Method for detecting pile load box based on multi-modal data

Abstract

The application discloses a kind of based on the detection method of pile foundation load box of multi-modal data, and the application relates to the technical field of multi-modal data, based on the use of pile foundation load box, corresponding multiple stress data and the overall form of pile foundation load box Determine the multi-modal data corresponding to pile foundation load box, according to the identification of each data combination Determine the load characteristics of pile foundation load box, based on each load characteristics, the overall form of pile foundation load box, corresponding stress data Determine the digital twin system of pile foundation load box, improve the accuracy of digital twin system of pile foundation load box.According to the area position of each abnormal compression area and the working stage of pile foundation load box Determine the force balance state of pile foundation load box;According to the thermal data of the pile foundation load box, the force balance state of the pile foundation load box and the acoustic data Determine the equilibrium state curve of pile foundation load box, improve the accuracy of equilibrium state curve of pile foundation load box.

Description

Technical Field

[0001] This invention relates to the technical field of multimodal data, and in particular to a method for detecting pile foundation load cells based on multimodal data. Background Technology

[0002] With the development of technology, pile foundation load cells, often referred to as self-balancing load cells, are specialized devices used for static load tests in pile foundation engineering. A pile foundation load cell is a disc-shaped, highly integrated hydraulic jack device that is pre-placed in a designated location inside the pile before the concrete is poured. In existing technologies, real-time monitoring of the pile foundation load cell and collection of multiple operational data points are used to determine its operational state. However, this reliance on only these data points neglects the multiple stress data points across different dimensions, affecting the accuracy of the digital twin system of the pile foundation load cell and resulting in low accuracy of its equilibrium curve. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of the prior art. This invention provides a method for detecting pile foundation load cells based on multimodal data.

[0004] This invention provides a method for detecting pile foundation load cells based on multimodal data, including: The usage status of the pile foundation load cell is determined based on its current location and corresponding usage scenario, and the stress detection of the pile foundation load cell is triggered to collect multiple stress data of the pile foundation load cell in different dimensions. Based on the usage of the pile foundation load cell, the corresponding multiple stress data, and the overall shape of the pile foundation load cell, the multimodal data corresponding to the pile foundation load cell are determined, and multiple data combinations are determined; the multimodal data includes images, data tables, and model files; The load characteristics of the pile foundation load cell are determined by identifying various data combinations. Based on these load characteristics, the overall shape of the pile foundation load cell, and the corresponding compression data, a digital twin system for the pile foundation load cell is determined. The digital twin system learns the nonlinear mapping relationship between state, performance, and final risk level, and outputs a decision item. The abnormal compression events of the pile foundation load cell are identified based on the identification of the digital twin system. Based on the detection of the abnormal compression events, multiple abnormal compression areas are identified. The force balance state of the pile foundation load cell is determined according to the regional location of each abnormal compression area and the working stage of the pile foundation load cell. Based on the thermal data, force balance state, and acoustic data of the pile foundation load cell, the equilibrium state curve of the pile foundation load cell is determined. Based on the equilibrium state curve of the pile foundation load cell, several key detection nodes are determined, and the dynamic control events of the pile foundation load cell are determined.

[0005] Compared with the prior art, the beneficial effects of the present invention are: (1) Determine the usage status of the pile foundation load box based on its current location and corresponding usage scenario, and trigger the stress detection of the pile foundation load box to collect multiple stress data of the pile foundation load box in different dimensions; determine the multimodal data corresponding to the pile foundation load box based on the usage status of the pile foundation load box, the corresponding multiple stress data and the overall shape of the pile foundation load box, and determine multiple data combinations, introduce multiple stress data of the pile foundation load box in different dimensions, and further control the multimodal data corresponding to the pile foundation load box.

[0006] (2) The load characteristics of the pile foundation load cell are determined based on the identification of each data combination. The digital twin system of the pile foundation load cell is determined based on each load characteristic, the overall shape of the pile foundation load cell, and the corresponding compression data. This system takes into account each load characteristic, the overall shape of the pile foundation load cell, and the corresponding compression data, thereby improving the accuracy of the digital twin system of the pile foundation load cell.

[0007] (3) Based on the identification of the digital twin system, the abnormal compression event of the pile foundation load box is determined. Based on the detection of the abnormal compression event, multiple abnormal compression areas are determined. Based on the regional location of each abnormal compression area and the working stage of the pile foundation load box, the force balance state of the pile foundation load box is determined. Based on the thermal data, force balance state and acoustic data of the pile foundation load box, the balance state curve of the pile foundation load box is determined. Based on the balance state curve of the pile foundation load box, multiple key detection nodes are determined, and the force balance state of the pile foundation load box is further controlled. The overall consideration of the thermal data, force balance state and acoustic data of the pile foundation load box is realized, the accuracy of the balance state curve of the pile foundation load box is improved, and the dynamic control event of the pile foundation load box is determined. Attached Figure Description

[0008] Figure 1 This is a flowchart illustrating the detection method for pile foundation load cells based on multimodal data in an embodiment of the present invention. Figure 2 This is a flowchart illustrating step S11 of the pile foundation load cell detection method based on multimodal data in an embodiment of the present invention. Figure 3 This is a flowchart illustrating step S12 in the pile foundation load cell detection method based on multimodal data in an embodiment of the present invention. Figure 4This is a flowchart illustrating step S13 in the pile foundation load cell detection method based on multimodal data in an embodiment of the present invention. Figure 5 This is a flowchart illustrating step S14 of the pile foundation load cell detection method based on multimodal data in an embodiment of the present invention. Figure 6 This is a flowchart illustrating step S15 of the pile foundation load cell detection method based on multimodal data in an embodiment of the present invention. Detailed Implementation

[0009] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0010] Please see Figures 1 to 6 A method for detecting pile foundation load cells based on multimodal data is proposed, applicable to multimodal data scenarios. The method includes: Step S11: Determine the usage status of the pile foundation load cell based on its current location and corresponding usage scenario, and trigger the stress detection of the pile foundation load cell to collect multiple stress data of the pile foundation load cell in different dimensions. Step S12: Based on the usage of the pile foundation load cell, the corresponding multiple stress data and the overall shape of the pile foundation load cell, determine the multimodal data corresponding to the pile foundation load cell, and determine multiple data combinations; Step S13: Determine the load characteristics of the pile foundation load cell based on the identification of each data combination, and determine the digital twin system of the pile foundation load cell based on each load characteristic, the overall shape of the pile foundation load cell, and the corresponding compression data. Step S14: Based on the identification of the digital twin system, determine the abnormal compression event of the pile foundation load cell, determine the corresponding multiple abnormal compression areas based on the detection of the abnormal compression event, and determine the force balance state of the pile foundation load cell according to the regional location of each abnormal compression area and the working stage of the pile foundation load cell. Step S15: Determine the equilibrium state curve of the pile foundation load cell based on the thermal data, force balance state, and acoustic data of the pile foundation load cell. Based on the equilibrium state curve of the pile foundation load cell, determine multiple key detection nodes and determine the dynamic control events of the pile foundation load cell.

[0011] Please see Figure 2 In step S11, the specific steps are as follows: S111: Mark the current position of the pile foundation load box, trigger the surrounding detection based on the current position of the pile foundation load box, determine multiple scene features during the surrounding detection process, determine the usage scenario of the pile foundation load box based on the multiple scene features and the working task of the pile foundation load box, and determine the usage status of the pile foundation load box based on the current position of the pile foundation load box and the corresponding usage scenario. S112: Determine the corresponding usage status based on the usage of the pile foundation load cell, and trigger the stress detection of the pile foundation load cell in the usage status to dynamically monitor the stress detection of the pile foundation load cell, and determine multiple stress data of the pile foundation load cell in different dimensions based on the stress detection of the pile foundation load cell.

[0012] In the embodiments of this application, a reference anchor point for a digital twin is established. This is not a simple coordinate record, but rather a high-precision positioning technology (such as total station, UWB, or association with BIM model coordinates) is used to assign each load cell an immutable ID and three-dimensional spatial coordinates (X,Y,Z) in the project's global coordinate system. These coordinates are then mapped to macroscopic GIS maps, mesoscopic BIM structural models, and even microscopic equipment CAD models, forming multi-scale and multi-level data associations, laying the foundation for all subsequent location-based intelligent analyses.

[0013] After establishing the precise location, the system will proactively trigger surrounding detection. This process is not physical detection, but rather uses API interfaces and database queries to retrieve raw data strongly related to the location from multiple heterogeneous information systems such as geological exploration, BIM, project management (PMS), and environmental monitoring. The system then cleans and structures this data, extracting it into machine-readable scene feature key-value pairs with clear physical meaning, such as geological parameters, design indicators, project progress, and real-time environmental data.

[0014] The system matches and infers the set of scene features extracted in the previous step with a pre-built knowledge base of usage scenarios. This knowledge base is defined by domain experts and covers all standard and non-standard working conditions in the project. Each scenario is described by a set of feature conditions. The system usually uses a rule-based expert system or a lightweight machine learning classification model (such as a decision tree) to compare the input features with the rules of the knowledge base, thereby accurately determining the most specific usage scenario of the payload box. The location identifier and usage scenario are semantically combined to form a highly condensed state definition. Its output will serve as the direct decision basis for all subsequent detection actions (S112).

[0015] Specifically, in a cross-sea bridge project, a pile foundation load cell was hoisted into pile hole No. 12 of the P3 main pier. The on-site engineer scanned the RFID tag on the load cell with a handheld device, and the system automatically read its unique serial number CLB-P3-12-001. At the same time, the total station uploaded the three-dimensional coordinates of the center point of the load cell (X=352145.67, Y=2876543.21, Z=-40.00) to the central control platform. The platform immediately dynamically bound this physical entity to the virtual load cell component at the -40m elevation of pile hole No. 12 of the P3 main pier in the BIM model, completing the precise mapping between the physical world and the digital world.

[0016] After completing the location mapping, the system immediately triggers surrounding detection. It retrieves data from the geological database and extracts the following features: {Soil layer 1: silty soft soil, thickness 15m, compression modulus 2.5MPa; Soil layer 2: silty clay, thickness 25m, compression modulus 6.8MPa; Groundwater: high water level, confined}; it retrieves the task from the PMS and extracts the following features: {Current process: static load test, test age: 28 days after pile concrete pouring, test type: single pile vertical compressive static load test}; it retrieves data from the marine environmental monitoring station and extracts the following features: {Current tide: rising tide, expected to reach highest tide level in 1 hour, tidal range: 4.5m, wind speed: level 5}.

[0017] The system inputs features such as {geology: soft soil / clay, groundwater: high, process: static load test, environment: strong tide} into the scene classification engine. The engine traverses and compares the features in the knowledge base, excluding conventional land-based static load test scenarios with low matching degree, and finally matches scenario C: static load test of soft soil foundation affected by tide with the defined features {geology: soft soil, groundwater: high, environment: periodic water level change}, and determines this to be the most accurate use scenario.

[0018] The system performs final information integration, combining location information P3 main pier No. 12 pile foundation, -40m elevation, and scene information soft soil foundation static load test affected by tides, and finally outputs an accurate usage description: The pile foundation load box located at the -40m elevation of P3 main pier No. 12 pile foundation is currently in the loading stage of the soft soil foundation static load test affected by strong tides.

[0019] Furthermore, based on the usage of the pile foundation load cell, the corresponding usage state is determined, and the stress detection of the pile foundation load cell is triggered in the usage state to dynamically monitor the stress detection of the pile foundation load cell. Based on the stress detection of the pile foundation load cell, multiple stress data of the pile foundation load cell in different dimensions are determined.

[0020] At this point, the system typically uses a finite state machine (FSM) model to manage the various states of the load cell, such as standby, loading, holding load, unloading, and abnormal pause. The state transition is driven by explicit triggering conditions, which can be external commands from the operator, automatic judgments by the system based on preset thresholds, or event-driven events triggered by other modules detecting anomalies. The system outputs a precise usage status, such as dynamic loading, which serves as the direct input for the next step, guiding the system on the priority and frequency of its operation.

[0021] This system employs a deterministic finite state machine (FSM) to manage and analyze the various states of the pile foundation load cell. The FSM model is formally defined by a quintuple M=(Q,Σ,δ,q0,F), where: Q (set of states): defines all discrete operating modes of the load cell throughout the entire test lifecycle; Σ (set of input events): the conditions triggering state transitions, derived from the parsed "usage status" text collected by S111, real-time sensor data streams, and decision commands; δ (transition function): defines how the system performs state transitions and associated data processing actions after receiving a specific input event in a specific state; q0 (initial state): the default state after the system is powered on or initialized; and F (set of final states): the state after the test ends normally or abnormally.

[0022] Optionally, the on-site engineer scans the RFID tag to read the serial number "CLB-P3-12-001". Combined with the 3D coordinates uploaded by the total station, this triggers the entity binding request in the input event set. The transfer function δδ executes the logical judgment accordingly, confirming that the physical entity and the virtual component at "P3 main pier No. 12 pile foundation -40m elevation" in the BIM model are successfully matched. The drive state jumps to q1 "registered state", completing the accurate mapping between the physical world and the digital world. Subsequently, the system triggers surrounding detection in the "registered state", automatically extracting multi-source feature vectors such as geology, process and marine environment. After the scene classification engine traverses and compares them, conventional land-based scenes are excluded, and "Scene C: Static load test of soft soil foundation affected by tides" is matched, thus generating a scene update event. After receiving this configuration event, the system state jumps from q1 to q2 "ready to load state" and outputs the final usage description. Through the operation of the FSM model, the system transforms isolated RFID signals, BIM coordinates, and tidal environment data into an ordered logical state flow. This not only enables precise positioning of the load cell but also presets specific environmental boundary conditions for subsequent "force balance state" calculations through scene recognition, ensuring the adaptability and accuracy of the intelligent detection system in complex marine environments.

[0023] The system does not collect all data in a fixed pattern. Instead, based on the usage state determined in the previous step, it calls the matching detection configuration file from the preset strategy library and triggers the corresponding detection task accordingly. Each state is bound to an independent configuration file, which defines in detail the sampling frequency, the list of sensors to be activated, the real-time data processing strategy (such as filtering and calibration), and the differentiated alarm thresholds for that state. For example, the dynamic loading state requires high-frequency sampling of 10Hz and full sensor activation, while the standby state only requires low-frequency sampling of 0.1Hz to save resources. The system sends the configuration file parameters to the data acquisition hardware (DAQ), activates the corresponding channel, and starts synchronizing the data stream according to the set frequency.

[0024] We introduce a high-precision, highly synchronous, and multi-dimensional set of stress data. This multi-dimensionality completely surpasses the traditional single PS curve. It covers macroscopic mechanical dimensions (total load P, total displacement S), local stress dimensions (strain distribution at key structural points), contact pressure dimensions (load transfer balance), spatial attitude dimensions (tilt, deflection), and physical effect dimensions (temperature rise, acoustic emission). Crucially, all dimensions of data must be timestamped by a unified high-precision clock (such as GPS timing or PTP protocol) to ensure absolute time synchronization. This is an absolute prerequisite for subsequent multimodal data fusion and causal analysis, ultimately forming a digital information that provides a comprehensive and three-dimensional description of the stress state of the load cell under specific usage conditions.

[0025] Specifically, when the system receives the output of S111, it parses the keyword "loading phase" and combines it with the start loading command just issued by the operator on the control software; the internal state machine switches from standby state to dynamic loading state; the system determines the usage state as dynamic loading, which tells the system that data acquisition needs to be performed with the highest priority and frequency.

[0026] The system is currently in a dynamic loading state, so it retrieves a configuration file named Loading_Profile.conf from the policy library. This file specifies: a sampling frequency of 10Hz; activation of {main pressure sensor, displacement sensor, 16-channel strain gauge, 8-channel earth pressure gauge, dual-axis tilt sensor, temperature sensor, acoustic emission sensor}; enabling real-time Kalman filtering; and setting alarm thresholds for strain > 300με, tilt angle > 0.1°, and acoustic emission event rate > 50 times / second. The system then sends instructions to the DAQ hardware, activating all specified sensor channels and starting operation with a 10Hz synchronous clock. The data stream is transmitted to the central processing unit in real time.

[0027] During dynamic loading, at a certain instant when the load reaches 5000kN (t=2025-10-0710:30:15.001), the system generates the following multi-dimensional data packet. This data packet is no longer an isolated number, but a digital content that provides a comprehensive and three-dimensional description of the stress situation of the load cell under a specific usage state.

[0028] Please see Figure 3 In step S12, the specific steps are as follows: S121: Acquire panoramic images of the pile foundation load cell, determine the overall shape of the pile foundation load cell based on the panoramic images, determine the first set of data based on the overall shape of the pile foundation load cell and its usage, determine the second set of data based on the overall shape of the pile foundation load cell and multiple stress data of the pile foundation load cell, and determine the multimodal data corresponding to the pile foundation load cell based on the first set of data and the second set of data. S122: In the multimodal data corresponding to the pile foundation load cell, the multimodal data is detected, and multiple data dimensions are determined during the detection process. The corresponding data combinations are determined by tracing along each data dimension.

[0029] In the embodiments of this application, given that the load cell is buried deep underground, a distributed optical fiber shape sensor array is deployed along the critical path on the surface of the load cell. Before pouring concrete, the optical fibers are lowered to the predetermined position along with the reinforcing cage. During the test, the bending strain data of the optical fibers are collected in real time using an optical fiber demodulator, input into a three-dimensional morphology reconstruction algorithm, and a digital panoramic three-dimensional model of the load cell is generated in real time, reflecting its morphological changes (such as dents, bulges, bending, etc.) during the compression process. Then, through meshing and texture mapping, a digital morphology model with a realistic appearance and accurate geometric dimensions is finally generated.

[0030] The system uses the usage information (such as location, working condition, and timestamp) determined in S111 as metadata and binds it strongly to the three-dimensional morphological model. This is usually achieved by embedding custom fields in the model file or establishing a relationship in the database. The purpose is to establish a baseline state, which provides a crucial reference system for subsequent anomaly detection and status assessment, so that the morphology is no longer isolated.

[0031] To ensure the synchronization of timestamps for all data, abstract data is loaded onto the 3D model using methods such as scalar field mapping (e.g., temperature and strain displayed as pseudo-color images), vector / tensor field mapping (e.g., displacement and stress displayed as arrows or contour maps), and dynamic deformation mapping (driving the displacement of model vertices to generate a 4D model). At this point, the system establishes an absolute synchronization mechanism, employing a unified network time protocol (NTP / PTP) to ensure that all acquisition nodes (e.g., high-frequency thermal imagers and low-frequency displacement gauges) carry UTC timestamps with microsecond-level precision. An extremely short time window is constructed centered on the rendering reference time (trender). Linear interpolation is used to calculate the precise values ​​of low-frequency data at the reference time, or the characteristic mean of high-frequency data within the window is statistically analyzed, thereby forcibly aligning all heterogeneous data to the same logical time slice.

[0032] For scalar data such as temperature, strain, and acoustic emission event rate, the system uses texture mapping and vertex shading techniques for loading. The system associates each sensor node with the nearest mesh vertex of the 3D model. For internal sensors, the system uses radial basis function (RBF) or inverse distance weighted (IDW) algorithms to interpolate the values ​​onto the model surface. The real-time acquired scalar value V is normalized to obtain vnorm, and this is used to index a preset pseudo-color spectrum (such as a Blue-Green-Red gradient). Finally, the corresponding RGB color values ​​are passed to the fragment shader in the GPU rendering pipeline to update the colors of the vertices on the model surface in real time, thereby generating a cloud map that intuitively reflects the stress or temperature distribution.

[0033] For directional vector or tensor data such as displacement, velocity, and principal stress, the system employs geometric primitive overlay and tensor visualization techniques for loading. For the displacement field, the system generates directional arrow primitives at key nodes of the model. The direction vector of the arrow is strictly consistent with the data direction, and the length is scaled proportionally according to the data size, deforming in real time as the data is refreshed to display the trend. For the stress tensor, the system calculates the principal stresses σ1 and σ3 based on strain rosette data. On the one hand, these can be converted into Von Mises equivalent stresses as scalars for cloud plot rendering. On the other hand, a "cross" rectangular icon can be drawn at the measurement point location. The long axis of the icon indicates the direction of the principal stress, the length of the axis indicates the stress magnitude, and the color distinguishes between tension and compression properties.

[0034] The system extracts three-dimensional displacement data (Δx, Δy, Δz) from the full-field displacement sensor and constructs a displacement vector field D covering the entire field. Read the original mesh vertex coordinates Pinitial of the 3D morphological model, and based on the displacement field data at the current moment, use the formula Pcurrent=Pinitial+D (Pinitial,trender) calculates the new coordinates of each vertex after compression; the GPU recalculates the normals based on the updated vertex coordinates Pcurrent and performs rasterization rendering, so that the model presents a dynamic geometric shape on the screen that "bulges", "compresses" or "skews" in sync with the compression process.

[0035] Typically, a data lake or time-series database is used as the underlying storage, with the first set (context information) serving as the metadata layer and the second set (spatiotemporal data) serving as the core content layer. The system establishes a unified index with spatiotemporal coordinates (x, y, z, t) as the primary key, allowing analysts or intelligent models to query all information at a specified time and spatial location. The final output is a highly structured, semantically rich multimodal data set, which includes images, data tables, and model files.

[0036] Specifically, in the static load test of pile foundation No. 12 of the P3 main pier of a cross-sea bridge, when the load was increased to 5000kN and entered the load holding stage, the system triggered S121; the distributed optical fiber shape sensing array started working, and the bending strain data of the optical fiber was collected in real time through the optical fiber demodulator; after processing by the background server, a digital panoramic three-dimensional model of the load box was generated in real time, generating a three-dimensional mesh model containing about 5 million vertices. The model clearly presents the overall structure of the pile foundation load box and quantitatively shows that: in the web area opposite the C-shaped opening, there is an inward concavity with a maximum depth of 2.3mm. This three-dimensional model is the overall shape of the load box under a 5000kN load.

[0037] The system associates the newly generated 3D morphological model (ID: MESH_CLB_P3-12_T1) with the output of S111; it writes its asset ID (CLB-P3-12-001), precise location (P3 main pier No. 12 pile foundation -40m), usage scenario (static load test of soft soil foundation affected by tides) and timestamp into the model's metadata. This 3D model, carrying complete identity and background information, constitutes the first layer of data set.

[0038] The system acquired the stress data of S112 at t=10:30:00 and fused it with the three-dimensional morphological model. Based on the predefined sensor coordinates in the BIM model, the system rendered the strain value of 155 με at point B of the web as yellow and the high strain value of 210 με on the opposite side of the opening web as a striking red stress cloud map. At the same time, a blue thermal map was superimposed based on the temperature data. This three-dimensional model, which superimposes the stress cloud map, temperature field and displacement vector, constitutes the second set of data.

[0039] The system integrates the first and second datasets and stores them in a dedicated multimodal database, creating an index with spatiotemporal coordinates as the primary key. Now, users can execute the query: SELECT * FROM Multimodal_Data WHERE Escape_ID='CLB-P3-12-001' AND Time. The query result will return multimodal data, which includes the geometry of the region, visual images, strain value of 210 με, temperature of 20.1°C, and the background of the soft soil static load test.

[0040] Furthermore, in the multimodal data corresponding to the pile foundation load cell, the multimodal data is detected, and multiple data dimensions are determined during the detection process. The corresponding data combinations are determined by tracing along each data dimension, thus determining multiple data combinations. This approach takes into account the overall consideration of tracing each data dimension, ensuring the accuracy of the corresponding data combinations. At the same time, multiple stress data of the pile foundation load cell in different dimensions are introduced to further control the multimodal data corresponding to the pile foundation load cell.

[0041] At this point, the system will scan the entire data volume, identify its internal data patterns, types, and relationships, and summarize them into several core data dimensions. This is a high-level data classification, which typically includes: a spatial dimension describing the three-dimensional spatial distribution and topological relationships; a temporal dimension describing the time evolution sequence; a physical / modal dimension representing different physical quantities such as force, displacement, and strain; and a semantic / contextual dimension carrying engineering significance and label information.

[0042] Tracing and drilling along one or more dimensions extracts data highly relevant to a specific analytical objective, forming multiple data combinations. The core of the tracing strategy is to select a primary dimension and aggregate relevant data from all other dimensions under that primary dimension. Different tracing strategies can generate different types of data combinations: tracing along the spatial dimension generates regional life history combinations for analyzing the evolutionary process of a specific region; tracing along the temporal dimension generates event information combinations for analyzing the overall picture of key events; tracing along the physical dimension generates pure modal relationship combinations for studying intrinsic physical relationships; and tracing along the semantic dimension generates stage archive combinations for reconstructing causal chains.

[0043] Specifically, the system detects the multimodal data volume of the pile foundation load cell generated by S121. After scanning, the system identifies that it contains a 3D mesh file in .obj format, a time series data table in .csv format, and a metadata file in .json format. Based on this, the system determines that the data volume contains the following four dimensions: spatial dimension (vertex coordinates and triangular facets in the .obj file), temporal dimension (millisecond-level timestamp column in the .csv file), physical dimension (Load, Displacement, Strain_B, etc. columns in the .csv file), and semantic dimension ("Scenario":"Tidal-Soft-Soil_Static_Load_Test", etc. information in the .json file).

[0044] The system generates data by combining data based on the four identified dimensions: Combination 1 (Spatial Tracing): By analyzing the stress cloud map, the system automatically identifies the web opposite the opening as a high-strain region (ROI-HighStrain). It traces along this spatial dimension and extracts all strain and temperature data of this region from the start of loading to the current moment, forming a data package named Combo-ROI-HighStrain_LifeHistory, which is used to analyze the damage evolution process of this region. Combination 2 (Time Tracing): The system detects a sharp increase in the acoustic emission event rate at t=2500s. It traces back along this time dimension, extracts all sensor data and corresponding panoramic images from the entire field within a 5-second time window before and after t=2500s, and forms a data package named Combo-Event-AE_Peak for in-depth analysis of the physical causes of the sudden change in acoustic emission. Combination 3 (Physical Tracing): In order to accurately evaluate the overall performance of the load cell, the system decides to analyze its load-displacement relationship. It traces along the two physical dimensions of total load and average displacement, extracts the corresponding data pairs throughout the loading process, and forms a classic two-dimensional dataset for plotting PS curves, named Combo-Physics_P_S_Curve. Combination 4 (Semantic Tracing): When the system receives the semantic instruction that the loading phase has ended, it traces along this semantic dimension, packages the time-series data of the entire loading phase, the final morphological model, and the context information of the phase into a data package named Combo-Phase-Loading_Complete, which serves as a complete record for archiving and subsequent reproduction of this phase.

[0045] Please see Figure 4 In step S13, the specific steps are as follows: S131: In multiple data combinations, identify multiple data combinations, determine the load characteristics of the corresponding pile foundation load cell during the identification process, and at the same time, monitor the compression status of the compression data in real time and collect the compression data corresponding to the compression data. S132: Determine the first level of intelligent detection information based on the characteristics of each load and the overall shape of the pile foundation load cell; determine the second level of intelligent detection information based on the overall shape of the pile foundation load cell and the corresponding compression data; and determine the digital twin system of the pile foundation load cell based on the first level of intelligent detection information and the second level of intelligent detection information.

[0046] In the embodiments of this application, multiple data combinations are identified among multiple resistors and multiple data combinations. During the identification process, the load characteristics of the corresponding pile foundation load cell are determined. At the same time, the compression status of the compression data is monitored in real time, and the compression data corresponding to the compression data is collected. The collection of the compression data corresponding to the compression data is introduced.

[0047] At this point, the system intelligently calls the matching feature extraction algorithm library according to the type of data combination (such as time series, spatial field, event combination). For time series, digital signal processing technology is applied to extract statistical, frequency domain, and time-frequency domain features. For spatial fields, computational geometry and image processing technology are applied to extract field distribution and geometric features. For event combinations, acoustic emission analysis technology is applied to extract event and sequence features. All low-order features are aggregated and engineered to form load features with clear physical meaning, such as mechanical response features reflecting load-bearing capacity (tangential stiffness, creep rate), damage evolution features assessing the degree of damage (plastic strain accumulation, crack propagation rate), and stability features judging the risk of instability (strain field inhomogeneity).

[0048] The system selects one or more key pressure data points (such as total load P or key displacement S) that best represent the working state of the load cell as the heartbeat indicators for real-time monitoring, and presets multi-level trigger conditions for them, including threshold trigger (exceeding the elastic limit), trend trigger (creep rate exceeding the limit), and event trigger (other subsystems emit abnormal signals). Once the trigger conditions are met, the system will automatically execute the pre-programmed acquisition script and dynamically adjust the acquisition behavior, including increasing the sampling frequency of key sensors, activating dormant high-precision redundant sensors, adjusting the acquisition focus of the mobile device (performing a staring scan of abnormal areas), and locking the data windows before and after the trigger moment for special marking and backup to ensure the integrity and traceability of the event process.

[0049] Specifically, the system is processing the data combination generated by S122; the system identifies Combo-Physics_P_S_Curve as a time series combination, and then applies signal processing algorithms to calculate that as the load increases from 4000kN to 5000kN, the tangent stiffness of the PS curve decreases from 200kN / mm to 170kN / mm; at the same time, the system identifies Combo-ROI-HighStrain_LifeHistory as a spatial field combination, and applies image processing algorithms to extract the maximum plastic strain of the web opposite the opening from the strain contour map as 320με, and calculates that the stress concentration factor in this region reaches 1.9; the system identifies the above results as two key load characteristics: {stiffness degradation rate: 15%} and {local plastic damage index: 0.85}.

[0050] The system selected the total load P as the heartbeat indicator and set 4500kN as the threshold for the high stress zone. When the experiment reached t=1800s, the real-time monitoring showed that the total load P just exceeded 4500kN. The system immediately triggered the high stress zone acquisition script: sent instructions to all strain gauges and displacement sensors to increase the sampling rate from 1Hz to 20Hz; sent instructions to abandon the conventional generation of the digital panoramic 3D model and instead perform a 30-second local high-precision shape sensing on the previously identified ROI-HighStrain region; at the same time, the system automatically encrypted and backed up all multimodal data from t=1790s to t=1810s and tagged it with Event_HighStress_Onset.

[0051] Furthermore, the first layer of intelligent detection information is determined based on various load characteristics and the overall shape of the pile foundation load cell. The second layer of intelligent detection information is determined based on the overall shape of the pile foundation load cell and the corresponding compression data. The digital twin system of the pile foundation load cell is determined based on the first and second layers of intelligent detection information, ensuring the accuracy of the digital twin system of the pile foundation load cell. At the same time, it takes into account various load characteristics, the overall shape of the pile foundation load cell, and the corresponding compression data, thereby improving the accuracy of the digital twin system of the pile foundation load cell.

[0052] At this point, the load characteristics (such as stiffness reduction rate) are fused with the overall shape (such as the indentation on the 3D model) based on the first layer of intelligent detection information, so as to make a qualitative or quantitative assessment of the structural integrity of the load cell. This is usually achieved through a rule-based engine or a shallow machine learning classifier to establish a causal relationship between load characteristics and geometric shape. Its output is the first layer of intelligent detection information, which can be a discrete classification label (such as structural integrity: good / minor damage) or a continuous score (such as structural health: 85 / 100), which directly reflects the state of the load cell.

[0053] Simultaneously, a "measuring point-space" mapping matrix is ​​established using a finite element model to map one-dimensional time-series load data to three-dimensional space, achieving spatial topological alignment of heterogeneous features. Subsequently, time-space coupling verification is performed. Based on the physical law that "deformation lags behind force," the moment t1 when the stiffness decrease rate exceeds the threshold is spatiotemporally matched with the subsequent abrupt indentation detected by the three-dimensional model. If the two highly overlap, it is determined to be a strong causal relationship of "structural plastic damage," otherwise it is identified as a weak relationship such as sensor drift. On this basis, the system constructs a fusion feature vector containing normalized stiffness decrease rate, indentation depth, damage area, and causal relationship strength coefficient. Qualitative classification labels are generated by a rule engine based on preset thresholds, or a shallow machine learning model is used to calculate a quantitative health score. The causal relationship strength coefficient serves as a penalty term to effectively filter false signals. This process successfully couples the abstract "stiffness decrease" with the concrete "model indentation," giving the output integrity assessment results a solid physical and mechanical basis and providing reliable initial values ​​for subsequent force equilibrium state calculations.

[0054] The system outputs structured and standardized "first-level intelligent detection information," which includes qualitative labels, quantitative scores, and key attribution descriptions. Regarding qualitative labels, the system maps the evaluation results to discrete engineering state labels, specifically including: Level 1 - Structural integrity (load characteristics are within the elastic range, and the overall shape shows no significant plastic deformation); Level 2 - Minor damage (local yielding is detected, i.e., a slight decrease in stiffness, but the overall shape remains stable, with no macroscopic distortion); Level 3 - Significant damage (significant stiffness degradation, measurable indentations or bulges appear on the 3D model, i.e., geometric changes, but not yet unstable); Level 4 - Critical failure (mechanical characteristics diverge exponentially, geometric shape is severely distorted such as buckling instability, and the structure is on the verge of collapse); in terms of quantitative scoring, the system outputs a continuous value between 0 and 100 to quantify the remaining health margin of the structure; in terms of causal attribution, the main basis of the evaluation results is explained in natural language, for example: "Since the stiffness reduction rate (12%) and the indentation depth of the web opposite the opening (1.8mm) show a strong positive correlation, the load cell is judged to be in a state of slight damage, mainly affected by local plastic buckling."

[0055] For the second layer of intelligent detection information, a direct causal correlation analysis is performed between the overall shape and the compression data to make a quantitative assessment of the working efficiency and coordination of the load cell. This usually requires combining structural mechanics principles and numerical simulation technology, calculating the actual stiffness distribution through reverse analysis, or comparing the measured shape with the prediction results of the ideal finite element model (FEM). Its output is the second layer of intelligent detection information, which is a series of quantitative performance indicators, such as load transfer efficiency: 92% and structural coordination coefficient: 0.88.

[0056] At this point, the displacement data of the "overall shape" collected by S121 and the "compression data" collected by S131 are input into the model for iterative calibration. The system automatically adjusts the boundary conditions (such as soil resistance) and contact stiffness parameters in the model to minimize the residual between the calculated displacement field and the measured shape field, thereby reproducing the "actual contact stiffness distribution" and "actual load distribution vector" that reflect the true state.

[0057] Based on this inversion result, the system outputs quantified "second-level intelligent detection information", which specifically includes: load transfer efficiency for measuring effective conversion of hydraulic capacity and uniformity, structural collaborative working coefficient for quantifying the deformation consistency of each component, and effective bearing area utilization rate calculated by identifying "suspended" areas. These indicators accurately reflect the internal force distribution mechanism and energy transfer efficiency of the load box.

[0058] Using the first and second layers of intelligent detection information as inputs, a core decision-making model generates a final, action-guided comprehensive judgment. This digital twin system is essentially a complex machine learning model (such as a deep neural network (DNN) or gradient boosting tree (XGBoost)) pre-trained using massive amounts of historical data, simulation data, and expert experience. The digital twin system learns the nonlinear mapping relationship between state, performance, and final risk level, and outputs a decision item, usually a health index between 0 and 1 or a discrete risk level (e.g., green - normal, yellow - warning, orange - dangerous, red - failure). This output will directly drive subsequent anomaly diagnosis and dynamic control.

[0059] At this point, the core decision-making model is a serial hybrid architecture of a deep neural network (DNN) based on a multimodal attention mechanism and a gradient boosting tree (XGBoost). It deeply fuses the "first-level intelligent detection information" and the "second-level intelligent detection information" through nonlinear mapping, outputting a quantitative risk decision. The model employs a three-layer architecture of "feature extraction - attention weighting - risk regression": In the input layer, channel A receives structural integrity data (such as health scores and damage levels), and channel B receives operational performance data (such as transmission efficiency and synergy coefficients); in the hidden layer, deep features are extracted through a fully connected layer, and the attention mechanism dynamically adjusts the weights of the two types of features according to different operating conditions (such as the initial loading stage or high-load stage), generating a fused feature vector; in the output layer, the sigmoid activation function is used to map the fused features to a comprehensive health index (HI) between 0 and 1, and a discrete risk level is output through a softmax classifier.

[0060] To explain the nonlinear mapping relationship, the Health Index (HI) output by the model is linked to the risk level: HI ∈ [0.85, 1.0] is considered green (normal), and the system maintains its strategy; HI ∈ [0.6, 0.85) is considered yellow (warning), and the system triggers enhanced monitoring; HI ∈ [0.3, 0.6) is considered orange (dangerous), and the system triggers dynamic control intervention; HI ∈ [0, 0.3) is considered red (failure), and the system performs emergency unloading. Furthermore, the model is trained using a massive dataset containing historical experiments, simulated failure cases, and expert labels. It employs a composite loss function with risk bias, assigning higher penalty weights to "high-risk" samples to ensure extremely high sensitivity to danger signals, thereby constructing an intelligent decision-making component capable of dynamic trade-offs.

[0061] Specifically, the system receives the load characteristics of S131 {stiffness degradation rate: 15%, local plastic damage index: 0.85} and the overall shape of S121 (the 3D model shows a 2.3mm indentation in the web opposite the C-shaped opening); the system inputs these data into a pre-trained decision tree classifier; the classifier makes judgments according to rules and outputs the first level of intelligent detection information: structural state assessment: it has entered the elastoplastic working stage, there is local plastic deformation, and the structural integrity is 85%.

[0062] The system continues to analyze the same set of data, inputting the overall shape of S121 (the 3D model shows asymmetrical deformation of the two arms of the C-shape) and the compression data of S112 (the current total load is 5000kN). The system superimposes and compares the measured deformation field with the simulation results of an ideal pile foundation load cell finite element model under 5000kN. The system calculates that the measured load transfer path deviates from the ideal model, and the forces on the two arms are unbalanced. Finally, it outputs the second layer of intelligent detection information: Performance analysis: The load transfer path has deviated, the collaborative workability of the C-shape structure has decreased, and the bearing efficiency is 90%.

[0063] The system merges the output information from the first two sub-steps: {structural integrity: 85%} and {bearing capacity: 90%}. These two values ​​are then input into a pre-trained deep neural network model. This model learns that when both integrity and capacity are below 90% and below 95%, the risk increases significantly. After complex nonlinear calculations, the model outputs: Health Index: 0.78, Risk Level: Yellow (Medium Risk). This model and its decision-making logic together constitute the digital twin system of the pile foundation load cell, accurately triggering subsequent anomaly diagnosis and dynamic control, truly achieving intelligent detection.

[0064] Please see Figure 5 In step S14, the specific steps are as follows: S141: Monitor the digital twin system in real time, identify multiple abnormal nodes based on the identification of the digital twin system, and mark multiple abnormal data of each abnormal node. In each abnormal node, determine the abnormal compression event of the pile foundation load cell based on the node location of the abnormal node, multiple abnormal data and the compression condition of the pile foundation load cell. S142: Based on the identification of abnormal compression events in the pile foundation load cell, multiple sub-abnormal compression contents are determined, and the corresponding abnormal compression areas are determined based on the detection of each sub-abnormal compression contents, so as to collect multiple abnormal compression areas and mark the regional location of each abnormal compression area. S143: Collect multiple working data of the pile foundation load cell, determine the working stage of the pile foundation load cell based on the multiple working data, determine the force balance frame based on the regional location of each abnormal compression area and the stress condition of the pile foundation load cell, and determine the force balance state of the pile foundation load cell based on the working stage of the pile foundation load cell and the force balance frame.

[0065] In the embodiments of this application, the digital twin system is monitored in real time, and multiple abnormal nodes are identified based on the identification of the digital twin system. Multiple abnormal data of each abnormal node are marked. In each abnormal node, the abnormal compression event of the pile foundation load cell is determined based on the node position of the abnormal node, multiple abnormal data and the compression condition of the pile foundation load cell. This approach takes into account the overall consideration of the node position of the abnormal node, multiple abnormal data and the compression condition of the pile foundation load cell, ensuring the accuracy of the abnormal compression event of the pile foundation load cell.

[0066] At this point, the system continuously and frequently monitors not the raw sensor data, but the final output of the digital twin system built by S13, namely the health index or risk level; more intelligent algorithms are used to identify abnormal nodes, such as applying change point detection algorithms such as CUSUM or EWMA to capture significant changes in the statistical characteristics of the health index time series, or to identify accelerated deterioration trends by calculating its rate of change, or even to determine abnormal nodes by matching with historical failure precursor patterns.

[0067] The system defines a time window centered on the abnormal node and locks all data within it. The system applies multiple anomaly detection algorithms in parallel to all data dimensions: statistical or density-based methods are used to identify abrupt changes in time-series data; inter-frame differencing, edge detection, or curvature analysis are used to identify newly generated displacements, cracks, or depressions in image or point cloud data; FFT transformation is performed on acoustic signals to detect abnormal high-frequency components; all identified abnormal data points, regions, or spectral peaks are tagged with timestamps and strongly correlated with the abnormal node.

[0068] The system performs spatiotemporal correlation analysis. If multiple anomalous data from different dimensions appear simultaneously within a very short time window and in the same or adjacent spatial locations, they are considered to point to the same physical cause. Through a rule engine or a lightweight causal inference model, highly correlated anomalous data are aggregated to generate a structured event description containing event ID, occurrence time, event type (such as local yielding, cracking), severity, and evidence summary. This complete process from node to data to event ensures that the system can quickly and accurately pinpoint the problem and present it in a way that engineers can understand.

[0069] Specifically, the system monitors the health index output by S13 at a frequency of 10Hz; at t=2495s, the index is 0.78 and has been slowly decreasing; at t=2500s, the system's CUSUM algorithm detects a sharp negative shift in the mean of the health index, with the index dropping sharply from 0.78 to 0.45 within 2 seconds; the algorithm is immediately triggered, marking the time point t=2500s as an abnormal node.

[0070] After marking t=2500s as an anomalous node, the system immediately locked all data within the 10-second time window from t=2495s to t=2505s for anomaly detection and marking. The system found that the strain value in the ROI-HighStrain region jumped instantaneously from 220με to 350με at t=2500s, far exceeding the 3σ range, and was marked as anomaly; the acoustic emission event rate surged from 10 times / second to 120 times / second at the same moment, and was marked as anomaly; the temperature in this region rose abnormally by 0.8°C within 5 seconds, and was also marked as anomaly.

[0071] The system performed a final analysis on all the abnormal data collected at t=2500s. The system found that the three abnormal phenomena of sudden strain increase, sudden acoustic emission surge, and temperature rise were completely synchronized in time and all pointed to the ROI-HighStrain region in space. The system's rule engine matched a rule: IF(strain increase AND sudden acoustic emission surge AND temperature rise) THEN(event type = local plastic instability). The system identified an abnormal compression event E-001, whose structured description is: at t=2500s, local plastic instability occurred in the web region of the pile foundation load box, accompanied by severe microcrack activity and temperature rise, with a severity of high. This clear event provides a clear starting point for subsequent precise location and stability analysis.

[0072] Furthermore, based on the identification of abnormal compression events in the pile foundation load cell, multiple sub-abnormal compression contents are determined. Based on the detection of each sub-abnormal compression content, the corresponding abnormal compression area is determined. Multiple abnormal compression areas are collected, and the location of each abnormal compression area is marked. This approach takes into account the overall consideration of detecting each sub-abnormal compression content and ensures the accuracy of the corresponding abnormal compression area.

[0073] At this point, the system maintains an engineering event ontology knowledge base, which defines the relationship between high-level events and underlying physical phenomena. When S141 identifies an abnormal event, the system queries this knowledge base and decomposes it into multiple sub-abnormal pressure contents with clear physical meanings, such as plastic yielding and microcrack initiation at the material level, or geometric buckling and local instability at the structural level. This process is achieved through rule-based mapping, transforming a vague event into a clear and quantifiable analysis target.

[0074] Optionally, the engineering event ontology knowledge base is a structured semantic network built using OWL (Web Ontology Language) or similar descriptive logic. It is not merely a data aggregation, but a formal definition of all possible physical phenomena and their logical relationships during the compression process of the pile foundation load cell. The core function of this knowledge base is to act as a "semantic translator," deconstructing the high-level decisions output by S141 (such as "orange - danger") into a set of physical parameters required by the specific analysis algorithms in S151. The knowledge base adopts a three-layer architecture model of "event-phenomenon-parameter," ensuring layer-by-layer penetration from macro to micro: High-level event layer: corresponding to the discrete risk levels or comprehensive decision items output by S141; Physical phenomenon layer: defining the specific physical mechanisms leading to the occurrence of high-level events; Quantitative parameter layer: specific physical indicators used to verify the existence of the above physical phenomena and extraction algorithm indexes.

[0075] For each sub-anomaly under pressure identified in the previous step, the system will invoke the most suitable detection algorithm to locate the specific physical location of the sub-content in the multimodal data, thereby determining the anomaly under pressure region. Each sub-content is associated with an optimal detection algorithm: for material plastic strain concentration, the system will analyze the strain contour map, identify sensor points exceeding the yield limit, and apply a spatial clustering algorithm to aggregate them into a continuous region; for microcrack activity surge, the system will use an acoustic emission sensor array to locate the sound source and perform density clustering on the sound source points; for geometric buckling, the system will compare the current and initial three-dimensional morphological models and identify the anomaly region by calculating the curvature change.

[0076] The system digitally collects data on each detected area, fully defining its spatial attributes such as boundaries, geometric center, and area. It also associates the sub-anomalies that caused the area to be identified as tags. For visual visualization, the system renders the boundaries and interiors of the anomaly areas in a striking color on a 3D morphological model and attaches the collected geometric attributes and tags as metadata. Finally, it outputs a list containing all identified and tagged anomaly areas. Each entry is a structured data object containing area ID, spatial coordinates, size, associated content, and severity, providing accurate geographic information for subsequent mechanical analysis and control decisions.

[0077] Specifically, the system receives the abnormal compression event E-001 output by S141: at t=2500s, local plastic instability occurs in the web area of ​​the pile foundation load box...; the system queries the knowledge base according to the event type local plastic instability, and determines that the event consists of the following two sub-abnormal compression contents according to the preset mapping rules: Content 1: material plastic strain concentration and Content 2: microcrack activity surge.

[0078] The system performs parallel region detection for two sub-contents. For the concentration of plastic strain in the material, the system filters out all grid points with strain values ​​>250με in the strain cloud map and applies the DBSCAN clustering algorithm to aggregate these points into an elliptical region of approximately 25cm x 20cm, which is initially defined as the abnormal compression region A. For the surge in microcrack activity, the system locates the sound source of acoustic emission events before and after t=2500s and finds that more than 90% of the sound source points are located near the geometric center of the abnormal compression region A. This greatly enhances the accuracy of the region A location and verifies that microcrack activity did indeed occur in this region.

[0079] After completing the area detection, the system performs data acquisition and labeling. The system acquires data from the abnormally compressed area A, recording its boundary vertex coordinates and center point coordinates (x_A, y_A, z_A), and associates them with the tags {material plastic strain concentration, microcrack activity surge}. On the 3D model of the pile foundation load cell, the system renders the precise extent of area A with semi-transparent red highlighting. The system generates a list of abnormal areas, containing one entry: {Region_ID:A,Location:(x_A,y_A,z_A),Size:25cmx20cm,Associated_Content:[Plastic_Yielding,Microcracking],Severity:High}, This precisely marked region will serve as a key input for introducing damage during subsequent mechanical analysis, achieving a precise leap from problem discovery to problem quantification.

[0080] Therefore, multiple working data points of the pile foundation load cell are collected, the working stage of the pile foundation load cell is determined based on the multiple working data points, the force balance frame is determined based on the regional location of each abnormal compression area and the stress condition of the pile foundation load cell, and the force balance state of the pile foundation load cell is determined based on the working stage of the pile foundation load cell and the force balance frame. This approach takes into account the overall consideration of the working stage of the pile foundation load cell and the force balance frame, ensuring the accuracy of the force balance state of the pile foundation load cell.

[0081] At this point, the system will activate a high-sensitivity infrared thermal imager and acoustic emission (AE) sensor to focus on and monitor the abnormally compressed area identified in S142, collecting temperature field data (temperature rise rate) and acoustic event flow data (event rate, energy) respectively. Based on a large amount of historical experimental data and material physics models, the system pre-trains a prediction model. By inputting the current {force balance state, thermodynamic / acoustic data} into this model, the corresponding instability load can be calculated, and a curve with physical symptoms as variables and instability load as dependent variable can be plotted. This curve shows how much load the structure can withstand without instability at the current damage rate.

[0082] The system utilizes acoustic data to construct a second equilibrium curve based on acoustic damage, similar to the S151 method. Employing the envelope method, the system selects the more conservative (lower value) of the two instability load prediction curves as the final equilibrium state curve to ensure the reliability of safety decisions. On this final curve, the system predefines several key detection nodes with different decision-making significance, such as early warning nodes prompting close operator attention, intervention nodes requiring control measures, and emergency nodes requiring immediate and stringent measures. At this point, the system defines a force balance state vector containing the current load, damage state quantities, acoustic emission energy rate, and local temperature rise rate, used to quantify the dynamic critical relationship between internal forces and external loads during damage evolution. Based on this state vector, a predictive model plots a monotonically decreasing curve with physical symptoms as variables and instability load as the dependent variable, intuitively reflecting the decreasing trend of the structure's remaining load-bearing capacity as damage increases.

[0083] The system employs the envelope method, selecting the smaller of two predicted curves based on thermodynamic and acoustic damage as the final equilibrium curve. Based on this, three key detection nodes are preset for step-by-step control: an early warning node is triggered when the curve value drops to 110%–120% of the current load, prompting close monitoring; an intervention node is triggered when the curve approaches the current load (105%–110%) and shows an accelerated decline inflection point, executing graded unloading; and an emergency node is triggered when the value falls below the current load, forcibly executing an emergency full unloading. Through this mechanism, the system transforms the abstract safety margin into a visualized, quantitative monitoring tool.

[0084] Meanwhile, to ensure the absolute reliability of safety decisions, the system constructs two independent prediction curves and fuses them using a conservative "envelope method": Curve 1 is a thermodynamic equilibrium curve, focusing on the input temperature rise rate and thermal damage factor, using infrared thermal imager data based on thermoelasticity and heat dissipation principles to predict strength loss due to material softening; Curve 2 is an acoustic damage equilibrium curve, focusing on the input acoustic emission energy rate and acoustic damage factor, using acoustic emission sensor data based on fracture mechanics and damage mechanics to predict structural fracture due to microcrack propagation. When generating the final equilibrium state curve, the system takes the smaller (more conservative) value of the two curves as the decision basis. This means that as long as any physical signal (thermal or acoustic) predicts impending instability, the system determines that a risk exists, thus effectively avoiding the potential for missed detections or blind spots that may exist with a single sensor.

[0085] At this point, curve one, the thermodynamic equilibrium curve, aims to capture the energy dissipation and macroscopic softening process of materials at the continuous medium level. Based on temperature field data collected by an infrared thermal imager, it extracts core feature vectors such as the temperature rise rate, maximum temperature difference, and temperature gradient. Based on the principle of irreversible thermodynamics, it uses a thermo-mechanical coupled constitutive model to invert and calculate the "thermal damage variable," establishing a mapping relationship between the temperature rise rate and the reduction of the effective stress area. This curve exhibits a smooth, gradually decreasing characteristic due to the hysteresis of heat conduction, and is highly sensitive to long-term creep, overall buckling, or large-scale plastic deformation. Curve two, the acoustic damage equilibrium curve, aims to capture the discrete fracture and crack propagation process of materials at the microstructural level. It uses an acoustic emission sensor array to collect... Based on the stress wave signal, acoustic emission energy rate, impact count rate, and waveform characteristic parameters are extracted. Based on fracture mechanics and damage accumulation theory, the "acoustic damage variable" is calculated using the Kaiser effect and Felicity effect to establish the relationship between acoustic emission energy release and stiffness degradation. This curve exhibits a step-like decrease or oscillating downward characteristic due to the randomness of crack propagation, and has extremely high sensitivity to brittle fracture, weld cracking, or local stress concentration. These two curves form a complementarity between "surface" and "point" in the sensitive domain, and constitute a combination of "hysteresis trend" and "real-time early warning" in terms of time response. They can also map different failure modes, thus providing a solid physical basis and mathematical diversity for subsequent conservative fusion using the "lower envelope method".

[0086] It should be noted that after constructing the thermodynamic equilibrium curve and the acoustic equilibrium curve separately, the two need to be merged into the final equilibrium state curve. Although the original physical quantities of the thermodynamic and acoustic data are different, after the damage variable transformation, the predicted instability loads output by the two curves are unified into force dimensions, so they can be directly compared and merged.

[0087] The system continuously collects the latest dynamic data and compares the current state point with the key detection nodes determined in S152 in real time. Once a certain condition is matched, the system's built-in decision engine generates a high-level, semantic dynamic control event (such as graded unloading to a safe load). The system translates this high-level instruction into a specific control instruction sequence that the underlying hardware can understand and sends it to the PLC or hydraulic control system of the loading system through the industrial bus to drive the actuator to move, thereby completing a complete closed loop from perception to decision-making to control.

[0088] Specifically, after determining the elastoplastic critical equilibrium in S14, the system immediately executes S151; the system focuses on the abnormally compressed region A, and the infrared thermal imager measures that its temperature rise rate has reached 0.2°C / min; the system inputs {force equilibrium state: critical equilibrium, temperature rise rate: 0.2°C / min} into the pre-trained thermodynamic damage model; the model calculates that at this damage rate, the structural instability load limit is 4850kN; based on this, the system generates the first equilibrium curve, which indicates that the current thermodynamic activity has reduced the structure's load-bearing capacity to below 4850kN.

[0089] At this point, the thermodynamic damage model and the acoustic damage model predict structural instability from different physical dimensions: The thermodynamic damage model is based on the principles of thermoelasticity and irreversible thermodynamics. By capturing the energy dissipation of the material during the elastic-plastic deformation process, it regards the abnormal temperature rise rate of the monitored area as the "thermal fingerprint" of internal damage dissipation. Using the coupled thermo-mechanical finite element correction model, it inverts the equivalent damage variable dT based on the input temperature rise rate, temperature rise range and historical plastic strain, and reduces the nominal ultimate bearing capacity through the logic of Fcriticalthermal=(1−dT)⋅Fultimatenominal.

[0090] The acoustic damage model, based on fracture mechanics and acoustic emission signal processing principles, captures transient elastic waves released during the initiation, propagation, and fracture of micro-cracks within the material. It uses the acoustic emission energy rate as a measure of the "fracture activity density" within the structure. Utilizing a strength degradation algorithm based on damage mechanics constitutive relations, it maps the damage variable dA using the acoustic emission energy rate, event count rate, and waveform characteristic parameters, and then predicts the instability load threshold based on the damage growth rate. In summary, the thermodynamic damage model focuses on reflecting the "surface" effect of overall or localized averaged plastic dissipation and is sensitive to slow plastic deformation; the acoustic damage model focuses on reflecting the "point" effect of dynamic propagation of micro-defects and is sensitive to brittle fracture or rapid instability precursors. Both models operate in parallel, simultaneously monitoring the risks in both "macro-softening" and "micro-fracture" dimensions, ensuring that the generated equilibrium curve comprehensively covers all possible failure modes.

[0091] The system inputs {force equilibrium state: critical equilibrium, acoustic emission event rate: 150 times / second} into the acoustic damage model; the model calculates that under this acoustic activity level, the structural instability load limit is only 4800kN; the system compares 4850kN (thermodynamic) and 4800kN (acoustic), and takes the more conservative 4800kN to construct the final equilibrium state curve; the system marks the intervention node on this curve: {load: 4800kN, acoustic emission event rate: 150 times / second}; the current load is 5000kN, and the state has seriously exceeded this node.

[0092] The system confirms that the current state {load: 5000kN,…} has exceeded {intervention node: 4800kN,…}; the decision engine matches the rules and generates a dynamic control event: graded unloading to a safe load; the system translates this event into a specific sequence of control commands: 1. Send a command to the hydraulic pump station to switch the loading mode to the unloading mode; 2. Send a command to the servo valve to smoothly reduce the load at a rate of -50kN / min; 3. When the load is reduced to 4500kN (a preset safe load), pause unloading and maintain it. These commands are sent and executed in real time, and the load on the pile foundation load cell begins to decrease smoothly. A potential instability accident is successfully avoided by the intelligent system.

[0093] Please see Figure 6 In step S15, the specific steps are as follows: S151: Anomaly detection is performed on the pile foundation load cell, and the corresponding thermal data is determined during the detection process. At the same time, the corresponding acoustic data is determined based on the acoustic detection of the pile foundation load cell. The acoustic data covers various crack sounds. The first equilibrium curve is determined based on the thermal data of the pile foundation load cell and the force balance state of the pile foundation load cell. S152: Determine the second equilibrium curve based on the acoustic data and force balance state of the pile foundation load cell, determine the equilibrium state curve of the pile foundation load cell based on the first and second equilibrium curves, and determine multiple key detection nodes based on the identification of the equilibrium state curve. S153: In multiple key detection nodes, multiple dynamic data of the pile foundation load cell are determined based on the tracing of each key detection node. Based on the multiple dynamic data, the current state of the pile foundation load cell and the corresponding working stage, the dynamic control event of the pile foundation load cell is determined to trigger the dynamic control measures of the pile foundation load cell.

[0094] In the embodiments of this application, anomaly detection is performed on the pile foundation load cell, and corresponding thermal data is determined during the detection process. At the same time, corresponding acoustic data is determined based on the acoustic detection of the pile foundation load cell. The acoustic data covers various crack sounds. A first equilibrium curve is determined based on the thermal data and force balance state of the pile foundation load cell. This approach takes into account both the thermal data and force balance state of the pile foundation load cell, ensuring the accuracy of the first equilibrium curve.

[0095] At this point, the system no longer focuses on the macroscopic mechanical state but delves into the material level. Through high-precision thermal imaging technology, it performs high-precision, high-frequency staring monitoring of the abnormally stressed areas identified in S14. The system will mobilize a high-resolution infrared thermal imager to collect a dynamic temperature field data stream and extract key indicators from it: the absolute temperature field to eliminate environmental interference, the temperature gradient field to identify the area with the most concentrated heat, and the temperature rise rate (°C / min), which is the most critical indicator and is used to quantify the severity of damage evolution. The scientific basis for this is that the plastic deformation and microcrack friction of the material will generate heat due to energy dissipation. Therefore, thermal data is the direct temperature content of the damage activity inside the material.

[0096] The system utilizes acoustic emission (AE) technology to capture transient elastic waves generated within materials due to stress release, converting them into analyzable acoustic data. The system activates high-sensitivity piezoelectric acoustic emission sensors arranged on the surface of the load cell and processes the acquired raw acoustic signals in real time, extracting a series of characteristic parameters describing crack sounds: acoustic emission event rate reflecting the degree of damage activity, event energy characterizing the severity of damage, ringing count related to signal duration, and dominant frequencies obtained through FFT analysis that can be used to distinguish different damage mechanism types. These data provide direct evidence for qualitative damage analysis.

[0097] By combining the collected thermodynamic data with the force equilibrium state of S14, a dynamic, physical damage-based first equilibrium curve is constructed through a prediction model. This curve is not a traditional load-displacement curve; its construction relies on a prediction model pre-trained based on a large amount of historical experimental data and a material physics model. The model's input is {force equilibrium state, thermodynamic data}, and its output is the predicted value of the thermodynamic instability load. Under the current operating conditions, the system inputs the real-time collected thermodynamic data into the model to obtain a predicted value of the instability load that dynamically decreases over time. This dynamic curve shows that the faster the temperature rises, the faster the safety margin is eroded.

[0098] Specifically, after determining that the system is in elastoplastic critical equilibrium in S14, the system immediately executes S151; the system activates the infrared thermal imager and continuously scans the abnormally compressed area A marked in S142; the system analyzes the temperature field data of area A and finds that its average temperature is steadily rising at a rate of 0.2°C / min, which is significantly higher than the 0.02°C / min of the background area; the system identifies {area A, temperature rise rate: 0.2°C / min, peak temperature gradient: 5°C / cm} as the current key thermodynamic data.

[0099] While performing thermodynamic monitoring, the system also performs acoustic detection in parallel. The system increases the sampling rate of the acoustic emission sensor to the highest level and analysis reveals that acoustic emission activity increases sharply near the abnormally stressed region A. The system extracts key acoustic data: {event rate: 150 times / second, average event energy: 80 aJ (at to Joules), dominant frequency: 150 kHz}. These data indicate that region A is experiencing intense microcrack propagation activity.

[0100] The system integrates the collected thermal data with the conclusions of S14; the system inputs {force equilibrium state: elastoplastic critical equilibrium, temperature rise rate: 0.2°C / min} into the pre-trained thermodynamic damage prediction model; the model calculates a prediction result based on the complex mapping relationship learned internally; the model outputs that, at the current temperature rise rate of 0.2°C / min, the thermodynamic instability load limit of the pile foundation load cell has decreased to 4850kN; the system generates the first equilibrium curve based on this, with its current point being (current load: 5000kN, predicted instability load: 4850kN). This result indicates that although the external load is 5000kN, due to the aggravation of internal thermal damage, the structure has actually exceeded its dynamic stability boundary.

[0101] Furthermore, a second equilibrium curve is determined based on the acoustic data and force balance state of the pile foundation load cell. The equilibrium state curve of the pile foundation load cell is determined based on the first and second equilibrium curves. Multiple key detection nodes are determined based on the identification of the equilibrium state curve. This approach incorporates the overall considerations of the first and second equilibrium curves, ensuring the accuracy of the equilibrium state curve of the pile foundation load cell.

[0102] At this point, the system uses acoustic emission data to construct a second, independent stability boundary curve based on the material fracture activity intensity. This curve shows how much load the structure can withstand without instability under the current level of acoustic damage activity. Its construction relies on a prediction model pre-trained based on a large amount of historical experimental data, especially data recording the complete acoustic emission activity from stability to instability. The input of this model is {force equilibrium state, acoustic data}, and the output is the predicted value of the acoustic instability load. The system inputs the acoustic data collected in real time into the model to obtain a predicted value of the instability load that dynamically decreases over time, forming the second equilibrium curve.

[0103] The thermodynamic and acoustic equilibrium curves are fused to form the most conservative and reliable final safety boundary. The system uses the lower envelope method for fusion, that is, the more conservative curve (predicting a lower instability load) is selected in real time as the final equilibrium state curve. This worst-case strategy is a standard practice in engineering safety-critical systems because structural instability is determined by its weakest link. Only when all safety boundaries are met simultaneously is the structure truly safe. The final equilibrium state curve represents the boundary defined by the mode that first reaches instability among all monitored damage modes.

[0104] At this point, regarding the fusion of the thermodynamic and acoustic equilibrium curves, the core of the system lies in constructing a unified spatiotemporal reference system and dynamically selecting the minimum load-bearing boundary based on the "shortest board effect" principle: establishing a high-resolution time reference axis, interpolating and aligning the two curves with different sampling frequencies to ensure that the thermodynamic and acoustic predicted instability loads can be obtained synchronously at any time $t$; performing minimum value comparison operations point by point on the unified time axis and drawing the lower envelope to ensure that the system always selects the value representing the pessimistic prediction as the effective boundary, and realizing the automatic switching of dominance between the thermodynamic and acoustic models based on the damage evolution characteristics; to prevent sensor noise interference, the system performs short-time window moving average filtering on the fused curve and implements consistency verification, locking the value only when a single curve does not experience a physical sudden drop to avoid misjudgment; the final generated equilibrium state curve not only contains the load value, but also includes a "dominant source identifier" that records the decisive factors, thereby fusing heterogeneous data into a continuous, smooth, and extremely conservative safety boundary, ensuring that the structural safety margin assessment is always based on the most stringent physical reality.

[0105] On the equilibrium state curve, key detection nodes with clear decision-making significance are defined. These nodes are signal points that trigger subsequent control measures. These nodes are not time points, but points in the state space, usually defined by a load value and a safety margin, and are pre-set on the final equilibrium state curve. The system typically defines multiple levels of nodes to achieve graded response: when the system state approaches but does not touch the curve, an early warning node is triggered to remind the operator; when the system state touches or just crosses the curve, an intervention node is triggered as a mandatory trigger point for automatic control measures; when the system state not only crosses the curve but also moves towards the instability region at an extremely rapid speed, an emergency node is triggered to trigger the highest level of emergency measures.

[0106] At this point, the key detection nodes are not fixed values, but rather decision boundaries dynamically generated in the multi-dimensional state space of {load-time} or {load-damage} based on the final equilibrium state curve. Key detection nodes are defined as a quadruple containing the trigger load value, safety margin, evolution rate, and response level. The system calculates node positions in real time based on the curve's geometric characteristics: early warning nodes are set in the safety buffer; intervention nodes correspond to zero safety margin; and emergency nodes target catastrophic precursors. Furthermore, the system introduces a time window sliding prediction mechanism, using historical data to fit short-term trends. When a rapid decline in carrying capacity is detected, the trigger positions of early warning and intervention nodes are automatically moved forward, thus reserving control execution lag time to ensure that these intelligent triggers based on real-time curve morphology and physical meaning can achieve accurate and timely graded responses.

[0107] Specifically, while executing S151, the system executes the first part of S152 in parallel. The system inputs {force balance state: elastoplastic critical balance, acoustic data: {event rate: 150 times / second, average event energy: 80aJ}} into the pre-trained acoustic damage prediction model. The model calculates based on the mapping relationship learned internally and outputs: at the current acoustic activity level of 150 times / second, the acoustic instability load limit of the pile foundation load cell has decreased to 4800kN. Based on this, the system generates a second equilibrium curve, whose current point is (current load: 5000kN, predicted instability load: 4800kN).

[0108] The system compares the instability load of 4850 kN predicted by the first equilibrium curve (thermodynamics) with the instability load of 4800 kN predicted by the second equilibrium curve (acoustics), and adopts the more conservative 4800 kN; the final equilibrium state curve of the system is determined as the boundary based on acoustic damage; the current state (5000 kN) has significantly exceeded this safe boundary defined by acoustic activity.

[0109] The system has a preset intervention node on the unstable load line of 4800kN. The system compares the current state (current load: 5000kN) with the intervention node (load: 4800kN). The system identifies that the current state has seriously exceeded the intervention node. This identification result will be transmitted to S153 as a high-priority event to trigger dynamic control measures.

[0110] Therefore, based on the tracing of each key detection node, multiple dynamic data of the pile foundation load cell are determined. Based on these multiple dynamic data, the current state of the pile foundation load cell, and the corresponding working stage, dynamic control events for the pile foundation load cell are determined to trigger dynamic control measures. This approach integrates multiple dynamic data, the current state of the pile foundation load cell, and the corresponding working stage, ensuring the accuracy of the dynamic control events. Simultaneously, it further controls the force balance state of the pile foundation load cell, achieving a holistic consideration of the thermal data, force balance state, and acoustic data of the pile foundation load cell. This improves the accuracy of the equilibrium curve of the pile foundation load cell and determines its dynamic control events.

[0111] At this point, the system defines an extremely short data tracing window centered on that moment and collects all real-time data streams within the window with the highest priority and sampling rate, including load data, deformation data, thermal data, acoustic data, and equilibrium state data. The system then precisely matches the real-time state points formed by the collected dynamic data with the theoretical conditions for triggering key detection nodes. This is not only to confirm the effectiveness of the triggering but also to quantify the degree of exceedance, such as calculating the difference between the current load and the predicted instability load.

[0112] The quantified status quo from the previous step is integrated with the system's known macroscopic state (such as force balance state) and working condition background (such as working stage). Through a decision engine, a semantic, high-level dynamic control event is ultimately generated. The input of the decision engine is a multi-dimensional set of information, including dynamic data that provides quantitative evidence, the current state that provides macroscopic background, and the working stage that defines the working condition. The decision engine is an expert system based on rules or machine learning models. Its output is not a specific hardware instruction, but an event with clear engineering significance, such as graded unloading to a safe load, immediate cessation of loading and holding, or emergency rapid unloading.

[0113] The system maintains an event-instruction mapping table. When a dynamic control event is determined, the system queries this mapping table and translates it into one or more specific, ordered sequences of control instructions, such as mode switching instructions (switching from load holding to unloading), rate control instructions (setting a precise unloading rate), and target value instructions (setting a target load and automatically stopping). The translated instructions are encapsulated into data packets conforming to industrial communication protocols and sent in real time to the execution unit of the load control system via buses such as EtherCAT or Modbus TCP / IP to drive hardware devices such as PLCs and servo valves to perform actions, thereby completing the closed loop of the entire intelligent detection and control.

[0114] Specifically, after the system identifies the intervention node as triggered in S152, it immediately executes the first step of S153. The system locks t=2505s as the trigger time and defines a 3-second data traceability window. The system collects key data within the window: {Current load: 5000kN, load change rate: 0kN / min (load maintenance state), temperature rise rate: 0.25°C / min, acoustic emission event rate: 160 times / second, predicted instability load: 4800kN}. The system confirms that the current load of 5000kN does indeed exceed the predicted instability load of 4800kN, with an excess of 200kN. The state matching is confirmed, the trigger is effective, and the situation is serious.

[0115] The decision engine receives the input: {Dynamic data: 200kN more than predicted instability load, current state: elastoplastic critical equilibrium, working stage: high stress load protection}; the engine matches a core safety rule: in the high stress load protection stage, once it enters critical equilibrium and exceeds the predicted instability, the load must be reduced immediately to restore the safety margin; the system finally determines a dynamic control event: graded unloading to the safe load.

[0116] The system translates the graded unloading to safe load event into specific actions; the system queries the mapping table and generates a sequence of instructions: 1. Send instructions to the hydraulic station PLC, setting Mode=Unload; 2. Send instructions to the servo valve controller, setting Target_Rate=-50kN / min; 3. Send instructions to the PLC, setting Target_Load=4500kN, Action=Hold. These instructions are sent to the loading control system via the EtherCAT bus; the hydraulic system responds, the loading jack begins to return oil, and the load borne by the pile foundation load box begins to decrease steadily at a rate of 50kN / min; a potential instability accident is successfully avoided by the intelligent system, and the entire detection and control closed loop is completed.

[0117] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A method for detecting pile foundation load cells based on multimodal data, characterized in that, include: The usage status of the pile foundation load cell is determined based on its current location and corresponding usage scenario, and the stress detection of the pile foundation load cell is triggered to collect multiple stress data of the pile foundation load cell in different dimensions. Based on the usage of the pile foundation load cell, the corresponding multiple stress data, and the overall shape of the pile foundation load cell, the multimodal data corresponding to the pile foundation load cell are determined, and multiple data combinations are determined; the multimodal data includes images, data tables, and model files; The load characteristics of the pile foundation load cell are determined by identifying each data combination. Based on each load characteristic, the overall shape of the pile foundation load cell, and the corresponding compression data, the digital twin system of the pile foundation load cell is determined. The digital twin system learns the non-linear mapping relationship between state, performance and final risk level, and outputs a decision item; The abnormal compression events of the pile foundation load cell are identified based on the identification of the digital twin system. Based on the detection of the abnormal compression events, multiple abnormal compression areas are identified. The force balance state of the pile foundation load cell is determined according to the regional location of each abnormal compression area and the working stage of the pile foundation load cell. Based on the thermal data, force balance state, and acoustic data of the pile foundation load cell, the equilibrium state curve of the pile foundation load cell is determined. Based on the equilibrium state curve of the pile foundation load cell, several key detection nodes are determined, and the dynamic control events of the pile foundation load cell are determined.

2. The method for detecting pile foundation load cells based on multimodal data according to claim 1, characterized in that, The process involves determining the usage status of the pile foundation load cell based on its current location and corresponding usage scenario, and triggering stress detection of the pile foundation load cell to collect multiple stress data from different dimensions, including: Mark the current position of the pile foundation load box, trigger the surrounding detection based on the current position of the pile foundation load box, determine multiple scene features during the surrounding detection process, determine the usage scenario of the pile foundation load box based on the multiple scene features and the working task of the pile foundation load box, and determine the usage status of the pile foundation load box based on the current position of the pile foundation load box and the corresponding usage scenario. Based on the usage of the pile foundation load cell, the corresponding usage state is determined, and the stress detection of the pile foundation load cell is triggered in the usage state to dynamically monitor the stress detection of the pile foundation load cell. Based on the stress detection of the pile foundation load cell, multiple stress data of the pile foundation load cell in different dimensions are determined.

3. The method for detecting pile foundation load cells based on multimodal data according to claim 1, characterized in that, The multimodal data corresponding to the pile foundation load cell are determined based on its usage, corresponding multiple stress data, and overall shape, and multiple data combinations are determined, including: A panoramic image of the pile foundation load cell is acquired. The overall shape of the pile foundation load cell is determined based on the panoramic image. The first set of data is determined based on the overall shape of the pile foundation load cell and its usage. The second set of data is determined based on the overall shape of the pile foundation load cell and multiple stress data of the pile foundation load cell. The multimodal data corresponding to the pile foundation load cell is determined based on the first and second sets of data. In the multimodal data corresponding to the pile foundation load cell, the multimodal data is detected, and multiple data dimensions are determined during the detection process. The corresponding data combinations are determined by tracing along each data dimension.

4. The method for detecting pile foundation load cells based on multimodal data according to claim 1, characterized in that, The process of determining the load characteristics of the pile foundation load cell based on the identification of various data combinations, and determining the digital twin system of the pile foundation load cell based on each load characteristic, the overall shape of the pile foundation load cell, and the corresponding compression data, includes: In multiple data combinations, multiple data combinations are identified, and the load characteristics of the corresponding pile foundation load cell are determined during the identification process. At the same time, the compression status of the compression data is monitored in real time, and the corresponding compression data is collected.

5. The method for detecting pile foundation load cells based on multimodal data according to claim 4, characterized in that, The process of determining the load characteristics of the pile foundation load cell based on the identification of various data combinations, and determining the digital twin system of the pile foundation load cell based on each load characteristic, the overall shape of the pile foundation load cell, and the corresponding compression data, further includes: The first layer of intelligent detection information is determined based on the characteristics of each load and the overall shape of the pile foundation load cell. The second layer of intelligent detection information is determined based on the overall shape of the pile foundation load cell and the corresponding compression data. The digital twin system of the pile foundation load cell is determined based on the first layer of intelligent detection information and the second layer of intelligent detection information.

6. The method for detecting pile foundation load cells based on multimodal data according to claim 1, characterized in that, The process of identifying abnormal compression events of the pile foundation load cell based on the digital twin system, determining multiple corresponding abnormal compression regions based on the detection of these abnormal compression events, and determining the force balance state of the pile foundation load cell based on the location of each abnormal compression region and the working stage of the pile foundation load cell includes: The digital twin system is monitored in real time. Multiple abnormal nodes are identified based on the identification of the digital twin system, and multiple abnormal data of each abnormal node are marked. In each abnormal node, the abnormal compression event of the pile foundation load cell is determined based on the node location of the abnormal node, multiple abnormal data and the compression condition of the pile foundation load cell. Based on the identification of abnormal compression events in the pile foundation load cell, multiple sub-abnormal compression contents are determined. Based on the detection of each sub-abnormal compression content, the corresponding abnormal compression area is determined, so as to collect multiple abnormal compression areas and mark the location of each abnormal compression area.

7. The method for detecting pile foundation load cells based on multimodal data according to claim 6, characterized in that, The process of identifying abnormal compression events of the pile foundation load cell based on the digital twin system, determining multiple corresponding abnormal compression regions based on the detection of these abnormal compression events, and determining the force balance state of the pile foundation load cell based on the location of each abnormal compression region and the working stage of the pile foundation load cell, further includes: Multiple working data points of the pile foundation load cell are collected. The working stage of the pile foundation load cell is determined based on the multiple working data points. The force balance frame is determined based on the regional location of each abnormal compression zone and the stress condition of the pile foundation load cell. The force balance state of the pile foundation load cell is determined based on the working stage of the pile foundation load cell and the force balance frame.

8. The method for detecting pile foundation load cells based on multimodal data according to claim 1, characterized in that, The equilibrium state curve of the pile foundation load cell is determined based on its thermal data, force balance state, and acoustic data. Multiple key detection nodes are then identified based on this equilibrium state curve, and the dynamic control events of the pile foundation load cell are determined, including: Anomaly detection was performed on the pile foundation load cell, and the corresponding thermal data was determined during the detection process. At the same time, the corresponding acoustic data was determined based on the acoustic detection of the pile foundation load cell. The acoustic data covered various crack sounds. The first equilibrium curve was determined based on the thermal data of the pile foundation load cell and the force balance state of the pile foundation load cell. The second equilibrium curve is determined based on the acoustic data and force balance state of the pile foundation load cell. The equilibrium state curve of the pile foundation load cell is determined based on the first and second equilibrium curves. Several key detection nodes are determined based on the identification of the equilibrium state curve.

9. The method for detecting pile foundation load cells based on multimodal data according to claim 8, characterized in that, The process of determining the equilibrium state curve of the pile foundation load cell based on its thermal data, force balance state, and acoustic data; determining multiple key detection nodes based on the equilibrium state curve; and determining the dynamic control events of the pile foundation load cell also includes: In multiple key detection nodes, multiple dynamic data of the pile foundation load cell are determined based on the tracing of each key detection node. Based on the multiple dynamic data, the current state of the pile foundation load cell and the corresponding working stage, the dynamic control event of the pile foundation load cell is determined to trigger the dynamic control measures of the pile foundation load cell.

Images