Transmission tower construction scheme automatic generation method based on search enhancement and constraint generation

Abstract

The application discloses a power transmission tower construction scheme automatic generation method based on search enhancement and constraint generation, which comprises the following steps: 1) collecting power transmission tower field authoritative materials to construct a hierarchical knowledge base; 2) collecting true construction schemes to construct a standardized data set; 3) constructing a construction scheme intelligent generation model comprising an input analysis module, a field adaptive hybrid knowledge search module, a multi-task constraint generation and hierarchical compliance decoding module and a multi-dimensional reward modeling module; 4) obtaining a final model through two-stage training; and 5) inputting construction parameters to automatically output a structured construction scheme. The field adaptive hybrid knowledge search module acquires knowledge fragments such as standard specifications from the authoritative knowledge base through hybrid scoring search; the multi-task constraint generation and hierarchical compliance decoding module adopts a shared encoder and a task-specific decoder cooperative architecture to jointly generate scheme components, and the hierarchical constraint set is used to enforce compliance in the decoding stage; and the multi-dimensional reward modeling module performs quality evaluation.

Description

Technical Field

[0001] This invention relates to an automatic generation method for power transmission tower construction schemes based on retrieval enhancement and constraint generation. Background Technology

[0002] The erection of transmission towers is a critical process in power grid construction. Its complex working environment, frequent high-altitude operations, and dense overlapping of lifting and installation activities have historically made it a high-risk operation. A complete construction plan must systematically cover modules such as work steps, equipment selection and layout, safety protection measures, and risk assessment. These modules must be closely interconnected technically and strictly comply with current national standards (GB), power industry standards (DL), and the real-time operating conditions of the construction site.

[0003] Currently, the preparation of such construction plans generally relies on experienced engineering technicians to complete them manually. The preparation process requires the comprehensive extraction and interpretation of heterogeneous information from multiple sources, including geological and meteorological reports, equipment technical manuals, standard specifications, and past project cases. This information integration is challenging, and the applicability of clauses and the matching of parameters must be repeatedly verified. Practical experience shows that the preparation of a medium-sized tower erection plan typically takes tens of hours, and the final quality is significantly limited by the experience and subjective judgment of the preparers, making it difficult to consistently guarantee the standardization, completeness, and consistency of the plan.

[0004] In recent years, Large Language Models (LLMs) based on the Transformer architecture have demonstrated superior performance in the field of general text generation. However, when directly applied to the automatic generation of construction schemes, they still face significant technical bottlenecks. On the one hand, the models are prone to "illusion" problems such as factual errors, misuse of standard clauses, or missing structural logic, making it difficult to meet the rigid requirements of accuracy and compliance in power engineering. On the other hand, existing Retrieval Enhanced Generation (RAG) methods are mostly designed for general corpora and lack the ability to deeply model the hierarchical structure, version timeliness, and normative wording of the power industry standard system. The retrieval results often confuse obsolete clauses with current clauses and cannot effectively distinguish between mandatory and recommended clauses, making it difficult for the generated content to pass the actual engineering review. Therefore, it is urgent to construct a scheme generation method for transmission tower erection construction that integrates domain knowledge constraints and normative logical reasoning to overcome the current technical limitations. Summary of the Invention

[0005] The present invention aims to overcome the above-mentioned shortcomings of the prior art and provide an automatic generation method for transmission tower construction schemes based on retrieval enhancement and constraint generation, so as to solve the deficiencies of the prior art in terms of compliance, logic and operability when generating transmission tower construction schemes.

[0006] The present invention solves the technical problem by adopting the following technical solution: An automatic generation method for power transmission tower construction schemes based on retrieval enhancement and constraint generation includes the following steps: Step 1: Collect standards, specifications, equipment manuals, and historical solutions in the field of power transmission tower construction, and form a hierarchical domain knowledge base through standardization and indexing operations; Step 2: Collect real construction plans as sample data, and after preprocessing, construct a dataset, which is then divided into training set, validation set and test set according to a set ratio; Step 3: Construct a Transmission Tower Construction Scheme Generation Model Based on Retrieval-Augmented and Constrained Generation (TTCSGM). The structure of the TTCSGM model is as follows: starting from the input parsing module, it is sequentially connected to the domain adaptive hybrid knowledge retrieval module, the multi-task constraint generation and hierarchical compliance decoding module, and the multi-dimensional reward modeling module. The input parsing module extracts construction parameters and scene categories; the domain-adaptive hybrid knowledge retrieval module performs a four-dimensional hybrid retrieval based on the hierarchical domain knowledge base, integrating semantic similarity, vocabulary matching, source authority, and document timeliness in the retrieval score, and dynamically weights and fuses multiple knowledge fragments obtained through a gating attention mechanism to generate enhanced prompts; the multi-task constraint generation and hierarchical compliance decoding module, based on a shared encoder and a multi-task decoder, conditionally generates construction procedures, equipment lists, safety measures, and risk assessments; the generation of the equipment list is conditional on the already generated construction procedures, and the generation of safety measures and risk assessments is conditional on the construction procedures and equipment lists; a coverage-aware attention mechanism is introduced during the generation process to reduce procedure repetition, and hierarchical constraint decoding is adopted, dynamically limiting the feasible token set at each decoding step to achieve real-time enforcement of terminology compliance, procedure sequence compliance, and semantic security compliance; the multi-dimensional reward modeling module outputs scores based on five dimensions: terminology accuracy, process compliance, safety integrity, operability, and document standardization, and optimizes module parameters through joint training of list sorting loss, regression loss, and consistency regularization term; Step 4: Input the training set into the power transmission tower construction scheme generation model and perform two-stage training: In the first stage, supervised learning is used to train the multi-dimensional reward modeling module with the goal of minimizing the weighted sum of regression loss, ranking loss, and consistency regularization term; in the second stage, reinforcement learning is used to train the construction scheme generation model with the goal of maximizing the overall reward output by the multi-dimensional reward modeling module, and the parameters of the trained multi-dimensional reward modeling module are frozen to obtain the trained power transmission tower construction scheme generation model. Step 5: Input the test set into the trained power transmission tower construction scheme generation model, and automatically generate construction schemes corresponding to different construction conditions in the test set.

[0007] More specifically, step 1 includes: Step 1.1 Collect national standards, power industry standards, equipment operation manuals, technical manuals, historical construction plans and typical construction methods in the field of power transmission tower construction, and clean the collected raw text data to remove duplicate, outdated and invalid content. Step 1.2 The cleaned text data is hierarchically split into chapters and clauses to form independent knowledge units. The knowledge units are labeled with types, including standard types, equipment parameters, construction procedures, safety measures and risk points. Step 1.3 uses the domain-adapted Sentence-BERT model to encode knowledge units and generate 768-dimensional vector representations; Step 1.4 Construct a knowledge structure using a graph neural network: Build a hierarchical index with knowledge units as nodes and relationships as edges, and assign timeliness weights and authority weights to each knowledge unit; Step 1.5 Construct semantic vector indexes and keyword indexes to build a vector knowledge base that supports hybrid retrieval.

[0008] More specifically, step 2 includes: Step 2.1 Collect real transmission tower construction project plan data from large power infrastructure companies over the past ten years, and remove sensitive identifying information after anonymization; Step 2.2 Filter out effective solutions covering various tower models, terrain types, and voltage levels to construct the original dataset; Step 2.3 The selected construction plan is divided into four modules according to the standard structure: construction procedures, equipment list, safety measures and risk assessment. The processed dataset is then divided into training set, validation set and test set according to the preset ratio.

[0009] In step 3, the input parsing module adopts a Transformer-based dual-encoder architecture. The construction parameter encoder is a BERT model fine-tuned from the power engineering corpus to perform entity extraction, while the scene encoder performs scene intent classification. The domain-adaptive hybrid knowledge retrieval module consists of a knowledge base encoder, a hybrid retrieval scorer, and a gated fusion unit connected in series. The knowledge base encoder uses Sentence-BERT output vectors based on a Siamese network architecture. The hybrid retrieval scorer integrates semantic similarity, BM25 matching, learnable scores based on metadata, and exponentially decaying timeliness scores. The gated fusion unit utilizes a graph attention network structure... The fully connected layer performs weighted fusion of the retrieved fragments; the multi-task constraint generation and hierarchical compliance decoding module consists of a shared Transformer encoder and four task-specific decoders. The process decoder introduces a coverage vector to calculate the coverage loss, the equipment list decoder connects to a multi-label classification head, and the cross-attention input of the safety and risk decoders originates from the output projection of the process and equipment decoders; the multi-dimensional reward modeling module consists of a hierarchical Transformer encoder combined with an attention pooling layer, outputting a document-level representation and feeding it into five independent multilayer perceptron heads. Each head is processed by a fully connected layer and an activation function to output a score in the [0,1] interval. More specifically, step 3 includes: Step 3.1: Construct an input parsing module to extract construction parameters and scene categories; Step 3.2: Construct a domain-adaptive hybrid knowledge retrieval module, which performs hybrid retrieval and evidence fusion based on a hierarchical domain knowledge base, and outputs enhanced hints; Step 3.3: Construct a multi-task constraint generation and hierarchical compliance decoding module, which uses a shared encoder and a multi-task decoder to jointly generate construction procedures, equipment lists, safety measures and risk assessments, and applies three levels of constraints on terminology, procedure sequence and semantic security during the decoding process; Step 3.4: Construct a multi-dimensional reward modeling module to score and select the best construction plan from five dimensions: terminology accuracy, process compliance, safety integrity, operability, and document standardization.

[0010] More specifically, step 3.1 includes: Step 3.1.1: Define output parameters and labels; the output parameter set includes tower type, terrain, construction method and voltage level, and scene labels are used to constrain the engineering scope of subsequent knowledge retrieval and solution generation; Step 3.1.2: Construct a dual encoder architecture; adopt a Transformer-based dual encoder to perform entity extraction and scene intent classification respectively; Step 3.1.3: Encoding Input; Input the construction requirements, which are then encoded by dual encoders to output the construction parameters and scene labels.

[0011] More specifically, step 3.2 includes: Step 3.2.1: Construct a multidimensional retrieval scoring function; the multidimensional retrieval scoring function integrates four dimensions: semantic similarity, BM25 word matching, source authority, and document timeliness, and optimizes the weight coefficients of each dimension through grid search; Step 3.2.2: Perform hybrid retrieval; based on the hierarchical domain knowledge base, apply the above multidimensional retrieval scoring function to retrieve and score knowledge fragments in the knowledge base; Step 3.2.3: Filter highly matching knowledge fragments; sort the knowledge fragments according to the search score and select the top 10 knowledge fragments with the highest scores; Step 3.2.4: Dynamically weighted fusion of evidence; The selected 10 knowledge fragments are dynamically weighted and fused using a gated attention mechanism to integrate them into coherent knowledge evidence; Step 3.2.5: Generate enhanced prompts; combine the fused knowledge evidence with the original user input to form enhanced prompts, providing authoritative standard basis for subsequent scheme generation.

[0012] More specifically, step 3.3 includes: Step 3.3.1: Unified context encoding and multi-task framework; given the enhanced input, a unified context representation is obtained through a shared Transformer encoder; the construction scheme generation is decomposed into four tasks: construction procedures, equipment list, safety measures and risk assessment. Each task is generated in an autoregressive manner using a dedicated decoder and jointly optimized under a weighted multi-task objective. Step 3.3.2: Construction sequence generation and coverage awareness mechanism; use the sequence decoder to generate the construction sequence; introduce coverage awareness attention during the generation process, construct the coverage vector by accumulating the historical attention distribution, and record the content that has been paid attention to in the generated steps; at the same time, add a coverage penalty term to the loss function to guide the model to pay attention to the construction sequence that has not been fully covered, and avoid repeated generation. Step 3.3.3: Joint generation of equipment list; The equipment list is generated by first predicting the required equipment categories through a multi-label classification head, and then generating the corresponding equipment models and quantity descriptions based on the selected categories and the generated construction procedures. Step 3.3.4: Generation of conditions for safety measures and risk assessment; The generation of safety measures and risk assessment is based on the output of construction procedures and equipment lists, realizing logical alignment between components, and generating safety protection measures and risk control suggestions corresponding to the construction procedures; Step 3.3.5: Hierarchical Constraint Decoding; Constraint-aware decoding is used to ensure the domain compliance of the generated content. A hierarchical constraint set is defined, which includes terminology compliance constraints, process sequence compliance constraints, and semantic security compliance constraints. At each step of decoding, the range of tokens is constrained to the compliance vocabulary based on the currently generated prefix dynamic restriction feasible token set. Constraint bundle search is used for decoding, and a delayed constraint activation strategy is introduced to balance fluency and strict compliance. Step 3.3.6: Output candidate solutions; the final decoder outputs a complete, compliant, and executable set of construction solution candidates. More specifically, step 3.4 includes: Step 3.4.1: Module architecture; A hierarchical Transformer architecture is adopted to synchronously encode text and context. Token-level representations are aggregated layer by layer into sentence-level and paragraph-level embeddings through hierarchical attention pooling to capture the document's hierarchical structure. Step 3.4.2: Five-dimensional scoring system; five independent evaluation dimensions are set: terminology accuracy, process compliance, security integrity, operability, and document standardization. The score of each dimension ranges from 0 to 1; the overall reward is obtained by weighted sum of the scores of each dimension, and the weight of each dimension is a learnable parameter. Step 3.4.3: Joint training loss function; the module training objective is composed of three loss terms: list sorting loss to incorporate expert preferences, regression loss to fit expert labeled scores, and consistency regularization term to enhance evaluation stability; the loss terms are coordinated by a balancing coefficient. Step 3.4.4: Ranking and Regression Loss; The list ranking loss assigns a higher probability to the expert-preferred solution than to other candidate solutions; The regression loss uses mean squared error to align the model's predicted scores with the expert-labeled values. Step 3.4.5: Consistency Regularization and Evaluation Output; Constraints are imposed on the scoring differences between semantically similar construction schemes, and a maximum allowable threshold is set to suppress abnormal fluctuations; Through multi-objective joint training, the module can stably output evaluation results that meet expert standards, guiding the generation module to output construction schemes.

[0013] More specifically, step 4 includes: Step 4.1: Input the construction parameter samples from the training set into the model in batches according to the preset batch size. Each sample includes tower type, terrain, equipment type and construction method, as well as the corresponding expert-written reference construction plan. Step 4.2: Extract entities and classify scene intents from construction parameters through the input parsing module to obtain a structured parameter set and scene tags; obtain a preset number of authoritative knowledge fragments from the knowledge base based on hybrid scoring retrieval, and then concatenate them with the original input after gating attention fusion to generate enhanced prompts; Step 4.3: The enhanced prompts are input into the multi-task constraint generation module, which generates a unified context representation through the shared Transformer encoder. Then, each task's dedicated decoder generates four types of solution components in sequence under the constraint-aware decoding mechanism: construction procedures, equipment list, safety measures, and risk assessment. Step 4.4: With the goal of minimizing the weighted sum of regression loss, ranking loss, and consistency regularization term, supervised learning is used to train the multi-dimensional reward modeling module. After each preset number of training epochs, the consistency between the overall reward score output by the current module and the expert-annotated score is calculated using the validation set. When the consistency on the validation set no longer increases after a preset number of consecutive iterations, the first stage of training is stopped, and the parameters of the reward module with the highest consistency are saved; otherwise, training continues. Step 4.5: Load and freeze the parameters of the reward module trained in the first stage. With the goal of maximizing the overall reward, train the construction scheme generation model using a reinforcement learning algorithm. After each preset number of training cycles, use the validation set to calculate the overall reward score obtained by the construction scheme generated by the current model. When the overall reward score on the validation set no longer increases for a preset number of consecutive cycles, stop training and save the model parameters with the highest overall reward score as the final model; otherwise, update the network parameters and return to step 4.5. More specifically, step 5 includes: Step 5.1: Input the construction parameter samples from the test set into the final trained model in batches. Each sample includes tower type, terrain and equipment type. Step 5.2: The model completes the entire process of input parsing, knowledge retrieval, multi-task constraint generation and constraint-aware decoding, and outputs a structured construction plan containing four components: construction procedures, equipment list, safety measures and risk assessment. Step 5.3: Submit the generated construction plan to the multi-dimensional reward modeling module, score it according to five dimensions: terminology accuracy, process compliance, safety integrity, operability, and document standardization, and calculate the overall reward as a comprehensive evaluation index of the plan quality; Step 5.4: Output the final construction plan document and its five-dimensional quality assessment report.

[0014] The innovation of this invention is: (1) A domain-adaptive hybrid retrieval enhancement mechanism is proposed. Addressing the problems of existing retrieval enhancement generation relying on general corpora and single semantic matching, and the difficulty in capturing the differences in the hierarchy and timeliness of engineering standard clauses, this invention employs a graph neural network to hierarchically organize the domain knowledge base. The retrieval scoring integrates four-dimensional signals: semantic similarity, BM25 vocabulary matching, source authority, and document timeliness. Furthermore, it dynamically weights and fuses retrieval fragments through gated attention, thereby elevating the positioning of authoritative knowledge from "semantic matching" to "engineering adaptation," effectively solving the problem of insufficient retrieval accuracy in safety-critical domains.

[0015] (2) A two-stage generation mechanism combining multi-task condition generation and hierarchical compliance decoding was constructed. Unlike the existing technology that generates solution components independently or relies on post-verification, this invention uses a shared encoder and a task-specific decoder to collaboratively generate multiple types of components. The components are generated sequentially according to their conditional dependencies, and the process is suppressed by covering perception attention. During the decoding process, a three-level constraint set of "terminology compliance → process sequence compliance → semantic security compliance" is constructed to dynamically limit the feasible token space, and the compliance guarantee is changed from "post-generation verification" to "mandatory execution during decoding", which avoids component conflicts and specification omissions from the mechanism level.

[0016] (3) A multi-dimensional reward modeling and expert preference-aligned evaluation and feedback mechanism was designed. In view of the problem that existing evaluations rely on general indicators such as BLEU and ROUGE and cannot comprehensively evaluate the quality of the solution from an engineering perspective, this invention establishes a five-dimensional evaluation system of terminology accuracy, process compliance, safety integrity, operability, and document standardization. The reward model is jointly trained by ranking loss, regression loss and consistency regularization term, and the generation strategy is fine-tuned by reinforcement learning with this model as the optimization target. For the first time, "expert engineering experience" is transformed into a computable reward signal, so that the generated solution continuously approaches the expert level in terms of compliance and executability.

[0017] The present invention has the following beneficial effects: (1) By using the enhanced generation mechanism, authoritative knowledge such as national standards, industry norms and typical construction schemes are dynamically integrated into the scheme generation process, which greatly shortens the construction scheme preparation time and significantly improves the compliance of the scheme. This effectively solves the problems of traditional manual preparation relying on engineers' experience, being time-consuming and having poor consistency.

[0018] (2) By integrating multi-task constraint generation and hierarchical constraint decoding, it can jointly generate multiple components such as construction procedures, equipment lists, safety measures and risk assessments under a unified framework, and enforce compliance constraints during the decoding stage to ensure logical consistency and standard compliance among the components of the solution, thereby reducing the workload of manual verification and revision.

[0019] (3) Through multi-dimensional rewards, the generated schemes are evaluated in detail in terms of terminology accuracy, process compliance, safety integrity, operability and document standardization. Combined with expert feedback, the quality of the generated schemes is continuously optimized, and the compilation of power transmission tower construction schemes is transformed from experience-driven to data and knowledge-driven. Attached Figure Description

[0020] Figure 1 This is the overall flowchart of the present invention.

[0021] Figure 2 This is a structural diagram of the power transmission tower construction scheme generation model in this invention.

[0022] Figure 3 This is a structural diagram of the input parsing module in this invention.

[0023] Figure 4 This is a structural diagram of the domain-adaptive hybrid knowledge retrieval module in this invention.

[0024] Figure 5 This is a structural diagram of the multi-task constraint generation and hierarchical compliance decoding module in this invention.

[0025] Figure 6 This is a structural diagram of the multi-dimensional reward modeling module in this invention. Detailed Implementation

[0026] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0027] Example 1

[0028] This embodiment relates to a method for automatically generating transmission tower construction plans based on retrieval enhancement and constraint generation, such as... Figure 1 As shown, the method includes the following steps: Step 1: Collect standards, specifications, equipment manuals, and historical solutions in the field of power transmission tower construction, and form a hierarchical domain knowledge base through standardization and indexing operations; Step 2: Collect real construction plans as sample data, and after preprocessing, construct a dataset, which is then divided into training set, validation set and test set according to a set ratio; Step 3: Construct a power transmission tower construction scheme generation model based on retrieval enhancement and constraint generation; the structure of the generation model is as follows: starting from the input parsing module, it is sequentially connected to the domain adaptive hybrid knowledge retrieval module, the multi-task constraint generation and hierarchical compliance decoding module, and the multi-dimensional reward modeling module; The input parsing module extracts construction parameters and scene categories; the domain-adaptive hybrid knowledge retrieval module performs a four-dimensional hybrid retrieval based on the hierarchical domain knowledge base, integrating semantic similarity, vocabulary matching, source authority, and document timeliness in the retrieval score, and dynamically weights and fuses multiple knowledge fragments obtained through a gating attention mechanism to generate enhanced prompts; the multi-task constraint generation and hierarchical compliance decoding module, based on a shared encoder and a multi-task decoder, conditionally generates construction procedures, equipment lists, safety measures, and risk assessments; the generation of the equipment list is conditional on the already generated construction procedures, and the generation of safety measures and risk assessments is conditional on the construction procedures and equipment lists; a coverage-aware attention mechanism is introduced during the generation process to reduce procedure repetition, and hierarchical constraint decoding is adopted, dynamically limiting the feasible token set at each decoding step to achieve real-time enforcement of terminology compliance, procedure sequence compliance, and semantic security compliance; the multi-dimensional reward modeling module outputs scores based on five dimensions: terminology accuracy, process compliance, safety integrity, operability, and document standardization, and optimizes module parameters through joint training of list sorting loss, regression loss, and consistency regularization term; Step 4: Input the training set into the power transmission tower construction scheme generation model and perform two-stage training: In the first stage, supervised learning is used to train the multi-dimensional reward modeling module with the goal of minimizing the weighted sum of regression loss, ranking loss, and consistency regularization term; in the second stage, reinforcement learning is used to train the construction scheme generation model with the goal of maximizing the overall reward output by the multi-dimensional reward modeling module, and the parameters of the trained multi-dimensional reward modeling module are frozen to obtain the trained power transmission tower construction scheme generation model. Step 5: Input the test set into the trained power transmission tower construction scheme generation model, and automatically generate construction schemes corresponding to different construction conditions in the test set.

[0029] Step S1 includes: S1.1: Collect authoritative engineering data in the field of power transmission tower construction, including national standards (GB), power industry standards (DL), equipment operation and technical manuals, historical construction plans and typical construction methods, etc. Clean the collected raw text data, remove duplicate, outdated and invalid content, and convert the cleaned text data into plain text format and mark the source metadata.

[0030] S1.2: The cleaned text data is structurally broken down according to the chapter number of the standard specification, the functional module of the equipment manual, the task stage of the construction plan, etc., to form knowledge units with uniform granularity. The knowledge units are labeled with type, including standard type, equipment parameters, construction procedures, safety measures and risk points.

[0031] S1.3: The domain-adaptive Sentence-BERT model is used to encode knowledge units, mapping each knowledge unit to a 768-dimensional dense vector, which contains both semantic information and implicit logical relationships between terms.

[0032] S1.4: Construct a directed heterogeneous graph with knowledge units as nodes and their relationships as edges. The relationships are determined based on explicit references, implicit dependencies, and logical attributions in the text. Use a graph neural network (GNN) to pass messages between nodes, aggregate neighborhood information, and enhance the context sensitivity of node representation. Assign timeliness weights and authority weights to each knowledge unit to ensure that the latest and most authoritative knowledge is prioritized during retrieval.

[0033] S1.5: Based on the processed knowledge units, a hybrid index system is established to form a vector knowledge base. The semantic vector index uses the FAISS library to build an IVF-PQ structure, and performs clustering and compression on the 768-dimensional vector to support efficient approximate nearest neighbor retrieval. The keyword index uses Elasticsearch to build an inverted index for the title, clause number and core terms of the knowledge unit and supports Boolean queries and phrase matching.

[0034] Step S2 includes: S2.1 Collect data on actual power transmission tower construction projects from large power infrastructure companies over the past decade, and anonymize the original data by removing sensitive identifiers such as project names, construction units, geographical locations, and personnel information. S2.2 Select effective schemes from the collected schemes that cover a variety of tower types (such as goblet type, cat head type, and T-shaped type), terrain types (flat land, mountainous land, hilly land, etc.), and voltage levels (110kV, 220kV, 500kV, etc.) to ensure that the dataset has broad representativeness and avoid overfitting of the model to specific working conditions.

[0035] S2.3 breaks down the selected construction plans into four core modules according to a unified structure: construction procedures, equipment list, safety measures, and risk assessment. These modules are then organized in a clause-based format, and the processed dataset is divided into training, validation, and test sets according to a preset ratio.

[0036] The input parsing module described in step 3 adopts a Transformer-based dual-encoder architecture. The construction parameter encoder is a BERT model fine-tuned from the power engineering corpus, which performs entity extraction and outputs construction parameters. The scene encoder performs scene intent classification and outputs scene labels. The domain-adaptive hybrid knowledge retrieval module consists of a knowledge base encoder, a hybrid retrieval scorer, and a gated fusion unit connected in series. The knowledge base encoder adopts a Siamese network-based Sentence-BERT architecture, with a Transformer encoder as the backbone. The output end is connected to a mean pooling layer to output a 768-dimensional vector. The hybrid retrieval scorer consists of a cosine similarity calculation unit, a BM25 matching unit, a learnable fully connected layer with metadata features as input, and an exponential decay function generator connected in parallel to a dynamic weighted network. It extracts the top K segments in descending order of comprehensive score. The gated fusion unit uses the entity corresponding to the retrieval segment as the seed node and outputs a 256-dimensional embedding through a two-layer graph attention network. This embedding is concatenated with the 768-dimensional encoding of each segment to form a 1024-dimensional joint representation. This representation is compressed to 128 dimensions by the first fully connected layer, activated by ReLU, and mapped to a single-dimensional attention vector by the second fully connected layer. The weights are normalized by Softmax and then weighted to output a 768-dimensional fusion vector, which is then concatenated with the query code. The multi-task constraint generation and hierarchical compliance decoding module consists of a shared Transformer encoder and four task-specific decoders. The shared encoder outputs a context representation by alternating stacks of multi-head self-attention layers and feedforward network layers. Each decoder consists of a masked multi-head self-attention layer, a cross-attention layer, and a feedforward network layer. The cross-attention layer of the process decoder accumulates the historical distribution as a coverage vector and sums it with the minimum value of the current step attention as the coverage loss. The masked multi-head self-attention layer of the equipment list decoder is followed by a multi-label classification head. The cross-attention key-value pairs of the security decoder and the risk decoder are respectively derived from the linear projections of the outputs of the process and equipment decoders. The multi-dimensional reward modeling module consists of a hierarchical Transformer encoder connected to sentence-level and paragraph-level attention pooling layers. The output document-level representation is fed into five independent multilayer perceptron heads. Each head consists of two fully connected layers connected to a ReLU activation function and activated by a Sigmoid function to output a score in the [0,1] interval.

[0037] Among them, such as Figure 2 As shown, step S3 includes: Step 3.1: Construct an input parsing module to extract construction parameters and scene categories; Step 3.2: Construct a domain-adaptive hybrid knowledge retrieval module, which performs hybrid retrieval and evidence fusion based on a hierarchical domain knowledge base, and outputs enhanced hints; Step 3.3: Construct a multi-task constraint generation and hierarchical compliance decoding module, which uses a shared encoder and a multi-task decoder to jointly generate construction procedures, equipment lists, safety measures and risk assessments, and applies three levels of constraints on terminology, procedure sequence and semantic security during the decoding process; Step 3.4: Construct a multi-dimensional reward modeling module to score and select the best construction plan from five dimensions: terminology accuracy, process compliance, safety integrity, operability, and document standardization. Among them, such as Figure 3 As shown, step S3.1 includes: S3.1.1: Define output parameters and labels; the output parameter set covers tower type (such as goblet type, cat head type, dry character type, etc.), terrain type (such as flat land, mountain, hill, etc.), equipment type (such as pole erection, lifting machinery, etc.), construction method and voltage level (such as 110kV, 220kV, 500kV, etc.), and sets scene labels to constrain the engineering applicability of the entire reasoning process, ensuring that the solution generation is only carried out within the matching working condition boundary; S3.1.2: Construct a dual encoder architecture; this architecture consists of two Transformer encoders that share the underlying pre-trained parameters. One encoder focuses on sequence labeling-based entity extraction, while the other encoder performs text classification-based scene intent determination.

[0038] S3.1.3: Encoded Input; When the user inputs construction requirement text... At that time, dual encoders Joint encoding is performed, and the structured entity parameter set is output by the entity extraction encoder. Scene labels are output by the scene classification encoder. The calculation formula is as follows: (1) The entity parameter set includes the aforementioned engineering attributes, while the scene tag automatically determines the engineering category based on the semantic requirements.

[0039] Among them, such as Figure 4 As shown, step S3.2 includes:

[0040] S3.2.1: Construct a multi-dimensional retrieval scoring function; this function integrates four dimensions: semantic similarity, BM25 vocabulary matching, source authority, and document timeliness, to comprehensively evaluate the degree of matching between the query and the knowledge fragment. Semantic similarity is obtained by calculating the cosine similarity between the query vector and the knowledge fragment vector; the BM25 score is used to capture precise matching at the keyword level; authority is indicated by an indicator function to determine whether the knowledge fragment comes from authoritative sources such as national or industry standards; and timeliness is weighted based on the publication year and current status of the knowledge fragment. The scores of the four dimensions are linearly weighted and summed using adjustable weight coefficients. Each weight coefficient employs a grid search strategy, and its calculation formula is as follows: (2) in, This represents the cosine similarity between the query and the document vector. To score the vocabulary matching, This is an authoritative source indicator function. A function to weight the timeliness of documents. These are the weighting coefficients optimized through grid search.

[0041] S3.2.2: Perform hybrid retrieval; Based on a hierarchical domain knowledge base, the multidimensional retrieval scoring function described above is applied to calculate the comprehensive score between each knowledge fragment in the knowledge base and the user query. The retrieval process executes two paths in parallel: vector semantic retrieval and keyword retrieval. The semantic retrieval part uses the FAISS index to quickly locate candidate fragments with similar semantics, while the keyword retrieval part uses the Elasticsearch inverted index to locate candidate fragments with exact word matching. The candidate sets returned by the two paths are then uniformly reordered by the multidimensional scoring function.

[0042] S3.2.3: Filter highly matching knowledge fragments; based on the comprehensive score calculated by the multidimensional scoring function, sort all candidate knowledge fragments from high to low score, and select the top-10 knowledge fragments as input for the subsequent fusion process.

[0043] S3.2.4: Dynamically weighted fusion of evidence; a gated attention mechanism is used to integrate the top-10 knowledge fragments. The weight value depends on the relevance of the fragment to the current query and its own authority and timeliness score. Then, the vector representations of each fragment are weighted and summed according to these weights to integrate them into coherent knowledge evidence.

[0044] S3.2.5: Generate enhanced prompts; map the fused integrated knowledge evidence vector into a text-based contextual knowledge summary, and then concatenate it with the original user input to form a structured enhanced prompt, providing authoritative standard basis for subsequent scheme generation.

[0045] Among them, such as Figure 5 As shown, step S3.3 includes:

[0046] S3.3.1: Unified Context Coding and Multitasking Framework; Given Enhanced Input By encoding it using a shared Transformer encoder, a unified context representation is obtained. The calculation formula is: (3) Based on this, the construction plan generation is decomposed into four condition generation tasks: construction sequence, equipment list, safety measures, and risk assessment. Each task is equipped with a dedicated decoder, which generates its own output for each token in an autoregressive manner. Each decoder uses a shared context representation when predicting the current token. Using previously generated content as a condition, the calculation formula is as follows: (4) in These correspond to the four tasks mentioned above; All tasks are jointly optimized under a weighted multi-task learning objective. The total loss is the weighted sum of the negative log-likelihood losses of each task, calculated using the following formula: (5) in For the task The negative log-likelihood, Assign task weights; S3.3.2: Construction sequence generation and coverage awareness mechanism; a construction sequence generation is achieved using a sequence decoder; coverage awareness attention is introduced during the generation process, and a coverage vector is constructed by accumulating historical attention distributions to record the content of interest in the generated steps. The calculation formula is as follows: (6) in Indicates the first Attention weight distribution for each generation step.

[0047] Based on this, a coverage penalty term is added to the loss function. A penalty is applied to the overlapping portion between the attention weights and the coverage vector, guiding the decoder to pay more attention to knowledge content that has not yet been fully covered in subsequent generation steps. The calculation formula is as follows: (7) The total loss function of the construction process decoder is composed of a weighted combination of the basic generation loss and the coverage penalty, and the computer formula is as follows: (8) in The standard negative log-likelihood loss generated for the process. The weighting coefficients for the penalty terms are used to adjust the strength of the diversity constraints.

[0048] S3.3.3: Joint generation of the equipment list; the equipment list generation adopts a strategy combining multi-label classification prediction and conditional quantity generation; the required equipment category is predicted after sequence pooling of the shared encoder output through the multi-label classification head, and the calculation formula is as follows: (9) in express The probability of selecting a particular equipment type For the context representation of the shared encoder output, For sequence pooling operations, For learnable weight matrix, This is the Sigmoid activation function.

[0049] After determining the equipment category, based on the selected category set and the generated construction procedure sequence, the specific model and technical parameter description of each type of equipment are further generated.

[0050] S3.3.4: Generation of conditions for safety measures and risk assessment; the generation of safety measures and risk assessment both use the output construction procedures and equipment list as prerequisites, and are modeled as follows: and During the decoding process, security measures and risk assessments are strictly anchored to specific operating steps and equipment usage scenarios to ensure that each critical operation has corresponding security protection measures and risk control recommendations.

[0051] S3.3.5: Hierarchical Constraint Decoding; Using constraint-aware decoding to ensure domain compliance of generated content, defining compliance constraints including terminology. Compliance constraints on process sequence and semantic security compliance constraints Hierarchical constraint set The formula is defined as follows: (10) At each step of decoding, the range of tokens is constrained within the compliant vocabulary based on the currently generated prefix dynamic limit set. The calculation formula is as follows: (11) in This is a complete list of constraint terms.

[0052] Subsequently, a constraint bundle search is performed within the feasible token space, employing a delayed constraint activation strategy to balance the fluency of the initial text generation with the strict compliance of the later content. The calculation formula is as follows: (12) in To constrain the penalty function for violations, it is taken as infinity for hard constraints. Hyperparameters used to control the strength of constraint enforcement.

[0053] S3.3.6: Output candidate solutions; After the above multi-task decoding and hierarchical constraint verification, the decoder finally outputs a set of several complete, compliant and executable construction solution candidates.

[0054] Among them, such as Figure 6 As shown, step S3.4 includes:

[0055] S3.4.1: Module Architecture; A hierarchical Transformer architecture is adopted as the basic encoder to synchronously encode the candidate construction scheme text and its contextual input information, so as to make full use of the complete semantic environment of the generation stage; During the encoding process, the token-level fine-grained representation is aggregated into sentence-level embeddings layer by layer through hierarchical attention pooling operations, and then further aggregated into paragraph-level embeddings to form an effective capture of the hierarchical structure of the scheme document, enabling the model to understand the quality of the scheme from local details to the overall structure at multiple levels.

[0056] S3.4.2: Five-Dimensional Scoring System; Five independent evaluation dimensions are set to comprehensively reflect the quality of the construction plan, namely: Terminology Accuracy (assessing whether the use of professional terminology is standardized and consistent), Process Compliance (assessing whether the arrangement of procedures conforms to technical specifications), Safety Integrity (assessing whether safety measures cover all risk points and are sufficient), Operability (assessing the feasibility and clarity of the plan at the on-site execution level), and Document Standardization (assessing the format structure, chapter completeness, and rigor of the plan's expression); Scoring for each dimension. The range of values ​​is The total reward is predicted by the module using a multi-task regression head. The formula is: The scores from each dimension are weighted and summed. (13) in For the first Scoring across multiple dimensions These are the learnable weight coefficients for the corresponding dimensions.

[0057] S3.4.3: Joint Training Loss Function; The training objective of a module is composed of three loss functions, calculated as follows: (14) in The list sorting loss is used to incorporate expert preferences; The regression loss is used to fit the expert-annotated scores; This is a consistency regularization term used to enhance the stability of the evaluation. and This represents the balance coefficient for each loss term.

[0058] S3.4.4: Ranking and Regression Loss; List Ranking Loss The optimal solution selected by experts from multiple candidate solutions. Other candidate solutions A comparison list is constructed, and Softmax normalization is used to maximize the probability of the model assigning a score to the optimal solution. The calculation formula is as follows: (15) in The model scores the quality of the proposed solution.

[0059] Regression loss For schemes with expert annotation and scoring, mean squared error is used for supervision, and the calculation formula is as follows: (16) in For experts to review the first The overall score of each solution's annotation. This loss term is used to predict the model's score, ensuring that the model's scoring scale aligns with the expert scoring system.

[0060] S3.4.5: Consistency Regularization and Evaluation Output; To ensure that the model provides similar quality assessments for semantically similar construction schemes and to avoid abnormal fluctuations in scores due to minor input perturbations, a consistency regularization term is introduced. The calculation formula is: (17) in and For semantically similar scheme pairs, This is the threshold for the maximum allowed score difference.

[0061] This regularization term only penalizes score differences that exceed the tolerance range, effectively suppressing instability while maintaining assessment sensitivity.

[0062] Step S4 includes:

[0063] S4.1: Input the construction parameter samples from the training set into the model in batches according to the preset batch size. Each sample includes tower type, terrain, equipment type and construction method, as well as the corresponding expert-written reference construction plan. S4.2: The input parsing module extracts entities and classifies scene intents from construction parameters to obtain a structured parameter set and scene tags; based on hybrid scoring retrieval, a preset number of authoritative knowledge fragments are obtained from the knowledge base, which are then combined with the original input after gating attention fusion to generate enhanced prompts; S4.3: Enhance the input prompt multi-task constraint generation module, generate a unified context representation through a shared Transformer encoder, and then have each task's dedicated decoder generate four types of solution components in sequence under the constraint-aware decoding mechanism: construction procedures, equipment list, safety measures, and risk assessment. S4.4: To minimize the weighted sum of regression loss, ranking loss, and consistency regularization term, supervised learning is used to train the multi-dimensional reward modeling module. After each preset number of training epochs, the consistency between the overall reward score output by the current module and the expert-annotated score is calculated using the validation set. When the consistency on the validation set no longer increases after a preset number of consecutive iterations, the first stage of training is stopped, and the parameters of the reward module with the highest consistency are saved; otherwise, training continues. S4.5: Load and freeze the parameters of the reward module trained in the first stage. With the goal of maximizing the overall reward, train the construction scheme generation model using a reinforcement learning algorithm. After each preset number of training cycles, use the validation set to calculate the overall reward score obtained by the construction scheme generated by the current model. When the overall reward score on the validation set no longer increases for a preset number of consecutive cycles, stop training and save the model parameters with the highest overall reward score as the final model. Otherwise, update the network parameters and return to step 4.5. Step S5 includes: S5.1: Input the construction parameter samples from the test set into the final trained model in batches. Each sample includes tower type, terrain and equipment type. S5.2: The model completes the entire process of input parsing, knowledge retrieval, multi-task constraint generation and constraint-aware decoding, and outputs a structured construction plan containing four components: construction procedures, equipment list, safety measures and risk assessment. S5.3: The generated construction plan is sent to the multi-dimensional reward modeling module and scored according to five dimensions: terminology accuracy, process compliance, safety integrity, operability and document standardization. The overall reward is calculated as a comprehensive evaluation index of the plan quality. S5.4: Output the final construction plan document and its five-dimensional quality assessment report.

[0064] Table 1 shows the performance comparison results of the generation method of this application with other generative large models on the dataset.

[0065] Table 1

[0066] Example 2

[0067] This embodiment relates to an automatic generation device for transmission tower construction schemes based on retrieval enhancement and constraint generation, including a memory and one or more processors. The memory stores executable code, and when the one or more processors execute the executable code, they are used to implement the automatic generation method for transmission tower construction schemes based on retrieval enhancement and constraint generation described in Embodiment 1.

[0068] Example 3

[0069] A computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the automatic generation method for power transmission tower construction schemes based on retrieval enhancement and constraint generation as described in Embodiment 1.

Claims

1. An automatic generation method for transmission tower construction schemes based on retrieval enhancement and constraint generation, including: Step 1: Collect standards, specifications, equipment manuals, and historical solutions in the field of power transmission tower construction, and form a hierarchical domain knowledge base through standardization and indexing operations; Step 2: Collect real construction plans as sample data, and after preprocessing, construct a dataset, which is then divided into training set, validation set and test set according to a set ratio; Step 3: Construct the Transmission Tower Construction Scheme Generation Model TTCSGM based on retrieval enhancement and constraint generation; The structure of the TTCSGM model is as follows: starting from the input parsing module, it is sequentially connected to the domain adaptive hybrid knowledge retrieval module, the multi-task constraint generation and hierarchical compliance decoding module, and the multi-dimensional reward modeling module; among them, the input parsing module extracts construction parameters and scene categories; the domain adaptive hybrid knowledge retrieval module performs four-dimensional hybrid retrieval based on a hierarchical domain knowledge base, and the retrieval score integrates semantic similarity, word matching, source authority, and document timeliness, and uses a gated attention mechanism to dynamically weight and fuse multiple knowledge fragments obtained from the retrieval to generate enhanced prompts; the multi-task constraint generation and hierarchical compliance decoding module is based on a shared encoder and a multi-task decoder to conditionally generate construction procedures, equipment lists, safety measures, and risk assessments; the multi-dimensional reward modeling module outputs scores based on five dimensions: terminology accuracy, process compliance, safety integrity, operability, and document standardization, and optimizes the module parameters through joint training of list sorting loss, regression loss, and consistency regularization term; The input parsing module adopts a Transformer-based dual-encoder architecture. The construction parameter encoder is a BERT model fine-tuned from the power engineering corpus to perform entity extraction, and the scene encoder performs scene intent classification. The domain-adaptive hybrid knowledge retrieval module consists of a knowledge base encoder, a hybrid retrieval scorer, and a gated fusion unit connected in series. The knowledge base encoder uses Sentence-BERT output vectors based on a Siamese network architecture. The hybrid retrieval scorer integrates semantic similarity, BM25 matching, learnable scores based on metadata, and exponential decay timeliness scores. The gated fusion unit uses a graph attention network combined with a fully connected layer to perform weighted fusion of retrieval fragments. The multi-task constraint generation and hierarchical compliance decoding module consists of a shared Transformer encoder and four task-specific decoders. The process decoder introduces a coverage vector to calculate the coverage loss, the equipment list decoder connects to a multi-label classification head, and the cross-attention input of the safety and risk decoders originates from the output projections of the process and equipment decoders. The multi-dimensional reward modeling module consists of a hierarchical Transformer encoder combined with an attention pooling layer. It outputs a document-level representation and feeds it into five independent multilayer perceptron heads. Each head is processed by a fully connected layer and an activation function to output a score in the [0,1] interval. Step 4: Input the training set into the power transmission tower construction scheme generation model and perform two-stage training: In the first stage, supervised learning is used to train the multi-dimensional reward modeling module with the goal of minimizing the weighted sum of regression loss, ranking loss, and consistency regularization term; in the second stage, reinforcement learning is used to train the construction scheme generation model with the goal of maximizing the overall reward output by the multi-dimensional reward modeling module, and the parameters of the trained multi-dimensional reward modeling module are frozen to obtain the trained power transmission tower construction scheme generation model. Step 5: Input the test set into the trained power transmission tower construction scheme generation model, and automatically generate construction schemes corresponding to different construction conditions in the test set.

2. The automatic generation method for transmission tower construction schemes based on retrieval enhancement and constraint generation as described in claim 1, characterized in that, Step 1 specifically includes: Step 1.1 Collect national standards, power industry standards, equipment operation manuals, technical manuals, historical construction plans and typical construction methods in the field of power transmission tower construction, and clean the collected raw text data to remove duplicate, outdated and invalid content. Step 1.2 The cleaned text data is hierarchically split into chapters and clauses to form independent knowledge units. The knowledge units are labeled with types, including standard types, equipment parameters, construction procedures, safety measures and risk points. Step 1.3 uses the domain-adapted Sentence-BERT model to encode knowledge units and generate 768-dimensional vector representations; Step 1.4 Construct a knowledge structure using a graph neural network: Build a hierarchical index with knowledge units as nodes and relationships as edges, and assign timeliness weights and authority weights to each knowledge unit; Step 1.5 Construct semantic vector indexes and keyword indexes to build a vector knowledge base that supports hybrid retrieval.

3. The automatic generation method for transmission tower construction schemes based on retrieval enhancement and constraint generation as described in claim 1, characterized in that, The generation of the equipment list in step 2 is based on the already generated construction procedures, and the generation of safety measures and risk assessments is based on the construction procedures and equipment list. During the generation process, an overlay-aware attention mechanism is introduced to reduce the repetition of procedures, and a hierarchical constraint decoding is adopted. At each step of decoding, the set of feasible tokens is dynamically restricted to achieve real-time enforcement of terminology compliance, procedure sequence compliance, and semantic security compliance.

4. The automatic generation method for transmission tower construction schemes based on retrieval enhancement and constraint generation as described in claim 3, characterized in that, Step 2 specifically includes: Step 2.1 Collect real transmission tower construction project plan data from large power infrastructure companies over the past ten years, and remove sensitive identifying information after anonymization; Step 2.2 Filter out effective solutions covering various tower models, terrain types, and voltage levels to construct the original dataset; Step 2.3 The selected construction plan is divided into four modules according to the standard structure: construction procedures, equipment list, safety measures and risk assessment. The processed dataset is then divided into training set, validation set and test set according to the preset ratio.

5. The automatic generation method for transmission tower construction schemes based on retrieval enhancement and constraint generation as described in claim 1, characterized in that... Step 3 includes: Step 3.1: Construct an input parsing module to extract construction parameters and scene categories; Step 3.2: Construct a domain-adaptive hybrid knowledge retrieval module, which performs hybrid retrieval and evidence fusion based on a hierarchical domain knowledge base, and outputs enhanced hints; Step 3.3: Construct a multi-task constraint generation and hierarchical compliance decoding module, which uses a shared encoder and a multi-task decoder to jointly generate construction procedures, equipment lists, safety measures and risk assessments, and applies three levels of constraints on terminology, procedure sequence and semantic security during the decoding process; Step 3.4: Construct a multi-dimensional reward modeling module to score and select the best construction plan from five dimensions: terminology accuracy, process compliance, safety integrity, operability, and document standardization.

6. The automatic generation method for transmission tower construction schemes based on retrieval enhancement and constraint generation as described in claim 5, characterized in that, Step 3.1 specifically includes: Step 3.1.1: Define output parameters and labels; the output parameter set includes tower model, terrain, equipment type, construction method and voltage level, and scene labels are used to constrain the engineering scope of subsequent knowledge retrieval and solution generation; Step 3.1.2: Construct a dual encoder architecture; adopt a Transformer-based dual encoder to perform entity extraction and scene intent classification respectively; Step 3.1.3: Encoding Input; Input the construction requirements, which are then encoded by dual encoders to output the construction parameters and scene labels.

7. The automatic generation method for transmission tower construction schemes based on retrieval enhancement and constraint generation as described in claim 5, characterized in that, Step 3.2 specifically includes: Step 3.2.1: Construct a multidimensional retrieval scoring function; the multidimensional retrieval scoring function integrates four dimensions: semantic similarity, BM25 word matching, source authority, and document timeliness, and optimizes the weight coefficients of each dimension through grid search; Step 3.2.2: Perform hybrid retrieval; based on the hierarchical domain knowledge base, apply the above multidimensional retrieval scoring function to retrieve and score knowledge fragments in the knowledge base; Step 3.2.3: Filter highly matching knowledge fragments; sort the knowledge fragments according to the search score and select the top 10 knowledge fragments with the highest scores; Step 3.2.4: Dynamically weighted fusion of evidence; The selected 10 knowledge fragments are dynamically weighted and fused using a gated attention mechanism to integrate them into coherent knowledge evidence; Step 3.2.5: Generate enhanced prompts; combine the fused knowledge evidence with the original user input to form enhanced prompts, providing authoritative standard basis for subsequent scheme generation.

8. The automatic generation method for transmission tower construction schemes based on retrieval enhancement and constraint generation as described in claim 5, characterized in that, Step 3.3 specifically includes: Step 3.3.1: Unified context encoding and multi-task framework; given the enhanced input, a unified context representation is obtained through a shared Transformer encoder; the construction scheme generation is decomposed into four tasks: construction procedures, equipment list, safety measures and risk assessment. Each task is generated in an autoregressive manner using a dedicated decoder and jointly optimized under a weighted multi-task objective. Step 3.3.2: Construction sequence generation and coverage awareness mechanism; use the sequence decoder to generate the construction sequence; introduce coverage awareness attention during the generation process, construct the coverage vector by accumulating the historical attention distribution, and record the content that has been paid attention to in the generated steps; at the same time, add a coverage penalty term to the loss function to guide the model to pay attention to the construction sequence that has not been fully covered, and avoid repeated generation. Step 3.3.3: Joint generation of equipment list; The equipment list is generated by first predicting the required equipment categories through a multi-label classification head, and then generating the corresponding equipment models and quantity descriptions based on the selected categories and the generated construction procedures. Step 3.3.4: Generation of conditions for safety measures and risk assessment; The generation of safety measures and risk assessment is based on the output of construction procedures and equipment lists, realizing logical alignment between components, and generating safety protection measures and risk control suggestions corresponding to the construction procedures; Step 3.3.5: Hierarchical Constraint Decoding; Constraint-aware decoding is used to ensure the domain compliance of the generated content. A hierarchical constraint set is defined, which includes terminology compliance constraints, process sequence compliance constraints, and semantic security compliance constraints. At each step of decoding, the range of tokens is constrained to the compliance vocabulary based on the currently generated prefix dynamic restriction feasible token set. Constraint bundle search is used for decoding, and a delayed constraint activation strategy is introduced to balance fluency and strict compliance. Step 3.3.6: Output candidate solutions; the final decoder outputs a complete, compliant and executable set of construction solution candidates.

9. The automatic generation method for transmission tower construction schemes based on retrieval enhancement and constraint generation as described in claim 5, characterized in that, Step 3.4 specifically includes: Step 3.4.1: Module architecture; A hierarchical Transformer architecture is adopted to synchronously encode text and context. Token-level representations are aggregated layer by layer into sentence-level and paragraph-level embeddings through hierarchical attention pooling to capture the document's hierarchical structure. Step 3.4.2: Five-dimensional scoring system; five independent evaluation dimensions are set: terminology accuracy, process compliance, security integrity, operability, and document standardization. The score of each dimension ranges from 0 to 1; the overall reward is obtained by weighted sum of the scores of each dimension, and the weight of each dimension is a learnable parameter. Step 3.4.3: Joint training loss function; the module training objective is composed of three loss terms: list sorting loss to incorporate expert preferences, regression loss to fit expert labeled scores, and consistency regularization term to enhance evaluation stability; the loss terms are coordinated by a balancing coefficient. Step 3.4.4: Ranking and Regression Loss; The list ranking loss assigns a higher probability to the expert-preferred solution than to other candidate solutions; The regression loss uses mean squared error to align the model's predicted scores with the expert-labeled values. Step 3.4.5: Consistency Regularization and Evaluation Output; Constraints are imposed on the scoring differences between semantically similar construction schemes, and a maximum allowable threshold is set to suppress abnormal fluctuations; Through multi-objective joint training, the module can stably output evaluation results that meet expert standards, guiding the generation module to output construction schemes.

10. The automatic generation method for transmission tower construction schemes based on retrieval enhancement and constraint generation as described in claim 1, characterized in that, Step 4 includes: Step 4.1: Input the construction parameter samples from the training set into the model in batches according to the preset batch size. Each sample includes tower type, terrain, equipment type and construction method, as well as the corresponding expert-written reference construction plan. Step 4.2: Extract entities and classify scene intents from construction parameters through the input parsing module to obtain a structured parameter set and scene tags; obtain a preset number of authoritative knowledge fragments from the knowledge base based on hybrid scoring retrieval, and then concatenate them with the original input after gating attention fusion to generate enhanced prompts; Step 4.3: The enhanced prompts are input into the multi-task constraint generation module, which generates a unified context representation through the shared Transformer encoder. Then, each task's dedicated decoder generates four types of solution components in sequence under the constraint-aware decoding mechanism: construction procedures, equipment list, safety measures, and risk assessment. Step 4.4: With the goal of minimizing the weighted sum of regression loss, ranking loss, and consistency regularization term, supervised learning is used to train the multi-dimensional reward modeling module. After each preset number of training epochs, the consistency between the overall reward score output by the current module and the expert-annotated score is calculated using the validation set. When the consistency on the validation set no longer increases after a preset number of consecutive iterations, the first stage of training is stopped, and the parameters of the reward module with the highest consistency are saved; otherwise, training continues. Step 4.5: Load and freeze the parameters of the reward module trained in the first stage. With the goal of maximizing the overall reward, train the construction scheme generation model using a reinforcement learning algorithm. After each preset number of training cycles, use the validation set to calculate the overall reward score obtained by the construction scheme generated by the current model. When the overall reward score on the validation set no longer increases after a preset number of consecutive cycles, stop training and save the model parameters with the highest overall reward score as the final model. Otherwise, update the network parameters and return to step 4.

5.

11. The automatic generation method for transmission tower construction schemes based on retrieval enhancement and constraint generation as described in claim 1, characterized in that, Step 5 includes: Step 5.1: Input the construction parameter samples from the test set into the final trained model in batches. Each sample includes tower type, terrain and equipment type. Step 5.2: The model completes the entire process of input parsing, knowledge retrieval, multi-task constraint generation and constraint-aware decoding, and outputs a structured construction plan containing four components: construction procedures, equipment list, safety measures and risk assessment. Step 5.3: Submit the generated construction plan to the multi-dimensional reward modeling module, score it according to five dimensions: terminology accuracy, process compliance, safety integrity, operability, and document standardization, and calculate the overall reward as a comprehensive evaluation index of the plan quality; Step 5.4: Output the final construction plan document and its five-dimensional quality assessment report.

Images