Real-time monitoring and early warning method for slope deformation based on interferometric radar images

Abstract

This invention relates to the field of geological disaster monitoring and distance sensor technology, specifically disclosing a real-time monitoring and early warning method for slope deformation based on interferometric radar images. The method includes: acquiring multi-time-series interferometric radar images and extracting high-precision surface deformation field sequences; constructing a physical information neural network model integrating the constitutive relations of rock and soil mechanics; using the deformation field as input to invert the internal stress, pore water pressure, and potential sliding surface of the slope in real time; determining the instability risk level based on preset thresholds and triggering a three-level early warning response mechanism to generate a visualized early warning report. This invention achieves highly timely and interpretable early warnings at the minute or even second level by deeply coupling millimeter-level deformation observation with physical constraints, thus improving the scientific rigor and reliability of slope disaster monitoring.

Description

Technical Field

[0001] This invention belongs to the field of geological disaster monitoring and distance sensor technology, specifically relating to a method for real-time monitoring and early warning of slope deformation based on interferometric radar images. Background Technology

[0002] With the continuous expansion of geological engineering projects and the increasing complexity of natural environmental conditions, the prevention and control of slope geological hazards has become a key area for ensuring the safe operation of transportation, mining, and water conservancy infrastructure. Interferometric synthetic aperture radar (IRSAR), as an advanced remote sensing measurement method, plays a central role in slope displacement monitoring due to its advantages of high precision, large area coverage, and all-weather Earth observation. Through continuous acquisition and phase analysis of multi-time-series radar images, the spatiotemporal evolution characteristics of minute surface displacements can be captured, providing crucial data support for the early identification and risk assessment of geological hazards.

[0003] Real-time monitoring and early warning of slope deformation based on interferometric radar images is a key technology for proactive prevention and control of geological disasters. This technology aims to quantitatively assess the evolution trend of slopes using high-frequency deformation field observation data and promptly release risk information in conjunction with early warning logic. The monitoring and early warning process requires accurate extraction of surface deformation sequences and, more importantly, establishing the intrinsic connection between deformation characteristics and deep mechanical mechanisms to achieve rapid perception and scientific assessment of slope instability risks in complex geological environments.

[0004] Existing monitoring methods mostly focus on the objective description of surface deformation values, making it difficult to effectively invert the stress distribution and mechanical evolution state inside the slope. Although traditional numerical simulation methods can perform mechanical mechanism analysis, they are limited by huge computational costs, and the processing time is too long, making it difficult to meet the timeliness requirements of real-time early warning for sudden landslide monitoring.

[0005] A severe coupling gap exists between monitoring data and physical equations, making it difficult for early warning models to scientifically calibrate the critical thresholds of physical indicators that trigger instability. Traditional early warning models, lacking deep integration of physical constraints, often suffer from feedback lag or insufficient accuracy when facing nonlinear dynamic processes such as accelerated slope instability. Therefore, a real-time monitoring and early warning method for slope deformation based on interferometric radar images is desired. Summary of the Invention

[0006] The purpose of this invention is to provide a real-time monitoring and early warning method for slope deformation based on interferometric radar images, which can effectively solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is: a real-time monitoring and early warning method for slope deformation based on interferometric radar images, comprising the following specific steps: Step 1: Acquire multi-temporal interferometric radar image data and extract high-precision surface deformation field sequences from them; Step 2: Construct a physical information neural network model that integrates the constitutive relations of slope rock and soil mechanics; Step 3: Use the surface deformation field sequence as input to drive the physical information neural network model to perform real-time inversion of the internal mechanical state of the slope; Step 4: Based on the mechanical state parameters obtained from the inversion, and combined with preset thresholds, determine the slope instability risk level; Step 5: Trigger the corresponding early warning response mechanism based on the risk level and generate a visual early warning report.

[0008] Preferably, the process of acquiring multi-time-series interferometric radar image data in step 1 includes imaging processing, phase unwrapping and atmospheric error correction of the raw echo signal of synthetic aperture radar, and finally generating surface deformation time series data with millimeter-level accuracy. The surface deformation time series data is organized in the form of a spatial grid, and each grid cell corresponds to a geographical location and its cumulative deformation at multiple observation times.

[0009] Preferably, the physical information neural network model constructed in step 2 includes an input layer, a hidden layer, and an output layer. Its loss function is composed of a data fitting term and a physical constraint term. The data fitting term is used to measure the difference between the model output and the measured deformation data. The physical constraint term embeds the basic governing equations describing the stress-strain behavior of slope soil and rock, including but not limited to the Mohr-Coulomb strength criterion and the unsaturated soil seepage continuity equation.

[0010] Preferably, the physical information neural network model adopts an adaptive weight adjustment strategy during training, so that the contribution ratio of physical constraints to the total loss function can change dynamically under different geological conditions or rainfall events, ensuring the model's generalization ability and physical consistency under complex working conditions.

[0011] Preferably, the specific process of real-time inversion of mechanical state in step 3 is as follows: the current time and the sequence of surface deformation field at several historical times are input into the trained physical information neural network model. The model synchronously outputs the stress distribution, pore water pressure evolution trend and potential sliding surface location information of key areas inside the slope through internal mapping relationship.

[0012] Preferably, the mechanical state parameters used to determine the slope instability risk level in step 4 include pore water pressure ratio, safety factor, and shear strain increment rate. When any parameter exceeds its corresponding preset threshold, the system determines that the slope has entered the corresponding risk state. The preset threshold is determined based on regional geological characteristics and statistical analysis of historical landslide cases.

[0013] Preferably, the early warning response mechanism triggered in step 5 is divided into three levels: attention level, alert level and danger level. Each level corresponds to different information release methods and emergency response suggestions. The attention level only records and alerts abnormal trends, the alert level pushes early warning notices to relevant management departments, and the danger level automatically links with the emergency command platform to activate the evacuation plan.

[0014] Preferably, the visualized early warning report is based on a geographic information system platform, integrating surface deformation heat maps, mechanical state profile maps, and risk level zoning maps, and supports multi-scale browsing and timeline backtracking functions, making it easy for technicians to quickly grasp the overall stability evolution process of the slope.

[0015] Preferably, the physical information neural network model is deployed on an edge computing node. Through lightweight structural design and model compression technology, it meets the millisecond-level response requirements while ensuring inversion accuracy, and is suitable for low-power operation in unattended monitoring scenarios in the field.

[0016] Preferably, the present invention also includes a mechanism for online updating of the physical information neural network model. When new landslide event samples or supplementary geological survey data are received, the system can automatically trigger the model fine-tuning process and use incremental learning algorithms to improve the model's adaptability to new working conditions without retraining all parameters.

[0017] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention deeply integrates high-precision surface deformation observation data provided by interferometric radar images with the physical constitutive relationship of slope rock and soil, and constructs a deep learning model with physical interpretability. It breaks through the bottleneck of low computational efficiency of traditional numerical simulation, and can complete the inversion of the internal mechanical state of the slope within minutes or even seconds, thereby improving the timeliness and scientificity of geological disaster early warning.

[0018] 2. Because the model of this invention incorporates strict physical equation constraints, its output results not only conform to the laws of engineering mechanics, but also provide a clear physical basis for setting the warning threshold, thus avoiding the uncertainty brought about by empirical criteria.

[0019] 3. This invention supports multi-source data fusion and online model evolution, possesses good robustness and scalability, and is suitable for constructing intelligent monitoring systems for slope safety in various geological environments. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the overall technical solution according to the present invention; Figure 2 This is a schematic diagram of data flow according to the present invention; Figure 3A flowchart illustrating the construction of a physical information neural network model that includes data fitting terms and physical constraint terms and has an adaptive weight adjustment strategy according to the present invention; Figure 4 This is a flowchart illustrating the logical process of inverting the internal stress distribution and potential sliding surface of a slope using a multi-temporal surface deformation field sequence driven model according to the present invention. Figure 5 This is a flowchart illustrating the generation of multi-level early warning response and visualization report based on the threshold determination of mechanical state parameters according to the present invention. Detailed Implementation

[0021] Example 1: Please refer to the appendix Figure 1 To be continued Figure 5 To make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments.

[0022] In the real-time monitoring and early warning method for slope deformation based on interferometric radar images, step 1 specifically involves acquiring multi-temporal interferometric radar image data and extracting a high-precision surface deformation field sequence from it. This process begins with high-precision imaging processing of the raw echo signal captured by the synthetic aperture radar sensor.

[0023] In this embodiment, a synthetic aperture radar sensor is configured on a satellite orbit or a fixed ground platform to scan the target slope area at a specific revisit period. The acquired raw echo signals are first converted into single-scatter complex image data using a range-Doppler algorithm or a line-frequency scaling algorithm. Each single-scatter complex image contains amplitude and phase information, with the phase information being the core foundation for extracting deformation data.

[0024] Step 1, which extracts the high-precision surface deformation field sequence, involves co-registering multiple single-scattered complex images. One image is selected as the master image, and the remaining images are used as slave images. The offsets of the slave images relative to the master image in the range and azimuth directions are calculated using cross-correlation functions, achieving sub-pixel-level spatial overlap. The registered image pairs are then subjected to differential interferometry to generate an interferometric phase map.

[0025] The interferometric phase map contains multiple phase components, namely, flat terrain phase, topographic phase, atmospheric delay phase, orbital residual phase, noise phase, and target deformation phase. To extract the pure deformation phase, the system introduces external digital elevation model data, simulates a reference terrain phase, and removes it from the interferometric phase.

[0026] In the phase unwrapping stage, step 1 employs the branching method, minimum cost flow algorithm, or least squares method to restore the phase values ​​wrapped between negative and positive pi to continuous phase values, eliminating the multivalued nature of the phase. To address atmospheric delay errors, the system utilizes the statistical characteristics of stable point targets or introduces synchronous meteorological observation data to establish an atmospheric correction model, compensating for ionospheric and tropospheric delays.

[0027] The resulting deformation field sequence is organized in a spatial grid, with each grid cell assigned a unique geographic coordinate identifier and recording the cumulative deformation at that location relative to a reference time across multiple observation times, achieving millimeter-level deformation detection accuracy. This sequence data not only reflects the geometric displacement of the land surface but also reveals key characteristics of the internal mechanical evolution of the slope.

[0028] In step 2, constructing a physical information neural network model that integrates the constitutive relations of slope rock and soil mechanics is the core step in realizing deep mechanics inversion. This physical information neural network model includes an input layer, multiple hidden layers, and an output layer. The input layer receives coordinate parameters representing spatial location and time parameters representing observation time. The hidden layers consist of multiple neurons connected by adjustable weights and bias parameters, and use a hyperbolic tangent function or a modified linear unit as the activation function to capture the highly nonlinear characteristics of the slope deformation process.

[0029] The key to building the model in step 2 lies in its unique loss function design. This loss function consists of a data fitting term and a physical constraint term, and the network parameters are trained by minimizing the total loss function. Specifically, the data fitting term is used to calculate the sum of squares of deviations between the predicted surface displacement output by the neural network and the measured deformation data obtained in step 1. The physical constraint term transforms the basic governing equations describing the stress-strain behavior of the slope soil and rock mass into residual terms. These governing equations include, but are not limited to, equilibrium equations, geometric equations, and constitutive equations.

[0030] In the equilibrium equations, it is required that, neglecting body forces, the sum of the partial derivatives of the stress tensor in all directions remains zero at every moment. In the geometric equations, strain components are defined as combinations of the partial derivatives of displacement components with respect to spatial coordinates. In the constitutive equations, the Mohr-Coulomb strength criterion and the continuity equation for seepage in unsaturated soil are introduced.

[0031] Specifically, regarding the implementation of physical constraints, the system utilizes automatic differentiation techniques to calculate the derivatives of the neural network output with respect to the input coordinates and time, without requiring explicit differentiation formulas. For example, for the Mohr-Coulomb criterion, the physical constraints require that the relationship between shear strain and normal stress must conform to the material's internal friction angle and cohesion properties. When the internal state predicted by the model violates these physical laws, the physical constraints will generate large residual values, thereby guiding the optimization algorithm to adjust the weights.

[0032] The introduction of the continuity equation for seepage in unsaturated soil allows the model to account for changes in pore water pressure caused by rainfall infiltration and its impact on effective stress. The proportion of physical constraint terms is dynamically managed through an adaptive weight adjustment strategy. In the early stages of training, the system assigns high weights to data fitting terms to quickly fit the observed trends. In the later stages of training, the weights of physical constraint terms are gradually increased to ensure that the inversion results not only mathematically fit the surface data but also conform to the laws of geomechanics in terms of physical logic.

[0033] The specific process of real-time inversion of the mechanical state in step 3 involves inputting the current and several historical time-series surface deformation field sequences into a trained physical information neural network model. The model, through internal multi-layered nonlinear mapping relationships, synchronously outputs the stress distribution, pore water pressure evolution trend, and potential sliding surface location information of key areas within the slope. Unlike numerical simulation methods that rely on mesh generation and matrix iteration, the physical information neural network in this embodiment performs direct function mapping calculations during the inference phase. This means that once the model parameters are determined and surface observations are input, the mechanical state variables at any point underground can be calculated in a very short time.

[0034] During the inversion process, the model first identifies regions with large displacement gradients based on the input deformation sequence and then infers the stress concentration within these regions by incorporating physical constraints. The inversion of pore water pressure relies on the rate of displacement change over time; when an abnormal acceleration in the deformation rate is observed, the model, combined with seepage physical constraints, determines the degree of increase in internal pore water pressure. The location of the potential sliding surface is determined by searching for the continuous trajectory line with the largest shear strain increment, which typically corresponds to the geometric boundary with the lowest safety factor. The inversion results are output in the form of a high-dimensional tensor, encompassing the three-dimensional stress tensor field, fluid pressure field, and cumulative damage field within the slope.

[0035] In step 4, the slope instability risk level is determined based on the inverted mechanical state parameters and a preset threshold. This step represents a shift from simple displacement monitoring to quantitative evaluation of mechanical stability. Key mechanical state parameters extracted by the system include pore water pressure ratio, safety factor, and shear strain increment rate. The safety factor is defined as the ratio of the slope's anti-sliding force to its sliding force; this mechanical state parameter is calculated based on the inverted stress distribution and material strength parameters.

[0036] The judgment logic is as follows: The system compares the current safety factor obtained from the inversion with the preset stability threshold. If the safety factor is greater than 1.5, it is judged to be in a stable state; if the safety factor is between 1.2 and 1.5, a comprehensive judgment is made in conjunction with the pore water pressure ratio. When the pore water pressure ratio exceeds the preset critical ratio value, and the shear strain increment rate shows a continuous upward trend, the system determines that the slope has entered the corresponding level of risk state. The preset threshold is not fixed, but is finely calibrated based on the specific geological characteristics of the area and the statistical analysis results of historical landslide cases in the region.

[0037] In step 5, the corresponding early warning response mechanism is triggered based on the risk level, and a visual early warning report is generated. The early warning response mechanism is divided into three levels: Attention Level, Alert Level, and Danger Level. When the system determines it to be at the Attention Level, the response mechanism automatically records the current abnormal trend, adds the slope to the key monitoring list, and increases the frequency of acquisition and processing of interferometric radar data. When it is determined to be at the Alert Level, the system automatically generates an early warning notification and sends alarm information to relevant management departments and safety production personnel via SMS, email, and industry-specific networks, suggesting that on-site inspections be strengthened. When it is determined to be at the Danger Level, the system automatically links with the emergency command platform, activates the personnel evacuation plan, triggers the ground-based audible and visual alarm facilities in the affected area, and simultaneously sends a road closure request to the traffic management department.

[0038] The visualized early warning report is generated based on a Geographic Information System (GIS) platform. The report includes a surface deformation heat map, a mechanical state profile, and a risk level zoning map. The surface deformation heat map visually displays the cumulative displacement at different locations on the slope surface, with color depth representing the deformation level. The mechanical state profile shows the distribution of internal stress and pore water pressure perpendicular to the slope's strike, clearly indicating the depth and orientation of potential sliding surfaces. The risk level zoning map divides the slope into high-risk, medium-risk, and low-risk zones based on calculation results. The report supports multi-scale browsing; users can switch from a macroscopic regional perspective to a microscopic grid cell perspective and use the timeline backtracking function to view the dynamic process of stability evolution over a past period.

[0039] This embodiment also includes a mechanism for online updating of the physical information neural network model. Considering the dynamic changes in the geological environment and the continuous accumulation of monitoring data, the system automatically triggers a model fine-tuning process upon receiving new landslide event samples or supplementary geological survey data. This model fine-tuning process employs an incremental learning algorithm, using new data to locally update the hidden layer weights of the existing model.

[0040] The system maintains the structure of the physical constraints unchanged, adjusting the weights only through gradient descent, enabling the model to quickly adapt to new operating conditions, such as the weakening of soil parameters caused by long-duration rainfall. This online update mechanism avoids the massive computational cost of retraining the model from scratch, ensuring the accuracy and robustness of the monitoring and early warning system in long-term operation.

[0041] To meet the requirements of unattended monitoring scenarios in the field, the physical information neural network model is deployed on edge computing nodes. These edge computing nodes employ low-power, high-performance embedded processors. Through model pruning, quantization, and lightweight structural design, the number of neural network parameters and computational complexity are reduced. Model quantization converts floating-point parameters to integer parameters, improving inference speed and reducing memory usage with minimal loss of inversion accuracy. The edge computing nodes directly receive preprocessed data from the interferometric radar, perform mechanical inversion and risk assessment locally, and send the assessment results and simplified visualization reports to a remote server. This architecture ensures millisecond-level response requirements while reducing bandwidth pressure on wireless communication links.

[0042] Example 2: Based on Example 1, this example further details the specific mathematical logic textual implementation of the physical constraint terms in the physical information neural network model and the deep strategy of data fusion.

[0043] The core of constructing the physical constraint term in step 2 is establishing a loss assessment logic based on multi-physics coupling. For the equilibrium state of soil and rock, the equilibrium equation components in the loss function require that, at any spatial point, the sum of the partial derivatives of the normal stress component with respect to the corresponding coordinate, plus the sum of the partial derivatives of the shear stress component with respect to the cross coordinate, must approach 0. This logic is implemented through an automatic differentiation framework. In addition to displacement values ​​in three directions, the output layer of the neural network also contains six independent stress tensor components. During training, the system not only requires the displacement output to match the InSAR observation data but also forces the stress tensor to satisfy the equilibrium constraint of spatial derivatives.

[0044] In this embodiment, for the key unsaturated soil characteristics in slope instability, the physical constraint term further integrates the textual logic of the Richards equation. This textual logic is described as follows: the rate of change of soil moisture content equals the negative of the spatial divergence of the seepage velocity. The seepage velocity follows Darcy's law and is expressed as the product of the permeability coefficient and the total head gradient. The total head consists of the position head and the pressure head.

[0045] The model incorporates pore water pressure as an intermediate variable, embedding the physical laws of water transport into a neural network. When the externally monitored deformation time series exhibits nonlinear creep characteristics highly correlated with rainfall intensity, the physical constraints guide the network to interpret this accelerated displacement as an increase in pore water pressure and a decrease in effective stress, thus achieving an accurate inversion of the intrinsic mechanism of rainfall-induced landslides.

[0046] Furthermore, the mechanical state inversion in step 3 incorporates a multi-source heterogeneous data fusion strategy. In addition to the surface deformation field sequence extracted by interferometric radar, the system also integrates real-time displacement data from fixed monitoring stations of the Global Navigation Satellite System (GNSS) deployed on the slope surface. GNSS data offers higher temporal resolution, providing point displacement information at the minute or even second level. A local calibration term for the fixed station data is added to the loss function of the physical information neural network.

[0047] During training or inference, the system complements the area-based low-frequency observations from radar with the point-based high-frequency observations from satellite positioning. Specifically, the real-time displacement of the Global Navigation Satellite System (GNSS) is used as a hard constraint, forcing the neural network's output at that specific coordinate point to maintain a high degree of consistency with the satellite positioning result. Meanwhile, the radar deformation field serves as a soft constraint, providing a spatial continuity background. This fusion strategy ensures the temporal accuracy of the inversion results at key monitoring points and the spatial accuracy across the entire system.

[0048] In step 4, risk level determination, this embodiment introduces energy dissipation rate as an auxiliary indicator. Energy dissipation rate is defined as the rate of conversion of plastic work during deformation of the soil and rock mass. The system calculates the cumulative energy state inside the slope by multiplying the stress tensor increment obtained from the inversion by the strain tensor increment and performing time integration. When the energy dissipation rate undergoes a sudden change accompanied by a rapid drop in the safety factor, the slope is determined to have entered the pre-instability stage from the creep stage.

[0049] This embodiment employs a dynamic evolution strategy for determining the early warning threshold. The system incorporates an initial parameter library based on geological survey reports, including the saturation, cohesion, internal friction angle, and permeability coefficient of the soil and rock mass in the area. As monitoring progresses, the physical information neural network automatically corrects these mechanical parameters by continuously fitting the observed deformation. If the inverted mechanical parameters consistently show a softening trend, the system automatically lowers the early warning threshold for the safety factor. For example, the safety factor threshold for the hazardous level is increased from 1.05 to 1.10 to provide a more sufficient emergency response time window.

[0050] The visualization report has been enhanced in terms of presentation, introducing 3D transparent display technology. Users can adjust the transparency of the slope surface within the GIS interface to directly observe the inverted internal stress cloud map and pore water pressure isosurface. The system uses spatial interpolation technology to convert the discrete-point mechanical variables output by the neural network into continuous 3D volume data.

[0051] Volume rendering technology can clearly identify stress concentration zones and dominant seepage channels. The report also integrates a risk evolution trend prediction module, which uses the physical laws learned by the model to extrapolate the deformation development trend over the next 24 to 72 hours under the assumption of sustained rainfall intensity.

[0052] In this embodiment, the online update mechanism is refined into model drift monitoring and selective fine-tuning. The system calculates the residual distribution between the displacement predicted by the model and the newly observed displacement in real time. If the residuals exhibit a random spatial distribution and a small amplitude, the current model status is maintained; if the residuals show a systematic deviation in a specific region, model drift is determined to have occurred. The system automatically extracts historical and latest data for that region and retrains the local connection weights of the model. During training, a learning rate decay strategy is adopted, with the initial learning rate set to a small value to avoid destroying the learned global features, and the residuals are brought back to a normal level through multiple iterations.

[0053] Example 3: Based on the above examples, this example focuses on describing a specific implementation method for complex rock slopes, especially the processing logic in the presence of obvious joint surfaces and fault fracture zones.

[0054] In step 1, to address the potential vegetation cover issue on rock slopes, the system employs polarimetric interferometric radar technology. By transmitting and receiving electromagnetic waves with different polarization states, the vegetation scattering contribution is separated from the surface scattering contribution based on differences in polarization coherence. This process includes constructing a polarization coherence matrix and estimating the vegetation layer height using a polarimetric interferometric phase decomposition algorithm. The phase component caused by the vegetation layer is then removed to obtain the true deformation information of the underlying rock layer. This is of significant importance for slope monitoring in mountainous and forested areas.

[0055] In step 2, the loss function of the physical information neural network was structurally adjusted to address the discontinuous characteristics of rock slopes. The physical constraint term is no longer solely based on continuum mechanics, but incorporates the virtual displacement principle describing joint slippage. The system pre-defines several potential structural surfaces in the neural network model based on slope surface structural characteristics obtained from ground-penetrating radar or UAV aerial surveys.

[0056] The physical constraints require that the relative displacement of the rock masses on both sides of these structural surfaces must comply with the shear strength criterion. If the shear stress does not reach the shear strength of the structural surface, the displacement on both sides remains continuous; once the shear stress exceeds the upper limit of the strength, discontinuous slip displacement is allowed, and this discontinuity is penalized for mechanical consistency in the loss function.

[0057] To address the instability characteristics of rock slopes, the mechanical state parameters in step 4 incorporate inversion logic based on acoustic emission characteristics. Although interferometric radar cannot directly acquire acoustic emission signals, the model identifies high-frequency pulse components in the deformation sequence, equating them to the macroscopic manifestation of microcrack development. A rock damage evolution equation is embedded in the physical constraints, describing the growth of damage variables with cumulative plastic strain. Risk level determination considers not only the safety factor but also the inverted cumulative damage index. When the damage index reaches a critical value, even if the safety factor remains greater than 1, the system triggers a warning level alert, indicating a potential local collapse.

[0058] In the early warning response mechanism of step 5, this embodiment adds an "expert collaborative decision-making" step. While generating a visual report, the system pushes the inverted mechanical profile, displacement sequence, and real-time on-site monitoring video to a cloud-based expert system. The expert system uses knowledge graph technology to compare the current characteristics of the slope with similar cases in the historical landslide database. If the system's judgment result is inconsistent with the risk assessment given by the expert system, a multi-parameter weighted re-judgment process is automatically triggered.

[0059] In terms of deployment, edge computing nodes are integrated into a unified intelligent monitoring terminal. This intelligent monitoring terminal incorporates a solar power management module, a 4G / 5G communication module, and a high-precision inertial navigation auxiliary sensor. The inertial navigation sensor monitors the terminal's own attitude changes and further improves the reliability of deformation data by correcting radar line-of-sight errors caused by device tilt in real time. The intelligent monitoring terminal performs preliminary phase unwrapping and atmospheric correction locally, uploading only high-value deformation sequences and preliminary early warning signals to the data center, significantly reducing data transmission latency.

[0060] During online updates, the system incorporates a self-supervised learning mechanism. When real mechanical parameters provided by geological surveys are lacking, the physical information neural network automatically seeks the optimal combination of physical parameters by assessing the consistency of deformation data predictions across different time periods.

[0061] The system attempts multiple sets of different values ​​for cohesion and internal friction angle, calculates the convergence rate of the physical constraint terms under each set of parameters, and finally selects the parameter set that causes the fastest decrease in the residuals of the physical constraint terms as the effective attributes of the current slope. This adaptive parameter identification capability enables the early warning method to be applied to field slopes with scarce data.

[0062] The visualized early warning report also integrates an "emergency resource dispatch map." When a dangerous level warning is issued locally, the map automatically marks the locations of the nearest evacuation points, emergency warehouses, hospitals, and emergency rescue teams to the slope. Based on the estimated landslide impact range, the system automatically calculates the affected road sections and generates multiple evacuation route suggestions. This information is synchronized in real time to the handheld terminals of rescue personnel, achieving a closed-loop linkage between monitoring, early warning, and response.

[0063] For landslides caused by prolonged rainfall, this embodiment also provides a time-series prediction function based on the coupling of a long short-term memory network and a physical information neural network. The long short-term memory network layer is responsible for capturing the hysteresis effect between rainfall history and deformation, and its output serves as the driving input for the physical information neural network. Through this combination, the system can simulate the evolution trajectory of pore water pressure inside the slope over the next week under the current continuous rainfall intensity, extending the warning time lead from "at the time of occurrence" to "before occurrence".

[0064] Example 4: This example further illustrates the application of this method in slope monitoring of open-pit mines in large mines, focusing on massive data processing and multi-level model collaboration.

[0065] Due to the massive scale of open-pit mine slopes, step 1 involves deformation monitoring of hundreds of thousands of grid cells. To improve processing efficiency, a distributed computing architecture is adopted. The interferometric processing task is distributed to multiple parallel computing nodes, each responsible for phase unwrapping and trajectory correction in a specific region. The extracted deformation field sequence is aggregated in real time to the central processing unit via a high-performance data bus.

[0066] In the model construction of step 2, a two-level physical information neural network, "global-local," was employed. The global model provides a coarse model of the entire mining slope, primarily constraining the overall structural balance and gravity loads of the slope. The local model, on the other hand, refines the model for key areas where deformation anomalies are detected. The local model receives the boundary conditions from the global model and embeds more detailed stratigraphic structure information and mine excavation sequence information.

[0067] The physical constraints incorporate the logic for the redistribution of the initial stress field caused by excavation unloading. Specifically, each excavation step results in an unloading displacement of the slope towards the free face, and the neural network must fit the measured displacement while satisfying the unloading mechanical boundaries.

[0068] The inversion process in step 3 achieves deep integration with the mine production plan. The system receives real-time access to the operating position of the excavators in the mining area and the blasting plan. When the model detects a sudden change in instantaneous displacement caused by blasting, the physical constraints automatically identify it as a dynamic load disturbance rather than a slope instability trend. Through this information fusion, the system can filter out normal production disturbances and improve the accuracy of early warnings.

[0069] In step 4, the risk level determination incorporates a probabilistic assessment model for slope stability. Considering the spatial variability of soil and rock parameters, the system employs a combination of Monte Carlo simulation and a physical information neural network. Based on the mechanical parameters output by the neural network, and given a reasonable range of parameter fluctuations, multiple inversion calculations are performed to derive an instability probability distribution map. If the instability probability exceeds 30%, an early warning is triggered. This probability-based determination method better meets the risk control requirements of mine safety management.

[0070] The visualized early warning report generated in step 5 is integrated into the mine's digital twin platform. By overlaying 3D panoramic images with monitoring data, managers can roam in virtual space and view the stability status of various levels of ramps on the slope. The system also has the function of automatically generating "weekly reports" and "monthly reports." By summarizing the evolution trend of the deformation field over a period of time, it automatically identifies deformation acceleration zones, settlement stability zones, and potential risk points, providing a basis for decision-making in the mine's slope reinforcement projects.

[0071] In mining applications, the online update mechanism manifests as a dynamic update of the current excavation geometry. As excavation progresses, the slope geometry continuously changes. The system uses drones for periodic inspections to acquire the latest digital surface model and automatically updates the geometric constraints of the input layer of the physical information neural network. This geometric adaptability ensures that the physical constraint equations always act on the correct spatial boundaries.

[0072] To address the unique working conditions of open-pit mines, edge computing nodes are ruggedized into industrial-grade devices that are explosion-proof, waterproof, and earthquake-resistant. The devices integrate local storage units, ensuring continuous monitoring and data recording even in the event of a temporary interruption of the mine's wireless network, with data automatically synchronized to the cloud once the network is restored.

[0073] The visual early warning report also supports "virtual excavation simulation". Technicians can simulate future excavation plans in the report interface. The physical information neural network predicts changes in slope stability under different excavation plans based on learned mechanical properties, optimizes mining design plans, and prevents geological disasters.

[0074] Example 5: This example details a monitoring scheme for reservoir bank slopes or slopes affected by water level changes.

[0075] In step 1, the acquired interferometric radar image data requires special processing to address phase noise caused by reflections from the reservoir surface. The system utilizes coherence masking technology to automatically identify and remove invalid pixels within the water area. Simultaneously, to monitor potential deformation in the underwater portion, the system incorporates data from a multi-segment inclinometer deployed underwater. Although interferometric radar cannot penetrate water, the physical information neural network uses the deep displacement data acquired by the underwater inclinometer as additional training labels, forcing the model to satisfy physical constraints even in the subsurface region.

[0076] In step 2, the physical constraint term primarily integrates the fluid-structure interaction effect caused by reservoir water level fluctuations. Its textual logical description is as follows: When the reservoir water level rises, it generates a normal pressure perpendicular to the slope surface. Simultaneously, the increased water head pressure penetrates into the slope body through seepage, reducing effective stress. The physical constraint term in the loss function explicitly includes this external water pressure load. When the reservoir water level drops rapidly, the pore water pressure discharge within the slope body lags behind, and the resulting reverse seepage force leads to a rapid decrease in stability. The model dynamically calculates the permeability of the soil and rock mass by capturing the sensitivity of the deformation sequence to water level changes.

[0077] The inversion output in step 3 now includes a real-time inversion of the "saturation line position." The saturation line is the free surface of groundwater within the slope. The physical information neural network analyzes the deformation differences at different heights of the slope and, combined with the unsaturated seepage physical equations, infers the specific shape of the saturation line at the current moment. The position of the saturation line directly affects the search logic of the sliding surface and is a core parameter for reservoir bank slope risk assessment.

[0078] In step 4, the early warning triggering logic is made more sensitive to the rate of change in reservoir water level. The system acquires real-time scheduling data for the reservoir area, and when the rate of water level decline exceeds the preset daily limit and the cumulative displacement of the slope increases, the early warning level is automatically raised by one level. The risk level determination matrix includes not only mechanical parameters but also hydrological triggering factors.

[0079] The visualized early warning report in step 5 now includes a dedicated page for "Reservoir Water Level-Displacement Correlation Analysis." This page uses statistical charts to display the synchronous evolution of water level fluctuation curves and deformation curves at key monitoring points, and calculates the cross-correlation coefficient between the two. In this way, managers can clearly distinguish whether the deformation is caused by seasonal fluctuations induced by water level or by a precursor to a trend-based instability.

[0080] The early warning response mechanism has been enhanced to include the ability to link vessel evacuation signals in reservoir-bank scenarios. When a danger-level warning is issued, the system sends navigation warnings to vessels around the reservoir area via the Automatic Identification System (AIS) to prevent secondary surges caused by landslides from threatening waterway safety.

[0081] The online update mechanism performs segmented fine-tuning based on the hydrological characteristics of different seasons. The system identifies whether it is currently in a "dry season," "wet season," or "drawdown period," and retrieves the initial values ​​of physical parameters for the corresponding season from the historical database for fine-tuning. This seasonally adaptive fine-tuning strategy further improves the accuracy of mechanical inversion under complex hydrological conditions.

[0082] When deploying edge computing nodes on the reservoir bank, the corrosive effects of moisture on electronic equipment were taken into account. Fully sealed, nitrogen-filled protective enclosures were used, and satellite links were used as a backup communication method to ensure real-time data transmission and risk warnings even under extreme environments such as dam flood discharge.

[0083] The visualization report also provides a "surge simulation module". Once the risk of a landslide is determined to be extremely high, the system simulates the surge height and propagation range after the landslide enters the water, based on the potential sliding volume, sliding speed, and reservoir topography obtained through inversion, providing more accurate emergency avoidance references for downstream threatened areas.

[0084] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for real-time monitoring and early warning of slope deformation based on interferometric radar images, characterized in that, The method includes the following steps: acquiring multi-temporal interferometric radar image data and extracting a high-precision surface deformation field sequence from it; A physical information neural network model integrating the constitutive relationship of slope rock and soil mechanics is constructed. The loss function of the physical information neural network model is composed of a data fitting term that measures the sum of squares of the deviations between the model output displacement and the measured deformation data, and a physical constraint term that embeds the basic control equations describing the stress and strain behavior of slope rock and soil. The surface deformation field sequence is used as input to drive the physical information neural network model to perform real-time inversion of the internal mechanical state of the slope through internal multi-layer nonlinear mapping relationships, and simultaneously output the stress distribution field, pore water pressure evolution trend and potential sliding surface location trajectory of key areas inside the slope. Based on the mechanical state parameters obtained from the inversion, the slope instability risk level is determined by comparing the current mechanical indicators with the preset threshold. Based on the risk level, the corresponding early warning response mechanism is triggered, and a visual early warning report is generated based on the geographic information system platform, which integrates surface deformation heat map, mechanical state profile map and risk level zoning map.

2. The method for real-time monitoring and early warning of slope deformation based on interferometric radar images according to claim 1, characterized in that, The steps of acquiring multi-time-series interferometric radar image data and extracting high-precision surface deformation field sequences specifically include: acquiring the original echo signal through a synthetic aperture radar sensor, and using an imaging algorithm to convert the original echo signal into single-scatter complex image data containing amplitude and phase information; Co-registration of multiple single-scattered complex images is achieved by selecting one image as the main image and calculating the offsets of the remaining sub-images relative to the main image in the range and azimuth directions, thus realizing sub-pixel level spatial overlap. Differential interferometry is performed on the registered image pairs to generate an interferometric phase map, and external digital elevation model data is introduced to remove the reference terrain phase component; Phase unwrapping is performed using the branch cutting method, the minimum cost flow algorithm, or the least squares method to restore the phase values ​​wrapped between negative and positive pi to continuous phase values ​​in order to eliminate multivaluedness. An atmospheric correction model is established using the statistical characteristics of stable point targets or synchronous meteorological observation data to compensate for ionospheric and tropospheric delays, generating a surface deformation field sequence organized in the form of a spatial grid. Each grid cell records the millimeter-level cumulative deformation relative to the reference time.

3. The method for real-time monitoring and early warning of slope deformation based on interferometric radar images according to claim 1, characterized in that, The steps of constructing a physical information neural network model that integrates the constitutive relationship of slope rock and soil mechanics specifically include: setting the architecture of the physical information neural network model, the architecture including an input layer that receives spatial coordinate parameters and time parameters, a hidden layer containing multiple neurons, and an output layer that outputs displacement components and stress tensor components. The neurons in the hidden layer are connected with weights and bias parameters, and hyperbolic tangent function or modified linear unit is used as activation function to handle nonlinear features. The residual evaluation logic in the physical constraint term is constructed by calculating the derivatives of the neural network output with respect to the input coordinates and time parameters using automatic differentiation techniques. An adaptive weight adjustment strategy is adopted during model training. In the early stage of training, the data fitting term is given a high contribution ratio. As the number of iterations increases, the contribution ratio of the physical constraint term to the total loss function is dynamically increased to ensure that the model output conforms to the surface observation trend while satisfying the laws of engineering mechanics.

4. The method for real-time monitoring and early warning of slope deformation based on interferometric radar images according to claim 3, characterized in that, The basic governing equations embedded in the physical constraint terms are defined in literal logic as: equilibrium equation logic, which requires that, ignoring body forces, the sum of the partial derivatives of the normal stress components with respect to the corresponding spatial coordinates, plus the sum of the partial derivatives of the shear stress components with respect to the cross coordinates, should approach zero at any given time. The logic of geometric equations requires that the strain components of soil and rock be defined as the combination of the partial derivatives of the displacement components with respect to spatial coordinates; The constitutive equation logic requires that the relationship between stress and strain follows the Mohr-Coulomb strength criterion, that is, the proportional relationship between shear stress and normal stress is limited by the internal friction angle property and cohesion property of the material. The logic of the seepage continuity equation requires that the rate of change of soil moisture content be equal to the negative value of the spatial divergence of the seepage velocity, where the seepage velocity is equal to the product of the permeability coefficient and the total head gradient. The weakening effect of rainfall infiltration on stress is reflected through the pore water pressure variable.

5. The method for real-time monitoring and early warning of slope deformation based on interferometric radar images according to claim 1, characterized in that, The steps for real-time inversion of the internal mechanical state of the slope specifically include: inputting the current time and the surface deformation field sequence of several historical observation times into the trained physical information neural network model; The model identifies regions with large displacement gradients based on the input deformation sequence and, in conjunction with the physical constraint terms, infers the degree of stress concentration within the region. By analyzing the rate characteristics of displacement change over time, and combining this with seepage physical constraints, the degree of increase in internal pore water pressure is determined. By searching for the continuous trajectory line where the shear strain increment reaches its peak value, the geometric boundary of the potential sliding surface with the lowest safety factor inside the slope is determined. The inversion results are output in the form of high-dimensional tensors, covering the three-dimensional stress tensor field, fluid pressure field, and cumulative damage field inside the slope.

6. The method for real-time monitoring and early warning of slope deformation based on interferometric radar images according to claim 5, characterized in that, The inversion process also includes a multi-source heterogeneous data fusion strategy, specifically: accessing real-time point displacement data provided by fixed monitoring stations of the Global Navigation Satellite System deployed on the slope surface; The loss function of the physical information neural network model is increased with a local calibration term, and the real-time displacement data of the global navigation satellite system is used as a hard constraint to force the output value of the neural network at the corresponding coordinate point to be consistent with the satellite positioning result. The deformation field sequence extracted from the interferometric radar image serves as a soft constraint, providing background features for spatial continuity to the neural network, thereby enabling point-to-surface complementary mechanical state inversion.

7. The method for real-time monitoring and early warning of slope deformation based on interferometric radar images according to claim 1, characterized in that, The steps for determining the risk level of slope instability specifically include: extracting the safety factor, pore water pressure ratio, and shear strain increment rate as core evaluation indicators. The safety factor is defined as the ratio of slope anti-sliding force to sliding force. A probability-based judgment logic is established. By considering the fluctuation range of given soil and rock parameters and combining multiple inversion calculations in Monte Carlo simulation, a probability distribution map of slope instability is obtained. If the safety factor is greater than 1.5, the slope is determined to be in a stable state. If the safety factor is between 1.2 and 1.5, the pore water pressure ratio exceeds the preset critical ratio value, and the shear strain increment rate shows a continuous upward trend, the slope is determined to enter the corresponding level of risk state based on the probability of instability. The preset threshold is calibrated based on the regional geological characteristics and the statistical analysis results of historical landslide cases.

8. The method for real-time monitoring and early warning of slope deformation based on interferometric radar images according to claim 1, characterized in that, The specific steps of the early warning response mechanism include: dividing the early warning level into three levels: attention level, alert level, and danger level; When a slope is identified as a level of concern, the current abnormal trend is automatically recorded and the slope is added to the list of key monitoring areas, thereby increasing the frequency of interferometric radar data acquisition and processing. When the alert level is determined, an early warning notification is automatically generated and alarm information is sent to relevant management departments and persons responsible for production safety, triggering instructions to strengthen on-site inspections; When the situation is determined to be dangerous, the emergency command platform will automatically activate the personnel evacuation plan, trigger the ground-based audible and visual alarm facilities in the affected area, send a road closure request to the traffic management department, and simultaneously activate the emergency resource dispatch map to mark safe points and evacuation routes.

9. The method for real-time monitoring and early warning of slope deformation based on interferometric radar images according to claim 1, characterized in that, The specific methods for generating and deploying the visualization early warning report include: using a geographic information system platform to display the cumulative displacement at different locations on the earth's surface through varying shades of color to generate a heat map; The discrete mechanical variables output by the neural network are converted into continuous three-dimensional volume data by spatial interpolation technology, generating a cross-sectional view showing the distribution of internal stress and pore water pressure on the slope section. The physical information neural network model is deployed on edge computing nodes. The number of parameters is reduced through model pruning technology, and floating-point parameters are converted into integer parameters using model quantization technology. This improves inference speed while ensuring inversion accuracy and meets the millisecond-level response requirements. The edge computing node performs deformation processing, mechanical inversion, and risk assessment locally, and sends the results to a remote server to reduce the bandwidth pressure on the wireless communication link.

10. The method for real-time monitoring and early warning of slope deformation based on interferometric radar images according to claim 1, characterized in that, The method also includes an online update mechanism for the physical information neural network model, specifically: calculating the residual distribution between the model-predicted displacement and the newly observed displacement in real time, and determining that model drift has occurred when a systematic deviation in the residual is detected in the region; The latest observation data and supplementary geological exploration data of the region are extracted, and the hidden layer weights of the model are locally updated using an incremental learning algorithm. During the update process, the structure of the physical constraint terms remains unchanged, and the connection weights are adjusted by gradient descent. A self-supervised learning mechanism is introduced to automatically identify the optimal combination of physical parameters by predicting the consistency of deformation data over different time periods. This enables the model to adaptively reflect the weakening of soil parameters caused by long-duration rainfall or the changes in geometric boundaries caused by mining.

Images