Intelligent landslide early warning method and system based on multi-source data fusion and dynamic threshold

Abstract

The invention discloses an intelligent landslide early warning method and system based on multi-source data fusion and a dynamic threshold value, and belongs to the technical field of geological disaster monitoring and early warning. The method comprises the following steps: collecting monitoring data through a multi-source sensor network; carrying out adaptive weighted fusion to generate a space-time fusion feature vector; constructing an initial early warning model based on historical data, coupling physical and mechanical parameters inversed in real time, performing time sequence prediction through a recurrent neural network, and dynamically generating an early warning decision boundary; carrying out risk assessment and graded early warning by calculating a multi-source collaborative instability index of the real-time features in a high-dimensional feature space; and performing closed-loop optimization on fusion and prediction parameters according to an early warning verification result. The system correspondingly comprises five function modules. According to the method, an intelligent closed loop of perception-fusion-prediction-decision-optimization is constructed, so that the problems of single data source, rigid threshold and poor adaptive capacity of a traditional method are solved, and the early warning success rate, the early warning time and the reliability are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of geological disaster monitoring and early warning technology, and in particular to a landslide intelligent early warning method and system based on multi-source data fusion and dynamic threshold. Background Technology

[0002] Landslides are a common and highly destructive geological hazard in my country and globally. Traditional landslide early warning methods mainly rely on single or a few monitoring indicators (such as surface displacement and cumulative rainfall) and use fixed thresholds based on historical experience or statistical analysis for judgment. These methods have significant limitations: First, a single data source cannot comprehensively and reliably reflect the complex coupling state of the landslide body from its surface to its depths, and from triggering factors to deformation response. The system has poor robustness and is prone to false alarms or missed alarms when sensors malfunction or there is noise interference in the local environment. Second, static early warning thresholds cannot adapt to the dynamic changes in slopes caused by seasonal changes, vegetation growth, long-term creep, and human engineering activities. For example, a sensitive threshold used during the rainy season may trigger numerous false alarms during the dry season, while thresholds set based on long-term stable states may be slow to respond to accelerated deformation, resulting in delayed early warnings. Third, existing early warning logic is mostly a simple "threshold exceeding the limit triggers an alarm," lacking in-depth analysis and intelligent comprehensive judgment of the nonlinear and temporal co-evolutionary laws among multiple monitoring parameters, resulting in insufficient scientific rigor and accuracy in early warning decisions.

[0003] In recent years, with the development of IoT and AI technologies, multi-source monitoring and intelligent early warning have become research hotspots. However, existing improvements often treat data fusion and threshold analysis as two independent processes, simply piling on technologies. For example, some inventions merely mechanically aggregate multi-source data and input it into a classification model, ignoring the dynamic differences in data quality and the constraints of physical mechanisms; others attempt to adjust thresholds, but these are mostly periodic static corrections based on historical data, lacking the ability to predict current trends over time. These inventions fail to fundamentally construct an integrated adaptive intelligent system of "perception-decision-optimization," resulting in limited performance improvements. Furthermore, the long-term operation and optimization of the system still heavily rely on human experience, making large-scale reliable application in practical engineering difficult. Summary of the Invention

[0004] In view of this, the present invention addresses the deficiencies of the existing technology, and its main objective is to provide a landslide intelligent early warning method and system based on multi-source data fusion and dynamic thresholds. By breaking down the barriers between data perception, decision-making and system optimization, it constructs an intelligent early warning system that is deeply coupled and operates in a closed loop, thereby significantly improving the accuracy, timeliness and adaptability of the early warning system.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A landslide intelligent early warning method based on multi-source data fusion and dynamic thresholds includes the following steps: S1. Through a multi-source heterogeneous sensor network deployed in the landslide body and surrounding environment, real-time monitoring data including surface displacement, deep displacement, soil moisture content, rainfall and rock and soil vibration signals are collected. S2. The multi-source monitoring data collected in step S1 is cleaned, interpolated, time-synchronized and spatially registered. An adaptive weighted fusion algorithm based on information entropy and predefined geomechanical rules is used to fuse the multi-source data at the feature level to generate a spatiotemporal fusion feature vector that can comprehensively and robustly characterize the overall stability state of the landslide. S3. Based on the spatiotemporal fusion feature vector of the historical long time series, an initial early warning model describing the spatial boundary of the normal state of the slope is constructed using a type of support vector machine; and based on the real-time fusion features within the sliding time window, the slope physical and mechanical state parameters obtained by real-time inversion are fused, and the time series prediction is performed through a recurrent neural network model to dynamically generate an early warning decision boundary that is adaptive to the current environment and the creep state of the slope. S4. Compare the spatiotemporal fusion feature vector generated in real time with the dynamic early warning decision boundary obtained in step S3. Calculate the "multi-source collaborative instability index" in the high-dimensional feature space jointly constructed by historical instability samples and normal samples to conduct a quantitative comprehensive assessment of landslide risk, and issue graded early warning information according to the preset index level. S5. Based on the subsequent verification and accuracy assessment of the early warning results, feedback data containing the comparison information between the early warning results and the actual slope status is generated. The fusion weight allocation strategy in step S2 and the internal parameters of the recurrent neural network model in step S3 are dynamically optimized to achieve adaptive evolution and improvement of the overall early warning performance of the system.

[0006] As a preferred option: In step S2, the weight coefficients of the adaptive weighted fusion algorithm are dynamically determined, specifically based on the signal-to-noise ratio and data missing rate of each data source in real time, as well as the gray correlation degree between each data source and the core deformation monitoring sequence in the time domain, to ensure the stability of the system fusion results when some sensors fail or are interfered with.

[0007] As a preferred option: In step S3, the slope physical and mechanical state parameters obtained by real-time inversion are integrated, specifically: using the soil moisture content and rainfall infiltration data within the sliding window, based on a simplified infinite slope model or an unsaturated soil mechanics model, the transient value or rate of change of the slope stability coefficient Fs is calculated in real time, and the output of the infinite slope model or the unsaturated soil mechanics model is used as one of the key features and input into the recurrent neural network model.

[0008] As a preferred option: In step S3, the recurrent neural network model adopts a long short-term memory network structure, and its training process introduces a reinforcement learning framework. The model's reward function is designed to be positively correlated with the early warning accuracy and effective early warning duration, and negatively correlated with the false alarm rate, guiding the model to learn the optimal threshold dynamic adjustment strategy.

[0009] As a preferred approach, step S4 involves the following specific calculation method for the multi-source collaborative instability index: First, principal component analysis is used to reduce the dimensionality of the fused feature vectors of historical normal and unstable states and construct the feature space; second, the Mahalanobis distance from the real-time fused feature vectors to the cluster center of the normal state and the shortest projection distance to the known instability evolution path are calculated in the feature space; finally, the multi-source collaborative instability index is a weighted sum of the Mahalanobis distance and the shortest projection distance, used to simultaneously quantify the degree of anomaly of the state and the tendency to evolve towards an unstable state.

[0010] As a preferred approach: In step S5, the system automatically records each early warning event and its subsequent verification results to establish an early warning effect sample library; based on this sample library, the system periodically fine-tunes the parameters of the data fusion weight calculation function and the threshold prediction network using a backpropagation algorithm or an evolutionary algorithm to reduce the probability of misjudgment under similar working conditions in the future.

[0011] A landslide intelligent early warning system based on multi-source data fusion and dynamic thresholds for implementing the method includes: The data sensing and acquisition module consists of various physical sensors and data transmission units, and is responsible for acquiring and remotely transmitting multi-source monitoring data. The intelligent fusion and processing center receives data from the perception layer, has a built-in data processing unit and fusion algorithm library, executes step S2, and outputs a spatiotemporal fusion feature vector. The dynamic threshold analysis engine, connected to the intelligent fusion and processing center, includes a historical feature database, an initial early warning model library, and a dynamic threshold prediction unit. It executes step S3 as described in claim 1 and outputs the dynamic early warning decision boundary. The intelligent early warning decision module is connected to the intelligent fusion and processing center and the analysis engine. It is equipped with a risk assessment unit and executes step S4 to generate and issue graded early warning instructions. The closed-loop self-optimization module is connected to the early warning decision module and feeds back to the intelligent fusion and processing center and the dynamic threshold analysis engine. Step S5 is executed to realize the system's self-learning and performance optimization.

[0012] As a preferred solution, the data sensing and acquisition module adopts a hybrid network topology combining star and mesh topologies. In this topology, GNSS receivers and rain gauges serve as backbone nodes, while soil moisture sensors, crack gauges, and other devices serve as subordinate nodes. The microseismic monitoring network is independently networked, and data is initially aggregated and converted into protocols through an edge computing gateway before being uploaded.

[0013] As a preferred option: the intelligent fusion and processing center and the dynamic threshold analysis engine are deployed on cloud servers or high-performance edge computing devices, adopting a containerized microservice architecture; the intelligent early warning decision module and the visual human-computer interaction platform are deployed on web application servers, supporting access from multiple terminals.

[0014] As a preferred embodiment, the system also includes a 3D visualization and decision support platform. This platform can integrate a geographic information system to display the spatial distribution of multi-source data, fusion feature curves, dynamic early warning decision boundaries, 3D deformation field cloud maps of landslide bodies, and spatial overlay effects of early warning levels in real time, providing managers with a panoramic decision-making view.

[0015] This invention differs significantly from traditional static thresholding or post-hoc correction methods. It proposes a "feedforward dynamic thresholding technique coupled with physical mechanisms." The system first trains a support vector machine model using a large amount of historical normal-state data to define the initial normal-state boundary. Its innovation lies in the subsequent dynamic adjustment process: the system sets a sliding time window and continuously analyzes the statistical regularities and trends of the real-time fused feature vectors within the window. More importantly, the system, based on real-time hydrological data, inversely retrieves the changing trend of the slope stability coefficient using a simplified physical mechanics model (such as an infinite slope model). Then, the statistical features reflecting data regularity, trend features, and stability change features reflecting the physical mechanism are input into a pre-trained recurrent neural network (such as an LSTM). The network's task is to predict how the warning threshold boundary should be adjusted (e.g., tightened or loosened) to adapt to the latest evolutionary trend of the slope state in the near future. This achieves adaptive and forward-looking adjustment of the threshold based on the environment and the slope's own state, rather than a passive response.

[0016] Traditional data-driven models lack physical constraints, potentially leading to predictions that violate mechanical principles; while purely physical models struggle to handle complex environments and unknown parameters. This invention uses key state variables (Fs and their rate of change) output from the physical model as time-varying features input to the data-driven model. Essentially, it guides and constrains the learning process of the neural network with physical laws, ensuring that the threshold adjustment direction of its predictions not only conforms to data patterns but also to the inherent logic of slope mechanics evolution. This solves the problem of integrating the unreliability of purely data-driven "black box" methods with the poor adaptability of purely physical methods.

[0017] Based on dynamic thresholds, the early warning decision-making of this invention is more intelligent. Instead of simple scalar comparisons, the system introduces the concept of a "multi-source collaborative instability index." This index is obtained by calculating the position of the real-time fused feature vector in a high-dimensional feature space constructed from historical data. Specifically, it simultaneously measures the "anomaly" of the current state deviating from the center of the normal state cluster, and the "tendency" of its approximation of the evolution trajectory of historical instability samples. This trajectory approximation judgment in high-dimensional space is far more effective than one-dimensional threshold comparisons in capturing complex collaborative precursors before landslide instability, thus significantly improving early warning accuracy.

[0018] One of the significant advancements of this invention lies in the introduction of a closed-loop self-optimization mechanism. The system records the issuance of each warning and the verification results of whether an actual landslide has occurred, generating feedback. This feedback data is periodically used to adjust the weight calculation strategy in data fusion and the parameters of the dynamic threshold prediction model through optimization algorithms. This allows the system to continuously learn from practical operational experience, especially from false alarms and missed alarms, constantly optimizing its judgment criteria. Consequently, it possesses intelligent characteristics that evolve and improve over time, greatly reducing the reliance on manual parameter tuning in long-term operations.

[0019] This invention utilizes closed-loop self-optimization to enable the system to continuously track and adapt to "concept drift" (the properties of the landslide body, environmental background, and sensor performance all change slowly over time, causing models trained on historical data to gradually become ineffective). The system not only optimizes parameters but also dynamically updates its implicit definitions of "normal state" and "precursors of instability," thereby ensuring the stability of its early warning performance during long-term (years) monitoring—something that static systems or systems requiring periodic manual adjustments cannot achieve.

[0020] To implement the above method, the corresponding system of this invention includes a data sensing and acquisition module, an intelligent fusion and processing center, a dynamic threshold analysis engine, an intelligent early warning decision-making module, and a closed-loop self-optimization module. These modules, through software and hardware collaboration, form a complete technology chain from data acquisition to early warning issuance and self-optimization. The system can adopt a cloud-edge-device collaborative architecture to ensure the real-time performance of data processing and the feasibility of computation.

[0021] In summary, this invention creatively integrates multi-source adaptive fusion, dynamic threshold time-series prediction coupled with physical models, high-dimensional feature space collaborative instability assessment, and feedback-based closed-loop optimization into a unified intelligent system, forming a complete technical invention that is distinct from existing technologies and not readily apparent. This invention effectively overcomes the inherent defects of traditional early warning methods, significantly improving early warning performance and possessing outstanding substantive features and remarkable progress.

[0022] Compared with existing technologies, this invention has significant advantages and beneficial effects. Specifically, as can be seen from the above technical solution, by constructing a full-link intelligent closed-loop system of "perception-fusion-prediction-decision-optimization," it fundamentally overcomes the inherent defects of traditional landslide early warning methods, such as single data source, rigid thresholds, and lack of adaptive capabilities. Its beneficial effects are specifically manifested in: significantly improved early warning accuracy, with a substantial increase in the success rate of early warnings in comparative experiments; significantly enhanced early warning timeliness, with the average effective early warning time shortened, gaining crucial time for emergency response; outstanding system reliability and practicality, with an effective reduction in false alarm rate during long-term operation and backtesting. Simultaneously, its unique closed-loop self-optimization mechanism enables the system to continuously adapt to environmental changes and sensor performance drift, significantly reducing reliance on manual maintenance and achieving continuous autonomous evolution of early warning performance.

[0023] To more clearly illustrate the structural features and effects of the present invention, a detailed description is provided below in conjunction with the accompanying drawings and specific embodiments. Attached Figure Description

[0024] Figure 1 This is a schematic diagram illustrating the overall process and working principle of the closed-loop system of the method of the present invention.

[0025] Figure 2 This is a diagram showing the architecture and data interaction relationships of each module in the system of this invention.

[0026] Figure 3 This is a schematic diagram illustrating the fusion of physical model inversion features and data-driven features during the dynamic threshold generation process.

[0027] Figure 4 This is a schematic diagram illustrating the principle of calculating the "multi-source cooperative instability index" in a high-dimensional feature space. Detailed Implementation

[0028] The present invention is as follows Figures 1 to 4 As shown, a landslide intelligent early warning method and system based on multi-source data fusion and dynamic thresholds are presented, wherein... like Figure 1As shown, this illustrates the closed-loop workflow of the intelligent early warning method described in this invention, from perception to optimization. The process begins with the multi-source data perception module, where raw data collected by various sensors undergoes adaptive weighted fusion processing (including cleaning, interpolation, synchronization, registration, and dynamic weight calculation) to generate a core spatiotemporal fusion feature vector. This vector simultaneously drives two paths: first, constructing an initial early warning decision boundary based on a type of support vector machine and historical data; second, combining the results of sliding window feature extraction and physical model inversion (infinite slope model), inputting them into an LSTM time-series prediction network, and outputting dynamically adjusted parameters. These two paths converge at the decision boundary synthesis stage, generating a dynamic early warning decision boundary. This boundary, along with the spatiotemporal fusion feature vector, is input into the intelligent early warning decision module, where the multi-source collaborative instability index (I=α·D) is calculated. m +β·D t The system ultimately outputs a Level 4 warning. After the warning results are released and verified, feedback data is generated, which drives the closed-loop self-optimization module. The latter adjusts the fusion weights and LSTM network parameters in reverse through optimization algorithms, thereby forming a continuously self-evolving closed-loop intelligent warning system.

[0029] The early warning method specifically includes the following steps: S1. Multi-source data acquisition. First, based on the geological survey results, a multi-source heterogeneous sensor network was deployed at key locations of the landslide (such as the rear edge, main slip zone, and leading edge) and surrounding stable areas. The network specifically includes: 8 high-precision GNSS receivers for three-dimensional surface displacement monitoring; a multi-point deep tiltmeter stringer with 20 measuring points deployed in 4 boreholes to capture potential slip surface locations and deep deformations; 15 soil moisture sensors deployed at different depths and locations in the shallow layer; 1 automatic weather station for monitoring rainfall, temperature, and humidity; and a monitoring array consisting of 6 microseismic sensors to capture vibration signals generated by internal fractures or creep in the soil and rock mass. All sensors are networked using low-power IoT technology, and the data is aggregated at the edge computing gateway at the landslide site. After preliminary preprocessing and data encapsulation, the data is transmitted in real-time to the cloud-based early warning platform via a 4G / fiber optic network.

[0030] S2. Adaptive Data Fusion and Spatiotemporal Feature Vector Generation. After receiving multi-source asynchronous data streams, the cloud platform initiates the intelligent fusion and processing flow. The data first enters the preprocessing submodule for missing value interpolation, outlier filtering, and noise reduction. Since the sampling frequencies and timestamps of each sensor are not completely synchronized, the system uses linear interpolation to unify all data to a standard time series once per second and completes unified registration of the spatial coordinate system. Subsequently, the core adaptive weighted fusion is performed. The system uses the GNSS surface displacement sequence as the core reference sequence and dynamically calculates the gray correlation (ρ) between other data sequences (such as deep displacement, soil moisture content, and microseismic energy) and this reference sequence in the time domain every minute. iSimultaneously, the signal-to-noise ratio (SNR) of each data stream is calculated in real time. i ) and instantaneous data missing rate (η) i In this embodiment, the weighting coefficient w i The following relationship is dynamically calculated: w i = (1-η i ) *ρ i *SNR i / Σj[(1-η j )*ρ j *SNR j ],in: η i With η j These represent the real-time data missing rates of the i-th and j-th data sources, respectively. ρ i With ρ j These represent the real-time grey relational degree between the monitoring data sequences of the i-th and j-th data sources and the core deformation monitoring sequence (GNSS surface displacement), respectively. SNR i With SNR j These represent the real-time signal-to-noise ratios of the i-th and j-th data sources, respectively. n represents the total number of data sources participating in the fusion; Σ represents the summation of the corresponding product terms of all n data sources. The purpose of this denominator is to perform normalization processing to ensure that the sum of all weight coefficients is 1.

[0031] The calculation process is also constrained and driven by predefined geomechanical rules. For example, one predefined geomechanical rule is: "When the real-time rainfall intensity (I t Exceeding the soil saturation permeability coefficient (k) s When the duration (T) exceeds the threshold T0, it is determined to be a strong infiltration condition. At this time, the basic value of the grey relational degree between soil moisture content data and slope stability (ρ) is determined. base ) will be based on formula ρ base =f(I t ,k sThe system dynamically increases the weight of soil moisture content data (T) and uses it as a reinforcing factor in the weight calculation. This rule, together with the data quality assessment based on information entropy, drives the final weight allocation. For example, when a weather station detects continuous heavy rainfall, the system temporarily increases the basic weight factor of the soil moisture content data according to the rule to reflect the prior knowledge that rainfall is a key inducing factor. Finally, the system generates a "spatiotemporal fusion feature vector" every 10 minutes, which integrates multi-dimensional information such as displacement, moisture content, and vibration. This vector robustly represents the overall stability state of the landslide body during that period. The weight coefficients of the adaptive weighted fusion algorithm are dynamically calculated and determined based on the real-time signal-to-noise ratio, data missing rate, and gray relational degree with the core deformation sequence of each data source.

[0032] S3. Dynamic Early Warning Decision Boundary Generation. The spatiotemporal fusion feature vector is synchronously input into the dynamic threshold analysis engine. The engine has two built-in models: an initial early warning model and a dynamic prediction model. The initial model uses a one-class support vector machine (SVM) and has been trained offline using historical spatiotemporal fusion feature vectors accumulated during the landslide's stable period over the past three years, generated using the method described in S2. This forms an initial early warning decision boundary in the feature space that encloses the normal state points. The dynamic prediction part runs continuously online. The engine maintains a 30-day sliding time window. Every 10 minutes, it performs the following operations: First, it extracts the statistical features such as the mean, variance, and skewness of the real-time fusion feature vector within the window and uses wavelet transform to extract its low-frequency trend components. Second, it starts the physical model inversion unit in parallel. This unit uses minute-level rainfall sequences and soil volumetric water content data within the same sliding window (the most recent 6 hours in this embodiment) and applies the Green-Ampt infiltration model combined with the infinite slope stability formula to calculate the current transient value of the slope stability coefficient Fs and its average rate of change over the past 6 hours in real time. Next, the statistical features, trend features, and the rate of change of Fs obtained from the physical model inversion are combined into a dynamic feature array, which is then input into a pre-trained Long Short-Term Memory (LSTM) network. This LSTM network has been trained using a large amount of historical time-series data (with the past dynamic feature array as input and whether effective deformation acceleration occurs within a certain future time window as a binary supervision label). Its task is to predict, based on the current and recent states, where the initial One-Class SVM decision boundary should be adjusted (e.g., translation, scaling) and the magnitude of the adjustment in the next 10-minute period. Based on the output of the LSTM network, the system dynamically updates the early warning decision boundary, thereby generating a dynamic early warning decision boundary that adapts to the current hydrological conditions, seasonal factors, and slope creep stage.

[0033] like Figure 3As shown, this diagram illustrates the core innovation in the dynamic early warning decision boundary generation process—the feature fusion mechanism. The left side of the diagram shows the data-driven feature extraction process: the spatiotemporal fusion feature vectors within a sliding time window are fed in parallel into a statistical feature extraction unit (outputting mean, variance, etc.) and a trend feature extraction unit (obtaining low-frequency trend terms through wavelet transform), which are then merged into a data-driven feature group. The right side shows the physical model inversion process: using rainfall and soil moisture data within the same time window, a physical mechanics model (Green-Ampt infiltration and infinite slope stability model) is driven to invert the key physical feature—the rate of change of the slope stability coefficient (Fs)—in real time. The data-driven feature group and the physical features are fused using a "⊕" symbol to form a fused feature array, which is then input into a Long Short-Term Memory (LSTM) network. The dynamically adjusted parameter (Δ) output by the LSTM network ultimately acts on the initial early warning decision boundary, adaptively adjusting it to a dynamic early warning decision boundary through translation, scaling, or rotation.

[0034] S4. Risk Assessment and Graded Early Warning Based on the Cooperative Instability Index. The intelligent early warning decision module is triggered every 10 minutes. It receives the latest spatiotemporal fusion feature vector generated by S2 and the corresponding dynamic early warning decision boundary generated by S3. Its early warning logic does not perform a simple scalar boundary judgment, but instead initiates the "multi-source cooperative instability index" calculation subroutine. First, the subroutine calls the historical feature database, which contains long-term accumulated normal state samples and feature samples from before historical instability events. Principal component analysis is used to reduce the dimensionality of these historical samples, constructing a three-dimensional principal feature space for risk assessment, and determining the cluster centers of normal state samples and typical instability evolution paths. Next, the projection point P of the current spatiotemporal fusion feature vector in this principal feature space is calculated. The Mahalanobis distance D from point P to the normal state cluster center is calculated respectively. M (Quantitative state anomaly degree), and the shortest projected distance D from point P to the historical instability evolution path. T (Quantification of the tendency to evolve towards an unstable mode).

[0035] Finally, the multi-source cooperative instability index I=α*D is calculated. M +β*D T , Among them, D M The Mahalanobis distance from the real-time feature vector projection points to the historical normal state cluster centers is used to quantify the state anomaly degree. D T The shortest projection distance from the projection point to the historical instability evolution trajectory is used to quantify the tendency to evolve towards an instability state; α and β are weighting coefficients determined by optimization based on specific landslide types and historical data, and satisfy α+β=1.

[0036] In this embodiment, after optimization using historical data, α=0.7 and β=0.3 are selected. Based on the preset threshold of index I, the system automatically issues four levels of warnings: "Normal (Blue, I<0.5)", "Attention (Yellow, 0.5≤I<1.0)", "Warning (Orange, 1.0≤I<1.5)", and "Alarm (Red, I≥1.5)". The "Multi-Source Cooperative Instability Index" is obtained by weighting the Mahalanobis distance of real-time features in the historical feature space and the projected distance to the instability trajectory.

[0037] like Figure 4 As shown, this intuitively illustrates the calculation principle and risk assessment logic of the multi-source collaborative instability index. The space consists of three principal axes—the first principal component, the second principal component, and the third principal component—after dimensionality reduction by principal component analysis. Within this space, a large number of historical normal state samples cluster to form an ellipsoidal normal state cluster, with its center marked as the normal state cluster center. A historical instability evolution trajectory radiates outward from the edge of this cluster, representing the typical path pattern of slope evolution from stability to instability. Real-time monitoring data, after fusion and dimensionality reduction projection, is represented in the space as the current state projection point. Risk assessment is achieved by calculating two key distances: first, the red dashed line from the current state projection point to the normal state cluster center, with the "Mahathirmann distance" clearly marked within a red dashed box next to it, used to quantify the degree of anomaly in the current state; second, the blue dashed line from the current state projection point to the historical instability evolution trajectory (perpendicular to the trajectory), with the "shortest projection distance" marked within a blue dashed box next to it, used to quantify its tendency to evolve towards an instability mode. The figure prominently illustrates that the multi-source coordinated instability index is a weighted sum of these two types of distances, reflecting the comprehensive risk assessment results. This figure clearly demonstrates how, through positional relationships in a high-dimensional feature space, a more refined and advanced comprehensive assessment of landslide risk can be achieved compared to traditional one-dimensional threshold comparisons.

[0038] S5. Closed-Loop Self-Optimization. During continuous system operation, the closed-loop self-optimization module works in the background. It automatically records all contextual information for each warning event, including the warning time, level, fusion feature vector at the time of triggering, dynamic decision boundary state, and the actual slope state (stable, slightly deformed, locally unstable, etc.) confirmed subsequently by manual inspection, video surveillance, or higher-precision monitoring methods, forming a warning effect sample library. At the end of each month, the module starts an optimization cycle. Based on the new samples accumulated in the past month, especially false alarms and missed alarms, it uses the gradient descent algorithm to construct a loss function, fine-tunes the weight calculation parameters (such as coefficients in gray relational analysis) in the adaptive weighted fusion algorithm described in stage S2, and simultaneously performs incremental training on the LSTM dynamic prediction network described in stage S3. The purpose is to enable the system to continuously learn from actual operating experience, automatically adapt to challenges such as sensor performance degradation and slow changes in environmental background (concept drift), thereby continuously improving the warning accuracy and reducing the false alarm rate in long-term operation, achieving adaptive evolution of system performance. The closed-loop optimization is based on the warning effect sample library, fine-tuning the fusion weight calculation function and prediction network parameters.

[0039] To verify this invention, a rigorous backtesting and comparative experiment was conducted using historical monitoring data of reservoir landslides over the past five years. Two control groups were set up: control group A used the complete method proposed in this invention; control group B used the existing traditional "GNSS displacement rate threshold + 24-hour cumulative rainfall threshold" dual-control early warning method for the reservoir. The backtesting covered two actual small-scale landslide events that occurred during the period. Experimental results showed that: in terms of early warning success rate, control group A successfully issued warnings for both events (100% success rate), while control group B only successfully issued warnings for one event (50% success rate); in terms of early warning lead time, control group A had an average effective early warning lead time of 26 hours, while control group B only had 9 hours; in terms of false alarm control, control group A generated 4 false alarms during the five-year backtesting period, while control group B generated 19 false alarms. The above data demonstrate that this invention significantly improves the reliability, timeliness, and accuracy of early warning systems.

[0040] like Figure 2 As shown, the intelligent early warning system that implements this method adopts a "cloud-edge-device" collaborative architecture. Its core design purpose is to build a closed-loop system that is comprehensive in perception, intelligent in decision-making, and capable of self-evolution, so as to solve the fundamental defects of traditional early warning methods, such as single data source, rigid threshold, and lack of adaptive capability.

[0041] Figure 2The diagram illustrates the hardware and software system components for implementing the method of this invention. The bottom layer is the perception and acquisition layer, consisting of a data perception and acquisition module and an edge computing gateway, connecting various sensor nodes such as GNSS receivers, rain gauges, and soil moisture sensors to form a hybrid network topology. The upper layer is the data fusion and processing layer, i.e., the intelligent fusion and processing center, responsible for receiving raw data and executing fusion algorithms. The core intelligent analysis and decision-making layer comprises three key modules: a dynamic threshold analysis engine (containing a historical feature database and model library) receives fused features and generates decision boundaries; an intelligent early warning decision module performs risk assessment and early warning based on boundaries and features; and a closed-loop self-optimization module processes feedback information. The top layer is a three-dimensional visualization and decision support platform for the application and presentation layer. The arrows in the diagram clearly illustrate the bidirectional interaction between data uplink and optimization command downlink, reflecting the cloud-edge-device collaboration and the integrated design of "perception-decision-optimization".

[0042] The early warning and system includes the following modules: Data Sensing and Acquisition Module (Edge Side): This module consists of various physical sensors and data transmission units deployed at the landslide site, responsible for the real-time acquisition of multi-dimensional monitoring data. The network adopts a hybrid topology combining star and mesh topologies: 8 high-precision GNSS receivers and 1 rain gauge serve as reliable backbone nodes; 15 soil moisture sensors, 20 deep inclinometer measuring points, and several crack gauges serve as subordinate nodes, flexibly covering key areas through a self-organizing network; a monitoring array composed of 6 microseismic sensors is independently networked, specifically used to capture micro-fracture signals within the soil and rock mass. All data undergoes preliminary time alignment, filtering, and protocol conversion at the site via an edge computing gateway, achieving initial data aggregation. This effectively reduces data transmission load and cloud processing latency, laying the physical foundation for subsequent real-time dynamic early warning.

[0043] Intelligent Fusion and Processing Center (Cloud / Edge): Deployed on cloud servers or high-performance edge computing devices, it adopts a containerized microservice architecture to ensure elasticity and scalability. It receives data streams from the perception layer and has built-in data processing units and fusion algorithm libraries.

[0044] Dynamic Threshold Analysis Engine (Cloud): Deployed in collaboration with the fusion processing center, it includes a historical feature database, an initial model library, and dynamic prediction units.

[0045] Intelligent early warning decision-making module (cloud): It is equipped with a dedicated risk assessment algorithm unit and is usually deployed on a web application server together with a 3D visualization human-computer interaction platform, supporting managers to access and make decisions through multiple terminals.

[0046] Closed-loop self-optimization module (cloud): Acting as the "brain cortex" of the system, it is responsible for learning and optimization.

[0047] In addition, the system also constructs a three-dimensional visualization and decision support platform. This platform is deeply integrated with the geographic information system and can display the spatial distribution of multi-source data, the time series curves of fused features, the evolution of dynamic early warning decision boundaries, the three-dimensional deformation field cloud map of landslide bodies, and the spatial overlay effect of early warning levels in a unified real-time manner. This provides managers with an intuitive panoramic decision-making view, greatly improving the efficiency and scientific nature of emergency command.

[0048] The overall working principle of this invention is as follows: Figure 1 As shown, this is a closed loop of "perception → fusion → prediction → decision → optimization". Its design focuses on breaking down the barriers between different stages in the traditional process to achieve deep coupling. In the multi-source fusion stage, the design focus is not only on aggregating data, but also on dynamically evaluating and trusting high-quality, highly relevant data sources through a hybrid driving mechanism based on information entropy and geomechanical rules. The core is to solve the robustness problem of uneven data quality in sensor networks in complex field environments.

[0049] In the threshold generation stage, the design focuses on overcoming the limitations of static or periodic adjustments, pioneering the principle of "feedforward dynamic adjustment coupled with physical mechanisms." Its core idea is that the warning boundary should not be a simple summary of historical data, but rather an intelligent prediction of the slope's short-term future state. By inputting the stability trend (mechanistic knowledge) derived from the physical model inversion with the statistical characteristics (data patterns) of the data into the time-series prediction model, the threshold adjustment achieves both physical interpretability and data adaptability.

[0050] During the risk assessment phase, the design focuses on replacing simple one-dimensional threshold comparisons with trajectory approximation judgments in a high-dimensional feature space. The "multi-source collaborative instability index" simultaneously quantifies both state anomaly and instability tendency, aiming to more precisely capture the complex precursor patterns of multiple parameters co-evolving before landslide instability.

[0051] At the system operation and maintenance level, the design focuses on endowing the system with closed-loop self-optimization capabilities, enabling it to learn from each early warning practice and automatically adjust internal parameters. The ultimate goal is to solve the "concept drift" problem caused by environmental changes and sensor performance drift in long-term monitoring, reduce the reliance on manual parameter adjustment, and achieve intelligent evolution that becomes more accurate with use.

[0052] The above description is merely a preferred embodiment of the present invention and does not constitute any limitation on the technical scope of the present invention. Therefore, any minor modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention shall still fall within the scope of the technical solution of the present invention.

Claims

1. A landslide intelligent early warning method based on multi-source data fusion and dynamic threshold, characterized in that, The method includes the following steps: S1. Through a multi-source heterogeneous sensor network deployed in the landslide body and surrounding environment, real-time monitoring data including surface displacement, deep displacement, soil moisture content, rainfall and rock and soil vibration signals are collected. S2. The multi-source monitoring data collected in step S1 is cleaned, interpolated, time-synchronized and spatially registered. An adaptive weighted fusion algorithm based on information entropy and predefined geomechanical rules is used to fuse the multi-source data at the feature level to generate a spatiotemporal fusion feature vector that can comprehensively and robustly characterize the overall stability state of the landslide. S3. Based on the spatiotemporal fusion feature vector of the historical long time series, an initial early warning model describing the spatial boundary of the normal state of the slope is constructed using a type of support vector machine; and based on the real-time fusion features within the sliding time window, the slope physical and mechanical state parameters obtained by real-time inversion are fused, and the time series prediction is performed through a recurrent neural network model to dynamically generate an early warning decision boundary that is adaptive to the current environment and the creep state of the slope. S4. Compare the spatiotemporal fusion feature vector generated in real time with the dynamic early warning decision boundary obtained in step S3. Calculate the "multi-source collaborative instability index" in the high-dimensional feature space jointly constructed by historical instability samples and normal samples to conduct a quantitative comprehensive assessment of landslide risk, and issue graded early warning information according to the preset index level. S5. Based on the subsequent verification and accuracy assessment of the early warning results, feedback data containing the comparison information between the early warning results and the actual slope status is generated. The fusion weight allocation strategy in step S2 and the internal parameters of the recurrent neural network model in step S3 are dynamically optimized to achieve adaptive evolution and improvement of the overall early warning performance of the system.

2. The method according to claim 1, characterized in that, The weight coefficients of the adaptive weighted fusion algorithm described in step S2 are dynamically determined. Specifically, they are calculated based on the real-time signal-to-noise ratio and data loss rate of each data source, as well as the gray correlation degree between each data source and the core deformation monitoring sequence in the time domain. This ensures the stability of the system fusion results when some sensors fail or are interfered with.

3. The method according to claim 1, characterized in that, In step S3, the slope physical and mechanical state parameters obtained by real-time inversion are integrated. Specifically, the transient value or rate of change of the slope stability coefficient Fs is calculated in real time based on the simplified infinite slope model or unsaturated soil mechanics model using soil moisture content and rainfall infiltration data within the sliding window. The output of the infinite slope model or unsaturated soil mechanics model is then used as one of the key features and input into the recurrent neural network model.

4. The method according to claim 1, characterized in that, In step S3, the recurrent neural network model adopts a long short-term memory network structure. Its training process introduces a reinforcement learning framework. The model's reward function is designed to be positively correlated with the early warning accuracy and effective early warning duration, and negatively correlated with the false alarm rate, guiding the model to learn the optimal threshold dynamic adjustment strategy.

5. The method according to claim 1, characterized in that, The specific calculation method of the multi-source collaborative instability index in step S4 includes: First, using principal component analysis to reduce the dimensionality of the fused feature vectors of historical normal and unstable states and construct the feature space; second, calculating the Mahalanobis distance from the real-time fused feature vectors to the cluster center of the normal state in the feature space, as well as the shortest projection distance to the known instability evolution path; finally, the multi-source collaborative instability index is a weighted sum of the Mahalanobis distance and the shortest projection distance, used to simultaneously quantify the degree of anomaly of the state and the tendency to evolve towards an unstable state.

6. The method according to claim 1, characterized in that, In step S5, the system automatically records each early warning event and its subsequent verification results to establish an early warning effect sample library. Based on this sample library, the system periodically fine-tunes the parameters of the data fusion weight calculation function and the threshold prediction network using a backpropagation algorithm or an evolutionary algorithm to reduce the probability of misjudgment under similar working conditions in the future.

7. A landslide intelligent early warning system based on multi-source data fusion and dynamic threshold for implementing the method of any one of claims 1 to 6, characterized in that, The system includes: The data sensing and acquisition module consists of various physical sensors and data transmission units, and is responsible for acquiring and remotely transmitting multi-source monitoring data. The intelligent fusion and processing center receives data from the perception layer, has a built-in data processing unit and fusion algorithm library, executes step S2, and outputs a spatiotemporal fusion feature vector. The dynamic threshold analysis engine, connected to the intelligent fusion and processing center, includes a historical feature database, an initial early warning model library, and a dynamic threshold prediction unit. It executes step S3 as described in claim 1 and outputs the dynamic early warning decision boundary. The intelligent early warning decision module is connected to the intelligent fusion and processing center and the analysis engine. It is equipped with a risk assessment unit and executes step S4 to generate and issue graded early warning instructions. The closed-loop self-optimization module is connected to the early warning decision module and feeds back to the intelligent fusion and processing center and the dynamic threshold analysis engine. Step S5 is executed to realize the system's self-learning and performance optimization.

8. The system according to claim 7, characterized in that, The data sensing and acquisition module adopts a hybrid network topology combining star and mesh topologies. In this topology, GNSS receivers and rain gauges serve as backbone nodes, while soil moisture sensors, crack gauges, and other devices serve as subordinate nodes. The microseismic monitoring network is independently networked, and data is initially aggregated and converted into protocols through an edge computing gateway before being uploaded.

9. The system according to claim 7, characterized in that, The intelligent fusion and processing center and the dynamic threshold analysis engine are deployed on cloud servers or high-performance edge computing devices, adopting a containerized microservice architecture. The intelligent early warning decision module and the visual human-computer interaction platform are deployed on web application servers, supporting access from multiple terminals.

10. The system according to claim 7, characterized in that, The system also includes a 3D visualization and decision support platform, which can integrate geographic information systems to display the spatial distribution of multi-source data, fusion feature curves, dynamic early warning decision boundaries, 3D deformation field cloud maps of landslide bodies, and spatial overlay effects of early warning levels in real time, providing managers with a panoramic decision-making view.

Images