Safety monitoring method and system for open caisson construction

Abstract

This application discloses a safety monitoring method and system for caisson construction, relating to the field of civil engineering construction safety monitoring technology. The disclosed safety monitoring method and system for caisson construction constructs a full-link safety monitoring system from real-time data acquisition and multi-scale feature extraction to dynamic risk assessment through collaborative perception of multi-physics field sensing and edge computing, intelligent fusion of spatial features of physical information graphs and graph neural networks, and continuous evolution mechanism of parallel inference of digital twins and federated learning. It breaks through the limitations of traditional monitoring methods such as data fragmentation, delayed early warning, and ambiguous risk positioning. By deeply coupling the micro-activity signs of soil with the dynamic response of the structure, it realizes early identification, accurate source tracing, and dynamic handling of the instability risk of the caisson-soil system, thereby improving the safety management level and risk response capability of caisson construction under complex geological conditions.

Description

Technical Field

[0001] This application relates to the field of civil engineering construction safety monitoring technology, and in particular to safety monitoring methods and systems for caisson construction. Background Technology

[0002] Caisson construction, as a core technology in deep foundation engineering, involves dynamic interactions between the soil and structure during its sinking process. It is susceptible to instability risks such as tilting and sudden settlement due to abrupt changes in geological conditions or construction disturbances. Current engineering practices commonly employ sensors such as displacement gauges and stress gauges to monitor the macroscopic deformation of the caisson wall, triggering alarm mechanisms through preset thresholds. This approach has two limitations: First, the monitoring signals only reflect the macroscopic deformation of the soil. When the displacement rate or stress value exceeds the safety threshold, the structure has already entered a critical instability state, leaving insufficient time for on-site personnel to take emergency measures. Second, it is difficult to effectively trace the physical causes of abnormal monitoring data. For example, it is difficult to distinguish between different contributing factors such as localized soil loss below the cutting edge, soil softening caused by groundwater seepage, or frictional imbalance between the caisson wall and the soil, making it difficult for engineers to develop targeted response plans. This passive response mode based on macroscopic phenomena leads to low risk management efficiency and the risk of misjudgment. It fails to achieve early identification and mechanism tracing of precursors to instability in the caisson-soil system, making it difficult to meet the needs of modern deep foundation engineering for proactive safety control.

[0003] The above content is only used to help understand the technical solution of this application and does not represent an admission that the above content is prior art. Summary of the Invention

[0004] The main purpose of this application is to provide a safety monitoring method and system for caisson construction, which aims to improve the efficiency and accuracy of risk management.

[0005] To achieve the above objectives, this application proposes a safety monitoring method for caisson construction, the method comprising: Multiple integrated sensing units deployed in the caisson structure and surrounding soil are used to synchronously collect original multi-physics field monitoring data during the construction process; the original multi-physics field monitoring data includes acoustic emission signals, vibration signals and soil contact pressure signals. Edge computing processing is performed on the original multiphysics field monitoring data collected by each integrated sensing unit to extract local symptom feature vectors that characterize the soil micro-activity state and stress change trend at the monitoring location of the integrated sensing unit. The local symptom feature vectors from all integrated sensing units are uploaded to the cloud analysis platform. In the cloud analysis platform, a physical information map is constructed based on the three-dimensional model of the caisson structure, the spatial location information of each integrated sensing unit, and the preset geological stratification information. The nodes of the physical information map represent soil or structural units and carry the corresponding local symptom feature vectors. The edges of the map represent the spatial adjacency relationship between units and are assigned physical attribute weights based on the geological and construction conditions. The physical information map is input into a pre-trained graph neural network model for message passing and feature aggregation, so as to fuse the spatially associated local symptom feature vectors and generate a symptom map reflecting the current dynamic behavior pattern of the caisson-soil system. The symptom map is matched and compared with the preset well instability mode knowledge base in real time to obtain the first comparison result. The system state reflected by the symptom map is used as the initial condition to drive the parallel digital twin model to perform forward inference, predict the evolution trend of the system state within the preset inference time, and obtain the inference prediction result. A risk assessment is performed by combining the first comparison result and the inference and prediction results to generate a corresponding intelligent diagnostic report; the intelligent diagnostic report includes the risk level, a comprehensive judgment of the instability mode, an inferred risk root cause, and targeted handling suggestions.

[0006] In one embodiment, the step of simultaneously collecting raw multi-physics field monitoring data during construction by deploying multiple integrated sensing units in the caisson structure and surrounding soil further includes: Each integrated sensing unit is controlled to synchronously collect the original monitoring data of the multi-physics field at a first preset acquisition frequency. Monitor the rate of change of the newly acquired multiphysics field raw monitoring data of each integrated sensing unit relative to the data of the previous period; When the rate of change of any physical field data of any integrated sensing unit exceeds its corresponding first threshold, it is determined that an initial abnormality has occurred at that location. The data acquisition frequency of the integrated sensing unit is switched to a second preset acquisition frequency higher than the first preset acquisition frequency. Based on the preset adjacency relationship of the physical information map, the adjacent integrated sensing units are notified to also switch their data acquisition frequency to the second preset acquisition frequency. Data is acquired at the second preset acquisition frequency, and the acquired preset frequency band signals related to soil micro-fractures or plastic flow are subjected to directional analysis. When all units that have switched to the second preset acquisition frequency have a data change rate that is lower than their corresponding second threshold and this continues for a preset time, it is determined that the abnormal symptoms have subsided, and the data acquisition frequency of the relevant units is switched back to the first preset acquisition frequency; wherein, the second threshold is lower than the first threshold.

[0007] In one embodiment, the step of performing edge computing processing on the raw multiphysics monitoring data collected by each integrated sensing unit to extract local symptom feature vectors characterizing the soil micro-activity state and stress change trend at the monitoring location of the unit includes: The acoustic emission signal is subjected to event detection and waveform analysis to extract the acoustic emission event rate, cumulative event energy value and main frequency offset to obtain an acoustic feature set; The vibration signal is subjected to spectrum analysis and energy calculation to extract the vibration main frequency, the energy ratio of the preset frequency band and the signal entropy value, so as to obtain the vibration feature set; Spatial gradient and time series analysis were performed on the soil contact pressure signal to calculate the rate of change of the current pressure value relative to the historical baseline, the pressure gradient value between adjacent sensors, and the pressure distribution entropy, thereby obtaining a pressure feature set. The acoustic feature set, the vibration feature set, and the pressure feature set are spliced ​​together in a predetermined order to form the local symptom feature vector.

[0008] In one embodiment, the step of uploading the local symptom feature vectors from all integrated sensing units to a cloud analysis platform, and constructing a physical information map on the cloud analysis platform based on the three-dimensional model of the well structure, the spatial location information of each integrated sensing unit, and preset geological stratification information includes: Based on the preset geological stratification information, determine the first weight coefficient of the edge between the corresponding nodes of the integrated sensing unit at the junction of different soil layers; Based on the sinking depth and real-time attitude data of the caisson during the current construction phase, calculate the second weighting coefficient of the edge between the corresponding nodes of the integrated sensing unit in the contact area between the well wall and the soil. The first weight coefficient and the second weight coefficient are combined to obtain the physical attribute weight; The nodes carrying the local symptom feature vectors and the edges assigned the weights of the physical attributes are combined to form the physical information graph.

[0009] In one embodiment, the step of inputting the physical information map into a pre-trained graph neural network model for message passing and feature aggregation to fuse the spatially correlated local symptom feature vectors and generate a symptom map reflecting the current dynamic behavior pattern of the caisson-soil system includes: For each node in the physical information graph, the local symptom feature vector information transmitted from neighboring nodes is weighted and aggregated according to the physical attribute weight assigned to each edge to obtain aggregated information. The local symptom feature vector carried by each node is fused with the aggregated information, and the feature representation of the node is updated using a nonlinear transformation function to obtain the updated node feature representation; wherein, the fusion process introduces momentum conservation constraints so that the updated node feature representation conforms to the balance trend of action and reaction forces between adjacent units. Repeat the weighted aggregation and fusion update steps at least once; After repeated updates, the final feature representations of all nodes are globally pooled and integrated to generate the symptom map.

[0010] In one embodiment, the steps of using the system state reflected by the symptom map as initial conditions to drive a parallel digital twin model to perform forward inference, predict the evolution trend of the system state within a preset inference period, and obtain the inference prediction results include: Based on the symptom map, multiple parallel digital twin instances with parameters perturbed within a preset range are generated. The symptom map is decoded into the initial state parameters for each of the parallel digital twin instances; Within the preset simulation time, all the parallel digital twin instances are run synchronously to simulate the dynamic evolution of the caisson-soil system under their respective parameters. At each simulation time step, the virtual symptom characteristics generated by each of the parallel digital twin instances are recorded; After the simulation is completed, based on the virtual symptom features recorded at all simulation time steps, all virtual symptom features generated by each of the parallel digital twin instances during the entire simulation duration are summarized to form the virtual symptom evolution sequence of that instance; the virtual symptom evolution sequences of all instances constitute the simulation prediction result.

[0011] In one embodiment, the method further includes: During the digital twin model derivation process, the virtual symptom characteristics generated by the parallel digital twin instance simulation are monitored in real time; Based on the first comparison result, the theoretical instability mode with the highest similarity ranking obtained by matching the symptom map with the knowledge base is set as the target verification mode. In the simulation and prediction results, the parallel digital twin instances whose simulated endpoint state is unstable and whose instability mode is consistent with the target verification mode are selected as the verification instance set; The common parameter perturbations of the verification instance set during the inference process are analyzed in reverse. These common parameter perturbations are used to corroborate or correct the speculation on the current risk root cause, and the analysis results are used for the risk assessment.

[0012] In one embodiment, reverse analysis of the common parameter perturbations of the verification instance set during the inference process includes: Extract the parameter perturbation applied to each instance in the verification instance set at the initial moment of the simulation. The parameter perturbation includes at least one of the following: local strength reduction coefficient of soil, change in friction coefficient between soil and well wall, and change in pore water pressure in a preset area. Cluster analysis is performed on all the extracted parameter perturbations to identify at least one common perturbation cluster that is spatially clustered and numerically consistent. Each identified common disturbance cluster is mapped to its corresponding spatial location in the physical information map; Based on the type, magnitude, and spatial location of the parameter disturbances contained in each common disturbance cluster, the most likely physical risk source leading to the current symptom map and the target verification mode is queried and inferred from the preset engineering disease cause comparison table. The inferred physical risk sources and their corresponding confidence levels are used as the analysis results.

[0013] In one embodiment, the step of performing a risk assessment by combining the first comparison result and the inference prediction result to generate a corresponding intelligent diagnostic report includes: From the first comparison results, the instability pattern with the highest similarity is selected as the first candidate pattern; Analyze the virtual symptom evolution sequence of all parallel digital twin instances in the inference and prediction results, identify the instances that cause the structural state to exceed the safety threshold, and count the instability modes corresponding to these instances. The instability mode with the highest frequency is selected as the second candidate mode. Compare the first candidate pattern with the second candidate pattern: If the two are consistent, the mode is determined as the comprehensive judgment instability mode, and the corresponding similarity is combined with the inferred risk ratio to determine the risk level; If the two are inconsistent, check the consistency between the simulation parameters of the parallel digital twin instances corresponding to the first candidate mode and the second candidate mode and the current construction record, and refer to the significance of the preset symptoms in the symptom map to finally determine the comprehensive judgment instability mode and the risk level. Based on the determined comprehensive judgment instability mode, a pre-set list of potential risk sources is retrieved from the instability mode knowledge base; Based on the spatial distribution intensity of abnormal symptoms in the symptom map and the perturbation parameter type of the parallel digital twin instance whose simulated instability mode is determined to be the second candidate mode, the inferred risk source is obtained by filtering and sorting from the potential risk source list. Based on the inferred risk root causes and the comprehensive determination of the instability mode, at least one of the targeted treatment suggestions is matched from the measure library, and the system response after executing each of the targeted treatment suggestions is simulated using the parallel digital twin instance that has determined the comprehensive determination of the instability mode. The simulation results are attached to the report as an explanation of the expected effect of the suggestion. The intelligent diagnostic report is generated by integrating the risk level, the comprehensive judgment of the instability mode, the inferred risk root cause, and the targeted treatment suggestions with expected effects.

[0014] Furthermore, to achieve the above objectives, this application also proposes a safety monitoring system for caisson construction, the safety monitoring system for caisson construction comprising: a memory, a processor, and a safety monitoring program for caisson construction stored in the memory and executable on the processor, the safety monitoring program for caisson construction being configured to implement the steps of the safety monitoring method for caisson construction.

[0015] The safety monitoring method and system proposed in this application for caisson construction constructs a full-link safety monitoring system from real-time data acquisition and multi-scale feature extraction to dynamic risk assessment through collaborative perception of multi-physics field sensing and edge computing, intelligent fusion of spatial features of physical information graphs and graph neural networks, and continuous evolution mechanism of parallel inference of digital twins and federated learning. It breaks through the limitations of traditional monitoring methods such as data fragmentation, early warning lag, and ambiguous risk positioning. By deeply coupling the micro-activity signs of soil with the dynamic response of the structure, it realizes early identification, accurate source tracing, and dynamic handling of the instability risk of the caisson-soil system, thereby improving the safety management level and risk response capability of caisson construction under complex geological conditions. Attached Figure Description

[0016] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1This is a flowchart illustrating an embodiment of the safety monitoring method for caisson construction provided in this application; Figure 2 This is a structural schematic diagram of an embodiment of the safety monitoring system for caisson construction provided in this application.

[0019] Explanation of icon numbers: 10. Memory; 20. Processor.

[0020] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0021] The technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of this application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0022] It should be understood that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, the terms "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0023] In existing technologies, safety monitoring methods for caisson construction mostly rely on post-event alarms triggered by exceeding thresholds of macroscopic physical quantities. This approach suffers from problems such as monitoring lag, short emergency response windows, difficulty in diagnosing the root causes of abnormal data, lack of targeted response measures, and low efficiency. This passive response mode is insufficient to meet the needs of caisson construction for advanced early warning and accurate diagnosis.

[0024] Based on this, the embodiments of this application provide a safety monitoring method for caisson construction, referring to... Figure 1 The safety monitoring method for caisson construction includes steps S100 to S600, wherein: Step S100: Simultaneously collect multi-physics field raw monitoring data during construction by deploying multiple integrated sensing units in the caisson structure and surrounding soil; the multi-physics field raw monitoring data includes acoustic emission signals, vibration signals and soil contact pressure signals. Step S200: Perform edge computing processing on the original multiphysics field monitoring data collected by each integrated sensing unit to extract local symptom feature vectors that characterize the soil micro-activity state and stress change trend at the monitoring location of the integrated sensing unit. Step S300: Upload the local symptom feature vectors from all integrated sensing units to the cloud analysis platform. In the cloud analysis platform, based on the three-dimensional model of the caisson structure, the spatial location information of each integrated sensing unit, and the preset geological stratification information, construct a physical information map. The nodes of the physical information map represent soil or structural units and carry the corresponding local symptom feature vectors. The edges of the map represent the spatial adjacency relationship between units and are assigned physical attribute weights based on the geological and construction conditions. Step S400: Input the physical information map into a pre-trained graph neural network model for message passing and feature aggregation, so as to fuse the spatially associated local symptom feature vectors and generate a symptom map reflecting the current dynamic behavior pattern of the caisson-soil system. Step S500: The symptom map is matched and compared with the preset well instability mode knowledge base in real time to obtain the first comparison result. The system state reflected by the symptom map is used as the initial condition to drive the parallel digital twin model to perform forward inference, predict the evolution trend of the system state within the preset inference time, and obtain the inference prediction result. Step S600: Perform a risk assessment by combining the first comparison result and the inference and prediction result to generate a corresponding intelligent diagnostic report; the intelligent diagnostic report includes the risk level, a comprehensive judgment of the instability mode, an inferred risk root cause, and targeted handling suggestions.

[0025] In this embodiment, the integrated sensing unit refers to an intelligent monitoring device that integrates multiple types of sensors, designed to simultaneously collect raw data of different physical quantities. In the caisson construction scenario, these units are deployed in the caisson structure itself and the surrounding soil to obtain comprehensive environmental information. Multiphysics raw monitoring data refers to various physical signals collected by the integrated sensing unit that reflect the state of the caisson-soil system. Specifically, acoustic emission signals can reflect micro-fractures or frictional activity within the material; vibration signals can reveal the dynamic response and energy release of the soil or structure; and soil contact pressure signals directly quantify the force distribution of the soil on the caisson structure. These data collectively constitute a comprehensive description of the system's microscopic behavior. Edge computing processing refers to the technology of performing preliminary processing and analysis of raw data at the data source (i.e., the integrated sensing unit or its vicinity). Its purpose is to reduce data transmission volume, lower the cloud computing load, and achieve faster response speeds. Through edge computing, key information can be extracted from massive amounts of raw data.

[0026] In this embodiment, the local symptom feature vector refers to a set of numerical features extracted from the raw multiphysics monitoring data collected from a single integrated sensing unit through edge computing processing. These features are designed to characterize the micro-activity state (such as micro-fractures, plastic deformation) and stress change trends of the soil at the monitoring location, and are a quantitative representation of anomalous behavior in a local area. The cloud analysis platform refers to a centralized data processing and analysis system based on a cloud computing architecture. This platform receives local symptom feature vectors from all integrated sensing units and utilizes its powerful computing capabilities for advanced analysis, such as building complex models, performing large-scale data fusion, and pattern recognition. The physical information graph is a data model that abstracts the caisson-soil system into a graph structure. In this graph, nodes represent physical entities in the system, such as soil units or structural units, and carry their corresponding local symptom feature vectors. Edges represent the spatial adjacency relationships between these physical entities and are assigned physical attribute weights based on geological conditions and construction status to reflect the strength or importance of different connections.

[0027] In this embodiment, the graph neural network model is a deep learning model specifically designed for processing graph-structured data. Its core mechanisms include message passing and feature aggregation. Message passing refers to the process of nodes exchanging information through edges; feature aggregation refers to each node integrating received neighbor information and its own features to update its feature representation. Through these operations, the model can capture the complex spatial relationships within the graph structure. The symptom map is an abstract representation of the current overall dynamic behavior pattern of the caisson-soil system, generated after processing the physical information graph using the graph neural network model. This map integrates the spatial correlation information of all local symptom feature vectors, revealing possible abnormal patterns or potential precursors to instability at the macroscopic level.

[0028] In this embodiment, the caisson instability mode knowledge base is a pre-built database containing various known caisson instability types and their typical symptom patterns. This knowledge base is used to compare with the real-time generated symptom map to identify the instability risk that may correspond to the current system state. The digital twin model refers to a virtual, dynamic, high-fidelity model of the caisson-soil system. This model can reflect the state of the physical entity in real time and can perform forward extrapolation to simulate the system's behavioral evolution under different conditions. In safety monitoring, it is used to predict the evolution trend of the system state over a future period. The intelligent diagnostic report is a comprehensive document automatically generated by the system after completing the risk assessment. This report not only includes the risk level, a comprehensive judgment of the instability mode, and a predicted risk root cause, but also provides targeted treatment suggestions, providing a scientific basis for engineering decisions.

[0029] In this embodiment, the safety monitoring method for caisson construction firstly involves synchronously collecting raw multi-physics monitoring data during the construction process using multiple integrated sensing units deployed in the caisson structure and surrounding soil. This raw multi-physics monitoring data includes acoustic emission signals, vibration signals, and soil contact pressure signals. For example, the integrated sensing unit can be designed to include independent acoustic emission sensors, vibration sensors, and pressure sensors, each connected to a data acquisition module via its own wiring. The data acquisition module can sequentially poll each sensor at preset time intervals to acquire its output signal. For example, it can first collect all acoustic emission signals, then all vibration signals, and finally all pressure signals, and then package and send this data.

[0030] In this embodiment, edge computing processing is performed on the raw multiphysics monitoring data collected by each integrated sensing unit to extract local symptom feature vectors characterizing the soil micro-activity state and stress change trend at the monitoring location of the integrated sensing unit. Specifically, edge computing processing can be executed on the microcontroller inside the integrated sensing unit. For example, for acoustic emission signals, the event count per unit time can be directly calculated; for vibration signals, their root mean square value can be calculated; and for soil contact pressure signals, their current value can be recorded. These simple statistics can be combined into a preliminary feature vector.

[0031] In this embodiment, the local symptom feature vectors from all integrated sensing units are uploaded to a cloud analysis platform. On this platform, a physical information map is constructed based on the 3D model of the caisson structure, the spatial location information of each integrated sensing unit, and pre-defined geological stratification information. Nodes in this physical information map represent soil or structural units and carry the corresponding local symptom feature vectors. Edges represent spatial adjacency relationships between units and are assigned physical attribute weights based on geological and construction conditions. For example, the local symptom feature vectors can be periodically sent to the cloud analysis platform via a wireless communication module. The cloud can pre-store the 3D geometric model of the caisson structure, the coordinates of each integrated sensing unit, and a static geological survey report. The physical information map can be constructed by using the location of each sensing unit as a node and determining adjacency based on their Euclidean distance in 3D space. For example, two nodes with a distance less than a certain threshold are considered to have an edge. The physical attribute weights of the edges can be simply set to a fixed value, or different fixed weights can be assigned based on the type of soil layer in the geological report.

[0032] Furthermore, the physical information map is input into a pre-trained graph neural network model for message passing and feature aggregation to fuse the spatially correlated local symptom feature vectors, generating a symptom map reflecting the current dynamic behavior pattern of the caisson-soil system. For example, the pre-trained graph neural network model can be a simple graph convolutional network. This model receives the physical information map as input, where the node features are local symptom feature vectors, and the edge weights are physical attribute weights. Message passing can be achieved by simply summing or averaging the features of neighboring nodes, then concatenating the aggregated information with the node's own features. Feature aggregation can be performed using a single-layer perceptron to perform a non-linear transformation on the concatenated features, thereby updating the node's feature representation. After several iterations, the final feature representations of all nodes can be flattened and connected to form the symptom map.

[0033] In this embodiment, the symptom map is matched and compared in real time with a preset knowledge base of well instability modes to obtain a first comparison result. Using the system state reflected in the symptom map as initial conditions, a parallel digital twin model is driven to perform forward inference, predicting the evolution trend of the system state within a preset inference period, and obtaining the inference prediction result. For example, the matching between the symptom map and the well instability mode knowledge base can be achieved by calculating the Euclidean distance or cosine similarity between the symptom map and the typical maps of each instability mode stored in the knowledge base. The mode with the highest similarity is considered the first comparison result. The digital twin model can be a simplified mechanical model based on the finite element or discrete element method, whose initial conditions are roughly mapped from the key features in the symptom map. Within the preset inference period, the model performs simulation calculations at fixed time steps, recording the macroscopic physical quantities at each time step; these records constitute the inference prediction result.

[0034] In this embodiment, a risk assessment is performed by combining the first comparison result and the inference and prediction result to generate a corresponding intelligent diagnostic report. This intelligent diagnostic report includes a risk level, a comprehensive assessment of the instability mode, a hypothesized root cause of the risk, and targeted remedial suggestions. For example, the risk assessment could directly use the instability mode with the highest similarity in the first comparison result as the comprehensive assessment of the instability mode. The inference and prediction result can be used to determine whether the system will reach a preset danger threshold within the inference period; if it does, the risk level is increased. The hypothesized root cause of the risk can be simply selected from a preset list of root causes associated with the comprehensive assessment of the instability mode in the instability mode knowledge base. The targeted remedial suggestions can also be general suggestions associated with the instability mode in the knowledge base. This information is integrated into a text template to generate the intelligent diagnostic report.

[0035] In this embodiment, by combining multiphysics data acquisition, edge computing, physical information graph construction, graph neural network analysis, and digital twin inference, advanced perception and accurate diagnosis of microscopic activities and macroscopic instability precursors of the caisson-soil system are achieved. This effectively overcomes the shortcomings of traditional monitoring methods, such as monitoring lag and ambiguous diagnosis, providing timely and targeted risk warnings and handling suggestions for caisson construction, thereby improving construction safety and efficiency.

[0036] In one feasible implementation, the step of synchronously collecting multi-physics field raw monitoring data during construction by deploying multiple integrated sensing units in the caisson structure and surrounding soil further includes: controlling each integrated sensing unit to synchronously collect the multi-physics field raw monitoring data at a first preset acquisition frequency; monitoring the rate of change of the newly collected multi-physics field raw monitoring data of each integrated sensing unit relative to the data of the previous period; when the rate of change of any physical field data of any integrated sensing unit exceeds its corresponding first threshold, determining that an initial abnormality has occurred at that location, and switching the data acquisition frequency of that integrated sensing unit to a frequency higher than the first preset acquisition frequency. The system uses a second preset acquisition frequency and, centered on this unit, notifies adjacent integrated sensing units to switch their data acquisition frequency to the second preset acquisition frequency based on the pre-defined adjacency relationship in the physical information map. Data is acquired at the second preset acquisition frequency, and the acquired signals in preset frequency bands related to soil micro-fractures or plastic flow are analyzed in a directional manner. When the data change rate of all units that have switched to the second preset acquisition frequency is lower than their corresponding second threshold and remains so for a preset time, the abnormal signs are determined to have subsided, and the data acquisition frequency of the relevant units is switched back to the first preset acquisition frequency. The second threshold is lower than the first threshold.

[0037] In this embodiment, each integrated sensing unit is controlled to synchronously acquire the raw multiphysics monitoring data at a first preset acquisition frequency, aiming to establish a normalized, low-power data acquisition mode. In this mode, the integrated sensing units are typically configured to operate in a low-power state and periodically wake up at a preset low frequency (e.g., every few minutes or at a low Hertz rate) to acquire raw multiphysics monitoring data such as acoustic emission signals, vibration signals, and soil contact pressure signals. Synchronous acquisition can be achieved through a unified clock synchronization mechanism or Network Time Protocol (NTP) to ensure that the data acquired by different sensing units are time-aligned, laying the foundation for subsequent joint analysis.

[0038] In this embodiment, monitoring the rate of change of the newly acquired multiphysics monitoring data of each integrated sensing unit relative to the data of the previous period is for the purpose of real-time assessment of the stability of the system state. This rate of change can be calculated on the edge processor of the integrated sensing unit or after the data is uploaded to the local gateway. For each type of physical field data, the system calculates the difference or percentage change between the data of the current acquisition period (e.g., the root mean square value of the vibration signal, the absolute value of the soil contact pressure) and the corresponding data of the previous acquisition period. For example, the rate of change can be expressed as (current value - previous period value) / previous period value × 100%, or directly as the absolute difference, to reflect the severity of data fluctuations.

[0039] In this embodiment, when the rate of change of any physical field data of any integrated sensing unit exceeds its corresponding first threshold, an initial abnormality is determined to have occurred at that location. The system then switches the data acquisition frequency of that integrated sensing unit to a second preset acquisition frequency higher than the first preset acquisition frequency. Furthermore, based on the pre-defined adjacency relationships in the physical information map, neighboring integrated sensing units are notified to also switch their data acquisition frequencies to the second preset acquisition frequency. This step is a crucial mechanism for the system to respond to anomalies. The first threshold, determined based on historical data analysis or engineering experience, represents the upper limit of the normal fluctuation range. Once the rate of change of data of a sensing unit exceeds this threshold, it indicates that there may be abnormal activity at that location. At this time, the sensing unit will immediately increase its data acquisition frequency to a higher second preset acquisition frequency (e.g., from the Hertz level to the kilohertz level) to capture more refined transient information. Simultaneously, the system uses the pre-constructed physical information map to query the neighboring units of the abnormal sensing unit and sends instructions to these neighboring units via wireless communication protocols (such as wireless mesh networks, LoRa, or Zigbee) to also switch them to the second preset acquisition frequency. This localized, adaptive high-frequency monitoring strategy helps to capture the spatial propagation characteristics and detailed evolution of anomalous signs.

[0040] In this embodiment, data is acquired at the second preset acquisition frequency, and the acquired signals in preset frequency bands related to soil micro-fractures or plastic flow are subjected to directional analysis to delve deeper into the physical nature of abnormal signs. At higher acquisition frequencies, the integrated sensing unit can acquire richer and more detailed raw data. Directional analysis typically involves preset digital signal processing techniques. For example, for acoustic emission signals, bandpass filtering can be performed to separate preset high-frequency bands (e.g., 50-500kHz) related to microcrack propagation, or event detection algorithms can be applied to identify micro-fracture events. For vibration signals, their spectral characteristics can be analyzed to identify low-frequency vibration modes or preset frequency band energy changes related to soil liquefaction or plastic deformation. These analyses are typically performed at the edge to achieve rapid response and reduce data transmission volume.

[0041] In this embodiment, when the data change rate of all units that have switched to the second preset acquisition frequency is lower than their corresponding second threshold and remains so for a preset time, the abnormal symptoms are determined to have subsided, and the data acquisition frequency of the relevant units is switched back to the first preset acquisition frequency; wherein, the second threshold is lower than the first threshold. This step ensures that the system can recover to an efficient normal monitoring mode after the abnormal situation is alleviated. The second threshold is set to be lower than the first threshold, introducing a hysteresis effect and avoiding frequent switching of acquisition frequencies due to small fluctuations. The preset time ensures the stability and continuity of the system state, rather than a temporary recovery. When the data change rate of all affected sensing units meets this condition and remains so for a period of time, the system sends an instruction to them to switch their data acquisition frequency back to the first preset acquisition frequency, thereby optimizing the efficiency of system resource utilization.

[0042] In this embodiment, through the above-described technical solution, this application achieves dynamic adaptive data acquisition for caisson construction safety monitoring. Under normal conditions, the system operates at a lower frequency, effectively saving energy and computing resources and reducing data transmission burden. Once a local abnormality is detected, the system can quickly and locally increase the monitoring frequency of the relevant area, ensuring that high-resolution transient data is captured at critical moments, thereby avoiding early warning information that may be missed due to fixed low-frequency acquisition. At the same time, directional analysis performed at a second preset acquisition frequency can more accurately identify preset signal features related to soil micro-fractures or plastic flow, providing more targeted raw data for subsequent symptom feature vector extraction. When the abnormality subsides, the system can intelligently adjust the acquisition frequency back to normal, avoiding unnecessary resource waste. This dynamic adjustment mechanism enables the system to improve operational efficiency and resource utilization while ensuring monitoring sensitivity, providing high-quality and timely data support for the cloud analysis platform to build more accurate physical information maps and symptom maps, thereby improving the response speed and diagnostic accuracy of the entire safety monitoring method.

[0043] In one feasible implementation, the step of performing edge computing processing on the original multiphysics monitoring data collected by each integrated sensing unit to extract a local symptom feature vector characterizing the soil micro-activity state and stress change trend at the monitoring location of the unit includes: performing event detection and waveform analysis on the acoustic emission signal to extract the acoustic emission event rate, cumulative event energy value, and dominant frequency offset to obtain an acoustic feature set; performing spectrum analysis and energy calculation on the vibration signal to extract the vibration dominant frequency, the energy proportion of a preset frequency band, and the signal entropy value to obtain a vibration feature set; performing spatial gradient and time series analysis on the soil contact pressure signal to calculate the rate of change of the current pressure value relative to the historical baseline, the pressure gradient value between adjacent sensors, and the pressure distribution entropy to obtain a pressure feature set; and concatenating the acoustic feature set, the vibration feature set, and the pressure feature set in a predetermined order to form the local symptom feature vector.

[0044] In this embodiment, event detection and waveform analysis are performed on the acoustic emission signal to identify effective transient events generated by activities such as micro-fractures, crack propagation, or particle friction within the soil from the continuous raw acoustic emission signal, and to quantify their waveform characteristics. Event detection can be achieved by setting adaptive threshold triggering, short-time energy to long-time energy ratio (STA / LTA) algorithms, or wavelet transform to effectively distinguish real events from environmental noise. Waveform analysis can extract parameters such as rise time, peak amplitude, and duration of the event. Based on this, the acoustic emission event rate is extracted, which is the number of acoustic emission events detected per unit time. This parameter directly reflects the frequency and activity level of soil micro-fractures. The cumulative event energy value refers to the sum of energy released by all acoustic emission events within a certain monitoring period, which can characterize the degree of accumulated damage and energy dissipation within the soil. The dominant frequency offset refers to the frequency change of the main frequency component of the acoustic emission signal relative to its normal or reference state. This change may indicate changes in soil material properties, crack propagation modes, or stress states. These parameters together constitute an acoustic feature set, which characterizes the microscopic damage behavior of soil from the perspectives of energy release and frequency response.

[0045] In this embodiment, spectral analysis and energy calculation are performed on the vibration signal to convert the time-domain vibration signal into a frequency-domain representation, thereby revealing its inherent frequency components and energy distribution. Spectral analysis typically employs methods such as Fast Fourier Transform (FFT) or power spectral density estimation to obtain the energy distribution of the signal at different frequencies. Based on this, the dominant vibration frequency, i.e., the frequency component with the most concentrated energy in the vibration signal, is extracted. Its changes may reflect the inherent frequency changes of the soil or caisson structure, thus indicating changes in its stiffness or constraint conditions. The preset frequency band energy ratio refers to the proportion of vibration energy to total energy within a preset frequency range related to phenomena such as soil loosening, structural resonance, or abnormal deformation. This helps to focus on vibration modes related to potential risks. The signal entropy value is used to measure the complexity and randomness of the vibration signal. When abnormal changes occur in the soil or structure, the regularity of the vibration signal may be broken, leading to a significant change in the entropy value. These parameters together constitute the vibration feature set, reflecting the overall stability of the soil from the perspective of macroscopic dynamic response.

[0046] In this embodiment, spatial gradient and time series analysis are performed on the soil contact pressure signal to capture the dynamic changes and local inhomogeneities of the soil stress state from the pressure data. Time series analysis focuses on the trend of pressure change over time, calculating the rate of change of the current pressure value relative to the historical baseline. This rate of change intuitively reflects whether the soil stress is increasing, decreasing, or remaining stable, as well as the rate and magnitude of change. The historical baseline can be determined by the average value of long-term monitoring data, a trend line, or a stable value for a preset construction stage. Spatial gradient analysis focuses on the pressure difference between adjacent sensors, calculating the pressure gradient value between adjacent sensors. This gradient value can reveal local stress concentration areas or stress unloading areas, helping to determine the soil deformation pattern, such as whether there is local heave or settlement. Pressure distribution entropy is used to measure the uniformity of pressure distribution in space. When local instability or plastic flow occurs in the soil, the uniformity of pressure distribution may be disrupted, leading to changes in entropy values. These parameters together constitute a pressure feature set, reflecting the macroscopic mechanical behavior of the soil from the perspective of stress distribution and change.

[0047] In this embodiment, the acoustic feature set, the vibration feature set, and the pressure feature set are concatenated in a predetermined order to form the local symptom feature vector. This step aims to integrate features with different physical meanings extracted from different physical field data to form a unified, structured data representation. The predetermined concatenation order ensures that the feature vectors generated by each integrated sensing unit have consistent dimensions and feature arrangement, which is crucial for subsequent message passing and feature aggregation in the graph neural network model. In this way, each local symptom feature vector can comprehensively and compactly encode the soil micro-activity state and stress change trend at the monitoring location of the integrated sensing unit, providing high-quality input for subsequent system-level analysis.

[0048] In this embodiment, event detection and waveform analysis of acoustic emission signals can capture the frequency of micro-fractures, energy release, and changes in damage mechanisms within the soil. Spectral analysis and energy calculation of vibration signals reflect the overall dynamic response and local loosening of the soil or structure. Spatial gradient and time series analysis of soil contact pressure signals reveal stress concentration, unloading, and non-uniformity of pressure distribution. These features complementarily characterize the micro-activity state and stress change trends of the soil from different physical perspectives, avoiding the limitations of a single data source. By splicing these refined acoustic, vibration, and pressure feature sets in a predetermined order, a comprehensive and structured local symptom feature vector is formed. This provides high-quality input for message transmission and feature aggregation in the subsequent construction of physical information maps and graph neural network models on the cloud analysis platform, thereby improving the accuracy of caisson-soil system state identification and the reliability of early warning.

[0049] In one feasible implementation, the steps of uploading the local symptom feature vectors from all integrated sensing units to a cloud analysis platform, and constructing a physical information map on the cloud analysis platform based on the three-dimensional model of the caisson structure, the spatial location information of each integrated sensing unit, and the preset geological stratification information, include: determining a first weight coefficient for the edges between corresponding nodes of the integrated sensing units at the interface of different soil layers according to the preset geological stratification information; calculating a second weight coefficient for the edges between corresponding nodes of the integrated sensing units in the contact area between the well wall and the soil according to the sinking depth and real-time attitude data of the caisson in the current construction stage; combining the first weight coefficient and the second weight coefficient to obtain the physical attribute weight; and combining the nodes carrying the local symptom feature vectors and the edges assigned the physical attribute weights to form the physical information map.

[0050] In this embodiment, based on preset geological stratification information, the first weighting coefficient of the edges between corresponding nodes of integrated sensing units at the boundaries of different soil layers is determined. This step aims to quantify the influence of the interface between different soil layers on the physical connection between soil units. Geological stratification information reflects the heterogeneity of the soil; the mechanical properties of different soil layers (such as sand, clay, silt, etc.) differ significantly, and their boundaries are often stress concentration or deformation-sensitive areas. By assigning preset weighting coefficients to the edges between nodes (representing integrated sensing units) at these boundaries, the physical connection strength or sensitivity of these key areas can be highlighted. In specific implementation, different first weighting coefficient values ​​can be set in advance for different types of soil layer boundaries (e.g., sand and clay boundary, soft soil and hard soil boundary) based on geological survey reports and soil mechanics parameters. For example, for soil layer boundaries with large differences in mechanical properties, a higher weighting coefficient can be assigned, indicating that the physical correlation of the area is stronger or more likely to interact; conversely, for soil layer boundaries with similar properties, a lower weighting coefficient can be assigned. These coefficients can be stored in a database and queried and assigned values ​​when constructing a physical information map based on the correspondence between the spatial location of the integrated sensing units and geological stratification information.

[0051] In this embodiment, based on the sinking depth and real-time attitude data of the caisson at the current construction stage, a second weighting coefficient is calculated for the edges between corresponding nodes of the integrated sensing unit in the contact area between the caisson wall and the soil. This step aims to dynamically reflect the intensity and pattern of the interaction between the caisson wall and the surrounding soil during the caisson construction process. The sinking depth and real-time attitude (such as tilting or deviation) directly affect the contact state between the caisson wall and the soil, the distribution of frictional force, and the redistribution of soil stress. By introducing a second weighting coefficient, the connection strength of the caisson wall-soil interface area in the physical information diagram can be adjusted in real time, making it more accurately reflect the current construction state and mechanical response. In specific implementation, an empirical or numerical model based on construction stage parameters (such as sinking depth, sinking speed, and correction measures) and real-time attitude data (such as data obtained through tilt sensors and displacement sensors) can be established to calculate the second weighting coefficient of the contact area between the caisson wall and the soil. For example, when the caisson sinks at a faster speed or when there is a large deviation in attitude, the frictional force or squeezing effect between the caisson wall and the soil may increase, and a higher second weighting coefficient can be assigned in this case; conversely, in the stable sinking stage, a relatively lower weight can be assigned. The calculation process can be a real-time function that takes current construction parameters and attitude data as input and outputs corresponding weight values.

[0052] In this embodiment, the first weighting coefficient and the second weighting coefficient are synthesized to obtain the physical attribute weight. This step aims to comprehensively consider the dual influence of geological conditions and construction status on the edge weights of the physical information map, forming a unified weight that fully reflects the physical correlation of the caisson-soil system. A single weighting coefficient may not be able to fully capture complex interactions. Through synthesis, it can be ensured that the physical attribute weight includes both static geological characteristics and dynamic construction influences. The synthesis method can adopt various approaches such as weighted average, product, maximum value selection, or rule-based fusion. For example, a weighted average method can be used: Physical attribute weight = α × first weighting coefficient + β × second weighting coefficient, where α and β are preset weighting factors that reflect the relative importance of geological factors and construction factors in a preset scenario, and α + β = 1. A product method can also be used, where the product can amplify the influence when both factors are important. The specific synthesis strategy should be optimized based on engineering experience and numerical simulation results to ensure that the synthesized weight can more accurately represent the physical attributes between units.

[0053] In this embodiment, nodes carrying the local symptom feature vectors and edges assigned the physical attribute weights are combined to form the physical information graph. This is the final step in constructing the physical information graph, integrating the previously processed node features and edge weights to form a complete graph structure. This graph forms the basis for subsequent analysis by the graph neural network model. It represents in a structured manner the local state (local symptom feature vector) of each monitoring point (node) in the caisson-soil system and the strength of their physical associations (physical attribute weights). In the cloud analysis platform, a graph database or graph computing framework can be used to store and manage the physical information graph. The spatial location of each integrated sensing unit corresponds to a node in the graph, and this node stores its corresponding local symptom feature vector. The edges between nodes are connected according to a preset adjacency relationship (e.g., spatial distance less than a certain threshold, or belonging to the same soil layer / structural unit) and assigned the previously calculated physical attribute weights. This graph structure is dynamic, and its edge weights are updated in real time as the construction status changes.

[0054] In this embodiment, through the above technical solution, this application can construct a more refined and dynamic physical information map. This map not only statically reflects the physical correlation differences brought about by geological stratification, but also dynamically incorporates real-time construction information such as the sinking depth and attitude of the caisson. This allows the edge weights of the map to accurately characterize the true physical connection strength of the caisson-soil system under different construction stages and complex geological conditions. This improves the physical information map's ability to represent the dynamic behavior patterns of the caisson-soil system, providing more reliable input for subsequent message passing and feature aggregation in the graph neural network model. This enables more accurate fusion of spatially correlated local symptom feature vectors, generating a symptom map reflecting the current dynamic behavior pattern of the system, thereby improving the accuracy and real-time performance of caisson instability mode identification and risk assessment.

[0055] In one feasible implementation, the step of inputting the physical information graph into a pre-trained graph neural network model for message passing and feature aggregation to fuse the spatially correlated local symptom feature vectors and generate a symptom map reflecting the current dynamic behavior pattern of the caisson-soil system includes: for each node in the physical information graph, weighted aggregation of the local symptom feature vector information transmitted from neighboring nodes is performed according to the physical attribute weights assigned to each edge to obtain aggregated information; the local symptom feature vector carried by each node is fused with the aggregated information, and the feature representation of the node is updated using a nonlinear transformation function to obtain an updated node feature representation; wherein, the fusion process introduces momentum conservation constraints so that the updated node feature representation conforms to the balance trend of action and reaction forces between adjacent units; the weighted aggregation and fusion update steps are repeated at least once; after repeated updates, the final feature representations of all nodes are globally pooled to integrate and generate the symptom map.

[0056] In this embodiment, for each node in the physical information graph, the local symptom feature vector information transmitted from adjacent nodes is weighted and aggregated according to the physical attribute weights assigned to each edge, to obtain aggregated information. In this step, the physical attribute weights are determined based on factors such as geological stratification information, caisson sinking depth, and real-time attitude data, reflecting the physical connection strength and influence range at the boundaries of different soil layers or in the contact area between the well wall and the soil. For example, when an edge connects different soil layers, its weight may be lower, indicating obstructed information transmission; while when an edge connects an area where the well wall and soil are in close contact, its weight may be higher, indicating more direct and significant information transmission. This weighted aggregation can be implemented through weighted averaging, weighted summation, or an aggregation function based on an attention mechanism, ensuring that the message transmission process truly reflects the heterogeneity of physical connections.

[0057] In this embodiment, the local symptom feature vector carried by each node is fused with the aggregated information, and the feature representation of the node is updated using a nonlinear transformation function to obtain the updated node feature representation. During this fusion process, a momentum conservation constraint is introduced, ensuring that the updated node feature representation conforms to the balance trend of action and reaction forces between adjacent units. This momentum conservation constraint can be implemented by adding a regularization term to the loss function of the graph neural network, which penalizes feature updates that do not meet physical equilibrium conditions. For example, a physical constraint layer can be designed to adjust the updated features after feature fusion, ensuring that they satisfy some form of mechanical equilibrium condition locally or globally, such as the discrete form of Newton's third law. This ensures that the features learned by the model are not only data-driven but also physically plausible, more accurately reflecting the stress transmission and response within the caisson-soil system. Nonlinear transformation functions, such as ReLU, Sigmoid, or Tanh, are used to enhance the model's expressive power, enabling it to capture complex nonlinear relationships.

[0058] In this embodiment, to capture a wider range of spatial correlations, the aforementioned weighted aggregation and fusion update steps will be repeated at least once. Through multi-layer message passing and feature updates, each node can receive information from its more distant neighbors, thereby gradually fusing the global information of the entire caisson-soil system. The number of repetitions is typically determined based on the diameter of the graph and the complexity of the task to ensure sufficient information propagation. After repeated updates, the final feature representations of all nodes are subjected to global pooling to generate the symptom map. The global pooling operation aims to aggregate the local features of all nodes into a global, fixed-dimensional vector, which is the symptom map, capable of comprehensively reflecting the current dynamic behavior pattern of the entire caisson-soil system. This operation can employ global average pooling, global max pooling, or more complex attention-based readout functions to effectively integrate the final feature vectors of all nodes.

[0059] In this embodiment, through the above technical solution, this application not only considers spatial correlation in the message passing and feature aggregation process of the graph neural network, but also innovatively introduces physical attribute weights for weighted aggregation, making information transmission more consistent with the heterogeneity of actual physical connections. Most importantly, a momentum conservation constraint is introduced during node feature fusion and updating, ensuring that the updated node feature representation is mechanically self-consistent and conforms to the balance trend of action and reaction forces between adjacent units. The introduction of this physical constraint enables the graph neural network model to learn more physically meaningful feature representations, thereby generating symptom maps that more accurately and reliably reflect the dynamic behavior patterns of the caisson-soil system. This effectively avoids the physically unreasonable predictions that may occur in purely data-driven models, improving the accuracy of caisson instability risk identification and the reliability of early warning.

[0060] In one feasible implementation, the steps of using the system state reflected by the symptom map as initial conditions to drive parallel digital twin models to perform forward extrapolation and predict the evolution trend of the system state within a preset extrapolation period to obtain the extrapolation prediction result include: based on the symptom map, generating multiple parallel digital twin instances with parameters perturbed within a preset range; decoding the symptom map to the initial state parameters of each of the parallel digital twin instances; synchronously running all the parallel digital twin instances within the preset extrapolation period to simulate the dynamic evolution process of the well-soil system under their respective parameters; recording the virtual symptom features simulated by each of the parallel digital twin instances at each extrapolation time step; after the extrapolation ends, summarizing all virtual symptom features generated by each of the parallel digital twin instances within the entire extrapolation period based on the virtual symptom features recorded at all extrapolation time steps to form the virtual symptom evolution sequence of that instance; the virtual symptom evolution sequences of all instances constitute the extrapolation prediction result.

[0061] In this embodiment, to address various uncertainties that may arise during the actual construction of the caisson-soil system, such as the heterogeneity of soil properties, deviations in construction operations, and fluctuations in environmental loads, this application generates multiple parallel digital twin instances with parameters perturbed within a preset range. A single digital twin model, when faced with these complex factors, may provide overly idealized predictions that fail to fully reflect potential risk scenarios. By introducing parameter perturbations, a prediction space that more closely approximates reality and possesses probabilistic distribution characteristics can be constructed. These parameter perturbations can target key physical parameters in the digital twin model, such as soil strength parameters (e.g., internal friction angle, cohesion), deformation parameters (e.g., elastic modulus, Poisson's ratio), permeability parameters, the interfacial friction coefficient between the well wall and the soil, and external loads (e.g., groundwater level changes, construction loads). The perturbation method can be random sampling (e.g., Monte Carlo method), Latin hypercube sampling, or sampling based on the statistical distribution of historical data. The preset range can be determined based on engineering experience, uncertainty analysis from geological survey reports, or sensitivity analysis results. For example, soil strength parameters can be randomly perturbed within a range of their mean ± one standard deviation to cover their possible variability.

[0062] In this embodiment, the symptom map is decoded into initial state parameters for each of the parallel digital twin instances. The symptom map is an abstract representation of the current dynamic behavior pattern of the caisson-soil system, containing local symptom feature vectors at the monitoring locations of each integrated sensing unit and their spatial relationships. To drive the digital twin model's simulation, this abstract, high-dimensional symptom map information needs to be transformed into specific physical initial state parameters that the digital twin model can understand and accept. The decoding process acts as a bridge between the monitoring data and the physical model, ensuring that the digital twin model starts its simulation from a starting point highly consistent with the current state of the actual system. The decoding process can be implemented using a pre-trained mapping function or a neural network. For example, a decoder network can be trained, taking the symptom map as input and outputting the initial stress field, initial displacement field, initial pore water pressure distribution, or correction values ​​for local material parameters required by the digital twin model. These initial state parameters should correspond to the soil micro-activity state and stress change trends reflected in the symptom map. For example, a high acoustic emission event rate and large vibration energy in a certain area of ​​the symptom spectrum may be decoded as the initial state of soil strength reduction, stress concentration, or local plastic deformation in that area.

[0063] In this embodiment, all parallel digital twin instances are run synchronously within a preset simulation period to simulate the dynamic evolution of the caisson-soil system under their respective parameters. After determining the initial state and disturbance parameters of each digital twin instance, they are simulated independently and simultaneously. Synchronous operation ensures that the diversity of system evolution under different parameter combinations and initial conditions can be observed at the same time scale, thereby comprehensively assessing potential risk paths. The digital twin model can be based on the finite element method, discrete element method, or a numerical simulation model coupled with a fluid-structure interaction mechanism. These models can simulate the elastoplastic deformation, failure, and seepage of the soil, as well as the response of the caisson structure. Synchronous operation can be achieved through a high-performance computing cluster or a distributed computing framework, with each computing node responsible for the simulation of one or more digital twin instances. The simulation period should be set according to the actual engineering needs and the evolution speed of potential instability modes; for example, it can be set to the next few hours, days, or even until the next critical construction phase.

[0064] In this embodiment, at each simulation time step, the virtual symptom features generated by the simulation of each parallel digital twin instance are recorded. During the simulation of the digital twin model, its internal state needs to be continuously monitored for subsequent analysis. The purpose of recording virtual symptom features is to transform the physical quantities (such as stress, strain, displacement, pore water pressure, etc.) within the digital twin model into a "virtual" form corresponding to the actual monitoring data (acoustic emission signals, vibration signals, soil contact pressure signals). This allows the simulation results to be compared with the actual monitoring data and provides a unified input for subsequent risk assessment. The generation of virtual symptom features can be achieved by simulating the response of integrated sensing units in the digital twin model. For example, stress wave propagation, local deformation, or contact pressure changes at sensor locations can be simulated in the digital twin model and converted into virtual acoustic emission signals, vibration signals, or soil contact pressure signals. These virtual signals can be further processed to extract virtual symptom features similar to the local symptom feature vector, such as virtual acoustic emission event rate, virtual vibration dominant frequency, virtual pressure change rate, etc.

[0065] In this embodiment, after the simulation is completed, based on the virtual symptom features recorded at all simulation time steps, all virtual symptom features generated by each parallel digital twin instance during the entire simulation duration are summarized to form the virtual symptom evolution sequence of that instance; the virtual symptom evolution sequences of all instances constitute the simulation prediction result. This step summarizes and standardizes the simulation results of a single digital twin instance, and ultimately forms a complete simulation prediction result. By summarizing the virtual symptom features of each instance during the entire simulation duration, a time series can be obtained, clearly showing the dynamic evolution trajectory of the caisson-soil system state under preset parameter perturbations. By combining the evolution sequences of all parallel instances, a comprehensive, multi-scenario prediction of the future system state evolution trend is formed. The summarization process can be a simple time series splicing, or it can be a statistical analysis of key symptom features (such as maximum value, average value, rate of change, etc.). The virtual symptom evolution sequence can be represented as multi-dimensional time series data, where each dimension corresponds to a virtual symptom feature. The inference and prediction results are a dataset containing the evolution sequences of virtual symptoms of all parallel instances. It not only provides a single prediction result, but also provides the probability distribution of future state evolution, providing richer information for risk assessment.

[0066] In this embodiment, the above-described technical solution overcomes the limitations of prediction using a single digital twin model. By introducing parameter perturbations and running multiple digital twin instances in parallel, various evolution paths of the caisson-soil system under different uncertainty conditions can be simulated, resulting in a more comprehensive and robust prediction of future states. This multi-scenario simulation not only reveals the most likely instability modes but also identifies extreme risks that may occur under preset unfavorable conditions, greatly improving the accuracy and reliability of risk assessment. Simultaneously, decoding the symptom map into initial state parameters ensures that the starting point of the digital twin model simulation is highly consistent with the actual monitored system state, making the prediction results more practically instructive. By recording and summarizing virtual symptom characteristics, the simulation results can be presented in a format similar to actual monitoring data, facilitating comparison with real-time data and subsequent risk analysis, providing richer and more convincing evidence for intelligent diagnostic reports. This method effectively addresses the uncertainties brought about by complex factors such as soil properties and construction environment during caisson construction, providing more reliable support for engineering safety decisions.

[0067] In one feasible implementation, the method further includes: during the digital twin model deduction process, real-time monitoring of the virtual symptom features generated by the parallel digital twin instance simulation; based on the first comparison result, setting the theoretical instability mode with the highest similarity ranking obtained by matching the symptom map with the knowledge base as the target verification mode; in the deduction prediction results, selecting the parallel digital twin instances whose simulated endpoint state is unstable and whose instability mode is consistent with the target verification mode as a verification instance set; reverse-analyzing the common parameter perturbations of the verification instance set during the deduction process, wherein the common parameter perturbations are used to corroborate or correct the speculation on the current risk root cause, and using the analysis results for the risk assessment.

[0068] In this embodiment, during the digital twin model simulation, the system continuously collects and records the virtual symptom features simulated by each parallel digital twin instance. These virtual symptom features are simulation data generated by the digital twin during the simulation of the dynamic evolution of the caisson-soil system, based on its internal physical model, initial conditions (decoded from the symptom map), and parameter perturbations. Examples include virtual acoustic emission event rates, vibration energy, and changes in soil contact pressure. The purpose of real-time monitoring is to capture subtle changes in the system state during the simulation, providing a detailed data foundation for subsequent screening and analysis. This can be achieved by setting up a data recording module within the digital twin model, automatically extracting and storing the virtual symptom features of the current state at the end of each simulation time step.

[0069] In this embodiment, based on the first comparison result, the theoretical instability mode with the highest similarity ranking obtained by matching the symptom map with the knowledge base is set as the target verification mode. The first comparison result is obtained by real-time matching and comparison of the symptom map with the preset well instability mode knowledge base. This comparison result usually includes multiple instability modes and their similarity ranking with the current symptom map. Setting the target verification mode means selecting the theoretical instability mode with the highest similarity (ranked first) with the current symptom map from these comparison results as the focus of this analysis and verification. This mode represents the most likely type of instability that the system will currently experience. For example, if the first comparison result shows that "local uneven settlement" has the highest similarity, then "local uneven settlement" is set as the target verification mode.

[0070] In this embodiment, parallel digital twin instances whose simulated final state is unstable and whose instability morphology is consistent with the target verification mode are selected from the simulation prediction results and used as the verification instance set. The simulation prediction results contain the virtual symptom evolution sequence of multiple parallel digital twin instances within a preset simulation time. The selection process aims to identify instances from these instances that not only exhibit instability in their final state, but also whose specific instability morphology (e.g., overall tilting, local uplift, or liquefaction of the soil around the well) is highly consistent with the previously set target verification mode. These selected instances constitute the verification instance set. For example, if the target verification mode is "local uneven settlement," then only those unstable instances whose simulation results show local uneven settlement will be included in the verification instance set. The selection can be achieved by performing pattern recognition or feature comparison on the final state of each instance.

[0071] In this embodiment, common parameter perturbations in the verification instance set during the deduction process are analyzed in reverse. These common parameter perturbations are used to corroborate or correct the inferences about the current risk root causes, and the analysis results are used for risk assessment. Reverse analysis refers to tracing back all instances in the verification instance set to identify common characteristics in the parameter perturbations applied to them in the initial stage. These parameter perturbations are introduced to simulate uncertainties when generating parallel digital twin instances, such as local reductions in soil strength parameters, changes in the soil-well wall friction coefficient, and anomalies in pore water pressure. By analyzing the common parameter perturbations of these instances leading to the same instability mode, the physical risk root cause most likely to cause the current system anomaly can be inferred. For example, if all instances in the verification instance set show a significant reduction in soil strength in a certain preset area, then insufficient soil strength in this area may be the risk root cause. These common parameter perturbations and their analysis results will be directly used to corroborate or correct the system's initial inferences about the current risk root causes and will be included as key inputs in the final risk assessment report, thereby improving the accuracy and reliability of the diagnosis.

[0072] In this embodiment, through the above-mentioned technical solution, this application introduces a refined screening and reverse analysis mechanism for the simulation results based on the digital twin model simulation. First, real-time monitoring of the virtual symptom characteristics of parallel digital twin instances ensures a comprehensive understanding of the system's evolution process. Second, setting a target verification mode based on the first comparison result provides clear guidance for subsequent screening work, avoiding blind analysis. By screening out instability instances consistent with the target verification mode, the focus can be placed on the instability path most relevant to the current actual symptom spectrum. More importantly, reverse analysis of these verification instance sets identifies their common parameter disturbances, directly revealing the deep physical mechanisms and potential risk roots leading to the preset instability mode. This not only corroborates or corrects the system's initial speculation on the risk roots, greatly improving the accuracy and reliability of risk assessment, but also provides a solid scientific basis for subsequent targeted treatment suggestions, thereby effectively improving the intelligence level and decision support capabilities of caisson construction safety monitoring.

[0073] In one feasible implementation, reverse analysis of the common parameter disturbances in the verification instance set during the simulation process includes: extracting the parameter disturbances applied to each instance in the verification instance set at the initial moment of the simulation, wherein the parameter disturbances include at least one of the following: local strength reduction coefficient of soil, change in friction coefficient between soil and well wall, and change in pore water pressure in a preset area; performing cluster analysis on all extracted parameter disturbances to identify at least one common disturbance cluster that is spatially clustered and numerically consistent; mapping each identified common disturbance cluster to its corresponding spatial location in the physical information map; based on the type, magnitude, and mapped spatial location of the parameter disturbances contained in each common disturbance cluster, querying and inferring the most likely physical risk root cause leading to the current symptom map and the target verification mode from a preset engineering disease cause comparison table; and using the inferred physical risk root cause and its corresponding confidence level as the analysis result.

[0074] In this embodiment, the parameter perturbations applied to each instance in the verification instance set at the initial moment of the simulation are extracted. These parameter perturbations are minor adjustments to the system's physical parameters introduced by the digital twin model when generating multiple parallel digital twin instances to simulate uncertainties in actual engineering. Examples include at least one of the following: local soil strength reduction coefficient, change in the friction coefficient between the soil and the well wall, and change in pore water pressure in a preset area. These perturbation parameters are precisely recorded at the start of the simulation and serve as key evidence for tracing the root causes of instability in subsequent analysis. For example, the local soil strength reduction coefficient can simulate the existence of locally weak soil areas, the change in the friction coefficient between the soil and the well wall reflects the actual changes in the contact interface between the well wall and the soil during construction, and the change in pore water pressure in the preset area can simulate groundwater level fluctuations or poor drainage.

[0075] In this embodiment, cluster analysis is performed on all extracted parameter perturbations to identify at least one common perturbation cluster that exhibits spatial aggregation and numerical consistency. Cluster analysis aims to identify sets of perturbations with similar characteristics or spatial correlations from a large number of potentially dispersed parameter perturbations, thereby revealing potential local or regional problems. Various clustering algorithms can be used, such as K-means, DBSCAN, or hierarchical clustering, with the initial parameter perturbations of all validation instances and their corresponding spatial location information as input. Clustering criteria typically include spatial proximity (e.g., perturbations occurring in adjacent nodes or regions in a physical information map) and numerical consistency (e.g., perturbation parameters of the same type and similar numerical variation trends). In this way, common problems such as "generally reduced soil strength on one side of the caisson" or "abnormally increased pore water pressure within a preset depth range" can be identified.

[0076] In this embodiment, each identified common disturbance cluster is mapped to its corresponding spatial location in the physical information map. Since each parameter disturbance is associated with a preset node or region in the physical information map within the digital twin model, once a common disturbance cluster is identified, it can be directly mapped onto a 3D model of the caisson structure or surrounding soil based on the original spatial information of these disturbance parameters. For example, a common disturbance cluster might correspond to a soil region at a corner of the caisson bottom, or a soil-structure interface at a certain depth on the caisson sidewall, thus enabling intuitive location of the root cause of the risk.

[0077] In this embodiment, based on the type, magnitude, and spatial location of the parameter disturbances contained in each common disturbance cluster, the system queries and infers the most likely physical risk root cause leading to the current symptom map and the target verification mode from a preset engineering defect cause lookup table. The engineering defect cause lookup table is a structured knowledge base containing association rules, thresholds, and typical characteristics between various engineering defects (such as uneven settlement, local uplift, piping, quicksand, etc.) and changes in physical parameters leading to these defects (such as reduced soil strength, decreased friction, increased pore water pressure). The system queries the lookup table based on the type (e.g., soil strength reduction), magnitude (e.g., reduction range) and spatial location (e.g., occurring in a weak soil layer or groundwater level fluctuation zone) of the common disturbance cluster. For example, if a significant decrease in the soil strength reduction coefficient is identified in a certain area at the bottom of the caisson, and this area is located in a weak clay layer, the lookup table might infer that "uneven settlement due to insufficient local bearing capacity" is a possible risk root cause.

[0078] In this embodiment, the inferred physical risk root causes and their corresponding confidence levels are used as the analysis results. There can be one or more inferred physical risk root causes, each with a confidence level. The confidence level can be calculated based on factors such as the degree of matching of rules in the lookup table, the tightness of clustering results, and the frequency of the perturbation cluster appearing in the validation instance set. For example, if multiple validation instances point to the same common perturbation cluster, and this perturbation cluster highly matches a certain disease cause in the lookup table, its confidence level will be higher. The final analysis results will be used as part of the intelligent diagnostic report to guide engineers in taking targeted measures.

[0079] In this embodiment, through the above-described technical solution, this application can trace back and identify the deep physical risk roots that cause the caisson-soil system to tend towards a preset instability mode from the inference results of the digital twin model. Specifically, by extracting the initial parameter disturbances of the verification instance and performing cluster analysis to identify common disturbance clusters, and then mapping these disturbance clusters to actual spatial locations, combined with a preset engineering defect cause comparison table, this application can accurately infer the physical risk roots most likely to lead to the current symptom map and target instability mode, as well as their confidence level. This not only makes up for the shortcomings of risk root cause identification based solely on symptom map matching and digital twin inference prediction results, but also enables risk assessment and subsequent handling suggestions to go from the surface to the essence, providing a more scientific, specific, and targeted basis for engineering decision-making, and improving the intelligence level and risk management capabilities of caisson construction safety monitoring.

[0080] In one feasible implementation, the step of combining the first comparison result and the inference prediction result to perform risk assessment and generate a corresponding intelligent diagnostic report includes: obtaining the instability mode with the highest similarity from the first comparison result as a first candidate mode; analyzing the virtual symptom evolution sequence of all the parallel digital twin instances in the inference prediction result, identifying instances that cause the structural state to exceed the safety threshold, and statistically analyzing the instability modes corresponding to these instances, selecting the instability mode with the highest frequency as a second candidate mode; comparing the first candidate mode and the second candidate mode: if they are consistent, the mode is determined as the comprehensive judgment instability mode, and its corresponding similarity is combined with the inferred risk ratio to determine the risk level; if they are inconsistent, the consistency between the simulation parameters of the parallel digital twin instances corresponding to the first candidate mode and the second candidate mode and the current construction record is checked, and the significance of preset symptoms in the symptom map is referenced to finally determine the risk level. The process involves: determining the comprehensive instability mode and the risk level; retrieving a pre-set list of potential risk sources from the instability mode knowledge base based on the determined comprehensive instability mode; filtering and sorting the potential risk sources from the list based on the spatial distribution intensity of abnormal symptoms in the symptom map and the perturbation parameter type of the parallel digital twin instance whose simulated instability mode is determined to be the second candidate mode, to obtain the inferred risk sources; matching at least one targeted treatment suggestion from the measure library based on the inferred risk sources and the comprehensive instability mode, and using the parallel digital twin instance with the determined comprehensive instability mode to simulate the system response after executing each targeted treatment suggestion, attaching the simulation results as an expected effect description of the suggestion to the report; and integrating the risk level, the comprehensive instability mode, the inferred risk sources, and the targeted treatment suggestions with expected effect descriptions to form and output the intelligent diagnostic report.

[0081] In this embodiment, the step of obtaining the first candidate mode aims to extract the most likely instability mode corresponding to the current system state from the matching results of real-time monitoring data and a preset instability mode knowledge base. The first comparison result typically includes multiple instability modes and their similarity scores with the current symptom map. The system selects the mode with the highest similarity score as the initial judgment basis, i.e., the first candidate mode. This provides an initial direction based on empirical knowledge for subsequent risk assessment.

[0082] In this embodiment, the analysis of simulation and prediction results identifies instances where the structural state exceeds a safety threshold. The steps of statistically analyzing the instability modes corresponding to these instances and selecting the most frequent instability mode as the second candidate mode utilize the predictive power of the digital twin model to provide another dimension for determining instability modes. Specifically, the system iterates through the virtual symptom evolution sequences generated by all parallel digital twin instances within a preset simulation period. For each instance, its key state parameters (e.g., settlement, tilt, soil stress, pore water pressure, etc.) are continuously monitored to ensure they do not exceed preset safety thresholds. Once an instance's state exceeds the threshold, it is determined to be unstable, and the simulated instability mode is recorded. After all instances have been simulated, the frequency of various instability modes among the instances leading to instability is statistically analyzed, and the most frequent mode is identified as the second candidate mode. This process verifies the possibility of potential instability modes by simulating future evolution and provides a judgment based on dynamic prediction.

[0083] In this embodiment, the core of risk assessment is the step of comparing the first candidate mode and the second candidate mode to determine the comprehensive instability mode and risk level. By combining real-time matching and future prediction, the accuracy of the judgment is improved. If both are consistent, meaning the first and second candidate modes match, it indicates that both real-time monitoring data and future evolution predictions point to the same instability mode, greatly increasing the likelihood of this mode occurring. In this case, the system identifies this mode as the comprehensive instability mode. The risk level determination comprehensively considers the similarity of the mode in the first comparison result (reflecting the degree of matching of the current state) and the predicted risk ratio of the mode in the prediction result (e.g., the proportion of instances simulating instability of this mode to the total number of instances, reflecting the likelihood of future occurrence). A comprehensive risk level is obtained through quantitative calculation using a preset risk assessment model. If the two are inconsistent, meaning the two candidate modes are inconsistent, it indicates a discrepancy in judgment, requiring deeper analysis. The system further checks the consistency between the simulation parameters of the parallel digital twin instances corresponding to each of the two candidate modes and the current construction records. For example, if the simulated parameters of a digital twin instance (such as soil strength and friction coefficient) deviate significantly from the parameters recorded during actual construction (such as geological survey data and construction logs), the reliability of its prediction results will decrease. Simultaneously, the system also considers the significance of preset symptoms in the symptom map. For instance, if an abnormal symptom strongly correlated with the first candidate mode is particularly prominent in the symptom map, the system will tend to adopt the first candidate mode. Through this multi-dimensional, cross-validation approach, the system ultimately determines the comprehensive instability mode and adjusts the risk level accordingly, ensuring the rigor of the assessment.

[0084] In this embodiment, the step of retrieving a pre-set list of potential risk sources from the instability mode knowledge base based on the determined comprehensive instability mode is intended to ensure that once the comprehensive instability mode is determined, the system immediately retrieves a list of all potential risk sources associated with that mode from the pre-set instability mode knowledge base. This list is constructed based on historical experience and engineering theory, providing a preliminary scope and direction for subsequent risk source prediction.

[0085] In this embodiment, the step of filtering and sorting from the potential risk root cause list based on the spatial distribution intensity of abnormal symptoms in the symptom map and the perturbation parameter types of parallel digital twin instances whose simulated instability modes are determined to be the second candidate modes, aims to further pinpoint the most likely risk root causes from the potential risk root cause list. Specifically, the system analyzes the spatial distribution of abnormal symptoms in the symptom map, such as whether the abnormal symptoms are concentrated in a certain area and the extent of their influence, reflecting the locality or wide-area nature of the risk occurrence. Simultaneously, it combines the perturbation parameter types (e.g., soil strength reduction, frictional changes, or pore water pressure anomalies) used by the parallel digital twin instances whose simulated instability modes are determined to be the second candidate modes when identifying the second candidate modes. These perturbation parameters directly point to the physical mechanisms leading to instability. By integrating this information, the system can filter and sort potential risk root causes, for example, prioritizing those risk root causes that highly match the spatial distribution of abnormal symptoms and are consistent with the key perturbation parameter types, thereby obtaining more targeted inferred risk root causes.

[0086] In this embodiment, the key step of matching at least one targeted treatment suggestion from the measure library based on the inferred risk root cause and the comprehensive determination of the instability mode, and simulating the system response after implementing each targeted treatment suggestion using parallel digital twin instances with determined comprehensive instability modes, and attaching the simulation results as an explanation of the expected effect of the suggestion in the report, is to provide actionable recommendations. The system will match at least one targeted treatment suggestion from a preset measure library based on the determined inferred risk root cause and comprehensive determination of the instability mode. These suggestions may include adjusting construction techniques, taking reinforcement measures, and drainage and pressure reduction. To verify the effectiveness of these suggestions, the system will use parallel digital twin instances with previously determined comprehensive instability modes (these instances represent the most dangerous state of the current system) to simulate the implementation of each matched treatment suggestion in these instances. For example, simulating the change in settlement rate after grouting reinforcement in a certain area. The system response after the simulation (such as settlement, stress distribution, stability coefficient, etc.) will be recorded as an explanation of the expected effect of the suggestion and attached to the final intelligent diagnostic report. This pre-simulation verification mechanism greatly improves the scientific validity and reliability of the treatment recommendations.

[0087] In this embodiment, the final step of integrating risk levels, comprehensively determining instability modes, speculative risk root causes, and targeted remedial suggestions with expected effects to form and output an intelligent diagnostic report involves integrating all assessment results, including the determined risk level, comprehensively determining instability modes, speculative risk root causes, and targeted remedial suggestions with expected effects, to form a comprehensive, clear, and actionable intelligent diagnostic report. This report provides engineering decision-makers with information across the entire chain from risk identification and root cause analysis to solution verification, supporting them in making timely and effective safety decisions.

[0088] In this embodiment, through the above-described technical solution, this application effectively solves the problem of accurately determining instability modes, quantifying risk levels, and providing validated treatment recommendations from multi-source information in the safety monitoring of caisson construction. Specifically, by introducing a dual verification mechanism of a first candidate mode and a second candidate mode, combined with real-time monitoring data matching and future trend prediction, the accuracy and reliability of judging the instability mode of the caisson are improved. When the two candidate modes are inconsistent, the consistency check between simulation parameters and construction records, as well as the significance of symptom maps, are introduced as auxiliary judgments, making the risk assessment more comprehensive and robust, and avoiding misjudgments that may be caused by a single judgment. In addition, by finely screening and ranking potential risk sources, combined with the spatial distribution intensity of symptom maps and the disturbance parameter analysis of digital twin models, the risk sources can be located more accurately, providing a clear direction for subsequent treatment. Most importantly, by using digital twin models to simulate the execution of treatment recommendations and provide explanations of expected effects, the proposed treatment recommendations are not only targeted, but their effectiveness is also pre-quantified and verified, greatly enhancing the practicality and decision support capabilities of the report, thereby providing a more scientific, reliable, and operable safety diagnosis and treatment plan for caisson construction.

[0089] In one feasible implementation, the method further includes: after the completion of the current caisson construction project, encrypting all the symptom maps generated during the construction process and their corresponding actual engineering status and handling effects to generate a local experience data package; based on the local experience data package, training the locally deployed pre-trained graph neural network model and / or the digital twin model to obtain local model parameter update amounts; uploading the local model parameter update amounts to the central server through a federated learning protocol, and downloading the globally evolved model parameters generated by aggregating the update amounts uploaded by other projects from the central server, to update the local pre-trained graph neural network model and / or the digital twin model for subsequent project safety monitoring.

[0090] In this embodiment, upon completion of a caisson construction project, the system collects all symptom maps generated during the project. These maps reflect the dynamic behavior patterns of the caisson-soil system at different times. Simultaneously, it records the actual engineering conditions corresponding to these symptom maps, such as settlement, tilt, and crack occurrence, as well as the remedial measures taken for these conditions and their final effects. To protect the privacy and security of the project data, this data is encrypted before generating the local experience data package. For example, advanced encryption algorithms are used to obfuscate the data, ensuring that only authorized parties can access and interpret it. This local experience data package is a pre-set, structured historical data set for the current project, providing valuable real-world feedback for subsequent model training.

[0091] In this embodiment, the locally deployed pre-trained graph neural network model and digital twin model are the core tools for real-time monitoring and prediction. Using the generated local experience data package, these models can be further trained or fine-tuned. For example, for the graph neural network model, its weights and biases can be adjusted using the backpropagation algorithm to improve its performance on new project data; for the digital twin model, its internal physical parameters or constitutive relations can be calibrated based on actual observation data. After training, the system does not directly upload the entire model or the original data, but instead calculates and extracts the update amount of the model parameters, i.e., the parameter changes that occurred during the training. This approach effectively reduces data transmission volume and further protects the privacy of the original data.

[0092] In this embodiment, the federated learning protocol is a distributed machine learning paradigm that allows multiple participants to collaboratively train a shared model without directly sharing raw data. In this method, each caisson construction project acts as an independent client, uploading its calculated local model parameter updates to a central server. The central server receives these updates from different projects and integrates them using an aggregation algorithm (e.g., weighted average) to generate a global evolutionary model parameter that represents the collective wisdom of all projects. Subsequently, the central server distributes these global evolutionary model parameters to each local project, allowing their pre-trained graph neural network models and / or digital twin models to update using these global parameters. Through this iterative process, the local model continuously absorbs experience from other projects, achieving continuous evolution.

[0093] In this embodiment, through the above technical solution, this application effectively solves the problem that pre-trained models lack generalization ability and are difficult to adapt to complex and changing working conditions when facing different construction projects. Specifically, after the completion of each caisson construction project, the system can encrypt the symptom maps, actual engineering status, and treatment effects generated by the project and form a local experience data package, which provides valuable real-world feedback for the model. Based on these local experience data packages, the locally deployed graph neural network model and / or digital twin model are trained, enabling the model to optimize for the geological conditions and construction characteristics of the preset project, improving the diagnostic accuracy and prediction accuracy of the model in the project environment. Furthermore, through the federated learning protocol, each project can share model parameter updates in a privacy-preserving manner. The central server aggregates these updates to generate global evolutionary model parameters, allowing each local model to absorb diverse experiences from other projects. This not only avoids the direct sharing of sensitive construction data and ensures data privacy, but also enables the local model to continuously evolve, enhancing its generalization ability and robustness. Therefore, the safety monitoring model used in subsequent projects will have higher accuracy and adaptability, be able to more accurately identify potential risks, and provide more reliable early warning and handling suggestions, thereby comprehensively improving the safety assurance level of caisson construction.

[0094] In the embodiments of this application, the safety monitoring method for caisson construction constructs a full-link safety monitoring system from real-time data acquisition and multi-scale feature extraction to dynamic risk assessment through collaborative perception of multi-physics field sensing and edge computing, intelligent fusion of spatial features of physical information graphs and graph neural networks, and continuous evolution mechanism of parallel inference of digital twins and federated learning. It breaks through the limitations of traditional monitoring methods such as data fragmentation, early warning lag, and ambiguous risk positioning. By deeply coupling the signs of soil micro-activity with the dynamic response of the structure, it realizes early identification, accurate tracing, and dynamic handling of the instability risk of the caisson-soil system, thereby improving the safety management level and risk response capability of caisson construction under complex geological conditions.

[0095] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the safety monitoring method for caisson construction. Any simple modifications based on this technical concept are within the protection scope of this application.

[0096] This application also provides a safety monitoring system for caisson construction, see reference. Figure 2 The safety monitoring system for caisson construction includes: a memory 10, a processor 20, and a safety monitoring program for caisson construction stored on the memory 10 and executable on the processor 20. The safety monitoring program for caisson construction is configured to implement the steps of the safety monitoring method for caisson construction.

[0097] The safety monitoring system for caisson construction provided in this application, employing the safety monitoring method for caisson construction described in the above embodiments, can improve the efficiency and accuracy of risk management. Compared with the prior art, the beneficial effects of the safety monitoring system for caisson construction provided in this application are the same as those of the safety monitoring method for caisson construction provided in the above embodiments, and other technical features of the safety monitoring system for caisson construction are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0098] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0099] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any equivalent structural transformations made under the technical concept of this application using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included within the scope of patent protection of this application.

Claims

1. A safety monitoring method for caisson construction, characterized in that, The method includes: Multiple integrated sensing units deployed in the caisson structure and surrounding soil are used to synchronously collect original multi-physics field monitoring data during the construction process; the original multi-physics field monitoring data includes acoustic emission signals, vibration signals and soil contact pressure signals. Edge computing processing is performed on the original multiphysics field monitoring data collected by each integrated sensing unit to extract local symptom feature vectors that characterize the soil micro-activity state and stress change trend at the monitoring location of the integrated sensing unit. The local symptom feature vectors from all integrated sensing units are uploaded to the cloud analysis platform. In the cloud analysis platform, a physical information map is constructed based on the three-dimensional model of the caisson structure, the spatial location information of each integrated sensing unit, and the preset geological stratification information. The nodes of the physical information map represent soil or structural units and carry the corresponding local symptom feature vectors. The edges of the map represent the spatial adjacency relationship between units and are assigned physical attribute weights based on the geological and construction conditions. The physical information map is input into a pre-trained graph neural network model for message passing and feature aggregation, so as to fuse the spatially associated local symptom feature vectors and generate a symptom map reflecting the current dynamic behavior pattern of the caisson-soil system. The symptom map is matched and compared with the preset well instability mode knowledge base in real time to obtain the first comparison result. The system state reflected by the symptom map is used as the initial condition to drive the parallel digital twin model to perform forward inference, predict the evolution trend of the system state within the preset inference time, and obtain the inference prediction result. A risk assessment is performed by combining the first comparison result and the inference and prediction results to generate a corresponding intelligent diagnostic report; the intelligent diagnostic report includes the risk level, a comprehensive judgment of the instability mode, an inferred risk root cause, and targeted handling suggestions.

2. The safety monitoring method for caisson construction as described in claim 1, characterized in that, The steps of simultaneously collecting raw multi-physics field monitoring data during construction by deploying multiple integrated sensing units in the caisson structure and surrounding soil also include: Each integrated sensing unit is controlled to synchronously collect the original monitoring data of the multi-physics field at a first preset acquisition frequency. Monitor the rate of change of the newly acquired multiphysics field raw monitoring data of each integrated sensing unit relative to the data of the previous period; When the rate of change of any physical field data of any integrated sensing unit exceeds its corresponding first threshold, it is determined that an initial abnormality has occurred at that location. The data acquisition frequency of the integrated sensing unit is switched to a second preset acquisition frequency higher than the first preset acquisition frequency. Based on the preset adjacency relationship of the physical information map, the adjacent integrated sensing units are notified to also switch their data acquisition frequency to the second preset acquisition frequency. Data is acquired at the second preset acquisition frequency, and the acquired preset frequency band signals related to soil micro-fractures or plastic flow are subjected to directional analysis. When all units that have switched to the second preset acquisition frequency have a data change rate that is lower than their corresponding second threshold and this continues for a preset time, it is determined that the abnormal symptoms have subsided, and the data acquisition frequency of the relevant units is switched back to the first preset acquisition frequency; wherein, the second threshold is lower than the first threshold.

3. The safety monitoring method for caisson construction as described in claim 1, characterized in that, The steps of performing edge computing processing on the raw multiphysics monitoring data collected by each integrated sensing unit to extract local symptom feature vectors characterizing the soil micro-activity state and stress change trend at the monitoring location of that unit include: The acoustic emission signal is subjected to event detection and waveform analysis to extract the acoustic emission event rate, cumulative event energy value and main frequency offset to obtain an acoustic feature set; The vibration signal is subjected to spectrum analysis and energy calculation to extract the vibration main frequency, the energy ratio of the preset frequency band and the signal entropy value, so as to obtain the vibration feature set; Spatial gradient and time series analysis were performed on the soil contact pressure signal to calculate the rate of change of the current pressure value relative to the historical baseline, the pressure gradient value between adjacent sensors, and the pressure distribution entropy, thereby obtaining a pressure feature set. The acoustic feature set, the vibration feature set, and the pressure feature set are spliced ​​together in a predetermined order to form the local symptom feature vector.

4. The safety monitoring method for caisson construction as described in claim 1, characterized in that, The steps of uploading the local symptom feature vectors from all integrated sensing units to the cloud analysis platform, and constructing a physical information map on the cloud analysis platform based on the three-dimensional model of the well structure, the spatial location information of each integrated sensing unit, and the preset geological stratification information include: Based on the preset geological stratification information, determine the first weight coefficient of the edge between the corresponding nodes of the integrated sensing unit at the junction of different soil layers; Based on the sinking depth and real-time attitude data of the caisson during the current construction phase, calculate the second weighting coefficient of the edge between the corresponding nodes of the integrated sensing unit in the contact area between the well wall and the soil. The first weight coefficient and the second weight coefficient are combined to obtain the physical attribute weight; The nodes carrying the local symptom feature vectors and the edges assigned the weights of the physical attributes are combined to form the physical information graph.

5. The safety monitoring method for caisson construction as described in claim 1, characterized in that, The steps of inputting the physical information map into a pre-trained graph neural network model for message passing and feature aggregation to fuse the spatially correlated local symptom feature vectors and generate a symptom map reflecting the current dynamic behavior pattern of the caisson-soil system include: For each node in the physical information graph, the local symptom feature vector information transmitted from neighboring nodes is weighted and aggregated according to the physical attribute weight assigned to each edge to obtain aggregated information. The local symptom feature vector carried by each node is fused with the aggregated information, and the feature representation of the node is updated using a nonlinear transformation function to obtain the updated node feature representation; wherein, the fusion process introduces momentum conservation constraints so that the updated node feature representation conforms to the balance trend of action and reaction forces between adjacent units. Repeat the weighted aggregation and fusion update steps at least once; After repeated updates, the final feature representations of all nodes are globally pooled and integrated to generate the symptom map.

6. The safety monitoring method for caisson construction as described in claim 1, characterized in that, Using the system state reflected in the symptom map as initial conditions, the steps of driving a parallel digital twin model to perform forward inference, predicting the evolution trend of the system state within a preset inference period, and obtaining the inference prediction results include: Based on the symptom map, multiple parallel digital twin instances with parameters perturbed within a preset range are generated. The symptom map is decoded into the initial state parameters for each of the parallel digital twin instances; Within the preset simulation time, all the parallel digital twin instances are run synchronously to simulate the dynamic evolution of the caisson-soil system under their respective parameters. At each simulation time step, the virtual symptom characteristics generated by each of the parallel digital twin instances are recorded; After the simulation is completed, based on the virtual symptom features recorded at all simulation time steps, all virtual symptom features generated by each of the parallel digital twin instances during the entire simulation duration are summarized to form the virtual symptom evolution sequence of that instance; the virtual symptom evolution sequences of all instances constitute the simulation prediction result.

7. The safety monitoring method for caisson construction as described in claim 6, characterized in that, The method further includes: During the digital twin model derivation process, the virtual symptom characteristics generated by the parallel digital twin instance simulation are monitored in real time; Based on the first comparison result, the theoretical instability mode with the highest similarity ranking obtained by matching the symptom map with the knowledge base is set as the target verification mode. In the simulation and prediction results, the parallel digital twin instances whose simulated endpoint state is unstable and whose instability mode is consistent with the target verification mode are selected as the verification instance set; The common parameter perturbations of the verification instance set during the inference process are analyzed in reverse. These common parameter perturbations are used to corroborate or correct the speculation on the current risk root cause, and the analysis results are used for the risk assessment.

8. The safety monitoring method for caisson construction as described in claim 7, characterized in that, The reverse analysis of the common parameter perturbations of the verification instance set during the inference process includes: Extract the parameter perturbation applied to each instance in the verification instance set at the initial moment of the simulation. The parameter perturbation includes at least one of the following: local strength reduction coefficient of soil, change in friction coefficient between soil and well wall, and change in pore water pressure in a preset area. Cluster analysis is performed on all the extracted parameter perturbations to identify at least one common perturbation cluster that is spatially clustered and numerically consistent. Each identified common disturbance cluster is mapped to its corresponding spatial location in the physical information map; Based on the type, magnitude, and spatial location of the parameter disturbances contained in each common disturbance cluster, the most likely physical risk source leading to the current symptom map and the target verification mode is queried and inferred from the preset engineering disease cause comparison table. The inferred physical risk sources and their corresponding confidence levels are used as the analysis results.

9. The safety monitoring method for caisson construction as described in claim 8, characterized in that, The steps for conducting a risk assessment by combining the first comparison result and the inference and prediction results to generate a corresponding intelligent diagnostic report include: From the first comparison results, the instability pattern with the highest similarity is selected as the first candidate pattern; Analyze the virtual symptom evolution sequence of all parallel digital twin instances in the inference and prediction results, identify the instances that cause the structural state to exceed the safety threshold, and count the instability modes corresponding to these instances. The instability mode with the highest frequency is selected as the second candidate mode. Compare the first candidate pattern with the second candidate pattern: If the two are consistent, the mode is determined as the comprehensive judgment instability mode, and the corresponding similarity is combined with the inferred risk ratio to determine the risk level; If the two are inconsistent, check the consistency between the simulation parameters of the parallel digital twin instances corresponding to the first candidate mode and the second candidate mode and the current construction record, and refer to the significance of the preset symptoms in the symptom map to finally determine the comprehensive judgment instability mode and the risk level. Based on the determined comprehensive judgment instability mode, a pre-set list of potential risk sources is retrieved from the instability mode knowledge base; Based on the spatial distribution intensity of abnormal symptoms in the symptom map and the perturbation parameter type of the parallel digital twin instance whose simulated instability mode is determined to be the second candidate mode, the inferred risk source is obtained by filtering and sorting from the potential risk source list. Based on the inferred risk root causes and the comprehensive determination of the instability mode, at least one of the targeted treatment suggestions is matched from the measure library, and the system response after executing each of the targeted treatment suggestions is simulated using the parallel digital twin instance that has determined the comprehensive determination of the instability mode. The simulation results are attached to the report as an explanation of the expected effect of the suggestion. The intelligent diagnostic report is generated by integrating the risk level, the comprehensive judgment of the instability mode, the inferred risk root cause, and the targeted treatment suggestions with expected effects.

10. A safety monitoring system for caisson construction, characterized in that, The safety monitoring system for caisson construction includes: a memory, a processor, and a safety monitoring program for caisson construction stored in the memory and executable on the processor, wherein the safety monitoring program for caisson construction is configured to implement the steps of the safety monitoring method for caisson construction as described in any one of claims 1 to 9.

Images