An AI technology consultation service system based on deep learning

Abstract

The application discloses an AI technical consultation service system based on deep learning, which comprises a data acquisition and coding module, a structure constraint latent variable module, a structured decoding module, a chapter structure construction module, a chapter generation module, a consistency checking module and a regenerating control module.The data acquisition and coding module acquires multi-source data in the engineering consultation field and performs text coding, image block coding and table unit coding to form a multi-modal input sequence.The structure constraint latent variable module inputs the multi-modal input sequence into an improved Perceiver IO model and outputs a corrected latent variable representation.The structured decoding module generates a parameter matrix, a parameter relationship set and a chapter slot binding result.The chapter structure construction module constructs a chapter structure constraint graph and determines a chapter corresponding parameter set and a clause set.The chapter generation module generates chapter texts.The consistency checking module outputs a conflict chapter identifier.The regenerating control module outputs a consultation report text.The application improves the structural consistency and numerical accuracy of the technical consultation report.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence and information processing technology, and in particular to an AI technology consulting service system based on deep learning. Background Technology

[0002] With the rapid development of artificial intelligence technology and the ongoing digital transformation of the engineering consulting industry, utilizing large language models and deep learning techniques to assist in the generation of engineering consulting reports has gradually become a research hotspot. Existing technologies typically rely on general-purpose large language models or retrieval-enhanced generation frameworks, using policy texts, standards and specifications, and historical cases as external knowledge input to generate report text through prompting engineering or vector retrieval methods. However, in actual engineering consulting scenarios, report content often involves a large number of structured parameters, chapter-level dependencies, and clause citation constraints, highlighting the significant shortcomings of existing technologies.

[0003] The general model lacks the structural modeling capability for engineering consulting standard outlines, often resulting in generated texts with chaotic chapter hierarchy or uneven parameter distribution. Secondly, existing methods mainly rely on simple, piecemeal input prompts, lacking a unified latent space modeling and structural consistency constraint mechanism for multimodal data, making it difficult to synchronously constrain chapter hierarchy relationships, parameter calculation relationships, and clause coverage relationships at the latent variable level. Thirdly, the generation process lacks a parameter relationship-driven structural control mechanism, making the model prone to numerical deviations, parameter omissions, or incorrect combination relationships, and most do not establish a computable verification path oriented towards engineering parameter conservation relationships. Furthermore, existing report generation systems typically separate the generation and verification processes, lacking a local regeneration control strategy based on conflicting chapters, resulting in the need for complete regeneration when some chapters contain errors, leading to significant computational resource consumption and insufficient stability.

[0004] Therefore, how to provide an AI technology consulting service system based on deep learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an AI technology consulting service system based on deep learning. This invention constructs a multimodal input representation system, combines an improved Perceiver IO model with a structural consistency tensor field constraint mechanism, to achieve unified modeling of chapter hierarchy, parameter relationships, and clause coverage at the latent variable level; it achieves structurally constrained chapter text generation through multi-query structured decoding and parameter relationship-driven generation control; and it constructs a closed-loop mechanism for numerical verification and conflict location regeneration to improve the structural consistency, numerical accuracy, and generation stability of engineering consulting reports.

[0006] According to an embodiment of the present invention, an AI technology consulting service system based on deep learning includes:

[0007] The data acquisition and encoding module is used to acquire multi-source data in the field of engineering consulting, and perform text encoding, image block encoding and table cell encoding respectively, and output multimodal input sequences;

[0008] The structural constraint latent variable module is used to input the multimodal input sequence into the improved Perceiver IO model, introduce a structural consistency tensor field constraint mechanism during the latent variable update process, perform structural consistency correction on the latent variable representation, and output the corrected latent variable representation.

[0009] The structured decoding module is used to decode based on the modified latent variable representation and output a parameter matrix, a set of parameter relationships, and chapter slot binding results;

[0010] The chapter structure construction module is used to construct a chapter structure constraint diagram based on the engineering consulting standard outline, and to determine the parameter set and clause set corresponding to each chapter node based on the chapter slot binding result;

[0011] The chapter generation module is used to input the parameter set and clause set of the corresponding chapter node into the locally deployed and fine-tuned large language model according to the node order of the chapter structure constraint diagram, and generate the text of each chapter.

[0012] The consistency verification module is used to perform numerical consistency verification on the text of each chapter based on the parameter relationship set, and to perform clause matching verification based on the clause coverage relationship, and output conflicting chapter identifiers;

[0013] The regeneration control module is used to call the chapter generation module to regenerate the corresponding chapter text based on the conflicting chapter identifier and output the consultation report text.

[0014] Optionally, the data acquisition and encoding module includes:

[0015] Data type identification is performed on multi-source data in the field of engineering consulting, and the multi-source data in the field of engineering consulting is divided into text data, image data and tabular data;

[0016] The text data is segmented into semantic word sequences. Each semantic word is mapped into a vector, and a position code is generated according to the order of the word in the original text. The semantic word vector is combined with the corresponding position code to form a text vector sequence.

[0017] The image data is divided into blocks according to a preset size, the image pixel matrix is ​​divided into image blocks of a preset size, a linear mapping is performed on each image block to obtain an image block vector, and a two-dimensional position code is generated according to the row and column position of the image block in the original image. The image block vector and the two-dimensional position code are combined to form an image vector sequence.

[0018] The table data is traversed by row index and column index, word segmentation vector mapping is performed on the text content in the cell, and table position code is generated according to the row index and column index. The cell vector and the table position code are combined to form a table vector sequence.

[0019] The text vector sequence, image vector sequence, and table vector sequence are concatenated in a preset order to generate a multimodal input sequence.

[0020] Optionally, the structural constraint latent variable module includes:

[0021] The multimodal input sequence is input into the improved Perceiver IO model. The improved Perceiver IO model sets a preset number of latent variable arrays. The latent variable arrays and the multimodal input sequence interact with each other through a cross-attention structure. In each cross-attention layer, the latent variable array is used as the query vector, and the multimodal input sequence is used as the key vector and value vector to calculate the attention weights. The value vectors are then weighted and aggregated to obtain the intermediate latent variable representation.

[0022] The improved Perceiver IO model sets up a latent variable partitioning structure, dividing the latent variable array into a structural latent variable subarray, a parameter latent variable subarray, and a clause latent variable subarray. Each subarray corresponds to chapter-level information, parameter calculation information, and clause coverage information, respectively.

[0023] In the latent variable update process of the improved Perceiver IO model, a structural consistency tensor field constraint mechanism is introduced. The structural consistency tensor field consists of a set of third-order tensors, including chapter-level tensors, parameter conservation tensors, and clause coverage tensors.

[0024] The chapter hierarchy tensor establishes an index mapping based on the hierarchical relationship between parent and child nodes in the engineering consulting standard outline, and records the association identifier for chapter node pairs with hierarchical dependencies.

[0025] The parameter conservation tensor establishes an index mapping based on the calculation combination relationship between engineering parameters, and records the combination identifier for parameter sets participating in the same calculation relationship;

[0026] The clause coverage tensor establishes an index mapping based on the reference correspondence between chapters and policy clauses, and records coverage identifiers for chapter nodes and their corresponding clause sets.

[0027] After the intermediate latent variable representations are generated at each level, the intermediate latent variable representations are mapped to the structural latent variable subarray, the parameter latent variable subarray, and the clause latent variable subarray, respectively. For each subarray, a consistency scan operation is performed according to the index mapping relationship in the corresponding tensor to identify the latent variable positions that do not satisfy the hierarchical dependency, parameter combination, and clause coverage relationships.

[0028] A structural rearrangement operation is performed on the identified inconsistent latent variable locations, remapping the relevant latent variables to a structurally feasible subspace that satisfies the common definition of the chapter-level tensor, parameter conservation tensor, and clause coverage tensor, to obtain the corrected latent variable representation.

[0029] Optionally, the structured decoding module includes:

[0030] Attention is calculated for each parameter query vector and the modified latent variable representation in the parameter query vector set to obtain the corresponding parameter embedding vector, and then arranged in a preset parameter order to form a parameter matrix;

[0031] Attention is calculated for each relation query vector in the relation query vector set and the modified latent variable representation to obtain the association score vector between parameter pairs. The parameter index pairs are then filtered based on the association score to form a parameter relation set.

[0032] Attention is calculated for each chapter slot query vector and the modified latent variable representation in the chapter slot query vector set to obtain the chapter embedding vector. Based on the similarity calculation results between the chapter embedding vector and the parameter embedding vector, the parameter index is assigned to the corresponding chapter slot to form the chapter slot binding result.

[0033] Optionally, the chapter structure construction module includes:

[0034] Read the engineering consulting standard outline document, extract the chapter number, chapter name and chapter level identifier, parse the membership relationship between parent and child nodes according to the hierarchical coding rules in the chapter number, and establish a set of chapter nodes;

[0035] Based on the membership relationship between parent and child nodes, directed edges are constructed in the chapter node set to form a chapter structure constraint graph, in which the upper-level chapter node points to the lower-level chapter node;

[0036] Match the chapter index in the chapter slot binding result with the chapter node set, establish a mapping relationship between chapter nodes and parameter indexes, and convert the mapped parameter index set into the parameter set of the corresponding chapter node;

[0037] Based on the association identifier between the chapter index and the clause index recorded in the clause coverage tensor, the clause index is mapped to the corresponding chapter node, forming the clause set corresponding to the chapter node;

[0038] Write the parameter set and clause set into the node attributes of the corresponding chapter node to complete the construction of the chapter structure constraint diagram.

[0039] Optionally, the chapter generation module includes:

[0040] The chapter generation order is determined according to the topological sorting result of the chapter structure constraint diagram;

[0041] For the current chapter node, construct a parameter relationship subgraph based on the corresponding parameter set, and arrange the parameter nodes with computational associations in an ordered manner according to the connection order in the parameter relationship set to form a parameter relationship sequence;

[0042] The parameter relationship sequence is converted into a structured control tag sequence. Each parameter node generates a unique parameter identifier, and the parameter value and associated parameter index tag are appended to the identifier.

[0043] The clause numbers in the clause set are mapped to clause control tags, and then inserted into the specified positions in the parameter relationship sequence according to the clause order corresponding to the chapter node, forming the chapter control input sequence;

[0044] The chapter control input sequence is input into the locally deployed and fine-tuned large language model. During the generation process, parameter identifier matching is scanned on the output text, and resampling is performed on the corresponding positions of parameter identifiers that are not matched or have abnormal order.

[0045] After generating the text for the current chapter node, process each chapter node in the order of chapter generation to generate the text for each chapter.

[0046] Optionally, the consistency verification module includes:

[0047] The text of each chapter is segmented and scanned. The parameter names and values ​​in the text are extracted according to the preset parameter identifiers to form a set of chapter parameters.

[0048] The parameter values ​​in the chapter parameter set are compared with the corresponding parameter values ​​in the parameter matrix, and the parameter indices that differ are marked.

[0049] Based on the parameter index pairs recorded in the parameter relationship set, read the corresponding parameter values ​​in the chapter parameter set, perform the corresponding operations according to the combination order in the parameter relationship set, compare the operation results with the corresponding target parameter values ​​in the chapter text, and mark the parameter index pairs that do not satisfy the combination relationship;

[0050] Based on the clause coverage tensor in the structural consistency tensor field, the mapping relationship between the chapter index and the clause index is read, the clause numbers in the chapter text are compared one by one, and the missing clause indexes are marked.

[0051] The chapter indexes corresponding to the difference markers, combination markers, and missing markers are summarized to form conflict chapter identifiers.

[0052] Optionally, the regeneration control module includes:

[0053] Receive conflict chapter identifiers output by the consistency verification module, and determine the corresponding chapter node index based on the conflict chapter identifiers;

[0054] Read the parameter set and clause set of the corresponding chapter node in the chapter structure constraint diagram;

[0055] Based on the difference markers, combination markers, and missing markers recorded in the conflict chapter identifiers, the values ​​of the corresponding parameter indices in the parameter set are replaced with the corresponding values ​​in the parameter matrix. Then, based on the parameter relationship set, the parameter relationship subgraph is reconstructed for the parameter index pairs with combination markers to obtain the updated parameter set.

[0056] The updated parameter set and clause set are passed to the chapter generation module, which then regenerates the corresponding chapter nodes according to the node order in the chapter structure constraint diagram to obtain the regenerated chapter text.

[0057] The regenerated chapter text will replace the original chapter text record of the corresponding node in the chapter structure constraint diagram;

[0058] After correcting all conflicts in all chapter nodes, the text of each chapter is spliced ​​together according to the topological order in the chapter structure constraint diagram to form the engineering consulting report text.

[0059] The beneficial effects of this invention are:

[0060] This invention constructs a multimodal input representation system that integrates multi-source data from engineering consulting. It combines an improved Perceiver IO model with a structural consistency tensor field constraint mechanism to address the problems in existing technologies, such as difficulty in uniformly modeling chapter levels, susceptibility to deviations in parameter calculation relationships, and lack of structural constraints in clause citations. In the latent variable update stage, a structural consistency tensor field composed of chapter-level tensors, parameter conservation tensors, and clause coverage tensors is introduced to perform consistency scanning and structural rearrangement on the latent variable representation, ensuring that the latent space expression simultaneously satisfies chapter-level dependencies, parameter combination relationships, and clause coverage relationships. In the decoding stage, a multi-query structured decoding mechanism is constructed. Through parameter query vectors, relationship query vectors, and chapter slot query vectors, the modified latent variable representation is mapped to different channels, generating a parameter matrix, a set of parameter relationships, and chapter slots. The binding results achieve precise alignment between structured information and chapter structure. During chapter generation, a control tag sequence driven by a parameter relationship subgraph is introduced, embedding parameter identifiers and clause control tags into the generation process. Combined with parameter identifier matching and resampling strategies, this enables structure-constrained chapter text generation. In the verification phase, numerical calculations and clause matching verification are performed based on the parameter matrix, parameter relationship set, and clause coverage tensor. A generation-verification-regeneration closed loop is formed through conflict chapter location and local regeneration control mechanisms, effectively improving the stability and reliability of engineering consulting reports in terms of structural consistency, numerical accuracy, and clause coverage completeness. This achieves structured, computable, and intelligent generation of engineering consulting reports. Attached Figure Description

[0061] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0062] Figure 1 This is a schematic diagram of the structure of an AI technology consulting service system based on deep learning proposed in this invention;

[0063] Figure 2 This is a schematic diagram of the processing flow of the improved Perceiver IO model and the structural consistency tensor field constraint mechanism in this invention.

[0064] Figure 3 This is a flowchart illustrating the generation-verification-regeneration closed-loop control mechanism in this invention. Detailed Implementation

[0065] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0066] refer to Figures 1-3A deep learning-based AI technology consulting service system includes:

[0067] The data acquisition and encoding module is used to acquire multi-source data in the field of engineering consulting, and perform text encoding, image block encoding and table cell encoding respectively, and output multimodal input sequences;

[0068] The structurally constrained latent variable module is used to input multimodal input sequences into the improved Perceiver IO model. During the latent variable update process, a structural consistency tensor field constraint mechanism is introduced to correct the latent variable representation for structural consistency and output the corrected latent variable representation.

[0069] The structured decoding module is used for decoding based on the modified latent variable representation, and outputs the parameter matrix, parameter relationship set, and chapter slot binding results;

[0070] The chapter structure construction module is used to construct a chapter structure constraint diagram based on the engineering consulting standard outline, and to determine the parameter set and clause set corresponding to each chapter node based on the chapter slot binding result;

[0071] The chapter generation module is used to generate the text of each chapter by inputting the parameter set and clause set of the corresponding chapter node into the locally deployed and fine-tuned large language model according to the node order of the chapter structure constraint diagram.

[0072] The consistency verification module is used to perform numerical consistency verification on the text of each chapter based on the parameter relationship set, and to perform clause matching verification based on the clause coverage relationship, and output conflicting chapter identifiers;

[0073] The regeneration control module is used to call the chapter generation module to regenerate the corresponding chapter text based on the conflicting chapter identifier and output the consultation report text.

[0074] In this embodiment, the data acquisition and encoding module includes:

[0075] Data type identification is performed on multi-source data in the field of engineering consulting, and the multi-source data in the field of engineering consulting is divided into text data, image data and tabular data;

[0076] The text data is segmented into semantic word sequences. Each semantic word is mapped into a vector, and a position code is generated according to the order of the word in the original text. The semantic word vectors are combined with the corresponding position codes to form a text vector sequence.

[0077] The image data is divided into blocks according to a preset size. The image pixel matrix is ​​divided into image blocks of a preset size. A linear mapping is performed on each image block to obtain an image block vector. A two-dimensional position code is generated based on the row and column position of the image block in the original image. The image block vector and the two-dimensional position code are combined to form an image vector sequence.

[0078] The table data is traversed by row and column indices, word segmentation vector mapping is performed on the text content in the cells, and table position codes are generated based on row and column indices. The cell vectors and table position codes are combined to form a table vector sequence.

[0079] Text vector sequences, image vector sequences, and table vector sequences are concatenated in a preset order to generate a multimodal input sequence.

[0080] During data type identification, a dual determination is made by reading the file header format identifier and content field features. Text data is identified based on character encoding format, image data is identified based on pixel matrix format, and table data is identified based on row and column structure markers. Text segmentation uses a pre-trained vocabulary for sub-word segmentation, and positional encoding is generated by numerical mapping based on the sequential index of the word in the text. Image block size is divided according to the image resolution at a preset ratio, and two-dimensional positional encoding is generated based on the row and column coordinates of the block. Table positional encoding is generated based on the row and column index of the cell in the table. Multimodal concatenation is arranged in the order of text, image, and table types, and a data type identifier vector is attached.

[0081] In this embodiment, the structural constraint latent variable module includes:

[0082] The multimodal input sequence is input into the improved Perceiver IO model. The improved Perceiver IO model sets a preset number of latent variable arrays. The latent variable arrays and the multimodal input sequence interact with each other through a cross-attention structure. In each cross-attention layer, the latent variable array is used as the query vector, and the multimodal input sequence is used as the key vector and value vector to calculate the attention weights. The value vectors are then weighted and aggregated to obtain the intermediate latent variable representation.

[0083] The improved Perceiver IO model sets up a latent variable partitioning structure, dividing the latent variable array into a structural latent variable subarray, a parameter latent variable subarray, and a clause latent variable subarray. Each subarray corresponds to chapter-level information, parameter calculation information, and clause coverage information, respectively.

[0084] In the latent variable update process of the improved Perceiver IO model, a structural consistency tensor field constraint mechanism is introduced. The structural consistency tensor field consists of a set of third-order tensors, including chapter-level tensors, parameter conservation tensors, and clause coverage tensors.

[0085] The chapter hierarchy tensor establishes an index mapping based on the hierarchical relationship between parent and child nodes in the chapters of the engineering consulting standard outline, and identifies the associated records for chapter nodes with hierarchical dependencies.

[0086] The parameter conservation tensor establishes an index mapping based on the calculation combination relationship between engineering parameters, and records the combination identifier for parameter sets participating in the same calculation relationship;

[0087] The clause coverage tensor establishes an index mapping based on the reference correspondence between chapters and policy clauses, and records coverage identifiers for chapter nodes and their corresponding clause sets;

[0088] After the intermediate latent variable representations are generated at each level, the intermediate latent variable representations are mapped to the structural latent variable subarray, the parameter latent variable subarray, and the clause latent variable subarray, respectively. For each subarray, a consistency scan operation is performed according to the index mapping relationship in the corresponding tensor to identify the latent variable positions that do not satisfy the hierarchical dependency, parameter combination, and clause coverage relationships.

[0089] A structural rearrangement operation is performed on the identified inconsistent latent variable locations, remapping the relevant latent variables to a structurally feasible subspace that satisfies the common definition of the chapter-level tensor, parameter conservation tensor, and clause coverage tensor, to obtain the corrected latent variable representation.

[0090] The improved Perceiver IO model continues the basic framework of the Perceiver IO model in its overall structure. It uses a preset number of latent variable arrays as information compression carriers and realizes information interaction between the latent variable arrays and multimodal input sequences through a cross-attention structure. In each layer, the latent variable array is used as the query vector, and the input sequence is used as the key vector and value vector for attention calculation. The latent variables are updated through multi-layer stacking, and finally the latent variable representation with a unified dimension is output.

[0091] Unlike the original Perceiver IO model, the improved Perceiver IO model sets up a partitioning structure within the latent variable array, dividing the latent variables into structural latent variable subarrays, parameter latent variable subarrays, and clause latent variable subarrays. After each layer of latent variables is updated, a structural consistency tensor field constraint mechanism is introduced. The structural consistency tensor field consists of a chapter-level tensor, a parameter conservation tensor, and a clause coverage tensor. Through the latent variable index mapping relationship, consistency scanning and structural rearrangement processing are performed on the intermediate latent variable representations to generate corrected latent variable representations.

[0092] Through the above improvements, the latent variable update process is coupled with the chapter hierarchy, parameter calculation, and clause reference relationships of the engineering consulting report, thereby achieving structural constraint control at the latent space level, reducing interference from unstructured information, improving the stability of multimodal information fusion, and enhancing the structural consistency of subsequent decoding results.

[0093] In this embodiment, the structured decoding module includes:

[0094] Attention is calculated for each parameter query vector and the modified latent variable representation in the parameter query vector set to obtain the corresponding parameter embedding vector, and then arranged in a preset parameter order to form a parameter matrix;

[0095] Attention is calculated for each relation query vector in the relation query vector set and the modified latent variable representation to obtain the association score vector between parameter pairs. The parameter index pairs are then filtered based on the association score to form a parameter relation set.

[0096] Attention is calculated for each chapter slot query vector and the modified latent variable representation in the chapter slot query vector set to obtain the chapter embedding vector. Based on the similarity calculation results between the chapter embedding vector and the parameter embedding vector, the parameter index is assigned to the corresponding chapter slot to form the chapter slot binding result.

[0097] In this implementation, the number of parameter query vector sets in the multi-query decoding structure is set to 64, the number of relation query vector sets is set to 128, and the number of chapter slot query vector sets is consistent with the number of chapter nodes in the engineering consulting standard outline; each query vector dimension is set to 256; attention calculation adopts the scaling dot product method, and the calculated association score is normalized and compared with the preset threshold of 0.6, retaining parameter index pairs with a value greater than 0.6 to form parameter relation sets; the similarity between chapter embedding vectors and parameter embedding vectors is calculated by the vector dot product method, and sorted from high to low similarity, the top 3 parameter indices are selected as the binding results of the corresponding chapter slots.

[0098] In this embodiment, the chapter structure construction module includes:

[0099] Read the engineering consulting standard outline document, extract the chapter number, chapter name and chapter level identifier, parse the membership relationship between parent and child nodes according to the hierarchical coding rules in the chapter number, and establish a set of chapter nodes;

[0100] Based on the membership relationship between parent and child nodes, directed edges are constructed in the chapter node set to form a chapter structure constraint graph, in which the upper-level chapter node points to the lower-level chapter node;

[0101] Match the chapter index in the chapter slot binding result with the chapter node set, establish a mapping relationship between chapter nodes and parameter indexes, and convert the mapped parameter index set into the parameter set of the corresponding chapter node;

[0102] Based on the association identifier between the chapter index and the clause index recorded in the clause coverage tensor, the clause index is mapped to the corresponding chapter node, forming the clause set corresponding to the chapter node;

[0103] Write the parameter set and clause set into the node attributes of the corresponding chapter node to complete the construction of the chapter structure constraint diagram.

[0104] In this implementation, chapter numbering adopts a hierarchical coding format, such as "1", "1.1", and "1.1.1". The hierarchical depth is determined by counting the number of separators in the number, and the number after removing the last level code is used as the parent node index to establish a hierarchical mapping. The chapter structure constraint graph uses an adjacency list structure to store nodes and directed edge relationships. During the matching process between the chapter index and the parameter index, a one-to-one mapping is achieved by establishing an index lookup table. The clause index is quickly located through a pre-constructed chapter-clause mapping dictionary, and the parameter set and clause set are stored in the graph structure in the form of node attribute key-value pairs, realizing the computable expression of chapter structure data.

[0105] In this embodiment, the chapter generation module includes:

[0106] The chapter generation order is determined according to the topological sorting result of the chapter structure constraint diagram;

[0107] For the current chapter node, construct a parameter relationship subgraph based on the corresponding parameter set, and arrange the parameter nodes with computational associations in an ordered manner according to the connection order in the parameter relationship set to form a parameter relationship sequence;

[0108] The parameter relationship sequence is converted into a structured control tag sequence. Each parameter node generates a unique parameter identifier, and the parameter value and associated parameter index tag are appended to the identifier.

[0109] The clause numbers in the clause set are mapped to clause control tags, and then inserted into the specified positions in the parameter relationship sequence according to the clause order corresponding to the chapter node, forming the chapter control input sequence;

[0110] The chapter control input sequence is input into the locally deployed and fine-tuned large language model. During the generation process, parameter identifier matching is scanned on the output text, and resampling is performed on the corresponding positions of parameter identifiers that are not matched or have abnormal order.

[0111] After generating the text for the current chapter node, process each chapter node in the order of chapter generation to generate the text for each chapter.

[0112] In this implementation, the locally deployed and fine-tuned large language model is trained for domain adaptation based on historical report corpora in the engineering consulting field, regional policy texts, and industry standard texts. During the training process, chapter-level sample alignment data is introduced to enable the model to learn the correspondence between chapter titles, parameter expressions, and clause citation formats. The chapter control input sequence is segmented and encoded before being input into the model, and chapter boundary markers are kept out of probability normalization calculations during the model decoding stage. The resampling process limits the candidate lexical set to only include parameter identifiers and clause control markers associated with the current chapter node, thereby achieving structured generation for engineering consulting reports.

[0113] In this embodiment, the consistency verification module includes:

[0114] The text of each chapter is segmented and scanned. The parameter names and values ​​in the text are extracted according to the preset parameter identifiers to form a set of chapter parameters.

[0115] The parameter values ​​in the chapter parameter set are compared with the corresponding parameter values ​​in the parameter matrix, and the parameter indices that differ are marked.

[0116] Based on the parameter index pairs recorded in the parameter relationship set, read the corresponding parameter values ​​in the chapter parameter set, perform the corresponding operations according to the combination order in the parameter relationship set, compare the operation results with the corresponding target parameter values ​​in the chapter text, and mark the parameter index pairs that do not satisfy the combination relationship;

[0117] Based on the clause coverage tensor in the structural consistency tensor field, the mapping relationship between the chapter index and the clause index is read, the clause numbers in the chapter text are compared one by one, and the missing clause indexes are marked.

[0118] The chapter indexes corresponding to the difference markers, combination markers, and missing markers are summarized to form conflict chapter identifiers.

[0119] In this implementation, parameter identifiers are encoded using a combination of a preset prefix and parameter index, and are precisely extracted from the chapter text through regular expressions. During numerical comparison, the extracted parameter values ​​are uniformly converted to floating-point data and two decimal places are retained before item-by-item comparison. The parameter relationship set records both the operation type and the operation order identifier, and the operations are performed sequentially according to the record order during verification. The clause number matching adopts a full string comparison method to ensure that the numbers are completely consistent before determining the matching result, thereby realizing computable consistency verification.

[0120] In this embodiment, the regeneration control module includes:

[0121] Receive conflict chapter identifiers output by the consistency verification module, and determine the corresponding chapter node index based on the conflict chapter identifiers;

[0122] Read the parameter set and clause set of the corresponding chapter node in the chapter structure constraint diagram;

[0123] Based on the difference markers, combination markers, and missing markers recorded in the conflict chapter identifiers, the values ​​of the corresponding parameter indices in the parameter set are replaced with the corresponding values ​​in the parameter matrix. Then, based on the parameter relationship set, the parameter relationship subgraph is reconstructed for the parameter index pairs with combination markers to obtain the updated parameter set.

[0124] The updated parameter set and clause set are passed to the chapter generation module, which then regenerates the corresponding chapter nodes according to the node order in the chapter structure constraint diagram to obtain the regenerated chapter text.

[0125] The regenerated chapter text will replace the original chapter text record of the corresponding node in the chapter structure constraint diagram;

[0126] After correcting all conflicts in all chapter nodes, the text of each chapter is spliced ​​together according to the topological order in the chapter structure constraint diagram to form the engineering consulting report text.

[0127] In this implementation, the conflict chapter identifier is formed by combining the chapter index and the conflict type code to form a unique location identifier; during the parameter value replacement process, the target value in the parameter matrix is ​​directly located through the index lookup table and the corresponding position in the original parameter set is overwritten; when reconstructing the parameter relationship, the parameter relationship subgraph is regenerated based on the parameter index pairs and their operation order recorded in the parameter relationship set; chapter regeneration is only performed on conflict chapter nodes, and unmarked chapters retain their original text; during the splicing stage, preset chapter separators are inserted between chapters, and the text of each chapter is linearly connected according to the hierarchical order of the chapter structure constraint diagram.

[0128] Example 1: To verify the feasibility and effectiveness of this invention in a real-world engineering consulting scenario, it was applied to the preparation of a feasibility study report for a comprehensive energy project in a new energy industrial park in a certain province. This project includes a distributed photovoltaic power generation system, an energy storage system, and a power distribution renovation project. The report covers multiple chapters, including construction scale calculation, power load balance calculation, investment estimation, energy conservation evaluation, and policy compliance analysis, involving 28 policy documents, 15 industry standards, and over 300 engineering parameters. In traditional manual preparation methods, engineering consultants need to repeatedly search for data across different documents, manually calculate parameters, and integrate chapters, which easily leads to problems such as inconsistent parameters, missing clause references, and disorganized chapter structures. Especially after multiple revisions, numerical deviations and missing clauses become more prominent.

[0129] In this experiment, relevant policy texts, standards and specifications, historical case reports, engineering drawings, and structured calculation tables were first input into the data acquisition and encoding module. A unified multimodal input sequence was formed through text segmentation encoding, image block encoding, and table cell encoding. This multimodal input sequence was then input into an improved Perceiver IO model. During the latent variable update phase, a structural consistency tensor field, consisting of chapter-level tensors, parameter conservation tensors, and clause coverage tensors, was introduced. Consistency scanning and structural rearrangement were performed on the latent variable representations to ensure that the latent space representation simultaneously satisfies chapter-level dependencies, parameter calculation combination relationships, and clause coverage relationships. The structured decoding module further outputs a parameter matrix, a set of parameter relationships, and chapter slot binding results. The parameter matrix contains 312 engineering parameters, the set of parameter relationships contains 84 sets of calculation relationships, and the chapter slot binding results cover 16 first-level chapters and 42 second-level chapters in the standard outline.

[0130] The chapter structure construction module establishes a chapter structure constraint diagram based on the engineering consulting standard outline, and writes the parameter set and clause set into the corresponding chapter node attributes. The chapter generation module generates a control mark sequence based on the parameter relationship subgraph, embedding parameter identifiers and clause control marks into the generation process. During the generation process, identifier matching checks and resampling control are performed. After generation, the consistency verification module calculates and verifies the numerical consistency and clause coverage, identifies conflicting chapters, and hands them over to the regeneration control module for partial regeneration processing, ultimately forming a complete engineering consulting report text.

[0131] To verify the effectiveness of the method of the present invention, the system of the present invention was compared with the traditional manual compilation method and the general large model generation method without introducing structural consistency tensor field constraints. The generation efficiency, parameter consistency error rate, clause coverage completeness rate and chapter structure matching accuracy were statistically analyzed. The comparative experiments are shown in Table 1.

[0132] Table 1 Comparison of Consulting Report Generation Results for New Energy Projects

[0133] Comparison indicators Traditional manual weaving Generate general large models This invention system Average report generation time (hours) 26.5 4.8 3.2 Number of parameters 312 312 312 Number of times parameter values ​​are inconsistent 27 18 2 Parameter consistency error rate (%) 8.65 5.77 0.64 Error count in parameter combination calculation 19 14 1 The terms should cover the quantity 43 43 43 Number of missing clauses 6 4 0 Complete coverage rate of terms (%) 86.05 90.70 100 Chapter structure matching accuracy (%) 92.1 88.4 99.3 Average number of conflict chapters Not counted 7 1 Local regeneration count Not counted Not counted 1

[0134] As can be seen from the comparison results in Table 1 above, the present invention reduces report generation time by approximately 87.9% compared to traditional manual methods and by approximately 33.3% compared to ordinary large model generation methods; in terms of parameter consistency, the error rate is reduced from 8.65% to 0.64%; in terms of clause coverage, 100% coverage is achieved with no omissions; and in terms of chapter structure matching, an accuracy rate of 99.3% is achieved. Especially in scenarios involving multiple rounds of revisions, the present invention, through conflict chapter location and local regeneration mechanisms, only requires local regeneration of one chapter to complete the correction, while the traditional method requires overall modification, significantly reducing repetitive workload.

[0135] This embodiment demonstrates that by introducing structural consistency tensor field constraints at the latent variable layer, introducing a parameter relationship-driven control mechanism at the generation stage, and forming a generation-verification-regeneration closed-loop mechanism at the verification stage, the present invention effectively solves the problems of unstable chapter structure, error-prone parameter calculation, and incomplete clause citation during the generation of engineering consulting reports, thereby achieving structured, computable, and highly stable generation of engineering consulting reports.

[0136] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A deep learning-based AI technology consulting service system, characterized in that, include: The data acquisition and encoding module is used to acquire multi-source data in the field of engineering consulting, and perform text encoding, image block encoding and table cell encoding respectively, and output multimodal input sequences; The structural constraint latent variable module is used to input the multimodal input sequence into the improved Perceiver IO model, introduce a structural consistency tensor field constraint mechanism during the latent variable update process, perform structural consistency correction on the latent variable representation, and output the corrected latent variable representation. The structured decoding module is used to decode based on the modified latent variable representation and output a parameter matrix, a set of parameter relationships, and chapter slot binding results; The chapter structure construction module is used to construct a chapter structure constraint diagram based on the engineering consulting standard outline, and to determine the parameter set and clause set corresponding to each chapter node based on the chapter slot binding result; The chapter generation module is used to input the parameter set and clause set of the corresponding chapter node into the locally deployed and fine-tuned large language model according to the node order of the chapter structure constraint diagram, and generate the text of each chapter. The consistency verification module is used to perform numerical consistency verification on the text of each chapter based on the parameter relationship set, and to perform clause matching verification based on the clause coverage relationship, and output conflicting chapter identifiers; The regeneration control module is used to call the chapter generation module to regenerate the corresponding chapter text based on the conflicting chapter identifier and output the consultation report text.

2. The AI ​​technology consulting service system based on deep learning according to claim 1, characterized in that, The data acquisition and encoding module includes: Data type identification is performed on multi-source data in the field of engineering consulting, and the multi-source data in the field of engineering consulting is divided into text data, image data and tabular data; The text data is segmented into semantic word sequences. Each semantic word is mapped into a vector, and a position code is generated according to the order of the word in the original text. The semantic word vector is combined with the corresponding position code to form a text vector sequence. The image data is divided into blocks according to a preset size, the image pixel matrix is ​​divided into image blocks of a preset size, a linear mapping is performed on each image block to obtain an image block vector, and a two-dimensional position code is generated according to the row and column position of the image block in the original image. The image block vector and the two-dimensional position code are combined to form an image vector sequence. The table data is traversed by row index and column index, word segmentation vector mapping is performed on the text content in the cell, and table position code is generated according to the row index and column index. The cell vector and the table position code are combined to form a table vector sequence. The text vector sequence, image vector sequence, and table vector sequence are concatenated in a preset order to generate a multimodal input sequence.

3. The AI ​​technology consulting service system based on deep learning according to claim 1, characterized in that, The structural constraint latent variable module includes: The multimodal input sequence is input into the improved Perceiver IO model. The improved Perceiver IO model sets a preset number of latent variable arrays. The latent variable arrays and the multimodal input sequence interact with each other through a cross-attention structure. In each cross-attention layer, the latent variable array is used as the query vector, and the multimodal input sequence is used as the key vector and value vector to calculate the attention weights. The value vectors are then weighted and aggregated to obtain the intermediate latent variable representation. The improved Perceiver IO model sets up a latent variable partitioning structure, dividing the latent variable array into a structural latent variable subarray, a parameter latent variable subarray, and a clause latent variable subarray. Each subarray corresponds to chapter-level information, parameter calculation information, and clause coverage information, respectively. In the latent variable update process of the improved Perceiver IO model, a structural consistency tensor field constraint mechanism is introduced. The structural consistency tensor field consists of a set of third-order tensors, including chapter-level tensors, parameter conservation tensors, and clause coverage tensors. The chapter hierarchy tensor establishes an index mapping based on the hierarchical relationship between parent and child nodes in the engineering consulting standard outline, and records the association identifier for chapter node pairs with hierarchical dependencies. The parameter conservation tensor establishes an index mapping based on the calculation combination relationship between engineering parameters, and records the combination identifier for parameter sets participating in the same calculation relationship; The clause coverage tensor establishes an index mapping based on the reference correspondence between chapters and policy clauses, and records coverage identifiers for chapter nodes and corresponding clause sets. After the intermediate latent variable representations are generated at each level, the intermediate latent variable representations are mapped to the structural latent variable subarray, the parameter latent variable subarray, and the clause latent variable subarray, respectively. For each subarray, a consistency scan operation is performed according to the index mapping relationship in the corresponding tensor to identify the latent variable positions that do not satisfy the hierarchical dependency, parameter combination, and clause coverage relationships. A structural rearrangement operation is performed on the identified inconsistent latent variable locations, remapping the relevant latent variables to a structurally feasible subspace that satisfies the common definition of the chapter-level tensor, parameter conservation tensor, and clause coverage tensor, to obtain the corrected latent variable representation.

4. The AI ​​technology consulting service system based on deep learning according to claim 1, characterized in that, The structured decoding module includes: Attention is calculated for each parameter query vector and the modified latent variable representation in the parameter query vector set to obtain the corresponding parameter embedding vector, and then arranged in a preset parameter order to form a parameter matrix; Attention is calculated for each relation query vector in the relation query vector set and the modified latent variable representation to obtain the association score vector between parameter pairs. The parameter index pairs are then filtered based on the association score to form a parameter relation set. Attention is calculated for each chapter slot query vector and the modified latent variable representation in the chapter slot query vector set to obtain the chapter embedding vector. Based on the similarity calculation results between the chapter embedding vector and the parameter embedding vector, the parameter index is assigned to the corresponding chapter slot to form the chapter slot binding result.

5. The AI ​​technology consulting service system based on deep learning according to claim 1, characterized in that, The chapter structure construction module includes: Read the engineering consulting standard outline document, extract the chapter number, chapter name and chapter level identifier, parse the membership relationship between parent and child nodes according to the hierarchical coding rules in the chapter number, and establish a set of chapter nodes; Based on the membership relationship between parent and child nodes, directed edges are constructed in the chapter node set to form a chapter structure constraint graph, with upper-level chapter nodes pointing to lower-level chapter nodes; Match the chapter index in the chapter slot binding result with the chapter node set, establish a mapping relationship between chapter nodes and parameter indexes, and convert the mapped parameter index set into the parameter set of the corresponding chapter node; Based on the association identifier between the chapter index and the clause index recorded in the clause coverage tensor, the clause index is mapped to the corresponding chapter node, forming the clause set corresponding to the chapter node; Write the parameter set and clause set into the node attributes of the corresponding chapter node to complete the construction of the chapter structure constraint diagram.

6. The AI ​​technology consulting service system based on deep learning according to claim 1, characterized in that, The chapter generation module includes: The chapter generation order is determined according to the topological sorting result of the chapter structure constraint diagram; For the current chapter node, construct a parameter relationship subgraph based on the corresponding parameter set, and arrange the parameter nodes with computational associations in an ordered manner according to the connection order in the parameter relationship set to form a parameter relationship sequence; The parameter relationship sequence is converted into a structured control tag sequence. Each parameter node generates a unique parameter identifier, and the parameter value and associated parameter index tag are appended to the identifier. The clause numbers in the clause set are mapped to clause control tags, and then inserted into the specified positions in the parameter relationship sequence according to the clause order corresponding to the chapter node, forming the chapter control input sequence; The chapter control input sequence is input into the locally deployed and fine-tuned large language model. During the generation process, parameter identifier matching is scanned on the output text, and resampling is performed on the corresponding positions of parameter identifiers that are not matched or have abnormal order. After generating the text for the current chapter node, process each chapter node in the order of chapter generation to generate the text for each chapter.

7. The AI ​​technology consulting service system based on deep learning according to claim 1, characterized in that, The consistency verification module includes: The text of each chapter is segmented and scanned. The parameter names and values ​​in the text are extracted according to the preset parameter identifiers to form a set of chapter parameters. The parameter values ​​in the chapter parameter set are compared with the corresponding parameter values ​​in the parameter matrix, and the parameter indices that differ are marked. Based on the parameter index pairs recorded in the parameter relationship set, read the corresponding parameter values ​​in the chapter parameter set, perform the corresponding operations according to the combination order in the parameter relationship set, compare the operation results with the corresponding target parameter values ​​in the chapter text, and mark the parameter index pairs that do not satisfy the combination relationship; Based on the clause coverage tensor in the structural consistency tensor field, the mapping relationship between the chapter index and the clause index is read, the clause numbers in the chapter text are compared one by one, and the missing clause indexes are marked. The chapter indexes corresponding to the difference markers, combination markers, and missing markers are summarized to form conflict chapter identifiers.

8. The AI ​​technology consulting service system based on deep learning according to claim 1, characterized in that, The regeneration control module includes: Receive conflict chapter identifiers output by the consistency verification module, and determine the corresponding chapter node index based on the conflict chapter identifiers; Read the parameter set and clause set of the corresponding chapter node in the chapter structure constraint diagram; Based on the difference markers, combination markers, and missing markers recorded in the conflict chapter identifiers, the values ​​of the corresponding parameter indices in the parameter set are replaced with the corresponding values ​​in the parameter matrix. Then, based on the parameter relationship set, the parameter relationship subgraph is reconstructed for the parameter index pairs with combination markers to obtain the updated parameter set. The updated parameter set and clause set are passed to the chapter generation module, which then regenerates the corresponding chapter nodes according to the node order in the chapter structure constraint diagram to obtain the regenerated chapter text. The regenerated chapter text will replace the original chapter text record of the corresponding node in the chapter structure constraint diagram; After correcting all conflicts in all chapter nodes, the text of each chapter is spliced ​​together according to the topological order in the chapter structure constraint diagram to form the engineering consulting report text.

Images