Highway construction quality common fault diagnosis and prevention system based on knowledge graph

Abstract

The application relates to the technical field of intelligent management and control of engineering quality, and discloses a highway construction quality common fault diagnosis and prevention system based on a knowledge graph; the system extracts quality phenomena, inducing factors and prevention measure entities of engineering documents, matches construction specifications to obtain quantitative thresholds to construct a static graph; a mask matrix is generated according to sensor data flow and the quantitative thresholds, and a dynamic topological graph is generated in combination with the static graph; a progress file is converted into a process directed acyclic graph, and a multi-agent Markov decision process is initialized; hidden danger investigation data is mapped to the static graph to generate an expert demonstration trajectory and pre-train a strategy network; the strategy network is updated according to the process directed acyclic graph to constrain the walking time sequence of the agent, and environmental rewards are calculated; a query instruction is mapped into a starting node, the strategy network is called to perform forward inference, and a control message is output. The application combines field knowledge and on-site sensing data, and improves the adaptability and engineering feasibility of the prevention strategy.

Description

Technical Field

[0001] This invention relates to the field of intelligent management and control technology for engineering quality, specifically a knowledge graph-based system for diagnosing and preventing common quality problems in highway construction. Background Technology

[0002] Common quality defects such as roadbed settlement and pavement cracking are prone to occur during highway construction, and their diagnosis and prevention are of great significance to project quality and operational safety. Existing quality defect diagnosis systems largely rely on the experience accumulated by on-site engineers or on matching judgments based on static rules extracted from construction specifications. With the increasing intelligence of on-site equipment, a large amount of continuous time-series sensor data is generated during construction. However, traditional diagnostic models typically process domain knowledge from construction specifications and on-site sensor data independently, lacking a mechanism to fuse static knowledge such as specification thresholds with on-site sensor characteristics. This results in static diagnostic systems being unable to dynamically adjust judgment boundaries according to real-time changes in the on-site environment, and diagnostic results often lag behind actual on-site conditions.

[0003] Existing methods fail to effectively integrate project schedules when reasoning about prevention and control strategies. Because they do not incorporate process dependencies, resource availability limits, and time constraints into the reasoning model, the prevention and control strategies generated by the system are prone to conflict with the current on-site processes or resource scheduling, resulting in low feasibility of strategy implementation.

[0004] Furthermore, the factors contributing to common quality defects at construction sites are complex. If reinforcement learning algorithms are directly introduced for collaborative diagnosis, the initial exploration cost of the computational model is high and the convergence is slow in the absence of prior knowledge extraction and pre-training from experts, such as historical hazard investigation records. This makes it difficult to meet the business needs of quickly generating clear control instructions at construction sites. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a knowledge graph-based system for diagnosing and preventing common quality defects in highway construction. This system solves the problem that existing methods for diagnosing common quality defects in highway construction mainly rely on manual experience or static rules, lacking a mechanism for dynamically associating and integrating real-time on-site perception data, process sequence constraints, and professional domain knowledge. This results in the process of identifying and diagnosing hidden dangers in complex construction environments lagging behind the actual construction progress, making it difficult to automatically generate prevention and control strategies that adapt to dynamic changes on-site.

[0006] To achieve the above objectives, this invention provides the following technical solution: a knowledge graph-based system for diagnosing and preventing common quality defects in highway construction, comprising: a knowledge graph construction module for processing engineering documents, extracting entities of quality phenomena, inducing factors, and prevention measures, matching them with a construction specification database to obtain quantitative thresholds, establishing directed connections, and constructing a static knowledge graph; a mask generation module for generating a mask matrix based on sensor data streams and quantitative thresholds, and generating a dynamic topology graph through operations between the mask matrix and the static knowledge graph; and a state initialization module for converting schedule documents into process-specific documents. The system employs a directed acyclic graph (DAG) model, which constructs a multi-agent network based on a dynamic topology graph and a process directed acyclic graph, and initializes the Markov decision process for the multi-agent network. A policy pre-training module maps hazard investigation form data to a static graph to generate expert demonstration trajectories and pre-trains the policy network. A temporal reasoning module constrains the multi-agent network's movement timeline within the dynamic topology graph based on the process directed acyclic graph, calculates the environmental reward function, and updates the policy network. A policy generation module maps query commands to starting entity nodes in the dynamic topology graph, invokes the policy network to perform forward inference, and outputs control messages.

[0007] Preferably, the graph construction module is specifically used for: extracting fuzzy semantic features from fuzzy semantic modifiers in engineering documents; constructing a construction specification database based on construction specification text, and converting specification clauses in the construction specification text into specification entries containing physical quantity identifier fields, fuzzy modifier fields, numerical lower limit fields, numerical upper limit fields, and standard unit of measurement fields; calculating the semantic similarity between fuzzy semantic features and specification entries; and extracting the numerical lower limit field, numerical upper limit field, and standard unit of measurement field from the matched specification entries based on the semantic similarity to obtain a quantitative threshold.

[0008] Preferably, the graph construction module is further configured to: perform named entity recognition on the engineering document to extract quality phenomenon entities, inducing factor entities, and prevention and control measure entities; based on the extracted quality phenomenon entities, inducing factor entities, and prevention and control measure entities, extract the relationships between entities and generate structured triples, and identify directed connecting edges; when the same inducing factor entity corresponds to multiple quantitative thresholds, perform weighted fusion of multiple quantitative thresholds according to the data source authority level and publication time of the corresponding engineering document to obtain the fused quantitative threshold; construct a static graph using structured triples, directed connecting edges, and the fused quantitative thresholds, and convert the static graph into an entity node feature matrix, a global adjacency matrix, and an edge feature tensor for persistent storage.

[0009] Preferably, the mask generation module is specifically used for: receiving a sensor data stream containing sensor output parameters; performing time synchronization, spatial alignment, outlier removal, and noise smoothing on the sensor data stream to obtain a continuous temporal feature tensor; traversing the directed connection edges associated with quantitative thresholds in the static graph; obtaining the continuous temporal feature tensor; comparing and determining the continuous temporal feature tensor with the quantitative thresholds associated with the directed connection edges; and generating a mask matrix corresponding to the current time step based on the comparison and determination results.

[0010] Preferably, the mask generation module is further configured to: calculate the normalized boundary offset based on the difference between the continuous temporal feature tensor and the quantitative threshold corresponding to the target entity node in the static graph; map the calculated normalized boundary offset to dynamic node mask weights through a smoothing nonlinear activation function; determine the connectivity weights of directed edges based on the dynamic node mask weights of adjacent entity nodes, and generate a dynamic edge mask matrix, wherein the mask matrix includes at least dynamic node mask weights and a dynamic edge mask matrix; perform a Hadamard product operation on the generated dynamic edge mask matrix and the global adjacency matrix of the static graph to obtain a dynamic global adjacency matrix, and combine it with the edge feature tensor to generate the dynamic topology graph at the current time step.

[0011] Preferably, the state initialization module is specifically used for: parsing the schedule file of the building information model, converting the schedule file into a directed acyclic graph of processes, the directed acyclic graph of processes including independent process nodes and process sequence dependency edges; extracting the process identifier, baseline duration, and resource consumption of independent process nodes, and constructing a graph node feature matrix; performing connectivity verification and loop elimination on the constructed directed acyclic graph of processes; dividing the multi-agent into multiple agent sets based on the material dimension, environmental dimension, and process dimension in the construction dimension; defining the graph diagnosis process as a Markov decision process, at any decision time step, concatenating the structural features of the dynamic topology graph, the features of the entity nodes currently occupied by the multi-agent, the features of the sensor data stream at the corresponding time, and the process constraint features representing the execution progress of independent process nodes, the remaining available amount of construction resources, and the time difference between the actual progress and the baseline plan, to obtain a global environment state tensor; and determining the state transition relationship based on the directed topology structure of the dynamic topology graph and the process constraints of the verified directed acyclic graph of processes.

[0012] Preferably, the strategy pre-training module is specifically used for: extracting quality phenomenon records, inducing factor records, prevention and control measure records, environmental condition records, and rectification result records from the hazard investigation form data; mapping the extracted hazard investigation form data to corresponding entity nodes in the static graph, and reconstructing the historical walking path based on the directed connection edges between the corresponding entity nodes to form an expert demonstration trajectory; converting the formed expert demonstration trajectory into state samples and action labels; using a behavior cloning algorithm, inputting the state samples and action labels into the multi-agent policy network, and pre-training the parameters of the policy network by calculating and minimizing the cross-entropy loss function.

[0013] Preferably, the temporal reasoning module is specifically used for: constraining the activation order and walk sequence of multi-agents based on the sequential dependency edges in the directed acyclic graph of the process; based on the walk sequence, concatenating the endpoint node features output by the multi-agents corresponding to the preceding process to the Markov state feature vector of the multi-agents corresponding to the following process; calculating an environmental reward function containing temporal constraints, the environmental reward function including a schedule benefit term determined by the number of days the critical path is shortened, a resource utilization rate term determined by the ratio of the actual consumption to the capacity limit within the remaining available construction resources during the action execution period, and a delay penalty term determined by the actual time exceeding the baseline plan; when calculating the environmental reward function, if the multi-agent action state conflicts with the rules of the directed acyclic graph of the process, a negative reward is calculated, and if the multi-agent walk path reaches the prevention and control measure entity, a positive reward is calculated.

[0014] Preferably, the temporal inference module is further configured to: perform local gradient calculation and policy network update using an asynchronous advantage action evaluation algorithm; construct a value network and calculate temporal difference error and advantage function value based on the value network and the environment reward function; calculate the gradient of the policy network objective function and the gradient of the value network loss function based on the temporal difference error and advantage function value; and calculate the gradients separately and asynchronously update the policy network through parallel worker threads to achieve model convergence.

[0015] Preferably, the strategy generation module is specifically used for: receiving a user's query command, parsing the query command for intent and extracting entities, and mapping it to a starting entity node in a dynamic topology graph; obtaining the dynamic topology graph corresponding to the moment the query command is received, and extracting the structural features of the dynamic topology graph, the features of the starting entity node, and the state features corresponding to the relevant sensor data stream at the current moment, and using the feature combination as input for the strategy network to perform forward inference; calling the strategy network to perform forward inference on the dynamic topology graph to obtain the cooperative walking trajectory of multiple agents in the dynamic topology graph; extracting the control parameters of the prevention and control measure entities at the corresponding endpoint positions based on the obtained cooperative walking trajectory; converting the extracted control parameters into control messages and outputting them to the construction equipment or field terminal, wherein the control message is a structured message and includes at least a unique instruction identifier, the target receiving end network address, the working face position, the operation effective time, the operation failure time, and the control parameters.

[0016] This invention provides a knowledge graph-based system for diagnosing and preventing common quality defects in highway construction. It offers the following advantages: 1. This invention uses a mask generation module to compare and calculate a dynamic edge mask matrix by linking continuous time-series sensor data with a quantitative threshold associated with a static graph. It then combines this with the global adjacency matrix of the static graph to generate a dynamic topological graph. By integrating static knowledge from the engineering field with real-time sensing data at the underlying matrix dimension, the system can dynamically adjust the weight structure of graph nodes and edges according to the actual environment on site. This solves the technical problem that traditional diagnostic models rely on static rules and are difficult to adapt to dynamic changes on site.

[0017] 2. This invention converts the schedule file into a directed acyclic graph of processes through a state initialization module and a temporal reasoning module. Based on the sequential dependency edges of this graph, it constrains the activation order and walk sequence of multiple agents in the dynamic topological graph. The temporal constraints of construction processes and resource availability are coupled with the graph computation model, ensuring that the diagnostic reasoning process is synchronized with the actual construction progress. The final output prevention and control measures meet the on-site schedule requirements and resource capacity constraints, improving the engineering feasibility of the generated strategy.

[0018] 3. This invention uses a strategy pre-training module to transform historical hazard investigation form data into expert demonstration trajectories to pre-train the policy network, and combines this with an asynchronous advantage action evaluation algorithm for parameter updates; it utilizes historical investigation records to provide initial boundary conditions for reinforcement learning, reducing ineffective exploration by multiple agents in the graph space and accelerating the convergence speed of the network model; at the same time, the system automatically infers and outputs structured messages containing the working face location and specific control parameters, reducing the reliance on human experience in the diagnosis of common quality defects. Attached Figure Description

[0019] Figure 1This is a system architecture diagram of the highway construction quality common defect diagnosis and prevention system according to an embodiment of the present invention; Figure 2 This is a flowchart of a method for diagnosing and preventing common quality defects in highway construction according to an embodiment of the present invention; Figure 3 This is a timeline diagram showing the evolution of real-time monitoring and system intervention of compaction speed in an embodiment of the present invention. Detailed Implementation

[0020] 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.

[0021] See attached document Figure 1 This invention provides a knowledge graph-based system for diagnosing and preventing common quality defects in highway construction. This system may include: The system includes a graph construction module, a mask generation module, a state initialization module, a policy pre-training module, a temporal reasoning module, and a policy generation module.

[0022] The graph construction module receives engineering documents. It extracts quality phenomenon entities, inducing factor entities, and prevention and control measure entities from the engineering documents and establishes directed connections between these entities.

[0023] The graph construction module constructs a set of fuzzy attributes for fuzzy semantic modifiers in engineering documents. It then uses a construction specification database to calculate the text feature vector similarity between the fuzzy attributes and the specification entries.

[0024] After matching the standard entries, the graph construction module extracts the corresponding precise numerical thresholds to generate a set of quantitative attributes. Based on entities, directed edges, and the set of quantitative attributes, the graph construction module constructs a static graph with quantitative thresholds for relational edges.

[0025] The mask generation module receives sensor data streams containing ambient temperature and humidity, wind speed, and device output parameters. The module iterates through directed edges with quantitative thresholds in the static graph.

[0026] The mask generation module compares the sensor data stream with the quantitative thresholds of the directed connections. Based on the comparison result, the mask generation module generates the mask matrix corresponding to the current time step.

[0027] The mask generation module performs a Hadamard product operation between the mask matrix and the global adjacency matrix of the static graph. Based on the operation result, the mask generation module generates a dynamic topological graph for the current time step.

[0028] The state initialization module is used to parse the schedule file of the Building Information Model (BIM). The state initialization module converts the BIM schedule file into a directed acyclic graph (DAG). The DAG contains independent process nodes and process dependencies along edges.

[0029] The state initialization module establishes a multi-agent set corresponding to the construction dimensions (material dimension, environmental dimension, and process dimension). The state initialization module defines the graph diagnosis process as a Markov decision process. In this invention, the state of the multi-agent Markov decision process is composed of the structural features of the dynamic topology graph, the features of the entity node where the agent is currently stationed, the sensor data stream features at the corresponding time, and the process constraint features representing the execution progress of the process node, the remaining available construction resources, and the time difference between the actual progress and the baseline plan. Its actions are defined as the agent's path selection among knowledge graph nodes; its rewards are calculated based on whether the temporal constraints conflict and whether the preventive measure node is reached. The state initialization module concatenates the graph topology features, the features of the currently stationed entity node, and the sensor data stream into a Markov state feature vector. The state initialization module determines the state transition relationship based on the directed topology structure of the dynamic topology graph and the process constraints of the directed acyclic graph of the process.

[0030] The policy pre-training module acquires data from hazard identification forms. It maps this data to a static graph to generate a set of walking paths serving as expert demonstration trajectories. The module employs a behavior cloning algorithm to pre-train the policy network parameters of the multi-agent system. The policy network is a deep neural network (e.g., a combination of graph neural networks and reinforcement learning policy network architecture) used to control the walking decisions of the multi-agent system. During the temporal inference phase, it is updated through the local gradients of each agent, thereby achieving iteration of the policy network (i.e., the global level) of the entire multi-agent system. The policy pre-training module updates the policy network parameters by calculating and minimizing the cross-entropy loss function.

[0031] The temporal reasoning module is used to constrain the activation order of multiple agents. It controls the timing of the agents' walks on the dynamic topological graph based on the sequential dependency edges in the directed acyclic graph of processes.

[0032] The temporal reasoning module concatenates the endpoint node features output by the agent corresponding to the preceding process into the Markov state feature vector of the agent corresponding to the following process.

[0033] The temporal reasoning module calculates the environmental reward function, which includes temporal constraints. It calculates negative rewards when there is a conflict between the agent's action state and the directed acyclic graph rules of the process, and positive rewards when the agent's path reaches a prevention and control measure node.

[0034] The temporal inference module employs an asynchronous advantage action evaluation algorithm for local gradient calculation and policy network updates. It calculates the multi-step temporal difference advantage function and the gradients of the policy network's objective function and the value network's loss function, respectively. The temporal inference module uses parallel worker threads to calculate these gradients and asynchronously update the policy network to achieve model convergence.

[0035] The policy generation module receives user query commands (such as natural language queries) through an interactive interface. It maps these queries to starting entity nodes in a dynamic topology graph. The module obtains the dynamic topology graph corresponding to the moment the natural language query command is received and invokes the policy network to perform forward inference. Based on the multi-agent cooperative walk trajectory, the module extracts control parameters for the corresponding endpoint prevention and control entities, converts these parameters into control messages, and outputs them.

[0036] See attached document Figure 2 Based on the above system, the macroscopic workflow of the present invention includes the following steps: Step S1: Process heterogeneous engineering data through the graph construction module, align structured construction specification data, and construct a static graph with numerical constraints. Step S2: Receive sensor data stream through mask generation module, generate mask matrix based on quantitative threshold of static map, and output dynamic topology map; Step S3: Extract the schedule file of the building information model through the state initialization module, construct the directed acyclic graph of the process, and initialize the state space and action space of the multi-agent Markov decision process. Step S4: The expert demonstration trajectory is generated by mapping the hidden danger investigation form data through the policy pre-training module, and the cold start pre-training of the multi-agent policy network parameters is completed by using the behavior cloning algorithm. Step S5: The temporal reasoning module calculates the environmental reward function based on the process sequence dependency edge constraint of the directed acyclic graph of the process, and updates the policy network using the asynchronous advantage action evaluation algorithm to complete the model training. Step S6: Receive the query command in natural language through the policy generation module, map the query command in natural language to the starting entity node in the dynamic topology graph, obtain the dynamic topology graph corresponding to the moment the query command in natural language is received, call the policy network to perform forward inference, and output a control message with control parameters of the entity for prevention and control measures.

[0037] In this embodiment, the map construction module performs deep analysis and structured processing on multi-source heterogeneous engineering data. The specific implementation steps are as follows: Step S101: Obtain unstructured engineering documents and perform data cleaning and word segmentation. The engineering documents include construction technical disclosure documents and construction organization design documents. The graph construction module filters out illegal characters such as non-Chinese characters, non-numerical characters, and non-standard punctuation marks in the documents using regular expressions, and uniformly converts the text encoding format to UTF-8. The graph construction module uses a forward maximum matching algorithm based on a domain dictionary to segment the cleaned text. To ensure the integrity of engineering terminology boundaries, the graph construction module loads proper nouns such as "spring soil" and "soft soil road section" into a custom construction domain dictionary recorded by historical engineering experts, preventing excessive segmentation by the basic word segmenter. For part-of-speech tagging and stop word removal during the word segmentation process, those skilled in the art can use open-source natural language processing toolkits; the underlying hidden Markov model mechanism is a well-known technology in this field and will not be elaborated upon here.

[0038] Step S102: Define the basic node types and construct the static graph architecture. Based on the causal chain between physical states and engineering behaviors, the graph construction module classifies entity categories into quality phenomenon entities, inducing factor entities, and prevention measure entities. Quality phenomenon entities represent negative physical states that occur during construction, such as longitudinal cracks in the pavement; inducing factor entities represent the preconditions that lead to this state, such as improper material mix proportions; prevention measure entities correspond to engineering operations that eliminate potential quality hazards, such as installing a drainage system.

[0039] Step S103: Perform named entity recognition to extract specific entities from the text. The graph construction module employs a sequence labeling model combining a bidirectional long short-term memory network (LSTM) and a conditional random field (CRF). During the inference phase, the input text sequence is converted into word vectors in continuous real-valued matrix format via a word embedding layer. These vectors are then input into the bidirectional LSTM network to extract context sequence features, and the output is concatenated to form a hidden state matrix of artificial neurons with matched channel dimensions. The CRF layer calculates the globally optimal label sequence based on this hidden state matrix and state transition matrix. The formula for calculating the conditional probability of the label sequence for a given input sequence is as follows: ; In the formula, A sequence of word feature vectors representing the input text; The corresponding entity classification label sequence; Characterizes the total length of the input sequence; Characterizes the number of characteristic functions; Characterizing the first Weight parameters of each feature function; Characterized at sequence position The local feature function used to evaluate the rationality of state transitions between adjacent labels; Characterize the normalization factor. Consider the total sequence length. The exponential accumulation of the normalization factor during addition may lead to loss of floating-point precision. Therefore, the underlying calculation employs a logarithmic summation exponential technique for equivalent algebraic transformation. The model is pre-trained using BIO annotation based on historical construction text and iteratively converges using the negative log-likelihood function as the loss function. The graph construction module extracts specific entities based on the solution of maximizing conditional probabilities.

[0040] Step S104: Extract the relationships between entities to generate structured triplet data. The graph construction module pre-defines two types of core directed connections: causal relationships where the starting point is the inducing factor entity and the ending point is the quality phenomenon entity; and prevention relationships where the starting point is the prevention measure entity and the ending point is the quality phenomenon entity. For regular dependencies, the graph construction module calculates the semantic dependency distance between co-existing entities using a dependency parsing algorithm, generating structured triples such as <improper soft foundation treatment at the bridgehead transition section leads to, bridgehead jump>. For long-distance dependencies across segments, the graph construction module uses a convolutional neural network with an attention mechanism for inference. The network input concatenates the word vectors and relative position vectors of the two entities, internally uses a one-dimensional convolutional kernel to extract local phrase features, and after feature dimensionality reduction, dynamically weights them through an attention mechanism layer, finally outputting the confidence probability through a fully connected layer. The graph construction module performs a weighted sum of the confidence probability and the inverse of the syntactic distance output by the dependency parsing. When the weighted total score exceeds the connection confidence threshold (ranging from 0.75 to 0.90) determined by historical benchmark data, a definite directed connection edge is determined between the two entities. After parsing and extraction, the unstructured document is transformed into a static graph containing entity nodes and directed connection edges, and stored in the graph database.

[0041] In this embodiment, the graph construction module performs attribute-level quantization alignment operations on the qualitative descriptions in the engineering text based on the previously extracted entity nodes, completing the mapping from fuzzy semantics to precise numerical ranges. The specific implementation steps are as follows: Step S105: Construct the specification database and define its internal organization. The graph construction module parses the current national highway construction specification text, transforming the mandatory requirements in the specification clauses into a relational data table containing fields for physical quantity identification, fuzzy modifiers, lower limits of values, upper limits of values, and standard units of measurement. For the clause in the specification that considers low-temperature construction when the daily average temperature is below 5℃ but above 0℃ as low-temperature construction, the graph construction module structures and stores it as a standard mapping record containing temperature, low temperature, 0℃, and 5℃, thereby centralizing and aggregating the scattered specification thresholds to form an underlying benchmark architecture that can be retrieved and used for numerical analysis.

[0042] Step S106: Extract fuzzy semantic features containing qualitative descriptions from engineering documents. In actual engineering documents, technicians often use qualitative terms lacking numerical constraints, such as strong winds and low temperatures. The graph construction module traverses the word sequences obtained from previous word segmentation using a dependency parsing algorithm, and selectively extracts adjective and noun combination nodes with modifying dependency relationships as fuzzy semantic entities to be mapped. This embodiment introduces a bidirectional encoder representation model based on the Transformer architecture to extract semantic features. To enable the model to have professional semantic understanding capabilities, the graph construction module pre-collects historical specifications and engineering handover documents as an unsupervised corpus, and pre-trains the underlying parameters through a masked language model mechanism; further, it fine-tunes the network weights using a contrastive loss function through manually constructed positive sample pairs (e.g., <low temperature, cold climate>) and negative sample pairs (e.g., <low temperature, heavy rainfall>). During the inference phase, the model receives a fixed-length character sequence as input. Internally, it is mapped to an initial dense vector via a word embedding layer. Then, a multi-head attention mechanism layer dynamically weights the contextual dependencies between characters, ultimately extracting a 768-dimensional continuous real-valued vector representing the semantic feature at the pooling output. For the feature concatenation and linear dimensionality reduction operations of the multi-head attention mechanism, those skilled in the art can implement them using built-in operators in existing deep learning frameworks. The forward propagation computation graph logic is a well-known technique in this field and will not be elaborated upon here.

[0043] Step S107: Calculate the semantic similarity between the entity to be mapped and the standard records in the canonical database. The graph construction module traverses the fuzzy modifier field in the canonical database, synchronously inputting it into the feature extraction model to transform it into a 768-dimensional standard feature vector. The graph construction module uses a cosine similarity algorithm to measure the directional consistency between the feature vector to be mapped and the standard feature vector in multidimensional space. To prevent the feature vector from having many zero elements due to the extremely short text, which could lead to floating-point overflow with the denominator approaching 0, the graph construction module introduces a very small constant in the denominator term of the basic cosine formula for arithmetic smoothing. The specific calculation formula is as follows: ; In the formula, The cosine similarity score represents the relationship between the feature vector to be mapped and the standard feature vector. Characterize the feature vector to be mapped; Characteristic eigenvectors; The total dimension of the feature vector is 768. and Representing the two eigenvectors at the th... Component values ​​in each dimension; A very small constant representing the minimum value that prevents division by zero, with a range of 10. -8 Up to 10 -6The specific value is determined by the lower limit of the floating-point precision of the underlying processor. This mechanism, by calculating the cosine of the angle between high-dimensional vectors, can eliminate the interference of absolute word frequency on semantic evaluation, and objectively assess the degree of closeness of text expression to normative standards in the semantic space dimension.

[0044] Step S108 executes the mapping mechanism and threshold determination for precise numerical intervals. The graph construction module introduces multi-dimensional weighted determination logic, combining the semantic hierarchical similarity of feature vectors with the normalized Levenstein edit distance at the character level to calculate literal structural similarity. The graph construction module assigns weight factors of 0.75 and 0.25 to the cosine similarity score and literal structural similarity, respectively, and performs linear weighted summation. The specific value of this weight factor is determined based on two-dimensional grid search optimization of historical validation datasets, aiming to use shallow literal structural features for auxiliary correction while prioritizing deep semantic features. When the comprehensive link score is greater than the preset entity alignment threshold, the graph construction module determines that the semantic link is successful. As a preferred method, the value range of this entity alignment threshold is set between 0.82 and 0.88, and the specific value is determined by backtracking from test points that achieve the optimal solution of harmonic mean between precision and recall during cross-validation. After the determination is completed, the graph construction module retrieves the corresponding upper and lower limits of values ​​and standard units of measurement from the canonical database, directly mapping the fuzzy semantics to precise numerical intervals. The graph construction module ultimately uses the generated precise numerical range as a quantitative attribute label and rigidly attaches it to the corresponding inducing factor entity node in the static graph, completing the transformation of qualitative construction experience into quantitative standard benchmarks, and providing numerical constraint support for the subsequent quantitative reasoning of abnormal states between nodes.

[0045] In this embodiment, the graph construction module, based on the entity nodes, logical relationships, and quantized numerical ranges extracted in the previous analysis, performs global topology assembly and fusion conflict resolution, and finally generates a static graph with numerical constraints. The specific implementation steps are as follows: Step S109 involves instantiating the attribute graph of the graph topology. The graph construction module uses the underlying data architecture of the attribute graph model, mapping the extracted quality phenomenon entities, inducing factor entities, and prevention and control measure entities to independent vertices in the static graph, and mapping the causal relationships and prevention and control relationships to directed edges connecting the vertices. To achieve the graph representation of numerical constraints, the graph construction module uses the aforementioned generated precise numerical ranges (including lower and upper limits and standard units of measurement) as attribute key-value pairs, and stores them as associated values ​​within the corresponding inducing factor entity nodes or relation edges with defined conditions. Through this mapping logic, the discrete text information network is transformed into a multi-dimensional attribute graph topology containing node classification, directed association logic, and quantification parameters.

[0046] When establishing a prevention and control measure entity, the graph construction module simultaneously parses the engineering specifications and emergency response plan, extracts the corresponding guidance recovery values ​​or adjustment step sizes (such as target set values, acceleration / deceleration ratios, etc.) as quantitative attributes, and attaches them to the prevention and control measure entity node so as to generate control messages in the future.

[0047] Step S110 involves performing knowledge fusion and conflict resolution for multi-source numerical constraints. In actual engineering, the same inducing factor entity (such as temperature anomaly) may correspond to multiple different quantitative thresholds in multiple engineering documents. In this case, the system obtains the data source authority level and publication time of each engineering document and performs weighted fusion accordingly. When processing multi-source heterogeneous engineering documents, the same inducing factor parameter (such as pavement compaction degree or loose paving thickness) often corresponds to different numerical ranges in different levels of specifications. If only the extreme values ​​of a single data source are relied upon for one-sided judgment, it can easily lead to logical inconsistencies in subsequent diagnostic reasoning. The atlas construction module introduces a weighted fusion algorithm based on spatiotemporal features and authority level to comprehensively evaluate the relative reliability of each data source to calculate the optimal benchmark numerical threshold. The specific calculation formula is as follows: ; In the formula, The quantitative threshold for characterizing fusion and digestion; It represents the total number of data sources retrieved for the same engineering physical quantity; Characterizing the first The original numerical threshold provided by each data source; Characterizing the first The basic authority weighting factor for each data source. As a preferred approach, the weighting factor is set at 0.9 for national mandatory standards, 0.7 for industry standards, and 0.5 for enterprise internal control documents. Characterizing the first The absolute span of the data source publication time from the current year is used to reflect the timeliness of the standard; The time decay coefficient, ranging from 0.1 to 0.3, is determined by linear regression analysis of the impact of historical specification changes on core construction parameters. A very small constant representing the prevention of division by zero, with a value set to 10. -6 To ensure the stability of division operations under extreme single-source data or weight distortion conditions, this formula integrates the authority of the standard with the timeliness of the data to form a comprehensive confidence weight, which smoothly weights the multi-source conflict constraint threshold. A negative exponential time decay model is introduced to ensure that the weights of older standard guidelines decrease exponentially, outputting a quantitative benchmark that balances minimum requirements with the latest technological advancements.

[0048] Step S111 involves performing matrix-based serialization and persistent storage of the static graph. The wandering storage format of the graph database cannot directly participate in tensor operations in deep learning networks. To support tensor-based reinforcement learning operations in subsequent modules, the graph construction module converts the assembled static graph into a standard mathematical matrix format. The graph construction module generates a dimension based on the topological connectivity of the graph. The global adjacency matrix ( (Representing the total number of entity nodes). To optimize the underlying computational memory, the graph construction module uses a compressed sparse row format to encode the global adjacency matrix: if there is a directed connection between two nodes, the corresponding matrix coordinate element is assigned a value of 1; otherwise, it is 0. For directed edges with numerical constraints, the graph construction module synchronously generates an edge feature tensor and serializes the quantitative threshold of the fused output into continuous real-valued components on a specific channel of this tensor. During this process, the graph construction module ensures that the spatial coordinate indices of the edge feature tensor are strictly aligned with the coordinates of the non-zero elements of the global adjacency matrix to prevent feature misalignment in subsequent graph convolution operations. After serialization, the graph construction module submits the complete data structure containing the entity node feature matrix, the sparse global adjacency matrix, and the edge feature tensor to the underlying graph database. At this point, the static graph with numerical constraints is completed, serving as a standardized basic environment input to support subsequent real-time sensor data stream verification and topology pruning operations.

[0049] In this embodiment, the mask generation module performs multi-source time-series data acquisition, synchronization, and noise reduction based on the underlying hardware sensing network, converting physical monitoring signals into tensor inputs that can be used for graph calculation. The specific implementation steps are as follows: Step S201 involves the hardware deployment and time-series data acquisition of multi-source heterogeneous sensing devices. Addressing the complex physical environment of the construction site, the mask generation module deploys various IoT sensing devices on-site, including a roadbed compaction monitor mounted on the road roller, temperature and humidity sensors embedded within the roadbed structure layer, and a paver operation parameter acquisition terminal. The mask generation module uses industrial Ethernet and wireless radio frequency communication technology to collect real-time raw continuous time-series data covering physical states such as compaction pass count, soil moisture content, ambient temperature, and paving speed, and simultaneously acquires high-precision spatial positioning coordinates of each sensing terminal. For the interface protocol conversion and basic analog-to-digital conversion of the hardware sensors, those skilled in the art can use existing programmable logic controllers; the underlying communication handshake mechanism is a well-known technology in the field and will not be elaborated upon here.

[0050] Step S202 involves performing time synchronization and spatial condition feature alignment operations on multi-source sensor data. The sampling frequencies of different types of equipment at the construction site exhibit heterogeneity. If asynchronous time sequences are not strictly aligned, physical misalignment of data on the time axis will directly lead to time-series reversal in the subsequent dynamic mask topology trimming logic. The mask generation module establishes a globally unified time reference axis, uniformly formatting the absolute time of the incoming data into millisecond-level timestamps. For low-frequency sampling sequences, the mask generation module uses a time-distance weighted cubic spline interpolation algorithm for upsampling and completion; for high-frequency sampling sequences, a mean downsampling mechanism based on the construction displacement interval is introduced. To accurately reflect the physical working surface of the equipment, the mask generation module combines real-time dynamic differential positioning data to map the sampling points after unified timestamps onto a preset two-dimensional construction grid, which is composed of vertical mileage markers and horizontal offsets. If multiple overlapping sensing data trajectories exist within the same grid, the mask generation module performs state coverage in chronological order. Through the aforementioned spatiotemporal resampling and coordinate alignment logic, frequency-heterogeneous multi-source sensing sequences are mapped and transformed into structured continuous temporal feature tensors with a unified time step and a rigorous spatial physical correspondence.

[0051] Step S203 involves performing local outlier removal and noise smoothing on the aligned continuous temporal feature tensors. Considering that strong electromagnetic interference or physical vibrations of equipment can easily cause transient data spikes, the mask generation module introduces a robust sliding window smoothing algorithm combining absolute median difference to clean the continuous temporal feature tensors. This mechanism utilizes the median's insensitivity to extreme isolated noise, calculating the absolute deviation between the current time step sample value and the local median within a set sliding window, and using the absolute median difference within the window as a dynamic judgment benchmark. The specific formula for calculating the local robust outlier score is as follows: ; In the formula, Characterization at time step Local robustness anomaly score; The original sensor sample value representing the current time step; This represents the median of the sampled values ​​within the current sliding window; This represents the absolute median difference within the current sliding window; The scaling factor is used to approximate the absolute median deviation as a standard deviation statistic. As a preferred approach, this factor is set to 1.4826, determined based on the asymptotic consistency derivation under the assumption of a normal distribution. A very small constant representing the prevention of division by zero, with a value set to 10. -6This is designed to prevent overflow in the underlying division operation when the absolute median difference is 0 due to completely identical sampled values ​​within the window. The physical meaning of this formula is that, by comparing the median fluctuation range, it removes destructive noise while preserving the true physical trend of the engineering data to the greatest extent possible.

[0052] In the anomaly detection stage, to avoid biased misjudgments caused by relying solely on a single extreme value, the mask generation module assigns anomaly scores. The data is compared with a preset anomaly truncation threshold and a multi-dimensional judgment is performed in conjunction with the physical rate of change. The anomaly truncation threshold is set between 3.0 and 3.5, with the specific value determined by backtracking from the noise normal distribution information interval of the field equipment calibration experiment. The mask generation module synchronously calculates the slope of the numerical change of the current time step relative to the previous time step. Only when the local robust anomaly score is greater than the anomaly truncation threshold, and the slope of the numerical change exceeds the maximum rate boundary set by the mechanical response limit of the equipment, the mask generation module determines that the corresponding sample value is hardware glitch noise and removes it. Then, the forward linear interpolation result of adjacent sampling points is used for numerical reconstruction. After the above acquisition, alignment and noise reduction operations, the unstructured monitoring data is transformed into a numerically stable continuous temporal feature tensor, providing reliable real-time numerical constraint support for the subsequent dynamic mask pruning of the map topology.

[0053] In this embodiment, the mask generation module receives the preprocessed continuous temporal feature tensor, compares it with the quantization benchmark in the static graph through a logic decision engine, and generates a dynamic mask matrix to characterize the topological activation state. The specific implementation steps are as follows: Step S204: Calculate the normalized boundary offset of the physical state. After acquiring the multi-source aligned sensor data stream, the mask generation module extracts the upper and lower numerical limits of the target entity nodes from the static map and compares them with the continuous temporal feature tensor input at the current time step. Since the physical measurement units of different inducing factors vary significantly, directly using the absolute difference can easily lead to the gradient update during backpropagation of the neural network being dominated by large-scale physical quantities. Therefore, the mask generation module calculates the dimensionless normalized boundary offset. The normalized boundary offset refers to the absolute value of the difference between the sensor's continuous temporal feature tensor (e.g., current real-time temperature) and the quantitative threshold (e.g., upper / lower limit of the specification) extracted from the static map, after normalization. The larger the deviation from the threshold, the higher the mask weight of the node. This parameter objectively quantifies the severity of the real-time construction data exceeding the specification benchmark in a multi-dimensional continuous space. The specific calculation formula is as follows: ; In the formula, Characterizing target entity nodes At time step Normalized boundary offset; The real-time sensing feature value corresponding to the target entity node at the current time step; and These respectively characterize the upper and lower limits of the normative values ​​corresponding to the target entity node retrieved from the static graph; A very small constant representing the prevention of division by zero, with a value set to 10. -6 This mechanism aims to prevent overflow in division operations caused by overlapping upper and lower threshold values, resulting in a denominator of 0. It extracts out-of-bounds numerical components and scales them proportionally according to the tolerance range, outputting a standardized dimensionless index reflecting the degree of deterioration in construction conditions.

[0054] Step S205 generates a continuously differentiable topological node mask vector. Traditional binary mask hard truncation mechanisms can easily lead to zero partial derivatives during gradient descent operations, blocking the error backpropagation path of subsequent temporal inference modules. To support end-to-end training of the deep learning framework, the mask generation module introduces a continuous mapping mechanism based on a smooth nonlinear activation function, transforming the normalized boundary offsets into node mask weights with continuous gradients. The mapping formula is as follows: ; In the formula, Characterizing target entity nodes At time step The dynamic node mask weights are projected onto a continuous real number field from 0 to 1. The slope amplification factor is used to control the steepness of the activation mapping curve. Characterizing target entity nodes At time step The normalized boundary offset. As a preferred approach, this factor is set to a value between 10 and 15 to ensure sufficient numerical discrimination near the boundary critical point; The soft activation threshold is set to 0.15, optimized based on the statistical distribution of historical false alarm samples. This threshold aims to filter out minor permissible operational errors or residual hardware jitter, preventing local noise from falsely triggering abnormal nodes in the graph. After mapping operations, the mask generation module generates a one-dimensional dynamic node mask vector for all entity nodes in the static graph.

[0055] Step S206: Perform matrix Hadamard product operation and global topology dynamic pruning. After obtaining the node-level activation weights, the mask generation module extracts the previously constructed global adjacency matrix and derives the activation state of directed edge relationships. The mask generation module constructs a dynamic edge mask matrix of the same dimension as the global adjacency matrix through a tensor broadcast mechanism. Considering that common engineering quality defects often extend from single-point anomalies to adjacent entity nodes, the mask generation module uses the logic of taking the maximum value to determine the connectivity activation state of adjacent entity nodes: the mask generation module compares the target entity nodes associated at both ends of the directed edge. With adjacent entity nodes The dynamic node mask weights are calculated, and the maximum value is selected as the connectivity weight of the edge with that relationship. Subsequently, the mask generation module performs a Hadamard product operation (i.e., multiplying corresponding elements of the matrix) on the extracted connectivity weights and the inherent topological connectivity states (values ​​of 0 or 1) in the global adjacency matrix, thereby generating a trimmed dynamic global adjacency matrix. This operation preserves the canonical topological constraints of the original graph's underlying structure while eliminating irrelevant nodes and edge relationships in the current time step that are in a normal construction state (decaying their connectivity weights to near 0), highlighting quality-abnormal subgraphs in the static network. Finally, the mask generation module combines the generated dynamic global adjacency matrix with the real-time edge feature tensor to form a dynamic topological graph reflecting the actual working conditions of the current physical site, and then transfers the dynamic topological graph to the subsequent network for deep feature propagation and temporal inference.

[0056] In this embodiment, after obtaining the topological activation state of the map, the mask generation module performs tensor fusion of multimodal features and synthesis of map snapshots. The specific implementation steps are as follows: Step S207 involves performing vectorized concatenation and nonlinear fusion of multimodal node features. Based on the high-frequency sensing tensor and static knowledge base, the mask generation module aligns the sensor signals collected on-site with the inherent canonical attributes in the static map at the feature level. The mask generation module extracts predefined basic entity attributes from the static map and concatenates them with the previously aligned sensor values ​​in terms of dimension. To eliminate the dimensional differences between heterogeneous data, the mask generation module introduces a multilayer perceptron architecture (including an input layer, a single-layer hidden layer, and an output layer) to perform feature reconstruction and dimensionality reduction fusion on the concatenated tensor. The total dimension of the input concatenation vector is set to... The target dimension of the output fusion feature is In terms of specific computational logic, the mask generation module concatenates the real-time sensing feature value corresponding to the target entity node at the current time step with the static attribute embedding vector of the target entity node extracted from the static graph; then, it uses the feature fusion weight matrix (whose dimensions are configured as follows) to perform the following calculations. Perform a linear mapping on the concatenated vectors and add a dimension of . The feature fusion bias vector is obtained; finally, the output is activated by a non-linear activation function (preferably ReLU), thus obtaining a dimension of The fused feature vector of the target entity node at the current time step. For the initialization of the weight matrix and the configuration of the bias vector of the neural network, those skilled in the art can use conventional parameter initialization mechanisms such as normal distribution. The parameters will be iteratively updated end-to-end through the error backpropagation algorithm in the subsequent network training stage, and the relevant underlying network parameter settings are well known in the art and will not be described in detail here.

[0057] Step S208 involves performing dynamic modulation of the edge feature tensor based on mask weights. The mutual influence of engineering conditions is constrained by the structural connectivity, spatial distance, and business association strength between nodes. The mask generation module extracts the pre-set static semantic relationship features from the static graph and dynamically modulates them using the previously derived edge connectivity weights. When the physical interaction between adjacent entities is determined to be safe (i.e., the connectivity weight approaches 0), the corresponding edge feature propagation channel is synchronously attenuated and blocked. In the specific feature modulation logic, the mask generation module first uses the weight matrix and bias vector of the relation feature mapping to perform linear spatial mapping on the inherent static semantic relationship embedding vector between the target entity node and adjacent entity nodes; then, it multiplies and fuses the vector output by the mapping with the edge connectivity weight of the corresponding coordinate in the dynamically generated global adjacency matrix after pruning, thereby obtaining the dynamic relation feature vector between the target entity node and adjacent entity nodes at the current time step. This computational mechanism ensures that the relation edges in the dynamic topology graph have topological scalars representing connectivity and contain multi-dimensional feature vectors reflecting physical attributes.

[0058] Step S209: Synthesize a dynamic topological graph sequence and construct a temporal tensor set for deep inference. After obtaining the node fusion features and modulated edge features, the mask generation module uses the current absolute timestamp as a reference to package and map the dynamic global adjacency matrix, the fusion feature vectors of all nodes, and the dynamic relation feature vectors of all relation edges, outputting a dynamic topological graph representing the actual working conditions of the current construction site. To support evolution trend prediction, the mask generation module introduces a sliding time window mechanism, stacking the graph snapshots within the current time step and its historical preset number of time steps in chronological order to construct a dimension-aligned three-dimensional spatiotemporal graph tensor sequence. This spatiotemporal graph tensor sequence preserves the numerical evolution trend of sensor data and the spatial propagation path of anomalies in the topological structure, providing feature support for subsequent coupled inference of the graph neural network and the temporal logic model.

[0059] In this embodiment, the state initialization module receives the building information model and initial schedule data of the engineering project. By parsing the topological dependencies between processes, it constructs a directed acyclic graph of processes to constrain the action space of the reinforcement learning agent. The specific implementation steps are as follows: Step S301: Perform joint parsing and node feature extraction of the building information model (BIM) schedule file. After obtaining the BIM schedule file for the project, the state initialization module traverses the task entries in the file, mapping them to independent entities in a directed acyclic graph (DAG). Based on the BIM data architecture, the state initialization module extracts baseline parameters for each construction operation (including operation identifier, baseline duration, and resource consumption). To provide a fixed-dimensional state observation space for the subsequent reinforcement learning environment, the state initialization module concatenates the extracted baseline parameters in node order to construct a dimensionless array. The graph node feature matrix. Wherein, The total number of process nodes contained in the directed acyclic graph representing the process; The baseline feature dimension is used to characterize the extracted features of a single node. For the basic format parsing and attribute mapping of Building Information Modeling (BIM), those skilled in the art can utilize the application programming interfaces (APIs) of conventional engineering software; this is well-known technology in the field and will not be elaborated upon here.

[0060] Step S302 involves performing a matrix representation of topological dependencies and constructing temporal constraints. Strict temporal constraints exist between various construction processes on-site. The state initialization module parses the flow logic relationships in the schedule file of the Building Information Model (BIM) and constructs a global adjacency matrix representing the process flow path. To impose strict action legality boundaries, the state initialization module establishes pre-constraint equations in the time dimension based on the global adjacency matrix. The specific temporal constraint judgment logic is as follows: the planned start time of any subsequent process node must not be earlier than the maximum value of the actual completion times of all its immediate predecessor process nodes; where the expected / predicted completion time (or actual completion time) of a single predecessor process node is equal to the sum of its actual start time (or planned start time), actual duration (or baseline duration), and the mandatory time interval between the predecessor and the subsequent process node. It should be noted that the range of this mandatory time interval is set to a set of real numbers, determined by the process overlap attribute in the BIM (negative values ​​are allowed when overlapping construction occurs before the predecessor process is completed, and positive values ​​are required when strict waiting is required). This determination mechanism physically restricts the triggering time of any subsequent process from being later than the actual completion time of all its dependent preceding processes, ensuring that the sequence of actions conforms to the spatiotemporal logic of construction.

[0061] Step S303: Perform connectivity verification and loop elimination for the directed acyclic graph (DAG) of the process. The schedule file of the Building Information Model (BIM) is prone to circular dependency conflicts due to human input, causing the agent's state transition to fall into a logical dead loop. The state initialization module uses matrix nullability to perform loop verification on the constructed DAG of the process. In the DAG of the process, the constant power of the global adjacency matrix of the schedule maps the reachability of any two nodes at a specific step length. To optimize computational overhead, the state initialization module performs matrix exponentiation to verify connectivity after converting the global adjacency matrix of the schedule to a sparse tensor format. The specific verification logic is as follows: calculate the power of the aforementioned global adjacency matrix of the schedule, and set the exponent to the total number of process nodes contained in the DAG of the process; if the result matrix output by this exponentiation is a zero matrix of the same dimension as the original global adjacency matrix of the schedule, then there is no closed loop in the verification graph. When this verification condition is not met (i.e., the result matrix contains non-zero elements), the state initialization module triggers a logic blocking instruction, outputting the node sequence that caused the circular conflict for correction. After the re-verification is successful, the state initialization module outputs a compliant directed acyclic graph of the process to the reinforcement learning state space construction stage.

[0062] In this embodiment, after completing the directed acyclic graph parsing of the process, the state initialization module further transforms the dynamic graph state information and process temporal constraint information into a Markov decision state space that can be processed by deep reinforcement learning algorithms. To support collaborative diagnostic and prevention reasoning by multiple agents on the dynamic topological graph, the state initialization module establishes a multi-agent heterogeneous observation model. The specific implementation steps are as follows: Step S304: Perform global environment state tensor encapsulation based on spatiotemporal constraints. In the multi-agent reinforcement learning framework, the state space of the Markov decision process needs to fully represent the system boundary conditions of the current decision time step. The state initialization module pre-establishes the association mapping relationship between process nodes and quality phenomenon entities, inducing factor entities, and prevention and control measure entities in the dynamic topology graph, and performs converged encoding on the corresponding dynamic topology graph features according to the order of process nodes.

[0063] At any decision time step, the state initialization module performs vector concatenation on the feature dimension of the dynamic topology graph, the features of the entity nodes currently occupied by the agent, the sensor data stream features at the corresponding time, and the constraint features representing the execution progress of the process nodes, the remaining available construction resources, and the time difference between the actual progress and the baseline plan, to obtain the global environment state tensor. Specifically, the execution progress feature represents the unstarted, in progress, and completed states of the process nodes; the resource feature represents the remaining available construction resources; and the delay constraint feature represents the time difference between the current actual progress and the baseline plan. Through the concatenation operation, the output dimension of the state initialization module is... The global environment state tensor. In the formula, The total number of process nodes contained in a directed acyclic graph representing the process. The comprehensive feature dimension characterizing a single process node is jointly determined by the feature dimensions of the dynamic topology graph, the entity node, the sensor data stream, and the process constraint features. This encapsulation mechanism primarily uses the dynamic topology graph state features and secondarily uses the process constraint features, projecting graph diagnostic information and construction process constraints into a unified tensor space, providing a structurally aligned comprehensive working condition feature base for multi-agent collaborative reasoning.

[0064] Step S305 involves performing topological mask pruning of the heterogeneous observation space for multi-agent systems. Directly inputting the full map features into each agent can easily lead to excessive expansion of the action search space due to redundant nodes. Since individual specialized agents only need to focus on their closely related local tasks, the state initialization module utilizes a service mask mechanism constructed using a reachability matrix for state pruning. For the first... For each agent, the state initialization module constructs a topological reachability mask matrix associated with its business scope (this matrix has the same dimension as the global environment state tensor). The elements of this mask matrix are determined by the preceding and subsequent values ​​within the agent's business scope. The weight of a node within a connected subgraph is determined by skipping associated nodes. Associated nodes within the connected subgraph are assigned a weight of 1, while non-associated nodes are assigned a weight of 0. The depth of the graph topology search is represented by a positive integer, ranging from 1 to 3. Its specific value is positively correlated with the maximum topological association span of the relevant entity nodes under the corresponding construction dimension of the agent. After obtaining the mask matrix, the state initialization module performs a Hadamard product operation (i.e., multiplies corresponding elements of the matrix) on the aforementioned global environment state tensor and the topology reachability mask matrix, filtering out state noise irrelevant to the current diagnosis and prevention decision, thereby outputting the... The local observation tensor received by an agent at the current time step.

[0065] Step S306: Perform graph convolutional dimensionality reduction and dense feature mapping on the local observation tensor. Since the output local observation tensor after pruning is sparse, and the local observation tensor already contains dynamic topological graph features, the state initialization module introduces a state embedding network module containing graph convolutional network layers and adaptive pooling layers to perform high-dimensional information compression and feature density reconstruction. During this process, the progress global adjacency matrix is ​​used to apply adjacency propagation under process timing constraints to the dynamic topological graph features. The specific feature mapping logic is as follows: The state initialization module derives the normalized Laplacian matrix based on the aforementioned global adjacency matrix and the node relationships in the dynamic topology graph at the current time step. To avoid computational overflow caused by isolated nodes resulting from the pruning operation having a zero degree denominator, the state initialization module forcibly injects a unit self-loop mechanism into the main diagonal of the adjacency matrix during calculation, ensuring that the degree constant of all nodes is strictly greater than or equal to 1.

[0066] Subsequently, the state initialization module will use the normalized Laplace matrix, the first... The local observation tensor of each agent, and the learnable weight matrix of the graph convolutional layer (its dimension is strictly limited to 1). Multiply the three together by a matrix and add the dimension to the result. The graph convolutional layer bias vector is used; finally, a non-linear activation function (preferably LeakyReLU to preserve negative gradient flow) is applied to output the dimensionality-reduced dense state embedding vector. The target dimension of this embedding vector is configured as follows: This is intended to be directly used as the input to the policy network of subsequent asynchronous advantage action evaluation algorithms. The underlying feature decomposition and frequency domain filtering mechanisms of graph convolutional networks are well-known technologies in this field and will not be elaborated here.

[0067] In this embodiment, to address the problem of blind exploration and low convergence efficiency of reinforcement learning agents in complex construction scheduling environments during the initial stage, the policy pre-training module introduces a behavior cloning mechanism based on hazard identification form data for cold-start pre-training of the model. This module initializes the parameters of the policy network using experience from hazard identification and prevention in similar past projects. The specific implementation steps are as follows: Step S401: Perform feature mapping and expert trajectory reconstruction of the hazard investigation form data. The strategy pre-training module performs structured parsing and expert trajectory reconstruction of the hazard investigation form data. This step extracts quality phenomenon records, inducing factor records, prevention and control measure records, environmental condition records, and rectification result records from the hazard investigation form data, transforming discrete historical handling experience into supervised learning ground truth labels. The strategy pre-training module maps the hazard investigation form data to corresponding entity nodes in the static graph, and reconstructs the historical walking path from inducing factor entities or quality phenomenon entities to prevention and control measure entities based on the directed connection relationships between entity nodes, forming a dataset of expert demonstration trajectories. The strategy pre-training module further encapsulates the environmental condition features, graph node state features, and graph topological association features corresponding to each time node into state samples, and encapsulates the walking actions or target prevention and control measure entity labels between graph nodes into action labels, thereby constructing state-action pairs for behavior cloning training. As a preferred approach, state samples are pre-transformed using a state embedding network module to ensure that their dimensions are consistent with the dense state embedding vectors in the subsequent reinforcement learning environment; the corresponding action labels are represented by one-hot encoding to indicate the node's walk decision or prevention measure selection results, serving as the physical truth for supervised learning.

[0068] Step S402: Perform feature mapping and action probability distribution inference for the policy network. For the constructed expert trajectory dataset, the policy pre-training module configures a multilayer perceptron as the policy network for behavior cloning. This policy network receives the state embedding vector output by the state embedding network module and outputs the probability distribution of each candidate walking action through nonlinear feature space transformation of the internal hidden layers. Specifically, the network forward propagation logic is as follows: the policy pre-training module maps the feature space of the model to a multilayer perceptron with dimensions of [missing information]. State embedding vectors and hidden layer weight matrices (dimension strictly limited to 1) ,in Multiply the matrix by the total number of hidden layer nodes, and add the matrix with dimension 1. The bias vector; the result is mapped by a non-linear activation function, and then compared with the output layer weight matrix (with dimensions configured as follows). ,in Multiply by the total number of optional walk actions in the reinforcement learning action space, and add the dimension of . The bias vector; finally, the output dimension is obtained by normalizing the exponential function (Softmax). The network architecture compresses and maps the high-dimensional engineering environment state into the execution tendency of graph node walking actions or prevention and control measure selection actions, so that each element of the output vector can quantitatively characterize the confidence of the agent in selecting the corresponding control decision action.

[0069] Step S403 involves performing cross-entropy-based loss calculation and network parameter updates. To ensure the output distribution of the policy network approximates the historical decision-making logic of human experts, the policy pre-training module sets a cross-entropy loss function with a structural risk penalty term to guide model training. The specific calculation logic is as follows: for each training batch (batch size...) (Set to a positive integer between 32 and 256), the policy pre-training module calculates the sum of the products of the true physical values ​​of the action labels of all samples in the batch (i.e., the aforementioned one-hot vector elements, taking values ​​of 0 or 1) and the logarithms of the predicted probability values ​​of the policy network, and takes the negative average as the base cross-entropy. To prevent the logarithmic function from causing negative infinity calculation anomalies due to the probability approaching zero, the policy pre-training module introduces a value fixed at 1×10 in the logarithmic calculation term. -8 The minimum constant. Based on this, the policy pre-training module combines the squared L2 norm of the network parameters with the regularization coefficient (preferably within the range of 10). -4 Up to 10 -3 The penalty term is obtained by multiplying the two (dynamically determined by cross-validation) and then adding it to the basic cross-entropy to finally output the comprehensive loss value of the current training batch.

[0070] After obtaining the comprehensive loss value, the policy pre-training module calls the backpropagation algorithm to calculate the target gradient and uses the Adaptive Momentum Optimization (Adam) algorithm to iteratively update the set of all learnable parameters in the policy network and state embedding network modules. To prevent the model from getting stuck in ineffective iterations in the later stages of training, the policy pre-training module introduces an early stopping mechanism, which monitors the relative decay rate of the comprehensive loss value on the independent validation set. When the decay rate is lower than 1×10 for five consecutive training epochs, the module stops training early. -4 When the tolerance threshold is reached, the policy pre-training module blocks the training process. After training, the policy pre-training module saves the converged network parameters as the weight basis for subsequent asynchronous relay exploration by the reinforcement learning agent in the simulation environment. The chain rule for differentiation and the gradient descent mechanism in backpropagation can be implemented by those skilled in the art using the underlying interfaces of conventional deep learning frameworks; these are well-known technologies in the field and will not be elaborated upon here.

[0071] In this embodiment, after completing the cold-start pre-training of the policy network, the temporal reasoning module loads multiple agents into a diagnostic and prevention reasoning environment constrained by a directed acyclic graph of processes for asynchronous reinforcement learning exploration. To ensure that control decision actions conform to objective physical temporal constraints, the temporal reasoning module establishes a temporal relay deduction and multi-dimensional reward mechanism based on the directed acyclic graph of processes. The specific implementation steps are as follows: Step S501 involves performing agent activation screening and action space masking based on the constraints of the directed acyclic graph of processes. This step determines the activation eligibility of each agent at the current time step through hard constraints of preconditions. The temporal reasoning module traverses all process nodes according to the edge relationships of the directed acyclic graph of processes, extracting the current state of the precondition nodes related to the corresponding process for each agent. Only when all precondition nodes have been completed and the remaining available construction resources meet the minimum requirements of the corresponding process, the temporal reasoning module marks the corresponding agent as an activatable state; otherwise, it marks it as an inactivatable state. Subsequently, the temporal reasoning module masks the output action probability distribution of agents in inactivatable states to filter out wandering actions that violate physical temporal constraints.

[0072] The temporal inference module constructs action masking vectors corresponding to the activation states of each agent, and performs a Hadamard product operation between these action masking vectors and the initial action probability distribution output by the policy network to obtain a product vector. To ensure that the sum of the filtered action probabilities is 1, the temporal inference module normalizes each element of the product vector by dividing it by the sum of all elements in the product vector. To prevent division-by-zero overflow caused by all-zero masks, the temporal inference module forcibly introduces a minimal constant representing the prevention of division by zero into the normalization denominator. Meanwhile, the temporal reasoning module pre-sets global, unmasked waiting empty action instructions within the action space, ensuring that the agent has legitimate output and coherent reasoning under extreme conditions.

[0073] Step S502: Execute event-driven asynchronous relay inference. Based on the instructions filtered by the mask, the timing inference module drives the timeline of the simulation environment. To avoid invalid idle time or instruction stacking caused by fixed time steps, the timing inference module introduces an event-driven simulation clock mechanism. When an agent issues a walk action, the environment adds the estimated duration of the corresponding process to the timeline and directly advances the simulation clock to the next discrete event trigger point (such as the acceptance of the preceding process or the material arrival node). Upon reaching this trigger point, the timing inference module re-evaluates the global environment state tensor, and the next agent to be activated, which currently meets the activation conditions and satisfies the topological preconditions, takes over to obtain the local observation tensor and outputs a decision, thereby binding the discrete actions to the actual state transition moments.

[0074] Step S503 involves calculating and allocating a multi-dimensional reinforcement learning reward function. To guide the multi-agent system towards global collaborative diagnosis and prevention optimization, and to avoid single-indicator-induced biased prevention decisions, the temporal reasoning module constructs a reward function that integrates project duration benefits, resource utilization, and delay penalties. At each discrete event trigger point, the temporal reasoning module calculates a comprehensive scalar reward value, using the following formula: ; In the formula, Characterizing the first An agent at time step The total scalar reward value obtained; Characterizing the first An agent at time step The time benefit item is the number of days the total time of the critical path is shortened after the action is executed (if the time is extended, it is a negative value). Characterization at time step The actual consumption of remaining available construction resources during the execution of the action; Characterizes the upper limit of the physical capacity of this type of resource at the current stage of the project; A very small constant representing the prevention of division by zero, with a value set to 10. -6 ; Characterization at time step The delay penalty item related to the current process constraint is configured as the product of the number of natural days the actual completion time exceeds the baseline planned time and the unit delay penalty coefficient, which is used to quantify the schedule loss caused by the conflict of timing constraints. , and These represent the labor option weight, resource weight, and penalty weight, respectively. The values ​​of these weights are all set between 0 and 1, satisfying the following conditions: The normalization conditions are determined dynamically using the analytic hierarchy process (AHP) in conjunction with the project's schedule and cost safety preferences.

[0075] When calculating the above comprehensive scalar reward value, if there is a conflict between the agent's action state and the directed acyclic graph rules of the process, a negative reward is added to the formula result; if the agent's traversal path reaches the prevention and control measure entity, a positive reward is added to the formula result.

[0076] After obtaining the comprehensive scalar reward value, the temporal inference module feeds it into the value network as the underlying truth benchmark for calculating the advantage function and subsequent policy gradient calculations. For the temporal difference error calculation and network parameter iteration rules in reinforcement learning, those skilled in the art can implement them by calling the corresponding interface modules according to the specific business framework; these are well-known technologies in the field and will not be elaborated upon here.

[0077] In this embodiment, to address the policy non-stationarity problem in the multi-agent collaborative diagnosis and prevention reasoning process, the temporal reasoning module constructs an asynchronous advantage action evaluation algorithm architecture that includes a policy network and multiple sets of independent working nodes. It utilizes a multi-threaded mechanism to explore the state space in parallel. The specific implementation steps are as follows: Step S504: Perform time-series difference error and advantage function calculation based on the value network. The time-series inference module uses the value network configured as a multilayer perceptron within each working node to calculate the expected long-term return of the current local observation state. Based on the comprehensive scalar reward value obtained from the preceding steps, the time-series inference module calculates the time-series difference error to measure the deviation between the actual reward and the expected value. The specific formula is as follows: ; In the formula, Characterizing the first An agent at time step The obtained timing difference error; Characterizing the first An agent at time step The total scalar reward value obtained; The discount factor, which is used to balance current and future returns, is set to a value between 0.90 and 0.99. The representation is given by the parameter set. The state value assessment scalar output of the value network; and Respectively characterize the first An agent at time step With time step Extracted local observation tensors. After obtaining the temporal difference error sequence, the temporal inference module accumulates and calculates the advantage function value based on a multi-step reward mechanism, which serves as the core evaluation index guiding strategy optimization.

[0078] Step S505: Calculate the local loss functions for the policy network and value network. Based on the aforementioned data, the temporal inference module calculates the loss gradients for the policy network and value network respectively, guiding the network parameters to iterate towards high-reward actions. To prevent the agent from getting trapped in local optima, the temporal inference module introduces an entropy regularization penalty term into the policy network loss and uses a minimal constant to prevent the logarithmic function's independent variable from approaching 0 and causing negative infinity computational anomalies. The comprehensive loss calculation formula is as follows: ; In the formula, The overall loss function value representing the policy network; The trajectory truncation step size, which represents a single local network training iteration, is set to a positive integer ranging from 10 to 50. Index representing cumulative summation; The representation is given by the parameter set. The corresponding action in the action probability distribution output by the policy network The probability of execution; Characterizing the first An agent at time step The actual control decision actions executed; A very small constant representing the prevention of division by zero, with a value set to 10. -6 ; Characterizing the first An agent at time step The calculated dominance function value; The entropy regularization coefficient is characterized, and its value is preferably between 0.01 and 0.05. The information entropy function characterizes the probability distribution of actions. Simultaneously, the temporal inference module directly calculates the mean square error of the aforementioned temporal difference error sequence within the trajectory truncation step size as the loss function value of the value network. By minimizing this mean square error, the network evaluation approximates the true cumulative expected return.

[0079] Step S506: Perform asynchronous gradient upload and parameter alignment for the policy network. After calculating the local loss, each worker node independently calls the backpropagation algorithm to calculate the local gradient for the network parameters and uploads it to the policy network in parallel through a lock-free asynchronous update mechanism. To avoid parameter divergence caused by concurrent writes, the temporal inference module configures a spinlock synchronization mechanism in the gradient upload channel and, combined with L2 norm pruning, forcibly limits the upper limit of the gradient tensor uploaded in a single instance to a threshold range of 0.5 to 1.0, thus preventing gradient explosion from the underlying numerical logic. After receiving the gradient, the policy network updates the global parameters using the adaptive momentum optimization algorithm. After the update is complete, the current worker node pulls the latest parameters from the policy network to overwrite its local parameters and clears the local trajectory cache to start a new round of simulation. For the chain rule of differentiation in backpropagation and the underlying tensor calculation logic of the optimizer, those skilled in the art can call the standard interfaces of existing deep learning frameworks to implement them, which are well-known technologies in this field and will not be elaborated here.

[0080] In this embodiment, the policy generation module constructs a control generation engine that integrates intent parsing and dynamic topology graph reasoning. By parsing unstructured natural language query commands, it maps the natural language query commands to starting entity nodes in the dynamic topology graph. Then, it calls the policy network to perform forward inference on the dynamic topology graph corresponding to the moment the natural language query command is received, so as to output the control message corresponding to the prevention and control measure entity. The specific implementation steps are as follows: Step S601: Perform intent parsing and entity extraction of the natural language query command.

[0081] The strategy generation module receives natural language query commands input by the user, performs word segmentation, intent parsing, and entity extraction on the natural language query commands, and obtains query entities and query intents related to the diagnosis and prevention of common construction quality defects. To improve the accuracy of recognizing engineering terminology, the strategy generation module can combine a pre-trained language model to semantically encode the natural language query commands and output the semantic feature representations corresponding to the query entities.

[0082] Step S602: Perform the initial entity node mapping.

[0083] The strategy generation module performs semantic matching between the query entity obtained in step S601 and the entity nodes in the dynamic topology graph at the current time to determine the starting entity node corresponding to the query entity. When there are multiple candidate entity nodes, the strategy generation module selects the entity node with the highest score as the starting entity node based on semantic similarity.

[0084] Step S603: Perform inference input construction.

[0085] The policy generation module obtains the dynamic topology map corresponding to the moment the natural language query command is received, and extracts the structural features, starting entity node features, and state features corresponding to the relevant sensor data stream at the current moment of the dynamic topology map. These features are then combined as inputs for the policy network to perform forward inference.

[0086] Step S604: Perform forward inference based on the policy network.

[0087] The strategy generation module calls the strategy network to perform forward inference on the dynamic topology graph. The system wakes up the corresponding master agent according to the construction dimension to which the starting entity node belongs, and transmits the observation state to other related agents through the connected edges of the dynamic topology graph. The system obtains the cooperative walking trajectory of multiple agents in the dynamic topology graph, and determines the prevention and control measure entity at the corresponding endpoint position based on the cooperative walking trajectory.

[0088] Step S605: Perform control parameter extraction and control message encapsulation.

[0089] The strategy generation module extracts the corresponding control parameters based on the entity of the prevention and control measure, and encapsulates the control parameters into a structured control message. The structured control message includes at least a unique instruction identifier, the target receiver network address, the working plane location, the operation effective time, the operation expiration time, and the control parameters.

[0090] Step S606: Perform control message output and terminal coordination.

[0091] The strategy generation module outputs structured control messages to construction equipment or field terminals. During the output process, the strategy generation module can perform timestamp alignment on the target receiving end and perform control message retransmission or backup terminal switching in case of communication failure, so as to ensure the timing consistency and execution reliability of control message distribution.

[0092] Specific application examples This embodiment takes the roadbed compaction operation of a certain highway section as an example to specifically illustrate the actual operation data flow process of the system of the present invention in combination with the aforementioned formula.

[0093] During the initialization phase of the map construction module, the system analyzes the "Technical Specifications for Highway Subgrade Construction" to extract and set the lower limit of the specification value for the compaction speed of the inducing factor. km / h, upper limit of numerical value km / h, and persistently store the above threshold in the static map.

[0094] At the 25-minute mark of construction, the on-site IoT sensors collected a real-time sensing characteristic value of 4.2 km / h for the target entity node (road roller speed) at the current time step. After receiving this data, the mask generation module called the aforementioned publicly available normalized boundary offset calculation formula, substituting the measured value and the static map threshold into a quantitative evaluation (ignoring minimal constants). ): ; Calculations show that the current compaction speed produces a dimensionless boundary offset of 0.80. Subsequently, the mask generation module substitutes this offset result into the aforementioned publicly disclosed formula for a continuously differentiable topological node mask vector. The system is pre-configured with a slope amplification factor. Soft activation threshold The specific calculation process is as follows: ; Calculation result When the value approaches 1, the mask generation module determines that the compaction speed node is strongly activated. After performing a Hadamard product operation on the mask weight and the global adjacency matrix, a dynamic topology graph is generated. This causes the weight of the directed edges in the graph pointing from excessive compaction speed to insufficient compaction to rise sharply, triggering a system anomaly alarm.

[0095] The strategy generation module performs forward inference based on the received high-risk dynamic topology map. According to the starting entity node mapped by the query command, it calls the policy network to generate a multi-agent cooperative walking trajectory and determines the corresponding prevention and control measure entity at the trajectory's endpoint. The system further extracts the control parameters corresponding to the prevention and control measure entity and encapsulates these parameters into a structured control message for distribution. This message forces the road roller control system to reduce its speed to the target baseline value of 2.5 km / h. The field equipment receives the message and responds by reducing its speed at 28 minutes. Subsequently, the sensor data falls back to the compliant range, and when substituted into the above formula again... Reduced to 0, mask weight It then decays to an inactive state, and the system resumes normal monitoring.

[0096] Reference Figure 3 , Figure 3 The horizontal axis represents the construction operation time (min), the left main vertical axis represents the roller compaction speed (km / h), and the right secondary vertical axis represents the normalized boundary offset (dimensionless) calculated and output by the system. Figure 3 The two horizontal dotted lines represent the upper limit (3.0 km / h) and lower limit (1.5 km / h) of the compaction speed set by the standard, respectively. The solid black line represents the actual compaction speed trajectory of the equipment, and the dashed black line represents the trajectory of the normalized boundary offset calculated by the system in real time.

[0097] As shown in the figure, during the first 20 minutes, the device speed remained within the specified range, and the normalized boundary offset was constant at 0. At the 25th minute, the device speed abnormally increased to 4.2 km / h, exceeding the upper limit; the normalized boundary offset calculated and output by the system surged to 0.8, triggering the threshold. After the system issued a deceleration control message at the 28th minute, the device responded and began to decelerate. By the 32nd minute, the speed had returned to the specified range of 2.5 km / h, and the corresponding normalized boundary offset had returned to 0. This figure objectively demonstrates the system's real-time tensor calculation response to the physical state and the closed-loop control effect.

Claims

1. A knowledge graph-based system for diagnosing and preventing common quality defects in highway construction, characterized in that: include: The graph construction module is used to process engineering documents, extract entities of quality phenomena, inducing factors, and prevention measures, match them with the construction specification database to obtain quantitative thresholds, establish directed connection edges, and construct a static graph. The mask generation module is used to generate a mask matrix based on the sensor data stream and the quantitative threshold, and to generate a dynamic topological map by operating the mask matrix with the static map. The state initialization module is used to convert the schedule file into a directed acyclic graph of processes, construct a multi-agent system based on the dynamic topology graph and the directed acyclic graph of processes, and initialize the Markov decision process of the multi-agent system. The strategy pre-training module is used to map the hazard investigation form data to the static graph to generate expert demonstration trajectories and pre-train the strategy network; The temporal reasoning module is used to calculate the environmental reward function and update the policy network based on the walking time sequence of the multi-agent in the dynamic topology graph constrained by the directed acyclic graph of the process. The strategy generation module is used to map query commands to starting entity nodes in the dynamic topology graph, call the policy network to perform forward inference, and output control messages. When the graph construction module extracts the entities of quality phenomena, inducing factors, and prevention and control measures, establishes directed connections, and constructs a static graph, it is specifically used for: For the aforementioned engineering document, named entity recognition is performed to extract the entity representing the quality phenomenon, the entity representing the inducing factor, and the entity representing the prevention and control measures. Based on the extracted entities of the quality phenomenon, the inducing factors, and the prevention and control measures, the relationships between the entities are extracted and structured triples are generated, and the directed connection edges are identified. The static graph is constructed using the structured triples, the directed connecting edges, and the quantitative threshold, and then the static graph is converted into an entity node feature matrix, a global adjacency matrix, and an edge feature tensor and stored persistently. When the state initialization module performs the steps of converting the schedule file into a directed acyclic graph of processes, constructing a multi-agent system based on the dynamic topology graph and the directed acyclic graph of processes, and initializing the Markov decision process of the multi-agent system, it is specifically used for: The schedule file of the building information model is parsed and converted into the directed acyclic graph of the process, which includes independent process nodes and process sequence dependency edges. Extract the process identifier, baseline duration, and resource consumption of the independent process nodes, and construct a graph node feature matrix; The constructed directed acyclic graph of the aforementioned processes is subjected to connectivity verification and loop elimination. Based on the material, environmental, and process dimensions within the construction dimension, the multi-agent system is divided into multiple agent sets. The graph diagnosis process is defined as the Markov decision process. At any decision time step, the structural features of the dynamic topology graph, the features of the entity nodes where the multi-agent is currently staying, the features of the sensor data stream at the corresponding time, and the process constraint features that characterize the execution progress of the independent process nodes, the remaining available amount of construction resources, and the time difference between the actual progress and the baseline plan are concatenated to obtain the global environment state tensor. The state transition relationship is determined based on the directed topology of the dynamic topology graph and the process constraints of the verified directed acyclic graph of the process. When the strategy pre-training module performs the process of mapping the hazard investigation form data to the static atlas to generate expert demonstration trajectories and pre-training the strategy network, it is specifically used for: Extract the quality phenomenon records, inducing factor records, prevention and control measure records, environmental condition records, and rectification result records from the data in the hazard investigation form; The extracted hazard investigation form data is mapped to the corresponding entity nodes in the static graph, and the historical walking path is reconstructed based on the directed connection edges between the corresponding entity nodes to form the expert demonstration trajectory; The resulting expert demonstration trajectory is converted into state samples and action labels; The behavior cloning algorithm is used to input the state samples and action labels into the policy network of the multi-agent, and the parameters of the policy network are pre-trained by calculating and minimizing the cross-entropy loss function. When the temporal reasoning module executes the walking sequence of the multi-agent in the dynamic topology graph based on the directed acyclic graph constraint of the process and calculates the environmental reward function, it is specifically used for: The activation order of the multi-agent system and the traversal sequence are constrained by the process sequence dependency edges in the directed acyclic graph of the process. Based on the walking sequence, the endpoint node features output by the multi-agent corresponding to the previous process are concatenated into the Markov state feature vector of the multi-agent corresponding to the subsequent process. Calculate the environmental reward function including time constraints, the environmental reward function including a schedule benefit term determined by the number of days the critical path is shortened, a resource utilization term determined by the ratio of the actual consumption to the upper limit of the remaining available construction resources during the execution of the action, and a delay penalty term determined by the actual time exceeding the baseline plan. When calculating the environmental reward function, if the action state of the multi-agent conflicts with the directed acyclic graph rule of the process, a negative reward is calculated; if the multi-agent's traversal path reaches the prevention and control measure entity, a positive reward is calculated. When the policy generation module executes the process of mapping the query instruction to the starting entity node in the dynamic topology graph, calling the policy network to perform forward inference, and outputting a control message, it is specifically used for: The system receives the user's query instruction, performs intent parsing and entity extraction on the query instruction, and maps it to the starting entity node in the dynamic topology graph. The dynamic topology map corresponding to the moment the query instruction is received is obtained, and the structural features of the dynamic topology map, the features of the starting entity node, and the state features corresponding to the sensor data stream at the current moment are extracted. The feature combination is then used as the input for the policy network to perform forward inference. The policy network is invoked to perform forward inference on the dynamic topology graph to obtain the cooperative walking trajectory of the multi-agent in the dynamic topology graph. Based on the obtained cooperative walking trajectory, the control parameters of the prevention and control measure entity at the corresponding endpoint position are extracted; The extracted control parameters are converted into control messages and output to the construction equipment or on-site terminal.

2. The knowledge graph-based highway construction quality defect diagnosis and prevention system according to claim 1, characterized in that, When the atlas construction module obtains a quantitative threshold by matching the construction specification database, it is specifically used for: Extract fuzzy semantic features from the fuzzy semantic modifiers in the project document; The construction specification database is constructed based on the construction specification text, and the specification clauses in the construction specification text are transformed into specification entries containing physical quantity identification fields, fuzzy modifier fields, numerical lower limit fields, numerical upper limit fields, and standard unit of measurement fields. Calculate the semantic similarity between the fuzzy semantic features and the canonical entries; Based on the semantic similarity, the lower limit field, the upper limit field, and the standard unit of measurement field are extracted from the matched normative entries to obtain the quantitative threshold.

3. The knowledge graph-based highway construction quality defect diagnosis and prevention system according to claim 1, characterized in that, When the mask generation module performs the process of generating a mask matrix based on the sensor data stream and the quantitative threshold, it is specifically used for: Receive the sensor data stream containing sensor output parameters, and perform time synchronization, spatial alignment, outlier removal and noise smoothing on the sensor data stream to obtain a continuous temporal feature tensor; Traverse the directed connection edges in the static graph that are associated with the quantitative threshold; Obtain the continuous temporal feature tensor, compare the continuous temporal feature tensor with the quantitative threshold associated with the directed connection edge, and generate the mask matrix corresponding to the current time step based on the comparison result.

4. The knowledge graph-based highway construction quality defect diagnosis and prevention system according to claim 1, characterized in that, When the mask generation module performs the operation of generating a dynamic topological map by combining the mask matrix with the static map, it is specifically used for: The normalized boundary offset is calculated based on the difference between the continuous temporal feature tensor and the quantitative threshold corresponding to the target entity node in the static graph. The calculated normalized boundary offset is mapped to dynamic node mask weights through a smooth nonlinear activation function. The connectivity weight of the directed connection edge is determined based on the dynamic node mask weights of adjacent entity nodes, and a dynamic edge mask matrix is ​​generated. The mask matrix includes at least the dynamic node mask weights and the dynamic edge mask matrix. The generated dynamic edge mask matrix is ​​multiplied by the global adjacency matrix of the static graph to obtain the dynamic global adjacency matrix, and the dynamic topology graph at the current time step is generated by combining the edge feature tensor.

5. The knowledge graph-based highway construction quality defect diagnosis and prevention system according to claim 1, characterized in that, The time-series inference module is also used for: An asynchronous advantage action evaluation algorithm is used for local gradient calculation and policy network update. Construct a value network, and calculate the time-series difference error and advantage function value based on the value network and the environmental reward function; Based on the temporal difference error and the advantage function value, the gradient of the objective function of the policy network and the gradient of the loss function of the value network are calculated respectively. The model converges by calculating the gradients separately in parallel working threads and updating them asynchronously to the policy network.

Images