Rock and soil construction quality monitoring and diagnosing method

Abstract

The invention provides a rock and soil construction quality monitoring and diagnosing method, which comprises the following steps of: modeling historical engineering events and expert rules, constructing an event causal atlas prototype containing weights, and realizing event sequence feature extraction and real-time causal atlas dynamic updating in combination with on-site multi-modal sensing data and construction logs; neural symbol reasoning and tensor completion technologies are adopted to predict a novel causal relationship, and a causal atlas structure is perfected through space, time and logic consistency verification; a multi-layer risk early warning mechanism is set, a causal map local risk assessment and event chain propagation are combined, spatial positioning early warning signals are generated in a grading manner, closed-loop backtracking optimization is supported, causal reasoning accuracy and early warning efficiency are improved, construction process risk identification and dynamic early warning can be realized, and the engineering safety management level is improved.

Description

Technical Field

[0001] This invention relates to the field of geotechnical engineering construction monitoring and intelligent early warning technology, and in particular to a method for monitoring and diagnosing the quality of geotechnical construction. Background Technology

[0002] In the field of geotechnical engineering construction quality monitoring and intelligent early warning, most current mainstream early warning technologies are based on static rules, experience templates, or limited statistical modeling. These technologies periodically analyze multi-dimensional information such as displacement, stress, and pore water pressure collected by sensors, and then set thresholds and event triggering logic based on typical process events (such as blasting, excavation, grouting, and support). These technologies typically employ pre-defined causal links or fault diagnosis processes to screen front-end monitoring information, thereby detecting anomalies and outputting early warning signals. In recent years, some research has attempted to improve risk identification capabilities using data-driven machine learning, Bayesian networks, or finite state machines, with development trends including expansion towards intelligence, automation, and real-time response. However, in actual geotechnical construction scenarios, current typical technologies have the following shortcomings: Most existing methods are based on fixed causal models or rule bases and lack the ability to reason about the dynamic causal evolution of construction event time chains. When non-preset event combinations or abnormal response patterns occur, misjudgment or omission is likely to occur, and they cannot adapt to engineering variations and complex interactions on site. Existing technologies primarily process data streams through periodic statistics or simple threshold triggers, failing to achieve real-time, automated causal learning and graph self-correction during the construction process. They also lack the ability to effectively generate, verify, and dynamically incorporate novel coupling relationships between events, leading to lags in knowledge accumulation and model evolution, and reducing the intelligence level and robustness of the technology in the long run. Due to the lack of information integration oriented towards multi-source data and spatial-temporal multi-scale information, the commonly used risk assessment and response processes are mostly serial and hierarchical, with weak linkage. Early warning signals are only directed at a single risk point and cannot upgrade risks, issue multi-level warnings, or push responsibilities based on causal chain propagation or changes in the overall situation, making it difficult to promote information closure and efficient handling. Traditional causal reasoning models rely heavily on expert experience or manual intervention for model structure and threshold adjustment, lacking periodic backtesting and adaptive optimization mechanisms. Once the environment or construction process changes, the original early warning model parameter configuration is prone to failure, resulting in high maintenance costs and poor adaptability. Summary of the Invention

[0003] In order to solve the above-mentioned technical problems, the present invention provides a method for monitoring and diagnosing the quality of geotechnical construction.

[0004] The technical solution of this invention is implemented as follows: a method for monitoring and diagnosing the quality of geotechnical construction, comprising: S1: Construct a prototype of a causal graph of basic events based on historical engineering data and expert knowledge base. The nodes of the graph prototype represent typical construction events, and the edges represent known causal relationships and their initial strength weights, forming an initial causal reasoning framework.

[0005] S2: Collect real-time sensor data, construction log timestamps, and operation type information during the construction process. The sensor data includes multi-dimensional parameters such as displacement, stress, and pore water pressure, forming an event sequence raw dataset.

[0006] S3: Perform a temporal similarity matching algorithm on the original dataset of the event sequence to identify the degree of deviation between the current event sequence and the existing paths in the basic causal graph prototype, and dynamically adjust the causal strength of the relevant edges based on the Bayesian online learning framework to generate a dynamic updated version of the causal graph.

[0007] S4: A lightweight neural symbolic reasoning model that integrates logical rule constraints and low-rank tensor completion techniques is used to generate hypotheses for novel event associations not recorded in the dynamic causal graph, and outputs potential causal links and their confidence scores. The logical rule constraints are based on the logic of geotechnical engineering procedures.

[0008] S5: Execute the context consistency verification module to perform spatial proximity verification, time reachability analysis, and process logic compliance judgment on newly added causal links, generate a causal relationship correction set that passes the verification, and write it into the dynamic causal graph master database.

[0009] S6: Construct a hierarchical early warning triggering mechanism: The bottom-level risk assessment unit calculates the direct risk probability of the current working condition based on local fragments of the dynamic causal graph; the middle-level event chain propagation simulator tracks high-confidence causal paths and predicts subsequent event sets in a forward traversal manner; and the top-level response decision engine generates hierarchical early warning signals based on the number of key nodes and binds them to BIM spatial coordinates.

[0010] S7: Activate the periodic backtracking verification module to compare the early warning results with the actual event development trajectory. Based on the deviation analysis results, optimize the structural parameters of the dynamic causal graph and the threshold configuration of the hierarchical early warning triggering mechanism to form a closed-loop evolution mechanism.

[0011] The geotechnical construction quality monitoring and diagnosis method provided by this invention has the following beneficial effects: (1) This scheme realizes the continuous modeling and evolution tracking of dynamic causal relationships between construction events by constructing an initial causal graph prototype and introducing an incremental update mechanism driven by online data flow. It uses a temporal similarity matching algorithm to identify the deviation between the actual event sequence and the historical path, and combines a Bayesian online learning framework to correct the causal edge weights in real time. This significantly improves the risk identification sensitivity and inference accuracy when facing atypical construction processes or sudden operation combinations, effectively overcoming the defects of missed and false alarms caused by the rigidity of traditional methods. At the same time, by generating potential causal hypotheses through a lightweight neural symbolic reasoning model that integrates logical rule constraints and low-rank tensor completion, the coverage of the graph is expanded under the premise of ensuring semantic rationality, so that the method has the ability to discover new risk transmission paths, thereby enhancing the adaptability and foresight of the overall early warning system. (2) The present invention designs a hierarchical early warning trigger architecture, which covers the complete chain from local anomaly assessment to global event chain simulation and hierarchical linkage response. The bottom single-point risk assessment unit relies on local map fragments to quickly calculate the direct risk probability under the current working condition, ensuring the timeliness of the response; the middle event chain propagation simulator tracks high confidence causal paths in a forward traversal manner, dynamically predicts the chain event set that may be activated and its propagation trend, breaking through the information island limitation of the traditional point-to-point alarm mode; the top multi-level response decision engine comprehensively considers the key node density, impact space range and closed-loop control possibility of the predicted path, automatically generates differentiated early warning levels, and accurately binds them with the geographical coordinates in the BIM model, realizing the spatial push of risk information and intelligent matching of responsible entities, greatly improving the pertinence and organizational efficiency of emergency response. This hierarchical structure not only enhances the method's ability to analyze complex causal networks, but also improves the situational awareness level and decision support quality of on-site management personnel through semantic enhancement and spatial mapping of early warning signals. (3) This invention constructs a closed-loop evolutionary mechanism of "perception-reasoning-early warning-verification". Through a periodic backtracking module, it continuously compares the prediction results with the actual development trajectory, and optimizes the topology and parameter configuration of the causal graph, enabling the method to have self-calibration and continuous learning capabilities. Compared with traditional static techniques that model once and remain unchanged for a long time, this invention can evaluate the effectiveness and prediction accuracy of the causal path after each early warning execution, thereby guiding the reconstruction and strengthening of weak links in the graph, significantly improving the long-term stability and generalization performance of the model. In addition, the context consistency test mechanism effectively filters out unreasonable or infeasible causal assumptions, ensuring the physical interpretability and engineering credibility of the graph evolution. The entire system achieves fully automated operation from data input to intelligent early warning and then to self-optimization without frequent human intervention, and has good scalability and engineering deployment value. Attached Figure Description

[0012] Figure 1This is a flowchart of a geotechnical construction quality monitoring and diagnosis method according to the present invention; Figure 2 This is a sub-flowchart of a geotechnical construction quality monitoring and diagnosis method according to the present invention; Figure 3 This is another sub-flowchart of a geotechnical construction quality monitoring and diagnosis method according to the present invention. Detailed Implementation

[0013] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0014] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0015] like Figure 1 As shown, the present invention provides a method for monitoring and diagnosing the quality of geotechnical construction, specifically including: S1: Construct a basic event causal graph prototype based on historical engineering data and expert knowledge base. The nodes of the graph prototype represent typical construction events, and the edges represent known causal relationships and their initial strength weights, forming an initial causal reasoning framework. S2: Collect real-time sensor data, construction log timestamps, and operation type information during the construction process. The sensor data includes multi-dimensional parameters such as displacement, stress, and pore water pressure to form an event sequence raw dataset. S3: Execute a temporal similarity matching algorithm on the original dataset of the event sequence to identify the degree of deviation between the current event sequence and the existing paths in the basic causal graph prototype, and dynamically adjust the causal strength of the relevant edges based on the Bayesian online learning framework to generate a dynamic updated version of the causal graph. S4: A lightweight neural symbolic reasoning model that integrates logical rule constraints and low-rank tensor completion technology is used to generate hypotheses for novel event associations not recorded in the dynamic causal graph, and output potential causal links and their confidence scores. The logical rule constraints are based on the logic of geotechnical engineering procedures. S5: Execute the context consistency check module to perform spatial proximity verification, time reachability analysis and process logic compliance judgment on newly added causal links, generate a causal relationship correction set that passes the test and write it into the dynamic causal graph main database; S6: Construct a hierarchical early warning triggering mechanism: The bottom-level risk assessment unit calculates the direct risk probability of the current working condition based on local fragments of the dynamic causal graph; the middle-level event chain propagation simulator tracks high-confidence causal paths and predicts subsequent event sets in a forward traversal manner; and the top-level response decision engine generates hierarchical early warning signals based on the number of key nodes and binds them to BIM spatial coordinates. S7: Activate the periodic backtracking verification module to compare the early warning results with the actual event development trajectory. Based on the deviation analysis results, optimize the structural parameters of the dynamic causal graph and the threshold configuration of the hierarchical early warning triggering mechanism to form a closed-loop evolution mechanism.

[0016] Step S1: Construct a prototype causal graph of basic events based on historical engineering data and an expert knowledge base. The nodes of the prototype graph represent typical construction events, and the edges represent known causal relationships and their initial strength weights, forming an initial causal reasoning framework. Specifically, this includes: S1.1: Structured extraction of geotechnical construction event records from historical engineering databases. Using construction event type, occurrence time, spatial location, operation parameters, and associated response characteristics as input conditions, a set of event nodes is generated based on graph database modeling technology to form the initial node library of the graph. A structured data extraction algorithm (parameters: event type, occurrence time, spatial location, operation parameters, and associated response characteristics) is used to extract geotechnical construction event records from historical engineering databases, converting unstructured or semi-structured records into a unified field format. Furthermore, by using an event type encoding method (parameter: predefined construction event encoding dictionary), the event type field is standardized and an event type dataset with unified encoding rules is obtained. Furthermore, a spatial coordinate analysis algorithm (parameters: BIM model spatial reference system, three-dimensional coordinate transformation matrix) is adopted to realize the three-dimensional reconstruction of the spatial location of construction events and generate spatial location vectors that meet the spatial index requirements of the graph database; Furthermore, by using the operation parameter normalization processing method (parameters: minimum-maximum normalization range [0,1], outlier removal rules), the dimensionality of operation parameters in construction records from different sources is unified, and a standardized operation parameter vector set is obtained; Furthermore, an associated response feature extraction algorithm (parameters: displacement response curve, stress time series, pore water pressure change rate) is adopted to extract the response information associated with the event from the physical monitoring data into numerical features and generate a multidimensional response feature set. By using graph database modeling (parameters: node type = construction event, attribute fields = {type, time, location, operation parameters, response characteristics}), the results of the previous step are transformed into a set of event nodes, thus forming the initial node library of the graph; For example, in a historical engineering database containing tunnel excavation and slope protection data, the original record field "Excavation Start" is parsed as event type code E003, the occurrence time field is parsed as 2021-07-16 08:30:00, the spatial location is parsed as {X=125.6, Y=87.3, Z=-12.5} meters after three-dimensional coordinate analysis, the operation parameter "Cutterhead speed = 3.2 rpm" is normalized to 0.64, and the displacement change rate is extracted as a correlation response feature. mm / hour, stress increment is MPa, pore water pressure gradient is MPa / hour. After the graph database nodes are constructed, the event is stored as a node object with attributes such as type, time, location, operation parameters, and response characteristics, realizing the accurate recording in the initial node library and laying the data structure foundation for subsequent causal reasoning; S1.2: Based on the causal rules extracted from the expert knowledge base, perform causal relationship reasoning processing on the node pairs in the event node set, and use the causal graph modeling algorithm to generate an initial causal edge set. Each edge in the causal edge set represents the causal relationship between two construction events, so as to construct the initial edge set of the causal graph. Based on the set of causal rules for geotechnical construction stored in the expert knowledge base, the set of event nodes generated in the previous step is used as the input object for reasoning. A rule parsing algorithm (parameters: rule type classification label, logical constraint template index) is used to realize the matching and retrieval function of event node pairs, so as to identify the event combinations that meet the causal conditions defined by the expert knowledge. Furthermore, by using a pattern matching algorithm (parameters: event type encoding vector, inference matching threshold), the semantic structure of event node pairs is aligned with the causal rule set, and a list of matching node pairs that meet the conditions is obtained. Furthermore, a causal graph modeling algorithm (parameters: node identifier ID set, edge attribute initialization template) is adopted to generate an initial causal edge set from the matching list, and add the start node and end node associated with each edge to the edge attribute mapping table; Furthermore, a causal direction determination algorithm (parameters: process time series scalar, spatial dependency matrix) is used to perform directional assignment of causal edges and generate a direction identifier field to ensure the unidirectionality and reasonability of causal relationships. By optimizing the edge set, the results of the previous step are transformed into a structured initial causal edge dataset, achieving a technical effect that can be used for subsequent intensity quantization and graph structure integration. For example, in a foundation pit support construction scenario, the node set includes event nodes such as "excavation operation," "shotcrete support," and "reinforcing mesh installation." The causal rule set in the expert knowledge base defines that "excavation operation" necessarily precedes "shotcrete support," and there is a direct causal relationship between the two. The rule parsing algorithm retrieves node pairs that satisfy this logic from the node set. A pattern matching algorithm encodes the node types into event type vectors of length 8, setting the matching threshold to 0.85, resulting in a list of successfully matched node pairs. The causal graph modeling algorithm transforms the matching list into an edge data structure, with the starting ID corresponding to "excavation operation" and the ending ID corresponding to "shotcrete support." The initial attribute template includes edge types of "direct causality" and time differences in hours. The causal direction determination algorithm uses the process time series matrix to determine that "excavation operation" occurs before "shotcrete support," thus assigning the edge direction label as positive. The final initial causal edge dataset includes three edges, each with a unique ID, starting and ending node IDs, a causal direction identifier, and initial attribute template parameters. This dataset is then ready for the strength quantification step. S1.3: Perform causal strength quantization processing on each causal edge in the initial causal edge set. Based on the co-occurrence frequency and temporal consistency index in the historical event sequence data, use the causal strength evaluation algorithm to calculate the initial strength weight of each causal edge to obtain a weighted causal graph edge set. S1.4: Perform graph structure integration processing on the weighted causal graph edge set and event node set, and construct a basic causal graph prototype based on the graph database storage structure. The basic causal graph prototype includes event nodes, causal edges and initial strength weights on the edges to form a graph structure model that can perform causal reasoning. Using a weighted causal graph edge set and event node set as input conditions, a graph structure integration algorithm (parameters: node attribute pattern, edge weight matrix, graph pattern definition file) is used to realize the topological merging of the edge set and node set and generate a unified graph structure cache. Furthermore, by using a pattern mapping method based on graph database storage structure (parameters: node label definition, edge type template, weight field constraint), the graph structure after topology merging is mapped into a persistent database storage model, and a graph pattern instance containing event nodes, causal edges, and initial strength weights is obtained. Furthermore, by utilizing the node attribute normalization algorithm (parameters: attribute range standardization coefficient, spatial coordinate transformation matrix), a unified encoding of numerical attributes and category attributes in the event node set is achieved, and a standardized node description set compatible with the input of the causal reasoning engine is generated. Furthermore, an edge weight index construction method (parameters: index key generation rules, weight retrieval optimization strategy) is adopted to achieve efficient index configuration of causal edge weight fields and generate a weight index table that can be used for fast causal strength query; By optimizing the graph structure, the integrated event nodes and causal edges are redistributed according to the master-slave node storage strategy of the graph database, forming a basic causal graph prototype with causal reasoning capabilities, thus achieving a significant improvement in structural consistency and reasoning executability. For example, in a basic tunnel engineering scenario, the input event node set contains 500 nodes, each with attributes such as event type (blasting, excavation, support), occurrence time (accurate to the second), and 3D spatial coordinates. The edge set contains 1200 causal edges, each with an initial strength weight (range 0.1 to 0.9). A graph structure integration algorithm is used, setting the node label in the graph pattern definition file to "construction_event", the edge type template to "causal_relation", and the weight field constraint to a floating-point range [0,1]. Through a pattern mapping method, the integrated graph structure is mapped to the Neo4j graph database. The node attribute normalization process uses a spatial coordinate scaling factor of 0.001 for 3D coordinate standardization, and the time attribute is uniformly converted to a Unix timestamp. In the weight index construction method, the index key generation rule is the "sourceID_targetID" composite key, and the weight retrieval optimization strategy is a B+ tree index to support range queries. The final database forms a basic causal graph prototype with event nodes, causal edges, and initial strength weights. After the inference engine calls the consistency check, it can be directly used for subsequent dynamic evolution and multi-level early warning triggering. The verification results show that the graph prototype completes a full graph traversal inference within 500 milliseconds, which significantly improves the response speed of the early warning system to complex construction event sequences. S1.5: Perform context consistency verification on the basic causal graph prototype, verify the compliance of causal paths in the graph based on the logic rules of geotechnical engineering procedures and space-time constraints, and use the logic reasoning engine to identify and correct causal edges that violate engineering logic in order to generate an optimized initial causal graph as the basic framework for subsequent dynamic evolution and reasoning. The weighted basic causal graph prototype is input into the context consistency verification processing module, which calls the preset set of geotechnical engineering process logic rules and space-time constraint conditions as verification conditions to achieve compliance screening of each causal path. A process logic matching algorithm (parameters: construction event type coding table, standard process sequence matrix) is used to compare the causal path with the engineering process specifications edge by edge and output the path logic consistency matrix. Furthermore, a spatial constraint verification algorithm is used (parameters: event node spatial coordinate set, spatial proximity threshold). This enables the determination of physical reachability between nodes and obtains a set of spatial compliance tags; Furthermore, a time constraint analysis algorithm (parameters: event timestamp list, minimum job cycle) is used. Maximum allowable delay This enables the determination of the time reachability of causal paths and generates a set of time-consistent labels; Furthermore, through a logical reasoning engine (reasoning mode: forward chain reasoning, rule set input: process logic consistency matrix, spatial compliance label set, temporal consistency label set), the causal edge that violates any constraint condition is located and corrected. The logical reasoning engine is a tool for verifying the compliance of causal paths in the causal graph. Based on the process logic rules and spatial-temporal constraints of geotechnical engineering, it identifies and corrects causal edges that violate engineering logic through the forward chain reasoning mode, ensuring the structural consistency and engineering rationality of the causal graph, and outputs the corrected causal path set. By using the causal edge weight re-estimation method (parameters: corrected path set, historical co-occurrence frequency data), the corrected paths are transformed into weighted optimized causal edge sets, thereby achieving context consistency optimization of the initial causal graph. For example, a prototype of a basic causal graph containing 100 event nodes and 300 causal edges is input into the consistency verification module. The process logic rule set includes the standard sequence of blasting → excavation → support → grouting, and a spatial proximity threshold. Set to 15 meters. Set to 30 minutes. The timeframe was set to 48 hours. Through process logic matching, 25 causal edges violating the order were identified; spatial constraint verification removed 12 causal edges with a distance exceeding 30 meters between nodes; and time constraint analysis removed 9 causal edges with a time interval of 15 minutes or exceeding 72 hours. The logic reasoning engine synthesized the three types of tags to form a set of 254 corrected paths. It then invoked a causal edge weight re-estimation algorithm, calculating weights based on historical co-occurrence frequency and temporal consistency. For example, the weight values ​​of some edges were adjusted from... Adjust to Finally, it is written into the optimized initial causal graph, which has compliance under multiple constraints of space, time and logic, and meets the accuracy requirements of subsequent dynamic evolution and reasoning.

[0017] Step S2: Collect real-time sensor data, construction log timestamps, and operation type information during construction. The sensor data includes multi-dimensional parameters such as displacement, stress, and pore water pressure, forming a raw dataset of event sequences. Specifically, this includes: S2.1: Based on a multimodal sensor network deployed at the geotechnical construction site, the raw signals output by displacement sensors, stress gauges and pore water pressure gauges are acquired to construct a physical response data stream during the construction process; Based on the multimodal sensor network already deployed at the geotechnical construction site, displacement sensors, stress gauges, and pore water pressure gauges are used as data acquisition units to obtain the initial signal streams of various physical responses during the construction process; A displacement sensor array is used to collect structural deformation information (parameters: sampling frequency 50Hz, resolution 0.01mm) to achieve real-time capture of instantaneous position offset at different monitoring points and form a raw displacement signal set; Furthermore, the stress state changes inside the component and on the slope surface are collected through the stress gauge channel (parameters: sampling frequency 100Hz, range 0~50MPa), and the original stress signal set is obtained. Furthermore, a pore water pressure gauge (parameters: sampling frequency 20Hz, range 0~1MPa) was used to collect changes in groundwater dynamic pressure and generate a raw pore water pressure signal set. A distributed data acquisition gateway is used to identify and bind displacement signals, stress signals, and pore water pressure signals according to sensor IDs, and package them in a unified communication protocol to construct a multi-dimensional physical response data stream for the construction process. The gateway's timestamp annotation module adds sampling time information to the raw signal streams collected by each sensor, forming a physical response raw data packet with time reference, thus achieving seamless integration with subsequent filtering and time alignment processing. By using multimodal sensor network signal acquisition and gateway protocol encapsulation processing, the results of the previous step are transformed into a comprehensive physical response data stream identified by monitoring point, signal type and acquisition time, thus providing complete and synchronous raw data input for subsequent filtering, noise reduction and feature extraction; For example, at a tunnel excavation site, five displacement sensors are deployed along the arch crown, arch waist, and sidewalls, with a sampling frequency of 50Hz, to record structural displacement changes per unit time; four stress gauges are embedded in the support structure of the arch crown and arch waist, with a sampling frequency of 100Hz, to reflect local stress changes in real time; and three pore water pressure gauges are deployed in the soil layer near the construction face, with a sampling frequency of 20Hz, to record groundwater pressure fluctuations. During data acquisition, the gateway assigns a unique ID to each sensor and adds a timestamp to the output signal, for example... Used to indicate the moment the signal is generated. The output of the displacement sensor at a certain moment is... The corresponding time output of the stress gauge is The output of the pore water pressure gauge is The raw data of this type of multidimensional physical response is encapsulated by the gateway to form a data stream in a unified format, which is then directly fed into the subsequent noise reduction and time alignment modules. This ensures that the signals from different sensors can be effectively matched in both time and space, enabling multi-parameter synchronous analysis based on the same construction state and significantly improving the real-time accuracy of causal map updates. S2.2: Perform filtering and noise reduction processing on the original signal, and use a Butterworth low-pass filter to remove high-frequency interference components to obtain more stable displacement, stress and pore water pressure measurement values; S2.3: Based on the time synchronization protocol, the filtered sensor data is timestamped and aligned. The NTP network time protocol is used to unify the multi-source data to a global time reference in order to generate a multi-dimensional data sequence with time consistency. S2.4: Collect operation type information from the construction log, including the start and end timestamps of construction events such as blasting, grouting, excavation, and support, to form structured construction event metadata; The data collected consisted of original construction log files from geotechnical construction sites. The logs contained descriptive text of construction events, event category identifiers, and start and end timestamps. An event text parsing algorithm (parameter: event keyword matching rules based on regular expressions) is used to automatically identify operation type information in construction logs and generate a set of event type tags. Furthermore, by using a timestamp parsing method (parameter: time format parsing rule is yyyy-MM-dd HH:mm:ss), the start and end times corresponding to the identified event tags are accurately extracted, and a set of event timestamp pairs is obtained; Furthermore, through a data structuring processing algorithm (parameter: JSON data mapping template), the event type label, start timestamp, and end timestamp are combined and stored as key-value pairs, and a structured construction event record unit is generated. Furthermore, by using the event classification and identification coding method (parameter: event category coding table), the event categories of the structured construction event recording units are standardized and coded, and an event record set with category codes is output. By using structured processing methods, the results of the previous step are transformed into structured construction event metadata, thereby achieving a unified expression of construction event information and standardized input for subsequent spatiotemporal correlation matching. For example, at a basic tunnel construction site, 120 original construction log files were collected. Each record contains an event description and a timestamp. The parsing algorithm sets the regular expression matching mode to match four keyword categories: "blasting," "grouting," "excavation," and "support," mapping the log text to the corresponding event type tag. For example, "blasting" is mapped to code B001. The timestamp parsing method converts "2024-05-10 09:15:32" in the original text into a numerical UNIX time format and combines them into <start timestamp, end timestamp> pairs. The structured processing algorithm organizes the records according to a preset JSON template into the format {"event_code":"B001","start_time":1715313332,"end_time":1715316932} and writes it into the event record set. The event classification coding table also includes spatial location fields (such as BIM coordinate system X=120.5, Y=45.0, Z=-10.3). After coding, each record has a unified category code and spatial label, forming a complete structured construction event metadata dataset. This can significantly improve matching efficiency and accuracy when performing spatiotemporal matching with sensor output data in the future. S2.5: Perform spatiotemporal correlation matching between structured construction event metadata and multidimensional sensor data, and use a time window sliding algorithm to align the event occurrence time with physical response data to generate an event-response coupled dataset; S2.6: Perform feature extraction on the event-response coupled dataset, calculate displacement change rate, stress increment and pore water pressure gradient based on sliding window statistical method to generate event-driven feature vector sequence; S2.7: Bind the feature vector sequence with the construction event label to generate a structured event sequence raw dataset and store it in the real-time data buffer to provide input for subsequent time series matching and causal graph updates.

[0018] like Figure 2 As shown, step S3 involves performing a temporal similarity matching algorithm on the original dataset of the event sequence to identify the degree of deviation between the current event sequence and existing paths in the prototype of the basic causal graph. Based on a Bayesian online learning framework, the causal strength of related edges is dynamically adjusted to generate a dynamically updated version of the causal graph. Specifically, this includes: S3.1: Perform sliding time window segmentation on the original dataset of construction event sequences to extract continuous event sequence segments and obtain a set of event subsequences with time continuity constraints, which serve as input data for time series similarity matching; S3.2: Based on the preset event type encoding rules, the event subsequence is converted into a low-dimensional temporal embedding vector to represent the semantic and temporal sequence features of the event sequence, forming an event sequence feature vector set as the input representation for subsequent similarity matching; S3.3: The Dynamic Time Warping (DTW) algorithm is used to perform pairwise matching between the feature vector set of the event sequence and the historical path vector in the prototype of the basic causal graph to calculate their temporal similarity and output the similarity matrix as the matching metric result between the event sequence and the existing causal path. For the event sequence feature vector set and the historical path vector in the basic causal graph prototype, the Dynamic Time Warping (DTW) algorithm (distance metric parameter: Euclidean distance; window constraint parameter: radius is 0.1 of the total sequence length) is used to achieve nonlinear time alignment matching across sequences. Furthermore, the scale of each feature dimension is unified by normalization (method: extreme value normalization, range: [0,1]) to ensure that the matching distance under different feature scales is comparable, and the preprocessed event sequence feature matrix and historical path feature matrix are obtained. Furthermore, the optimal alignment path between each event sequence segment and the historical path segment is calculated using the DTW cumulative distance matrix, and the cell distance is calculated using the following formula:

[0019] in, The unit distance at position ( , The value at ) The feature vector of the event sequence at the th Value at position, The historical path feature vector at the th The value at the position; Furthermore, the local cumulative distance is calculated using a dynamic programming recursive formula:

[0020] in, For the cumulative distance matrix at position ( , The value at ) The function is used to select the cumulative value with the smallest distance among three paths; Furthermore, a temporal similarity metric is generated by dividing the normalized cumulative distance by the matching path length:

[0021] in, The optimal matching path length; By using matrix filling, the similarity index of all event sequences and historical paths is written into the similarity matrix, realizing the quantitative output of the matching between event sequences and existing causal paths; The result of the DTW algorithm or the above formula is used to convert the feature vector matching output of the previous step into a quantifiable similarity matrix, thereby achieving accurate measurement of dynamic causal graph path matching. For example, in a geotechnical tunnel construction monitoring project, the event sequence contains six coded typological features within a 60-second time window; the historical path vector is 55 seconds long, with the same feature dimension. Setting the DTW window radius to 6, and using Euclidean distance to calculate cell distances, the maximum cell distance matrix is ​​2.35, and the minimum is 0.15. Using a recursive formula to calculate the cumulative distance matrix, the optimal path length is 58, and the normalized cumulative distance is 1.12, resulting in a temporal similarity Sim=1. (1.12 / 58)≈0.9807. This similarity score is filled into the corresponding position in the similarity matrix. The high similarity entries in the matrix are used for the generation of the deviation feature vector in the downstream step S3.4, which ultimately achieves high-precision matching in this scenario and ensures the effectiveness and stability of the causal graph update stage. S3.4: Based on the matching results in the similarity matrix that are higher than a set threshold, identify the degree of deviation between the current event sequence and the existing paths in the basic causal graph prototype, so as to generate an event path deviation feature vector, which serves as the input observation data for the Bayesian online learning framework; Based on the similarity matrix data output from step S3.3, a threshold filtering method (parameter: similarity threshold τ, which is set comprehensively based on the stability of historical models and the detection accuracy of actual scenarios) is adopted to select elements in the matrix that meet the high matching degree conditions, and retain the set of event path pairs with matching values ​​greater than τ as the deviation detection benchmark. Furthermore, through the event path index mapping method (parameters: event encoding mapping table, basic causal graph path index structure), the current event sequence index in the high matching degree path pair is matched one by one with the corresponding historical path index in the graph, and a path difference set is generated, which contains event location difference information and time interval deviation information. Furthermore, a path deviation quantization algorithm (parameters: time interval Δt, event order difference δe, node coverage ρn) is employed to numerically represent the path difference set, where the deviation characteristics... The calculation formula is:

[0022] in, , , These are the weighting coefficients for the three deviation dimensions: time, sequence, and node coverage. Furthermore, a normalization processing algorithm (parameters: maximum deviation value Pmax, minimum deviation value Pmin) is adopted to map the deviation feature values ​​to the [0,1] interval, thereby obtaining a normalized event path deviation feature vector to eliminate the influence of different path lengths and event complexities on the deviation values. By using the deviation vector packaging processing method (parameters: vector dimension d, sequence number identifier), the normalized event path deviation feature vector and its corresponding event sequence identifier are structurally integrated to generate input observation data for the Bayesian online learning framework, realizing the transformation from time-series matching results to feature data that can be used for dynamic updating of causal strength; For example, during tunnel excavation, the similarity matrix output value between the current event sequence and path P45 in the basic causal graph is detected to be 0.87. A similarity threshold τ = 0.85 is set, and this path pair is retained after filtering. Using an index mapping method, the current sequence [excavation-support-grouting] is matched with the historical path [excavation-support-secondary support] based on positional differences, obtaining a time interval Δt = 180s, an event sequence difference δe = 0.25, and a node coverage rate ρn = 0.8. Substituting these values ​​into the calculation formula:

[0023] The calculated bias value is P = 72.45. Setting Pmax = 200 and Pmin = 0, the normalized result is (72.45 - 0) / (200 - 0) = 0.36225, forming a 3D normalized bias feature vector [0.36225, 0.25, 0.8]. This vector is structurally bound to the sequence number ID = E20230206, serving as the input observation data for the recursive update of causal edge strength in the Bayesian online learning framework. In this scenario, the value of the bias feature vector significantly reflects the temporal deviation between the current sequence and the historical path, ensuring that subsequent intensity adjustment processes are performed under conditions containing high-confidence bias data, thus improving the accuracy and robustness of causal inference. S3.5: Based on the Bayesian online learning framework, the intensity weights of relevant causal edges are recursively updated using the event path deviation feature vector to integrate new observation information and retain historical knowledge, outputting an updated version of the dynamic causal graph and realizing real-time adaptive adjustment of causal intensity. Based on the event path deviation feature vector and the existing edge attributes of the basic causal graph prototype, a Bayesian online learning algorithm is adopted (parameters: prior distribution type is Beta distribution, initial hyperparameters...). and (Obtained from historical event co-occurrence frequency statistics), enabling recursive updating of causal edge strength weights; Furthermore, the weights of each relevant causal edge are dynamically adjusted using a Bayesian recursive update formula, which is expressed as:

[0024] in and These are the supporting and counter-evidence quantities of the current observations, respectively, which are mapped from the event path deviation feature vector. Furthermore, by combining the eigenvectors and the confidence function, and using the Gaussian kernel density estimation method (with the bandwidth parameter adaptively calculated according to the Silverman rule), the likelihood value of the new observation data is obtained, and adjustments are made accordingly. and This ensures that the update process remains smooth and stable within the probability space; Furthermore, a recursive update strategy is used to retain historical weight information, and a discount factor is introduced in each update. (Value range 0-1), the update formula is:

[0025] in The weight is the weight of the previous time step. For the updated weights, To provide updated supporting evidence, For the updated amount of counter-evidence, the and The updated formula is:

[0026]

[0027] Furthermore, the uncertainty of the weight of each edge is assessed through variance analysis of the Bayesian posterior distribution, and the calculation formula is as follows:

[0028] This variance value is used to determine whether additional weight correction is needed in the next round of data input; By using the recursively updated Bayesian online learning processing method, the event path deviation feature vector of the previous step is transformed into real-time weight adjustment data of the dynamic causal graph edge set, so as to achieve the expected technical effect of adaptive evolution of causal intensity during construction. For example, in a tunnel construction scenario, the historical co-occurrence frequency between a certain excavation event and a subsequent support event is 37 times, with an average time interval of 15 minutes, corresponding to the initial Beta distribution hyperparameter. It is 37. The value is 10. In the current monitoring data, this event sequence exhibits a 5-minute time delay and a significant increase in displacement response in terms of path deviation, corresponding to a supporting evidence value Δα of 4 and a counter-evidence value Δβ of 1. According to the update formula... The calculated expected value of the new weights is 0.754, and a discount factor is introduced. After reaching 0.95, update the formula. The new weight is 0.752, calculated using the variance formula. The variance was 0.0039, indicating low uncertainty, and no additional correction is needed in the next round of data input. This process significantly improves the response speed and matching accuracy of causal weights under dynamic field conditions, ensuring that the early warning module can make accurate linkage responses based on the latest spectral structure.

[0029] like Figure 3 As shown, step S4 involves using a lightweight neural symbolic reasoning model that integrates logical rule constraints and low-rank tensor completion techniques to generate hypotheses for novel event associations not recorded in the dynamic causal graph, outputting potential causal links and their confidence scores. The logical rule constraints are constructed based on geotechnical engineering process logic. Specifically, this includes: S4.1: Based on the logic of geotechnical engineering construction procedures, a set of logical rule constraints is constructed. The logical rules include the sequence of construction events, spatial dependencies, and time window constraints. The rules are formally encoded using first-order logical expressions to generate an executable rule knowledge graph, providing prior logical constraints for the neural symbolic reasoning process. Using the updated version of the dynamic causal graph based on the sequence of geotechnical construction events as input, and employing the logical relationship analysis method of geotechnical engineering construction procedures (parameters: construction procedure type library, spatial association matrix, time window constraint set), the sequential relationship of common construction events in the field is extracted and summarized. Furthermore, by using a spatial dependency modeling method (parameters: BIM spatial coordinate data, work surface proximity threshold), the interdependence of construction events in the spatial dimension is defined, and a set of spatial dependency rules is obtained. Furthermore, by using a time window constraint extraction algorithm (parameters: operation type, standard job cycle, maximum allowable delay time), a reasonable reachability range for the event sequence in the time dimension is defined, and a set of time constraint conditions is generated. Furthermore, a formal encoding method for first-order logic expressions is adopted (parameters: set of logical predicates, set of relational operators) to transform three types of rules—sequence order rules, spatial dependency rules, and time window constraints—into formal logical predicate structures, and to generate a set of machine-readable and parsable rule expressions; Furthermore, by using a rule knowledge graph construction algorithm (parameters: node type definition, edge type definition, attribute set), the formal logical predicates mentioned above are associated and mapped with construction event nodes to generate a structured rule knowledge graph, thereby providing prior logical constraints for the subsequent neural symbol reasoning process. By using formal encoding of logical rule sets and knowledge graph construction, the logical relationships of construction events in the previous step are transformed into standardized and structured rule sets, thereby ensuring the semantic legitimacy of the constraint inputs to the reasoning module and the generation of causal assumptions. For example, in a complex slope reinforcement project, the construction event type library includes four types of events: blasting (BL), excavation (EX), grouting (GR), and support (SP). The spatial correlation matrix is ​​calculated from the BIM model to show pairs of adjacent events with an Euclidean distance of no more than 15m. The proximity threshold for work surfaces is set to 20m. The standard work cycles for operation types are set as follows: blasting 0.5 hours, excavation 2 hours, grouting 4 hours, and support 6 hours. The maximum allowable delay time is set to 24 hours. In the spatial dependency rule, if the distance between a blasting event node and an excavation event node is less than 20m, a spatial dependency relationship exists. The time window constraint formula is expressed as:

[0030] in, The time interval between construction events. This represents the minimum work cycle for the corresponding process. For the rules of excavation and support, a first-order logic expression encoding method is used to formally represent the sequence of processes. After processing by a rule knowledge graph construction algorithm, the above rule set forms an initial rule knowledge graph containing 4 types of event nodes, 12 spatial dependency edges, and 8 temporal order constraint edges. In the subsequent reasoning stage, when constraining reasoning on uncovered event combinations using this rule set, the logical consistency and domain applicability of assumed causal relationships are significantly improved. S4.2: Input the event combinations not covered in the dynamic causal graph into a lightweight neural symbolic reasoning model. The model includes a symbolic reasoning layer and a neural network embedding layer. The symbolic reasoning layer performs logical consistency reasoning based on the rule knowledge graph to generate a set of candidate causal paths to ensure that the assumed causal relationship conforms to the logic of the geotechnical engineering field. S4.3: Use low-rank tensor completion technology to predict and complete the missing causal relationships in the candidate causal path set. Construct a third-order tensor from construction events, spatial locations, and timestamps. Perform tensor decomposition and reconstruction based on the known subset of causal relationships to obtain the initial confidence matrix of potential causal links. The input data includes a set of candidate causal paths obtained by the symbolic reasoning layer of step S4.2. The set elements contain construction event identifiers, spatial location coordinates and timestamp information, and have met the logical constraints of geotechnical engineering procedures. A three-dimensional tensor construction method (dimension settings: event type dimension, spatial location dimension, and timestamp dimension) is adopted to encode the known causal relationships in the candidate causal path set into non-zero elements in the tensor, thereby realizing the structured representation of multidimensional causal data; Furthermore, the third-order tensor is decomposed using a low-rank tensor decomposition algorithm (parameter: the rank r is selected based on the sparsity and information retention rate of the historical causal relationship matrix) to obtain three low-dimensional factor matrices, corresponding to event pattern factors, spatial pattern factors and temporal pattern factors, respectively, in order to characterize the multidimensional pattern structure of potential relationships. Furthermore, by using a tensor reconstruction algorithm (based on the product operation between the factor matrix obtained from decomposition and the kernel tensor), the missing elements in the original tensor are predicted and filled in to obtain the complete causal relationship tensor after filling. Furthermore, the initial confidence level calculation formula is used to numerically evaluate the completed elements:

[0031] in This is the initial confidence level. For event pattern factor similarity, For spatial pattern factor similarity, For the temporal pattern factor similarity, the similarity of the event pattern factor, spatial pattern factor, and temporal pattern factor is calculated using the cosine similarity formula. Through the above calculations, the predicted causal relationships are assigned to the corresponding positions in the initial confidence matrix to form an initial confidence matrix that covers new event associations, thereby achieving quantitative completion of uncovered causal relationships. The low-rank tensor completion algorithm described above transforms the result of the previous step into a matrix containing all candidate causal paths and their initial confidence scores, thereby achieving the expected technical effect of multi-dimensional feature fusion confidence mapping in subsequent S4.4. For example, in a construction scenario, a known subset of causal relationships includes event types E1 (blasting), E2 (grouting), and E3 (support) and their corresponding spatial locations S1 and S2, and timestamps T1 and T2. The matrix sparsity is 0.65, the information retention target is 0.9, and the tensor rank r is set to 3. A 3×2×2 third-order tensor is constructed, where E1→E2 is assigned a value of 1 between positions (S1,T1) and (S2,T2), and the remaining unknown relationships are assigned empty values. Using the CP decomposition method, the tensor is decomposed into an event factor matrix (3×3), a spatial factor matrix (2×3), and a time factor matrix (2×3). Tensor reconstruction is then performed, predicting unknown elements as values ​​such as 0.68 and 0.74. For the predicted value of E2→E3 at (S2,T2) of 0.74, the similarity of the event pattern factor is calculated to be 0.82, the similarity of the spatial pattern factor is 0.9, and the similarity of the temporal pattern factor is 0.85, corresponding to an initial confidence level C of [value missing]. This process yields an initial confidence matrix covering all event combinations, significantly improving the comprehensiveness and reliability of subsequent causal hypothesis scoring. S4.4: Based on the neural network embedding layer, the initial confidence matrix is ​​nonlinearly mapped, and the event feature vector, spatial distance factor and time decay function are introduced as inputs. The output is a causal link confidence score that integrates multi-dimensional information, so as to enhance the engineering applicability and data-driven nature of causal hypothesis. S4.5: Perform threshold filtering and sorting on the confidence scores of the fused causal links, set a dynamic confidence threshold, retain only potential causal links that are higher than the threshold, generate a candidate set of hypothetical causal relationships, and attach confidence labels and causal direction identifiers as input for the subsequent contextual consistency test module.

[0032] Step S5: Execute the context consistency verification module to perform spatial proximity verification, temporal reachability analysis, and process logic compliance judgment on newly added causal links, generate a causal relationship correction set that passes the verification, and write it into the dynamic causal graph main database. Specifically, this includes: S5.1: Based on the newly added causal link event node pairs, obtain their three-dimensional spatial coordinate information in the BIM model, use the spatial proximity calculation function to calculate the Euclidean distance between nodes, if the distance exceeds the preset threshold, it is determined that the spatial proximity is not satisfied, and the causal link assumption is removed to ensure that the newly added causal relationship has physical reachability in space. S5.2: Based on the timestamp of the event node, combined with the standard operating cycle of each operation type in the geotechnical construction process, the time reachability model is used to calculate whether the time interval between two events meets the reasonable time sequence requirements of the operation. If the time interval is less than the minimum operating cycle or greater than the maximum allowable delay time, it is determined that the time reachability is not met, and the causal link is removed to ensure the rationality of causal reasoning in the time dimension. S5.3: Based on the knowledge base of logical rules for geotechnical construction procedures, extract the process sequence constraints related to the newly added causal links, and use a logical consistency verification algorithm to judge the compliance of the event pairs of causal links. If the known process sequence is violated (such as support occurring before excavation), the causal link is marked as a logical conflict item and removed to ensure that the reasoning results comply with engineering specifications. S5.4: For the set of causal links that have passed spatial proximity verification, time accessibility analysis and process logic compliance judgment, perform multi-dimensional confidence fusion calculation, and use a weighted scoring model to integrate the test results of the three dimensions of space, time and logic to generate a comprehensive confidence score of the causal link correction set, so as to quantify its credibility in actual engineering scenarios. For the set of causal links verified by spatial proximity, time accessibility analysis and process logic compliance judgment, a weighted scoring model (parameters: spatial proximity score, time accessibility score, and process logic compliance score weight coefficients) is adopted to achieve numerical fusion processing of multi-dimensional test results; Furthermore, a normalization method (parameter: minimum-maximum normalization interval [0,1]) is used to unify the dimensions of scores from different dimensions and obtain a normalized score vector, ensuring the comparability of each score in the fusion calculation; Furthermore, a weighted summation algorithm is employed (parameter: weight coefficients). satisfy The linear combination of the three-dimensional scores is achieved, as shown in the following formula:

[0033] in, To calculate the overall confidence score, Scoring spatial proximity Rate time accessibility. Scoring for compliance of process logic; Furthermore, the uncertainty of the comprehensive score is quantified by the confidence interval assessment method (parameters: sample variance, sample size), and the score confidence interval is generated to determine the stability of the causal relationship under field conditions. Furthermore, a threshold determination algorithm (parameter: dynamic scoring threshold θ_d optimized based on historical bias rate) is used to transform the comprehensive score into a binary classification result. Causal links above the threshold are marked as trustworthy, and causal links below the threshold are marked as needing verification, thus obtaining the final trustworthiness label of the causal link correction set. By using weighted scoring fusion and threshold judgment processing, the multidimensional test results of the previous step are transformed into a quantitative comprehensive confidence score, thereby realizing the objective quantification of the credibility of causal links in actual engineering scenarios. For example, in a geotechnical construction project, the spatial proximity score is 0.85, the time accessibility score is 0.78, and the process logic compliance score is 0.92. The scores are calculated using weighted coefficients w_s=0.3, w_t=0.4, and w_l=0.3, and then normalized before weighted summation. ,get The comprehensive score was calculated. Based on a sample variance of 0.012 and a sample size of 50, the 95% confidence interval of the score was [0.83, 0.86]. The dynamic threshold θ_d was set to 0.80, and the causal link was determined to be reliable. In this scenario, the comprehensive score was significantly higher than the threshold and the confidence interval was stable, indicating that the causal relationship has high reliability in complex construction environments and can be directly written into the dynamic causal graph main database to provide reliable causal chain support for subsequent early warning triggering. S5.5: The causal link correction set that has passed the context consistency test and obtained a comprehensive confidence score is written into the main database of the dynamic causal graph based on the graph database update protocol. The attribute information of the corresponding nodes and edges in the graph is updated, and the update timestamp and confidence source are recorded to realize the continuous evolution and knowledge accumulation of the causal graph.

[0034] Step S6: Constructing a hierarchical early warning triggering mechanism: The bottom-level risk assessment unit calculates the direct risk probability of the current working condition based on local fragments of the dynamic causal graph; the middle-level event chain propagation simulator tracks high-confidence causal paths and predicts subsequent event sets using a forward traversal method; and the top-level response decision engine generates hierarchical early warning signals based on the number of key nodes and binds them to BIM spatial coordinates. Specifically, this includes: S6.1: Perform a risk assessment algorithm on local graph segments of the dynamic causal graph. Based on the strength weight of the current event node and its direct causal edge, calculate the direct risk probability under the current construction condition and output the local risk assessment result as the core output basis of the bottom risk assessment unit. S6.2: Based on the local risk assessment results, the graph traversal algorithm is used to trace the high-confidence causal path in the dynamic causal graph, identify potential event propagation links and calculate the propagation speed, and generate event chain propagation simulation results to support the prediction function of the mid-level event chain propagation simulator for subsequent construction events. Based on the set of event nodes and their direct risk probabilities contained in the local risk assessment results, a graph traversal algorithm (parameters: the starting point of traversal is the risk assessment node, and the traversal depth is determined by the risk threshold) is used to achieve forward tracing of high-confidence causal paths in the dynamic causal graph. Furthermore, by using a high-confidence path filtering algorithm (parameter: confidence threshold derived from the comprehensive confidence score generated in step S5), only causal edges that meet the confidence conditions are retained as components of the traversal reachable paths, thus obtaining a set of high-confidence paths; Furthermore, an event chain path expansion method is adopted (parameter: limit the number of loop repetitions to reduce computational complexity) to realize sequential access to each node in the high confidence path set and subsequent event identification, and generate potential event propagation link data; Furthermore, a propagation speed calculation algorithm (parameters: timestamp data sourced from the original dataset of real-time event sequences, edge weights derived from the strength weight attributes of the dynamic causal graph) is used to quantitatively calculate the propagation rate between adjacent nodes in the link. The calculation formula is as follows:

[0035] in, For the speed of transmission, The spatial distance between event nodes. For time intervals; Furthermore, based on the propagation speed sequence, a link propagation feature aggregation method (parameters: mean, variance, maximum value) is used to extract the propagation speed distribution features of the entire event chain and generate an event chain propagation simulation result index set; Through the above graph traversal and propagation feature calculation and processing method, the risk assessment results of the previous step are transformed into predictive data containing the propagation link structure and propagation speed indicators of potential events, so as to realize the simulation and prediction function of subsequent construction events. For example, in a tunnel excavation project, the local risk assessment results show that the direct risk probability of node E3 is 0.85. This node is selected as the starting point for traversal, and the traversal depth is set to 3 layers. The high-confidence path selection threshold is 0.75, resulting in paths E3→E5→E7 and E3→E6→E9. The link spatial distance parameter comes from the BIM coordinate system. The spatial distance Δs between E3 and E5 is 12.4 meters, and the time interval Δt is 4.0 seconds. Therefore, the propagation speed v is: The result was 3.1 m / s. The average propagation speed of the link was calculated and averaged for each segment of the path, yielding an average propagation speed of 2.9 m / s with a variance of 0.25 and a maximum speed of 3.4 m / s. The event chain propagation simulation results show that this risk chain may trigger a coordinated response from downstream nodes E7 and E9 within approximately 9 seconds. The prediction accuracy is significantly improved, supporting the accurate characterization of risk expansion trends by the mid-level event chain propagation simulator. S6.3: Based on the number of key nodes, the scope of influence, and the possibility of closed loop in the event chain propagation simulation results, a multi-dimensional decision-making algorithm is used to classify the early warning level and generate three-level early warning signals: yellow alert, orange alarm, and red linkage, in order to realize a multi-level linkage response mechanism. S6.4: Bind and map multi-level early warning signals to spatial coordinates in the Building Information Model (BIM), perform spatial positioning processing of early warning signals based on the spatial distribution information of construction events, and generate a hierarchical early warning information set with spatial labels to enhance the visualization of early warning information and the accuracy of responsibility attribution. S6.5: Based on the hierarchical early warning information set bound to spatial tags, a multi-channel early warning push mechanism is implemented. Early warning information of different levels is pushed to the corresponding responsible parties through mobile terminals, on-site displays and dispatch system interfaces to realize a hierarchical and role-based early warning response process, thereby improving the efficiency of on-site emergency response and linkage response capabilities. Based on a hierarchical early warning information set bound to spatial tags, a multi-channel information distribution algorithm (parameters: early warning level identifier, spatial coordinates, and responsible party ID mapping table) is used to load early warning information of different levels into the distribution queue according to the corresponding spatial region and responsible party mapping relationship. Furthermore, by using the mobile terminal push protocol (parameters: MQTT message format, QoS level, device subscription topic list), the system can push early warning information to the mobile terminals of each responsible party on site in real time and obtain the terminal reception confirmation dataset for subsequent delivery success rate statistics. Furthermore, through the on-site display screen control interface (parameters: display screen ID, display refresh rate, color encoding rules), different levels of warning information are presented on the on-site display screen in the form of a combination of colors and symbols, and display status feedback data is generated to confirm that the on-site visual status is consistent with the system warning information. Furthermore, through the scheduling system integration interface (parameters: API authentication key, information synchronization period, task queue priority), different levels of early warning information can be synchronized to the project scheduling and management system, and scheduling task trigger records can be generated for linkage with on-site emergency response procedures. By using a multi-channel early warning information distribution and processing method, the spatial tag hierarchical early warning information set from the previous step is transformed into real-time early warning push data that is synchronized across platforms and multiple terminals, thereby achieving a hierarchical and role-based closed-loop control effect for early warning response. For example, in a tunnel construction site under complex geological conditions, the system generates a set of spatial label warning information at three levels: yellow alert, orange alarm, and red linkage. The yellow alert is bound to spatial coordinates (X=35.2, Y=128.4, Z=12.7) corresponding to the tunneling face, with the on-site safety officer as the responsible party; the orange alarm is bound to spatial coordinates (X=22.6, Y=140.1, Z=8.3) corresponding to the support work area, with the emergency commander as the responsible party; and the red linkage is bound to spatial coordinates (X=10.1, Y=155.6, Z=5.0) corresponding to the hazardous area associated with the main ventilation system, with the on-site commander as the responsible party. The mobile terminal push protocol is configured as MQTT, with a QoS level of 2, and the terminal subscription topic is set to "Prewarning / RoleID," achieving a high push success rate. The on-site display screen refresh rate is 2 seconds. The yellow alert is presented as a yellow circle, the orange alarm as an orange triangle, and the red linkage as a flashing red square. The displayed status feedback is completely consistent with the warning signal. The dispatch system's integrated interface is called every 60 seconds, with a consistent synchronization cycle. A red alert automatically triggers the ventilation system shutdown. This multi-channel distribution process ensures rapid and accurate transmission of alert information at different levels among responsible parties and equipment, significantly improving emergency response efficiency and coordinated response capabilities.

[0036] Step S7: Activate the periodic backtracking verification module to compare the early warning results with the actual event trajectory. Based on the deviation analysis results, optimize the structural parameters of the dynamic causal graph and the threshold configuration of the hierarchical early warning triggering mechanism to form a closed-loop evolutionary mechanism. Specifically, this includes: S7.1: Based on historical early warning records and construction event retrospective logs, construct a retrospective verification dataset. The dataset includes early warning trigger time, actual event occurrence time, event type, spatial location information, and causal path evolution trajectory to form the data foundation for closed-loop verification. S7.2: Execute a path alignment algorithm on the backtracking verification dataset, based on time window sliding and event sequence editing distance calculation, identify the set of deviation nodes between the early warning prediction path and the actual event evolution path, and generate a path deviation map; S7.3: Based on the path deviation graph, a causal graph structure consistency evaluation model is used to quantitatively evaluate the node connection strength and directionality of related paths in the dynamic causal graph, and output the graph structure deviation index to identify the set of causal edges in the graph that need to be optimized. A causal graph structure consistency evaluation model (parameters: node connection strength threshold, directional consistency weight coefficient) is used to quantify the connection strength of related paths in the dynamic causal graph for the causal link set in the path deviation graph. Furthermore, by using a weighted degree centrality-based method (parameter: weights are taken from the initial causal strength value), the degree value of each node in the path is calculated, and the node connection strength distribution data is obtained, which is used to measure the degree of association between nodes in the path. Furthermore, a directional consistency evaluation algorithm (parameters: direction vector construction rules, timestamp order relationship) is adopted to calculate the matching degree between the causal link direction and the actual event time sequence, and generate a directional consistency score matrix to characterize the rationality of the causal direction in the path; Furthermore, based on the structural consistency comprehensive index formula, the node connection strength and directional consistency scores are fused and calculated, as follows:

[0037] in, It is a comprehensive index for structural consistency. This is the normalized value for connection strength. This is the normalized value for directional consistency. and These are the weighting coefficients for connection strength and directional consistency, respectively. Furthermore, by setting a comprehensive index deviation threshold, a set of causal edges below the threshold is selected as a candidate set of structural deviation causal edges; By integrating the output of the consistency evaluation model, the path deviation map results from the previous step are transformed into a structural deviation index library, thereby achieving the technical effect of locating the set of causal edges that need to be optimized. For example, in a geotechnical tunnel construction scenario, five critical paths from the path deviation graph are selected as evaluation objects. The node connection strength threshold is set to 0.6, the directional consistency weighting coefficient is 0.4, and the initial causal strength values ​​are derived from the normalized historical frequency values ​​of each edge in the dynamic causal graph. After calculating the weighted degree centrality, the mean node connection strength distribution is 0.55, and the normalized C value is 0.55. Using the directional consistency evaluation algorithm, the mean directional consistency is calculated to be 0.62, and the normalized D value is 0.62. Applying the comprehensive index formula, in... =0.6、 Under the condition of 0.4, the comprehensive index of structural consistency is obtained. = =0.578. Based on the deviation threshold of 0.6, this comprehensive index is below the threshold, indicating that the causal edges contained in this path need optimization, and it is ultimately marked into the causal edge candidate set for subsequent parameter updates. In this embodiment, the algorithm effectively identifies causal paths with insufficient structural consistency, providing accurate input for subsequent optimization of the dynamic causal graph structure parameters; S7.4: Based on the graph structure deviation index, the weight parameters of the causal edges are dynamically adjusted using the Bayesian parameter update algorithm. Combined with the actual occurrence frequency of events and the effectiveness of early warning response, the structural parameters of the dynamic causal graph are optimized to improve the accuracy of causal inference. S7.5: Perform gradient descent optimization algorithm on the risk threshold parameters in the hierarchical early warning triggering mechanism. Based on the historical statistical results of the early warning false alarm rate and false alarm rate, dynamically adjust the trigger thresholds of the bottom-level risk assessment unit, the middle-level event chain propagation simulator and the top-level response decision engine to improve the adaptability and stability of the early warning mechanism. S7.6: Write the optimized dynamic causal graph structure parameters and the threshold configuration of the hierarchical early warning triggering mechanism into the main database, and mark the update version number to form a closed-loop system update instance with self-evolution capability, so as to support the continuous optimization of the subsequent early warning inference process.

[0038] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0039] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and rules of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for monitoring and diagnosing the quality of geotechnical construction, characterized in that, Includes the following steps: S1: Construct a prototype of the basic event causal graph based on historical engineering data and expert knowledge base to form an initial causal reasoning framework; S2: Collect real-time sensor data, construction log timestamps, and operation type information during the construction process to form the raw dataset of the event sequence; S3: Execute a temporal similarity matching algorithm on the original dataset of the event sequence to identify the degree of deviation between the current event sequence and the existing paths in the basic causal graph prototype, and dynamically adjust the causal strength of the relevant edges based on the Bayesian online learning framework to generate a dynamic updated version of the causal graph. S4: A neural symbolic reasoning model that combines logical rule constraints with low-rank tensor completion technology generates hypotheses for novel event associations not recorded in the dynamic causal graph, and outputs potential causal links and their confidence scores. S5: Execute the context consistency check module to perform spatial proximity verification, time reachability analysis, and process logic compliance judgment on newly added causal links, generate a causal relationship correction set that passes the check, and write it into the dynamic causal graph main database; S6: Construct a hierarchical early warning triggering mechanism. The bottom-level risk assessment unit calculates the direct risk probability of the current working condition based on local fragments of the dynamic causal graph. The middle-level event chain propagation simulator tracks high-confidence causal paths and predicts subsequent event sets in a forward traversal manner. The top-level response decision engine generates hierarchical early warning signals based on the number of key nodes and binds them to BIM spatial coordinates.

2. The method for monitoring and diagnosing the quality of geotechnical construction according to claim 1, characterized in that, Following step S6, the following is also included: S7: Activate the periodic backtracking verification module to compare the early warning results with the actual event development trajectory. Based on the deviation analysis results, optimize the structural parameters of the dynamic causal graph and the threshold configuration of the hierarchical early warning triggering mechanism to form a closed-loop evolution mechanism.

3. The method for monitoring and diagnosing the quality of geotechnical construction according to claim 1, characterized in that, Step S1 specifically includes: Structured extraction of geotechnical construction event records from historical engineering databases is performed, and event node sets are generated based on graph database modeling technology; Based on the causal rules extracted from the expert knowledge base, causal relationship reasoning is performed on the node pairs in the event node set, and an initial causal edge set is generated using a causal graph modeling algorithm. For each causal edge in the initial causal edge set, causal strength quantification is performed. Based on the co-occurrence frequency and temporal consistency index in the historical event sequence data, the initial strength weight of each causal edge is calculated using the causal strength evaluation algorithm to obtain a weighted causal graph edge set. The weighted causal graph edge set and event node set are integrated into a graph structure, and a basic causal graph prototype is constructed based on the graph database storage structure. The basic causal graph prototype is subjected to context consistency verification. Based on the logic rules of geotechnical engineering procedures and space-time constraints, the causal paths in the graph are verified for compliance. The logic reasoning engine is used to identify and correct causal edges that violate engineering logic, and an optimized initial causal graph is generated.

4. The method for monitoring and diagnosing the quality of geotechnical construction according to claim 3, characterized in that, The event node set includes event type, occurrence time, spatial location, operating parameters, and associated response characteristics, including displacement response curve, stress time series, and pore water pressure change rate.

5. The method for monitoring and diagnosing the quality of geotechnical construction according to claim 1, characterized in that, Step S2 specifically includes: Based on a multimodal sensor network deployed at the geotechnical construction site, the raw signals output by displacement sensors, stress gauges and pore water pressure gauges are acquired to construct a physical response data stream during the construction process. The original signal is filtered and denoised, and a Butterworth low-pass filter is used to remove high-frequency interference components to obtain more stable displacement, stress and pore water pressure measurements. Based on the time synchronization protocol, the filtered sensor data is timestamped and aligned. The NTP network time protocol is used to unify the multi-source data to a global time base and generate a multi-dimensional data sequence. Collect operation type information from construction logs to form structured construction event metadata; The structured construction event metadata is spatiotemporally correlated and matched with the multidimensional data sequence, and the event occurrence time is aligned with the physical response data using a time window sliding algorithm to generate an event-response coupled dataset. Feature extraction is performed on the event-response coupled dataset. The displacement change rate, stress increment, and pore water pressure gradient are calculated based on the sliding window statistical method to generate an event-driven feature vector sequence. The event-driven feature vector sequence is bound to construction event labels to generate a structured event sequence raw dataset, which is then stored in a real-time data buffer.

6. The method for monitoring and diagnosing the quality of geotechnical construction according to claim 1, characterized in that, Step S3 specifically includes: A sliding time window segmentation process is performed on the original dataset of construction event sequences to extract continuous event sequence segments and obtain event subsequence sets; Based on preset event type encoding rules, the event subsequences are converted into low-dimensional temporal embedding vectors to form an event sequence feature vector set; The event sequence feature vector set is matched one by one with the historical path vectors in the basic causal graph prototype to calculate their temporal similarity and output the similarity matrix. Based on the matching results in the similarity matrix that are higher than a set threshold, the degree of deviation between the current event sequence and the existing paths in the basic causal graph prototype is identified, and an event path deviation feature vector is generated. Based on the Bayesian online learning framework, the intensity weights of relevant causal edges are recursively updated using the event path deviation feature vector, new observation information is fused while historical knowledge is retained, and an updated version of the dynamic causal graph is output.

7. The method for monitoring and diagnosing the quality of geotechnical construction according to claim 6, characterized in that, Step S3 further includes event sequence matching using a dynamic time warping algorithm, normalizing feature vectors to the [0,1] interval, using paths with similarity matrices higher than the similarity threshold to generate bias feature vectors, using a Beta distribution for the prior distribution of the Bayesian online learning weight parameters, obtaining the parameters statistically from historical event data, and introducing a discount factor for each weight update to smooth out historical influences.

8. The method for monitoring and diagnosing the quality of geotechnical construction according to claim 1, characterized in that, Step S4 specifically includes: Based on the logic of geotechnical engineering construction procedures, a set of logical rule constraints is constructed. First-order logical expressions are used to formally encode the rules, generating an executable rule knowledge graph. Inputting combinations of events not covered in the dynamic causal graph into a lightweight neural symbolic reasoning model generates a set of candidate causal paths; The missing causal relationships in the candidate causal path set are predicted and filled using low-rank tensor completion technology. The construction event, spatial location, and timestamp are constructed into a third-order tensor. Tensor decomposition and reconstruction are performed based on the known subset of causal relationships to obtain the initial confidence matrix of potential causal links. The initial confidence matrix is ​​nonlinearly mapped based on the neural network embedding layer. Event feature vectors, spatial distance factors and time decay functions are introduced as inputs, and causal link confidence scores are output. The confidence scores of the causal links are subjected to threshold filtering and sorting. A dynamic confidence threshold is set, and only potential causal links with scores higher than the threshold are retained. A candidate set of hypothetical causal relationships is generated and a confidence label and causal direction identifier are attached.

9. The method for monitoring and diagnosing the quality of geotechnical construction according to claim 8, characterized in that, The set of logical rules constraints includes the sequence of construction events, spatial dependencies, and time window constraints.

10. The method for monitoring and diagnosing the quality of geotechnical construction according to claim 1, characterized in that, Step S5 specifically includes: Based on the newly added causal link event node pairs, obtain their three-dimensional spatial coordinate information in the BIM model, and use the spatial proximity calculation function to calculate the Euclidean distance between nodes. If the distance exceeds the preset threshold, it is determined that the spatial proximity is not satisfied, and the causal link hypothesis is removed. Based on the timestamp of the event node, combined with the standard operating cycle of each operation type in the geotechnical construction process, the time reachability model is used to calculate whether the time interval between two events meets the reasonable time sequence requirements of the operation. If the time interval is less than the minimum operating cycle or greater than the maximum allowable delay time, it is determined that the time reachability is not met, and the causal link is removed. Based on the knowledge base of logical rules for geotechnical construction procedures, the sequence constraints of procedures related to newly added causal links are extracted. A logical consistency verification algorithm is used to judge the logical compliance of the event pairs of causal links. If the known sequence of procedures is violated, the causal link is marked as a logical conflict item and removed. For the set of causal links that have passed spatial proximity verification, time accessibility analysis and process logic compliance judgment, perform multi-dimensional confidence fusion calculation to generate a comprehensive confidence score for the causal link correction set; The causal link correction set, which has passed the context consistency test and obtained a comprehensive confidence score, is written into the main database of the dynamic causal graph based on the graph database update protocol. This updates the attribute information of the corresponding nodes and edges in the graph and records the update timestamp and confidence source.

Images