Construction process dynamic risk management and control method and system based on knowledge enhanced large model

Abstract

The present application relates to the technical field of building engineering safety informatization, and more particularly to a construction process dynamic risk management and control method and system based on a knowledge-enhanced large model, comprising: constructing a two-way index for unstructured safety data to form a safety knowledge base; receiving a question, reconstructing a prompt word using a historical dialogue when triggering state migration, retrieving a knowledge slice; deploying a multi-layer risk control knowledge graph in a graph database; retrieving control measures in the graph based on project data, triggering a feedback loop when the matching degree is insufficient, scheduling secondary retrieval and fusion of the knowledge base, and outputting a risk management and control list; extracting process nodes of a construction plan, dynamically generating and dispatching a time-sensitive task list through time sequence inclusion determination; pre-enhancing operation images, converting them into a site behavior graph and a compliance behavior graph respectively, triggering a thinking chain reasoning to perform compliance comparison according to a topological difference set, and outputting a structured data package.

Description

Technical Field

[0001] This invention relates to the field of information technology for safety in construction engineering, and in particular to a method and system for dynamic risk management of construction processes based on a knowledge-enhanced large-scale model. Background Technology

[0002] Construction sites are characterized by complex working environments, multiple overlapping and parallel work processes, a high density of high-risk points, and an explosive growth of multi-source heterogeneous safety data. Existing intelligent safety production management typically employs Retrieval Enhanced Generation (RAG) technology to integrate and process unstructured textual materials such as industry regulations, technical standards, and accident cases. By constructing a knowledge base for literal keyword retrieval or semantic vector retrieval, it enables on-site safety expertise consultation and risk prevention.

[0003] Existing safety data retrieval strategies typically employ single keyword matching or single vector retrieval. In practical applications, pure keyword retrieval cannot identify the deep semantics of construction site terminology, which can easily lead to missed detections of critical risks. Meanwhile, pure vector retrieval often results in false detections and the illusion of a large model due to the sparse semantic space when faced with highly specialized construction safety terms and special construction method provisions. This leads to serious distortions in the safety prevention technical solutions and standard references provided by the system. Summary of the Invention

[0004] To address the above shortcomings, this invention provides a method and system for dynamic risk management of construction processes based on a knowledge-enhanced large model, aiming to improve the problem of severely distorted references to safety prevention technical solutions and specifications provided by the system.

[0005] In a first aspect, the present invention provides the following technical solution: a dynamic risk management method for construction processes based on a knowledge-enhanced large-scale model, comprising:

[0006] S1. Build an index based on the input unstructured security data to form a security knowledge base;

[0007] S2. Receive the question to be processed. When the question to be processed is detected to trigger the state transition condition, reconstruct the prompt words using the associated historical dialogue fragments, and retrieve the knowledge slices in the security knowledge base accordingly.

[0008] S3. Deploy a risk control knowledge graph including multiple business-level nodes in a graph database, and establish link mapping relationships between the multiple business-level nodes;

[0009] S4. Based on the input project basic information, retrieve control measures in the risk control knowledge graph, and trigger a feedback loop when the retrieved control measures do not meet the preset conditions. Perform a secondary retrieval in the safety knowledge base and integrate and complete the control measures to output a risk management list.

[0010] S5. Extract process nodes from the input construction plan document, cross-compare the target time corresponding to the process node with the risk control list, extract the corresponding control measures, generate a task list and send it to the execution end;

[0011] S6. Receive the operation image and perform enhancement processing, convert the enhanced operation image into a field behavior map, convert the safety data retrieved from the safety knowledge base into a compliance behavior map, and perform compliance comparison based on the topological difference between the compliance behavior map and the field behavior map, and output a structured data packet.

[0012] Secondly, this invention provides the following technical solution: a dynamic risk management system for construction processes based on a knowledge-enhanced large-scale model, the system comprising:

[0013] Heterogeneous data storage is used to store unstructured security data and deploy indexes to form a security knowledge base, as well as a risk control knowledge graph in a graph database;

[0014] The user interaction terminal is connected to the heterogeneous data storage device and is used to receive input questions to be processed and to receive knowledge slices fed back by the security knowledge base. It is also used to receive input project basic information and construction plan documents.

[0015] Visual acquisition equipment, set up facing the construction site, is used to capture and output images of the construction site operations;

[0016] The central processing unit is electrically connected to the heterogeneous data storage, the user interaction terminal, and the visual acquisition device, respectively. It is used to retrieve control measures from the risk control knowledge graph in response to the project basic data to output a risk management list, and to compare the target time corresponding to the process node with the risk management list to generate a task list.

[0017] A network dispatch communication master controller, electrically connected to the central processing unit, is used to dispatch the task list generated by the central processing unit to the execution end;

[0018] The data packet output interface is electrically connected to the central processing unit (CPU) and is used to receive the structured data packets generated by the CPU in response to the job image, after performing dual-image conversion and topology difference compliance comparison, and to output the structured data packets to an external terminal.

[0019] The present invention has the following beneficial effects:

[0020] 1. In this invention, by associating sparse and dense dual-path indexes, and combining finite state machines and pruning algorithms to track the reference and omission of questions, accurate semantic retrieval and noun fusion are achieved, as well as accurate context completion for fragmented multi-turn dialogues. This solves the problems of severe illusion in professional question answering, distortion of normative references, and logical breakage in multi-turn dialogues in general large models.

[0021] 2. In this invention, by matching the four-layer subgraph of the graph, the feedback loop is activated when the matching degree is insufficient, and the cross-attention mechanism is used to fuse and complete the control measures. Combined with the construction plan, the sequence inclusion of the process is determined, realizing the dynamic two-level coupling and completion of rules and knowledge base, as well as the real-time dynamic distribution of control measures with the construction progress. This solves the problems of missing measures when the traditional static graph faces special construction methods, and the problem of personnel information overload caused by the full distribution of the list.

[0022] 3. In this invention, by denoising, histogram equalization, and super-resolution pre-enhancing of images, spatiotemporal interaction labels are extracted using a multimodal network to construct a scene behavior map, and the topological difference between the map and the compliance behavior map is calculated. When the difference is not empty, thought chain reasoning is triggered, thus realizing degraded image restoration and a closed loop from object recognition to qualitative judgment based on topological difference and multi-step logical judgment. This solves the problems of target loss caused by environmental interference and the fact that traditional CV can only recognize static objects and cannot recognize complex spatiotemporal violations. Attached Figure Description

[0023] Figure 1 This is a flowchart of the dynamic risk management method for construction process based on a knowledge-enhanced large model proposed in this invention;

[0024] Figure 2 This is a flowchart illustrating the construction process of a dual-index safety knowledge base for a dynamic risk management method for construction processes based on a knowledge-enhanced large model, as proposed in this invention.

[0025] Figure 3 This is a flowchart of the two-level retrieval and agent feedback completion control measures for the dynamic risk management method of construction process based on knowledge-enhanced large model proposed in this invention;

[0026] Figure 4 This is a flowchart of image enhancement and dual-graph topology difference comparison for the dynamic risk management method of construction process based on knowledge-enhanced large model proposed in this invention.

[0027] Figure 5 This is an architecture diagram of the dynamic risk management system for construction processes based on a knowledge-enhanced large model proposed in this invention. Detailed Implementation

[0028] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] Example 1:

[0030] In a first embodiment of the present invention, the present invention provides a dynamic risk management method for construction processes based on a knowledge-enhanced large model, such as... Figures 1-4 As shown, it includes:

[0031] S1. Build an index based on the input unstructured security data to form a security knowledge base;

[0032] Furthermore, in S1, an index is built based on the input unstructured security data to form a security knowledge base, including:

[0033] Perform text cleaning and slicing on unstructured secure data to obtain a set of text data blocks;

[0034] We use word segmentation operators to extract word frequency features from a set of text data blocks and construct a sparse inverted index.

[0035] Running the text representation model transforms a collection of text data blocks into dense feature vectors, and constructs a dense vector index;

[0036] The sparse inverted index is associated with the dense vector index and stored in the database to form a security knowledge base.

[0037] Specifically, unstructured safety data mainly originates from historical safety accident case records at construction sites and safety production regulations and standards issued by the state and industry. The system first performs text cleaning and slicing processing on this data. The text cleaning process includes using preset regular expressions to filter out garbled characters, redundant spaces, and invisible control characters in the text. Subsequently, the system uses a sliding window slicing algorithm based on fixed character length and combined with semantic punctuation marks to segment the cleaned long text into independent text data blocks. A 20% overlap area is set between adjacent text data blocks to ensure semantic continuity at the slice boundaries, and finally, the data is organized into a set of text data blocks.

[0038] To extract word frequency features and construct a sparse inverted index, the system calls a word segmentation operator to segment each text data block in the text data block set. The word segmentation operator uses a pre-defined general dictionary and a professional vocabulary database for the safety production field to perform forward maximum matching word segmentation, filtering out meaningless words from the stop word list, thus transforming the text data blocks into discrete keyword sequences. Next, the system calculates the word frequency of each keyword in the current text data block. To balance the impact of long text blocks on word frequency, the system uses a word frequency normalization matrix for mapping. For the construction of the sparse inverted index, the system traverses all keywords, using the keyword as the key and the unique identifier, word frequency, and position information of the text data block containing that keyword as linked list nodes to construct a global sparse inverted index table.

[0039] In the parallel data processing stream, the system runs a text representation model to transform the set of text data blocks into dense feature vectors. This text representation model employs a pre-trained language model based on a bidirectional Transformer encoder structure. The system converts the keyword sequence of each text data block into a corresponding token sequence, adding start and end markers to the beginning and end of the sequence. The model inputs the token sequence into its embedding layer to convert it into low-dimensional initial word vectors, and then performs feature extraction through a twelve-layer bidirectional attention mechanism network. During computation, the attention mechanism dynamically weights the query matrix, key matrix, and value matrix through self-attention, capturing long-distance semantic dependencies within the text. Finally, the model extracts the vector corresponding to the start marker from the top-level output; this vector serves as a fixed-dimensional dense feature vector, used to represent the global semantics of the text data block. The system sequentially transforms all data blocks in the text data block set into fixed-dimensional dense feature vectors, forming a dense feature matrix.

[0040] After transforming the feature vectors, the system uses a clustering quantization algorithm to perform spatial partitioning on the dense feature vectors to construct a dense vector index. The system calls the K-means clustering operator to iteratively calculate within the vector space, dividing the dense feature vector set into several spatial clusters and determining the center vector of each cluster. The dense vector index is implemented by constructing a multi-level inverted file index structure, which records the feature vector identifiers of the text data blocks covered by each cluster center vector.

[0041] The system ultimately associates the constructed sparse inverted index with the dense vector index, storing them together with the text data block set in a heterogeneous database, thus completing the physical storage and index foundation of the security knowledge base. In subsequent processing flows, when receiving an externally input question, the system will call the sparse inverted index for precise matching of literal keywords and the dense vector index for semantic-level approximate nearest neighbor retrieval. The preliminary candidate text data blocks generated by the dual-path retrieval will be aligned with confidence scores using a reciprocal ranking fusion algorithm, calculated as follows:

[0042] ;

[0043] In the above formula, the parameters Representative candidate text data blocks The overall evaluation value after dual-path fusion. Parameters Representative candidate text data blocks The positive integer rank in the sparse inverted index search results. Parameter Representative candidate text data blocks The positive integer rank in the dense vector index retrieval results. Parameters This represents a constant used for smoothing the ranking. In this embodiment, the parameter... The value is a fixed constant of 60. The system sorts all candidate text data blocks from highest to lowest based on their comprehensive evaluation scores, and selects the top five text data blocks as the final knowledge slices. These knowledge slices will serve as deterministic compliance and security evidence and will be input into the subsequent prompt word reconstruction process.

[0044] S2. Receive the question to be processed. When the state transition condition is triggered by the question to be processed, reconstruct the prompt words using the associated historical dialogue fragments, and retrieve the knowledge slices in the security knowledge base accordingly.

[0045] Furthermore, in S2, when the state transition condition is triggered by the question to be processed, the cue words are reconstructed using associated historical dialogue fragments, including:

[0046] The semantic features of the question to be processed are analyzed in real time. When the presence of pronouns or omitted components is detected, the state transition conditions are determined and the finite state machine is controlled to transition to the context tracking state.

[0047] The pruning algorithm is used to extract relevant dialogue fragments from the historical dialogue set as historical dialogue fragments.

[0048] Historical dialogue fragments are injected into preset prompt template slots as placeholders, thus completing the prompt reconstruction.

[0049] Specifically, the system receives the questions to be processed in real time through a multi-information input terminal and calls a semantic parsing operator to perform serialized feature extraction on the questions. The semantic parsing operator first uses dependency parsing to analyze the part-of-speech tags and subject-verb-object topology of the questions, calculating the association weights between words. In terms of feature construction, the system inputs the text sequence of the questions to be processed into a lightweight named entity recognition model to extract construction elements, process nodes, and safety restriction words from the text, generating an entity distribution probability matrix. To determine whether a state transition condition is triggered, the system runs a feature matching algorithm to retrieve this probability matrix and detect whether there are semantic gaps in the questions.

[0050] When the system detects the presence of pronouns such as "this," "the process," or "its" in the text, or omitted elements such as subjects or objects, and the missing preceding core node in the dependency relationship within the current sentence, the matching algorithm outputs a trigger flag. At this point, the system determines the trigger state transition condition, and the control state of the finite state machine transitions. By rewriting the current state register value of the finite state machine, the system controls the finite state machine to transition from the initial single-sentence parsing state to the context tracking state, activating the historical dialogue retrieval path.

[0051] In context-tracking mode, the system runs a pruning algorithm to extract relevant dialogue fragments from the locally cached historical dialogue set. The historical dialogue set consists of multiple rounds of historical input text and corresponding system output text, arranged in ascending order of timestamps. The pruning algorithm first uses the entity distribution probability matrix of the current question as a retrieval vector and calculates its cosine similarity to the feature vectors of each round of dialogue text in the historical dialogue set. Next, the pruning algorithm runs a weighted confidence calculation logic based on a time decay factor, performing time-dimensional correction on the similarity score. The confidence score is calculated using the following formula:

[0052] ;

[0053] In the above formula, the parameters This represents the combined confidence score of the association between the current round of historical dialogues and the questions to be processed. Parameters This represents the cosine similarity between the feature vector of the current round of historical dialogue text and the feature vector of the question to be processed. (Parameters) This represents the set time decay exponential constant. In this embodiment, the parameter... The value is a fixed constant of 0.15. Parameter The timestamp represents the positive integer difference between the timestamp of the current round of historical dialogue and the timestamp of the question to be processed, in seconds. The pruning algorithm sequentially calculates the comprehensive relevance confidence score of all historical dialogues and filters out and retains historical dialogue segments with scores greater than the set similarity threshold of 0.65, while removing other irrelevant rounds. Subsequently, the system reassembles the selected historical dialogue segments into a historical dialogue context string according to the order of their timestamps.

[0054] After obtaining the historical dialogue context string, the system injects it into the preset prompt word template slots using placeholder replacement, completing the prompt word reconstruction. The preset prompt word template consists of fixed constraint text and specific reserved character marker slots. The system calls the string retrieval and replacement operator to locate the context placeholder slot in the prompt word template, erases the slot, and replaces it with the reconstructed historical dialogue context string. Simultaneously, the system fills the current question to be processed into the current question slot of the prompt word template, assembling and generating reconstructed prompt word text containing complete context.

[0055] The system transforms the reconstructed prompt text into a text feature retrieval query and performs a two-way hybrid search within a safety knowledge base hosted in a heterogeneous database. First, the system uses a sparse inverted index to perform text keyword matching, recalling initial data blocks containing the same construction terms and performing preliminary ranking based on word frequency features. Simultaneously, the system uses a dense vector index to calculate the Euclidean distance between the reconstructed prompt text vector and each dense feature vector in the knowledge base within the vector space, recalling semantically adjacent candidate data blocks. The system employs a reciprocal ranking fusion algorithm to combine the sparse and dense search rankings, outputting a comprehensive evaluation value for the candidate text data blocks. The system then re-ranks the candidate text data blocks according to their comprehensive evaluation values ​​from highest to lowest, selecting the top five text data blocks as the final knowledge slices. These knowledge slices, as deterministic data, are input into subsequent risk control map comparisons or anomaly feedback interfaces, participating in the output of subsequent risk management lists.

[0056] S3. Deploy a risk control knowledge graph that includes multiple business-level nodes in the graph database, and establish link mapping relationships between multiple business-level nodes;

[0057] Furthermore, in S3, a risk control knowledge graph comprising multiple business-level nodes is deployed in the graph database, and link mapping relationships are established between these multiple business-level nodes, including:

[0058] The business-level nodes in the risk control knowledge graph are divided from top to bottom as follows: project nodes, risk point nodes, hazard source nodes, and control measure nodes.

[0059] Based on the dependency logic of the expert security system, bidirectional edge relationships are established between multiple business-level nodes as link mapping relationships to complete the entity deployment of the risk control knowledge graph in the graph database.

[0060] Specifically, the system first initializes the graph topology model in the graph database hosted by the heterogeneous data storage, establishing a mesh schema architecture for node labels and edge types. The system divides the business-level nodes in the risk control knowledge graph into four core levels from top to bottom: project nodes, risk point nodes, hazard source nodes, and control measure nodes. When deploying these business-level nodes, the system reads unstructured security data and project basic information using text parsing operators, extracts entity text using named entity recognition technology, and assigns a globally unique 64-bit integer ID as the node's primary key. Project nodes store the attribute characteristics of specific projects, including project name and geographical coordinates.

[0061] Risk point nodes are subordinate to project nodes, storing information about construction areas or sub-projects, including tunnel excavation faces or high-formwork areas. Hazard source nodes are further subordinate to risk point nodes, storing specific physical unsafe conditions or unsafe human behaviors, including rock instability or lack of edge protection. Control measure nodes are at the lowest level, storing specific technical means or management requirements for eliminating or reducing hazard source risks, including advanced support boreholes or the installation of double-layer horizontal safety nets. The system executes graph definition statements to transform the above-mentioned entities into graph nodes with attribute key-value pairs and persists them to the graph database.

[0062] After deployment across multiple business-level nodes, the system establishes bidirectional edge relationships between these nodes as link mapping relationships, based on the dependency logic of the expert safety system. The dependency logic of the expert safety system is digitally expressed through a pre-defined rule base for investigating and managing potential production safety hazards. The system extracts the entity features of project nodes, uses path retrieval algorithms to locate their corresponding risk point nodes, and establishes bidirectional edge relationships of inclusion and being included between project nodes and risk point nodes. Similarly, based on the attribution association between hazard sources and specific areas, the system establishes bidirectional edge relationships of triggering and evolution between risk point nodes and hazard source nodes, and bidirectional edge relationships of management and mapping between hazard source nodes and final control measure nodes.

[0063] To quantify the tightness of edge connections and filter weakly correlated links, the system executes dependency correlation calculation logic when establishing bidirectional edge relationships, assigning a floating-point weight value to each edge. This weight value comprehensively considers the correlation coefficient of historical incident statistics and the expert experience matrix, and the calculation formula is as follows:

[0064] ;

[0065] In the above formula, the parameters Represents the connection node With nodes The entity dependency weight value of the bidirectional edge. Parameter Representing nodes in historical security data Corresponding security events and nodes The positive integer frequency of the corresponding security events. Parameter Representative node The total positive integer frequency of the corresponding security event occurring alone in historical data. Parameter Representative node The total positive integer frequency of the corresponding security event occurring alone in historical data. Parameter Represents nodes read from a pre-defined expert experience matrix. With nodes The expert prior correlation coefficient between them, whose value is a floating-point number between 0 and 1. Parameter and parameters This represents a preset adjustment weight coefficient. In this embodiment, the parameter... The value is a fixed constant of 0.6, parameter The value is a fixed constant of 0.4. The system calculates the dependency association weight value and encapsulates it as an attribute field of the edge into the graph data. When the dependency association weight value is greater than the set pruning threshold of 0.35, the graph database officially executes the graph modification transaction, writing the bidirectional edge relationship to disk, thereby completing the entity deployment and link construction of the risk control knowledge graph in the graph database. This deployed graph will serve as the base data for graph matching and control measure retrieval when receiving basic project data.

[0066] S4. Based on the input project basic information, retrieve control measures in the risk control knowledge graph, and trigger a feedback loop when the retrieved control measures do not meet the preset conditions. Perform a secondary search in the safety knowledge base and integrate and complete the control measures to output a risk management list.

[0067] Furthermore, in S4, triggering the feedback loop when the retrieved control measure does not meet the preset conditions includes:

[0068] Extract entity tags from the project's basic data and use graph query statements to match and obtain the first control measure in the risk control knowledge graph;

[0069] Calculate the semantic matching score between the first control measure and the basic project data;

[0070] When the semantic matching score is less than the preset threshold, it is determined that the first control measure does not meet the preset conditions, and the abnormal feedback interface for the current project basic data is activated to trigger the feedback loop.

[0071] Furthermore, in S4, the secondary retrieval and fusion completion control measures performed in the security knowledge base include:

[0072] In response to the activation of the anomaly feedback interface, the basic project data is converted into a text feature retrieval formula, and a multi-path hybrid retrieval is performed in the security knowledge base to recall the associated second control measure text segment;

[0073] The global features of the first control measure are extracted using a cross-attention mechanism and the local features of the text segment of the second control measure are semantically aligned.

[0074] Based on the semantic alignment results, the text segment of the second control measure is merged and injected into the first control measure to complete the content and output a risk management list.

[0075] Specifically, in the implementation of a dynamic risk management method for construction processes based on a knowledge-enhanced large model, the process of performing two-level retrieval, anomaly feedback, and measure fusion and completion based on the input project basic data in the risk control knowledge graph and safety knowledge base has a clear data processing flow.

[0076] The system first receives basic project data through a multi-source information input terminal and uses a named entity recognition model to extract entity labels such as specific construction strata and foundation pit excavation depth. Then, the system converts these entity labels into structured graph query statements and inputs them into the graph database. It performs subgraph topology matching with a pre-deployed risk control knowledge graph, traversing project nodes, risk point nodes, and hazard source node links to retrieve and match the control measure nodes connected to its lowest level, thereby obtaining the text of the first control measure.

[0077] To determine whether the first control measure meets the preset conditions, the system runs a bidirectional encoding representation model to convert the text of the first control measure and the text of the project's basic data into fixed-dimensional first dense feature vectors and second dense feature vectors. Next, the system calls a similarity measurement operator to calculate the cosine similarity between these two feature vectors and adjusts it using a scaling factor to calculate the semantic matching score. The calculation formula is as follows:

[0078] ;

[0079] In the above formula, the parameters This represents the semantic matching score between the primary control measure and the project's basic data. (Parameter) This represents the fixed-dimensional first-dense feature vector obtained after encoding the text of the first control measure. Parameters The second dense feature vector of fixed dimension is obtained after encoding the basic data text of the representative project. Parameters and parameters These represent the magnitudes of the corresponding feature vectors. Parameters This represents a preset scaling correction factor. In this embodiment, the parameter... The value is a fixed constant of 0.95. When the semantic matching score is less than the preset threshold of 0.75, the system determines that the first control measure does not meet the preset conditions, activates the abnormal feedback interface and outputs an error control code, triggering the feedback loop.

[0080] In response to the activation of the exception feedback interface, the system assembles the project's basic information text with unmet entity tags into a text feature retrieval expression, and schedules the security knowledge base to perform a multi-path hybrid retrieval. In the retrieval stream, the retrieval expression is input in parallel into a sparse inverted index engine for exact keyword matching, and into a dense vector index engine for approximate nearest neighbor retrieval, recalling the initial list of text paragraphs. The system then runs a reciprocal ranking fusion algorithm to perform weighted alignment of the two rankings, ultimately recalling the second control measure text paragraph with the highest overall ranking.

[0081] After obtaining the text segments of the first and second control measures, the system performs semantic alignment using a cross-attention mechanism. The system inputs the text sequence of the first control measure into the query matrix of the cross-attention network to construct a query vector, and simultaneously inputs the text sequence of the second control measure into the key and value matrices to construct key and value vectors, respectively. During network computation, the query vector and key vector undergo a matrix dot product operation to calculate the association weight matrix between the two tokens. The network uses the Softmax function to normalize this weight matrix and then performs a weighted sum with the value vectors to extract complementary local features from the second control measure text segment, thus achieving precise semantic alignment.

[0082] Finally, based on the semantic alignment results, the system merges and injects high-weight text segments from the second control measure text segment into the designated text boundary slots of the first control measure to complete the content, eliminate semantic conflicts, and ultimately output a structured risk management list. This risk management list will be stored as deterministic safety technology foundation data in a heterogeneous data storage system for subsequent cross-timeliness comparison processing with construction plan document process nodes.

[0083] S5. Extract process nodes from the input construction plan document, cross-compare the target time corresponding to the process node with the risk control list, extract the corresponding control measures, generate a task list and send it to the execution end;

[0084] Furthermore, in S5, the cross-comparison of the target time corresponding to the process node with the risk control list includes:

[0085] Parse the construction plan document to obtain the current execution time span of the system, make a containment judgment between the time span and the target time execution, and filter out the process nodes that match the current time span;

[0086] Retrieve the risk management list and extract control measures that have a dependency mapping relationship with the selected process nodes;

[0087] Map aggregated process nodes to corresponding control measures to generate a time-sensitive task list.

[0088] Specifically, the system first receives the construction plan document through the terminal data input module. This document is typically a Gantt chart, network diagram, or project schedule management table. The system then uses a text parsing engine and a time entity extraction operator to perform a time-series structured transformation on the construction plan document. In terms of feature construction, the system converts the text stream in the document into an event sequence matrix, extracting key features including the process name, predecessor process identifier, planned start time, and planned completion time. This forms a set of process nodes with time intervals, where the target time for each process node is represented by a closed interval determined by its planned start time and planned completion time.

[0089] To perform cross-referencing, the system first parses the construction plan document to obtain the current execution time span. Specifically, the system reads the server system clock or the current progress status stamp from the scheduling plan file to determine a deterministic time span defined by the current start boundary timestamp and the current end boundary timestamp. Subsequently, the system performs an inclusion check on this time span against the target time of each process node in the process node set. In this decision flow, the system traverses the process node set, runs the interval overlap determination operator, and calculates the intersection of the target time interval of the process node and the current execution time span interval. If the intersection is not empty, the process node is determined to match the current time span. To quantify the urgency of the timing and perform precise filtering, the system runs the overlapping time sequence score calculation logic when performing the inclusion check. The calculation formula is as follows:

[0090] ;

[0091] In the above formula, the parameters This represents the overlap time sequence score between the current process node and the current execution time span of the system. Parameters The planned start timestamp for the current process node. Parameter The planned completion timestamp for the current process node. Parameter The current start boundary timestamp representing the current execution time span of the system. Parameter The current cutoff timestamp representing the current execution time span of the system. Operators This represents retrieving the maximum value from the input parameters. (Operator) This represents taking the minimum value among the input parameters. In this embodiment, when the timing score overlaps... When the time overlap threshold is greater than 0.10, the system determines that the process node is active in the current execution cycle, thus successfully filtering out the target process node that hits the current time span.

[0092] After obtaining the selected target process nodes, the system further retrieves the risk management list previously stored in the heterogeneous data storage. The system invokes the entity relationship dependency matching operator to extract control measures from the risk management list that have a dependency mapping relationship with the target process node. In this data processing stream, the system extracts the process name and construction technology tag of the target process node, using them as search keys to perform a hash key-value pair search in the dependency relationship association table of the risk management list. Because the risk management list pre-stores the mapping matrix between risk points, hazard sources, and specific processes, the matching operator can accurately extract the deep control measure text directly associated with the currently active process node, thereby eliminating redundant control interference from non-current time periods or non-current processes.

[0093] The system ultimately maps and aggregates the selected target process nodes with the extracted corresponding control measures. In the data assembly process, the system uses the target process node as the master record, encapsulating the corresponding control measure text, time-sensitive time interval information, and execution responsibility chain tags into a unified structured JSON data object, generating a time-sensitive task list. Subsequently, the main control processor activates the network dispatch communication master control, invoking a remote process call protocol or message queue push mechanism to dispatch the generated time-sensitive task list in the form of encrypted data packets to the execution-end intelligent terminal located at the construction site in real time for on-site response and status monitoring, thereby completing dynamic closed-loop management at the process sequence level.

[0094] S6. Receive the operation image and perform enhancement processing. Convert the enhanced operation image into a field behavior map. Convert the safety data retrieved from the safety knowledge base into a compliance behavior map. Perform compliance comparison based on the topological difference between the compliance behavior map and the field behavior map, and output a structured data packet.

[0095] Furthermore, in S6, receiving the job image and performing enhancement processing includes:

[0096] A noise reduction network is invoked to suppress dust and rain / fog scattering noise in the input job image;

[0097] Run the histogram equalization algorithm to perform lighting compensation on the noise-reduced image;

[0098] A super-resolution reconstruction network is used to improve the sharpness of target edges in the light-compensated work image, and the enhanced work image is output.

[0099] Furthermore, in S6, compliance comparison based on the topological difference between the compliance behavior diagram and the on-site behavior diagram includes:

[0100] Multimodal networks are used to extract spatiotemporal interaction labels of multiple entities in the enhanced operation images, and scene semantic graphs are constructed as on-site behavior graphs.

[0101] Analyze the entity and action constraints in security data to construct a compliance constraint graph as a compliance behavior graph;

[0102] The running graph matching operator calculates the topological difference between the compliance behavior graph and the field behavior graph. When the topological difference is not empty, the thinking chain reasoning operator is triggered to perform compliance comparison on the missing constraint nodes, and generates and outputs a structured data packet.

[0103] Specifically, the system first receives real-time images of the construction area using a field vision acquisition component. Due to the complex environment of the construction site, the input images often suffer from degradation caused by dust, rain, or dense fog. The system then calls a denoising network to suppress dust and rain / fog scattering noise from the input images. This denoising network employs an end-to-end convolutional neural network. The network receives the original image matrix as input, extracts noise distribution features using multiple degenerate kernel convolutional layers, estimates and subtracts scattering noise components through a residual connection network, and outputs a denoised intermediate image. Subsequently, the system runs a histogram equalization algorithm to compensate for the lighting in the denoised images. This algorithm first statistically analyzes the gray-level histogram probability distribution of the denoised image, calculates the cumulative gray-level distribution function, and then performs a non-linear mapping on each pixel of the image using the cumulative distribution function, stretching the dynamic range of low-contrast areas to compensate for visual loss caused by insufficient lighting.

[0104] Next, the system utilizes a super-resolution reconstruction network to enhance the sharpness of target edges in the light-compensated work image. The super-resolution reconstruction network employs a generative adversarial network based on an attention mechanism or a deep subpixel convolutional network. It takes the low-resolution image matrix after light compensation as input, reconstructs high-frequency texture details through subpixel upsampling operators, strengthens the edge contours of target entities such as construction workers and safety protection facilities, and finally outputs a high-definition enhanced work image.

[0105] In the graph topology transformation and comparison process, the system first uses a multimodal network to extract spatiotemporal interaction labels for multiple entities in the enhanced work images, constructing a scene semantic graph as the on-site behavior graph. The multimodal network, in conjunction with a visual perception backbone network and a spatiotemporal behavior detection operator, converts the enhanced work image sequence into entity spatial coordinates and temporal action trajectory matrices, identifying entities such as workers, electrical boxes, and safety belts on the construction site, and extracting their spatiotemporal interaction relationships, such as a worker approaching an electrical box or a worker not fastening their safety belt. The system uses the identified entities as nodes and the spatiotemporal interaction labels as edges to construct a directed attribute graph as the on-site behavior graph.

[0106] Simultaneously, the system constructs a compliance constraint graph as a compliance behavior graph by parsing the entity and action constraints in the security data. The data for this step comes from the security data retrieved from the aforementioned security knowledge base. The system calls the text entity parsing operator to extract the standard compliance behavior norms from the security data, such as the requirement to wear insulating gloves when operating a distribution box, thereby mapping the compliance rules into a compliance constraint graph containing the relationships between standard entity nodes and constraint edges.

[0107] After obtaining the compliance behavior graph and the field behavior graph, the system uses a graph matching operator to calculate their topological difference. The graph matching operator defines the compliance behavior graph as the standard graph and the field behavior graph as the graph to be checked, comparing their topological structures by performing graph isomorphism mapping or subgraph matching algorithms. In the computational flow, the matching operator uses the set of nodes and edges in the compliance behavior graph as a benchmark, traversing to search for corresponding equivalent nodes and interaction edges in the field behavior graph. When the field behavior graph lacks the standard constraint edges or entity nodes required by the compliance behavior graph, the graph matching operator outputs a non-empty topological difference matrix through set subtraction.

[0108] When the topological difference set is not empty, the system triggers the thought chain inference operator to perform compliance comparison on the missing constraint nodes. The thought chain inference operator encapsulates the topological difference matrix, the contextual attributes of the on-site behavior graph, and the compliance constraint rules into a structured hint vector, which is then input into the built-in large language model inference engine. The inference engine analyzes the specific violation nature of the missing constraint nodes in the current construction context step by step through a multi-step explicit inference chain, and calculates the risk confidence score of the violation behavior. The calculation formula is as follows:

[0109] ;

[0110] In the above formula, the parameters This represents the confidence score for the potential risk of violations related to a specific missing constraint node. Parameters Represents the first step in the reasoning process of the thought chain. The local compliance violation probability value output by each inference step, which is a floating-point number between 0 and 1. Parameters Represents the total number of reasoning steps. Parameter Representing a preset environmental risk correction coefficient, in this embodiment, the parameter... The value is a fixed constant of 1.25. This is used when the final risk confidence score is... When the risk threshold of 0.70 is exceeded, the system identifies the missing node as a confirmed violation hazard and automatically generates a structured data packet containing the violation type, occurrence time, risk level, and rectification suggestions. This data packet is then transmitted to an external terminal via the data packet output interface to participate in subsequent safety warnings and closed-loop rectification processes.

[0111] Example 2:

[0112] In the deep foundation pit excavation and support construction scenario of a large high-rise building project, due to the involvement of deep soft soil layers, multi-level slope excavation, and complex internal support installation procedures, on-site safety management faces severe technical challenges. On the one hand, traditional manually compiled safety briefings and risk control measures are too generalized. When faced with sudden emergencies such as localized water seepage in the foundation pit or sudden stress changes in support anchors, managers often fail to obtain precise special construction method countermeasures due to the severe distortion of standard references and the illusion of a large model caused by a single search strategy when searching through existing question-and-answer systems. On the other hand, static risk control lists are distributed in their entirety and cannot be dynamically updated according to the actual progress of daily earthwork excavation. Moreover, the dust on site blurs the visual monitoring images, and traditional video surveillance can only identify whether personnel are wearing safety helmets, completely failing to automatically identify and characterize logical violations with complex spatiotemporal constraints, such as mechanical over-excavation and reversed support removal procedures. To solve the above problems, this invention provides a dynamic risk management system for construction processes based on a knowledge-enhanced large model, the structure of which is as follows: Figure 5 As shown. The specific implementation process of this system is as follows:

[0113] Heterogeneous data storage is used to store unstructured security data and deploy indexes to form a security knowledge base, as well as a risk control knowledge graph in a graph database;

[0114] The user interaction terminal is connected to a heterogeneous data storage device. It is used to receive input questions to be processed and to receive knowledge slices fed back from the security knowledge base. It is also used to receive input project basic information and construction plan documents.

[0115] Visual acquisition equipment, set up facing the construction site, is used to capture and output images of the construction site operations;

[0116] The central processing unit is electrically connected to the heterogeneous data storage, user interaction terminal and visual acquisition device respectively. It is used to retrieve control measures in the risk control knowledge graph in response to the basic data of the project to output a risk management list, and to compare the target time corresponding to the process node with the risk management list to generate a task list.

[0117] The network dispatch communication master controller is electrically connected to the central processing unit and is used to dispatch the task list generated by the central processing unit to the execution end;

[0118] The data packet output interface is electrically connected to the central processing unit (CPU) and is used to receive structured data packets generated by the CPU in response to the job image, perform dual-image conversion and topology difference compliance comparison, and output the structured data packets to external terminals.

[0119] Specifically, the system first utilizes a heterogeneous data storage device to store unstructured safety data such as deep foundation pit excavation specifications and case studies. A sparse inverted index and a dense vector index are then constructed in parallel by the central processing unit, forming a deep foundation pit safety knowledge base. Simultaneously, a four-layer risk control knowledge graph is deployed in the graph database of the heterogeneous data storage device, sequentially dividing nodes into project nodes, risk point nodes, hazard source nodes, and control measure nodes.

[0120] When a dangerous situation occurs on-site, when the manager inputs a question containing references or omissions through the user interaction terminal, the central processing unit controls the finite state machine to transition to the context tracking state, runs the pruning algorithm to extract historical dialogue reconstruction prompts, performs a reciprocal sorting and fusion retrieval in the safety knowledge base, obtains accurate knowledge slices, and feeds them back to the user interaction terminal to eliminate model illusions.

[0121] The central processing unit (CPU) generates a list of basic project data input via a user terminal. The CPU extracts entity labels such as excavation depth and uses a graph query to match the data in the graph to obtain the first control measure. When the semantic matching score is less than a preset threshold, the CPU activates an anomaly feedback interface to trigger a feedback loop, retrieves the second control measure text segment from the knowledge base, and uses a cross-attention mechanism to stitch it into the first control measure for content completion, outputting a deep foundation pit risk management list.

[0122] The central processing unit (CPU) responds to construction plan documents input via user-interactive terminals to implement timeliness control. The CPU parses the document to obtain the current execution time span, extracts earthwork excavation process nodes, and calculates the overlap score between the process target time and the current time span. When the score meets a threshold, process nodes matching the current time span are selected, and control measures with dependencies in the risk management list are retrieved and mapped to generate a timeliness task list. The CPU sends this list to the network dispatch communication master, which then dispatches it in real-time to the execution-end intelligent terminals using encrypted data packets, enabling dynamic task updates.

[0123] Visual acquisition devices positioned at the excavation pit work surface capture visual signals in real time and output work images. To address dust interference, the central processing unit (CPU) utilizes noise reduction, histogram equalization, and super-resolution reconstruction networks to perform multi-level pre-enhancement of the images. Subsequently, the CPU uses a multimodal network to extract spatiotemporal interaction labels of multiple entities in the enhanced images (such as mechanical over-excavation and reversed support removal procedures), constructing a scene semantic graph as a site behavior graph; simultaneously, it analyzes safety data to construct a compliance constraint graph as a compliance behavior graph. The CPU runs a graph matching operator to calculate the topological difference between the two. When the topological difference is not empty, it triggers a thought chain reasoning operator to perform multi-step logical comparison and qualitative analysis of missing constraints, producing a structured data packet containing the qualitative analysis of violations and the legal basis. This data packet is transmitted to the data packet output interface and output to an external management terminal, achieving closed-loop management of potential hazards.

[0124] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A dynamic risk management method for construction processes based on a knowledge-enhanced large-scale model, characterized in that: include: S1. Build an index based on the input unstructured security data to form a security knowledge base; S2. Receive the question to be processed. When the question to be processed is detected to trigger the state transition condition, reconstruct the prompt words using the associated historical dialogue fragments, and retrieve the knowledge slices in the security knowledge base accordingly. S3. Deploy a risk control knowledge graph including multiple business-level nodes in a graph database, and establish link mapping relationships between the multiple business-level nodes; S4. Based on the input project basic information, retrieve control measures in the risk control knowledge graph, and trigger a feedback loop when the retrieved control measures do not meet the preset conditions. Perform a secondary retrieval in the safety knowledge base and integrate and complete the control measures to output a risk management list. S5. Extract process nodes from the input construction plan document, cross-compare the target time corresponding to the process node with the risk control list, extract the corresponding control measures, generate a task list and send it to the execution end; S6. Receive the operation image and perform enhancement processing, convert the enhanced operation image into a field behavior map, convert the safety data retrieved from the safety knowledge base into a compliance behavior map, and perform compliance comparison based on the topological difference between the compliance behavior map and the field behavior map, and output a structured data packet.

2. The method for dynamic risk management of construction process based on knowledge-enhanced large model according to claim 1, characterized in that, In S1, the step of constructing an index based on the input unstructured security data to form a security knowledge base includes: Text cleaning and slicing are performed on the unstructured security data to obtain a set of text data blocks; The word frequency features of the text data block set are extracted using the word segmentation operator, and a sparse inverted index is constructed. The text representation model is run to transform the set of text data blocks into dense feature vectors, and a dense vector index is constructed. The sparse inverted index is associated with the dense vector index and stored in the database to form a security knowledge base.

3. The method for dynamic risk management of construction process based on a knowledge-enhanced large model according to claim 1, characterized in that, In S2, when the state transition condition triggered by the question to be processed is detected, the reconstructing of the prompt words using associated historical dialogue fragments includes: The semantic features of the question to be processed are analyzed in real time. When the presence of pronouns or omitted components is detected, the state transition condition is determined and the finite state machine is controlled to transition to the context tracking state. The pruning algorithm is used to extract relevant dialogue fragments from the historical dialogue set as historical dialogue fragments. The historical dialogue fragments are injected into the preset prompt word template slots using placeholder replacement, thus completing the prompt word reconstruction.

4. The method for dynamic risk management of construction process based on knowledge-enhanced large model according to claim 1, characterized in that, In S3, deploying a risk control knowledge graph comprising multiple business-level nodes in a graph database and establishing link mapping relationships between the multiple business-level nodes includes: The business-level nodes in the risk control knowledge graph are divided from top to bottom as follows: project nodes, risk point nodes, hazard source nodes, and control measure nodes. Based on the dependency logic of the expert security system, bidirectional edge relationships are established between multiple business-level nodes as link mapping relationships to complete the entity deployment of the risk control knowledge graph in the graph database.

5. The method for dynamic risk management of construction process based on knowledge-enhanced large model according to claim 1, characterized in that, In S4, triggering the feedback loop when the retrieved control measures do not meet the preset conditions includes: Extract entity tags from the project's basic data and use graph query statements to match them in the risk control knowledge graph to obtain the first control measure; Calculate the semantic matching score between the first control measure and the project basic data; When the semantic matching score is less than a preset threshold, it is determined that the first control measure does not meet the preset conditions, and the abnormal feedback interface for the current project basic data is activated to trigger the feedback loop.

6. The method for dynamic risk management of construction process based on knowledge-enhanced large model according to claim 5, characterized in that, In S4, the secondary retrieval and fusion completion control measures performed in the security knowledge base include: In response to the activation of the anomaly feedback interface, the basic project data is converted into a text feature retrieval formula, and a multi-path hybrid retrieval is performed in the security knowledge base to recall the associated second control measure text segment; The global features of the first control measure are extracted using a cross-attention mechanism and the local features of the text segment of the second control measure are semantically aligned. Based on the semantic alignment results, the text segment of the second control measure is merged and injected into the first control measure to complete the content and output a risk management list.

7. The method for dynamic risk management of construction process based on a knowledge-enhanced large model according to claim 1, characterized in that, In S5, the cross-comparison of the target time corresponding to the process node with the risk control list includes: The construction plan document is parsed to obtain the current execution time span of the system. The execution time span is compared with the target time to determine the inclusion of the process nodes that match the current time span. Retrieve the risk management list and extract control measures that have a dependency mapping relationship with the selected process nodes; The process nodes are mapped and aggregated with their corresponding control measures to generate a time-sensitive task list.

8. The method for dynamic risk management of construction process based on knowledge-enhanced large model according to claim 1, characterized in that, In S6, receiving the job image and performing enhancement processing includes: A noise reduction network is invoked to suppress dust and rain / fog scattering noise in the input image of the operation. Run the histogram equalization algorithm to perform lighting compensation on the noise-reduced image; A super-resolution reconstruction network is used to improve the sharpness of target edges in the light-compensated work image, and an enhanced work image is output.

9. The method for dynamic risk management of construction process based on knowledge-enhanced large model according to claim 1, characterized in that, In S6, the compliance comparison based on the topological difference between the compliance behavior diagram and the on-site behavior diagram includes: A multimodal network is used to extract spatiotemporal interaction labels of multiple entities in the enhanced operation image, and a scene semantic map is constructed as the on-site behavior map. Analyze the entity and action constraints in security data to construct a compliance constraint graph as a compliance behavior graph; The running graph matching operator calculates the topological difference between the compliance behavior graph and the field behavior graph. When the topological difference is not empty, the thinking chain reasoning operator is triggered to perform compliance comparison on the missing constraint nodes, and generates and outputs a structured data packet.

10. A dynamic risk management and control system for construction processes based on a knowledge-enhanced large-scale model, characterized in that: The system for the dynamic risk management method for construction processes based on a knowledge-enhanced large model as described in any one of claims 1-9, the system comprising: Heterogeneous data storage is used to store unstructured security data and deploy indexes to form a security knowledge base, as well as a risk control knowledge graph in a graph database; The user interaction terminal is connected to the heterogeneous data storage device and is used to receive input questions to be processed and to receive knowledge slices fed back by the security knowledge base. It is also used to receive input project basic information and construction plan documents. Visual acquisition equipment, set up facing the construction site, is used to capture and output images of the construction site operations; The central processing unit is electrically connected to the heterogeneous data storage, the user interaction terminal, and the visual acquisition device, respectively. It is used to retrieve control measures from the risk control knowledge graph in response to the project basic data to output a risk management list, and to compare the target time corresponding to the process node with the risk management list to generate a task list. A network dispatch communication master controller, electrically connected to the central processing unit, is used to dispatch the task list generated by the central processing unit to the execution end; The data packet output interface is electrically connected to the central processing unit (CPU) and is used to receive the structured data packets generated by the CPU in response to the job image, after performing dual-image conversion and topology difference compliance comparison, and to output the structured data packets to an external terminal.

Images