Deep foundation pit construction accident early warning and emergency decision method and system

Abstract

The application discloses a deep foundation pit construction accident early warning and emergency decision method and system, and belongs to the technical field of deep foundation pit construction safety monitoring. In view of the problem that the system cannot make early warning and decision due to communication interruption or sensor failure, the method constructs a knowledge graph, deploys an edge node local cache subgraph and data, collects sensor and video data in real time, selects a cloud cooperation or local fault tolerance mode according to a communication state, uploads data for comparison to generate early warning and push decision in the cloud mode, generates early warning and decision by local inference of the edge node and broadcasts in the local mode, and synchronizes data after communication recovery. The application can maintain basic early warning and decision capability when communication is interrupted or sensors fail.

Description

Technical Field

[0001] This invention belongs to the field of deep foundation pit construction safety monitoring technology, specifically involving a method and system for early warning and emergency decision-making in deep foundation pit construction accidents. Background Technology

[0002] During the construction of deep foundation pit projects, the probability of various safety accidents is relatively high due to complex geological conditions, sensitive surrounding environment, and variable stress state of the support structure. To reduce construction risks, the industry generally adopts sensor network-based monitoring and early warning systems. Displacement sensors, stress sensors, and other equipment are deployed at key parts of the foundation pit support structure to collect deformation and stress data in real time and compare them with preset safety thresholds. An alarm is triggered when the monitored value exceeds the limit. At the same time, video surveillance is also introduced to assist in safety inspections at some project sites. However, existing early warning and emergency decision-making methods still have several shortcomings in practical applications, specifically in the following aspects.

[0003] First, the correlation between monitoring data and accident response measures is not sufficiently expressed. Traditional monitoring systems often use the exceeding of a single physical quantity's threshold as an alarm criterion, lacking a systematic description of the coupling relationship between multiple monitoring indicators and the logic of accident evolution. Since deep foundation pit accidents are usually caused by the combined effect of multiple factors, the instantaneous exceeding of a single indicator does not necessarily correspond to the formation of a specific accident type, and the abnormal changing trends of multiple indicators are difficult to accurately identify through simple threshold rules. When an early warning is generated, on-site managers need to rely on personal experience to judge the accident type and retrieve the corresponding response plan. This process is inefficient and prone to judgment errors under stress, affecting the timeliness and pertinence of emergency response.

[0004] Secondly, the operation of existing systems heavily relies on stable network connections with cloud servers. Monitoring data typically needs to be uploaded to a remote data center in real time for analysis and processing. Early warning assessment and emergency response plan generation are all completed in the cloud before the results are sent to on-site terminals. However, deep foundation pit construction sites are often located in densely populated urban areas with uneven coverage of temporary network facilities, and the lines are easily interrupted by mechanical disturbances during construction. Once a communication link fails, monitoring data cannot be uploaded, cloud analysis capabilities are lost, and the early warning function is essentially rendered inoperable. During this period, even if sensors collect obvious signs of danger, the site cannot obtain effective early warning prompts and handling guidance, thus widening the risk window. The communication interruption problem is more prominent in long-distance tunnel foundation pit or underground structure construction environments, and the duration of the interruption is uncertain. The existing system architecture, which relies on continuous cloud interaction, lacks fault-tolerant measures to address this.

[0005] Secondly, existing systems have limited utilization of real-time dynamic information in emergency decision-making. While some systems can push pre-set emergency response texts to responsible personnel after an alert is triggered, they rarely consider the actual distribution of personnel on-site, the real-time availability of emergency resources, and changes in external traffic conditions. When an accident occurs, emergency personnel may be unable to reach the designated assembly point within the scheduled time due to their different locations, and the dispatch routes of emergency supplies may be delayed due to road congestion. If decision-making plans cannot be dynamically adjusted based on real-time information, their practical operability will be significantly affected.

[0006] Furthermore, the heterogeneous nature of the monitoring data itself poses challenges to early warning assessment. Displacement sensors and stress sensors differ in sampling frequency and time reference, and video surveillance data, being unstructured information, is difficult to directly fuse and compare with structured monitoring data. Without unified preprocessing, data from different sources may reflect inconsistent foundation pit safety conditions, or even contradictory signals, increasing the probability of false alarms and missed alarms. Achieving effective fusion and status assessment of multi-source data at the edge of the field, where communication resources are limited, has always been a challenging engineering challenge in this field.

[0007] In summary, existing methods for early warning and emergency decision-making in deep foundation pit construction accidents have room for improvement in terms of formal expression of accident knowledge, continuous availability under communication interruption scenarios, dynamic information-driven scheme optimization, and multi-source data fusion processing. There is an urgent need for a technical means that can adapt to fluctuations in on-site communication and improve the accuracy of early warning and the timeliness of emergency response. Summary of the Invention

[0008] One object of the embodiments of the present invention is to solve at least the above-mentioned problems and / or defects, and to provide at least the advantages described below.

[0009] Another objective of this invention is to provide a method and system for early warning and emergency decision-making in deep foundation pit construction accidents.

[0010] This paper addresses the problem of existing deep foundation pit construction early warning systems failing to provide continuous early warning and emergency decision-making services after communication interruptions. Current systems rely on cloud-based data processing and decision generation; once network connectivity is lost due to site conditions, monitoring data cannot be uploaded, the early warning function becomes ineffective, and on-site personnel lose safety guidance. Furthermore, the correlation between accident causes, monitoring indicators, and response measures is not embedded in the system as structured knowledge, leading to reliance on human experience for emergency decision-making and limited response efficiency. Maintaining the continuity of early warning and emergency response functions in construction environments with unstable communication is a critical engineering challenge that needs to be overcome in deep foundation pit safety management.

[0011] This addresses the issue of existing emergency decision-making plans failing to optimize and adjust based on the real-time location of on-site personnel and external traffic dynamics. Conventional methods only push static plan texts or fixed lists of emergency measures after an alert is triggered, without considering the spatiotemporal constraints of the actual distribution of emergency personnel and resource accessibility. When an accident occurs, the distance difference between the location of the emergency response leader and the assembly point, as well as traffic congestion along the emergency resource transportation route, directly affect the feasibility of the plan. The lack of a route planning and prioritization mechanism based on real-time dynamic information leads to a disconnect between the emergency plan and the actual on-site situation, delaying the optimal response time.

[0012] This addresses the challenge of how edge computing can accurately generate early warnings and emergency decisions based on limited local cached data during communication or sensor data stream interruptions. During outages, cloud analytics are unavailable, and real-time monitoring data stops updating; edge nodes only possess historical data and incomplete subsets of knowledge graphs. Under these constraints, the core technical challenges of autonomous, fault-tolerant edge operation lie in how to fill data gaps through prediction, how to determine accident types and retrieve response measures based solely on local knowledge fragments, and how to develop usable emergency guidance plans without relying on external real-time traffic and personnel location data.

[0013] To achieve the above-mentioned objectives, the present invention employs the following technical solution: A method for early warning and emergency decision-making in deep foundation pit construction accidents includes the following steps: A knowledge graph of deep foundation pit construction accidents is constructed, which defines the relationships between accident types, accident causes, monitoring indicators, emergency response measures, emergency responsible persons, and emergency resources. Edge computing nodes are deployed at the construction site, and the edge computing nodes locally cache the subgraphs of the knowledge graph and the monitoring data sequences within the most recent time window; Real-time acquisition of monitoring data from displacement and stress sensors on the deep foundation pit support structure, as well as video data from the construction site; Based on the communication status between the edge computing node and the cloud, select to execute either cloud collaboration mode or local fault-tolerant mode; In the cloud-based collaborative mode, monitoring data is uploaded to the cloud and compared with the knowledge graph to generate accident early warning information. Emergency decision-making plans are generated and pushed based on real-time personnel location data and traffic condition data. In local fault-tolerant mode, edge computing nodes perform local inference based on locally cached data to generate accident warning information and emergency decision-making plans, and notify on-site personnel through local broadcast. Once communication is restored, the data from the local fault-tolerant mode will be synchronized to the cloud.

[0014] Preferably, in the deep foundation pit construction accident early warning and emergency decision-making method, the generation of emergency decision-making schemes in the cloud-based collaborative mode includes: Starting with the type of accident, the algorithm retrieves related emergency response measures, responsible persons, and emergency resources from the cloud-based knowledge graph using a graph traversal algorithm. Based on real-time on-site personnel location data and traffic condition data, the retrieved emergency response measures and emergency resources are dynamically sorted, and the optimal dispatch path for emergency resources is planned to generate an emergency decision-making plan that includes step-by-step response instructions, personnel assembly points, and resource arrival times.

[0015] Preferably, in the deep foundation pit construction accident early warning and emergency decision-making method, the local inference in the local fault-tolerant mode includes: When a communication interruption or sensor data stream interruption is detected for more than a preset time threshold, the edge computing node uses a time series prediction model to extrapolate and predict the predicted values ​​of key monitoring indicators during the data interruption period, and then splices the predicted values ​​after the last real monitoring data before the interruption to form a continuous data sequence. The continuous data sequence is compared in real time with the monitoring indicators corresponding to the accident causes in the local cached knowledge graph subgraph. The comparison includes comparing the data with a preset safety threshold and calculating the waveform similarity between the local change trend of the sequence and the preset accident evolution path trend in the subgraph. When the data exceeds the safety threshold and the waveform similarity exceeds the preset matching degree, an accident warning message is generated; Starting with the accident type, the system retrieves associated emergency response measures, responsible persons, and emergency resources from the locally cached knowledge graph subgraph, and generates a static emergency decision-making plan based on the latest personnel locations and preset default traffic conditions.

[0016] Preferably, the deep foundation pit construction accident early warning and emergency decision-making method further includes a knowledge graph dynamic update step: Once an accident warning is generated, real-time monitoring data sequences are continuously collected from the warning time to the end of the preset time window, and it is recorded whether an accident actually occurs within that time window. Calculate the waveform similarity between the real-time monitoring data sequence and the accident evolution path that triggered the early warning information; If an accident actually occurs and the waveform similarity is below the first threshold, the real-time monitoring data sequence will be added to the knowledge graph as a new accident evolution path, and the actual emergency response measures will be associated with this new accident evolution path. If no accident occurs and the waveform similarity is below the second threshold, the real-time monitoring data sequence will be marked as a false alarm mode to adjust the matching conditions of the preset safety threshold or accident evolution path.

[0017] Preferably, the deep foundation pit construction accident early warning and emergency decision-making method further includes a dynamic adjustment step for the emergency plan: After the emergency decision-making plan is pushed out, the system acquires the mobile terminal location data of emergency responders, the GPS trajectory data of emergency vehicles, and the real-time monitoring data continuously collected by on-site sensors. The deviation between the real-time GPS trajectory of the emergency vehicle and the planned optimal dispatch route is calculated. When the deviation exceeds the preset threshold or congestion is detected on the road ahead, the optimal route from the current location to the accident site is replanned and the resource arrival time is updated. At the same time, the monitoring data continuously collected by the on-site sensors is compared with the monitoring data when the accident warning information is generated. If the rate of deterioration of the monitoring data exceeds the preset rate threshold, the accident risk level is determined to have increased. Supplementary emergency response measures and additional emergency resources are automatically retrieved from the knowledge graph, and a reinforcement plan is generated and pushed out. The updated routes, arrival times, and reinforcement plans will be synchronized to the relevant terminals in real time.

[0018] Preferably, the deep foundation pit construction accident early warning and emergency decision-making method further includes a multi-source heterogeneous data preprocessing step: Before comparing the monitoring data with the monitoring indicators corresponding to the accident causes in the knowledge graph, the following preprocessing operations are performed: Timestamp interpolation and alignment of monitoring data from displacement and stress sensors are performed using a unified time reference; A lightweight target detection model is used to extract structural deformation features from keyframes of video data, which are then converted into structured deformation quantification indicators and spatially correlated with displacement sensor data. Monitoring data from different sensors are converted into dimensionless risk contribution values ​​according to their respective preset standardized formulas. The standardized formulas are dynamically determined based on the sensor range and the historical threshold range corresponding to the accident cause. The standardized multi-source data are fused using Kalman filtering, and the resulting integrated monitoring data sequence is used for comparison.

[0019] Preferably, the deep foundation pit construction accident early warning and emergency decision-making method further includes a step of dynamic conflict detection and reallocation of emergency resources: During the execution of the emergency decision-making plan, real-time status data of each emergency resource is continuously collected. The status data includes the current location of the resource, the currently occupied task identifier, the estimated release time, the availability indicator, and the resource type label. When a new accident warning or a change in the resource requirements of an existing task is received, a resource conflict detection is triggered: with a preset time window length as the period, the resource requirements of all unfinished tasks are compared with the occupied time of the currently allocated resources. If the same resource is allocated to two or more different tasks in the same time period, it is marked as a resource conflict event. For resources that are in conflict, extract the urgency score and resource suitability score of all associated tasks. The urgency score is dynamically calculated based on the accident risk level, the number of casualties, and the accident spread speed. The resource suitability score is based on the degree of matching between the resource type and the task requirements, as well as the distance between the current location of the resource and the task location. Using the Hungarian algorithm or genetic algorithm based on conflict resolution, a conflict-free resource reallocation scheme is generated with the goal of maximizing the weighted fitness of the total urgency of all tasks, under the premise that each task is allocated at least one available resource and the total resource limit is not exceeded. The resource allocation instructions affected by the reallocation plan will be pushed to the terminals of relevant emergency response personnel in real time, and the resource arrival time and disposal instructions in the original emergency decision-making plan will be updated simultaneously.

[0020] Preferably, the deep foundation pit construction accident early warning and emergency decision-making method further includes a static decision-making scheme optimization step based on personnel movement trajectory prediction under a local fault-tolerant mode: In local fault-tolerant mode, the edge computing node locally stores the historical location trajectory sequence of each emergency responder within a preset time period. The location trajectory sequence consists of the location data obtained during the last network communication before the interruption and the relative position increment obtained by local short-range wireless ranging after the interruption. When a static emergency decision-making scheme needs to be generated, the edge computing node takes each person's historical location trajectory sequence as input, uses Kalman filtering or particle filtering algorithms to predict the probability location distribution at the current moment, and outputs the estimated location with the highest confidence. For personnel whose positions cannot be predicted through historical trajectories or whose confidence level is below a preset threshold, the edge computing node sets their estimated location to a preset key point corresponding to their position based on the personnel's job responsibility label and the functional zoning map of the deep foundation pit construction area. The estimated location of each person is used as the latest personnel location in the local cache for personnel matching, sorting and assembly point setting in emergency response measures, generating an optimized static emergency decision-making plan; At the same time, an instruction will be added to the emergency information broadcast locally, stating that the personnel location is a predicted value and requesting all personnel to actively confirm their location via walkie-talkie or manually.

[0021] Preferably, the deep foundation pit construction accident early warning and emergency decision-making method further includes an adaptive calibration step for the prediction model parameters under local fault-tolerant mode: After switching to local fault-tolerant mode, the edge computing node acquires the real monitoring data sequence within the last preset time window before the data interruption, which is recorded as the calibration window data. Edge computing nodes employ a sliding window recursive identification method, using calibration window data as input, to estimate the model parameters of the time series prediction model online. The model parameters include autoregressive coefficients, moving average coefficients, and trend term coefficients. The original fixed parameters of the prediction model are replaced by the model parameters obtained from online estimation to form an adaptive prediction model for operating conditions. The aforementioned adaptive prediction model is used to extrapolate and predict the key monitoring indicators during data interruption periods. During the prediction process, after extrapolating a time step, if the prediction time point corresponding to that step does not exceed the upper limit of the interruption duration, the prediction value of that step is used as the input of the next time step and fed back to the prediction model to achieve rolling recursive prediction. Once communication is restored, the actual monitoring data during the interruption period is compared with the predicted values ​​to calculate the prediction error sequence. This error sequence is then used to update the selection strategy for calibration window data or adjust the forgetting factor of recursive identification.

[0022] A deep foundation pit construction accident early warning and emergency decision-making system includes: The knowledge graph construction module is used to construct a knowledge graph of deep foundation pit construction accidents. The knowledge graph defines the relationships between accident types, accident causes, monitoring indicators, emergency response measures, emergency responsible persons, and emergency resources. Edge computing nodes, deployed at the construction site, are used to locally cache the subgraphs of the knowledge graph and the monitoring data sequences within the most recent time window; The data acquisition module is used to collect monitoring data from displacement sensors and stress sensors on the deep foundation pit support structure in real time, as well as video data from the construction site. The mode selection module is used to select between cloud collaboration mode and local fault tolerance mode based on the communication status between the edge computing node and the cloud. The cloud collaboration module is used to upload monitoring data to the cloud and compare it with the knowledge graph in the cloud collaboration mode to generate accident early warning information, and generate and push emergency decision-making plans based on real-time personnel location data and traffic condition data. The local fault tolerance module is used in local fault tolerance mode to enable edge computing nodes to perform local inference based on locally cached data, generate accident warning information and emergency decision-making plans, and notify on-site personnel through local broadcast. The data synchronization module is used to synchronize the data from the local fault-tolerant mode to the cloud after communication is restored.

[0023] Compared with the prior art, the advantages and beneficial technical effects of the present invention are: This invention constructs a knowledge graph of deep foundation pit construction accidents and deploys it on both the cloud and edge computing platforms. This structuredly embeds the relationships between accident types, causes, monitoring indicators, and response measures within the system, replacing the traditional reliance on manual experience-based judgment. When communication is normal, the cloud can utilize the full knowledge graph and real-time external data for accurate comparison and dynamic decision-making. When communication is interrupted, edge nodes autonomously perform reasoning and alerts based on locally cached subgraphs and historical data, ensuring continuous availability of early warning and emergency response functions under fluctuating network conditions at the construction site and avoiding safety monitoring gaps caused by network failures.

[0024] This invention introduces a dynamic sorting and route planning mechanism based on real-time personnel location and traffic conditions in a cloud-based collaborative mode. This enables the generated emergency decision-making plan to accurately reflect the current actual distribution of emergency personnel and resource accessibility. Compared to fixed plan push methods, this method outputs step-by-step handling instructions, personnel assembly points, and resource arrival times that are spatiotemporally targeted. This effectively shortens the time spent on information transmission and manual dispatching in the emergency response chain, and improves the executability and response efficiency of emergency plans in complex on-site environments.

[0025] This invention employs a time-series prediction model in a local fault-tolerant mode to fill in the missing monitoring sequences during data interruptions. It then compares the predicted sequences with the pre-defined accident evolution paths in the local knowledge subgraph based on waveform similarity, enabling risk trend identification even in the absence of real-time data. This approach surpasses simple alarm logic that relies solely on a single threshold exceeding limits. It captures the evolutionary characteristics of multiple indicators changing synergistically, providing valuable early warning information and static emergency guidance even under extreme conditions where communication and sensor data flows are both constrained, significantly enhancing the system's robustness.

[0026] This invention establishes a closed-loop dynamic update mechanism for the knowledge graph based on actual feedback data after an early warning, enabling the system to continuously learn from operational experience. For accidents that actually occur after an early warning, if their evolution patterns differ from existing knowledge, the new patterns can be added to the knowledge base. For false alarms that do not evolve into accidents, the thresholds and matching conditions are optimized in reverse. This mechanism allows the knowledge graph to be continuously enriched and calibrated with engineering applications, gradually reducing the false alarm and false negative rates, and improving the adaptability and long-term accuracy of the early warning system to specific deep foundation pit engineering geology and construction conditions.

[0027] This invention achieves closed-loop monitoring and dynamic re-decision-making of the emergency execution process and accident situation by continuously tracking the trajectory of emergency vehicles and changes in on-site monitoring data after the emergency plan is pushed out. When the route deviates or becomes congested, the route is automatically replanned; when the monitoring data deteriorates rapidly, the risk level is automatically increased and reinforcement plans are retrieved. This enables emergency command to respond and adjust in real time according to the evolution of the situation, avoiding the risk that the initial plan will fail due to unforeseen factors and ensuring dynamic matching between emergency resource investment and the severity of the accident.

[0028] This invention addresses the inconsistencies in time reference, data structure, and physical dimensions between displacement and stress sensor data and video surveillance data through a multi-source heterogeneous data preprocessing step. Timestamp interpolation alignment eliminates comparison biases caused by asynchronous sampling, while structured extraction of video features allows image information to participate in fusion as quantifiable indicators. Standardization and Kalman filtering suppress the interference of single-sensor noise on comprehensive judgment. The comparison results between the fused comprehensive monitoring data sequence and knowledge graph monitoring indicators are more reliable, providing a high-quality data foundation for subsequent early warning inference.

[0029] This invention addresses the problem of globally optimal allocation of limited emergency resources in multi-task concurrent scenarios by introducing resource conflict detection and an optimization algorithm-based reallocation mechanism during emergency decision-making execution. By quantifying task urgency and resource suitability and constructing an objective function to maximize overall efficiency, it can automatically resolve conflicts arising from the repeated assignment of the same resource, generating a scheduling scheme that meets the basic requirements of each task and achieves optimal overall performance. This method avoids the risk of critical tasks being delayed due to resources being occupied by low-priority tasks, and improves the efficiency of emergency resource utilization in complex construction site scenarios with multiple concurrent risks.

[0030] This invention improves the timeliness of personnel distribution information in static emergency plans after communication interruptions by employing a personnel location prediction method based on historical trajectories and short-range ranging information in a local fault-tolerant mode. Kalman filtering or particle filtering algorithms are used to probabilistically infer the possible movement range of personnel during the interruption, and default locations associated with responsibilities are used to replace those with insufficient confidence levels, effectively reducing the problem of inaccurate task allocation caused by outdated location information. Simultaneously, additional prompts encourage on-site personnel to actively confirm their locations, further enhancing the reliability of emergency coordination.

[0031] This invention addresses the issue of decreased prediction accuracy in fixed-parameter prediction models when operating conditions change by performing online adaptive calibration of prediction model parameters when the edge node switches to local fault-tolerant mode. By utilizing the latest real monitoring data before the interruption to identify autoregressive, moving average, and trend term coefficients online, the prediction model can accurately reflect the specific stress and deformation characteristics at the time of the interruption, thereby improving the accuracy of data extrapolation. The prediction error feedback mechanism after communication is restored provides a basis for optimizing subsequent calibration window selection strategies, forming a self-sustaining prediction parameter management system.

[0032] This invention integrates knowledge graph construction, edge caching, pattern selection, cloud-based collaborative reasoning, local fault-tolerant reasoning, and data synchronization into a unified system architecture, providing a complete technical solution capable of adapting to the complex communication conditions at deep foundation pit construction sites. Each module has clearly defined responsibilities and collaborates effectively. When communication is normal, it fully leverages the advantages of cloud computing and full data; when communication is interrupted, it seamlessly switches to autonomous edge operation; and after communication is restored, it automatically completes data backhaul and status synchronization. This system architecture effectively balances the level of intelligent early warning with the requirements of operational reliability, providing a practical technical carrier for the engineering deployment of safety management in deep foundation pit construction.

[0033] Other advantages, objectives, and features of the embodiments of the present invention will be apparent in part from the following description, and in part will be understood by those skilled in the art through study and practice of the embodiments of the present invention. Detailed Implementation

[0034] To further illustrate the technical means and effects of this invention, the following embodiments are provided for further explanation. The specific embodiments described herein are merely illustrative and not intended to limit the scope of the invention.

[0035] It should be noted that, unless otherwise specified, the experimental methods described in the following implementation plan are all conventional methods, and the reagents and materials described are all commercially available unless otherwise specified.

[0036] According to one embodiment of the present invention, a method for early warning and emergency decision-making in deep foundation pit construction includes the following steps: A knowledge graph of deep foundation pit construction accidents is constructed, which defines the relationships between accident types, accident causes, monitoring indicators, emergency response measures, emergency responsible persons, and emergency resources. Edge computing nodes are deployed at the construction site, and the edge computing nodes locally cache the subgraphs of the knowledge graph and the monitoring data sequences within the most recent time window; Real-time acquisition of monitoring data from displacement and stress sensors on the deep foundation pit support structure, as well as video data from the construction site; Based on the communication status between the edge computing node and the cloud, select to execute either cloud collaboration mode or local fault-tolerant mode; In the cloud-based collaborative mode, monitoring data is uploaded to the cloud and compared with the knowledge graph to generate accident early warning information. Emergency decision-making plans are generated and pushed based on real-time personnel location data and traffic condition data. In local fault-tolerant mode, edge computing nodes perform local inference based on locally cached data to generate accident warning information and emergency decision-making plans, and notify on-site personnel through local broadcast. Once communication is restored, the data from the local fault-tolerant mode will be synchronized to the cloud.

[0037] In a specific implementation, a knowledge graph of deep foundation pit construction accidents is first constructed. Historical accident cases, design specifications, construction plans, and expert experience from deep foundation pit projects are collected. Accident types are defined as including support instability, pit bottom heave, seepage failure, and excessive settlement in the surrounding area; accident causes include groundwater level change rate, support axial force loss rate, and soil creep coefficient; monitoring indicators include horizontal displacement rate of the support structure, support axial force, groundwater level, and surrounding surface settlement; emergency response measures include steel support reinforcement, external dewatering, slope top unloading, and personnel evacuation; emergency response personnel are categorized by function as project manager, technical head, safety officer, and monitoring officer; emergency resources include grouting equipment, jacks, drainage pumps, and emergency vehicles. Through entity linking and relation extraction techniques, the above elements are organized into a graph database in the form of nodes and edges, forming the knowledge graph.

[0038] Edge computing nodes are deployed around the foundation pit at the construction site. These nodes possess local computing capabilities and storage units. Based on the functional division and risk level of the construction area, subgraphs closely related to the accident causes in this area are extracted from the complete knowledge graph and cached in the local storage of the edge nodes. Simultaneously, the edge nodes continuously cache monitoring data sequences within the most recent time window, such as the past 24 hours, including displacement and stress time history data.

[0039] For data acquisition, vibrating wire rebar stress gauges and guide wheel inclinometers were deployed at key sections of the foundation pit support structure, with a sampling frequency of once per minute. High-definition cameras were installed around the foundation pit to cover the support nodes and excavation face. Sensor data was aggregated to edge nodes via wired or wireless means, and the video stream was accessed using the RTSP protocol.

[0040] The mode selection module uses a heartbeat detection mechanism to determine the network connectivity between the edge node and the cloud server. If the heartbeat response time is less than 2 seconds for three consecutive times, the communication is considered normal, and the node enters the cloud collaboration mode; if the cumulative heartbeat timeout exceeds 30 seconds, the local fault-tolerant mode is switched.

[0041] In cloud-based collaborative mode, edge nodes upload collected monitoring data to the cloud server in real time. The cloud server performs knowledge graph comparison and reasoning: it matches the current value of the monitoring indicator with the threshold range of the corresponding accident cause in the knowledge graph, and uses a graph traversal algorithm to trace possible accident types starting from the current out-of-limit indicator. For example, when the support axial force decreases at a rate exceeding 0.1 kN / h and continues to decrease for three consecutive sampling points, it retrieves the associated support instability accident type and its confidence level from the knowledge graph. If the matching degree exceeds a preset threshold, it generates an early warning message containing the accident type, probability of occurrence, and scope of impact. The cloud server obtains UWB positioning tag data worn by personnel through the construction management platform interface and obtains real-time road conditions of surrounding roads through the traffic department API. Based on this, it retrieves the corresponding emergency response measures, responsible persons, and resource lists from the knowledge graph, calculates the shortest time to reach the assembly point based on the actual location of the personnel, plans the optimal route for emergency vehicles based on traffic conditions, and generates an emergency decision plan containing step-by-step handling instructions, coordinates of the personnel assembly point, and estimated arrival times of resources. This plan is then pushed to relevant personnel terminals via a mobile app.

[0042] In local fault-tolerant mode, edge nodes cease uploading data to the cloud and activate the local inference engine. Utilizing locally cached knowledge subgraphs and historical monitoring data, judgments are made based on preset accident evolution rules and thresholds within the subgraphs. For example, by comparing the current support axial force with the corresponding safety threshold in the subgraph, and considering the slope trend of the stress-time history curve, an early warning is generated if an instability risk is identified. Simultaneously, based on the associated handling measures and resource list in the subgraph, combined with the locally cached snapshot of the last personnel location and the default traffic conditions (estimated at a normal speed of 30 km / h), a static emergency decision-making plan is generated. The early warning information and plan are broadcast locally via the construction site broadcasting system and explosion-proof walkie-talkie repeaters connected to the edge nodes. After communication is restored, the edge nodes synchronize the early warning event logs, raw sensor data packets, and local inference intermediate results recorded during fault-tolerant operation to the cloud database for subsequent analysis and knowledge updates.

[0043] In existing technologies, deep foundation pit monitoring and early warning systems generally adopt a three-layer architecture of sensor-data collector-cloud platform, with all analysis and decision-making completed in the cloud, and the edge node only responsible for data transmission. In contrast, this implementation deploys knowledge subgraph caching and local inference capabilities at the edge nodes, achieving seamless switching between cloud collaboration and local fault tolerance modes. Existing technologies completely disable the early warning function during network interruptions, leaving on-site personnel without any warning notifications; however, this implementation can continuously generate early warning and emergency guidance information based on locally cached data during communication interruptions and notify personnel through local broadcasting, effectively filling the safety monitoring gap during network failures.

[0044] This implementation transforms scattered accident cases and expert experience into machine-readable structured knowledge by establishing a knowledge graph of deep foundation pit accidents, providing a knowledge foundation for automatic early warning reasoning. Deploying edge computing nodes at the construction site and implementing a mode-switching mechanism overcomes the problem of cloud service unavailability caused by network instability at deep foundation pit construction sites, ensuring the continuity of early warning and emergency decision-making functions. The cloud-based collaborative mode fully utilizes comprehensive knowledge and real-time dynamic data to generate accurate decisions, while the local fault-tolerant mode maintains basic early warning capabilities during communication interruptions. The combination of these two modes ensures high availability of the system in complex construction communication environments. The data synchronization mechanism after communication is restored ensures the integrity and traceability of data during offline periods.

[0045] According to one embodiment of the present invention, the method for early warning and emergency decision-making in deep foundation pit construction accidents, wherein the generation of an emergency decision-making scheme in the cloud-based collaborative mode includes: Starting with the type of accident, the algorithm retrieves related emergency response measures, responsible persons, and emergency resources from the cloud-based knowledge graph using a graph traversal algorithm. Based on real-time on-site personnel location data and traffic condition data, the retrieved emergency response measures and emergency resources are dynamically sorted, and the optimal dispatch path for emergency resources is planned to generate an emergency decision-making plan that includes step-by-step response instructions, personnel assembly points, and resource arrival times.

[0046] The specific implementation method for generating emergency decision-making solutions in the cloud-based collaborative mode is as follows: When the cloud-based early warning module determines that the accident type is support instability, it uses the node of this accident type in the knowledge graph as the starting point and executes a breadth-first graph traversal algorithm. It retrieves related measure nodes along the accident type-emergency response measure relationship, including immediately adding steel supports, removing the load at the pit edge, and suspending dewatering in the pit; it retrieves the responsible roles along the emergency response measure-emergency responsible person relationship, identifying specific personnel such as structural engineers and safety directors; and it retrieves the required equipment along the emergency response measure-emergency resource relationship, such as 500kN hydraulic jacks and H-beam support components.

[0047] The cloud server obtains the real-time coordinates of emergency personnel wearing location tags from the positioning system, calculates their walking distance and estimated time to the pre-set assembly point, such as the safe area north of the foundation pit. It also obtains real-time traffic conditions for the optimal route from the emergency resource storage point to the site from the traffic information interface. If a red congestion section is detected, the route planning engine is invoked to recalculate the detour route based on the current road network status. Based on the above dynamic information, multiple emergency response measures are prioritized: measures where the required resources can be dispatched nearby and the responsible personnel can arrive fastest are executed first. Simultaneously, the optimal dispatch route from the current location to the designated location is planned for each emergency resource, with estimated arrival time, forming a structured emergency decision-making plan. This plan includes: first, the safety officer organizes the workers in the foundation pit to evacuate to the ground assembly point along the east side passage within 5 minutes; second, the materials officer dispatches jacks and steel supports from the west storage yard, expected to arrive at the northeast corner of the foundation pit in 15 minutes; third, the structural engineer waits at the assembly point for the resources to arrive and directs the installation of temporary supports. The plan is pushed to the mobile terminals of relevant personnel in text and map annotation form.

[0048] This implementation uses a graph traversal algorithm to quickly retrieve the complete set of emergency elements associated with the accident type, avoiding the delays caused by manually consulting emergency plan manuals. The introduction of real-time personnel location and traffic data for dynamic sorting and route planning ensures that the generated emergency plan closely matches the actual spatiotemporal conditions of the scene, reducing decision-making biases caused by outdated information. The clear output of step-by-step disposal instructions, personnel assembly points, and resource arrival times provides clear and actionable guidelines for on-site emergency command, helping to shorten the response time from the issuance of warnings to the commencement of disposal actions.

[0049] According to one embodiment of the present invention, the local inference in the local fault-tolerant mode of the deep foundation pit construction accident early warning and emergency decision-making method includes: When a communication interruption or sensor data stream interruption is detected for more than a preset time threshold, the edge computing node uses a time series prediction model to extrapolate and predict the predicted values ​​of key monitoring indicators during the data interruption period, and then splices the predicted values ​​after the last real monitoring data before the interruption to form a continuous data sequence. The continuous data sequence is compared in real time with the monitoring indicators corresponding to the accident causes in the local cached knowledge graph subgraph. The comparison includes comparing the data with a preset safety threshold and calculating the waveform similarity between the local change trend of the sequence and the preset accident evolution path trend in the subgraph. When the data exceeds the safety threshold and the waveform similarity exceeds the preset matching degree, an accident warning message is generated; Starting with the accident type, the system retrieves associated emergency response measures, responsible persons, and emergency resources from the locally cached knowledge graph subgraph, and generates a static emergency decision-making plan based on the latest personnel locations and preset default traffic conditions.

[0050] In local fault-tolerant mode, the specific process of edge computing nodes performing local inference is as follows. Edge nodes have built-in communication status monitoring threads and sensor data stream monitoring threads. When a TCP long connection to the cloud is detected to be broken and reconnection attempts fail consecutively for more than a preset threshold of 60 seconds, or when displacement sensor data packets from the data acquisition unit are detected to be interrupted for more than 30 seconds, the edge node determines that it has entered a data interruption state.

[0051] At this point, the edge node invokes a locally deployed time-series forecasting model, which is trained offline using a seasonal autoregressive integral moving average model. Starting from the last real monitoring data point before the interruption, the forecasting model extrapolates to generate predicted values ​​for key monitoring indicators at subsequent time steps. For example, if the last three sampling points of the support axial force data before the interruption were 852kN, 848kN, and 843kN, the extrapolated values ​​after the forecasting model at four steps would be 838kN, 833kN, 829kN, and 825kN. These predicted values ​​are then concatenated with the real data to form a continuous data sequence.

[0052] Edge nodes compare continuous data sequences with preset accident evolution path trends in their locally cached knowledge subgraphs. This subgraph defines the evolution path trend of a support instability accident as a sustained decrease in support axial force exceeding 5% within one hour, with the rate of decrease gradually increasing. The corresponding monitoring indicator has a preset safety threshold of 80% of the design bearing capacity, i.e., 800kN. It also stores a normalized waveform template for this evolution path. The comparison process includes two dimensions: first, comparing the current axial force value with the 800kN threshold; second, calculating the dynamic time warping distance between the axial force decrease curve over the past 20 minutes and the accident evolution waveform template in the subgraph, serving as a waveform similarity metric. When the monitored value is below 800kN and the calculated waveform similarity value is greater than the preset matching degree of 0.75, the edge node generates a support instability accident early warning.

[0053] After an early warning is generated, starting with the accident type, the associated emergency response measures, responsible persons, and resources are retrieved from the local cache subgraph. Due to communication interruption, edge nodes cannot obtain real-time personnel location and traffic data; therefore, the last synchronized personnel location snapshot before the interruption is used as the latest personnel location, and the traffic status is preset to a default value to generate a static emergency decision-making plan. This plan is broadcast cyclically via voice through the broadcast system connected to the edge nodes, informing on-site personnel of the accident type, evacuation direction, and key points of temporary handling.

[0054] This implementation method maintains the ability to continuously assess the safety status of the foundation pit under extreme conditions of overlapping communication and data flow interruptions by extrapolating the changing trends of key monitoring indicators using a time series prediction model. By comparing the waveform similarity between the predicted sequence and the accident evolution path in the local knowledge subgraph, abnormal signs with dangerous evolution trends can be identified earlier than with a single threshold judgment, reducing the risk of missed reports. Although the locally cached knowledge subgraph is not as complete as the full knowledge in the cloud, it contains the most core accident types and handling measures for the region, sufficient to support the generation of basic emergency guidance plans. Local broadcasting ensures on-site coverage of early warning information, guaranteeing that on-site personnel still receive safety alerts even when public network communication fails.

[0055] According to one embodiment of the present invention, the deep foundation pit construction accident early warning and emergency decision-making method further includes a knowledge graph dynamic update step: Once an accident warning is generated, real-time monitoring data sequences are continuously collected from the warning time to the end of the preset time window, and it is recorded whether an accident actually occurs within that time window. Calculate the waveform similarity between the real-time monitoring data sequence and the accident evolution path that triggered the early warning information; If an accident actually occurs and the waveform similarity is below the first threshold, the real-time monitoring data sequence will be added to the knowledge graph as a new accident evolution path, and the actual emergency response measures will be associated with this new accident evolution path. If no accident occurs and the waveform similarity is below the second threshold, the real-time monitoring data sequence will be marked as a false alarm mode to adjust the matching conditions of the preset safety threshold or accident evolution path.

[0056] The knowledge graph dynamic update process is executed offline by the cloud server after the warning event ends. After the accident warning information generated during the local fault-tolerant mode is uploaded after the cloud collaboration mode or communication is restored, the system continuously collects the complete monitoring data sequence from the warning time T0 to the end time T0+2h of the preset time window, and at the same time obtains the markers of whether an accident actually occurred within this time window from the construction log system.

[0057] Calculate the dynamic time-warped similarity between the actual monitoring data sequence and the waveform template of the accident evolution path that triggered this early warning. The waveform similarity value ranges from 0 to 1, with a higher degree of similarity indicating a closer match.

[0058] If a corresponding type of accident does occur after the warning, but the calculated waveform similarity is below the first threshold (e.g., 0.6), it indicates a significant difference between the actual accident development process and the accident evolution path already stored in the knowledge graph. For example, in this accident, the support axial force decreased in a step-like manner rather than continuously. In this case, the system adds this actual monitoring data sequence as a new accident evolution path instance to the knowledge graph, as a new evolution branch of the support instability accident type, and establishes a connection between the actual and effective measures implemented during the emergency response and this new path. If an actual accident occurs and the similarity is greater than or equal to 0.6, it is considered as effective coverage of existing knowledge, and only the statistical frequency of this path is updated.

[0059] If no accident occurs after the warning, but the waveform similarity is below the second threshold (e.g., 0.4), it indicates that the warning signal was not generated due to typical accident precursors, but may be caused by instantaneous sensor disturbances or data fluctuations due to normal construction procedures such as prestressing. In this case, the system marks the monitored data sequence as a false alarm and stores it in the false alarm sample library. Subsequently, the safety threshold corresponding to the accident cause can be adjusted accordingly, for example, lowering the original threshold from 800kN to 780kN to reduce false alarms, or adjusting the matching conditions in the waveform matching algorithm, such as increasing the similarity threshold or modifying the template length.

[0060] This implementation establishes a post-early warning evaluation and knowledge feedback mechanism, making the knowledge graph no longer a static, pre-set knowledge set, but a living knowledge base that evolves with engineering applications. By identifying and adding new accident evolution paths, the system's ability to identify diverse accident precursor patterns is expanded, reducing the risk of missed alarms. By analyzing false alarm patterns and adjusting thresholds and matching conditions in reverse, false alarms caused by normal construction disturbances can be gradually reduced, improving the credibility of early warnings, alleviating the desensitization of on-site personnel to frequent false alarms, and ensuring that early warning signals truly play their due role in alerting the public.

[0061] According to one embodiment of the present invention, the deep foundation pit construction accident early warning and emergency decision-making method further includes a dynamic adjustment step for the emergency plan: After the emergency decision-making plan is pushed out, the system acquires the mobile terminal location data of emergency responders, the GPS trajectory data of emergency vehicles, and the real-time monitoring data continuously collected by on-site sensors. The deviation between the real-time GPS trajectory of the emergency vehicle and the planned optimal dispatch route is calculated. When the deviation exceeds the preset threshold or congestion is detected on the road ahead, the optimal route from the current location to the accident site is replanned and the resource arrival time is updated. At the same time, the monitoring data continuously collected by the on-site sensors is compared with the monitoring data when the accident warning information is generated. If the rate of deterioration of the monitoring data exceeds the preset rate threshold, the accident risk level is determined to have increased. Supplementary emergency response measures and additional emergency resources are automatically retrieved from the knowledge graph, and a reinforcement plan is generated and pushed out. The updated routes, arrival times, and reinforcement plans will be synchronized to the relevant terminals in real time.

[0062] After the emergency decision-making plan is pushed to the relevant personnel's terminals via the cloud, the system enters the emergency execution tracking phase. The cloud server continuously subscribes to the location data reported by the mobile terminals held by emergency responders, such as smartphones or smart helmets, from the location service platform, with a sampling frequency of once every 10 seconds; it obtains GPS trajectory data from the on-board terminals of emergency vehicles, with a frequency of once per second; and it continuously receives real-time monitoring data streams collected by on-site sensors.

[0063] The real-time GPS trajectory points of the emergency vehicle are matched against the previously planned optimal dispatch route from the current location to the accident site using a map, and deviation is calculated. Deviation is defined as the vertical distance from the vehicle's actual location to the nearest point on the planned route. If the deviation exceeds a preset threshold of 50m for three consecutive sampling points, or if a traffic information interface indicates sudden congestion 2km ahead of the vehicle (e.g., average speed dropping below 5km / h), a route replanning process is triggered. Route replanning uses the vehicle's current GPS location as the starting point and the accident site coordinates as the ending point, calling the navigation engine to avoid congested sections and generate a new route, and recalculates the estimated arrival time. The updated route information and arrival time are sent to the driver's mobile terminal and the command center interface via push notification service.

[0064] Meanwhile, the cloud server analyzes the monitoring data continuously transmitted from the on-site sensors. It extracts the characteristic rate of change of the monitoring data before and after the warning time; for example, if the initial rate of decrease in support axial force is 0.2 kN / min, it increases to 0.8 kN / min within the following 10 minutes. If the rate of deterioration exceeds a preset threshold of 0.5 kN / min, the accident risk level is determined to have increased from yellow to orange. The system automatically retrieves supplementary emergency response measures from the knowledge graph using the accident type and higher risk level as search criteria, such as expanding the warning area, calling up backup grouting equipment, and additional emergency resources. It then generates a reinforcement plan and pushes it to the terminals of the reinforcement personnel.

[0065] The updated route information, estimated arrival time, and reinforcement plan will be synchronized to all relevant terminals in real time to maintain the overall consistency of emergency command information.

[0066] This implementation introduces a real-time tracking and dynamic adjustment mechanism during the emergency response phase, making the emergency plan no longer a static instruction but a flexible plan that can be modified in real time according to changes on the ground. When emergency vehicles deviate from their routes or encounter congestion, they are automatically replanned, preventing resources from arriving late to the optimal response window due to route delays. Continuous analysis of the rate of deterioration in monitoring data allows for timely detection of signs of escalation and automatic triggering of reinforcement procedures, avoiding the potential lag in manual judgment of escalation timing. Real-time information synchronization ensures that all emergency personnel have access to the latest situation, improving the efficiency and success rate of collaborative response.

[0067] According to one embodiment of the present invention, the deep foundation pit construction accident early warning and emergency decision-making method further includes a multi-source heterogeneous data preprocessing step: Before comparing the monitoring data with the monitoring indicators corresponding to the accident causes in the knowledge graph, the following preprocessing operations are performed: Timestamp interpolation and alignment of monitoring data from displacement and stress sensors are performed using a unified time reference; A lightweight target detection model is used to extract structural deformation features from keyframes of video data, which are then converted into structured deformation quantification indicators and spatially correlated with displacement sensor data. Monitoring data from different sensors are converted into dimensionless risk contribution values ​​according to their respective preset standardized formulas. The standardized formulas are dynamically determined based on the sensor range and the historical threshold range corresponding to the accident cause. The standardized multi-source data are fused using Kalman filtering, and the resulting integrated monitoring data sequence is used for comparison.

[0068] The multi-source heterogeneous data preprocessing step is performed by edge computing nodes or cloud preprocessing modules before the monitoring data enters the knowledge graph comparison process. First, the raw sampling data from displacement and stress sensors are timestamped and aligned using a unified time base synchronized by the NTP network time protocol. Since there may be a second-level deviation in the sampling time of the two types of sensors, a linear interpolation method is used to resample the data to a unified 1-second interval time series.

[0069] Secondly, a lightweight object detection model is used to process the video data. This model is a simplified YOLO network optimized for construction scenarios, running on GPU or NPU acceleration units at edge nodes. The model infers from keyframes transmitted from the camera, identifies regions of interest such as the edges of the supporting structure and bolt connections, calculates the pixel displacement changes of feature points between adjacent frames, and outputs structured deformation quantification indices, such as the relative displacement value of the support endpoint, in mm, after conversion by camera calibration parameters. This deformation quantification index is spatially correlated with measured data from displacement sensors in the same spatial location as a basis for cross-validation of visual and physical sensing data.

[0070] Subsequently, the monitoring data from different sensors were standardized. Each sensor had a pre-defined standardized formula: the risk contribution value equals the difference between the current measured value and the lower safety limit, divided by the difference between the upper and lower safety limits, with the result limited to between 0 and 1. The upper and lower safety limits are dynamically determined based on the sensor's range and the historical threshold range of the corresponding accident triggers. For example, if a support axial force meter has a range of 1000kN and a design warning value of 800kN, then the risk contribution value corresponding to an axial force of 850kN is approximately 1.06 (850-0 divided by 800-0), which is taken as 1.0. If the displacement gauge's positive displacement warning value is 30mm and the current displacement is 22mm, then the risk contribution value is approximately 0.73 (22 divided by 30).

[0071] Finally, a Kalman filter was applied to fuse the standardized multi-source risk contribution value sequences. Using displacement and stress data as observations, a linear dynamic system model of the foundation pit's safety status was established. The fused comprehensive risk index sequence was recursively estimated using Kalman filtering. This sequence smoothed out the instantaneous noise and anomalous jumps from individual sensors, more stably reflecting the evolution trend of the overall safety situation of the foundation pit. The fused comprehensive monitoring data sequence replaced the original single-source data in the subsequent early warning comparison process.

[0072] This implementation eliminates the timing inconsistency problem caused by asynchronous sampling from multiple sensors through interpolation alignment with a unified time reference, ensuring the accuracy of subsequent trend comparisons. Unstructured video data is transformed into structured deformation indicators, allowing visual information to participate in fusion calculations in numerical form, compensating for the limitations of physical sensors in terms of spatial coverage density. A dynamic standardization formula based on sensor range and historical thresholds ensures the comparability of monitored values ​​for different physical quantities on a unified risk scale. Kalman filtering effectively filters out random noise interference, resulting in a smoother and more reliable response of the output comprehensive monitoring data sequence to changes in real security status. This provides high-quality data input for accurate knowledge graph comparisons and reduces the probability of false alarms due to data quality issues.

[0073] According to one embodiment of the present invention, the deep foundation pit construction accident early warning and emergency decision-making method further includes an emergency resource dynamic conflict detection and reallocation step: During the execution of the emergency decision-making plan, real-time status data of each emergency resource is continuously collected. The status data includes the current location of the resource, the currently occupied task identifier, the estimated release time, the availability indicator, and the resource type label. When a new accident warning or a change in the resource requirements of an existing task is received, a resource conflict detection is triggered: with a preset time window length as the period, the resource requirements of all unfinished tasks are compared with the occupied time of the currently allocated resources. If the same resource is allocated to two or more different tasks in the same time period, it is marked as a resource conflict event. For resources that are in conflict, extract the urgency score and resource suitability score of all associated tasks. The urgency score is dynamically calculated based on the accident risk level, the number of casualties, and the accident spread speed. The resource suitability score is based on the degree of matching between the resource type and the task requirements, as well as the distance between the current location of the resource and the task location. Using the Hungarian algorithm or genetic algorithm based on conflict resolution, a conflict-free resource reallocation scheme is generated with the goal of maximizing the weighted fitness of the total urgency of all tasks, under the premise that each task is allocated at least one available resource and the total resource limit is not exceeded. The resource allocation instructions affected by the reallocation plan will be pushed to the terminals of relevant emergency response personnel in real time, and the resource arrival time and disposal instructions in the original emergency decision-making plan will be updated simultaneously.

[0074] During the execution of the emergency response plan, the resource management module of the cloud server continuously collects real-time status data of various emergency resources. Emergency resources include emergency vehicles, large equipment, and specialized rescue teams, each with a unique identifier in the system. Status data includes the resource's current GPS coordinates, the task number of the currently occupied task, the estimated release timestamp, an availability Boolean flag, and resource type tags such as grouting machine, generator, and diver team. Status data is reported every minute via resource management terminals or vehicle-mounted terminals.

[0075] When the system receives new accident warning information and generates new emergency task requirements, or when an ongoing task requests additional resources due to changes in on-site conditions, the resource conflict detection module is triggered. The module uses a preset two-hour time window to extract a list of all incomplete emergency tasks and their required resource types and quantities, comparing each list with the currently allocated resource's usage period. If it detects that the same resource is allocated to two or more different tasks within its usage period—for example, if Grouting Machine No. 3 is simultaneously requested by Task A and Task B between 3 PM and 4 PM—it is marked as a resource conflict event.

[0076] All tasks associated with conflicting resources are scored for urgency and resource suitability. The urgency score is calculated by weighting and summing factors including the accident risk level, the number of reported trapped or injured persons, and the rate of deterioration of monitored data; a higher score indicates a more urgent task. The resource suitability score is calculated by weighting factors including the degree of match between the resource type label and the task requirements, and the inverse normalized value of the straight-line distance between the resource's current location and the task location; a higher score indicates that the resource is more suitable for performing the task.

[0077] Resource reallocation is performed using a Hungarian algorithm improved based on conflict resolution. A cost matrix is ​​constructed with all tasks to be assigned as rows and available resources as columns. The cost value is the reciprocal of the product of the task urgency score and the resource fitness score. Under the constraints that each task is allocated at least one available resource of the required type and the total number of each type of resource allocated does not exceed its available quantity, the algorithm finds the allocation scheme with the minimum total cost, which is equivalent to maximizing the total urgency-weighted fitness of all tasks. For the genetic algorithm implementation, chromosomes encode the mapping relationship between tasks and resources, and the fitness function is the total urgency-weighted fitness, which is iteratively optimized through selection, crossover, and mutation operations.

[0078] The algorithm outputs a conflict-free resource reallocation scheme that includes a list of resources to be reassigned for each task and their expected usage time. In real time, the algorithm pushes affected resource scheduling instructions, such as changing a grouting machine from task A to task B, to the mobile terminals of the resource operator and the person in charge of task B. Simultaneously, it updates the resource arrival time field in the emergency decision-making interface and the resource call portion of the handling instructions.

[0079] This implementation establishes a real-time perception and automatic conflict detection mechanism for emergency resource status, enabling resource allocation conflicts to be proactively detected by the system rather than relying on manual reports. Quantitative calculation of urgency and suitability scores provides an objective basis for prioritizing different tasks, avoiding subjective judgment bias. Conflict resolution based on the Hungarian algorithm or genetic algorithm seeks resource allocation schemes that maximize overall emergency benefits while meeting the basic resource requirements of each task, ensuring that the most urgent tasks receive the most suitable resources first. Real-time push update instructions ensure timely communication of scheduling changes, avoiding resource idleness or duplicate scheduling due to information lag. Overall, it improves the efficiency of emergency resource utilization and response performance in scenarios with multiple concurrent incidents or resource shortages.

[0080] According to one embodiment of the present invention, the deep foundation pit construction accident early warning and emergency decision-making method further includes a static decision scheme optimization step based on personnel movement trajectory prediction under a local fault-tolerant mode: In local fault-tolerant mode, the edge computing node locally stores the historical location trajectory sequence of each emergency responder within a preset time period. The location trajectory sequence consists of the location data obtained during the last network communication before the interruption and the relative position increment obtained by local short-range wireless ranging after the interruption. When a static emergency decision-making scheme needs to be generated, the edge computing node takes each person's historical location trajectory sequence as input, uses Kalman filtering or particle filtering algorithms to predict the probability location distribution at the current moment, and outputs the estimated location with the highest confidence. For personnel whose positions cannot be predicted through historical trajectories or whose confidence level is below a preset threshold, the edge computing node sets their estimated location to a preset key point corresponding to their position based on the personnel's job responsibility label and the functional zoning map of the deep foundation pit construction area. The estimated location of each person is used as the latest personnel location in the local cache for personnel matching, sorting and assembly point setting in emergency response measures, generating an optimized static emergency decision-making plan; At the same time, an instruction will be added to the emergency information broadcast locally, stating that the personnel location is a predicted value and requesting all personnel to actively confirm their location via walkie-talkie or manually.

[0081] In local fault-tolerant mode, edge computing nodes store the historical location trajectory sequence of each emergency responder over a preset 24-hour period. This trajectory sequence consists of two parts: personnel location data synchronized from the cloud during the last normal network communication before the interruption, recorded as absolute latitude and longitude or local coordinates at the construction site; and relative position increments calculated by triangulation based on the relative distance information of personnel tags obtained after the interruption through local short-range wireless ranging base stations such as UWB anchors deployed around the foundation pit. The two are then fused to form a continuous location estimation sequence.

[0082] When personnel location data is needed for local static emergency decision-making, the edge node initiates a personnel location prediction program. For each personnel, a kinematic state equation is established using their historical trajectory sequence as input, and a Kalman filter algorithm is used for recursive state prediction. The Kalman filter prediction step estimates the current position using the position and velocity from the previous moment, while the update step corrects the prediction result using the latest short-range ranging observations, outputting the estimated position coordinates with the highest posterior probability and the corresponding confidence ellipse size.

[0083] For certain personnel, if the historical trajectory sequence is too short, causing the prediction model to diverge, or if the output confidence level is lower than the preset threshold of 0.7, then the edge nodes will not use the prediction result. Instead, they will be set to a default position based on the responsibility label bound to the personnel in the knowledge graph subgraph. For example, for personnel whose responsibility label is concrete worker, their estimated position will be set to the center point of the formwork operation area by default, based on the functional zoning map of the deep foundation pit construction area; for personnel whose responsibility label is monitor, their position will be set to a key point of the monitoring channel around the foundation pit by default.

[0084] The predicted or default locations of each person are used as the latest personnel locations cached locally. These locations are then substituted into the personnel matching and sorting algorithm in the emergency response measures to calculate the estimated arrival time of each person at each assembly point. The person with the shortest estimated arrival time is selected and assigned the appropriate task, and reasonable assembly point coordinates are set accordingly. An optimized static emergency decision-making plan is generated. Simultaneously, when the edge node controls the local broadcast system to broadcast emergency information, it adds a voice and text prompt stating that the personnel locations are predicted values ​​and all personnel should actively confirm their locations via walkie-talkie or manually, reminding on-site personnel that location information may be inaccurate and requires active verification.

[0085] Under the constraint of communication interruption and the inability to obtain real-time positioning, this implementation method predicts location by fusing historical trajectory data and post-interruption short-range ranging incremental data. This makes the personnel distribution information used in the static emergency plan more realistic than simply relying on outdated locations. For cases with insufficient prediction confidence, a responsibility-associated default location is used as a substitute, ensuring that each person has a reasonable estimated location for decision-making calculations and avoiding personnel mismatch issues due to missing locations. Additional manual confirmation prompts encourage on-site personnel to actively participate in location verification, further reducing the risk of emergency coordination errors due to inaccurate location information and enhancing the reliability of local fault-tolerant emergency command.

[0086] According to one embodiment of the present invention, the deep foundation pit construction accident early warning and emergency decision-making method further includes an adaptive calibration step for the prediction model parameters under local fault-tolerant mode: After switching to local fault-tolerant mode, the edge computing node acquires the real monitoring data sequence within the last preset time window before the data interruption, which is recorded as the calibration window data. Edge computing nodes employ a sliding window recursive identification method, using calibration window data as input, to estimate the model parameters of the time series prediction model online. The model parameters include autoregressive coefficients, moving average coefficients, and trend term coefficients. The original fixed parameters of the prediction model are replaced by the model parameters obtained from online estimation to form an adaptive prediction model for operating conditions. The aforementioned adaptive prediction model is used to extrapolate and predict the key monitoring indicators during data interruption periods. During the prediction process, after extrapolating a time step, if the prediction time point corresponding to that step does not exceed the upper limit of the interruption duration, the prediction value of that step is used as the input of the next time step and fed back to the prediction model to achieve rolling recursive prediction. Once communication is restored, the actual monitoring data during the interruption period is compared with the predicted values ​​to calculate the prediction error sequence. This error sequence is then used to update the selection strategy for calibration window data or adjust the forgetting factor of recursive identification.

[0087] Once the edge computing node detects a communication interruption and switches to local fault-tolerant mode, the adaptive calibration step for the prediction model parameters is immediately initiated. The edge node extracts the last set of real monitoring data within a preset time window (e.g., 30 minutes) before the data interruption from its locally stored time-series database, using this data as the calibration window. This data segment is continuous in time and accurate in value, reflecting the specific stress and deformation characteristics of the foundation pit support structure in the period immediately preceding the interruption.

[0088] Edge nodes are identified using a sliding window method based on recursive least squares, with calibration window data as the input sequence to estimate various parameters of the time series prediction model online. The identified model parameters include: coefficients of the autoregressive component, describing the linear dependence of the current monitored value on the monitored values ​​at several past times; coefficients of the moving average component, describing the dependence of the current monitored value on the prediction error at several past times; and coefficients of the trend term, describing the slope of the deterministic trend of the monitored sequence. The identification process employs a recursive algorithm with a forgetting factor, set between 0.95 and 0.99, ensuring that recent data contributes more significantly to the parameter estimation than earlier data.

[0089] The online-estimated model parameters replace the fixed parameters obtained through offline training in the original prediction model, forming an adaptive prediction model for the current interruption condition. This adaptive model is then used to extrapolate and predict key monitoring indicators during the data interruption period. During the prediction process, a rolling recursive approach is employed: after extrapolating to the next time step, if the predicted time point corresponding to that step has not yet exceeded the preset interruption duration limit (e.g., 2 hours), the predicted value is fed back as a new known value to the model input to predict the value for the next step. This recursive approach continues until the interruption duration limit is reached or communication is restored.

[0090] Once communication is restored, the edge nodes compare the actual monitoring data stored locally during the interruption period—the retransmitted data uploaded after sensor recovery—with the predicted value sequence point by point, calculating the prediction error sequence composed of the root mean square error and the maximum absolute error. This error sequence is recorded and uploaded to the cloud analysis module for evaluating the accuracy of the prediction. Based on the statistical characteristics of the error sequence, the cloud updates the selection strategy for subsequent calibration window data; for example, it appropriately extends the calibration window length or adjusts the forgetting factor to improve identification accuracy when the error is large.

[0091] This implementation adaptively calibrates the prediction model parameters at the initial moment of switching to local fault-tolerant mode. This ensures that the prediction model used to fill data gaps accurately reflects the actual stress and deformation characteristics of the foundation pit support structure at the time of interruption, avoiding excessive prediction deviations caused by mismatched working conditions in fixed-parameter models. The sliding window recursive identification algorithm has low computational cost, making it suitable for the limited computing resources of edge nodes, and the forgetting factor mechanism ensures the model's ability to quickly track changes in working conditions. The rolling recursive prediction method allows the predicted values ​​to reasonably convey trend and periodic information in the data, rather than simply extending a straight line. Prediction error analysis and feedback updates after communication is restored provide data support for continuous improvement of the calibration strategy, forming a self-optimizing closed loop for prediction parameter management.

[0092] According to one embodiment of the present invention, a deep foundation pit construction accident early warning and emergency decision-making system includes: The knowledge graph construction module is used to construct a knowledge graph of deep foundation pit construction accidents. The knowledge graph defines the relationships between accident types, accident causes, monitoring indicators, emergency response measures, emergency responsible persons, and emergency resources. Edge computing nodes, deployed at the construction site, are used to locally cache the subgraphs of the knowledge graph and the monitoring data sequences within the most recent time window; The data acquisition module is used to collect monitoring data from displacement sensors and stress sensors on the deep foundation pit support structure in real time, as well as video data from the construction site. The mode selection module is used to select between cloud collaboration mode and local fault tolerance mode based on the communication status between the edge computing node and the cloud. The cloud collaboration module is used to upload monitoring data to the cloud and compare it with the knowledge graph in the cloud collaboration mode to generate accident early warning information, and generate and push emergency decision-making plans based on real-time personnel location data and traffic condition data. The local fault tolerance module is used in local fault tolerance mode to enable edge computing nodes to perform local inference based on locally cached data, generate accident warning information and emergency decision-making plans, and notify on-site personnel through local broadcast. The data synchronization module is used to synchronize the data from the local fault-tolerant mode to the cloud after communication is restored.

[0093] The specific implementation architecture of the deep foundation pit construction accident early warning and emergency decision-making system includes the following modules: The knowledge graph construction module is deployed on a cloud server, providing a graphical knowledge editing interface and batch import tools, supporting the extraction of entities and relationships from accident case texts, standard provisions and expert interviews, and constructing and storing the deep foundation pit construction accident knowledge graph in the Neo4j graph database.

[0094] The edge computing nodes utilize industrial-grade edge gateway devices deployed in containerized monitoring rooms at the construction site. Each device features a quad-core ARM processor, 8GB of RAM, and 256GB of solid-state storage, running a Linux operating system and a containerized application environment. The nodes run data receiving services, local cache management services, mode switching arbitration services, a local inference engine, and local broadcast control services. The knowledge graph subgraph caching module subscribes to change events in the cloud-based knowledge graph, synchronizing only the subgraph portions that match the risk characteristics of the local construction area. Monitoring data sequence caching uses a time-series database for time-partitioned storage.

[0095] The data acquisition module consists of front-end sensors and a convergence gateway. Displacement sensors are connected in series with inclinometers, and stress sensors are vibrating wire rebar gauges. Signals are collected via an RS485 bus to the data acquisition unit, which communicates with the edge nodes via a 4G router. Video data is transmitted from Hikvision network cameras to the edge nodes via the ONVIF protocol.

[0096] The mode selection module runs as an independent daemon on the edge node, periodically sending ICMP echo requests and TCP port connectivity tests to the cloud server, judging the communication status by combining packet loss rate and response latency, and executing mode switching logic.

[0097] The cloud-based collaboration module is deployed on a cloud server cluster and includes a knowledge graph query engine, a monitoring data stream processing engine, a personnel location and traffic information interface adapter, and a decision-making solution generation and push service. Under normal communication conditions, the cloud module is responsible for the main computationally intensive and data fusion tasks.

[0098] The local fault-tolerant module runs within an edge node container and includes a lightweight rule inference engine, an ARIMA prediction model library, and a local speech synthesis and broadcasting interface. It is activated when communication is interrupted and takes over the early warning and emergency decision-making functions.

[0099] The data synchronization module adopts the message queue telemetry transmission protocol. After communication is restored, it automatically synchronizes the data change logs, event records and original monitoring data blocks generated during the offline period of the edge node to the cloud database. The synchronization process supports breakpoint resume and integrity verification.

[0100] This implementation integrates knowledge graph construction, edge caching, mode selection, cloud collaboration, local fault tolerance, and data synchronization into a complete system, providing a highly adaptable and reliable early warning and emergency decision-making technology platform for deep foundation pit construction sites. The cloud module handles knowledge management and complex reasoning tasks, while the edge module handles real-time data collection and fault tolerance. The two modules collaborate adaptively through network status, balancing intelligence and operational reliability. The modular design allows for flexible deployment based on specific project scales; for example, small foundation pits can utilize only edge fault tolerance mode with lightweight cloud services, while large-scale key projects can be configured with a full-featured cluster. The data synchronization mechanism ensures the traceability of offline events and the continuous improvement of the knowledge base, laying the architectural foundation for long-term system operation and iterative optimization.

[0101] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. It can be applied to various fields suitable for the embodiments of the present invention. Other modifications can be readily implemented by those skilled in the art. Therefore, without departing from the general concept defined by the claims and their equivalents, the embodiments of the present invention are not limited to the specific details.

Claims

1. A method for early warning and emergency decision-making in deep foundation pit construction accidents, characterized in that, Includes the following steps: A knowledge graph of deep foundation pit construction accidents is constructed, which defines the relationships between accident types, accident causes, monitoring indicators, emergency response measures, emergency responsible persons, and emergency resources. Edge computing nodes are deployed at the construction site, and the edge computing nodes locally cache the subgraphs of the knowledge graph and the monitoring data sequences within the most recent time window; Real-time acquisition of monitoring data from displacement and stress sensors on the deep foundation pit support structure, as well as video data from the construction site; Based on the communication status between the edge computing node and the cloud, select to execute either cloud collaboration mode or local fault-tolerant mode; In the cloud-based collaborative mode, monitoring data is uploaded to the cloud and compared with the knowledge graph to generate accident early warning information. Emergency decision-making plans are generated and pushed based on real-time personnel location data and traffic condition data. In local fault-tolerant mode, edge computing nodes perform local inference based on locally cached data to generate accident warning information and emergency decision-making plans, and notify on-site personnel through local broadcast. Once communication is restored, the data from the local fault-tolerant mode will be synchronized to the cloud.

2. The method for early warning and emergency decision-making in deep foundation pit construction as described in claim 1, characterized in that, The emergency decision-making solutions generated under the cloud-based collaborative model include: Starting with the type of accident, the algorithm retrieves related emergency response measures, responsible persons, and emergency resources from the cloud-based knowledge graph using a graph traversal algorithm. Based on real-time on-site personnel location data and traffic condition data, the retrieved emergency response measures and emergency resources are dynamically sorted, and the optimal dispatch path for emergency resources is planned to generate an emergency decision-making plan that includes step-by-step response instructions, personnel assembly points, and resource arrival times.

3. The method for early warning and emergency decision-making in deep foundation pit construction accidents as described in claim 1, characterized in that, The local inference under the local fault-tolerant mode includes: When a communication interruption or sensor data stream interruption is detected for more than a preset time threshold, the edge computing node uses a time series prediction model to extrapolate and predict the predicted values ​​of key monitoring indicators during the data interruption period, and then splices the predicted values ​​after the last real monitoring data before the interruption to form a continuous data sequence. The continuous data sequence is compared in real time with the monitoring indicators corresponding to the accident causes in the local cached knowledge graph subgraph. The comparison includes comparing the data with a preset safety threshold and calculating the waveform similarity between the local change trend of the sequence and the preset accident evolution path trend in the subgraph. When the data exceeds the safety threshold and the waveform similarity exceeds the preset matching degree, an accident warning message is generated; Starting with the accident type, the system retrieves associated emergency response measures, responsible persons, and emergency resources from the locally cached knowledge graph subgraph, and generates a static emergency decision-making plan based on the latest personnel locations and preset default traffic conditions.

4. The method for early warning and emergency decision-making in deep foundation pit construction accidents as described in claim 1, characterized in that, It also includes the steps for dynamically updating the knowledge graph: Once an accident warning is generated, real-time monitoring data sequences are continuously collected from the warning time to the end of the preset time window, and it is recorded whether an accident actually occurs within that time window. Calculate the waveform similarity between the real-time monitoring data sequence and the accident evolution path that triggered the early warning information; If an accident actually occurs and the waveform similarity is below the first threshold, the real-time monitoring data sequence will be added to the knowledge graph as a new accident evolution path, and the actual emergency response measures will be associated with this new accident evolution path. If no accident occurs and the waveform similarity is below the second threshold, the real-time monitoring data sequence will be marked as a false alarm mode to adjust the matching conditions of the preset safety threshold or accident evolution path.

5. The method for early warning and emergency decision-making in deep foundation pit construction accidents as described in claim 1, characterized in that, It also includes steps for dynamically adjusting emergency plans: After the emergency decision-making plan is pushed out, the system acquires the mobile terminal location data of emergency responders, the GPS trajectory data of emergency vehicles, and the real-time monitoring data continuously collected by on-site sensors. The deviation between the real-time GPS trajectory of the emergency vehicle and the planned optimal dispatch route is calculated. When the deviation exceeds the preset threshold or congestion is detected on the road ahead, the optimal route from the current location to the accident site is replanned and the resource arrival time is updated. At the same time, the monitoring data continuously collected by the on-site sensors is compared with the monitoring data when the accident warning information is generated. If the rate of deterioration of the monitoring data exceeds the preset rate threshold, the accident risk level is determined to have increased. Supplementary emergency response measures and additional emergency resources are automatically retrieved from the knowledge graph, and a reinforcement plan is generated and pushed out. The updated routes, arrival times, and reinforcement plans will be synchronized to the relevant terminals in real time.

6. The method for early warning and emergency decision-making in deep foundation pit construction as described in claim 1, characterized in that, It also includes a multi-source heterogeneous data preprocessing step: Before comparing the monitoring data with the monitoring indicators corresponding to the accident causes in the knowledge graph, the following preprocessing operations are performed: Timestamp interpolation and alignment of monitoring data from displacement and stress sensors are performed using a unified time reference; A lightweight target detection model is used to extract structural deformation features from keyframes of video data, which are then converted into structured deformation quantification indicators and spatially correlated with displacement sensor data. Monitoring data from different sensors are converted into dimensionless risk contribution values ​​according to their respective preset standardized formulas. The standardized formulas are dynamically determined based on the sensor range and the historical threshold range corresponding to the accident cause. The standardized multi-source data are fused using Kalman filtering, and the resulting integrated monitoring data sequence is used for comparison.

7. The method for early warning and emergency decision-making in deep foundation pit construction accidents as described in claim 1, characterized in that, It also includes steps for dynamic conflict detection and reallocation of emergency resources: During the execution of the emergency decision-making plan, real-time status data of each emergency resource is continuously collected. The status data includes the current location of the resource, the currently occupied task identifier, the estimated release time, the availability indicator, and the resource type label. When a new accident warning or a change in the resource requirements of an existing task is received, a resource conflict detection is triggered: with a preset time window length as the period, the resource requirements of all unfinished tasks are compared with the occupied time of the currently allocated resources. If the same resource is allocated to two or more different tasks in the same time period, it is marked as a resource conflict event. For resources that are in conflict, extract the urgency score and resource suitability score of all associated tasks. The urgency score is dynamically calculated based on the accident risk level, the number of casualties, and the accident spread speed. The resource suitability score is based on the degree of matching between the resource type and the task requirements, as well as the distance between the current location of the resource and the task location. Using the Hungarian algorithm or genetic algorithm based on conflict resolution, a conflict-free resource reallocation scheme is generated with the goal of maximizing the weighted fitness of the total urgency of all tasks, under the premise that each task is allocated at least one available resource and the total resource limit is not exceeded. The resource allocation instructions affected by the reallocation plan will be pushed to the terminals of relevant emergency response personnel in real time, and the resource arrival time and disposal instructions in the original emergency decision-making plan will be updated simultaneously.

8. The method for early warning and emergency decision-making in deep foundation pit construction accidents as described in claim 1 or 3, characterized in that, It also includes optimization steps for static decision-making schemes based on personnel movement trajectory prediction in a local fault-tolerant mode: In local fault-tolerant mode, the edge computing node locally stores the historical location trajectory sequence of each emergency responder within a preset time period. The location trajectory sequence consists of the location data obtained during the last network communication before the interruption and the relative position increment obtained by local short-range wireless ranging after the interruption. When a static emergency decision-making scheme needs to be generated, the edge computing node takes each person's historical location trajectory sequence as input, uses Kalman filtering or particle filtering algorithms to predict the probability location distribution at the current moment, and outputs the estimated location with the highest confidence. For personnel whose positions cannot be predicted through historical trajectories or whose confidence level is below a preset threshold, the edge computing node sets their estimated location to a preset key point corresponding to their position based on the personnel's job responsibility label and the functional zoning map of the deep foundation pit construction area. The estimated location of each person is used as the latest personnel location in the local cache for personnel matching, sorting and assembly point setting in emergency response measures, generating an optimized static emergency decision-making plan; At the same time, an instruction will be added to the emergency information broadcast locally, stating that the personnel location is a predicted value and requesting all personnel to actively confirm their location via walkie-talkie or manually.

9. The method for early warning and emergency decision-making in deep foundation pit construction as described in claim 3, characterized in that, It also includes an adaptive calibration step for the prediction model parameters in local fault-tolerant mode: After switching to local fault-tolerant mode, the edge computing node acquires the real monitoring data sequence within the last preset time window before the data interruption, which is recorded as the calibration window data. Edge computing nodes employ a sliding window recursive identification method, using calibration window data as input, to estimate the model parameters of the time series prediction model online. The model parameters include autoregressive coefficients, moving average coefficients, and trend term coefficients. The original fixed parameters of the prediction model are replaced by the model parameters obtained from online estimation to form an adaptive prediction model for operating conditions. The aforementioned adaptive prediction model is used to extrapolate and predict the key monitoring indicators during data interruption periods. During the prediction process, after extrapolating a time step, if the prediction time point corresponding to that step does not exceed the upper limit of the interruption duration, the prediction value of that step is used as the input of the next time step and fed back to the prediction model to achieve rolling recursive prediction. Once communication is restored, the actual monitoring data during the interruption period is compared with the predicted values ​​to calculate the prediction error sequence. This error sequence is then used to update the selection strategy for calibration window data or adjust the forgetting factor of recursive identification.

10. A deep foundation pit construction accident early warning and emergency decision-making system, characterized in that, include: The knowledge graph construction module is used to construct a knowledge graph of deep foundation pit construction accidents. The knowledge graph defines the relationships between accident types, accident causes, monitoring indicators, emergency response measures, emergency responsible persons, and emergency resources. Edge computing nodes, deployed at the construction site, are used to locally cache the subgraphs of the knowledge graph and the monitoring data sequences within the most recent time window; The data acquisition module is used to collect real-time monitoring data from displacement and stress sensors on the deep foundation pit support structure, as well as video data from the construction site. The mode selection module is used to select between cloud collaboration mode or local fault-tolerant mode based on the communication status between the edge computing node and the cloud. The cloud collaboration module is used to upload monitoring data to the cloud and compare it with the knowledge graph in the cloud collaboration mode to generate accident early warning information, and generate and push emergency decision-making plans based on real-time personnel location data and traffic condition data. The local fault tolerance module is used in local fault tolerance mode to perform local inference based on locally cached data by edge computing nodes, generate accident early warning information and emergency decision-making plans, and notify on-site personnel through local broadcast. The data synchronization module is used to synchronize the data from the local fault-tolerant mode to the cloud after communication is restored.