A method and system for monitoring the stability of soil and rock masses based on high-density electrical resistivity tomography.

Abstract

This invention provides a method and system for monitoring the stability of soil and rock masses based on high-density electrical resistivity methods (EDS), comprising: acquiring apparent resistivity data of the target soil and rock mass monitoring area based on a high-density electrode array; using a parallelized three-dimensional resistivity inversion algorithm to invert the apparent resistivity data into a true three-dimensional resistivity model of the soil and rock mass, and storing the inversion results in time sequence to construct a four-dimensional resistivity database; spatiotemporally aligning and fusing the four-dimensional resistivity database with other monitoring data of the target soil and rock mass monitoring area to obtain an analysis dataset; wherein, the other monitoring data include rainfall, groundwater level, and surface displacement; training a neural network model based on the analysis dataset to construct a soil and rock mass stability monitoring model; and performing real-time monitoring of the soil and rock mass to be monitored based on the soil and rock mass stability monitoring model to obtain monitoring results. This invention elevates the traditional decentralized, delayed, and superficial monitoring mode into an integrated, real-time, and in-depth intelligent early warning system.

Description

Technical Field

[0001] This invention belongs to the field of geotechnical engineering, specifically relating to a method and system for monitoring the stability of soil and rock masses based on high-density electrical resistivity tomography. Background Technology

[0002] High-density electrical resistivity tomography (EDS) is an array-based exploration method that infers the structure and properties of underground media by measuring the apparent resistivity distribution of soil and rock masses. Its traditional applications have primarily focused on static, one-off geological surveys. The root causes of soil and rock instability (such as landslides, slope collapses, and dam seepage) are often closely related to groundwater migration, changes in soil and rock moisture content, and fissure development, all of which lead to significant changes in soil and rock resistivity. Currently, soil and rock stability monitoring mostly employs point-based monitoring methods such as displacement gauges, inclinometers, and GPS, which have the following limitations: 1. Point-to-area: It is difficult to comprehensively capture the overall changes in the entire monitoring area, potentially missing key potential hazard areas. 2. Lag: Alarms are usually triggered only after macroscopic deformation occurs, missing the optimal warning time. 3. Difficulty in revealing internal mechanisms: It cannot intuitively reflect the internal processes leading to instability, such as water infiltration and the formation of potential slip surfaces. Although some studies have attempted to use high-density EDS for monitoring, these are mostly limited to periodic manual repetitive measurements, resulting in low efficiency, poor continuity, reliance on human experience for data interpretation, and the inability to achieve real-time early warning. Summary of the Invention

[0003] To address the problems existing in the prior art, this invention provides a method and system for monitoring the stability of soil and rock masses based on high-density electrical resistivity methods. It enables continuous capture of the four-dimensional (three-dimensional spatial + time-dimensional) evolution of the resistivity field within the soil and rock mass, and extracts leading information on stability changes to achieve accurate early warning.

[0004] To achieve the above objectives, the present invention provides the following solution: A method for monitoring the stability of soil and rock masses based on high-density electrical resistivity tomography (EDT) includes: Electrodes with integrated micro-acquisition circuits and identification tags are deployed in the target soil and rock monitoring area, and each electrode is connected by the Internet of Things to build a high-density electrode array. Based on the high-density electrode array, the apparent resistivity data of the target soil and rock monitoring area are collected; Using a parallelized three-dimensional resistivity inversion algorithm, the apparent resistivity data is inverted into a true three-dimensional resistivity model of the soil and rock body, and the inversion results are stored in time sequence to construct a four-dimensional resistivity database. The resistivity four-dimensional database is spatiotemporally aligned and fused with other monitoring data from the target soil and rock monitoring area to obtain an analysis dataset; wherein, the other monitoring data includes rainfall, groundwater level and surface displacement. A neural network model was trained based on the analyzed dataset to construct a rock and soil stability monitoring model. Based on the aforementioned rock and soil stability monitoring model, the rock and soil mass to be monitored is monitored in real time to obtain monitoring results.

[0005] Preferably, the method for constructing a high-density electrode array includes: The miniature acquisition circuit is integrated into the sealed chamber at the tail of the electrode, and each electrode is given a unique identification. Based on the survey results of the target soil and rock monitoring area, an optimal three-dimensional electrode layout scheme is established; Based on the optimal three-dimensional electrode layout scheme, each electrode with a unique identifier is used as a slave node, and relay nodes are selected from all slave nodes according to a preset ratio to complete the construction of a high-density electrode array.

[0006] Preferably, the inversion method for the true resistivity three-dimensional model includes: The collected apparent resistivity data is compared with historical normal data, and the apparent resistivity data that does not meet the quality requirements is processed to obtain a complete apparent resistivity dataset. The three-dimensional model mesh of the target soil and rock monitoring area is divided into multiple sub-regions and assigned to different computing nodes for parallel computing to obtain the predicted apparent resistivity and sensitivity matrix. Construct a spatiotemporal constraint objective function that includes data fit difference, spatial constraints, and time constraints; The Gauss-Newton method is used for iterative solution to minimize the spatiotemporal constraint objective function, update the resistivity in the three-dimensional model, until the preset convergence condition is met, and obtain the true resistivity three-dimensional model.

[0007] Preferably, the method for obtaining the analysis dataset includes: The other monitoring data are treated as spatially discrete point data, while the true resistivity three-dimensional model is treated as spatially continuous volume data. Establish a unified coordinate system for the target soil and rock mass monitoring area. For each data point, based on its corresponding physical meaning and monitoring principle, find the spatial influence domain on the mass data to obtain a spatially aligned dataset. Based on the network time protocol, the timestamps of all data in the spatially aligned dataset are uniformly verified, synchronized and standardized to obtain the spatiotemporally aligned dataset; Each resistivity voxel is used as a basic unit, and a multidimensional feature vector is assigned; wherein, the elements of the multidimensional feature vector include the voxel's own resistivity value, resistivity change rate, and other monitoring data and derived features associated with the corresponding voxel through spatial alignment. Combine the feature vectors of all voxels at all time steps to construct a spatiotemporal data cube; Based on the spatiotemporal data cube, a structured analysis dataset indexed by spatiotemporal units is constructed.

[0008] Preferably, the method for constructing the soil and rock stability monitoring model includes: Based on a unified coordinate system, resistivity voxels, rainfall, groundwater level and ground displacement are defined as different types of nodes, and a heterogeneous graph with embedded physical knowledge is established according to the physical relationship between the data. Using a meta-path-aware heterogeneous graph attention network, information about each node and its neighboring nodes is aggregated to obtain node embeddings containing physical semantics. The embedding sequences of each node are input into a temporal convolutional network to capture the dynamic evolution patterns of each node. Through a gating mechanism, the node state is transferred in time and diffused in space to obtain the spatiotemporal evolution characteristics of the analysis dataset. Based on the aforementioned spatiotemporal evolution characteristics and combined with physical constraint losses, the stability monitoring model for the soil and rock mass is constructed.

[0009] The present invention also provides a rock and soil stability monitoring system based on high-density electrical resistivity tomography (EDT) for implementing the method, comprising: The electrode array construction module is used to deploy electrodes with integrated micro-acquisition circuits and identification tags in the target soil and rock monitoring area, and connect each electrode to the Internet of Things to build a high-density electrode array. The data acquisition module is used to acquire apparent resistivity data of the target soil and rock monitoring area based on the high-density electrode array. The inversion module is used to invert the apparent resistivity data into a true resistivity three-dimensional model of the soil and rock body using a parallelized resistivity three-dimensional inversion algorithm, and to store the inversion results in time sequence to build a resistivity four-dimensional database. The dataset construction module is used to perform spatiotemporal alignment and fusion of the resistivity four-dimensional database with other monitoring data in the target soil and rock monitoring area to obtain an analysis dataset; wherein, the other monitoring data includes rainfall, groundwater level and surface displacement; The model building module is used to train a neural network model based on the analysis dataset and build a rock and soil stability monitoring model. The stability monitoring module is used to perform real-time monitoring of the soil and rock mass to be monitored based on the soil and rock mass stability monitoring model, and to obtain monitoring results.

[0010] Preferably, the electrode array construction module includes: The identification unit is used to integrate the miniature acquisition circuitry into the sealed chamber at the tail of the electrode and to assign a unique identification to each electrode. The deployment scheme design unit is used to establish the optimal three-dimensional electrode deployment scheme based on the survey results of the target soil and rock monitoring area; The electrode deployment unit is used to construct a high-density electrode array by using each electrode with a unique identifier as a slave node based on the optimal three-dimensional electrode deployment scheme, and selecting relay nodes from all slave nodes according to a preset ratio.

[0011] Preferably, the inversion module includes: The preprocessing unit is used to compare the collected apparent resistivity data with historical normal data and process the apparent resistivity data that does not meet the quality requirements to obtain a complete apparent resistivity dataset. Parallel computing units are used to divide the three-dimensional model mesh of the target soil and rock monitoring area into multiple sub-regions and assign them to different computing nodes for parallel computing to obtain the predicted apparent resistivity and sensitivity matrix. The objective function construction unit is used to construct a spatiotemporal constraint objective function that includes data fit difference, spatial constraints, and time constraints. The iterative solution unit is used to perform iterative solution using the Gauss-Newton method, minimize the spatiotemporal constraint objective function, update the resistivity in the three-dimensional model, until the preset convergence condition is met, and obtain the true resistivity three-dimensional model.

[0012] Compared with existing technologies, the beneficial effects of this invention are as follows: It achieves large-scale, three-dimensional, full-section, non-destructive, and continuous visualization of the interior of soil and rock masses. Traditional methods (such as inclinometers and displacement gauges) only provide point-line monitoring, failing to fully grasp the formation and evolution of anomalies (such as potential slip surfaces, seepage channels, and weak interlayers) in deep three-dimensional space. This technology, through high-density electrode arrays and three-dimensional inversion, can intuitively "see" the water infiltration process, the spatial distribution of weak zones, and their dynamic changes, achieving deep localization and spatial delineation of stability hazards. It significantly advances the warning time window, achieving true advanced warning. Before macroscopic deformation occurs, this technology can detect anomalies through the temporal changes in the resistivity three-dimensional model (such as the expansion of low-resistivity zones and sudden drops in resistivity), thus issuing warnings in the early stages of deformation acceleration or even earlier, leaving valuable time for disaster relief. It reduces reliance on human labor and improves the objectivity, accuracy, and efficiency of assessment. It overcomes the potential for false alarms / missed alarms from a single information source. AI models comprehensively analyze multiple signals such as "abnormal resistivity + rising water level + continuous rainfall + slight displacement changes," and their conclusions are far more reliable than those of a single indicator. Neural network models can learn complex, nonlinear instability precursor patterns from massive amounts of four-dimensional data that are difficult for the human brain to detect, identifying dangerous states caused by the coupling of various factors and making more accurate judgments. This eliminates the need for complex cable deployment using traditional high-density electrical methods, solving the pain points of cable damage, high maintenance workload, and high costs, making large-scale, long-term deployment possible. Attached Figure Description

[0013] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 This is a flowchart of a method for monitoring the stability of soil and rock masses based on high-density electrical resistivity tomography (EDT) according to an embodiment of the present invention. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0017] Example 1 like Figure 1 As shown, a method for monitoring the stability of soil and rock masses based on high-density electrical resistivity tomography includes: S1: Electrodes integrating micro-acquisition circuits and identification tags are deployed in the target soil and rock monitoring area, and each electrode is connected via the Internet of Things to construct a high-density electrode array. A further implementation method for constructing the high-density electrode array includes: The miniature acquisition circuitry is integrated into a sealed chamber at the tail of the electrode, and each electrode is assigned a unique identifier. Special materials with low polarization, high corrosion resistance, and high mechanical strength, such as platinum-plated titanium alloy or stainless steel special alloys, are used to ensure long-term electrochemical stability and lifespan. The miniature acquisition circuitry is used for multi-parameter sensing and localized preliminary processing. The multi-parameter sensing component can not only measure traditional voltage / current but also simultaneously measure the electrode's own temperature (for temperature compensation) and grounding resistance (for self-diagnosis). An integrated microprocessor (MCU) with analog-to-digital conversion (ADC) and local computing capabilities allows for preliminary filtering, noise reduction, and packetization before data transmission, enabling edge computing. Each electrode is assigned a globally unique ID number at the factory and written to the chip. This ID is strongly bound to its physical location (such as borehole number and depth) and theoretical coordinates in the system database. The electrodes can periodically self-test, measuring their own loop impedance to determine if there are faults such as open circuits, short circuits, excessive corrosion, or poor contact, and reporting the status code along with the data. Low-power wide-area network (LPWAN) technology is used, specifically LoRa (long-range radio) or NB-IoT. These two technologies feature low power consumption, long range, and strong penetration, making them ideal for outdoor and underground environments.

[0018] Based on the survey results of the target soil and rock monitoring area, an optimal three-dimensional electrode layout scheme is established. Specifically, before construction, BIM or GIS software is used to design the optimal three-dimensional electrode layout scheme based on the survey results. A unique UID is pre-assigned to each planned electrode point in the software, and a deployment map is generated, essentially creating a "digital twin" of the array in the digital world. Construction personnel drill holes and bury smart electrode nodes at designated coordinates according to the deployment map. A dedicated handheld smart terminal (PDA) is brought close to the buried node. The PDA reads the node's UID via NFC (Near Field Communication) or Bluetooth, binds it with the point's GPS coordinates and design depth with a single click, and automatically uploads it to the cloud platform database. This process establishes a precise mapping relationship between the physical and digital worlds, ensuring absolute consistency between "identity" and "location," and eliminating errors that may occur with manual recording.

[0019] After all electrodes are installed and powered on, the gateway sends a global wake-up signal. Electrodes automatically perform neighbor discovery and link quality detection via LoRa signals, autonomously forming an optimal, multi-path Mesh network topology, and reporting the topology structure to the gateway.

[0020] During non-data acquisition periods, the electrodes are in a low-power sleep state, maintaining only signal listening, significantly extending battery life (perpetual operation is possible if powered by solar panels). The gateway broadcasts acquisition commands wirelessly according to a predetermined schedule. These commands include the electrode serial number (UID) and measurement mode to be activated. Only the designated electrodes are awakened and arranged in the specified measurement array (e.g., a Wenner array), quickly returning to sleep mode after data acquisition. When the gateway receives a heavy rainfall warning or earthquake vibration signal, it can immediately broadcast an emergency command, triggering all electrodes or electrodes in high-risk areas to perform high-frequency encrypted data acquisition.

[0021] It should be noted that the two identification identifiers mentioned in this embodiment are different. The factory-issued globally unique ID (UID) is the device's "ID number." It is physical, fixed, and unchangeable, and is only bound to the device itself. The software-pre-assigned UID is the device's "employee number" or "seat number." It is logical, configurable, and bound to specific responsibilities and locations. For example, an engineer designs a monitoring plan in BIM / GIS software and creates a logical UID, such as B01-05m, at specific coordinates in virtual space (e.g., 5 meters deep in borehole B01). At this time, B01-05m is an "empty space," without a corresponding physical device. Construction workers, according to the design drawings, bury a smart electrode at the actual physical location (5 meters deep in borehole B01). The electrode's hardware has a factory-issued UID, such as SN: ELE-2024-8888. Construction workers use a handheld smart terminal (PDA) to approach the electrode. The terminal reads its physical UID (ELE-2024-8888) via NFC or Bluetooth, and simultaneously uses the terminal's GPS or laser rangefinder to determine its actual installation location (confirmed to be 5m deep in hole B01). The operator clicks "confirm" on the terminal, and the software automatically performs a "registration" operation: binding the physical UID (ELE-2024-8888) with the logical UID (UIDB01-05m), and uploading this binding relationship to the cloud database. The system gateway issues the command: "Please measure the electrode at location B01-05m." The communication module, by searching the binding relationship database, knows that the physical device corresponding to B01-05m is ELE-2024-8888. Therefore, the wireless signal calls: "Device ELE-2024-8888, please report data." Device ELE-2024-8888 responds and uploads the data, which is automatically tagged with the location of B01-05m for storage and processing.

[0022] Based on the optimal 3D electrode deployment scheme, each electrode with a unique identifier is used as a slave node. Relay nodes are selected from all slave nodes according to a preset ratio to complete the construction of a high-density electrode array. Specifically, among all nodes, nodes located higher up or with strong signals are selected at a certain ratio (e.g., 10%) and given relay functions. Besides completing their own data acquisition tasks, these nodes are also responsible for forwarding data from neighboring nodes, forming an ad-hoc mesh network. This allows signals to be relayed, greatly expanding network coverage and improving reliability (a single link failure does not affect the overall network).

[0023] S2: Based on a high-density electrode array, collect apparent resistivity data of the target soil and rock monitoring area.

[0024] S3: Using a parallelized resistivity 3D inversion algorithm, apparent resistivity data is inverted into a true resistivity 3D model of the soil and rock mass, and the inversion results are stored in time sequence to construct a four-dimensional resistivity database. A further implementation method includes the following inversion method for the true resistivity 3D model: S31: The collected apparent resistivity data is compared with historical normal data, and data that does not meet quality requirements is processed to obtain a complete apparent resistivity dataset. Specifically, the core innovation of the entire inversion process lies in the introduction of a closed-loop optimization concept of "historical prior" and "spatiotemporal constraints," upgrading it from an independent mathematical calculation task into a continuous modeling system with memory and evolutionary capabilities. Specifically, in the data preprocessing stage, the innovation lies in not simply comparing with fixed "historical normal data," but intelligently comparing with theoretical predictions obtained from forward modeling based on the previous time-series inversion model. The system integrates the grounding resistance, noise level, and other status data reported by each electrode node through self-diagnosis to construct a dynamic quality assessment model. This model can intelligently distinguish which data deviations are caused by abnormal factors such as electrode faults and instantaneous electromagnetic interference, and which are genuine responses to changes in underground physical properties, thereby achieving accurate identification, labeling, and adaptive repair or removal of abnormal data, ensuring that the dataset input to the inversion algorithm is of high quality and reliability.

[0025] S32: The 3D model mesh of the target soil and rock monitoring area is divided into multiple sub-regions and distributed to different computing nodes for parallel computation to obtain the predicted apparent resistivity and sensitivity matrices. Specifically, within the parallel computing framework, this involves not only the use of a standard domain decomposition parallel strategy but also the intelligent provision of an optimal initial model for each inversion task. The system does not start from scratch but automatically retrieves the resistivity model obtained from the previous successful inversion from the 4D database, using it as the initial state for this inversion. This is equivalent to allowing the iterative solution process to start from a point very close to the true solution, greatly reducing the number of iterations required and improving computational efficiency. Simultaneously, this historical model is also transformed into spatial prior knowledge, guiding the solution process to make more reasonable inferences in unknown regions.

[0026] S33: Construct a spatiotemporal constraint objective function that includes data fitting difference, spatial constraints, and temporal constraints. Specifically, in addition to requiring the inversion results to fit the current observation data and be spatially smooth, this embodiment uniquely adds a temporal constraint term. This term mandates that there cannot be unnecessary or drastic abrupt changes between the inversion results at the current moment and the model at the previous moment, unless the observation data strongly supports such a change. This constraint cleverly transforms the continuous physical laws of time series into mathematical constraints, making the series of 3D models obtained from the inversion no longer isolated "snapshots," but a coherent, smooth, high signal-to-noise ratio "underground film," which can effectively suppress random noise interference and highlight the evolution of real physicochemical processes, such as the advancement of seepage fronts or the gradual formation of potential slip surfaces.

[0027] S34: The Gauss-Newton method is used for iterative solving to minimize the spatiotemporal constraint objective function, updating the resistivity in the 3D model until the preset convergence condition is met, thus obtaining the true resistivity 3D model. Specifically, in the final iterative solution using algorithms such as the Gauss-Newton method, the entire process forms a self-optimizing closed loop. After quality verification, the results of this inversion will be immediately stored in the 4D database, providing more accurate historical priors for the next inversion. Through continuous learning, the system can even adaptively adjust key configurations such as regularization parameters, making the inversion process increasingly efficient and robust over time. This completely changes the traditional inversion model that relies on manual intervention and processing of single sets of data, achieving fully automatic and sustainable temporal spatial insight, providing an incredibly accurate and clear data foundation for subsequent AI intelligent early warning.

[0028] S4: The resistivity four-dimensional database is spatiotemporally aligned and fused with other monitoring data from the target soil and rock monitoring area to obtain an analysis dataset; the other monitoring data include rainfall, groundwater level, and surface displacement. A further implementation method for obtaining the analysis dataset includes: S41: Treating other monitoring data as spatially discrete point data and the true resistivity 3D model as spatially continuous volume data; Step S41 involves redefining and abstracting the essence of the data. Instead of simply treating rainfall, water level, and other data as isolated values, it explicitly defines them as "point data" with clear spatial representativeness. For example, a rain gauge reading represents the rainfall intensity in a certain surrounding area, and the data from a water level well reflects the average water level of the aquifer where its filter pipe is located. Simultaneously, the resistivity model is elevated from an "image" to "spatially continuous volume data," that is, a point cloud composed of tens of thousands of voxels with precise 3D coordinates and property values. This abstraction lays the theoretical foundation for subsequent deep fusion within a unified spatial framework.

[0029] S42: Establish a unified coordinate system for the target soil and rock monitoring area. For each data point, based on its corresponding physical meaning and monitoring principles, find the spatial influence domain on the volumetric data to obtain a spatially aligned dataset. Specifically, introduce the concept of "physically guided spatial influence domain," which is far more complex than simple spatial nearest neighbor matching. The system dynamically and intelligently defines the influence range of different data points within the resistivity volumetric data according to their physical meaning. For example, for a water level well, its "influence domain" is all voxels within a certain range above and below the depth of its filter pipe, simulating the hydraulic connection between groundwater and the surrounding soil and rock. For a surface displacement monitoring point, its "influence domain" may be a set of voxels in an inverted cone-shaped region below its projection and above the potential slip surface, reflecting the control mechanism of deep deformation on surface displacement. This physical principle-based association method ensures the scientific nature and accuracy of data fusion.

[0030] S43: Based on network time protocols, the timestamps of all data in the spatially aligned dataset are uniformly verified, synchronized, and standardized to obtain a spatiotemporally aligned dataset. This approach abandons the previous method of requiring each field sensor to independently perform high-power clock synchronization. Instead, a more efficient and reliable method is adopted: all devices only need to record their own timestamps and upload them to the cloud platform. The platform's powerful time service performs unified reception verification, synchronization, and standardization of all data based on a high-precision time source (such as BeiDou / GPS). This process automatically identifies and corrects time deviations caused by device clock drift and resamples all heterogeneous data onto a unified, equally spaced time series. This constructs a strictly aligned, phase-difference-free data foundation in the time dimension, making it possible to capture temporal causal relationships across data.

[0031] S44: Each resistivity voxel is used as a basic unit, and a multidimensional feature vector is assigned. The elements of this multidimensional feature vector include the voxel's own resistivity value, resistivity change rate, and other monitoring data and derived features spatially aligned to the corresponding voxel. Specifically, the vector not only includes its own static and dynamic electrical parameters (resistivity value, rate of change), but more importantly, through the "spatial influence domain" defined in step S42, external driving and response variables (such as cumulative rainfall from associated rain gauges within the influence domain, water level change rate from water level orifices, and displacement increments from displacement monitoring points) are "injected" as new feature dimensions into the voxel. This makes the feature vector of each voxel a "convergence point" of information, encompassing both its internal physical state and the effects of the external environment, greatly enriching the information dimensions that the AI ​​model can learn.

[0032] S45: Combine the feature vectors of all voxels at all time steps to construct a spatiotemporal data cube; Step S45 constructs a "spatiotemporal data cube" data structure. This is not a visual 3D cube, but a giant, multidimensional feature matrix. Its row indices are "time" and "space" (voxel ID), and the column indices are "feature dimensions" (electrical, hydraulic, mechanical, etc.). The feature vectors of all voxels at all time steps are systematically organized in this cube, thus forming a "digital twin" that completely records the spatiotemporal evolution of the entire monitoring area, providing a unique source of truth for subsequent data mining based on big data and deep learning.

[0033] S46: Based on spatiotemporal data cubes, a structured analytical dataset indexed by spatiotemporal units is constructed. This dataset is directly addressed to meet the needs of machine learning models and has undergone a structured reconstruction. It can be easily converted into tensors in formats such as (sample, feature, time step), and directly input into advanced time-series deep learning models for end-to-end training, learning complex evolutionary patterns of subsurface systems, thus achieving a seamless transformation from raw data to usable fuel for AI.

[0034] S5: Train a neural network model based on the analyzed dataset to build a soil and rock stability monitoring model.

[0035] A further implementation method involves constructing a soil and rock stability monitoring model, including: S51: Based on a unified coordinate system, resistivity voxels, rainfall, groundwater level, and ground displacement are defined as different types of nodes. A heterogeneous graph embedding physical knowledge is established based on the physical relationships between the data. In step S51, abstract monitoring data is transformed into a concrete "knowledge graph" containing physical relationships. Different node types are clearly defined based on their physical nature (e.g., resistivity reflects physical properties, groundwater level reflects pore pressure, and displacement reflects mechanical response). More importantly, the connections (edges) between nodes are no longer based on simple spatial distance, but are defined by prior physical mechanisms. For example, a "rainfall" node will establish an "infiltration influence" edge with a "resistivity voxel" node in the surface area; a "deep displacement" node will establish a "deformation-driven" edge with a "resistivity voxel" node on a potential slip zone. This heterogeneous graph construction based on physical semantics injects professional domain knowledge into the model, allowing it to start from a higher cognitive level, avoiding the need to learn these basic physical laws from scratch, and greatly improving learning efficiency and the model's rationality.

[0036] S52: Using a meta-path-aware heterogeneous graph attention network, information about each node and its neighbors is aggregated to obtain node embeddings containing physical semantics. Step S52 employs a meta-path-aware heterogeneous graph attention network for information aggregation. Ordinary graph attention mechanisms cannot distinguish between different physical relationships such as "hydraulic connections" and "mechanical connections." This invention introduces the concept of "meta-path," which predefines a sequence of node types with clear physical meaning (e.g., the path "rainfall node → resistivity voxel node → displacement node" describes the process of rainfall infiltration softening the soil and causing deformation). The algorithm aggregates information along multiple different meta-paths, and then a semantic attention layer automatically learns which physical path is most important in the current system state (e.g., during a rainstorm) and dynamically adjusts its weight. This means that the model can not only aggregate information but also understand the physical paths of information transmission and their importance, achieving "interpretability" of information aggregation.

[0037] S53: The embedded sequences of each node are input into a temporal convolutional network to capture the dynamic evolution patterns of each node. Through a gating mechanism, the node states are iteratively transmitted over time and diffused spatially, obtaining the spatiotemporal evolution characteristics of the analysis dataset. Step S53 uses a temporal convolutional network (TCN) and a gating mechanism to capture and fuse spatiotemporal dynamics. Due to its parallelization capabilities and advantages in capturing long-range dependencies, TCN processes long-term temporal data more efficiently than traditional RNNs. Its core innovation lies in the iterative process of the "gating mechanism," which acts as a "valve" controlling how information propagates in the spatiotemporal dimension. For example, the model can learn that during droughts, information transmission along hydraulically related edges should be suppressed; while after heavy rainfall, the channel should be quickly opened. This makes the transmission of node states over time and the diffusion of them spatially (graph structure) not a fixed mathematical operation, but an intelligent process that adaptively adjusts according to the dynamic state of the system, thereby extracting spatiotemporal evolution characteristics that better reflect the essence of the system.

[0038] S54: Based on spatiotemporal evolution characteristics and combined with physical constraint loss, a rock and soil stability monitoring model is constructed. The final step, S54, introduces "physical constraints" as part of the loss function into the model training, constructing a stability monitoring model driven by physical mechanisms. This not only embeds knowledge into the model structure (graph network) but also imposes physical constraints on the optimization objective. For example, a constraint term is added to the loss function to penalize model predictions that violate common sense in rock and soil mechanics, such as "significantly rising water level but sharply increasing resistivity (meaning no softening)" or "drastic increase in displacement but no corresponding weak zone found." This physical constraint loss, together with the data fitting loss, guides model training, ensuring that the model output is not only statistically optimal but also physically reliable. This significantly enhances the model's generalization ability and prediction reliability under extreme conditions, ultimately forming a robust and interpretable rock and soil stability monitoring model that is both data-driven and knowledge-guided.

[0039] S6: Based on a soil and rock stability monitoring model, real-time monitoring of the soil and rock mass to be monitored is performed to obtain monitoring results. Through spatiotemporal data cubes and interpretable AI technology, key hazardous areas can be located, the development process of anomalies can be visualized (such as how seepage paths expand), and the importance of various influencing factors can be quantified (e.g., rainfall contributed 70% and engineering disturbance contributed 30% in this warning). This provides engineers with unprecedented scientific basis and in-depth insights for taking precise measures (such as where to drain water and where to reinforce).

[0040] Example 2 The present invention also provides a rock and soil stability monitoring system based on high-density electrical resistivity tomography (EDT) for implementing the method of Embodiment 1, comprising: The electrode array construction module is used to deploy electrodes with integrated micro-acquisition circuits and identification tags in the target soil and rock monitoring area, and connect each electrode to the Internet of Things to build a high-density electrode array. The data acquisition module is used to acquire apparent resistivity data of the target soil and rock monitoring area based on a high-density electrode array. The inversion module is used to invert apparent resistivity data into a true resistivity three-dimensional model of the soil and rock body using a parallelized resistivity three-dimensional inversion algorithm, and to store the inversion results in time sequence to build a four-dimensional resistivity database. The dataset construction module is used to perform spatiotemporal alignment and fusion of the resistivity four-dimensional database with other monitoring data in the target soil and rock monitoring area to obtain the analysis dataset; among which, other monitoring data include rainfall, groundwater level and surface displacement; The model building module is used to train a neural network model based on the analysis dataset and build a rock and soil stability monitoring model. The stability monitoring module is used to monitor the soil and rock mass in real time based on the soil and rock mass stability monitoring model and obtain the monitoring results.

[0041] A further embodiment of the embodiment is that the electrode array construction module includes: The identification unit is used to integrate the miniature acquisition circuitry into the sealed chamber at the tail of the electrode and to assign a unique identification to each electrode. The deployment scheme design unit is used to establish the optimal three-dimensional electrode deployment scheme based on the survey results of the target soil and rock monitoring area; The electrode deployment unit is used to construct a high-density electrode array by using each electrode with a unique identifier as a slave node based on the optimal three-dimensional electrode deployment scheme, and selecting relay nodes from all slave nodes according to a preset ratio.

[0042] A further implementation method is that the inversion module includes: The preprocessing unit is used to compare the collected apparent resistivity data with historical normal data and process the apparent resistivity data that does not meet the quality requirements to obtain a complete apparent resistivity dataset. Parallel computing units are used to divide the three-dimensional model mesh of the target soil and rock monitoring area into multiple sub-regions and assign them to different computing nodes for parallel computing to obtain the predicted apparent resistivity and sensitivity matrix. The objective function construction unit is used to construct a spatiotemporal constraint objective function that includes data fit difference, spatial constraints, and time constraints. The iterative solution unit is used to perform iterative solutions using the Gauss-Newton method, minimize the spatiotemporal constraint objective function, update the resistivity in the three-dimensional model, until the preset convergence condition is met, and obtain the true resistivity three-dimensional model.

[0043] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for monitoring the stability of soil and rock masses based on high-density electrical resistivity tomography (EDT), characterized in that, include: Electrodes with integrated micro-acquisition circuits and identification tags are deployed in the target soil and rock monitoring area, and each electrode is connected by the Internet of Things to build a high-density electrode array. Based on the high-density electrode array, the apparent resistivity data of the target soil and rock monitoring area are collected; Using a parallelized three-dimensional resistivity inversion algorithm, the apparent resistivity data is inverted into a true three-dimensional resistivity model of the soil and rock body, and the inversion results are stored in time sequence to construct a four-dimensional resistivity database. The resistivity four-dimensional database is spatiotemporally aligned and fused with other monitoring data from the target soil and rock monitoring area to obtain an analysis dataset; wherein, the other monitoring data includes rainfall, groundwater level and surface displacement. A neural network model was trained based on the analyzed dataset to construct a rock and soil stability monitoring model. Based on the aforementioned rock and soil stability monitoring model, the rock and soil mass to be monitored is monitored in real time to obtain monitoring results.

2. The method according to claim 1, characterized in that, Methods for constructing high-density electrode arrays include: The miniature acquisition circuit is integrated into the sealed chamber at the tail of the electrode, and each electrode is given a unique identification. Based on the survey results of the target soil and rock monitoring area, an optimal three-dimensional electrode layout scheme is established; Based on the optimal three-dimensional electrode layout scheme, each electrode with a unique identifier is used as a slave node, and relay nodes are selected from all slave nodes according to a preset ratio to complete the construction of a high-density electrode array.

3. The method according to claim 1, characterized in that, The inversion method for the true resistivity three-dimensional model includes: The collected apparent resistivity data is compared with historical normal data, and the apparent resistivity data that does not meet the quality requirements is processed to obtain a complete apparent resistivity dataset. The three-dimensional model mesh of the target soil and rock monitoring area is divided into multiple sub-regions and assigned to different computing nodes for parallel computing to obtain the predicted apparent resistivity and sensitivity matrix. Construct a spatiotemporal constraint objective function that includes data fit difference, spatial constraints, and time constraints; The Gauss-Newton method is used for iterative solution to minimize the spatiotemporal constraint objective function, update the resistivity in the three-dimensional model, until the preset convergence condition is met, and obtain the true resistivity three-dimensional model.

4. The method according to claim 1, characterized in that, Methods for obtaining the analysis dataset include: The other monitoring data are treated as spatially discrete point data, while the true resistivity three-dimensional model is treated as spatially continuous volume data. Establish a unified coordinate system for the target soil and rock mass monitoring area. For each data point, based on its corresponding physical meaning and monitoring principle, find the spatial influence domain on the mass data to obtain a spatially aligned dataset. Based on the network time protocol, the timestamps of all data in the spatially aligned dataset are uniformly verified, synchronized and standardized to obtain the spatiotemporally aligned dataset; Each resistivity voxel is used as a basic unit, and a multidimensional feature vector is assigned; wherein, the elements of the multidimensional feature vector include the voxel's own resistivity value, resistivity change rate, and other monitoring data and derived features associated with the corresponding voxel through spatial alignment. Combine the feature vectors of all voxels at all time steps to construct a spatiotemporal data cube; Based on the spatiotemporal data cube, a structured analysis dataset indexed by spatiotemporal units is constructed.

5. The method according to claim 1, characterized in that, The method for constructing the soil and rock stability monitoring model includes: Based on a unified coordinate system, resistivity voxels, rainfall, groundwater level and ground displacement are defined as different types of nodes, and a heterogeneous graph with embedded physical knowledge is established according to the physical relationship between the data. Using a meta-path-aware heterogeneous graph attention network, information about each node and its neighboring nodes is aggregated to obtain node embeddings containing physical semantics. The embedding sequences of each node are input into a temporal convolutional network to capture the dynamic evolution patterns of each node. Through a gating mechanism, the node state is transferred in time and diffused in space to obtain the spatiotemporal evolution characteristics of the analysis dataset. Based on the aforementioned spatiotemporal evolution characteristics and combined with physical constraint losses, the stability monitoring model for the soil and rock mass is constructed.

6. A rock and soil stability monitoring system based on high-density electrical resistivity tomography (EDT), used to implement the method described in any one of claims 1-5, characterized in that, include: The electrode array construction module is used to deploy electrodes with integrated micro-acquisition circuits and identification tags in the target soil and rock monitoring area, and connect each electrode to the Internet of Things to build a high-density electrode array. The data acquisition module is used to acquire apparent resistivity data of the target soil and rock monitoring area based on the high-density electrode array. The inversion module is used to invert the apparent resistivity data into a true resistivity three-dimensional model of the soil and rock body using a parallelized resistivity three-dimensional inversion algorithm, and to store the inversion results in time sequence to build a resistivity four-dimensional database. The dataset construction module is used to perform spatiotemporal alignment and fusion of the resistivity four-dimensional database with other monitoring data in the target soil and rock monitoring area to obtain an analysis dataset; wherein, the other monitoring data includes rainfall, groundwater level and surface displacement; The model building module is used to train a neural network model based on the analysis dataset and build a rock and soil stability monitoring model. The stability monitoring module is used to perform real-time monitoring of the soil and rock mass to be monitored based on the soil and rock mass stability monitoring model, and to obtain monitoring results.

7. The system according to claim 6, characterized in that, The electrode array construction module includes: The identification unit is used to integrate the miniature acquisition circuitry into the sealed chamber at the tail of the electrode and to assign a unique identification to each electrode. The deployment scheme design unit is used to establish the optimal three-dimensional electrode deployment scheme based on the survey results of the target soil and rock monitoring area; The electrode deployment unit is used to construct a high-density electrode array by using each electrode with a unique identifier as a slave node based on the optimal three-dimensional electrode deployment scheme, and selecting relay nodes from all slave nodes according to a preset ratio.

8. The system according to claim 6, characterized in that, The inversion module includes: The preprocessing unit is used to compare the collected apparent resistivity data with historical normal data and process the apparent resistivity data that does not meet the quality requirements to obtain a complete apparent resistivity dataset. Parallel computing units are used to divide the three-dimensional model mesh of the target soil and rock monitoring area into multiple sub-regions and assign them to different computing nodes for parallel computing to obtain the predicted apparent resistivity and sensitivity matrix. The objective function construction unit is used to construct a spatiotemporal constraint objective function that includes data fit difference, spatial constraints, and time constraints. The iterative solution unit is used to perform iterative solution using the Gauss-Newton method, minimize the spatiotemporal constraint objective function, update the resistivity in the three-dimensional model, until the preset convergence condition is met, and obtain the true resistivity three-dimensional model.

Images