Deep foundation pit low delay risk early warning method

Abstract

This invention discloses a low-latency risk early warning method for deep foundation pits, belonging to the field of construction safety monitoring technology. The method includes: collecting multi-source monitoring data through an on-site sensor network and transmitting it in real-time to an edge computing node for preprocessing to generate a standard time-series data stream; extracting spatiotemporal feature vectors from the data stream and inputting them into a pre-trained risk early warning model that incorporates physical constraints to calculate the overall risk probability value of the foundation pit in real-time; when the risk value exceeds a threshold, the edge node immediately generates and sends an early warning message to a remote monitoring center. This method solves the problems of long early warning delays and insufficient ability to identify complex risk patterns caused by remote data transmission and cloud processing in existing deep foundation pit safety monitoring. By completing core analysis and calculation at the data source and using a physical information neural network model for intelligent judgment, it significantly reduces early warning delays and improves the accuracy and reliability of early warnings for complex risks.

Description

Technical Field

[0001] This invention relates to the field of construction safety monitoring technology, and in particular to a low-latency risk early warning method for deep foundation pits. Background Technology

[0002] Deep foundation pit engineering is a common and high-risk construction phase in modern urban development. Its construction process disturbs the surrounding soil, causing a series of complex responses, including stress and deformation of the support structure, changes in groundwater levels, and settlement of adjacent buildings. If these changes exceed safe limits and are not addressed promptly, they can lead to serious safety accidents such as support structure instability and foundation pit collapse, resulting in significant loss of life and property. Therefore, real-time and effective safety monitoring and risk warning during deep foundation pit construction are indispensable technical means to ensure project safety.

[0003] Currently, deep foundation pit safety monitoring mainly relies on deploying various sensors on-site to collect key parameters such as displacement, stress, and water level. Traditional monitoring and early warning models typically follow a process of data acquisition, remote transmission, central processing, and manual assessment. Raw data collected by sensors is transmitted via wired or wireless means to a remote monitoring center server or cloud platform, where it is stored, processed, and analyzed. Finally, professionals assess the risk and issue warnings based on experience or simple threshold judgments. However, with the increasing scale of foundation pits, the number of monitoring points, and the growing demand for timely warnings, this traditional model has gradually revealed several prominent problems. First, the warning delay is significant. Massive amounts of raw monitoring data need to be transmitted over long distances to the center. The transmission process itself is affected by network conditions, resulting in delays. Furthermore, the central server needs to process data from multiple projects, leading to competition for computing resources. This results in a long chain from data acquisition to the generation of warning results, making it difficult to meet the requirement of a second-level response to sudden emergencies. Second, the risk identification capability is limited. Existing methods often rely on setting fixed thresholds for single monitoring indicators to trigger over-limit alarms, or on engineers making judgments based on experience and multiple indicators. The former method fails to identify the gradual and systemic risks predicted by the complex spatiotemporal correlations between multiple parameters; the latter relies heavily on personal experience and struggles to achieve consistent, objective, and real-time judgments around the clock. For complex risk patterns caused by the coupling of multiple factors, traditional methods are prone to false negatives or false negatives. Furthermore, the models lack adaptability. Some methods that attempt to introduce machine learning models typically deploy them in the cloud. These models are often static models trained on historical general data, making it difficult to adapt to the unique geological conditions, support forms, and dynamically changing construction conditions of specific foundation pits. The model's generalization ability and long-term prediction accuracy in new scenarios are difficult to guarantee. In addition, complete reliance on cloud processing paralyzes the system during network outages, resulting in reliability issues.

[0004] Therefore, the industry urgently needs to design an intelligent risk early warning method for deep foundation pits that can achieve lower latency, higher accuracy, stronger adaptability, and does not rely on a continuous and stable remote network. Summary of the Invention

[0005] This invention overcomes the problems of long early warning delays and insufficient ability to identify complex risk patterns caused by remote data transmission and cloud processing in existing deep foundation pit safety monitoring technologies. It provides a low-latency risk early warning method for deep foundation pits. By completing core analysis and calculation at the data source and using a physical information neural network model for intelligent judgment, it achieves a significant reduction in early warning delay and improves the accuracy and reliability of early warnings for complex risks.

[0006] To achieve the above objectives, the present invention adopts the following solution: A low-latency risk early warning method for deep foundation pits includes the following steps: S1: By deploying a sensor network at the deep foundation pit engineering site, data such as displacement of the support structure, axial force of the support, horizontal displacement of the deep soil, groundwater level, and settlement of surrounding buildings are collected to form the original multi-source monitoring dataset. S2: Transmit the original multi-source monitoring dataset to the edge computing node in real time. The received data is preprocessed in real time at the edge computing node. The preprocessing includes noise reduction, timestamp alignment of multi-source data, and unification of data format into time-series data with a set sampling interval, generating a preprocessed standard time-series data stream. S3: In the edge computing node, extract the spatiotemporal feature vector for risk assessment from the standard time-series data stream. The spatiotemporal feature vector includes the statistical characteristics of the same monitoring project within a continuously set time window, the difference characteristics between monitoring data at different spatial locations under the same time section, and the rate of change characteristics. S4: Input the spatiotemporal feature vector into the pre-trained risk warning model. The risk warning model is a physical information neural network model that integrates the differential equation of soil equilibrium in the foundation pit as a physical constraint. The risk warning model calculates the input spatiotemporal feature vector and outputs the risk probability value that represents the overall risk state of the foundation pit. S5: When the risk probability value exceeds the preset warning threshold, the edge computing node immediately generates a warning message containing the risk probability value, the identifier of the main anomaly monitoring item, and the time of the anomaly occurrence, and sends the warning message to the remote monitoring center.

[0007] Preferably, step S1 specifically includes: The sensor network deployed at the deep foundation pit project site is a multi-parameter sensing layer. The multi-parameter sensing layer includes fully automatic intelligent total stations deployed at a set density on the top of the foundation pit support piles and foundation pit support walls, vibrating wire stress sensors installed at key sections of each horizontal support, fixed inclinometer arrays buried at a set depth along the perimeter of the foundation pit, anti-clogging pore water pressure gauges distributed inside and outside the foundation pit, and static leveling instrument systems arranged at the foundations of surrounding buildings that need to be protected. The deployment points and depths of the fixed inclinometer array are designed based on the three-dimensional geological survey data and support structure before the excavation of the foundation pit. The potential maximum displacement area and sliding surface depth are determined through finite element numerical simulation analysis. Vibrating wire stress sensors are deployed in groups at the mid-span of each support and near the two supports to monitor the bending moment distribution of the support. All types of sensors in the sensor network are deployed using a dual backup strategy. Two sensors of the same model are installed at each key monitoring point. The measurement data from the two sensors are used for cross-validation and data quality assessment.

[0008] Preferably, step S2 specifically includes: The edge computing node is an industrial-grade embedded computing gateway deployed at the construction site. The industrial-grade embedded computing gateway establishes data connections with wired and wireless sensors in the sensor network through a fieldbus protocol interface and a wireless data receiving module, respectively. In the real-time preprocessing operation, the Kalman filter algorithm is used to reduce noise in the data. The Kalman filter algorithm adopts an adaptive process noise covariance matrix that includes prior knowledge of the foundation pit deformation. This prior knowledge is based on historical engineering data with similar geological conditions and support forms to the current foundation pit. When aligning timestamps of multi-source data based on a unified clock source, a time synchronization signal is periodically broadcast to all sensors in the network through an industrial-grade embedded computing gateway to complete hard synchronization. For minor time deviations caused by sensor sampling and transmission delays, a soft correction is performed using an interpolation algorithm based on the arrival sequence of data packets. The method to unify the data format into time-series data with a set sampling interval is as follows: after completing the timestamp alignment, the data stream of each monitoring item is resampled and aligned at fixed time intervals based on the system clock of the edge computing node, thereby generating a preprocessed standard time-series data stream that is strictly synchronized in the time dimension.

[0009] As a preferred option, a local lightweight digital twin model is established within the industrial-grade embedded computing gateway. The local lightweight digital twin model is a simplified version of the physical information neural network model, which is driven online in real time by receiving preprocessed standard time-series data streams. After generating a standard time-series data stream, the industrial-grade embedded computing gateway immediately inputs the data stream into the local lightweight digital twin model for forward computation and performs a rolling comparison between the predicted sequence output by the model and the monitoring data of the next actual sampling period. When the average error of the rolling comparison within a set number of consecutive times exceeds the preset model tolerance threshold, the industrial-grade embedded computing gateway determines that the local lightweight digital twin model is inaccurate and triggers a model update request. The model update request is encapsulated and sent to the remote monitoring center through the wireless communication network. After receiving a model update request, the remote monitoring center retrains and optimizes the physical information neural network model based on the complete historical database, and sends the updated model parameter set to the industrial-grade embedded computing gateway to replace its internal local lightweight digital twin model.

[0010] Preferably, step S3, which involves extracting the spatiotemporal feature vector for risk assessment from the standard time-series data stream, specifically includes: The time series data of each monitoring item are extracted with a set base time window length. The arithmetic mean, variance, and linear trend slope of the data within the window are calculated as statistical features. The difference in horizontal displacement of deep soil between adjacent inclinometer monitoring points deployed along the perimeter of the foundation pit at the same time section is calculated, as well as the difference between the readings of sensors deployed in groups at the support axial force monitoring section, as difference features. The rate of change of each monitoring data is calculated with the previous shortened time window as the rate of change feature. All the above-calculated statistical features, difference features, and rate of change features are spliced ​​together in a preset order to form a high-dimensional spatiotemporal feature vector.

[0011] Preferably, after constructing a high-dimensional spatiotemporal feature vector, it is input into a pre-trained feature dimensionality reduction module, which is a principal component analysis model. The principal component analysis model is used to perform a linear transformation on the input high-dimensional spatiotemporal feature vector, retaining the top N principal components whose cumulative contribution rate exceeds a set threshold, and outputting a low-dimensional principal component feature vector. The low-dimensional principal component feature vector is used as the direct input to the risk warning model in step S4.

[0012] Preferably, while generating low-dimensional principal component feature vectors, a distance-based local anomaly calculation process is performed in parallel. The obtained high-dimensional spatiotemporal feature vectors are input into a computational unit based on the local anomaly factor algorithm. This computational unit calculates a local anomaly score for each time step feature vector by comparing the local reachability density of each feature vector with its K nearest neighbor feature vectors in the feature space. The low-dimensional principal component feature vectors are then concatenated with the local anomaly scores to form the final input feature vector of the risk warning model in step S4.

[0013] Preferably, in step S4, the pre-trained risk warning model adopts a hybrid physical information neural network architecture that integrates transfer learning and online incremental learning mechanisms. The initial training of the hybrid physical information neural network architecture is based on a historical monitoring dataset of multiple engineering projects containing various geological conditions and foundation pit types. Through the transfer learning mechanism, the general feature knowledge learned from one or more source engineering projects is transferred to the risk warning model for the current target foundation pit. The physical constraints of the physical information neural network model are formed by adding a physical loss term to the loss function of the hybrid physical information neural network. This physical loss term is calculated based on the static equilibrium differential equation of the foundation pit soil. The static equilibrium differential equation of the foundation pit soil is: Where σ is the stress tensor predicted by the model, Here, b is the divergence operator, and b is the force vector; the physical loss term is... Specifically, it is the mean of the squared norms of the equation residuals calculated at a series of points within the computational domain, i.e.: in For the neural network at the placement point x i The predicted stress at the location, N p For the total number of configuration points, The norm is used; after deploying the risk warning model on edge computing nodes, an online incremental learning mechanism is activated. It uses the newly generated standard time-series data stream of the current foundation pit as incremental data and adopts a small-batch gradient descent algorithm to continuously fine-tune the model parameters. At the same time, to prevent catastrophic forgetting, an elastic weight consolidation algorithm is introduced to constrain the changes of important parameters.

[0014] As a preferred approach, the hybrid physical information neural network architecture running in the edge computing nodes implements its transfer learning mechanism through a pre-built source domain model pool. The source domain model pool stores multiple pre-trained physical information neural network sub-models under different typical engineering conditions. During initial deployment, based on the current three-dimensional geological survey data of the foundation pit and the design parameters of the support structure, at least two sub-models with the highest matching degree are selected from the source domain model pool, and some of their network layer parameters are loaded into the current target model as initialization values. The implementation of the online incremental learning mechanism includes an uncertainty-based data filtering step. Using the Monte Carlo Dropout method built into the physical information neural network model, the newly generated standard time-series data stream is propagated forward multiple times to calculate the variance of the model's predicted output. Only data samples with prediction uncertainty higher than a set threshold are filtered out as valuable incremental data for fine-tuning the model parameters.

[0015] Preferably, step S5 specifically includes: The warning information is encapsulated into a structured warning data packet, which includes three parts: a header, a body, and a footer. The header includes the data format version number and unique sequence number of the warning information. The body includes the risk probability value, the identifier of the main anomaly monitoring item that triggered the warning and its corresponding real-time monitoring data, the time of the anomaly, and the risk level code mapped according to the preset interval where the risk probability value is located. The footer includes a cyclic redundancy check code calculated based on the content of the body. Edge computing nodes synchronously send warning data packets to the remote monitoring center via at least two independent wireless communication links. These two independent wireless communication links include one link based on a cellular mobile network and one link based on a local wireless LAN. After receiving the warning data packet from either link, the remote monitoring center first verifies its cyclic redundancy check code. If the verification passes, it parses and stores the warning information and simultaneously returns an acknowledgment signal containing the corresponding sequence number to the edge computing node. If the edge computing node does not receive an acknowledgment signal from the remote monitoring center within a set time after sending the warning data packet, it retransmits the warning data packet via the other wireless communication link.

[0016] The present invention has at least the following beneficial effects: (1) By deploying data preprocessing, feature extraction and core risk judgment model at the edge computing node of the construction site, most of the computing tasks are completed near the data collection source, avoiding the network delay and queuing time caused by the remote transmission of massive raw data to the cloud, so that the entire process from data perception to early warning generation is extremely short, which can meet the near real-time response requirements for sudden risks in the foundation pit; (2) By systematically extracting spatiotemporal feature vectors containing statistical features, spatial difference features and rate of change features from multi-source monitoring data streams, it is possible to comprehensively characterize the overall behavior pattern of the foundation pit system in terms of time evolution and spatial correlation, overcoming the one-sidedness of traditional single threshold method or manual experience judgment, so that risk warning is based on multi-dimensional and deep-level data correlation analysis, significantly improving the identification accuracy and early detection capability of complex and hidden risk patterns; (3) Using a neural network that integrates physical constraints as the early warning model ensures that data-driven prediction conforms to basic mechanical principles, enhances the generalization of the model and the reliability under extrapolation conditions, and at the same time, the model set It has become a transfer learning and online incremental learning with anti-forgetting mechanism, which can quickly obtain prior knowledge from historical big data and can adaptively fine-tune with the progress of the current project, so that the early warning capability can remain optimal for a long time and adapt to the dynamic changes of the project; (4) From the dual backup and cross-validation of sensors at the hardware level to the dual-link redundant transmission, verification confirmation and retransmission mechanism at the data transmission stage, a multi-level full-chain reliability guarantee has been constructed. At the same time, the intelligent noise reduction processing, high-precision time synchronization and online self-checking and update process of the model at the edge side have jointly ensured the accuracy and synchronization of the input information and the continuous effectiveness of the computing engine, which greatly improves the robustness and credibility of the entire early warning system in the harsh construction site environment; (5) By introducing dimensionality reduction methods such as principal component analysis after feature extraction, the high-dimensional feature vector is purified and compressed. While retaining the core risk information, the input dimension and computational burden of the subsequent neural network model are greatly reduced, so that the complex intelligent early warning algorithm can run efficiently and stably in the resource-constrained industrial edge computing gateway, providing a feasible technical path for the engineering implementation of low-latency early warning. Attached Figure Description

[0017] Figure 1 The flowchart illustrates the principle of the low-latency risk early warning method for deep foundation pits provided by this invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings, so that those skilled in the art can implement it based on the description.

[0019] like Figure 1 As shown, the low-latency risk early warning method for deep foundation pits provided by the present invention includes the following steps: S1: By deploying a sensor network at the deep foundation pit engineering site, data such as displacement of the support structure, axial force of the support, horizontal displacement of the deep soil, groundwater level, and settlement of surrounding buildings are collected to form the original multi-source monitoring dataset. S2: Transmit the original multi-source monitoring dataset to the edge computing node in real time. The received data is preprocessed in real time at the edge computing node. The preprocessing includes noise reduction, timestamp alignment of multi-source data, and unification of data format into time-series data with a set sampling interval, generating a preprocessed standard time-series data stream. S3: In the edge computing node, extract the spatiotemporal feature vector for risk assessment from the standard time-series data stream. The spatiotemporal feature vector includes the statistical characteristics of the same monitoring project within a continuously set time window, the difference characteristics between monitoring data at different spatial locations under the same time section, and the rate of change characteristics. S4: Input the spatiotemporal feature vector into the pre-trained risk warning model. The risk warning model is a physical information neural network model that integrates the differential equation of soil equilibrium in the foundation pit as a physical constraint. The risk warning model calculates the input spatiotemporal feature vector and outputs the risk probability value that represents the overall risk state of the foundation pit. S5: When the risk probability value exceeds the preset warning threshold, the edge computing node immediately generates a warning message containing the risk probability value, the identifier of the main anomaly monitoring item, and the time of the anomaly occurrence, and sends the warning message to the remote monitoring center.

[0020] The low-latency risk early warning method for deep foundation pits relies on a meticulously deployed sensor network at the engineering site. This sensor network is a data acquisition layer integrating various types of monitoring equipment, used to comprehensively capture changes in key physical parameters of the deep foundation pit during construction. These parameters include, but are not limited to, the horizontal and vertical displacement of the support structure, the axial force borne by the internal support system, the horizontal movement of the surrounding soil at different depths, fluctuations in groundwater levels, and the settlement of surrounding protected buildings. This data collectively constitutes the original multi-source monitoring dataset, its multi-source nature reflected in the differences in monitoring objects, physical quantities, and spatial locations, providing rich information input for subsequent comprehensive risk analysis. The selection and deployment strategy of sensors can be optimized according to the specific geological conditions, support design, and surrounding environment of the project. For example, total stations or GNSS receivers can be used for displacement monitoring, vibrating wire or fiber optic grating sensors can be used for support internal forces, and fixed or portable inclinometers can be used for soil displacement. The specific deployment density and location should effectively reflect the overall deformation and internal force distribution trend of the structure.

[0021] The collected raw data is transmitted in real time to edge computing nodes deployed at the construction site via wired or wireless means. These edge computing nodes are industrial-grade embedded devices or gateways with strong computing capabilities. By setting up these nodes, data processing and analysis tasks can be moved forward to the data generation source, significantly reducing the time latency caused by data transmission, processing, and result return to remote servers. Within the edge computing nodes, the system performs real-time preprocessing on the received streaming data. The first step in preprocessing is data noise reduction. Due to the complex environment of the construction site, sensor signals are susceptible to electromagnetic interference, mechanical vibration, and other factors. Therefore, digital filtering algorithms (such as Kalman filtering and low-pass filtering) are needed to filter out high-frequency noise and retain useful signals reflecting the actual structural behavior. The second step is timestamp alignment of multi-source data. Since different sensors may have independent clocks or different sampling and transmission delays, the data may be out of sync. This requires hardware synchronization signal broadcasting or software interpolation correction algorithms to unify all data onto the same high-precision time reference. The third step is to resample and format the aligned data into a standard time-series data stream with a fixed sampling interval (e.g., once per second or once per minute) to ensure that the data points of each monitoring item are strictly corresponding sequences on the time axis, so as to facilitate subsequent feature extraction and model calculation.

[0022] The system extracts spatiotemporal feature vectors characterizing the risk state of the foundation pit from the preprocessed standard time-series data stream, transforming the raw data into more advanced and discriminative information. These spatiotemporal features consider not only the trend of data changes at individual monitoring points over time (time-series characteristics) but also the correlation between monitoring data from different spatial locations at the same time (spatial characteristics). Specifically, the system processes the time-series data of each monitoring item using a sliding time window, calculating statistics such as the mean (reflecting the overall level), variance (reflecting the degree of fluctuation), and linear trend slope (reflecting the rate of change) within the window. Simultaneously, the system calculates characteristics under specific spatial relationships. For example, it calculates the difference in soil displacement measured by inclinometers on adjacent sides of the foundation pit to determine whether the deformation is coordinated or if there are local abrupt changes; it calculates the difference in axial force at different sections of the same support to assess the uniformity of stress on the support. Furthermore, it calculates the rate of change of each monitoring data point over a short period, as a significant increase in instantaneous rate is often a precursor to a dangerous situation. All these calculated statistical features, spatial difference features, and rate of change features are combined in a predefined order into a high-dimensional feature vector, which comprehensively describes the overall behavior pattern of the foundation pit system in the spatiotemporal dimension at the current moment.

[0023] This high-dimensional spatiotemporal feature vector is input into a pre-trained risk warning model, which is a neural network model incorporating physical constraints—a physical information neural network. Unlike traditional black-box models that rely purely on data-driven approaches, this model, during training, not only minimizes prediction errors but also introduces physical equations (such as static equilibrium equations) reflecting the mechanical equilibrium of the foundation pit soil as constraints through a loss function. This means that the model's predictions not only need to fit historical data but must also conform to basic physical principles to a certain extent, thereby improving the model's generalization ability and prediction reliability in situations where data is scarce. After receiving the feature vector, the model performs complex nonlinear calculations internally and ultimately outputs a quantified risk probability value, which ranges from 0 to 1 or is divided into several levels, intuitively representing the likelihood that the foundation pit is currently in a dangerous state.

[0024] The system makes early warning decisions based on the risk probability values ​​output by the model. One or more early warning thresholds are set in the edge computing nodes. These thresholds can be set according to the safety level, stage, and experience of the project; for example, a yellow warning threshold of 0.6 and a red warning threshold of 0.8 can be set. When the real-time calculated risk probability value exceeds the preset warning threshold, the edge computing node immediately triggers the early warning process. It generates a structured early warning message, which includes at least the risk probability value, the identified main anomaly monitoring item (such as "abnormal horizontal displacement rate of deep soil"), and the precise timestamp of the anomaly. Subsequently, this early warning message is pushed to the platform of the remote monitoring center or the terminal devices of relevant personnel in near real-time via a reliable communication link (such as 4G / 5G, LoRa, Wi-Fi, or a combination thereof), thus completing the closed loop from data acquisition to risk alarm. The entire process completes the core computation at the edge, minimizing the latency caused by remote data transmission and cloud processing, achieving truly low-latency early warning.

[0025] Compared with existing technologies, this solution achieves a substantial improvement in the timeliness of early warning. Traditional solutions typically rely on uploading massive amounts of raw data to cloud servers for processing, resulting in lengthy network transmission and centralized computation times. This method, however, completes data preprocessing, feature extraction, and model inference at the edge of the construction site, reducing the time delay for early warning generation to the second or even sub-second level, thus saving valuable time for emergency response. Regarding the accuracy of early warning, by fusing multi-source heterogeneous monitoring data and extracting spatiotemporal correlation features, it can more comprehensively perceive the state of the foundation pit system compared to traditional methods that rely solely on a single type of data or simple thresholds, reducing false alarms and missed alarms. Simultaneously, the introduction of a physical information neural network model embeds domain knowledge into the data-driven model, enhancing the model's generalization ability to complex working conditions and its robustness to potential local missing data or noise in the monitoring data, making risk assessment more scientific and reliable. This method offers enhanced adaptability and maintainability. The deployment of edge computing nodes reduces its dependence on the network, making it suitable for application at construction sites with inadequate infrastructure. Furthermore, the model supports online updates and incremental learning, enabling adaptive optimization as new data accumulates and the model maintains its early warning performance over time, overcoming the problem of traditional static models becoming ineffective over time. Overall, this method constructs a closed-loop monitoring system for deep foundation pit safety, integrating real-time perception, intelligent analysis, and rapid early warning, significantly improving the proactiveness and intelligence of risk management in deep foundation pit engineering.

[0026] In another technical solution, step S1 specifically includes: The sensor network deployed at the deep foundation pit project site is a multi-parameter sensing layer. The multi-parameter sensing layer includes fully automatic intelligent total stations deployed at a set density on the top of the foundation pit support piles and foundation pit support walls, vibrating wire stress sensors installed at key sections of each horizontal support, fixed inclinometer arrays buried at a set depth along the perimeter of the foundation pit, anti-clogging pore water pressure gauges distributed inside and outside the foundation pit, and static leveling instrument systems arranged at the foundations of surrounding buildings that need to be protected. The deployment points and depths of the fixed inclinometer array are designed based on the three-dimensional geological survey data and support structure before the excavation of the foundation pit. The potential maximum displacement area and sliding surface depth are determined through finite element numerical simulation analysis. Vibrating wire stress sensors are deployed in groups at the mid-span of each support and near the two supports to monitor the bending moment distribution of the support. All types of sensors in the sensor network are deployed using a dual backup strategy. Two sensors of the same model are installed at each key monitoring point. The measurement data from the two sensors are used for cross-validation and data quality assessment.

[0027] The sensor network is a multi-parameter sensing layer integrated from various specialized sensors. This multi-parameter sensing layer is not a simple stacking of sensors, but a systematic configuration based on the multi-dimensional needs of risk monitoring in deep foundation pit engineering. Among them, fully automatic intelligent total stations, used to monitor the horizontal and vertical displacement of the top of the support structure, are deployed at a certain density on the top of the support piles or continuous walls of the foundation pit. This density can be adjusted according to the perimeter and shape complexity of the foundation pit; for example, a monitoring point is set every 10 to 20 meters to ensure continuous capture of the overall deformation profile of the support structure. Vibrating wire stress sensors installed at key sections of each horizontal support have a steel wire whose vibration frequency changes with the applied stress. By measuring the frequency, the axial force of the support can be calculated, thus directly reflecting the stress state of the support system. To monitor the stability of deep soil caused by foundation pit excavation, a fixed array of inclinometers is buried at a certain depth along the perimeter of the foundation pit. By measuring the tilt angle of the probe at different depths within the inclinometer tube, and performing integration calculations, the horizontal displacement curve of the soil along the depth can be obtained, which is a key tool for identifying potential sliding surfaces. The anti-clogging pore water pressure gauges, distributed inside and outside the foundation pit, employ a special filter design to prevent mud blockage and can continuously and accurately measure changes in pore water pressure in the soil. This is crucial for assessing the effectiveness of dewatering and providing early warning of piping or sudden surge risks. The hydrostatic leveling system, deployed at the foundations of surrounding buildings requiring protection, acquires minute settlement data of buildings with extremely high precision by measuring changes in the height of liquid levels in a series of interconnected level bowls. These sensors collectively constitute a three-dimensional physical quantity sensing system from the foundation pit itself to the surrounding environment, providing a comprehensive and accurate data source for risk early warning.

[0028] The deployment of fixed inclinometer arrays is not uniform or determined empirically, but rather based on detailed three-dimensional geological survey data and support structure design drawings before excavation. Pre-analysis is performed using professional geotechnical finite element numerical simulation software. The simulation process considers different excavation conditions, predicts potential displacement and stress fields in the soil, and identifies the potential maximum displacement area and the most dangerous sliding surface depth range. For example, the simulation may show concentrated displacement at the corner of the excavation pit or at the interface between soft and hard strata. The placement and depth of the inclinometer tubes (e.g., the depth may need to penetrate the potential sliding surface and enter the stable soil layer to a certain depth, with a total depth ranging from 20 to 50 meters or even deeper, depending on geological conditions) are based primarily on this analysis, allowing the monitoring network to directly locate the most likely areas of danger, greatly improving monitoring efficiency and effectiveness. Similarly, for monitoring the axial force of the supports, the deployment of vibrating wire stress sensors on each support is also based on precise direction, not only installed at the mid-span of the support but also grouped at key sections near the end supports. This arrangement allows the system to obtain the axial force at different locations of the support components and even calculate the bending moment distribution, thereby determining whether the support is under uniform compression or whether there is a risk of local buckling.

[0029] Considering the harsh environment of the construction site and the potential for accidental damage or temporary inaccuracy of sensors, a dual-backup strategy is implemented at all critical monitoring points. This means that at each selected monitoring point crucial to overall safety assessment, such as the maximum displacement prediction point or the section with the maximum support stress, two sensors of identical model and performance are installed side-by-side. These two sensors operate independently, simultaneously measuring the same physical quantity. During system operation, edge computing nodes or subsequent data processing units continuously compare the readings of these two sensors. Under normal circumstances, the readings should remain consistent within the allowable error range; if a significant deviation occurs, the system immediately triggers a data quality alarm, indicating that the monitoring data at that point may be unreliable. This allows for a comprehensive judgment based on data from other sensors in the same area, or notifies maintenance personnel for inspection. This strategy not only significantly reduces the risk of missing critical data due to a single sensor failure, providing backup for continuous monitoring, but more importantly, it provides a real-time, built-in data verification mechanism. Through cross-validation, it effectively identifies and eliminates abnormal data caused by sensor drift, interference, or damage, improving the quality of data entering subsequent analysis stages from the source.

[0030] By effectively setting up a multi-parameter, three-dimensional sensing layer, it overcomes the problems of single monitoring dimensions and one-sided data in traditional monitoring, achieving synchronous perception of all elements of deep foundation pit engineering, including "structure-soil-groundwater-environment," making the information foundation for subsequent risk analysis extremely comprehensive and three-dimensional. Secondly, it introduces a targeted deployment design concept based on geological exploration and numerical simulation, changing the previous extensive mode of relying heavily on experience or uniform distribution of monitoring points. This allows limited monitoring resources to be precisely focused on weak links and high-risk areas of the project, greatly improving the efficiency of the monitoring network and the targeting of early warnings, realizing a shift from broad coverage to precise deployment. The unique dual-backup strategy and cross-validation mechanism greatly enhance the robustness and reliability of the entire monitoring system from both hardware and data source perspectives, effectively resisting interference and damage to individual sensors from various uncertainties at the construction site, ensuring the continuous stability and high reliability of the data stream, and providing indispensable hardware support for achieving low-latency, high-reliability risk early warning.

[0031] In another technical solution, step S2 specifically includes: The edge computing node is an industrial-grade embedded computing gateway deployed at the construction site. The industrial-grade embedded computing gateway establishes data connections with wired and wireless sensors in the sensor network through a fieldbus protocol interface and a wireless data receiving module, respectively. In the real-time preprocessing operation, the Kalman filter algorithm is used to reduce noise in the data. The Kalman filter algorithm adopts an adaptive process noise covariance matrix that includes prior knowledge of the foundation pit deformation. This prior knowledge is based on historical engineering data with similar geological conditions and support forms to the current foundation pit. When aligning timestamps of multi-source data based on a unified clock source, a time synchronization signal is periodically broadcast to all sensors in the network through an industrial-grade embedded computing gateway to complete hard synchronization. For minor time deviations caused by sensor sampling and transmission delays, a soft correction is performed using an interpolation algorithm based on the arrival sequence of data packets. The method to unify the data format into time-series data with a set sampling interval is as follows: after completing the timestamp alignment, the data stream of each monitoring item is resampled and aligned at fixed time intervals based on the system clock of the edge computing node, thereby generating a preprocessed standard time-series data stream that is strictly synchronized in the time dimension.

[0032] Edge computing nodes are not ordinary computing devices, but industrial-grade embedded computing gateways specifically designed to withstand the harsh environments of construction sites. These gateways typically feature high protection levels, wide operating temperatures, and abundant industrial communication interfaces, enabling stable operation in high-humidity, dusty, and frequently vibrating construction environments. In terms of hardware connectivity, the gateway establishes reliable, high-speed data links with wired sensors in the sensor network through physical interfaces supporting various fieldbus protocols, such as RS-485, CAN, or industrial Ethernet. Simultaneously, its built-in wireless data receiving module supports LoRa, ZigBee, or specific industrial wireless protocols to receive data from wireless sensors that are more flexibly deployed or difficult to wire. This hybrid wired and wireless connectivity architecture balances data transmission reliability with deployment flexibility, ensuring that data from hundreds or thousands of sensor units distributed throughout the foundation pit can be stably and continuously collected at the edge computing node, forming the real-time data inflow port for the entire early warning system.

[0033] For noise reduction, this method does not employ a simple static filter but introduces a more adaptive Kalman filtering algorithm. Kalman filtering is a recursive estimation algorithm based on a system state-space model. It can optimally estimate the true state of the system by combining the system's prior dynamic model and current actual observations. In this scheme, the design of the process noise covariance matrix incorporates prior knowledge of the foundation pit deformation. This knowledge is not fixed but originates from historical engineering data similar to the current foundation pit project in terms of geological conditions, support methods, and excavation depth. It is obtained by analyzing the statistical characteristics of the deformation process in this historical data. This allows the filter to understand that foundation pit deformation usually has a certain inertia and smoothness, thus more intelligently distinguishing between real deformation signals and sudden noise, resulting in a significant improvement in noise reduction compared to filters with fixed parameters. Regarding timestamp alignment, the scheme adopts a dual mechanism combining hard synchronization and soft correction. Hard synchronization refers to the edge computing gateway acting as the master clock, broadcasting a high-precision time signal to all sensors in the network at a frequency of, for example, once every 10 seconds or once per minute, forcibly unifying the absolute clock reference of each sensor. However, due to delays in the sensor's internal sampling circuitry, data processing time, and random delays in wireless transmission, there may still be minute deviations ranging from several milliseconds to hundreds of milliseconds when data packets arrive at the gateway. To address this, the system employs a soft correction algorithm based on the data packet arrival sequence. This involves constructing a time mapping function based on the actual arrival time sequence of the data packets and their theoretical sampling timestamps, fine-tuning the timestamp of each data point, thereby logically achieving synchronization of all data streams with microsecond-level precision.

[0034] After achieving precise timestamp alignment, the edge computing nodes use their own system clocks as the absolute time axis to define a unified and fixed target sampling interval for the data sequence of each monitoring project. This interval can be set based on a trade-off between the urgency of monitoring needs and computing resources. For example, for rapidly changing support axial forces and soil displacements, it can be set to 1 or 2 seconds; for slowly changing groundwater levels and building settlements, it can be set to 10 or 30 seconds. The system then resamples each raw data stream according to this target interval. For cases where there is no raw data at the target time point, algorithms such as linear interpolation or spline interpolation are used to generate the corresponding data values. Ultimately, the data from all monitoring projects are normalized to the same set of strictly aligned time points (t1, t2, t3, ...), generating a multi-channel, fully synchronized standard time-series data stream. Each frame of this data stream contains a snapshot of the overall state of the foundation pit at the same moment, providing high-quality, formatted input for subsequent spatiotemporal feature analysis.

[0035] This solution translates the abstract concept of edge computing into concrete, reliable industrial-grade hardware and a sophisticated data processing pipeline. By employing adaptive Kalman filtering, the signal-to-noise ratio of the raw data is significantly improved, allowing subtle, early-stage risk signals to emerge clearly from environmental noise, thus enhancing monitoring sensitivity. The innovative hardware-software combined time synchronization scheme fundamentally solves the time consistency problem in multi-source heterogeneous sensor network data fusion, ensuring that causal relationships and correlation analyses between different physical quantities such as displacement, stress, and water level are based on precise, simultaneous moments, avoiding misjudgments caused by time misalignment. The resulting standardized time-series data stream provides a clean, well-organized, and reliable data source for all subsequent advanced analysis modules.

[0036] Within an industrial-grade embedded computing gateway, a local lightweight digital twin model is established. This local lightweight digital twin model is a simplified version of the physical information neural network model, which is driven online in real time by receiving preprocessed standard time-series data streams. After generating a standard time-series data stream, the industrial-grade embedded computing gateway immediately inputs the data stream into the local lightweight digital twin model for forward computation and performs a rolling comparison between the predicted sequence output by the model and the monitoring data of the next actual sampling period. When the average error of the rolling comparison within a set number of consecutive times exceeds the preset model tolerance threshold, the industrial-grade embedded computing gateway determines that the local lightweight digital twin model is inaccurate and triggers a model update request. The model update request is encapsulated and sent to the remote monitoring center through the wireless communication network. After receiving a model update request, the remote monitoring center retrains and optimizes the physical information neural network model based on the complete historical database, and sends the updated model parameter set to the industrial-grade embedded computing gateway to replace its internal local lightweight digital twin model.

[0037] The local lightweight digital twin model is a simplified version of a fully functional physical information neural network risk warning model deployed on a remote server or in the cloud. The simplification aims to adapt to the limited computing power and storage resources of edge gateways, achieved by reducing the number of network layers, neurons, or using quantization compression techniques. However, its core function of receiving time-series data input and outputting risk state predictions is retained. The local lightweight digital twin model is pre-installed in an industrial-grade embedded computing gateway and runs continuously. Its driving source is the previously obtained standard time-series data stream. The model uses forward computation to predict the state of the foundation pit system in the next one or several sampling periods based on current and past data, such as predicting displacement or stress change trends in the next few seconds. This essentially establishes a rapidly responding digital mirror of the actual foundation pit at the edge, providing a core computing engine for immediate risk assessment.

[0038] Each time a new standard time-series data stream segment is generated, it is not only used for the current risk assessment but also undergoes a rolling comparison process. Specifically, the gateway inputs data from a period prior to the current moment into the local lightweight model, allowing the model to predict the monitoring values ​​at the next (or several) sampling moments, forming a prediction sequence. Then, when the actual sensor data arrives in the next sampling period, the system compares this actual measurement value with the model's prediction value one by one, calculating the prediction error. This comparison process is continuous and rolling. The system sets an observation window, for example, performing 50 or 100 consecutive rolling comparisons, and calculates the average of all prediction errors within that window. Simultaneously, the system presets a model tolerance threshold, which is a reasonable error boundary set based on the model's average performance during historical data testing or engineering experience. If the average error within the continuous observation window exceeds this tolerance threshold—for example, if the average error suddenly rises from the usual 1-2% and continues to exceed 5%—the system determines that the current local model can no longer accurately describe the dynamic behavior of the actual foundation pit, indicating model inaccuracy. Inaccuracy may stem from the foundation pit entering a new excavation stage, encountering unforeseen geological conditions, or a fundamental change in the structural state.

[0039] Once an inaccurate judgment is made, the edge computing node does not need to wait for manual intervention. Instead, it automatically triggers a model update request. This request is encapsulated in a data packet containing a brief log of the inaccuracy judgment and fingerprint information of recent key data samples, and is then sent to the remote monitoring center at the backend via a wireless communication network. Upon receiving the request, the remote monitoring center initiates a model retraining process. This process does not start from scratch but retrains and optimizes the complete physical information neural network model based on a more complete historical monitoring database of the foundation pit stored at the center, and can incorporate new on-site data. After training is complete, the center sends the updated, performance-calibrated model parameter set (usually a file) back to the edge computing gateway that issued the request. After security verification, the gateway replaces its original lightweight model parameters with the new parameter set, thus completing a model upgrade. Subsequently, the system continues to run with the updated model, restarting rolling comparison monitoring, forming an autonomous closed loop of monitoring-judgment-update-remonitoring. By deploying lightweight models at the edge and implementing rigorous rolling prediction comparisons, deviations between model prediction capabilities and actual working conditions can be keenly detected, allowing for timely identification of model failure risks caused by changes in engineering conditions. More importantly, it establishes an automated pipeline from edge-sensing inaccuracies to cloud-triggered retraining and subsequent updates, achieving closed-loop management of the model throughout its lifecycle. This fundamentally overcomes the problem in traditional monitoring systems where analytical models become fixed once deployed, potentially gradually deviating from engineering realities and ultimately failing. It ensures that the early warning model dynamically evolves throughout the entire excavation process, maintaining a high degree of synchronization with real-world physical behavior, significantly improving the long-term reliability of the entire system and the continuous accuracy of early warnings.

[0040] In another technical solution, step S3, the process of extracting the spatiotemporal feature vector for risk assessment from the standard time-series data stream, specifically includes: The time series data of each monitoring item are extracted with a set base time window length. The arithmetic mean, variance, and linear trend slope of the data within the window are calculated as statistical features. The difference in horizontal displacement of deep soil between adjacent inclinometer monitoring points deployed along the perimeter of the foundation pit at the same time section is calculated, as well as the difference between the readings of sensors deployed in groups at the support axial force monitoring section, as difference features. The rate of change of each monitoring data is calculated with the previous shortened time window as the rate of change feature. All the above-calculated statistical features, difference features, and rate of change features are spliced ​​together in a preset order to form a high-dimensional spatiotemporal feature vector.

[0041] The system sets a basic sliding time window length for the data sequence of each monitoring item. This window length is the basic time unit for feature extraction, and its selection needs to strike a balance between reflecting short-term dynamics and maintaining trend stability. For example, for rapidly changing support axial forces or deep displacements, the window length may be set to 10 to 30 minutes; for slowly changing building settlements, the window may be extended to several hours. Whenever new data arrives, the system extracts a data segment of the monitoring item within the latest time window and calculates a set of statistics. The arithmetic mean represents the central level of the monitoring data within that time period, the variance quantifies the degree of fluctuation of the data around the mean, and the linear trend slope is obtained by linearly fitting the data points within the window; it intuitively indicates whether the physical quantity is rising, falling, or remaining stable, and at what rate of change. These three statistical features together constitute a quantitative description of the short-term behavior of the monitoring point. For example, a continuously increasing average displacement, accompanied by high variance and a positive slope, strongly suggests that the point is in an unstable state of accelerated deformation.

[0042] The system acquires readings from all spatially correlated sensors at the same precise timestamp. For example, it calculates the difference in horizontal displacement readings at the same depth from two adjacent fixed inclinometers deployed around the perimeter of the foundation pit. If the foundation pit deformation is uniform, this difference should be small and stable; a sudden increase in the difference indicates uneven displacement or shear deformation of the soil between the two points, potentially a sign of local slippage. Similarly, for axial force on supports, the system calculates the difference between sensor readings from groups deployed near the mid-span and abutments of the same support. Under ideal stress conditions, the axial force should be uniformly transmitted along the support, with a small difference; a significant difference suggests that the support may be subjected to a large bending moment or there is a risk of local instability. These spatial difference characteristics link previously isolated point monitoring data, constructing a spatial network reflecting the relative deformation and stress state within the foundation pit structure, enabling the system to detect local inconsistencies.

[0043] To capture more immediate changes, the system uses a shorter time window than the basic statistical window to calculate the rate of change. For example, using a data window of 1 minute or 5 minutes prior to the current moment, the instantaneous rate of change of the monitored data within that window (such as the rate of change of displacement or axial force) is calculated. Rate features can amplify sudden acceleration or deceleration inflection points during slow changes. For example, soil displacement may be increasing slowly, but if its rate of change suddenly jumps from 0.1 mm / hour to 1.0 mm / hour, this rate jump is often a key signal of impending instability, providing a better early warning than simple cumulative displacement values. Finally, the system concatenates all the calculated statistical features (mean, variance, slope) from each monitoring item, the difference features from spatially correlated points, and the rate of change features of each monitoring item in a predefined, fixed order to form a high-dimensional spatiotemporal feature vector. This vector is a multi-dimensional detection report of the overall safety status of the foundation pit at the current moment, integrating time trends, spatial correlations, and instantaneous dynamics, providing highly condensed and physically meaningful input for subsequent intelligent risk assessment models. Through systematic feature engineering, raw, massive, and directly meaningful monitoring data is transformed into structured features with high information density and strong discriminative power. This overcomes the limitations of traditional methods that often focus only on single indicators (such as final cumulative displacement) or simple thresholds. By constructing multi-dimensional, multi-scale features, the risk warning model can identify the spatiotemporal correlation patterns and dynamic evolution processes behind the data. This greatly enriches the information basis for the model's decision-making, enabling warnings to be based not only on "what the current situation is," but also on "how fast the changes are" and "whether the various parts are coordinated," thereby significantly improving the ability to identify complex risk patterns, especially early-stage, slowly changing risks, and enhancing the accuracy of warnings.

[0044] After constructing a high-dimensional spatiotemporal feature vector, it is input into a pre-trained feature dimensionality reduction module, which is a principal component analysis model. The principal component analysis model is used to perform a linear transformation on the input high-dimensional spatiotemporal feature vector, retaining the top N principal components whose cumulative contribution rate exceeds a set threshold, and outputting a low-dimensional principal component feature vector. The low-dimensional principal component feature vector is used as the direct input to the risk warning model in step S4.

[0045] Since the generated high-dimensional spatiotemporal feature vectors may contain dozens or even hundreds of feature dimensions (e.g., dozens of monitoring points, each point extracting multiple statistical and rate features, plus numerous spatial difference features), directly inputting them into complex neural network models may result in low computational efficiency, easy overfitting during model training, and sensitivity to noise. Therefore, this solution integrates a pre-trained feature dimensionality reduction module after the feature extraction process, specifically using the classic principal component analysis (PCA) model. PCA is an unsupervised linear dimensionality reduction technique that, through orthogonal transformation, converts potentially correlated high-dimensional variables into a set of linearly uncorrelated new variables, i.e., principal components. These principal components are arranged in descending order of the variance of the original data they can explain (i.e., importance). The first principal component retains the directional information of the largest variance in the data, the second principal component has the largest remaining variance and is orthogonal to the first principal component, and so on.

[0046] Before system deployment or during the initialization phase, a large amount of historical or simulated training data (which has already undergone previous data processing to generate high-dimensional feature vectors) is used to train the PCA model. The training process determines the transformation matrix from the original high-dimensional feature space to the low-dimensional principal component space. In actual operation, whenever a new high-dimensional spatiotemporal feature vector is generated, the system immediately inputs it into the pre-trained PCA model. The model performs a linear transformation on the vector and calculates its projection scores on each principal component. The system does not retain all principal components but filters them based on a preset cumulative contribution rate threshold, for example, setting the threshold to 95%. This means the system will select the top N principal components, ensuring that the sum of the variance contribution rates of these N principal components exceeds 95% of the total variance. This N value is much smaller than the original feature dimension. Ultimately, the PCA model outputs a low-dimensional principal component feature vector consisting only of the scores of these top N principal components. This process essentially achieves data compression and redundancy removal while retaining most of the original information (variance), condensing the core variation patterns scattered across numerous original features into a few comprehensive indicators. This low-dimensional principal component feature vector, as the essence of the high-dimensional original features after purification and condensation, is directly designated as the input to the risk warning model in the subsequent step S4. This significantly reduces the dimensionality of the input neural network, lowering the model's computational complexity and inference time, which is crucial for achieving low-latency warnings on resource-constrained edge computing nodes. Secondly, since the principal components are orthogonal and uncorrelated, this helps improve the model's training stability and generalization performance, reducing overfitting. More importantly, the principal components retaining the maximum variance often correspond to the most important change patterns in the data. These patterns are likely closely related to the overall stability of the foundation pit system and the main risk drivers, allowing the warning model to focus more on core risk signals and filter out secondary or noise-induced fluctuations.

[0047] This method cleverly adds an information purification step to the information flow. Through principal component analysis, it compresses the high-dimensional initial feature space, which may contain redundancy and noise, into a low-dimensional feature representation with highly concentrated information. This significantly improves the processing efficiency and robustness of the subsequent risk warning model without losing core risk discrimination information. It enables the entire system to process complex feature inputs more quickly and stably even with the limited computing power of edge devices, ensuring the feasibility of low-latency warnings. Simultaneously, this dimensionality reduction also automatically extracts comprehensive risk indicators, making the model's learning process more efficient and accurate.

[0048] While generating low-dimensional principal component feature vectors, a distance-based local anomaly calculation process is performed in parallel. The obtained high-dimensional spatiotemporal feature vectors are input into a computational unit based on the local anomaly factor algorithm. This computational unit calculates a local anomaly score for each time step feature vector by comparing the local reachability density of each feature vector with its K nearest neighbor feature vectors in the feature space. The low-dimensional principal component feature vectors are then concatenated with the local anomaly scores to form the final input feature vector of the risk warning model in step S4.

[0049] In addition to the main dimensionality reduction path, a parallel auxiliary analysis path is established. This auxiliary analysis path primarily calculates a local anomaly score. Its input is also the previously generated high-dimensional spatiotemporal feature vector, but the algorithm used is a density-based classic unsupervised anomaly detection algorithm, namely the Local Anomaly Factor (LOF) algorithm. This algorithm determines whether a data point is an anomaly by comparing its local density with that of other data points in its neighborhood. If a data point's local density is significantly lower than that of its neighbors, it is considered an outlier in a sparse region, i.e., an anomaly. The system continuously inputs the newly generated high-dimensional feature vector at each time step into a specially configured LOF calculation unit. This unit searches for the K nearest neighbors of the current feature vector in a reference feature space composed of historical normal operating condition feature vectors (or in the feature vector set within a sliding time window). The value of K is an important parameter that needs to be adjusted according to the amount of data, for example, set to 20 or 50. The algorithm then precisely calculates the local reachability density of the current point and the average local reachability density of its K nearest neighbors. The ratio of the two is the LOF score at that moment, also known as the local anomaly score. The larger this score is than 1, the more anomalous and isolated the current feature vector appears relative to its surrounding historical or neighboring states.

[0050] The system does not replace PCA dimensionality reduction with LOF scores, nor does it allow the two to work independently. Instead, it computes them in parallel and then fuses the results. Specifically, in each sampling period, the system performs two operations simultaneously: first, the high-dimensional feature vector is reduced in dimensionality using the PCA model, outputting a low-dimensional principal component feature vector; second, the same high-dimensional feature vector is processed by the LOF computation unit, outputting a scalar value, namely the local anomaly score. Then, the system concatenates this scalar score with the low-dimensional vector. For example, if the PCA output is a vector containing 5 principal components [PC1, PC2, PC3, PC4, PC5], and the LOF score is 1.8, then the final input feature vector after concatenation is [PC1, PC2, PC3, PC4, PC5, 1.8]. This new vector contains information from two perspectives: the principal component feature vector represents the projection of the current state onto the global main change pattern, reflecting the overall situation; while the LOF score reflects the degree of outlier of the current state in the subtle, local feature space, a direct measure of unconventional and unforeseen anomalous patterns.

[0051] The final input feature vector formed by the concatenation is used as the direct input to the risk warning model in step S4. This means that while the neural network model receives condensed information (principal components) reflecting the overall situation of the foundation pit, it also receives a clear signal about the differences in the current state. This design greatly enhances the model's perception capability. For example, when a completely new damage pattern that has not been fully reflected in historical data begins to appear in the foundation pit, the principal component features after PCA dimensionality reduction may not change significantly because they have not learned such patterns, but the LOF score can rise sharply due to its obvious outlier. This increased score, as a strong feature input model, can effectively remind the model to pay attention to the anomalies of the current state, thereby triggering a higher risk probability output and making up for the potential blind spots of models that are purely based on historical patterns when facing unknown risks. This scheme constructs a dual-track feature supply system that combines global situational awareness and local anomaly detection. By introducing parallel LOF calculation, a warning radar that is highly sensitive to novelty and local anomalies is added to the risk warning model. This mechanism significantly enhances the system's ability to detect unknown risk patterns, early weak anomaly signals, and localized sudden emergencies, reducing the risk of missed detections that may result from insufficient training data for the main model (physical information neural network). By organically integrating the sensitivity of unsupervised anomaly detection with the comprehensive judgment capabilities of supervised neural network models, the early warning system not only excels at identifying known risk patterns but also possesses greater resilience and adaptability in dealing with unknown anomalies.

[0052] In another technical solution, in step S4, the pre-trained risk warning model adopts a hybrid physical information neural network architecture that integrates transfer learning and online incremental learning mechanisms. The initial training of the hybrid physical information neural network architecture is based on a historical monitoring dataset of multiple engineering projects containing various geological conditions and foundation pit types. Through the transfer learning mechanism, the general feature knowledge learned from one or more source engineering projects is transferred to the risk warning model for the current target foundation pit. The physical constraints of the physical information neural network model are formed by adding a physical loss term to the loss function of the hybrid physical information neural network. This physical loss term is calculated based on the static equilibrium differential equation of the foundation pit soil. The static equilibrium differential equation of the foundation pit soil is: Where σ is the stress tensor predicted by the model, Here, b is the divergence operator, and b is the force vector; the physical loss term is... Specifically, it is the mean of the squared norms of the equation residuals calculated at a series of points within the computational domain, i.e.: in For the neural network at the placement point x i The predicted stress at the location, N p For the total number of configuration points, The norm is used; after deploying the risk warning model on edge computing nodes, an online incremental learning mechanism is activated. It uses the newly generated standard time-series data stream of the current foundation pit as incremental data and adopts a small-batch gradient descent algorithm to continuously fine-tune the model parameters. At the same time, to prevent catastrophic forgetting, an elastic weight consolidation algorithm is introduced to constrain the changes of important parameters.

[0053] The risk warning model is a hybrid physical information neural network architecture that integrates two advanced learning paradigms: transfer learning and online incremental learning. The model's initial training does not start from scratch and rely solely on limited data from a single foundation pit. Instead, it is built upon a vast historical monitoring dataset of multiple engineering projects, encompassing various geological conditions (such as soft soil, sand, and rock strata) and foundation pit types (such as deep foundation pits for subway stations and raft foundation pits for high-rise buildings). The transfer learning mechanism allows the model to learn general characteristics and patterns regarding foundation pit deformation, stress, and instability precursors from one or more completed, data-rich source projects. This learned knowledge is captured and stored as weight parameters at the bottom layer of the neural network. When building a warning model for a new target foundation pit, the system does not randomly initialize the network parameters but loads the universally applicable weights learned from the source projects as initial values. This is equivalent to giving the model rich engineering experience before it begins learning for a specific task, greatly accelerating model convergence and significantly improving prediction accuracy and stability in the early stages of target foundation pit monitoring when data is insufficient.

[0054] An additional physical loss term is added to the loss function used in model training. The overall training objective of the model is not only to minimize the error between the network's predicted values ​​and the actual monitored values ​​(data loss), but also to minimize this physical loss term. This physical loss term is directly derived from the static equilibrium differential equations in geotechnical engineering, which describe the equilibrium between internal stresses and external forces in the soil. During training, the system virtually arranges a series of sampling points (configuration points) within the computational domain (i.e., the soil region of interest in the foundation pit). For each configuration point, the neural network predicts the stress state at that point based on the input features, and then substitutes this predicted stress into the static equilibrium equation for calculation. Theoretically, if the prediction perfectly conforms to physical laws, the equation should hold strictly, and the residual should be zero; in practice, the magnitude of the calculated residual constitutes the physical loss. The value of the physical loss term is obtained by averaging the squares of these residuals across all configuration points. Therefore, during training, the optimization algorithm simultaneously drives the network in two directions: first, to fit historical monitoring data, and second, to ensure that its predictions satisfy basic mechanical equilibrium. This forces the network to follow physical common sense in its predictions even in areas where the data may contain noise, blind spots, or sparse regions, thereby significantly improving the model's generalization ability, interpretability, and reliability in extrapolation cases.

[0055] As the foundation pit project continues, new monitoring data are constantly generated, carrying the latest project status information. An online incremental learning mechanism is activated, using these newly generated standard time-series data streams as incremental data to continuously fine-tune the deployed model. This process typically employs a mini-batch gradient descent algorithm, using a small batch of new data to incrementally update the model parameters each time, enabling the model to dynamically adapt to the gradual changes in soil properties and the stress state of the support structure throughout the entire process from excavation to backfilling. However, continuous learning of new knowledge can lead to a serious problem: the model may overwrite or forget old knowledge learned from a large amount of historical data when learning new data patterns. To prevent this, an elastic weight consolidation algorithm is introduced. Before incremental learning begins, the importance of each parameter in the existing neural network to the learned task (i.e., accurately predicting historical conditions) is assessed. During subsequent fine-tuning, the allowable variation of parameters identified as very important is strictly limited or subject to a strong constraint; while relatively unimportant parameters are allowed larger adjustments. In this way, the model can flexibly absorb new information while firmly remembering core past knowledge, achieving stable knowledge accumulation and updating, and ensuring that early warning capabilities remain at a high level throughout the entire project cycle. This method fundamentally improves the performance and engineering practicality of risk early warning models. By integrating transfer learning, the model gains a strong starting point advantage and generalization ability, enabling it to quickly provide relatively reliable judgments even when faced with a completely new foundation pit, effectively overcoming the dependence of traditional machine learning models on large amounts of labeled data for the target scene. By embedding physical constraints, the model transforms from a pure data fitting tool into a physical information inference engine guided by physical laws. Its prediction results are not only accurate in areas with abundant data, but also more reasonable and robust when data is sparse or the operating conditions exceed historical ranges, significantly reducing false alarms and missed alarms caused by data-driven blindness. By introducing online incremental learning with an anti-forgetting mechanism, it can continuously evolve along with the project progress, always maintaining the best fit to the current operating conditions, achieving long-term reliability and adaptive improvement of early warning capabilities.

[0056] The hybrid physical information neural network architecture running in the edge computing node implements its transfer learning mechanism through a pre-built source domain model pool. The source domain model pool stores multiple pre-trained physical information neural network sub-models under different typical engineering conditions. During initial deployment, based on the current 3D geological survey data of the foundation pit and the design parameters of the support structure, at least two sub-models with the highest matching degree are selected from the source domain model pool, and some of their network layer parameters are loaded into the current target model as initialization values. The implementation of the online incremental learning mechanism includes an uncertainty-based data filtering step. Using the Monte Carlo Dropout method built into the physical information neural network model, the newly generated standard time-series data stream is forward-propagated multiple times to calculate the variance of the model's predicted output. Only data samples with prediction uncertainty higher than a set threshold are filtered out as valuable incremental data for fine-tuning the model parameters.

[0057] The source domain model pool is a knowledge base maintained by the remote monitoring center, storing multiple pre-trained physical information neural network sub-models under different typical engineering conditions. These typical conditions can be classified and constructed according to key dimensions such as geological type (e.g., silty clay foundation pit model, silty sand layer foundation pit model), support form (e.g., diaphragm wall + internal support model, pile anchor support model), and excavation depth (e.g., shallow, deep). Each sub-model is an expert model trained using rich historical data of the corresponding category. When an early warning system needs to be deployed for the current target foundation pit, the system does not blindly select a general model, but intelligently selects one or more sub-models with the highest matching degree from the source domain model pool based on key information such as the detailed 3D geological survey report and support structure design drawings of the current foundation pit. The matching degree can be calculated based on multi-dimensional features such as the similarity of geological parameters and the analogy of support structures. For example, for a deep foundation pit in a soft soil area using a diaphragm wall with five layers of support, the system can select a combination of a diaphragm wall deep foundation pit sub-model for soft soil areas and a multi-layer internal support system sub-model. After filtering, the system extracts the parameters of some network layers (usually the first few layers used to extract common features from the underlying layers) in these highly matched sub-models and uses them as the initialization parameters of the corresponding layer of the current target model.

[0058] Edge computing nodes continuously generate new monitoring data. However, indiscriminately using all new data for model fine-tuning is not only computationally inefficient but may also introduce significant redundancy or noise, potentially leading the model to learn irrelevant fluctuation patterns. To address this, this solution employs an intelligent filtering gating mechanism, utilizing the Monte Carlo Dropout method built into the physical information neural network model. During the inference phase, when a batch of new data is input into the model, the system does not perform a single forward propagation to obtain a single prediction. Instead, it activates the Dropout mechanism, performing multiple (e.g., 10 to 100) random forward propagations. Since Dropout randomly shuts down a portion of neurons during each propagation, this is equivalent to allowing the model to make multiple predictions from slightly different sub-network perspectives. For the same input, the outputs of these multiple predictions will form a distribution. The system calculates the variance of this prediction distribution. The magnitude of variance directly reflects the model's uncertainty about the current input data: a large variance indicates internal disagreement within the model regarding how to determine the current state, suggesting a lack of confidence. This may mean the current condition exceeds the model's comfort zone based on past experience, representing a new and valuable learning sample. Conversely, a small variance indicates high certainty in the model's prediction, suggesting the current data pattern may have been sufficiently learned and represents redundant information. The system sets an uncertainty threshold, for example, if the variance of the predicted probability exceeds 0.01. Only data samples with uncertainty higher than this threshold are selected and marked as valuable incremental data.

[0059] These data samples, marked as high-uncertainty, represent the weak links or gaps in the model's knowledge system. The system organizes this data into small batches and feeds them into the online incremental learning process. Under the constraints of the elastic weight consolidation algorithm, these small batches of data are used to update the model parameters using gradient descent. Because the data is filtered and high-value, each parameter update is more directional, resulting in higher learning efficiency. In this way, the model continuously and selectively compensates for its cognitive shortcomings, becoming increasingly familiar with new but potentially risky state patterns, and its early warning response becomes more sensitive. Simultaneously, the remote monitoring center can periodically or irregularly summarize the incremental data and model update status uploaded by each edge node, and use more powerful computing power in the background to perform global model retraining and optimization. Then, the improved model parameters are redistributed, forming a co-evolutionary cycle from edge-aware uncertainty perception to centralized optimization in the cloud, and then back to edge. By constructing a source domain model pool, it transforms transfer learning from an abstract concept into queryable and matchable standardized knowledge components, making the model initialization process for specific projects scientific, accurate, and efficient, maximizing the use of valuable experience assets accumulated from historical projects. By introducing Monte Carlo Dropout-based uncertainty data filtering, it installs an intelligent information filter for online incremental learning, ensuring that limited computational resources are used only to learn new knowledge that truly expands the model's cognitive boundaries. This avoids ineffective learning and noise interference, significantly improving the efficiency of incremental learning and the purity of model evolution. The combination of these two aspects gives the entire early warning system not only strong initial capabilities and continuous learning potential, but also an efficient and economical knowledge absorption and updating strategy. Thus, throughout the entire project lifecycle, it maintains a higher level of early warning performance that is more focused on real risks at a lower computational cost.

[0060] In another technical solution, step S5 specifically includes: The warning information is encapsulated into a structured warning data packet, which includes three parts: a header, a body, and a footer. The header includes the data format version number and unique sequence number of the warning information. The body includes the risk probability value, the identifier of the main anomaly monitoring item that triggered the warning and its corresponding real-time monitoring data, the time of the anomaly, and the risk level code mapped according to the preset interval where the risk probability value is located. The footer includes a cyclic redundancy check code calculated based on the content of the body. Edge computing nodes synchronously send warning data packets to the remote monitoring center via at least two independent wireless communication links. These two independent wireless communication links include one link based on a cellular mobile network and one link based on a local wireless LAN. After receiving the warning data packet from either link, the remote monitoring center first verifies its cyclic redundancy check code. If the verification passes, it parses and stores the warning information and simultaneously returns an acknowledgment signal containing the corresponding sequence number to the edge computing node. If the edge computing node does not receive an acknowledgment signal from the remote monitoring center within a set time after sending the warning data packet, it retransmits the warning data packet via the other wireless communication link.

[0061] When an edge computing node determines that the risk probability exceeds a threshold and triggers an alert, it doesn't simply send a piece of free text. Instead, it encapsulates the alert information into a highly structured alert data packet. This packet contains three logical parts: a header, a body, and a trailer. The header acts as the packet's identifier and format declaration. It must include a data format version number, such as version 1.0 or 2.0. This allows the receiving software at the remote monitoring center to recognize and be compatible with data formats generated by systems deployed at different times, achieving backward compatibility and smooth upgrades. Simultaneously, the header also contains a unique sequence number, typically generated by a combination of a timestamp, device ID, and an incrementing counter. This ensures that each issued alert data packet is uniquely identifiable globally, which is crucial for subsequent log tracking, confirmation, and deduplication. The message body is the core carrier area of ​​the data packet, systematically containing all the key information for the warning: the most crucial risk probability value, a quantified risk index; identifiers of the main anomaly monitoring items that triggered the warning, such as specifying whether it's "abnormal deep displacement rate of the CX-05 inclinometer" or "sudden change in axial force of the ZC-03 support," which directly guides engineers to focus on specific problem points; real-time monitoring data corresponding to these anomalies, providing concrete numerical evidence of the anomaly's occurrence; a precise timestamp of the anomaly's occurrence; and a risk level code automatically mapped based on a preset interval of the risk probability value. For example, a probability value between 0.6 and 0.8 can be mapped to a yellow warning (code YELLOW), and values ​​above 0.8 can be mapped to a red warning (code RED), allowing the receiver to quickly understand the urgency of the risk. The message tail contains a cyclic redundancy check (CRC) code calculated based on the entire message body. This is a widely used error detection code, allowing the receiver to recalculate the CRC and compare it with the value in the message tail to verify whether any bit errors or tampering occurred during data transmission, thus ensuring information integrity.

[0062] Given the complex and variable communication environment at construction sites, a single communication link may be interrupted due to signal interference, base station congestion, or equipment failure. Therefore, it is mandatory for edge computing nodes to simultaneously transmit encapsulated warning data packets via at least two independent wireless communication links. These two links are typically heterogeneous in terms of physical medium and technology to avoid shared failure risks. One typical link is based on a wide-area coverage cellular mobile network, such as 4G or 5G, which provides broad connectivity but may experience signal instability in localized basements or remote areas. The other complementary link is based on a local wireless LAN with local coverage, such as through industrial WiFi base stations or mesh self-organizing networks deployed at the project site. This link typically has lower latency but may fail when the node is too far from the base station. During transmission, the edge computing node simultaneously duplicates the warning data packet and sends it to the corresponding receiving address at the remote monitoring center via both independent links. This dual-transmission mechanism significantly increases the probability of successful first-time transmission of warning information; as long as either link is operational at any given moment, the warning can be issued.

[0063] Upon receiving an early warning data packet from any link, the remote monitoring center immediately initiates the processing procedure. First, a cyclic redundancy check (CRC) verification is performed. If the verification fails, the corrupted data packet is discarded and logged. If the verification succeeds, the sequence number in the header and the content in the body are parsed, stored in the database, and an alarm notification is triggered in the background (e.g., audible and visual alarm, SMS push notification). Next, the monitoring center must return a short acknowledgment signal to the edge computing node that sent the data packet. This signal contains the sequence number of the corresponding early warning data packet, clearly indicating which early warning is being acknowledged. This acknowledgment signal is typically sent from the most reliable link (such as the link that was successfully received earlier). After sending the early warning data packet, the edge computing node starts a timer to wait for the acknowledgment signal. This waiting time (i.e., the "set time") can be set according to network conditions, for example, 3 seconds or 5 seconds. If, within this time, the edge node receives an acknowledgment signal containing the correct sequence number through any means, the entire early warning transmission process successfully ends. If no acknowledgment is received after the timer expires, the edge node infers that the previous transmission attempt may have failed (although it's a dual transmission, in extreme cases both links may temporarily fail simultaneously). In this case, it will automatically initiate a retransmission process. However, the retransmission is not a blind repetition; instead, it retransmits through another wireless communication link. For example, if no acknowledgment is received after the initial dual transmission, the node may choose the link it deems more stable at the time (or alternate between links) to retransmit the warning data packet until acknowledgment is received. This combination of dual transmission + acknowledgment + timeout retransmission mechanism constructs a highly robust communication assurance system. By standardizing and structurally encapsulating the warning information, it ensures the unambiguity, machine readability, and verifiability of the information content, providing directly usable high-quality data input for the automated emergency response system and avoiding processing delays or misinterpretations caused by inconsistent information formats. By forcing the use of dual-link heterogeneous synchronous transmission, it greatly overcomes the risk of single-point failures in the harsh wireless communication environment of construction sites, increasing the probability of successful delivery of warning information to an extremely high level and ensuring communication resilience at critical moments. By introducing end-to-end confirmation and intelligent retransmission mechanisms, the originally unreliable wireless transmission is transformed into a near-reliable transmission service, ensuring that no critical early warning information is lost during transmission, and the remote monitoring center can be confident that it has obtained a complete on-site risk situation.

[0064] It should be noted that although the steps are described in a specific order above, this does not mean that they must be performed in that order. In fact, some of these steps can be executed concurrently, or even in a different order, as long as the required functionality is achieved. The number of devices and processing scale described herein are for simplification of the invention; applications, modifications, and variations of this invention will be readily apparent to those skilled in the art.

[0065] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and illustrations shown and described herein.

Claims

1. A low-latency risk early warning method for deep foundation pits, characterized in that, Includes the following steps: S1: By deploying a sensor network at the deep foundation pit engineering site, data such as displacement of the support structure, axial force of the support, horizontal displacement of the deep soil, groundwater level, and settlement of surrounding buildings are collected to form the original multi-source monitoring dataset. S2: Transmit the original multi-source monitoring dataset to the edge computing node in real time. The received data is preprocessed in real time at the edge computing node. The preprocessing includes noise reduction, timestamp alignment of multi-source data, and unification of data format into time-series data with a set sampling interval, generating a preprocessed standard time-series data stream. S3: In the edge computing node, extract the spatiotemporal feature vector for risk assessment from the standard time-series data stream. The spatiotemporal feature vector includes the statistical characteristics of the same monitoring project within a continuously set time window, the difference characteristics between monitoring data at different spatial locations under the same time section, and the rate of change characteristics. S4: Input the spatiotemporal feature vector into the pre-trained risk warning model. The risk warning model is a physical information neural network model that integrates the differential equation of soil equilibrium in the foundation pit as a physical constraint. The risk warning model calculates the input spatiotemporal feature vector and outputs the risk probability value that represents the overall risk state of the foundation pit. S5: When the risk probability value exceeds the preset warning threshold, the edge computing node immediately generates a warning message containing the risk probability value, the identifier of the main anomaly monitoring item, and the time of the anomaly occurrence, and sends the warning message to the remote monitoring center.

2. The method for low-delay risk early warning of deep foundation pits according to claim 1, characterized in that, Step S1 specifically includes: The sensor network deployed at the deep foundation pit engineering site adopts a multi-parameter sensing layer. The multi-parameter sensing layer includes fully automatic intelligent total stations deployed at a set density on the top of the foundation pit support piles and foundation pit support walls, vibrating wire stress sensors installed at the key sections of each horizontal support, fixed inclinometer arrays buried at a set depth along the perimeter of the foundation pit, anti-clogging pore water pressure gauges distributed inside and outside the foundation pit, and static leveling instrument systems arranged at the foundations of surrounding buildings that need to be protected. The deployment points and depths of the fixed inclinometer array are designed based on the three-dimensional geological survey data and support structure before the excavation of the foundation pit. The potential maximum displacement area and sliding surface depth are determined through finite element numerical simulation analysis. Vibrating wire stress sensors are deployed in groups at the mid-span of each support and near the two supports to monitor the bending moment distribution of the support. All types of sensors in the sensor network are deployed using a dual backup strategy. Two sensors of the same model are installed at each key monitoring point. The measurement data from the two sensors are used for cross-validation and data quality assessment.

3. The method for low-delay risk early warning of deep foundation pits according to claim 1, characterized in that, Step S2 specifically includes: The edge computing node is an industrial-grade embedded computing gateway deployed at the construction site. The industrial-grade embedded computing gateway establishes data connections with wired and wireless sensors in the sensor network through a fieldbus protocol interface and a wireless data receiving module, respectively. In the real-time preprocessing operation, the Kalman filter algorithm is used to reduce noise in the data. The Kalman filter algorithm adopts an adaptive process noise covariance matrix that includes prior knowledge of the foundation pit deformation. This prior knowledge is based on historical engineering data with similar geological conditions and support forms to the current foundation pit. When aligning timestamps of multi-source data based on a unified clock source, a time synchronization signal is periodically broadcast to all sensors in the network through an industrial-grade embedded computing gateway to complete hard synchronization. For minor time deviations caused by sensor sampling and transmission delays, a soft correction is performed using an interpolation algorithm based on the arrival sequence of data packets. The method to unify the data format into time-series data with a set sampling interval is as follows: after completing the timestamp alignment, the data stream of each monitoring item is resampled and aligned at fixed time intervals based on the system clock of the edge computing node, thereby generating a preprocessed standard time-series data stream that is strictly synchronized in the time dimension.

4. The method for low-delay risk early warning of deep foundation pits according to claim 3, characterized in that, Within an industrial-grade embedded computing gateway, a local lightweight digital twin model is established. This local lightweight digital twin model is a simplified version of the physical information neural network model, which is driven online in real time by receiving preprocessed standard time-series data streams. After each standard time-series data stream is generated, the industrial-grade embedded computing gateway immediately inputs the data stream into the local lightweight digital twin model for forward computation, and performs a rolling comparison between the predicted sequence output by the model and the monitoring data of the next actual sampling period. When the average rolling comparison error within a set number of consecutive times exceeds the preset model tolerance threshold, the industrial-grade embedded computing gateway determines that the local lightweight digital twin model is inaccurate and triggers a model update request. The model update request is encapsulated and sent to the remote monitoring center through the wireless communication network. After receiving a model update request, the remote monitoring center retrains and optimizes the physical information neural network model based on the complete historical database, and sends the updated model parameter set to the industrial-grade embedded computing gateway to replace its internal local lightweight digital twin model.

5. The method for low-delay risk early warning of deep foundation pits according to claim 1, characterized in that, Step S3, the process of extracting the spatiotemporal feature vector for risk assessment from the standard time-series data stream, specifically includes: The time series data of each monitoring item are extracted with a set base time window length. The arithmetic mean, variance, and linear trend slope of the data within the window are calculated as statistical features. The difference in horizontal displacement of deep soil between adjacent inclinometer monitoring points deployed along the perimeter of the foundation pit at the same time section is calculated, as well as the difference between the readings of sensors deployed in groups at the support axial force monitoring section, as difference features. The rate of change of each monitoring data is calculated with the previous shortened time window as the rate of change feature. All the above-calculated statistical features, difference features, and rate of change features are spliced ​​together in a preset order to form a high-dimensional spatiotemporal feature vector.

6. The method for low-delay risk early warning of deep foundation pits according to claim 5, characterized in that, After constructing a high-dimensional spatiotemporal feature vector, it is input into a pre-trained feature dimensionality reduction module, which is a principal component analysis model. The principal component analysis model is used to perform a linear transformation on the input high-dimensional spatiotemporal feature vector, retaining the top N principal components whose cumulative contribution rate exceeds a set threshold, and outputting a low-dimensional principal component feature vector. The low-dimensional principal component feature vector is used as the direct input to the risk warning model in step S4.

7. The method for low-delay risk early warning of deep foundation pits according to claim 6, characterized in that, While generating low-dimensional principal component feature vectors, a distance-based local anomaly calculation process is performed in parallel. The obtained high-dimensional spatiotemporal feature vectors are input into a computational unit based on the local anomaly factor algorithm. This computational unit calculates a local anomaly score for each time step feature vector by comparing the local reachability density of each feature vector with its K nearest neighbor feature vectors in the feature space. The low-dimensional principal component feature vectors are then concatenated with the local anomaly scores to form the final input feature vector of the risk warning model in step S4.

8. The method for low-delay risk early warning of deep foundation pits according to claim 1, characterized in that, In step S4, the pre-trained risk warning model adopts a hybrid physical information neural network architecture that integrates transfer learning and online incremental learning mechanisms; The initial training of the hybrid physical information neural network architecture is based on a historical monitoring dataset of multiple engineering projects containing various geological conditions and foundation pit types. Through transfer learning, the general feature knowledge learned from one or more source engineering projects is transferred to the risk warning model for the current target foundation pit. The physical constraints of the physical information neural network model are formed by adding a physical loss term to the loss function of the hybrid physical information neural network. This physical loss term is calculated based on the static equilibrium differential equation of the foundation pit soil. The static equilibrium differential equation of the foundation pit soil is: Where σ is the stress tensor predicted by the model, Here, b is the divergence operator, and b is the force vector; the physical loss term is... Specifically, it is the mean of the squared norms of the equation residuals calculated at a series of points within the computational domain, i.e.: in For the neural network at the placement point x i The predicted stress at the location, N p For the total number of configuration points, The norm is used; after deploying the risk warning model on edge computing nodes, an online incremental learning mechanism is activated. It uses the newly generated standard time-series data stream of the current foundation pit as incremental data and adopts a small-batch gradient descent algorithm to continuously fine-tune the model parameters. At the same time, to prevent catastrophic forgetting, an elastic weight consolidation algorithm is introduced to constrain the changes of important parameters.

9. The method for low-delay risk early warning of deep foundation pits according to claim 8, characterized in that, The hybrid physical information neural network architecture running in the edge computing node implements its transfer learning mechanism through a pre-built source domain model pool. The source domain model pool stores multiple pre-trained physical information neural network sub-models under different typical engineering conditions. During initial deployment, based on the current 3D geological survey data of the foundation pit and the design parameters of the support structure, at least two sub-models with the highest matching degree are selected from the source domain model pool, and some of their network layer parameters are loaded into the current target model as initialization values. The implementation of the online incremental learning mechanism includes an uncertainty-based data filtering step. Using the Monte Carlo Dropout method built into the physical information neural network model, the newly generated standard time-series data stream is forward-propagated multiple times to calculate the variance of the model's predicted output. Only data samples with prediction uncertainty higher than a set threshold are filtered out as valuable incremental data for fine-tuning the model parameters.

10. The method for low-delay risk early warning of deep foundation pits according to claim 1, characterized in that, Step S5 specifically includes: The warning information is encapsulated into a structured warning data packet, which includes three parts: a header, a body, and a footer. The header includes the data format version number and unique sequence number of the warning information. The body includes the risk probability value, the identifier of the main anomaly monitoring item that triggered the warning and its corresponding real-time monitoring data, the time of the anomaly, and the risk level code mapped according to the preset interval where the risk probability value is located. The footer includes a cyclic redundancy check code calculated based on the content of the body. Edge computing nodes synchronously send warning data packets to the remote monitoring center via at least two independent wireless communication links. These two independent wireless communication links include one link based on a cellular mobile network and one link based on a local wireless LAN. After receiving the warning data packet from either link, the remote monitoring center first verifies its cyclic redundancy check code. If the verification passes, it parses and stores the warning information and simultaneously returns an acknowledgment signal containing the corresponding sequence number to the edge computing node. If the edge computing node does not receive an acknowledgment signal from the remote monitoring center within a set time after sending the warning data packet, it retransmits the warning data packet via the other wireless communication link.

Images