Road slope risk analysis method and system based on multi-source data fusion

Abstract

The invention discloses a road slope risk analysis method and system based on multi-source data fusion, belongs to the technical field of road risk analysis, and aims to solve the problems of one-sided data support and insufficient fusion processing dynamics of traditional slope risk analysis. The system firstly collects geological parameters, meteorological data, deformation information and video images to form a multi-source heterogeneous original data set, then constructs a three-level fusion architecture of an original data layer, a feature representation layer and a risk decision-making layer through space-time alignment, feature extraction and semantic fusion, and then introduces a multi-modal feature weighting strategy of an attention mechanism to the feature representation layer. In a risk decision-making layer, a graph neural network is utilized to model a slope structure unit topology dependency relationship, and finally, based on a fused high-dimensional feature vector, an instability probability is calculated through a dynamic Bayesian network, and network parameters are updated online in combination with real-time data, so that the data fusion effectiveness is improved, the evaluation accuracy and dynamics are guaranteed, and the risk decision-making efficiency is improved. And a scientific decision basis is provided for slope risk prevention and control.

Description

Technical Field

[0001] This invention belongs to the field of road risk analysis technology, specifically a road slope risk analysis method and system based on multi-source data fusion. Background Technology

[0002] As a crucial component of highway engineering, the stability of road slopes directly impacts road traffic safety, the operational lifespan of engineering projects, and the safety of the surrounding ecological environment. Especially in mountainous and hilly areas with complex geological structures, steep, mixed soil-rock slopes and rock slopes are prone to landslides and collapses due to factors such as rainfall, geological disturbances, and topographic stress. These collapses not only cause direct losses such as road interruptions and facility damage but can also trigger secondary disasters, threatening the safety of people and property. With the expansion of my country's transportation infrastructure construction into remote and complex geological areas, the number and structural complexity of road slopes have increased significantly. This has placed higher demands on the industry for precise monitoring and scientific analysis of slope instability risks. Traditional single-indicator monitoring methods are no longer sufficient to meet engineering needs. The industry is gradually adopting a trend of integrating geological, meteorological, and deformation data to conduct slope risk analysis.

[0003] While current technologies for road slope risk analysis have begun to explore the collection and application of multi-source data, the overall technical system still has many imperfections, resulting in unsatisfactory practical effects in slope instability risk analysis. Existing analytical methods generally suffer from incomplete data support; some methods still rely on single-type monitoring data for risk assessment. Even when integrating multi-source heterogeneous data, the lack of a systematic fusion processing architecture prevents effective correction of the spatiotemporal differences in multi-source data, leading to insufficient logic and effectiveness in feature extraction and fusion. Furthermore, the utilization of multi-source data features lacks specificity, failing to differentiate based on the contribution of features to slope risk, and neglecting the structural characteristics of the slope itself. It fails to explore the topological dependencies such as physical connections and spatial proximity between different structural units of the slope, making it difficult to conduct analysis that closely reflects the actual slope condition. In addition, existing risk assessments are mostly qualitative, lacking quantitative probability calculation models, and existing analytical models are mostly static and fixed, unable to dynamically update model parameters based on real-time monitoring data, making it difficult to adapt to real-time changes in slope condition and the external environment. The aforementioned problems combine to result in significant deficiencies in the comprehensiveness, accuracy, and dynamism of traditional road slope risk analysis methods. This makes it difficult to accurately assess slope instability risks and to provide reliable technical support for early warning and emergency response to slope disasters. Summary of the Invention

[0004] The purpose of this invention is to provide a road slope risk analysis method and system based on multi-source data fusion, so as to solve the following technical problems mentioned in the background art: Traditional methods for analyzing road slope risks suffer from incomplete data support, a lack of systematic multi-source data fusion and processing, and insufficient targeted use of features. Furthermore, risk assessments are mostly qualitative analyses, and the models cannot adapt to real-time changes in slope conditions. As a result, the comprehensiveness, accuracy, and dynamism of slope instability risk analysis are insufficient, making it difficult to accurately assess slope risks.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A method for road slope risk analysis based on multi-source data fusion includes the following steps: S1 collects geological parameters, meteorological data, deformation information and video images to form a multi-source heterogeneous raw dataset; S2 addresses the dimensional differences of multi-source heterogeneous original datasets by sequentially performing spatiotemporal alignment, feature extraction, and semantic fusion on the original datasets. Based on the processed data, a three-level fusion architecture is constructed, consisting of the original data layer, the feature representation layer, and the risk decision layer. S3, based on a three-level fusion architecture, introduces a multimodal feature weighting strategy based on an attention mechanism in the feature representation layer, and uses a graph neural network to model the topological dependencies between slope structural units in the risk decision layer; S4, based on high-dimensional feature vectors fused after being weighted by attention mechanism and modeled by topological dependencies, calculates the slope instability probability through a dynamic Bayesian network, and updates the network parameters online by combining newly collected multi-source heterogeneous real-time data.

[0006] Furthermore, geological parameters are collected through geological sensors, meteorological data is collected through meteorological stations, deformation information is collected through deformation monitoring equipment, and video images are collected through video monitoring units. Among them, geological sensors include pore water pressure gauges, earth pressure cells, and inclinometers; meteorological data collected by meteorological stations include rainfall, temperature, humidity, and wind speed; deformation monitoring equipment includes BeiDou / GNSS positioning modules and fiber optic strain gauges; and the video monitoring unit is equipped with a dual-mode camera for visible light and infrared.

[0007] Furthermore, spatiotemporal alignment employs a joint correction method based on timestamp interpolation and geographic coordinate projection to unify multi-source heterogeneous data into the same spatiotemporal grid; feature extraction performs sliding window statistics on geological and meteorological data to generate mean, variance, and rate of change features respectively; displacement gradient and rate of change features are calculated for deformation data; and texture, edge, and motion vector features are extracted from video images through a convolutional neural network.

[0008] Furthermore, the multimodal feature weighting strategy dynamically allocates weight coefficients by calculating the contribution of each modal feature to historical risk events. The weight update formula is as follows: in, Let be the weight of the i-th mode at time t. Its feature significance score, To adjust the parameters, M is the total number of modes. This is the sum of the exponential scores of all modes at time t.

[0009] Furthermore, the graph neural network divides the slope profile into several structural units as nodes, and uses the physical connections or spatial proximity relationships between units as edges. The node feature vector is composed of multimodal features weighted and fused by the attention mechanism. Neighborhood information is aggregated through two layers of graph convolution operations, and the local risk embedding representation of each unit is output.

[0010] Furthermore, the dynamic Bayesian network includes state nodes and observation nodes, where state nodes are used to characterize the stability state of the slope, and observation nodes correspond to high-dimensional feature vectors fused after being weighted by the attention mechanism and modeled by topological dependencies. The conditional probability table is initialized by the maximum likelihood estimation method, and the network parameters are iteratively updated based on the newly acquired multi-source heterogeneous real-time observation data using the online expectation-maximization algorithm.

[0011] Furthermore, the formula for calculating the probability of slope instability is: In the formula, Let be the state variable at time t. Given the observation sequence from the initial time to the current time, the posterior probability... The forward-backward algorithm is used to recursively solve the problem, calculating the local instability probability of each structural unit of the slope and the overall comprehensive instability probability of the slope. The calculation accuracy is retained to four decimal places. The newly acquired observation data at the current moment, Let be the state variable at time t; This refers to historical observation data up to the previous moment. Let be the state variable at time t−1; For the observation likelihood probability, let represent the probability in the current state. Below, observed data The probability of; The one-step prediction probability means predicting the state at the current moment based on the observation data up to the previous moment. The probability of; The normalization constant is obtained by considering all possible states from the previous time step. Summation ensures that the sum of the posterior probabilities is 1.

[0012] A road slope risk analysis system based on multi-source data fusion, used to implement the method described in the preceding part, includes an acquisition module, a data fusion module, a multimodal feature fusion and topology-aware risk decision-making module, and an inference module; The acquisition module is used to collect geological parameters, meteorological data, deformation information and video images to form a multi-source heterogeneous raw dataset; The data fusion module is used to sequentially perform spatiotemporal alignment, feature extraction, and semantic fusion on the original dataset, constructing a three-level fusion architecture consisting of the original data layer, the feature representation layer, and the risk decision layer; The multimodal feature fusion and topology-aware risk decision module is used to introduce an attention-based multimodal feature weighting strategy in the feature representation layer and to use a graph neural network to model the topological dependencies between slope structural units in the risk decision layer. The inference module is used to calculate the slope instability probability through a dynamic Bayesian network based on the fused high-dimensional feature vector, and to update the network parameters online by combining the newly collected data.

[0013] A road slope risk analysis device based on multi-source data fusion includes a memory and a processor. The memory stores a computer program, and the processor executes the program to implement the road slope risk analysis method based on multi-source data fusion described above.

[0014] Compared with the prior art, the present invention has the following beneficial effects: This invention integrates multi-source heterogeneous data, overcoming the limitations of single-source data in representing slope conditions, and providing a comprehensive and rich data foundation for risk analysis.

[0015] This invention constructs a three-level fusion architecture, which eliminates data dimensional differences through spatiotemporal alignment, feature extraction, and semantic fusion, thereby achieving hierarchical and systematic processing of multi-source data and improving the effectiveness of data fusion.

[0016] This invention combines attention mechanisms with graph neural networks to dynamically strengthen the weights of key features, while also mining the topological dependencies of slope structural units, making risk analysis more closely aligned with the actual structural characteristics of slopes and improving the accuracy of feature utilization and analysis.

[0017] This invention utilizes a dynamic Bayesian network to quantitatively calculate the probability of slope instability and supports online updates of network parameters, allowing the model to adapt to real-time changes in slope status and ensuring the dynamism and accuracy of risk assessment. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall process of the present invention; Figure 2 This is a schematic diagram of the three-level converged architecture of the present invention; Figure 3 This is a schematic diagram of the feature weighting and topology modeling process of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0020] Example 1: A method for road slope risk analysis based on multi-source data fusion, such as Figure 1 As shown, it includes the following steps: S1, based on the slope geological characteristics, topographic distribution and key risk prevention and control points, distributed deployment of various monitoring and acquisition devices, directional collection of geological parameters, meteorological data, deformation information and video images, and basic preprocessing of the collected raw data including outlier removal, format standardization and timestamp calibration, to ensure data validity and consistency, forming a multi-source heterogeneous raw dataset covering four dimensions: geology, meteorology, deformation and vision.

[0021] Geological parameters are collected through geological sensors, meteorological data through weather stations, deformation information through deformation monitoring equipment, and video images through video monitoring units. Geological sensors, including pore water pressure gauges, earth pressure cells, and inclinometers, are deployed at key stress-changing areas such as the slope toe, slope waist, and slope top, with a sampling frequency uniformly set to 1 minute / time, and data transmitted via RS-485 bus. Rainfall, temperature, humidity, and wind speed data collected by weather stations are precisely matched to the slope monitoring area, with a time resolution of 5 minutes / time, and are uploaded via LoRa wireless protocol. Deformation monitoring equipment includes a BeiDou / GNSS positioning module and a fiber optic strain gauge. The BeiDou / GNSS module enables overall slope displacement monitoring, while the fiber optic strain gauge enables local micro-strain monitoring, with sampling frequencies of 10 minutes / time and 10Hz, respectively. The video monitoring unit is equipped with dual-mode visible light and infrared cameras, deployed at high points around the slope to achieve full coverage without blind spots, with a frame rate of 25fps. Key frames are uploaded only when the detected image change rate exceeds a threshold, reducing data transmission costs. All collected data are uniformly encoded into a standardized format, and the timestamp accuracy is calibrated to the UTC millisecond level. At the same time, the collection data with obvious abnormalities (such as extreme data caused by sensor failure) are filtered out by the threshold method to form a multi-source heterogeneous raw dataset that can be directly used for subsequent processing.

[0022] It should be noted that the selection of geological, meteorological, deformation, and video data is based on engineering considerations to establish a complete chain of evidence for slope instability, encompassing internal causes (geological structure and stress), external causes (meteorological loads), response (deformation), and manifestations (macroscopic deformation and failure). Ignoring any link (such as using only deformation) or including irrelevant data (such as traffic flow) will undermine the logical foundation and physical meaning of the subsequent integrated analysis.

[0023] S2 addresses the differences in data dimension, sampling frequency, and units of measurement among multi-source heterogeneous original datasets. It sequentially performs spatiotemporal alignment, feature extraction, and semantic fusion operations on the original datasets to eliminate spatiotemporal deviations and the influence of units of measurement between data. Based on the processed data, it constructs a three-level fusion architecture with clear hierarchy and defined functions. This architecture progresses from the original data layer and feature representation layer to the risk decision layer, realizing the entire process of data processing from original acquisition to feature fusion and then to risk decision.

[0024] The three-tiered fusion architecture is constructed as follows: the raw data layer receives and stores standardized preprocessed data from S1; the feature representation layer retrieves data from the raw data layer, sequentially performing spatiotemporal alignment, feature extraction, and semantic fusion operations, outputting a unified multimodal feature vector; the risk decision layer receives the feature vector output by the feature representation layer, performs feature weighting and topological modeling, preparing for final probabilistic inference. Data is transferred between the three layers through clearly defined interface protocols, ensuring the modularity and scalability of the processing flow.

[0025] The spatiotemporal alignment employs a joint correction method based on timestamp interpolation and geographic coordinate projection: the time dimension uses a 10-minute basic time step, performs linear interpolation or nearest neighbor filling on data with sampling frequencies higher than this step, and performs forward hold on data with sampling frequencies lower than this step, thus achieving time synchronization of all data; the spatial dimension projects the WGS-84 coordinates of all monitoring points to the UTM coordinate system, and then maps them to a regular 10m×10m spatiotemporal grid. If multiple sensors of the same type exist in a grid, their arithmetic mean is taken as the representative value of the grid cell, ultimately unifying multi-source heterogeneous data into the same spatiotemporal grid. Feature extraction employs differentiated processing methods for the feature attributes of different data types: For geological and meteorological data, sliding window statistics are performed (window length set to 6 time steps, i.e., 60 minutes, step size set to 1 time step), generating mean, variance, and rate of change features respectively; for deformation data, displacement gradient (displacement difference between adjacent grid cells divided by distance) and velocity accumulation (integral of displacement velocity over the past 60 minutes) features are calculated; for video images, texture, edge, and motion vector features are extracted using a lightweight convolutional neural network (MobileNetV3). All extracted features are normalized to eliminate dimensional differences. Semantic fusion concatenates the normalized feature vectors to form a unified multimodal feature representation, completing the feature layer fusion process. The functions of each layer in the three-level fusion architecture are as follows: the raw data layer implements the classification, unified retrieval, and basic management of multi-source heterogeneous raw data; the feature representation layer receives the output of the raw data layer, completing feature extraction, normalization, and semantic fusion, providing a foundation for subsequent feature weighting; the risk decision layer, as the core layer of the fusion architecture, realizes slope topological relationship modeling, instability probability calculation, and risk level determination. Figure 2 As shown.

[0026] The key to this invention lies in recognizing that for spatial geographic objects like slopes, a unified spatiotemporal grid coordinate system must first be established, and all subsequent feature extraction and fusion must be performed on this grid. This spatiotemporal alignment is an absolute prerequisite for subsequent spatial topology modeling by the GNN. Simply stitching features together at the feature level cannot provide the GNN with inputs that have clear spatial correspondences, and therefore cannot achieve true spatial dependency modeling.

[0027] S3, based on the constructed three-level fusion architecture, introduces a multimodal feature weighting strategy based on attention mechanism in the feature representation layer. It dynamically allocates weights according to the contribution of each modal feature to the slope risk analysis, highlighting the role of highly correlated features. In the risk decision layer, graph neural networks are used to model the topological dependencies between slope structural units, explore the physical and spatial connections between different areas of the slope, and realize the local and overall collaborative assessment of slope risk.

[0028] The steps for modeling the topological dependencies between slope structural units using graph neural networks include: Node division: Based on the slope geological survey report and monitoring grid, the slope is divided into several "structural units" with engineering significance, and each unit corresponds to a graph node.

[0029] Edge construction: Based on the geological continuity (e.g., belonging to the same rock stratum), physical connectivity (e.g., connected by support structures), or spatial proximity (e.g., distance less than a set threshold) between units, edges are established between nodes to form a slope topology map.

[0030] Feature aggregation: The initial features of each node are the weighted fused feature vectors obtained in S3. Through the message passing mechanism of the graph convolutional layer (GCN), each node aggregates the feature information of its neighboring nodes. After multiple layers of propagation, the final feature representation of each node (i.e., the "local risk embedding representation") not only contains its own state, but also contains its local topological structure information.

[0031] The multimodal feature weighting strategy dynamically assigns weight coefficients by calculating the contribution of each modality feature to historical risk events. The weight update formula is as follows: In the formula, Let be the weight of the i-th mode at time t. Its feature significance score, To adjust the parameters, M represents the total number of modes (in this embodiment, M=4, corresponding to four modes: geological, meteorological, deformation, and video). This is the sum of the exponential scores of all modes at time t. This formula enables dynamic updates of the weights, giving higher weights to features that have a greater impact on slope instability.

[0032] The topological dependency modeling of graph neural networks is as follows: First, the slope profile is divided into several structural units according to the geological stratification, topographic features, and spatiotemporal grid division rules of the slope, and these units serve as nodes of the graph neural network. The node feature vectors are composed of multimodal features weighted and fused by the attention mechanism. Then, the physical connection relationship between units (such as the same rock layer or the same soil structure) or spatial proximity relationship (spatial distance less than 20 meters) is used as the edge of the graph neural network to construct the topological graph of the slope structure. Finally, neighborhood information is aggregated through two layers of graph convolution operations. The first layer aggregates the feature information of the node's direct neighbors, and the second layer aggregates the feature information of the node's second-order neighbors. The number of convolution kernels is set to 64, and the mean aggregation method is used to achieve neighborhood information fusion. Finally, a 64-dimensional local risk embedding representation of each unit is output, which contains the state characteristics of the unit itself and the topological association characteristics of its surrounding area.

[0033] It's important to note that the attention mechanism is placed in the feature representation layer, while the graph neural network (GNN) is placed in the risk decision layer. This placement is based on functional isolation: the attention mechanism is responsible for addressing the issue of feature importance, which is a global evaluation relatively independent of specific spatial locations; the GNN, on the other hand, is responsible for addressing the issue of spatial information propagation. If the order is reversed, or if they are used interchangeably in the same layer, the model will confuse the two different issues of which features are important and how important features propagate, thereby reducing the model's performance and interpretability.

[0034] S4 takes the high-dimensional feature vector (i.e. the local risk embedding representation output by the graph neural network) fused after being weighted by the attention mechanism and modeled by topological dependencies as input, calculates the instability probability of each structural unit and the whole slope through a dynamic Bayesian network, and updates the parameters of the dynamic Bayesian network online at a preset frequency in combination with newly collected multi-source heterogeneous real-time data to achieve adaptive optimization of the model and ensure the dynamism and accuracy of slope risk analysis.

[0035] The computation and update process of dynamic Bayesian networks includes: Network construction: Define state variables (such as slope stability state) and observation variables (i.e., high-dimensional feature vectors output by S3), and determine the network structure (dependence between state transition probabilities and observation probabilities) based on engineering experience or data learning.

[0036] Parameter initialization: The conditional probability table (CPT) of the network is initialized using the historical case database and methods such as maximum likelihood estimation.

[0037] Online inference: When new observation data (real-time feature vectors) arrive, the forward-backward algorithm is used to recursively update the posterior probability distribution of the state variables, i.e., the slope instability probability.

[0038] Online parameter updates: The Online Expectation-Maximization (Online EM) algorithm is employed, using new observation sequences as incremental data to iteratively optimize the network's CPT parameters, enabling the model to adapt to changes in slope condition and environment. The update process can be set to a fixed period or triggered based on data freshness.

[0039] The dynamic Bayesian network comprises state nodes and observation nodes. State nodes characterize the slope stability state, categorized into four levels: stable, slightly unstable, moderately unstable, and severely unstable. Observation nodes correspond to high-dimensional feature vectors fused after weighting via an attention mechanism and modeling topological dependencies. The conditional probability table of the dynamic Bayesian network is initialized using maximum likelihood estimation. The initial sample set consists of over 5000 labeled sets of full lifecycle data on slope disaster events, ensuring the accuracy of the initial model. The formula for calculating the slope instability probability is: In the formula, Let be the state variable at time t. Given the observation sequence from the initial time to the current time, the posterior probability... The forward-backward algorithm is used to recursively solve the problem, calculating the local instability probability of each structural unit of the slope and the overall comprehensive instability probability of the slope. The calculation accuracy is retained to four decimal places. The newly acquired observation data at the current moment, Let be the state variable at time t; This refers to historical observation data up to the previous moment. Let be the state variable at time t−1; For the observation likelihood probability, let represent the probability in the current state. Below, observed data The probability of; The one-step prediction probability means predicting the state at the current moment based on the observation data up to the previous moment. The probability of; The normalization constant is obtained by considering all possible states from the previous time step. Summation ensures that the sum of the posterior probabilities is 1.

[0040] The online parameter update of the dynamic Bayesian network employs an online expectation-maximization algorithm. The trigger condition is the receipt of a new batch of multi-source heterogeneous real-time data (updated every 10 minutes). The algorithm fine-tunes the parameters of the conditional probability table using sufficient statistics from the new data, with each iteration taking less than 30 seconds to ensure real-time model updates. Simultaneously, parameter updates are only applied to network nodes highly correlated with the newly collected data, avoiding the computational cost of updating the entire network's parameters. This achieves lightweight online optimization of the model, enabling it to quickly adapt to new slope instability patterns under sudden environments such as extreme rainfall and seismic disturbances. Figure 3 As shown.

[0041] It's important to note that the application of dynamic Bayesian networks in this context is irreplaceable. Not only can it output probabilities, but its state-observation model structure perfectly aligns with the nature of slope risk analysis, where hidden states need to be inferred through multi-source observations. More importantly, its online parameter update mechanism transforms the entire system from a static empirical model into a dynamic learning system. After a small landslide, the system can update its CPT (Conditional Perception Level) to remember the precursory patterns of the event, thus becoming more sensitive to similar patterns in the future.

[0042] This embodiment achieves deep fusion of multi-source heterogeneous data and intelligent analysis of slope risk through the above steps. It forms a complete technical closed loop from data acquisition and fusion processing to risk calculation and model optimization, effectively improving the accuracy, real-time performance and interpretability of road slope instability risk analysis.

[0043] Example 2: This embodiment applies to an all-weather risk monitoring and early warning scenario for a typical steep, mixed rock and soil slope of a highway in the southwestern mountainous region. The slope is approximately 1.2 kilometers long, with a maximum height of 45 meters. It has a complex geological structure, has experienced multiple shallow landslides in the past, and is significantly affected by monsoon rainfall, making it a key monitoring target. To achieve accurate, dynamic, and interpretable assessment of the slope's instability risk, a road slope risk analysis system based on multi-source data fusion was deployed.

[0044] First, from a system architecture perspective, this system adopts an edge-cloud collaborative computing paradigm, consisting of three layers: a front-end perception layer, an edge processing layer, and a cloud-based decision-making layer. The front-end perception layer comprises four types of heterogeneous sensing units: (1) Geological sensing subsystem, including distributed pore water pressure gauges (range 0.1MPa, nonlinear error <1%) and biaxial inclinometers (measurement range ±30°, resolution 0.01°). All devices are connected to the local data acquisition unit (DAQ) via RS-485 bus, and the sampling frequency is uniformly set to once per minute.

[0045] (2) Meteorological monitoring subsystem, equipped with an integrated six-element automatic weather station (integrated tipping bucket rain gauge (resolution 0.1mm), temperature and humidity sensor (accuracy ±2%RH / ±0.2℃), ultrasonic anemometer (starting wind speed <0.1m / s), data output time resolution of 5 minutes, and uploaded to edge nodes via LoRa wireless protocol.

[0046] (3) Deformation monitoring subsystem consists of a Beidou / GNSS high-precision positioning terminal (supporting multiple systems such as BDS / GPS / GLONASS, with static positioning accuracy better than 2cmRMS) and a fiber optic grating (FBG) strain sensing network. The latter uses a wavelength demodulator (16 channels, 100Hz sampling rate, strain measurement range 0~5000με, accuracy ±1με) to transmit the strain signal to the edge processing unit through a single-mode optical fiber.

[0047] (4) Video surveillance subsystem, deploying a dual-spectrum intelligent PTZ camera with IP67 protection level (visible light 1 / 1.8" CMOS, infrared night vision distance ≥200m, frame rate 25fps), and accessing the edge computing gateway via gigabit Ethernet.

[0048] The core of the edge processing layer is an industrial-grade edge computing gateway (built with an Intel Xeon D-1500 series processor, 32GB DDR4 ECC memory, and a 2TB NVMe solid-state drive, and integrated with 4G / 5G, Wi-Fi 6, and Gigabit Ethernet interfaces). This gateway runs a customized Linux real-time operating system, on which a data preprocessing module, a spatiotemporal alignment engine, and a lightweight feature extraction model are deployed. All front-end sensing data is aggregated here via TCP / IP or Modbus / TCP protocols, and a hardware watchdog and power management unit ensure stable operation 24 / 7. A secure tunnel is established between the edge nodes and the cloud via IPSec VPN, and the MQTT protocol is used for low-latency data synchronization, ensuring end-to-end transmission latency of less than 2 seconds.

[0049] The cloud-based decision-making layer is deployed on a public cloud platform (such as Alibaba Cloud ECS instance cluster), configured with at least 8 computing nodes (each node: Intel Xeon Gold 6248 RCPU, 192GB RAM, NVIDIA A100 GPU), running Kubernetes containerized services. The cloud carries three core functional modules: (1) a three-level deep fusion engine, responsible for performing feature weighting and graph neural network inference; (2) a dynamic Bayesian network (DBN) evaluation and online learning module; and (3) a visualization and early warning service module, providing web and mobile API interfaces. In addition, the system also integrates a historical case database, which adopts a hybrid architecture of distributed time series database (such as InfluxDB) and relational database (PostgreSQL), storing more than 5,000 sets of labeled slope disaster event full life cycle data, including original monitoring sequences, final instability state labels and post-event investigation reports, for model pre-training and transfer learning.

[0050] With the support of the above system architecture, the system's workflow strictly follows step S in the invention, as specifically illustrated below: After system startup, each front-end sensing unit operates continuously at a preset frequency. The pore water pressure gauge, earth pressure cell, and inclinometer convert analog signals into Modbus RTU messages via RS-485 bus. These messages are packaged by the DAQ module every 60 seconds and forwarded to the edge gateway via the CAN bus. The weather station generates a JSON data packet every 5 minutes, containing fields such as rainfall, temperature, humidity, and wind speed. This packet is aggregated by the LoRa gateway and pushed to the edge node. The GNSS terminal outputs NMEA-0183 format positioning data at a frequency of 1Hz. After being smoothed by Kalman filtering, the data is downsampled to one valid point every 10 minutes. The FBG demodulator collects wavelength drift signals at a rate of 10Hz, converts them into micro-strain values ​​in real time using a peak detection algorithm, and performs statistical compression in 10-minute windows. The dual-spectrum camera continuously records video streams at 25fps, while simultaneously enabling H.265 encoding and motion detection triggering mechanisms. Keyframes (one frame extracted every 10 minutes) are only uploaded to the edge node when the detected image change rate exceeds a threshold (e.g., 5% pixel change). Thus, the system forms a multi-source heterogeneous original dataset covering four dimensions: geology, meteorology, deformation, and vision, with timestamp accuracy unified to UTC millisecond level at the edge.

[0051] The raw dataset first enters the spatiotemporal alignment engine of the edge gateway. This engine employs a joint correction strategy: in the temporal dimension, with a basic time step of 10 minutes, linear interpolation or nearest neighbor filling is performed on data with sampling frequencies higher than this step size (such as FBG and video frames), while forward preserving is performed on data with shorter time steps (such as meteorological data); in the spatial dimension, all monitoring points obtain precise WGS-84 coordinates via RTK-GNSS, project them onto the UTM coordinate system, and then map them onto a 10m×10m regular grid. If multiple sensors of the same type exist within a grid, their arithmetic mean is taken as the representative value of that grid cell. After this processing, all data are unified into the same spatiotemporal grid system, forming a standardized input tensor.

[0052] Subsequently, the data enters the feature extraction stage. Geological and meteorological data are processed using a sliding window (window length = 6 time steps, i.e., 60 minutes) to calculate statistical features, including mean, standard deviation, and first-order difference mean (representing the rate of change). Deformation data are processed by calculating displacement gradient (displacement difference between adjacent grid cells divided by distance) and velocity accumulation (integral of displacement velocity over the past 60 minutes). Video image frames are fed into a lightweight convolutional neural network (MobileNetV3 backbone) deployed on an edge GPU to extract 128-dimensional texture features (based on LBP histogram), 64-dimensional edge intensity features (Canny operator response), and 32-dimensional motion vector features (calculated using optical flow). All feature vectors are concatenated to form a preliminary multimodal feature representation.

[0053] This feature representation is then uploaded to the cloud and enters the second feature representation layer of the three-level fusion architecture. At this layer, the system introduces a multimodal feature weighting strategy based on an attention mechanism. Specifically, the system maintains a feature saliency scorer, whose input is the cosine similarity between the current modality feature vector and the positive sample features in the historical risk event database. For the i-th modality (e.g., geology, meteorology, deformation, video), its saliency score is calculated. Defined as the average similarity between this modal feature and the features of known instability events over the past 24 hours. Weighting coefficients. Update dynamically using the following formula: in, Let exp represent the dynamic weights at time t in the i-th mode, and let exp represent the natural exponential function, with the adjustment parameter being... =0.8, total number of modes M=4, This represents the significance score at time t in the i-th mode.

[0054] Weighted eigenvectors The node features that constitute each 10m × 10m grid cell, wherein, This represents the total feature vector after fusion at time t. Let represent the original feature vector at time t of the i-th mode.

[0055] Next, the system enters the risk decision layer. In this layer, the slope profile is divided into N structural units (N=120 in this example), each unit corresponding to a grid node. The Graph Neural Network (GNN) uses these units as nodes, establishing undirected edges between units that are physically connected (e.g., in the same rock stratum) or whose spatial distance is less than 20 meters. The GNN employs a two-layer Graph Convolutional Network (GCN) architecture: the first layer aggregates direct neighbor information, and the second layer aggregates second-order neighbor information. After processing by the GCN, the node features output a local risk embedding representation for each unit. , where d=64, which represents the state of the cell itself and its neighborhood topological dependencies.

[0056] The fused high-dimensional feature vector (i.e., the embedded representation of the GNN output) serves as the observation input, driving the Dynamic Bayesian Network (DBN) for risk assessment. The DBN contains hidden state nodes. With observation nodes The set of hidden states is as follows: The initial parameters of the conditional probability table (CPT) were pre-trained using maximum likelihood estimation (MLE) with over 5000 samples from a historical case database. During system runtime, the posterior probability is recursively calculated using a forward-backward algorithm. Thus, the overall slope instability probability is obtained: .

[0057] To achieve model adaptation, the system initiates the Online Expectation-Maximization (OnlineEM) algorithm to fine-tune the CPT parameters every time it receives a new batch of observation data (once every 10 minutes). This process updates the parameters using sufficient statistics from the new data, with each iteration taking less than 30 seconds, ensuring that the model can quickly adapt to new instability modes following extreme rainfall or seismic disturbances.

[0058] The assessment results are processed by the visualization module and then pushed to the management terminal. Risk level is based on... The risk factors are divided into three levels: Level 1 (<0.1, green), Level 2 (0.1-0.7, orange), and Level 3 (>0.7, red). Key risk factors are assigned modal weights at the current time. Sort in descending order, and take the top 3 items (such as "sudden increase in pore water pressure", "cumulative rainfall exceeding the threshold", and "abnormal local displacement rate") as the dominant factors for output. The spatial distribution map renders the risk embedding value of each structural unit using color gradations (0~1 linearly mapped to blue-red gradient). It visually displays high-risk areas.

[0059] This embodiment fully demonstrates the entire technological implementation of the system, from hardware deployment to algorithm execution, showcasing its core innovations: multi-source fusion, three-level deep processing, dynamic evaluation, and precise early warning. In actual deployment, the system processes over 2TB of data daily, supports concurrent monitoring of 100+ slopes, has an overall response time of less than 60 seconds, and maintains a false alarm rate consistently below 12%, significantly outperforming traditional single-index methods.

[0060] This embodiment demonstrates, through a specific engineering example, the complete implementation process of the method of the present invention, from hardware deployment, data acquisition, processing and fusion to risk reasoning and early warning output. The specific equipment parameters, algorithm models, calculation formulas, and threshold settings given are merely examples; those skilled in the art can make adaptive adjustments based on different slope types, geological conditions, monitoring requirements, and computing resources without departing from the core principles of the present invention.

[0061] Example 3: A road slope risk analysis system based on multi-source data fusion is provided to implement the road slope risk analysis method based on multi-source data fusion described above. This system is adaptable to various road slope risk monitoring and analysis scenarios, including rock slopes, soil-rock mixed slopes, and soil-rock slopes. It supports a distributed deployment architecture with edge-cloud collaboration and can automate the entire process of multi-source heterogeneous data acquisition, fusion, feature processing, and risk inference. The system's functional modules are progressively arranged and coordinated according to the data processing flow, specifically including an acquisition module, a data fusion module, a multi-modal feature fusion and topology-aware risk decision-making module, and an inference module. The hardware compatibility, functional implementation logic, and data interaction methods of each module are as follows: The data acquisition module serves as the system's front-end data sensing unit, adapting to the communication protocols and data formats of various monitoring hardware devices. It is used for the targeted acquisition of four core data types from roadside slopes: geological parameters, meteorological data, deformation information, and video images. The module performs preliminary format standardization, outlier filtering, and timestamp calibration on the acquired raw data, forming a structured, multi-source, heterogeneous raw dataset. It also supports pushing the raw dataset to the data fusion module at a preset transmission frequency. This module includes geological sensor acquisition submodules, meteorological data acquisition submodules, deformation information acquisition submodules, and video image acquisition submodules. These submodules are adapted to geological sensors (pore water pressure gauges, earth pressure cells, and inclinometers), meteorological stations, deformation monitoring equipment (BeiDou / GNSS positioning modules and fiber optic strain gauges), and visible light and infrared dual-mode video monitoring units, respectively. It is compatible with mainstream communication protocols such as RS-485, LoRa, Gigabit Ethernet, and Modbus / TCP. It can flexibly adopt wired / wireless data transmission methods depending on the deployment location of the monitoring equipment and can customize the acquisition frequency of each data type, precisely matching the sampling frequency of the front-end monitoring equipment to ensure the integrity and timeliness of data acquisition.

[0062] The data fusion module is the core data processing unit of the system. It receives the multi-source heterogeneous raw datasets output by the acquisition module. Addressing the inherent differences in the time, space, data volume, and feature dimensions of the raw datasets, it sequentially performs three core processing steps: spatiotemporal alignment, feature extraction, and semantic fusion. This eliminates the spatiotemporal bias and volume influence between the multi-source data. Based on the processed standardized fused data, a three-level fusion architecture is constructed, consisting of the raw data layer, the feature representation layer, and the risk decision layer. This architecture provides structured data support for subsequent feature weighting and risk decision-making, and enables the classified storage, rapid retrieval, and real-time updating of data at each level. Spatiotemporal alignment employs a joint correction method combining timestamp interpolation and geographic coordinate projection to unify all multi-source heterogeneous data into the same spatiotemporal grid. Feature extraction utilizes differentiated feature mining strategies for the attributes of different data types, generating statistical features of mean, variance, and rate of change for geological and meteorological data, calculating displacement gradient and rate accumulation features for deformation data, and extracting texture, edge, and motion vector visual features from video images using convolutional neural networks. Semantic fusion integrates the normalized feature vectors of various types to form a unified multimodal feature representation. The three-level fusion architecture has clear functional boundaries and orderly data flow at each level. The raw data layer stores preprocessed multi-source heterogeneous raw data, the feature representation layer stores extracted and normalized multimodal feature data, and the risk decision layer receives the weighted feature data from the feature representation layer and provides high-dimensional feature input for risk inference.

[0063] The multimodal feature fusion and topology-aware risk decision-making module serves as the system's feature optimization and topology modeling unit. Based on a three-level fusion architecture built by the data fusion module, it performs targeted processing. At the feature representation layer, an attention-based multimodal feature weighting strategy is introduced. At the risk decision layer, a graph neural network is used to model the topological dependencies between slope structural units. The final output is a high-dimensional feature vector that integrates modal weight features and topological correlation features, which is then pushed to the inference module for slope instability probability calculation. Specifically, the multimodal feature weighting strategy dynamically assigns weight coefficients by calculating the contribution of each modal feature to historical slope risk events, assigning higher weights to features highly correlated with slope instability to achieve differentiated feature strengthening and weakening. The weight coefficients are updated in real-time according to a preset time step. The graph neural network modeling first divides the slope profile into several structural units as network nodes, using the physical connections or spatial proximity between units as network edges. Then, the attention-weighted multimodal features are used as node feature vectors. Two layers of graph convolution operations aggregate neighborhood node information, outputting a local risk embedding representation for each structural unit, fully exploring the inherent correlation features between different areas of the slope.

[0064] The inference module serves as the system's risk assessment and model optimization unit. Based on the high-dimensional feature vectors output by the multimodal feature fusion and topology-aware risk decision module, which are weighted by the attention mechanism and fused after topological dependency modeling, the module accurately calculates the local instability probability of each structural unit of the slope and the overall comprehensive instability probability of the slope through a dynamic Bayesian network. It also supports online iterative updates of the parameters of the dynamic Bayesian network by combining newly acquired multi-source heterogeneous real-time data, thereby achieving adaptive optimization of the model and ensuring that the slope risk analysis results are consistent with the real-time state changes of the slope. This module includes a Bayesian network inference submodule and an online parameter update submodule. The Bayesian network inference submodule uses high-dimensional feature vectors as input to the observation nodes and recursively solves the posterior probability through a forward-backward algorithm to obtain the slope instability probability value, and can classify the risk level according to the instability probability value. The online parameter update submodule adopts an online expectation-maximization algorithm, using a preset time step as the trigger condition, and uses newly collected multi-source heterogeneous real-time data as update samples to fine-tune the conditional probability table parameters of the dynamic Bayesian network. The time consumption of each parameter update is controlled within a preset threshold to ensure the real-time performance and lightweight nature of the model update, adapting to the computing power deployment requirements of edge computing.

[0065] The modules mentioned above adopt a loosely coupled modular design, which can flexibly adjust the deployment location (edge / cloud) of each module according to the actual slope monitoring scenario and computing power deployment conditions. The modules interact with each other through standardized API interfaces, and the data transmission process adopts an encryption protocol to ensure data security and integrity. The operating status of each module can be monitored in real time, and fault alarms and redundant backups are supported to ensure stable operation of the system 24 / 7.

[0066] During system operation, the data flow and logical control flow are as follows: The acquisition module pushes raw data to the data fusion module; after completing the construction of the three-level fusion architecture and basic feature extraction, the data fusion module sends the multimodal feature representation to the multimodal feature fusion and topology-aware risk decision-making module; after completing feature weighting and GNN modeling, this module sends the high-dimensional feature vector to the inference module; the inference module calculates the instability probability and updates the model parameters, finally outputting the risk result. The execution logic of all modules is defined by computer programs stored in the device's memory and executed by the processor.

[0067] A road slope risk analysis device based on multi-source data fusion is disclosed. This device serves as the hardware carrier for the aforementioned road slope risk analysis method and system. It supports edge, cloud, or edge-cloud collaborative deployment methods and is adaptable to the environmental requirements of various industrial application scenarios. It includes a memory and processor, and can be configured with communication, power, display, and interface modules as needed. All hardware components work together to provide full-process hardware support for road slope risk analysis. The memory stores a computer program containing all the algorithmic logic for implementing any of the aforementioned road slope risk analysis methods based on multi-source data fusion. The memory also stores various types of data, including multi-source heterogeneous raw datasets, feature processing data, slope risk analysis results, historical case data, and model parameters. The memory includes both volatile and non-volatile memory. The storage device can use solid-state drives, flash memory, hard disks, or other storage media to ensure persistent data storage. The volatile memory can use DDR4 / DDR5 memory to provide high-speed cache space for the processor's program execution. The storage capacity of the memory can be flexibly expanded according to actual monitoring needs. The processor is electrically connected to the memory. When the processor executes the computer program, it implements a road slope risk analysis method based on multi-source data fusion as described in any of the preceding claims. The processor can use an industrial-grade multi-core CPU, embedded microprocessor, or GPU, supporting single-core / multi-core parallel computing. Intel Xeon series, ARM series, or other processor models can be selected according to computing power requirements. It can efficiently execute algorithms such as spatiotemporal alignment, feature extraction, attention mechanism weighting, graph neural network modeling, and dynamic Bayesian network inference to ensure the real-time performance of data processing and risk inference.

[0068] When the device's processor executes the risk assessment program, it schedules the communication module to receive data, calls the in-memory algorithm models (such as CNN feature extraction models, attention weight calculation modules, GNN models, and DBN inference engines) for processing, and writes the intermediate and final results into memory. Collaboration between the edge and cloud devices is achieved through the communication module: the edge device handles high-frequency, low-latency local preprocessing and simple inference, while the cloud device handles complex model training, global optimization, and deep analysis. Both synchronize models and data through a secure network protocol.

[0069] Furthermore, the device's communication module is compatible with various communication methods such as RS-485, LoRa, 4G / 5G, Wi-Fi 6, and Gigabit Ethernet, enabling data transmission between the device and front-end monitoring equipment, between devices, and between the device and terminal display devices. The power module uses an industrial-grade power supply, supporting dual power supply from mains and backup power, and features overvoltage, overcurrent, and short-circuit protection to ensure stable power supply for the device. The interface module includes data input interfaces, data output interfaces, and expansion interfaces, enabling physical connections with various monitoring hardware, storage devices, and display devices. The expansion interfaces support flexible access and expansion of hardware devices. The display module can be configured as needed to display key information such as slope instability probability, risk level, and equipment operating status in real time, while also supporting touch operation for visualized parameter configuration.

[0070] The device can be divided into an edge analysis device and a cloud analysis device according to the actual application scenario. The edge analysis device adopts a lightweight hardware configuration and is deployed at the slope monitoring site. It mainly realizes functions such as data acquisition, preliminary preprocessing, spatiotemporal alignment, and lightweight feature extraction. The cloud analysis device adopts a high-performance hardware cluster configuration and is deployed in the cloud data center. It mainly realizes functions such as multimodal feature weighting, graph neural network modeling, dynamic Bayesian network inference, and global update of model parameters. The edge analysis device and the cloud analysis device realize data synchronization and interaction through encrypted network communication protocols, forming an edge-cloud collaborative road slope risk analysis hardware system that balances the real-time performance of on-site data processing with the high efficiency of cloud model optimization.

[0071] In practical use, when the processor of the device executes the computer program in the memory, it can call the various hardware components and the functional modules of the aforementioned road slope risk analysis system to achieve linkage, automatically completing the entire process from multi-source data acquisition to slope instability probability calculation and model parameter update. The analysis results can be pushed to mobile terminals, Web terminals and other terminal devices through the communication module to realize real-time monitoring and early warning of road slope risks. The hardware components of the device are all designed with industrial grade, and have the characteristics of dustproof, waterproof and anti-interference, which can be adapted to complex road slope monitoring environments such as mountainous areas and the wild.

[0072] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for road slope risk analysis based on multi-source data fusion, characterized in that, Includes the following steps: S1 collects geological parameters, meteorological data, deformation information and video images to form a multi-source heterogeneous raw dataset; S2 addresses the dimensional differences of multi-source heterogeneous original datasets by sequentially performing spatiotemporal alignment, feature extraction, and semantic fusion on the original datasets. Based on the processed data, a three-level fusion architecture is constructed, consisting of the original data layer, the feature representation layer, and the risk decision layer. S3, based on a three-level fusion architecture, introduces a multimodal feature weighting strategy based on an attention mechanism in the feature representation layer, and uses a graph neural network to model the topological dependencies between slope structural units in the risk decision layer; S4, based on high-dimensional feature vectors fused after being weighted by attention mechanism and modeled by topological dependencies, calculates the slope instability probability through a dynamic Bayesian network, and updates the network parameters online by combining newly collected multi-source heterogeneous real-time data.

2. The road slope risk analysis method based on multi-source data fusion according to claim 1, characterized in that: Geological parameters are collected through geological sensors, meteorological data is collected through meteorological stations, deformation information is collected through deformation monitoring equipment, and video images are collected through video monitoring units. Among them, geological sensors include pore water pressure gauges, soil pressure cells, and inclinometers; meteorological data collected by meteorological stations include rainfall, temperature, humidity, and wind speed; deformation monitoring equipment includes BeiDou / GNSS positioning modules and fiber optic strain gauges; and the video monitoring unit is equipped with a dual-mode camera for visible light and infrared.

3. The road slope risk analysis method based on multi-source data fusion according to claim 1, characterized in that: Spatiotemporal alignment employs a joint correction method based on timestamp interpolation and geographic coordinate projection to unify multi-source heterogeneous data into the same spatiotemporal grid; Feature extraction uses sliding window statistics on geological and meteorological data to generate mean, variance, and rate of change features, respectively. For deformation data, displacement gradient and rate of change features are calculated. For video images, texture, edge, and motion vector features are extracted using a convolutional neural network.

4. The road slope risk analysis method based on multi-source data fusion according to claim 1, characterized in that: The multimodal feature weighting strategy dynamically assigns weight coefficients by calculating the contribution of each modality feature to historical risk events. The weight update formula is as follows: in, Let be the weight of the i-th mode at time t. Its feature significance score, To adjust the parameters, M is the total number of modes. This is the sum of the exponential scores of all modes at time t.

5. The road slope risk analysis method based on multi-source data fusion according to claim 1, characterized in that: The graph neural network divides the slope profile into several structural units as nodes, and uses the physical connections or spatial proximity relationships between units as edges. The node feature vector is composed of multimodal features weighted and fused by the attention mechanism. Neighborhood information is aggregated through two layers of graph convolution operations, and the local risk embedding representation of each unit is output.

6. The road slope risk analysis method based on multi-source data fusion according to claim 1, characterized in that: The dynamic Bayesian network consists of state nodes and observation nodes. The state nodes are used to represent the stability state of the slope, and the observation nodes correspond to high-dimensional feature vectors that are weighted by the attention mechanism and fused after modeling topological dependencies. The conditional probability table is initialized by the maximum likelihood estimation method, and the network parameters are iteratively updated based on the newly acquired multi-source heterogeneous real-time observation data using the online expectation-maximization algorithm.

7. The road slope risk analysis method based on multi-source data fusion according to claim 1, characterized in that: The formula for calculating the probability of slope instability is: In the formula, Let be the state variable at time t. Given the observation sequence from the initial time to the current time, the posterior probability... The forward-backward algorithm is used to recursively solve the problem, calculating the local instability probability of each structural unit of the slope and the overall comprehensive instability probability of the slope. The calculation accuracy is retained to four decimal places. The newly acquired observation data at the current moment, Let be the state variable at time t; This refers to historical observation data up to the previous moment. Let be the state variable at time t−1; For the observation likelihood probability, let represent the probability in the current state. Below, observed data The probability of; The one-step prediction probability means predicting the state at the current moment based on the observation data up to the previous moment. The probability of; The normalization constant is calculated by considering all possible states from the previous time step. Summation ensures that the sum of the posterior probabilities is 1.

8. A road slope risk analysis system based on multi-source data fusion, used to implement the method described in any one of claims 1-7, characterized in that: It includes a data acquisition module, a data fusion module, a multimodal feature fusion and topology-aware risk decision-making module, and an inference module; The acquisition module is used to collect geological parameters, meteorological data, deformation information and video images to form a multi-source heterogeneous raw dataset; The data fusion module is used to sequentially perform spatiotemporal alignment, feature extraction, and semantic fusion on the original dataset, constructing a three-level fusion architecture consisting of the original data layer, the feature representation layer, and the risk decision layer; The multimodal feature fusion and topology-aware risk decision module is used to introduce an attention-based multimodal feature weighting strategy in the feature representation layer and to use a graph neural network to model the topological dependencies between slope structural units in the risk decision layer. The inference module is used to calculate the slope instability probability through a dynamic Bayesian network based on the fused high-dimensional feature vector, and to update the network parameters online by combining the newly collected data.

Images