Slope stability monitoring method and system based on multi-sensor time series data fusion

Abstract

The application discloses a slope stability monitoring method and system based on multi-sensor time series data fusion, and belongs to the technical field of slope stability monitoring. The method comprises the following steps: multi-source heterogeneous data acquisition and preprocessing: deploying and collecting original time series data from at least three types of sensor nodes, the sensor types including a first type of sensor node for monitoring surface deformation, a second type of sensor node for monitoring internal deformation, and a third type of sensor node for monitoring environmental triggers, and preprocessing the original time series data of each node. Through an adaptive space-time fusion network, the application combines the time series characteristics of multi-source heterogeneous sensors, cross-sensor attention weights, and dynamic spatial dependency, realizes efficient deep fusion of slope monitoring data, significantly improves the comprehensiveness and accuracy of state representation, and overcomes the limitations of low information utilization rate and insufficient space-time correlation mining of traditional single or simple weighted fusion methods.

Description

Technical Field

[0001] This invention relates to the field of slope stability monitoring technology, and more specifically, to a slope stability monitoring method and system based on multi-sensor time-series data fusion. Background Technology

[0002] Slope stability monitoring is a key technical means to prevent geological disasters and ensure the safety of infrastructure. Traditional monitoring methods mostly rely on single or a few types of sensors (such as GNSS, inclinometers, etc.), and rely on threshold judgment or simple statistical analysis for early warning. This makes it difficult to comprehensively and dynamically reflect the complex behavior of slopes under the coupled effects of multiple factors. With the development of sensor technology and the Internet of Things, multi-sensor collaborative monitoring has become a trend. However, how to effectively integrate multi-source heterogeneous time-series data, explore their spatiotemporal correlations, and achieve intelligent, adaptive, and interpretable stability assessment remains a major challenge.

[0003] In existing technologies, multi-sensor data fusion often employs simple methods such as weighted averaging and Kalman filtering, lacking in-depth modeling of temporal dynamics and spatial correlations. Monitoring models are mostly trained statically based on historical data, making it difficult to adapt to the long-term evolution of slope conditions, which can easily lead to delayed early warnings or false alarms. System architectures are often centralized in cloud processing, resulting in poor real-time performance, high bandwidth costs, and a lack of fault tolerance mechanisms for sensor failures. Furthermore, existing methods generally suffer from a "black box" problem, with early warning decisions lacking interpretability, which is detrimental to engineers' trust and emergency decision-making.

[0004] Based on this, the present invention designs a slope stability monitoring method and system based on multi-sensor time-series data fusion to solve the above problems. Summary of the Invention

[0005] The purpose of this invention is to provide a slope stability monitoring method and system based on multi-sensor time-series data fusion, so as to solve the problems mentioned in the background art.

[0006] A slope stability monitoring method based on multi-sensor time-series data fusion includes the following steps: S1. Multi-source heterogeneous data acquisition and preprocessing: Deploy and acquire raw time-series data from at least three types of sensor nodes, including a first type of sensor node for monitoring surface deformation, a second type of sensor node for monitoring internal deformation, and a third type of sensor node for monitoring environmental factors. Preprocess the raw time-series data of each node to form a standardized time-series data stream. S2. Adaptive Spatiotemporal Feature Fusion: The standardized time-series data stream is input into an adaptive spatiotemporal fusion network. The adaptive spatiotemporal fusion network includes a time-series feature extraction module, a cross-sensor attention fusion module, and a dynamic graph convolution module. The time-series feature extraction module extracts time-series features from time-series data of various sensors. The cross-sensor attention fusion module calculates dynamic attention weights between features of different sensors to achieve fusion between sensor features. The dynamic graph convolution module dynamically constructs a sensor relationship graph based on the correlation coefficient between real-time data and captures the spatial dependency features between sensor nodes through graph convolution operations. The adaptive spatiotemporal fusion network outputs a high-order feature vector that fuses the spatiotemporal characteristics of multiple sensors. S3. Online Learning and Model Update: An online learning mechanism is set up, which includes a stability feature drift detector and an incremental learning unit. The stability feature drift detector monitors the statistical distribution changes of the high-order feature vector. When the distribution drift exceeds a preset threshold, the incremental learning unit is activated. The incremental learning unit uses the latest historical data to fine-tune the key parameters of the adaptive spatiotemporal fusion network. S4. Stability Assessment and Explainable Early Warning: The high-order feature vector is input into a stability state evaluator, which outputs a comprehensive slope stability index. At the same time, an interpretability analysis module generates a heat map of the contribution of each sensor feature to the current comprehensive stability index, and sets a dynamic dual-threshold early warning rule based on the trend and rate of change of the comprehensive stability index. S5. Cloud-edge collaborative decision-making and management: When the edge computing device executes steps S2 to S4, the stability comprehensive index, the contribution heatmap, early warning events and feature summaries are synchronized to the cloud management platform. The cloud management platform integrates the data and performs trend analysis and model optimization.

[0007] Preferably, in step S2, when the dynamic graph convolution module constructs the sensor relationship graph, it slides to calculate the Pearson correlation coefficient between the temporal features of any two sensor nodes within a time window, and connects the node pairs whose absolute value of the correlation coefficient is greater than a preset connection threshold, with the weight of the edge being the correlation coefficient.

[0008] Preferably, in step S3, the stability feature drift detector employs a sliding window-based method. The verification method involves using an elastic weight consolidation algorithm for the incremental learning unit.

[0009] Preferably, in step S4, the stability state evaluator is a fully connected neural network, and the dynamic dual-threshold warning rule includes a slow change warning threshold and a fast mutation warning threshold. The slow change warning threshold is dynamically adjusted based on the historical rolling average and standard deviation of the stability composite index.

[0010] Preferably, the method further includes a sensor fault diagnosis and data reconstruction step: real-time monitoring of the data streams of each sensor; when a sensor is determined to be suspected of being faulty, a data reconstruction model is started, and the estimated value of the sensor data is generated using the characteristics of the other normal sensors to replace the abnormal data stream input step S2.

[0011] Preferably, the criteria for sensor fault diagnosis include: the variance of data for multiple consecutive periods is close to zero, the average cross-correlation between the data and other sensors suddenly drops to a low level, or the residual sequence after filtering exceeds the statistical control limit.

[0012] Preferably, in step S5, the communication between the edge computing device and the cloud management platform adopts an adaptive throttling strategy: when the overall stability index is stable, the data synchronization cycle is extended; when an early warning is triggered or model drift is detected, the system switches to a high-frequency synchronization mode.

[0013] Preferably, the method further includes the step of constructing a digital twin slope model on the cloud management platform: using geographic information system data, survey data and monitoring data to drive a slope numerical model; and mapping the real-time acquired comprehensive stability index and the contribution heat map to the digital twin slope model for visualization.

[0014] The slope stability monitoring system, which integrates multi-sensor time-series data, includes an edge sensing and computing subsystem deployed on the slope site, and a cloud-based intelligent management subsystem located remotely. The edge sensing and computing subsystem includes: The sensor array consists of multiple heterogeneous sensor nodes, including at least a first type of sensor node for monitoring surface displacement, a second type of sensor node for monitoring internal deformation, and a third type of sensor node for monitoring environmental parameters. An edge intelligent fusion terminal has a built-in edge computing module; the edge computing module is configured to execute steps S1 to S4. The cloud-based intelligent management subsystem includes: A cloud server cluster is deployed with a cloud-based analysis and management platform, which is configured to execute step S5. User interaction terminal.

[0015] Preferably, the edge computing module in the edge intelligent fusion terminal specifically includes: Preprocessing unit; The adaptive spatiotemporal fusion network unit integrates a temporal feature extraction subunit, a cross-sensor attention fusion subunit, and a dynamic graph convolution subunit. The online learning engine unit includes a stability feature drift detector and an incremental learner; The stability assessment and early warning unit includes a stability state evaluator and an interpretability analysis module; Fault diagnosis and reconfiguration unit.

[0016] Compared with the prior art, the advantages of this invention are: 1. This invention combines the temporal characteristics, cross-sensor attention weights, and dynamic spatial dependencies of multi-source heterogeneous sensors through an adaptive spatiotemporal fusion network, achieving efficient and deep fusion of slope monitoring data. This significantly improves the comprehensiveness and accuracy of state characterization and overcomes the limitations of traditional single or simple weighted fusion methods, which suffer from low information utilization and insufficient spatiotemporal correlation mining.

[0017] 2. By introducing an online learning mechanism and utilizing stability feature drift detection and incremental learning units, this invention can continuously adapt to the dynamic evolution of slope conditions and changes in the external environment, automatically update model parameters, effectively alleviate the performance degradation problem caused by data distribution drift in traditional static models, and improve the stability and adaptability of long-term monitoring.

[0018] 3. This invention, through an interpretability analysis module and dynamic dual-threshold early warning rules, not only outputs a quantitative comprehensive stability index, but also generates a heatmap of the contribution of each sensor feature, making the early warning decision-making process transparent and traceable; combined with dynamic threshold adjustment of trends and rates, it achieves hierarchical and accurate early warning, enhancing the credibility and emergency response capability of the monitoring system.

[0019] 4. This invention completes real-time data fusion and early warning calculation at the edge through a cloud-edge collaborative architecture and adaptive communication strategy, reducing data transmission latency and bandwidth pressure. At the same time, it performs big data integration, model optimization and digital twin construction in the cloud, realizing optimized resource allocation and system scalability, and is suitable for large-scale, distributed slope monitoring scenarios.

[0020] 5. By integrating a sensor fault diagnosis and data reconstruction module, this invention can detect sensor anomalies in real time and reconstruct missing values ​​using normal sensor data, ensuring continuous and reliable operation of the system even when some sensors fail. This improves the robustness and fault tolerance of the monitoring system and reduces the risk of data interruption due to equipment failure. Attached Figure Description

[0021] Figure 1This is a flowchart of the slope stability monitoring method based on multi-sensor time-series data fusion proposed in this invention. Detailed Implementation

[0022] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Please see Figure 1 A slope stability monitoring method based on multi-sensor time-series data fusion includes the following steps: S1. Multi-source heterogeneous data acquisition and preprocessing: Deploy and acquire raw time-series data from at least three types of sensor nodes. The sensor types include a first type of sensor node for monitoring surface deformation, a second type of sensor node for monitoring internal deformation, and a third type of sensor node for monitoring environmental factors. Preprocess the raw time-series data of each node to form a standardized time-series data stream. S2. Adaptive Spatiotemporal Feature Fusion: Standardized time-series data streams are input into an adaptive spatiotemporal fusion network. The adaptive spatiotemporal fusion network includes a time-series feature extraction module, a cross-sensor attention fusion module, and a dynamic graph convolution module. The time-series feature extraction module extracts time-series features from time-series data of various sensors. The cross-sensor attention fusion module calculates dynamic attention weights between features of different sensors to achieve fusion between sensor features. The dynamic graph convolution module dynamically constructs a sensor relationship graph based on the correlation coefficient between real-time data and captures the spatial dependency features between sensor nodes through graph convolution operations. The adaptive spatiotemporal fusion network outputs a high-order feature vector that fuses the spatiotemporal characteristics of multiple sensors. S3. Online Learning and Model Update: An online learning mechanism is set up, which includes a stability feature drift detector and an incremental learning unit. The stability feature drift detector monitors the statistical distribution changes of high-order feature vectors. When the distribution drift exceeds a preset threshold, the incremental learning unit is activated. The incremental learning unit uses the latest historical data to fine-tune the key parameters of the adaptive spatiotemporal fusion network. S4. Stability Assessment and Explainable Early Warning: Input the high-order feature vector into a stability state evaluator. The stability state evaluator outputs a comprehensive slope stability index. At the same time, the interpretability analysis module generates a heat map of the contribution of each sensor feature to the current comprehensive stability index and sets a dynamic dual-threshold early warning rule based on the trend and rate of change of the comprehensive stability index. S5. Cloud-Edge Collaborative Decision-Making and Management: When the edge computing device executes steps S2 to S4, the stability comprehensive index, contribution heatmap, early warning events and feature summaries are synchronized to the cloud management platform. The cloud management platform integrates the data and performs trend analysis and model optimization.

[0024] In step S2, when the dynamic graph convolution module constructs the sensor relationship graph, it slides to calculate the Pearson correlation coefficient between the temporal features of any two sensor nodes within a time window, and connects the node pairs whose absolute value of the correlation coefficient is greater than the preset connection threshold, with the weight of the edge being the correlation coefficient.

[0025] The length of the time window is configurable, typically ranging from 24 hours to 7 days, depending on the slope deformation response time and monitoring requirements. The preset connection threshold ranges from 0.3 to 0.7, preferably 0.5, to filter noise correlation. The constructed sensor relationship graph is a weighted undirected graph, where nodes represent sensors and edges represent the dynamic correlation between sensors. The graph structure updates as the time window slides.

[0026] In step S3, the stability feature drift detector adopts a sliding window-based method. The verification method uses an elastic weight consolidation algorithm for incremental learning units.

[0027] The length of the sliding window is 7 to 30 days, preferably 14 days; The significance level of the test was set to 0.05; the elastic weight consolidation algorithm applies regularization constraints based on the Fisher information matrix to the key network parameters during fine-tuning to mitigate catastrophic forgetting; the incremental learning unit is triggered only when a significant distribution drift is detected, and the amount of data used for each fine-tuning is the monitoring data from the most recent 1 to 3 months.

[0028] In step S4, the stability state evaluator is a fully connected neural network, and the dynamic dual-threshold warning rule includes a slow change warning threshold and a fast mutation warning threshold. The slow change warning threshold is dynamically adjusted based on the historical rolling average and standard deviation of the stability composite index.

[0029] The fully connected neural network consists of one input layer, two hidden layers, and one output layer. The number of neurons in the hidden layers are 64 and 32, respectively, and the activation function is ReLU. The output layer uses a linear activation function. The calculation window length for the historical rolling average and standard deviation is 30 days. The warning threshold for slow change is the rolling average plus or minus 2 standard deviations, and the warning threshold for rapid mutation is when the change in the stability composite index between two adjacent time steps exceeds 3 times the standard deviation of its historical change.

[0030] The method also includes sensor fault diagnosis and data reconstruction steps: real-time monitoring of the data stream of each sensor; when a sensor is identified as a suspected fault, a data reconstruction model is started, and the estimated value of the sensor data is generated using the characteristics of the other normal sensors to replace the abnormal data stream input step S2.

[0031] The data reconstruction model is a multivariate time-series prediction model based on an attention mechanism. The input is the features of other normal sensors at the current and historical time steps, and the output is the estimated value of the faulty sensor. The reconstruction model is pre-trained on historical normal data and deployed as a lightweight version in edge computing devices.

[0032] Criteria for sensor fault diagnosis include: the variance of data for several consecutive periods is close to zero, the average cross-correlation between the data and other sensors suddenly drops to a low level, or the residual sequence after filtering exceeds the statistical control limit.

[0033] Multiple consecutive periods typically refer to more than 6 consecutive sampling periods; variance close to zero means the variance is less than 0.1% of the sensor's range; the average cross-correlation suddenly drops to below 50% of the historical average; statistical control limits are set based on the 3σ principle of historical residuals.

[0034] In step S5, the communication between the edge computing device and the cloud management platform adopts an adaptive throttling strategy: when the overall stability index is stable, the data synchronization cycle is extended; when an early warning is triggered or model drift is detected, the system switches to a high-frequency synchronization mode.

[0035] Under stable conditions, the data synchronization cycle is 1 to 24 hours, preferably 4 hours; under high-frequency synchronization mode, the cycle is shortened to 1 to 10 minutes, preferably 5 minutes; in addition to the stability comprehensive index, contribution heat map, and early warning events, the synchronization content also includes abnormal sensor identifiers, reconstructed data markers, and real-time feature vector summaries.

[0036] The method also includes the step of building a digital twin slope model on a cloud management platform: using geographic information system data, survey data and monitoring data to drive a slope numerical model; and mapping the real-time acquired stability comprehensive index and contribution heat map to the digital twin slope model for visualization.

[0037] The digital twin slope model is constructed based on the finite element or discrete element method, which can reflect the slope geometry, soil and rock parameters and stress-strain relationship; real-time monitoring data updates model parameters through data assimilation technology; the visualization display supports 3D rendering, deformation animation, highlighting of early warning areas and overlay of contribution heatmaps, and supports multi-terminal access and interactive operation.

[0038] The slope stability monitoring system, which integrates multi-sensor time-series data, includes an edge sensing and computing subsystem deployed on the slope site, and a cloud-based intelligent management subsystem located remotely. The edge sensing and computing subsystem includes: The sensor array consists of multiple heterogeneous sensor nodes, including at least a first type of sensor node for monitoring surface displacement, a second type of sensor node for monitoring internal deformation, and a third type of sensor node for monitoring environmental parameters. The edge intelligent fusion terminal has a built-in edge computing module; the edge computing module is configured to execute steps S1 to S4. The cloud-based intelligent management subsystem includes: A cloud server cluster is deployed with a cloud-based analytics and management platform, which is configured to execute step S5. User interaction terminal; sensor nodes include, but are not limited to: GNSS receiver, crack gauge, inclinometer, earth pressure cell, piezometer, rain gauge, temperature and humidity sensor; edge intelligent fusion terminal adopts industrial-grade embedded device, has IP67 protection level, and supports 4G / 5G and LoRa communication; cloud analysis and management platform provides data warehouse, model management, early warning distribution and visualization screen functions; user interaction terminal includes web terminal and mobile app, supporting real-time monitoring, historical query, early warning reception and report export.

[0039] The edge computing module in the edge intelligent fusion terminal specifically includes: Preprocessing unit; The adaptive spatiotemporal fusion network unit integrates a temporal feature extraction subunit, a cross-sensor attention fusion subunit, and a dynamic graph convolution subunit. The online learning engine unit includes a stability feature drift detector and an incremental learner; The stability assessment and early warning unit includes a stability state evaluator and an interpretability analysis module; The system comprises several sub-units: a fault diagnosis and reconstruction unit; a preprocessing unit responsible for data cleaning, interpolation, normalization, and time alignment; a temporal feature extraction sub-unit employing a one-dimensional convolutional neural network or a long short-term memory network; a cross-sensor attention fusion sub-unit using a multi-head self-attention mechanism; a dynamic graph convolution sub-unit supporting real-time graph construction and graph convolution operations; an interpretability analysis module generating contribution heatmaps based on gradient-weighted class activation mapping; and a fault diagnosis and reconstruction unit integrating fault detection logic and a lightweight reconstruction model, supporting real-time switching and data replacement.

[0040] The workflow of this invention: I. Data Acquisition and Preprocessing Stage Deploy multiple types of sensors: Deploy at least three types of sensors on the slope site to monitor surface deformation (such as GNSS, crack gauges), internal deformation (such as inclinometers, earth pressure cells), and environmental factors (such as rain gauges, temperature and humidity sensors).

[0041] Data standardization processing: The collected raw time series data undergoes preprocessing operations such as cleaning, interpolation, and normalization to form a standardized time series data stream with a unified format for subsequent fusion and analysis.

[0042] II. Adaptive Spatiotemporal Feature Fusion Stage Temporal feature extraction: Using one-dimensional convolutional neural networks or long short-term memory networks, representative temporal features are extracted from temporal data of various sensors.

[0043] Cross-sensor attention fusion: The weights between features from different sensors are dynamically calculated through a multi-head self-attention mechanism to achieve adaptive fusion of multi-source information.

[0044] Dynamic graph convolution fusion: Based on the Pearson correlation coefficient between sensor features within a sliding time window, a sensor relationship graph is dynamically constructed, and the spatial dependencies between sensor nodes are captured through graph convolution operations.

[0045] Output high-order feature vector: Finally, a high-order feature vector that integrates temporal and spatial features is generated to comprehensively represent the slope state.

[0046] III. Online Learning and Model Update Phase Feature drift detection: The statistical distribution changes of high-order feature vectors are continuously monitored through the Kolmogorov-Smirnov test based on a sliding window.

[0047] Incremental learning optimization: Once the distribution drift is detected to exceed the threshold, the elastic weight consolidation algorithm is triggered to fine-tune the key parameters of the fusion network using the latest historical data to avoid model degradation.

[0048] IV. Stability Assessment and Explainable Early Warning Phase Stability comprehensive index calculation: Input the high-order feature vector into a fully connected neural network, and output a quantified slope stability comprehensive index.

[0049] Contribution visualization: By using the gradient-weighted class activation mapping method, a heatmap of the contribution of each sensor feature to the current stability index is generated, improving the interpretability of decision-making.

[0050] Dynamic dual-threshold early warning: Based on the historical rolling mean and standard deviation of the stability index, dual thresholds for slow change and rapid mutation are set to achieve graded early warning.

[0051] V. Cloud-Edge Collaborative Decision-Making and Management Phase Edge computing execution: Feature fusion, evaluation and early warning tasks are performed in real time at the edge intelligent terminal on the slope site, reducing data transmission latency and bandwidth pressure.

[0052] Cloud data synchronization and optimization: Edge terminals synchronize key data (such as stability index, early warning events, and contribution heatmaps) to the cloud platform for long-term trend analysis, model optimization, and digital twin construction.

[0053] Adaptive communication strategy: Extend the synchronization period in a stable state, and switch to high-frequency synchronization when there is an early warning or model drift, so as to realize intelligent scheduling of communication resources.

[0054] VI. Fault Diagnosis and Data Reconstruction Real-time monitoring of sensor status: The sensor is detected in real time for faults by using criteria such as variance, cross-correlation, and residual control limits.

[0055] Data reconstruction compensation: Once a faulty sensor is detected, a lightweight reconstruction model based on an attention mechanism is activated to estimate the missing value using data from other normal sensors, ensuring continuous system operation.

[0056] VII. Digital Twins and Visualization Constructing a digital twin of a slope: Integrating geographic information, survey data, and real-time monitoring data to establish a dynamically updatable numerical model of the slope.

[0057] 3D visualization and interaction: Mapping stability indices, heatmaps, etc., into the twin model, supporting 3D rendering, animation display and multi-terminal access, to assist in decision-making and emergency command.

[0058] In summary, this invention achieves intelligent, real-time, and reliable monitoring of slope stability through a closed-loop process involving multi-sensor heterogeneous data acquisition, spatiotemporal adaptive fusion, online learning and updating, interpretable assessment and early warning, and cloud-edge collaborative management. The system possesses adaptability, interpretability, robustness, and scalability, making it suitable for long-term health monitoring and risk early warning in various slope engineering projects.

[0059] 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 preferred examples and are not intended to limit 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 the present invention is defined by the appended claims and their equivalents.

Claims

1. A slope stability monitoring method based on multi-sensor time-series data fusion, characterized in that, Includes the following steps: S1. Multi-source heterogeneous data acquisition and preprocessing: Deploy and acquire raw time-series data from at least three types of sensor nodes, including a first type of sensor node for monitoring surface deformation, a second type of sensor node for monitoring internal deformation, and a third type of sensor node for monitoring environmental factors. Preprocess the raw time-series data of each node to form a standardized time-series data stream. S2. Adaptive Spatiotemporal Feature Fusion: The standardized time-series data stream is input into an adaptive spatiotemporal fusion network. The adaptive spatiotemporal fusion network includes a time-series feature extraction module, a cross-sensor attention fusion module, and a dynamic graph convolution module. The time-series feature extraction module extracts time-series features from time-series data of various sensors. The cross-sensor attention fusion module calculates dynamic attention weights between features of different sensors to achieve fusion between sensor features. The dynamic graph convolution module dynamically constructs a sensor relationship graph based on the correlation coefficient between real-time data and captures the spatial dependency features between sensor nodes through graph convolution operations. The adaptive spatiotemporal fusion network outputs a high-order feature vector that fuses the spatiotemporal characteristics of multiple sensors. S3. Online Learning and Model Update: An online learning mechanism is set up, which includes a stability feature drift detector and an incremental learning unit. The stability feature drift detector monitors the statistical distribution changes of the high-order feature vector. When the distribution drift exceeds a preset threshold, the incremental learning unit is activated. The incremental learning unit uses the latest historical data to fine-tune the key parameters of the adaptive spatiotemporal fusion network. S4. Stability Assessment and Explainable Early Warning: The high-order feature vector is input into a stability state evaluator, which outputs a comprehensive slope stability index. At the same time, an interpretability analysis module generates a heat map of the contribution of each sensor feature to the current comprehensive stability index, and sets a dynamic dual-threshold early warning rule based on the trend and rate of change of the comprehensive stability index. S5. Cloud-edge collaborative decision-making and management: When the edge computing device executes steps S2 to S4, the stability comprehensive index, the contribution heatmap, early warning events and feature summaries are synchronized to the cloud management platform. The cloud management platform integrates the data and performs trend analysis and model optimization.

2. The slope stability monitoring method based on multi-sensor time-series data fusion according to claim 1, characterized in that, In step S2, when the dynamic graph convolution module constructs the sensor relationship graph, it slides to calculate the Pearson correlation coefficient between the temporal features of any two sensor nodes within a time window, and connects the node pairs whose absolute value of the correlation coefficient is greater than a preset connection threshold, with the weight of the edge being the correlation coefficient.

3. The slope stability monitoring method based on multi-sensor time-series data fusion according to claim 1, characterized in that, In step S3, the stability feature drift detector employs a sliding window-based Kolmogorov algorithm. The Smirnov test method is used, and the incremental learning unit employs the elastic weight consolidation algorithm.

4. The slope stability monitoring method based on multi-sensor time-series data fusion according to claim 1, characterized in that, In step S4, the stability state evaluator is a fully connected neural network, and the dynamic dual-threshold warning rule includes a slow change warning threshold and a fast mutation warning threshold. The slow change warning threshold is dynamically adjusted based on the historical rolling average and standard deviation of the stability composite index.

5. The slope stability monitoring method based on multi-sensor time-series data fusion according to claim 1, characterized in that, The method also includes a sensor fault diagnosis and data reconstruction step: real-time monitoring of the data streams of each sensor; when a sensor is determined to be suspected of being faulty, a data reconstruction model is started, and the estimated value of the sensor data is generated using the characteristics of the other normal sensors to replace the abnormal data stream input step S2.

6. The slope stability monitoring method based on multi-sensor time-series data fusion according to claim 5, characterized in that, The criteria for sensor fault diagnosis include: the variance of data for several consecutive periods is close to zero, the average cross-correlation between the data and other sensors suddenly drops to a low level, or the residual sequence after filtering exceeds the statistical control limit.

7. The slope stability monitoring method based on multi-sensor time-series data fusion according to claim 1, characterized in that, In step S5, the communication between the edge computing device and the cloud management platform adopts an adaptive throttling strategy: when the overall stability index is stable, the data synchronization cycle is extended; when an early warning is triggered or model drift is detected, the system switches to a high-frequency synchronization mode.

8. The slope stability monitoring method based on multi-sensor time-series data fusion according to claim 1, characterized in that, The method also includes the step of constructing a digital twin slope model on the cloud management platform: using geographic information system data, survey data and monitoring data to drive a slope numerical model; and mapping the real-time acquired comprehensive stability index and the contribution heat map to the digital twin slope model for visualization.

9. A slope stability monitoring system based on multi-sensor time-series data fusion, used to execute the slope stability monitoring method based on multi-sensor time-series data fusion as described in any one of claims 1-8, characterized in that, This includes an edge sensing and computing subsystem deployed on the slope site, and a cloud-based intelligent management subsystem located remotely; The edge sensing and computing subsystem includes: The sensor array consists of multiple heterogeneous sensor nodes, including at least a first type of sensor node for monitoring surface displacement, a second type of sensor node for monitoring internal deformation, and a third type of sensor node for monitoring environmental parameters. An edge intelligent fusion terminal has a built-in edge computing module; the edge computing module is configured to execute steps S1 to S4. The cloud-based intelligent management subsystem includes: A cloud server cluster is deployed with a cloud-based analysis and management platform, which is configured to execute step S5. User interaction terminal.

10. The slope stability monitoring system based on multi-sensor time-series data fusion according to claim 9, characterized in that, The edge computing module in the edge intelligent fusion terminal specifically includes: Preprocessing unit; The adaptive spatiotemporal fusion network unit integrates a temporal feature extraction subunit, a cross-sensor attention fusion subunit, and a dynamic graph convolution subunit. The online learning engine unit includes a stability feature drift detector and an incremental learner; The stability assessment and early warning unit includes a stability state evaluator and an interpretability analysis module; Fault diagnosis and reconfiguration unit.

Images