is correct in form, the overall conclusion is unreliable.

[0043] 2-4. Adaptively weight the enhanced signal using the confidence level, and perform semantic correction and relation supplementation on the initial joint features based on the weighted enhanced signal to obtain the target joint features with confidence level.

[0044] Specifically, using an uncertainty metric function, the intensity of the enhanced signal is adaptively scaled based on the confidence level to obtain a weighted logical enhanced signal; A feature correction and supplementation unit is adopted, which uses the weighted logical enhancement signal to perform targeted correction and information injection processing on the semantic and relational dimensions of the initial joint features through vector addition and gating mechanism, so as to realize the semantic correction and relational supplementation of features; The corrected initial joint features are concatenated with their corresponding confidence scores to output a target joint feature vector with quantized confidence scores.

[0045] The confidence level of the enhanced signal here is used to assess the credibility of the operation of injecting domain knowledge into the initial joint features.

[0046] In step S3, based on the target joint features, the semantic and procedural relevance between the target joint features and the standard clauses in the construction process diagram is calculated in parallel to generate construction standard matching results.

[0047] Specifically, the joint features of the target are input into the construction stage classification model to generate the current construction stage and the construction stage context representation. The construction stage context representation includes relevant constraint information and commonly used tools and materials in the construction stage. Based on the construction stage context representation, the process nodes of all standard clauses in the construction process graph are traversed and queried to obtain the process relevance of each standard clause to the current construction stage; and the semantic similarity between the target joint feature and the feature of each standard clause in the construction process graph is calculated. After adjusting the process relevance and semantic similarity of each standard clause based on the confidence level, a weighted fusion is performed to obtain the comprehensive matching degree of each standard clause. Based on the comprehensive matching degree of each standard clause, the construction standard matching results are selected from the construction process diagram.

[0048] The above method transforms initial joint features into machine-understandable, confidence-based engineering facts using a power distribution network primitive knowledge graph, fundamentally solving the problem of inaccurate extraction of professional features. Based on the engineering facts of the target joint features, a construction stage classification model is used to determine which construction stage the current activity belongs to. Based on the process correlation between the current construction stage and each standard clause node, as well as the semantic similarity between the target joint features and each standard clause, the most relevant results that strictly fit the current construction stage are selected.

[0049] The graph-based knowledge graph ensures accurate feature identification in images and text, fundamentally solving the problem of inaccurate feature extraction. The introduction of the construction process graph provides constraints and guidance for the matching process based on the construction workflow, filtering out a large number of semantically relevant but stage-irrelevant interference items. The output matching results have direct operability and business applicability.

[0050] This solution achieves accurate extraction of image and text features through a dynamic feature pyramid network and a gated attention mechanism, solving the fundamental problem of inaccurate feature extraction in general models. Furthermore, by integrating neural symbolic reasoning with a primitive knowledge graph, image and text features are enhanced into engineering semantic representations with confidence, ensuring accurate understanding of technical terms and symbols. Based on this, a construction process graph is introduced for parallel semantic and process relevance calculation, ensuring that the matching results not only conform to text semantics and image features but also strictly fit the current construction stage, effectively filtering out stage-irrelevant interference. Ultimately, a complete technical closed loop is formed, from accurate perception and semantic understanding to process-oriented decision-making, significantly improving the accuracy and practicality of standard matching.

[0051] Example 2 Based on the same inventive concept, this invention also provides an artificial intelligence-based standard text matching system for power distribution network construction, such as... Figure 2 As shown, it includes: The feature extraction module is used to extract semantic features from the text data of power distribution network construction standards using a gated attention mechanism to obtain semantic feature vectors, and to perform adaptive weighted fusion processing on the image data of power distribution network construction standards using a dynamic feature pyramid network to obtain image feature vectors. The feature analysis module is used to concatenate the image feature vector and the semantic feature vector to obtain the initial joint feature; using the uncertainty measurement mechanism and differentiable logic rules, based on the pre-constructed distribution network primitive knowledge graph, the initial joint feature is subjected to probabilistic knowledge query and logical reasoning enhancement to obtain the target joint feature with confidence. The standard matching module is used to calculate the semantic and procedural relevance between the target joint features and the standard clauses in the construction process diagram in parallel based on the target joint features, and generate construction standard matching results.

[0052] In one possible implementation, the feature analysis module includes: The probability association unit is used to perform semantic similarity matching between the initial joint features and the pre-constructed distribution network primitive knowledge graph to obtain the association probability between the initial joint features and each node in the distribution network primitive knowledge graph, and to collect the association probabilities of all the nodes to obtain the probability association vector of the distribution network primitive knowledge graph. The logic reasoning unit is used to perform multi-hop message propagation and feature aggregation on the probability association vector using the graph propagation mechanism to obtain the node feature vector of the distribution network graph element knowledge graph, perform semantic matching and logical reasoning on the node feature vector using differentiable logic rules, generate an enhancement signal for correcting the initial joint features, and record the activation intensity vector of each activated logic rule in the differentiable logic rules. The fusion unit is used to perform nonlinear fusion of the entropy value of the probabilistic correlation vector and the activation intensity vector of the logical rule using a lightweight regression network combined with an uncertainty measurement mechanism to obtain the confidence level of the enhanced signal. An adaptive correction unit is used to adaptively weight the enhanced signal using the confidence level, and to perform semantic correction and relational supplementation on the initial joint features based on the weighted enhanced signal to obtain target joint features with confidence level.

[0053] In one possible implementation, the aforementioned fusion unit is specifically used for: Based on the uncertainty measurement mechanism, the statistical characteristics of the activation intensity vector of the logical rule are calculated, and the statistical characteristics are the mean or variance of the activation intensity vector. The entropy value of the probabilistic association vector is concatenated with the statistical features of the activation intensity vector to obtain the uncertainty description vector of the initial joint features; The uncertainty description vector is transformed and fused using the nonlinear activation function of a lightweight regression network to obtain an implicit feature vector. The implicit feature vector is then mapped through a normalized probability transformation layer to obtain the confidence level of the enhanced signal.

[0054] In one possible implementation, the aforementioned logical reasoning unit is specifically used for: Based on the probabilistic association vector, the graph propagation mechanism of the multi-head graph attention network is used to perform multi-hop message propagation and feature aggregation on the topology of the power distribution network primitive knowledge graph to obtain the node representation vector of the power distribution network primitive knowledge graph. The semantic similarity is obtained by semantically matching the node representation vector with the preconditions of each rule in the differentiable logic rules. From the differentiable logic rules, select multiple rules with semantic similarity greater than the similarity threshold as applicable logic rules; By using logical operators to perform forward reasoning on multiple applicable logical rules, a unified enhancement signal is generated for correcting the initial joint features.

[0055] In one possible implementation, the system further includes a rule building module for: Using dependency parsing and a domain keyword database, we performed syntactic structure analysis and core extraction on the standard and specification texts for power distribution network construction to obtain preliminary corpus. Using a pre-trained language model, multiple sets of construction conditions, operational behaviors, and constraint parameters in the preliminary corpus are identified, and multiple structured triples based on construction conditions, operational behaviors, and constraint parameters are constructed. Using a production-based rule framework, each structured triple is converted into a predicate logic expression, and the predicate logic expressions are aggregated to obtain a logic rule library; The discrete operators in the predicate logic expression are replaced with continuously differentiable functions using fuzzy logic functions, and the continuously differentiable functions are encapsulated into callable computing units. The callable computing units are then fused with the vector space of the distribution network element knowledge graph to obtain a rule computing unit library. By using a lightweight pre-training strategy combined with historical power distribution network construction standard data, the link weights in the rule calculation unit library are subjected to supervised fine-tuning to obtain differentiable logic rules.

[0056] In one possible implementation, the aforementioned standard matching module is specifically used for: Input the joint features of the target into the construction stage classification model to generate the current construction stage and the construction stage context representation; Based on the construction stage context representation, the process nodes of all standard clauses in the construction process graph are traversed and queried to obtain the process relevance of each standard clause to the current construction stage; and the semantic similarity between the target joint feature and the feature of each standard clause in the construction process graph is calculated. After adjusting the process relevance and semantic similarity of each standard clause based on the confidence level, a weighted fusion is performed to obtain the comprehensive matching degree of each standard clause. Based on the comprehensive matching degree of each standard clause, the construction standard matching results are selected from the construction process diagram.

[0057] In one possible implementation, the feature extraction module described above is specifically used for: Using a deep convolutional neural network, the image data of power distribution network construction standards are processed by convolution and pooling layer by layer to obtain multi-level feature maps with decreasing resolution, forming the initial feature pyramid. The overall semantic information of the multi-level feature maps is obtained by global average pooling. A dynamic weight generation network is used to generate a set of dynamic weight coefficients for fusing features at each level based on the overall semantic information. Using a weighted fusion algorithm, based on the dynamic weight coefficients, adaptive weighted fusion processing is performed on the multi-level feature maps to obtain the image feature vector of the image data.

[0058] Example 3 like Figure 3 As shown, the present invention also provides an electronic device, which may be a computer device, a microcontroller device, a smart mobile device, etc. The electronic device in this embodiment may include a processor, a memory, a transceiver component, etc. The memory, processor, and transceiver component are connected via a bus; the memory can be used to store executable programs, and an exemplary executable program may include instructions; the processor is used to execute the instructions stored in the memory. The memory can also be used to store data, which can be accessed and / or modified when instructions are executed.

[0059] The processor may be a Central Processing Unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, and it is suitable for implementing one or more instructions. Specifically, it is suitable for loading and executing one or more instructions in the storage medium to realize the corresponding method flow or corresponding function, so as to realize the steps of the artificial intelligence-based power distribution network construction standard text matching method in the above embodiments.

[0060] Example 4 Based on the same inventive concept, this invention also provides a readable storage medium, specifically an electronic device readable storage medium (Memory). This readable storage medium is a memory device within an electronic device used to store programs and data. It is understood that the storage medium here can include both built-in storage media within the electronic device and extended storage media supported by the electronic device. The storage medium provides storage space, which stores the terminal's operating system. Furthermore, this storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more executable programs (including program code). It should be noted that the storage medium here can be high-speed RAM or non-volatile memory, such as at least one disk storage device. Loading and executing one or more instructions stored in the storage medium by the processor can implement the steps of the artificial intelligence-based power distribution network construction standard text matching method described in the above embodiments.

[0061] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention 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.

[0062] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. 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 illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0063] 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.

[0064] 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.

[0065] 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 its scope of protection. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that after reading the present invention, they can still make various changes, modifications or equivalent substitutions to the specific implementation methods of the application, but these changes, modifications or equivalent substitutions are all within the scope of protection of the claims pending approval.

Claims

1. A method for matching standard texts in power distribution network construction based on artificial intelligence, characterized in that, include: A gating attention mechanism is used to extract semantic features from the text data of power distribution network construction standards to obtain semantic feature vectors. A dynamic feature pyramid network is used to perform adaptive weighted fusion processing on the image data of power distribution network construction standards to obtain image feature vectors. The image feature vector and semantic feature vector are concatenated to obtain the initial joint feature; using the uncertainty measurement mechanism and differentiable logic rules, based on the pre-constructed distribution network primitive knowledge graph, the initial joint feature is enhanced by probabilistic knowledge query and logical reasoning to obtain the target joint feature with confidence. Based on the target joint features, the semantic and procedural relevance between the target joint features and the standard clauses in the construction process diagram is calculated in parallel to generate construction standard matching results.

2. The method as described in claim 1, characterized in that, The method utilizes an uncertainty measurement mechanism and differentiable logic rules, based on a pre-constructed distribution network primitive knowledge graph, to perform probabilistic knowledge querying and logical reasoning enhancement on the initial joint features, resulting in target joint features with confidence levels, including: The initial joint features are semantically similar to the pre-constructed distribution network primitive knowledge graph to obtain the association probability between the initial joint features and each node in the distribution network primitive knowledge graph. The association probabilities of all the nodes are aggregated to obtain the probability association vector of the distribution network primitive knowledge graph. The probability association vector is propagated and its features are aggregated through multi-hop message propagation using a graph propagation mechanism to obtain the node feature vector of the distribution network primitive knowledge graph. The node feature vector is then semantically matched and logically reasoned using differentiable logic rules to generate an enhancement signal for correcting the initial joint features. The activation intensity vector of each activated logic rule in the differentiable logic rules is also recorded. By using a lightweight regression network combined with an uncertainty measurement mechanism, the entropy value of the probabilistic correlation vector and the activation intensity vector of the logical rule are nonlinearly fused to obtain the confidence level of the enhanced signal. The enhanced signal is adaptively weighted using the confidence level, and the initial joint features are semantically corrected and relationally supplemented based on the weighted enhanced signal to obtain the target joint features with confidence level.

3. The method as described in claim 2, characterized in that, The method of using a lightweight regression network combined with an uncertainty measurement mechanism to nonlinearly fuse the entropy value of the probabilistic association vector and the activation intensity vector of the logical rule to obtain the confidence of the enhanced signal includes: Based on the uncertainty measurement mechanism, the statistical characteristics of the activation intensity vector of the logical rule are calculated, and the statistical characteristics are the mean or variance of the activation intensity vector. The entropy value of the probabilistic association vector is concatenated with the statistical features of the activation intensity vector to obtain the uncertainty description vector of the initial joint features; The uncertainty description vector is transformed and fused using the nonlinear activation function of a lightweight regression network to obtain an implicit feature vector. The implicit feature vector is then mapped through a normalized probability transformation layer to obtain the confidence level of the enhanced signal.

4. The method as described in claim 2, characterized in that, The process involves using a graph propagation mechanism to perform multi-hop message propagation and feature aggregation on the probabilistic association vectors to obtain node feature vectors of the distribution network primitive knowledge graph. Then, differentiable logic rules are used to perform semantic matching and logical reasoning on the node feature vectors to generate an enhancement signal for correcting the initial joint features. This includes: Based on the probabilistic association vector, the graph propagation mechanism of the multi-head graph attention network is used to perform multi-hop message propagation and feature aggregation on the topology of the power distribution network primitive knowledge graph to obtain the node representation vector of the power distribution network primitive knowledge graph. The semantic similarity is obtained by semantically matching the node representation vector with the preconditions of each rule in the differentiable logic rules. From the differentiable logic rules, select multiple rules with semantic similarity greater than the similarity threshold as applicable logic rules; By using logical operators to perform forward reasoning on multiple applicable logical rules, a unified enhancement signal is generated for correcting the initial joint features.

5. The method as described in claim 1, characterized in that, The pre-construction process of the differentiable logic rules is as follows: Using dependency parsing and a domain keyword database, we performed syntactic structure analysis and core extraction on the standard and specification texts for power distribution network construction to obtain preliminary corpus. Using a pre-trained language model, multiple sets of construction conditions, operational behaviors, and constraint parameters in the preliminary corpus are identified, and multiple structured triples based on construction conditions, operational behaviors, and constraint parameters are constructed. Using a production-based rule framework, each structured triple is converted into a predicate logic expression, and the predicate logic expressions are aggregated to obtain a logic rule library; The discrete operators in the predicate logic expression are replaced with continuously differentiable functions using fuzzy logic functions, and the continuously differentiable functions are encapsulated into callable computing units. The callable computing units are then fused with the vector space of the distribution network element knowledge graph to obtain a rule computing unit library. By using a lightweight pre-training strategy combined with historical power distribution network construction standard data, the link weights in the rule calculation unit library are subjected to supervised fine-tuning to obtain differentiable logic rules.

6. The method as described in claim 1, characterized in that, Based on the target joint features, the semantic and procedural relevance between the target joint features and the standard clauses in the construction process diagram is calculated in parallel to generate construction standard matching results, including: Input the joint features of the target into the construction stage classification model to generate the current construction stage and the construction stage context representation; Based on the construction stage context representation, the process nodes of all standard clauses in the construction process graph are traversed and queried to obtain the process relevance of each standard clause to the current construction stage; and the semantic similarity between the target joint feature and the feature of each standard clause in the construction process graph is calculated. After adjusting the process relevance and semantic similarity of each standard clause based on the confidence level, a weighted fusion is performed to obtain the comprehensive matching degree of each standard clause. Based on the comprehensive matching degree of each standard clause, the construction standard matching results are selected from the construction process diagram.

7. The method as described in claim 1, characterized in that, The method of using a dynamic feature pyramid network to adaptively weight and fuse image data of power distribution network construction standards to obtain image feature vectors includes: Using a deep convolutional neural network, the image data of power distribution network construction standards are processed by convolution and pooling layer by layer to obtain multi-level feature maps with decreasing resolution, forming the initial feature pyramid. The overall semantic information of the multi-level feature maps is obtained by global average pooling. A dynamic weight generation network is used to generate a set of dynamic weight coefficients for fusing features at each level based on the overall semantic information. Using a weighted fusion algorithm, based on the dynamic weight coefficients, adaptive weighted fusion processing is performed on the multi-level feature maps to obtain the image feature vector of the image data.

8. A text matching system for power distribution network construction standards based on artificial intelligence, characterized in that, include: The feature extraction module is used to extract semantic features from the text data of power distribution network construction standards using a gated attention mechanism to obtain semantic feature vectors, and to perform adaptive weighted fusion processing on the image data of power distribution network construction standards using a dynamic feature pyramid network to obtain image feature vectors. The feature analysis module is used to concatenate the image feature vector and the semantic feature vector to obtain the initial joint feature; using the uncertainty measurement mechanism and differentiable logic rules, based on the pre-constructed distribution network primitive knowledge graph, the initial joint feature is subjected to probabilistic knowledge query and logical reasoning enhancement to obtain the target joint feature with confidence. The standard matching module is used to calculate the semantic and procedural relevance between the target joint features and the standard clauses in the construction process diagram in parallel based on the target joint features, and generate construction standard matching results.

9. The system as described in claim 8, characterized in that, The feature analysis module includes: The probability association unit is used to perform semantic similarity matching between the initial joint features and the pre-constructed distribution network primitive knowledge graph to obtain the association probability between the initial joint features and each node in the distribution network primitive knowledge graph, and to collect the association probabilities of all the nodes to obtain the probability association vector of the distribution network primitive knowledge graph. The logic reasoning unit is used to perform multi-hop message propagation and feature aggregation on the probability association vector using the graph propagation mechanism to obtain the node feature vector of the distribution network graph element knowledge graph, perform semantic matching and logical reasoning on the node feature vector using differentiable logic rules, generate an enhancement signal for correcting the initial joint features, and record the activation intensity vector of each activated logic rule in the differentiable logic rules. The fusion unit is used to perform nonlinear fusion of the entropy value of the probabilistic correlation vector and the activation intensity vector of the logical rule using a lightweight regression network combined with an uncertainty measurement mechanism to obtain the confidence level of the enhanced signal. An adaptive correction unit is used to adaptively weight the enhanced signal using the confidence level, and to perform semantic correction and relational supplementation on the initial joint features based on the weighted enhanced signal to obtain target joint features with confidence level.

10. The system as described in claim 9, characterized in that, The fusion unit is specifically used for: Based on the uncertainty measurement mechanism, the statistical characteristics of the activation intensity vector of the logical rule are calculated, and the statistical characteristics are the mean or variance of the activation intensity vector. The entropy value of the probabilistic association vector is concatenated with the statistical features of the activation intensity vector to obtain the uncertainty description vector of the initial joint features; The uncertainty description vector is transformed and fused using the nonlinear activation function of a lightweight regression network to obtain an implicit feature vector. The implicit feature vector is then mapped through a normalized probability transformation layer to obtain the confidence level of the enhanced signal.

Images