Landslide disaster cloud-side-end monitoring and early warning scene simulation test platform

Abstract

The invention discloses a cloud-side-end monitoring and early warning scene simulation test platform for landslide disasters. The landslide simulation test device of the platform is used for simulating landslides in different gradients, different intensity rainfall and underground water seepage scenes; the end layer multi-source monitoring and collecting system is mounted in the landslide simulation test device and is used for monitoring landslide data; the boundary layer edge early warning monitoring system receives the landslide data and obtains landslide danger early warning information through model processing; the cloud layer cloud computing simulation system obtains landslide data and landslide danger early warning information for online processing and visualization, and carries out training and lightweight deployment on a model in the side layer edge early warning monitoring system. According to the invention, a sensor, a monitoring and early warning system and an early warning mechanism in the field of landslide disaster prevention and control can be tested and verified under laboratory conditions, iterative optimization is carried out in each landslide test, and the efficiency and reliability of the emerging technology in the field of landslide disaster prevention and control are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to a pre-warning scene simulation test platform, relates to landslide disaster monitoring and pre-warning technology, relates to the intersection field of geological disaster prevention, Internet of Things, computer vision, edge computing and artificial intelligence application, and particularly relates to a landslide disaster cloud-edge-end monitoring and pre-warning scene simulation test platform. BACKGROUND

[0002] The landslide is complex in origin and rapid in evolution, and early warning is very challenging. The research on landslide in real scene has long research period and strong uncontrollability, and the laboratory artificial landslide simulation has small scale and insufficient authenticity, which limits the development of the monitoring system. The traditional landslide monitoring mainly monitors a single physical quantity, lacks multi-source information fusion analysis, and the data processing mainly depends on a remote server, so that there is a problem of communication delay and insufficient real-time performance. Therefore, it is urgent to develop a cloud-edge-end integrated monitoring and pre-warning platform which can simulate the whole process of landslide disaster in a laboratory environment and comprehensively integrates multi-sensor, visual system and edge computing. SUMMARY

[0003] In order to solve the problems in the background art, the present application provides a landslide disaster cloud-edge-end monitoring and pre-warning scene simulation test platform. The present application is a landslide disaster cloud-edge-end simulation test platform combining visual monitoring, multi-sensor fusion, edge intelligent reasoning and cloud data management, integrating image crack monitoring, multi-sensor data aggregation, multi-source fusion pre-warning and edge intelligent reasoning function, aiming to solve the technical problems of single monitoring means, data processing lag and lack of intelligent experiment management in the existing landslide disaster experiment platform, and provide an efficient and intelligent test scheme.

[0004] The technical scheme adopted by the present application is: The landslide disaster cloud-edge-end monitoring and pre-warning scene simulation test platform of the present application comprises: A landslide simulation test device for simulating landslide under different slope, different intensity rainfall and underground water seepage scene.

[0005] An end layer multi-source monitoring and collecting system installed in the landslide simulation test device and used for multi-source monitoring data collection during landslide simulation to monitor landslide data.

[0006] An edge layer edge pre-warning monitoring system receiving landslide data of the end layer multi-source monitoring and collecting system and obtaining landslide danger degree through model processing.

[0007] A cloud layer cloud computing simulation system acquiring landslide data and landslide danger degree transmitted by the edge layer edge pre-warning monitoring system for online processing and visualization, and training and lightweight deployment of the model in the edge layer edge pre-warning monitoring system.

[0008] The landslide experiment process includes slope body filling and ramming, sensor layout, edge device initialization, artificial rainfall induced landslide, whole process monitoring and data recording, and device recovery and site restoration after the experiment; among them, the experiment preparation and landslide induction stage are identified by visual monitoring, crack evolution and soil block collapse automatic identification and analysis, which are used to ensure the safety of the experiment and enrich the surface deformation monitoring information.

[0009] The landslide simulation test device comprises a rock-soil bearing box, a lifting and tensioning mechanism, an artificial rainfall mechanism and a porous ceramic pipe. One end of the rock-soil bearing box is supported by the lifting and tensioning mechanism on the bottom surface, and the lifting and tensioning mechanism is lifted and lowered to change the slope of the rock-soil bearing box. A gravel layer and a rock-soil layer are laid in the rock-soil bearing box from bottom to top as a simulated soil body for constructing a simulated soil slope of a specific material. The artificial rainfall mechanism is installed directly above the rock-soil bearing box to perform rainfall of different intensities. The artificial rainfall mechanism can specifically adopt a water spraying array. The porous ceramic pipe is laid on the inner bottom surface of the rock-soil bearing box and is located below the gravel layer. The porous ceramic pipe is connected to an external water pump to simulate the seepage condition of underground water.

[0010] The end-layer multi-source monitoring and collecting system comprises a contact type slope body measuring unit and a non-contact type slope body measuring unit. A plurality of monitoring points away from the box body of the rock-soil bearing box are arranged on the simulated soil body in the rock-soil bearing box in an array. The contact type slope body measuring unit comprises a plurality of sensing arrays. Each sensing array is arranged in the rock-soil layer under a respective monitoring point. Each sensing array comprises a tensioned wire displacement meter, two soil moisture meters and a soil pressure meter arranged in an array from top to bottom. The tensioned wire displacement meter, the two soil moisture meters and the soil pressure meter monitor the surface tensioned wire displacement amount, the shallow soil moisture content, the deep soil moisture content and the soil pressure, respectively. The contact type slope body measuring unit further comprises a rain gauge, an underground three-dimensional displacement meter and a shear wave velocity meter. The rain gauge is arranged in the space between the simulated soil body and the artificial rainfall mechanism to monitor the rainfall amount. The underground three-dimensional displacement meter and the shear wave velocity meter are installed in the middle of the inner bottom surface of the rock-soil bearing box to obtain the mechanical and hydrological state parameters of the surface layer and the deep layer of the slope body in real time. A multi-protocol sensor network combining WiFi, Bluetooth, Lora and other communication modes is used to realize the data collection and transmission of the multi-source information of the test field. The state parameters of the sensors such as the water level of the storage tank of the artificial rainfall mechanism (i.e. the rainfall amount), the inclination angle of the box body of the rock-soil bearing box and the slope body information control group can also be obtained.

[0011] The non-contact type slope body measuring unit comprises a microprocessor and a network camera / recorder installed with an image crack meter. The microprocessor is electrically connected to the network camera. The network camera is installed on the side of the rock-soil bearing box and faces the upper surface of the simulated soil body. The image crack meter is installed in the microprocessor. The image crack meter processes the images of the simulated soil body surface during the landslide taken by the network camera to identify the crack information of the rock-soil cracks during the landslide.

[0012] The image crack meter is used in processing a simulated soil surface image, when a slope crack exists in the simulated soil surface image, first, the simulated soil surface image is converted into a gray image by using a mask, gray scale, and histogram equalization in sequence, then the gray image is further processed by using erosion and dilation morphological operations in sequence to optimize crack details to obtain a crack optimization image, then the contour of each slope crack in the crack optimization image is identified, the slope crack with a pixel point greater than a preset threshold in the contour is reserved, and the number of all pixel points of each reserved slope crack is taken as the crack size of the current simulated soil surface image, that is, as crack information.

[0013] The edge layer edge early warning monitoring system comprises an upper computer and an edge computing node, the upper computer receives and stores landslide data transmitted by a contact type slope body measurement unit of an end layer multi-source monitoring and collecting system, and performs preprocessing and landslide danger degree calculation, and then transmits to the edge computing node and a cloud layer cloud computing simulation system in sequence, the landslide danger degree is obtained through a landslide danger degree evaluation and early warning model arranged in the edge computing node; the edge computing node receives crack information transmitted by a microprocessor of a non-contact type slope body measurement unit and simulated soil surface images transmitted by a network camera, and then transmits to the cloud layer cloud computing simulation system.

[0014] When the upper computer pre-processes the landslide data, the upper computer performs outlier rejection, sensor error correction and filtering operation in sequence.

[0015] The landslide danger degree R is specifically as follows: R=w1×(r s ) n +w2×(r c ) n r b =(d b ×v b 2 ) / 2+d b ×a b ×v b , b=s、c Wherein, w1 and w2 represent first and second weight coefficients respectively; (r s ) n and (r c ) n are normalized values of the tension displacement danger degree r s and the slope crack danger degree r c ; d b , v b and a b represent current value, change speed and acceleration of the tension displacement or crack information.

[0016] When the landslide risk R exceeds the preset risk threshold, it is judged that the current simulated soil body has a landslide risk, and the landslide risk R can prompt the value to reach the peak in advance, reflecting the mutation of the sensor value.

[0017] The landslide risk assessment and early warning model further comprises a feedforward neural network, which inputs rainfall, soil moisture content and soil pressure data obtained by each monitoring point, and outputs a slope deformation related risk degree, so as to further warn the landslide risk.

[0018] The cloud computing simulation system comprises a plurality of independent hardware computing nodes, backs up and displays the data transmitted from the edge layer edge early warning monitoring system, trains and lightweightly deploys the landslide risk assessment and early warning model in the edge layer edge early warning monitoring system, and provides remote visualization function through a data dashboard. The cloud layer is based on a plurality of independent hardware computing nodes, simulates cloud service deployment, realizes time series database storage, online data dashboard, data cleaning, model training and lightweight deployment. The platform also has a landslide test management system, which realizes automatic and intelligent monitoring of the progress of the test task itself.

[0019] The beneficial effects of the present application are: 1. The experimental platform is consistent with the actual monitoring scene: the experimental platform of the present application efficiently reproduces the monitoring scene of actual landslide disasters in a laboratory environment, can simulate rainfall, slope deformation, groundwater seepage and other complex working conditions, provides reliable support for the development, testing and optimization of the landslide monitoring and early warning system, and significantly improves the practicality and deployment efficiency of the system.

[0020] 2. Multi-source information fusion and analysis of the correlation of landslide parameters: the present application comprehensively utilizes rainfall, shallow and deep water content, soil pressure, tension displacement, crack width and other sensor data, deeply analyzes the landslide evolution process, extracts precursor indicators, supports the design and verification of the landslide early warning mechanism, and ensures the scientificity and accuracy of the early warning judgment.

[0021] 3. Experimental-application closed loop to build a comprehensive test platform: the present application not only has a complete cloud-edge-end collaborative architecture, but also realizes an experimental-application closed loop from data acquisition, edge reasoning to cloud deep learning and visualization, forms an integrated test environment for landslide disaster prevention, and improves the development and verification efficiency of new technologies.

[0022] 4. Deep integration of computer vision and Internet of Things: the present application innovatively combines computer vision (such as crack detection and target recognition) with traditional sensor networks, supports data fusion and dynamic early warning in complex scenes, and provides an efficient and practical test field for the comprehensive application of computer vision, Internet of Things and edge computing in the field of landslide disasters.

[0023] 5. Intelligent experiment management improves safety and efficiency: The platform has built-in automation management functions, including personnel and equipment identification during the experiment process, liquid level and inclination detection, data backup and push function, which optimizes the experiment process, improves the safety protection level, significantly reduces the labor input and improves the integrity and reliability of the test data. BRIEF DESCRIPTION OF DRAWINGS

[0024] Figure 1 A flow chart of the landslide disaster process simulation test steps of the present application; Figure 2 A monitoring schematic diagram of the landslide simulation test of the present application; Figure 3 A schematic diagram of the cloud-edge-end monitoring scene of the landslide of the present application; Figure 4 A visual auxiliary cloud-edge-end landslide monitoring and early warning system architecture diagram of the present application; Figure 5 A slope crack evolution image recognition result diagram in the landslide simulation experiment of the present application; Figure 6 A contact and non-contact sensor measurement data diagram in the landslide simulation experiment of the present application; Figure 7 A landslide risk curve diagram of the present application fusing the tension displacement and image crack; Figure 8 A risk prediction model diagram based on the relationship between the monitoring point position and each sensor of the present application; Figure 9 A comparison diagram of the inference result of the lightweight risk model and the BP neural network model of the present application. DETAILED DESCRIPTION

[0025] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings and specific embodiments. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0026] The landslide disaster cloud-edge-end monitoring and early warning scene simulation test platform of the application comprises a landslide simulation test device, an end layer multi-source monitoring and collecting system, an edge layer edge early warning monitoring system and a cloud layer cloud computing simulation system. The landslide simulation test device is used to simulate landslides under different slope gradients, different intensity rainfall and underground water seepage conditions. The end layer multi-source monitoring and collecting system is installed in the landslide simulation test device and is used for multi-source monitoring data collection during landslide simulation to monitor landslide data. The edge layer edge early warning monitoring system receives landslide data from the end layer multi-source monitoring and collecting system and then obtains landslide risk degree through model processing. The cloud layer cloud computing simulation system obtains landslide data and landslide risk degree transmitted by the edge layer edge early warning monitoring system for online processing and visualization, and trains and lightweightly deploys the model in the edge layer edge early warning monitoring system. The landslide experiment process comprises slope body filling and ramming, sensor arrangement, edge device initialization, artificial rainfall induced landslide, full-process monitoring and data recording, and equipment recovery and site recovery after the experiment ends. In the experiment preparation and landslide induction stage, personnel and equipment are identified, crack evolution and soil block collapse are automatically identified and analyzed through visual monitoring, which is used to ensure experiment safety and enrich surface deformation monitoring information.

[0027] The landslide simulation test device comprises a rock-soil bearing box, a lifting and tensioning mechanism, an artificial rainfall mechanism and a porous ceramic pipe. One end of the rock-soil bearing box is supported by the lifting and tensioning mechanism at the bottom surface, and the lifting and tensioning mechanism is lifted and lowered to change the slope of the rock-soil bearing box. The inside of the rock-soil bearing box is paved with a gravel layer and a rock-soil layer from bottom to top as a simulated soil body, which is used to construct a simulated soil slope body of a specific material. The artificial rainfall mechanism is installed directly above the rock-soil bearing box to perform rainfall of different intensities. The artificial rainfall mechanism can specifically use a water spraying array. The porous ceramic pipe is laid on the inner bottom surface of the rock-soil bearing box and is located below the gravel layer. The porous ceramic pipe is connected to an external water pump to simulate underground water seepage conditions.

[0028] The multi-source monitoring and acquisition system includes a contact slope measurement unit and a non-contact slope measurement unit. Several monitoring points, not close to the soil bearing box, are arrayed at intervals on the simulated soil within the soil bearing box. The contact slope measurement unit includes several sensor arrays, each arranged in the soil layer below its respective monitoring point. Each sensor array includes, from top to bottom, a pull-wire displacement gauge, two soil moisture meters, and one earth pressure gauge, which monitor surface pull-wire displacement, shallow soil moisture content, deep soil moisture content, and earth pressure, respectively. The contact slope measurement unit also includes a rain gauge, a subsurface 3D displacement gauge, and a shear wave velocity meter. The rain gauge is positioned in the space between the simulated soil and the artificial rainfall mechanism to monitor rainfall. The subsurface 3D displacement gauge and shear wave velocity meter are installed in the center of the inner bottom surface of the soil bearing box to acquire real-time mechanical and hydrological parameters of the slope surface and deep layers. It can be combined with a multi-protocol sensor network using communication methods such as WiFi, Bluetooth, and LoRa to achieve multi-source data acquisition and transmission from the test site. It can also acquire the status parameters of sensors such as the water level in the storage tank of the artificial rainfall mechanism (i.e., the amount of precipitation), the tilt angle of the soil and rock bearing box, and the slope information control group.

[0029] The non-contact slope measurement unit includes a microprocessor equipped with an image crack gauge and a network camera / recorder. The microprocessor is electrically connected to the network camera. Specifically, the microprocessor is a Nuvoton NUC980 microprocessor, which reads the network video stream from the camera / recorder using the Real-Time Streaming Protocol (RTSP). The network camera is installed on the side of the soil-rock bearing box and faces the upper surface of the simulated soil. The image crack gauge is installed in the microprocessor and processes the images of the simulated soil surface captured by the network camera during a landslide to identify crack information in the soil-rock during the landslide.

[0030] When processing simulated soil surface images, the image-based crack meter first converts the simulated soil surface image into a grayscale image by sequentially applying masking, grayscale conversion, and histogram equalization. Then, it further processes the grayscale image using erosion and dilation morphological operations to optimize crack details and obtain an optimized crack image. Next, it identifies the contours of each slope crack in the optimized crack image, retaining slope cracks with pixels larger than a preset threshold. The total number of pixels for each retained slope crack is used as the crack size in the current simulated soil surface image, i.e., crack information. The image-based crack meter uses digital image processing and an embedded system to identify landslide crack evolution and soil collapse. Specifically, it uses the OpenCV computer vision library to perform Region of Interest (ROI) segmentation, morphological erosion / dilation, and contour recognition on the crack image, extracting the pixel set of the largest crack contour as the crack index.

[0031] The edge early warning and monitoring system includes a host computer and edge computing nodes. The host computer receives and stores landslide data transmitted from the contact slope measurement unit of the edge multi-source monitoring and acquisition system. Simultaneously, it performs preprocessing and landslide hazard calculation before transmitting the data sequentially to the edge computing nodes and the cloud computing simulation system. The landslide hazard is obtained through a landslide hazard assessment and early warning model deployed in the edge computing nodes. The specific landslide hazard R is as follows: R=w1×(r s ) n +w2×(r c ) n r b =(d b ×v b 2 ) / 2+d b ×a b ×v b b=s,c Where w1 and w2 represent the first and second weighting coefficients, respectively; (r s ) n and (r) c ) n The respective values ​​represent the risk level r of the wire displacement. s and slope crack risk level r c The normalized value of d; b v b and a b These represent the current value, rate of change, and acceleration of the wire displacement or crack information, respectively. The unit of wire displacement is mm.

[0032] When the landslide hazard R exceeds the preset hazard threshold, it is determined that the simulated soil has a landslide hazard. The landslide hazard R can provide an early warning that the value has reached its peak, reflecting the sudden change in the sensor reading.

[0033] Edge computing nodes receive crack information transmitted by the microprocessor of the non-contact slope measurement unit and simulated soil surface images transmitted by network cameras, and then transmit them to the cloud computing simulation system. A Raspberry Pi 2B single-board computer is used as the edge computing node, which includes two functions: multi-source information aggregation and preprocessing, and landslide hazard output. The edge layer aggregates multi-source information, enabling local database storage of data at the second level, preprocessing and sending data to the cloud at the minute level, lightweight model inference and early warning, and data webpage interaction. The edge computing node uses the TensorFlow Lite engine to execute lightweight model inference, which includes model pruning, quantization, computational graph optimization, and edge hardware adaptation.

[0034] The landslide hazard assessment and early warning model also includes a feedforward neural network, which takes into account rainfall and soil moisture content and earth pressure data obtained from various monitoring points, and outputs slope deformation-related hazard levels to provide further early warning of landslide risks.

[0035] When some sensors fail, edge computing nodes can run a landslide hazard assessment and early warning model to predict the current hazard level based on data from the remaining sensors. The model undergoes pruning, quantization, structural optimization, and edge adaptation. The landslide hazard assessment and early warning model is based on a feedforward neural network, constructing a nonlinear mapping through fully connected layers. The key innovation lies in dynamically injecting real state values ​​into the intermediate layers of the network, forming a conditional input mechanism with hybrid supervisory signals. A multi-channel neural network architecture is constructed based on external input, internal state, and deformed output, and measured state corrections are introduced into the model at the intermediate layers to improve assessment accuracy while maintaining interpretability of the landslide mechanism.

[0036] Specifically, for two monitoring points on the slope, the model simulates the forward propagation of external rainfall input along two parallel paths through the internal soil conditions. Rainfall (jy) measured by a rain gauge is used as the external input, then passes through stages of shallow moisture (q1, q2), deep moisture (s1, s2) measured by a soil moisture meter, and soil pressure (pr) measured by an earth pressure gauge. At each stage, real-world sensor readings are injected in a cascaded manner to integrate real-time observation data and enhance model fidelity. The first path simulates the chain jy-q1-s1-pr1, and the second path simulates jy-q2-s2-pr2. These two branches are then merged to generate the final deformation-related hazard output, representing the two monitoring pits on the slope.

[0037] When the host computer preprocesses the landslide data, it sequentially performs outlier removal, sensor error correction, and filtering operations. Edge computing nodes aggregate sensor data from different communication methods, including database queries and message queue protocol-based data subscriptions; they then perform preprocessing on the acquired sensor data within a time window, including outlier removal, sensor error correction, and filtering operations, before sending the processed data to the cloud.

[0038] The cloud-based cloud computing simulation system comprises several independent hardware computing nodes. It backs up and displays data transmitted from the edge early warning and monitoring system, trains and lightweight-deploys the landslide hazard assessment and early warning model within the edge early warning and monitoring system, and provides remote visualization through a data dashboard. Based on multiple independent hardware computing nodes, the system simulates cloud service deployment, enabling time-series database storage, online data dashboards, data cleaning, model training, and lightweight deployment. The platform also features a landslide test management system, automating and intelligently monitoring the progress of the test tasks.

[0039] The cloud-based cloud computing simulation system performs online processing and visualization analysis of landslide monitoring data, as well as training and lightweight deployment of early warning models. It also includes automated auxiliary tools for landslide simulation experiments. Based on a cloud computing environment, it supports data cleaning, deep learning model training, and lightweight model deployment, and provides remote visualization capabilities through a data dashboard. These include on-site database backup, remote video stream backup, human-machine interaction during experiment preparation, soil collapse identification, public network access to monitoring data, and online software updates. The cloud computing service, based on a time-series database and an online visualization dashboard, is used for the storage, querying, and visualization of multi-source monitoring data. The cloud-based automated auxiliary tools, based on the operating system, automatically back up the database and remote video streams and support remote updates of code and software.

[0040] The platform of this invention also features a landslide test management system, enabling automated and intelligent monitoring of the progress of the test tasks themselves. This allows for the testing and verification of sensors, monitoring and early warning systems, and early warning mechanisms in the field of landslide disaster prevention even under laboratory conditions, and enables iterative optimization in each landslide test, thereby improving the efficiency and reliability of the practical application of emerging technologies in the field of landslide disaster prevention.

[0041] like Figure 1 As shown, in specific implementation, the simulation of the platform landslide disaster process of this invention is divided into three stages: the preliminary stage, the preparatory stage, and the production stage, as detailed below: Step 1: Artificially Constructing the Slope. Using a bulldozer, fill the soil into a soil-supporting box (approximately 5m × 4m × 1.5m in size). Lay a 20cm thick layer of crushed stone at the bottom of the box to simulate a permeable underground layer. Then, fill the top with approximately 1 meter of cohesive soil to construct the slope. During this stage, keep the box door closed, and place the box horizontally or at a slight angle to meet the requirements of filling and compaction.

[0042] Step Two: Sensor Deployment and System Initialization. After manually compacting the slope, monitoring points are set up. Holes are excavated at predetermined locations to deploy monitoring sensors, including soil moisture meters, earth pressure gauges, and wire displacement gauges, etc. Figure 2 As shown, the soil was sealed and compacted to restore the slope's integrity. A porous ceramic pipe was connected to a water pump system to simulate groundwater seepage conditions. The enclosure door was then unlocked and opened, and the opening was adjusted using a hand-cranked hoist to tilt the enclosure, creating boundary conditions for subsequent slope slippage. This stage also included checking and initializing the edge computing devices, power system, and network communication.

[0043] Step 3: Landslide Disaster Simulation Begins. The artificial rainfall device is activated to raise the soil moisture content to between 25% and 30%. At this point, the container is tilted to a specified angle (adjustable within the range of 0–60°). After the initial soil slippage caused by the tilt, the subsequent changes in the slope are considered part of the simulated rainfall-induced landslide process. During this process, multiple sensors are used to monitor the external inputs (rainfall), internal conditions (soil moisture content, soil pressure), and deformation outputs (wire displacement, image cracks) of the slope. This data is used for analyzing landslide induction mechanisms and early warning criteria, and can also be used to test the reliability and accuracy of sensors, monitoring and early warning systems, and landslide early warning models.

[0044] Step 4: During the landslide simulation, intermittent rainfall is conducted. Once the slope at all monitoring points with embedded sensors has failed, the container is laid flat, and the landslide disaster simulation process ends. Further analysis is performed on the backed-uploaded data from each monitoring system, including data cleaning and model training.

[0045] Step 5: Once the slope fails and all sensors are removed from the monitoring points, begin retrieving the sensors, clearing away residual soil from the slope, closing the enclosure door, and entering the preparation phase for the next landslide disaster simulation.

[0046] like Figure 3 and Figure 4 As shown, in the specific implementation of this invention, the system is divided into three layers—cloud, edge, and terminal—based on the geographical location of the landslide disaster scenario simulation experimental platform. Considering that there may be many potentially hazardous slopes in an area, each with numerous sensors monitoring in real time (i.e., terminal nodes), if all sensors directly transmit data to cloud nodes via network access points and long-distance communication (such as fiber optics, satellite, etc.), it will place enormous pressure on the energy requirements, network bandwidth, and cloud servers of the terminal nodes. By adding an edge layer and edge nodes, the terminal data can be preliminarily calculated and processed near the site, thus reducing energy consumption and network bandwidth usage for medium- and long-distance communication (such as 4G, LoRa, etc.), reducing time latency, and enabling short-distance communication (such as WiFi, Bluetooth, etc.) between terminal nodes and edge nodes within a single potentially hazardous slope. Table 1 summarizes the resource usage information of the cloud-edge-terminal three layers.

[0047] Table 1 Comparison of Cloud-Edge-Device Resource Requirements and Data Volume

[0048] The vision-assisted cloud-edge-end landslide monitoring and early warning platform needs to include the following functions: expansion of the types and number of sensors at the "end" layer; multi-sensor data aggregation and lightweight model early warning at the "edge" layer; and real-time monitoring, online analysis, model training, and lightweight deployment of landslide data at the "cloud" layer. The "cloud" layer is simulated based on the existing computer hardware and network equipment in the research laboratory, and has advantages such as controllable hardware and software resources, low cost, and data security.

[0049] The "end" layer has both contact and non-contact measuring devices. For example... Figure 4 As shown, this includes contact-type measuring devices such as rain gauges, wire displacement gauges, soil moisture meters, and earth pressure gauges. Data monitoring is conducted using two methods: an STM32-NRF2401-x86 host computer and an ESP32-wifi-MQTT protocol, to adapt to the increased variety and number of nodes in the heterogeneous system. The non-contact slope crack visual measurement device (image crack meter) analyzes video streams from a network camera / network recorder, using image processing operations such as ROI-grayscale-binarization-erosion-dilation-contour recognition to extract crack areas and calculate the total number of pixels in the target contour area. This quantitatively reflects the size and expansion trend of the cracks, serving as a supplementary image indicator for slope surface deformation. For example... Figure 5 The image shown depicts the crack image recognition results from a landslide simulation experiment, illustrating the initial appearance, growth, and eventual disappearance of two cracks on the slope. Figure 6 As shown, the data includes rainfall (jy), shallow water (q1, q2), deep water (s1, s2), and soil pressure (pr) data obtained from monitoring points 1 and 2. Image cracks (cr) can be used as a supplement to the wire displacement (wy) to reflect the relationship between the surface deformation cr and the internal state of the slope (q2, s2) in a given area (monitoring point 2).

[0050] The "edge" layer uses a Raspberry Pi single-board computer as the edge computing node, and realizes the functions of multi-sensor data integration, lightweight early warning model operation, and online debugging of image processing parameters through the following four processes.

[0051] Process 1: Acquire data from the "end" layer sensors. Access the MySQL database on the x86 host computer and synchronize the latest data on moisture content, soil pressure, rainfall, and surface wire displacement every second to ensure compatibility with the existing landslide monitoring system.

[0052] Process 2: Acquire sensor data at the "end" layer. Based on the MQTT proxy server (Mosquitto software), subscribe to MQTT messages published by sensors, including ESP image crack gauges, soil moisture meters, wire displacement gauges, box inclinometers, and rainwater tank level gauges, and store the data in a local MariaDB database (an embedded version of a MySQL database).

[0053] Process 3: Preprocess and send landslide experiment data. Query the latest 60 seconds of data from each table, remove erroneous values, correct sensor dimensions, take the median, package in JSON format, and publish via MQTT protocol.

[0054] Process 4: Edge Model Inference. Based on the TensorFlow-Lite environment, the lightweight early warning model is inferred to achieve near-field early warning, avoiding network transmission delays and improving the timeliness of early warnings. Edge model inference includes two parts: hazard calculation, used to directly output the current hazard level based on surface deformation; and hazard assessment, used to predict the hazard level based on rainfall, water content, and soil pressure data when some surface deformation sensors fail.

[0055] This invention constructs a landslide hazard definition and assessment method based on multi-source sensing, realizing real-time quantification and intelligent early warning of slope stability. For example... Figure 7 The figure shows the results calculated using the hazard definition expression, describing the values ​​(wy_value, cr_value) of the guy wire displacement (wy) and the image crack (cr), and the energy curves (wy_energy, cr_energy) reflecting the trend of value changes. The method focuses on surface deformation characteristics, fusing the results of guy wire displacement sensors and visual crack identification, such as... Figure 6 As shown in the (cr ordinate subplot), two cracks appeared sequentially over time. The cracks initially grew slowly, then detached from the slope after reaching a critical value and disappeared. A weighted superposition method after unified normalization was used to define the current hazard index of the slope. Specifically, the hazard level r caused by the wire displacement was... s Risk r of image crack extraction c After normalizing the norms, a weighted average is performed to obtain the total hazard R, where the weights w1=w2=1. This hazard definition method has significant advantages: it can identify the changing trends and abrupt changes before the physical readings of sensors reach their extreme values, thus achieving proactive early warning of hazardous conditions. Especially in the early stages of landslide evolution, this method can effectively identify dangerous signs that "have not yet reached their peak but have already shown a sudden change trend."

[0056] Furthermore, such as Figure 8As shown, this invention also designs a landslide hazard assessment neural network model that integrates control theory. This model is built upon a feedforward neural network structure, achieving a nonlinear mapping from input (external causes and internal states) to output (surface hazard) through a multi-layer fully connected network. Based on this, a key innovative mechanism is proposed: dynamically injecting real-collected soil state quantities (such as moisture content and soil pressure) into the middle layer of the network, constructing a hybrid supervision mechanism of "measured state embedding," thereby improving the model's prediction accuracy and convergence speed. The model adopts a multi-channel input structure, receiving rainfall information (external input) and moisture content and soil pressure (internal states) respectively, and injecting state correction information on multiple paths, enabling the network to possess both data-driven characteristics and physical mechanism constraints during training and inference. Furthermore, to achieve stable operation on landslide experimental platforms or edge device environments, the system prunes and optimizes the landslide hazard assessment model designed in this paper, retaining the key encoding-decoding structure while compressing the parameter scale. Figure 9 As shown, by comparing the true value model with the lightweight control theory model and the BP model, the pruned model retains the accuracy of hazard inference, enhances the interpretability of the model for landslide evolution, and has good edge deployment capability.

[0057] The "cloud" layer employs multiple independent hardware devices combined with network infrastructure to simulate cloud service functions: landslide experimental data storage and interaction, online data processing and analysis, early warning model training and lightweighting, and auxiliary tools for the landslide experimental field. The specific implementation is as follows: Data storage and interaction in the cloud are achieved using a Linux operating system, the InfluxDB time-series database, and the Grafana data dashboard. MariaDB / InfluxDB databases, a Jupyter Notebook / Lab environment, Python, and Shell scripts are used to achieve online real-time processing and analysis of landslide experimental data. Data cleaning mainly includes sensor calibration, outlier and missing value removal, missing value imputation, time resampling, and time window filtering. The Tensorflow framework is used to design and train the model, Tensorflow Lite is used for model quantization and pruning, and the scp command is used to remotely deploy the model to edge computing nodes. The specific training environment is an Ubuntu system. Finally, landslide data and images are transmitted to the user's browser for display via an Apache web server and the Grafana data dashboard.

[0058] To ensure stable operation on landslide experimental platforms or edge device environments, the landslide hazard assessment model designed in this paper was tailored and optimized. While retaining the key encoder-decoder structure, the parameter scale was compressed, improving the model's operating efficiency on embedded platforms (such as Raspberry Pi and NVIDIA Jetson). Through structural compression and lightweight modification, the network can be deployed and run on edge computing nodes such as Raspberry Pi, enabling real-time data access and hazard assessment. It possesses rapid response and low-latency early warning capabilities under a cloud-edge-device collaborative architecture, making it suitable for the construction of intelligent monitoring systems in landslide simulation experimental platforms and actual high-risk landslide areas.

[0059] The above-described specific embodiments are used to explain and illustrate the present invention, not to limit it. Any modifications and alterations made to the present invention within the spirit and scope of the claims fall within the protection scope of the present invention. The above descriptions are merely preferred embodiments of the present invention; therefore, all equivalent changes or modifications made to the structures, features, and principles described in the claims of this patent application are included within the scope of this patent application.

Claims

1. A landslide disaster cloud-edge-end monitoring and early warning scenario simulation test platform, characterized in that, include: The landslide simulation test device is used to simulate landslides under different slopes, rainfall intensities, and groundwater seepage scenarios. The end-layer multi-source monitoring and acquisition system is installed in the landslide simulation test device and used for multi-source monitoring data acquisition during landslide simulation to monitor landslide data; The edge early warning and monitoring system receives landslide data from the multi-source monitoring and acquisition system at the end layer and then obtains the landslide hazard level through model processing. The cloud computing simulation system acquires landslide data and landslide hazard levels transmitted by the edge early warning and monitoring system for online processing and visualization, while also training and lightweight deployment of models in the edge early warning and monitoring system.

2. The landslide disaster cloud-edge-end monitoring and early warning scenario simulation test platform according to claim 1, characterized in that: The landslide simulation test device includes a soil-rock bearing box, a lifting and bracing mechanism, an artificial rainfall mechanism, and a porous ceramic pipe. One end of the bottom surface of the soil-rock bearing box is supported by the lifting and bracing mechanism, and the slope of the soil-rock bearing box is changed by raising and lowering the lifting and bracing mechanism. The interior of the soil-rock bearing box is lined with a layer of gravel and a layer of soil-rock as simulated soil from bottom to top. The artificial rainfall mechanism is installed directly above the soil-rock bearing box to carry out rainfall of different intensities. The porous ceramic pipe is laid on the inner bottom surface of the soil-rock bearing box and located below the gravel layer. The porous ceramic pipe is connected to an external water pump to simulate groundwater seepage conditions.

3. The landslide disaster cloud-edge-end monitoring and early warning scenario simulation test platform according to claim 2, characterized in that: The aforementioned multi-source monitoring and acquisition system includes a contact slope measurement unit and a non-contact slope measurement unit. Several monitoring points, not close to the soil bearing box, are arrayed at intervals on the simulated soil within the soil bearing box. The contact slope measurement unit includes several sensor arrays, each arranged in the soil layer below one of its monitoring points. Each sensor array includes, from top to bottom, a pull-wire displacement meter, two soil moisture meters, and one earth pressure meter. The contact slope measurement unit also includes a rain gauge, a three-dimensional underground displacement meter, and a shear wave velocity meter. The rain gauge is arranged in the space between the simulated soil and the artificial rainfall mechanism. The three-dimensional underground displacement meter and the shear wave velocity meter are installed in the middle of the inner bottom surface of the soil bearing box. The non-contact slope measurement unit includes a microprocessor equipped with an image crack meter and a network camera. The microprocessor is electrically connected to the network camera, which is installed on the side of the soil-rock bearing box and facing the upper surface of the simulated soil. The image crack meter is installed in the microprocessor and processes the images of the simulated soil surface captured by the network camera during a landslide to identify crack information of the soil-rock cracks during the landslide.

4. The landslide disaster cloud-edge-end monitoring and early warning scenario simulation test platform according to claim 3, characterized in that: When processing simulated soil surface images, if slope cracks exist in the simulated soil surface image, the image crack meter first converts the simulated soil surface image into a grayscale image by sequentially using masking, grayscale conversion, and histogram equalization. Then, it further processes the grayscale image by sequentially using erosion and dilation morphological operations to obtain an optimized crack image. Then, it identifies the contours of each slope crack in the optimized crack image, retains slope cracks with pixels in the contours greater than a preset threshold, and uses the total number of pixels of each retained slope crack as the crack size of the current simulated soil surface image, i.e., as crack information.

5. The landslide disaster cloud-edge-end monitoring and early warning scenario simulation test platform according to claim 3, characterized in that: The edge early warning monitoring system includes a host computer and edge computing nodes. The host computer receives and stores landslide data transmitted from the contact slope measurement unit of the edge multi-source monitoring and acquisition system. It also performs preprocessing and landslide hazard calculation before transmitting the data to the edge computing nodes and the cloud computing simulation system. The landslide hazard is obtained through the landslide hazard assessment and early warning model deployed in the edge computing nodes. The edge computing nodes receive crack information transmitted by the microprocessor of the non-contact slope measurement unit and simulated soil surface images transmitted by the network camera, and then transmit them to the cloud computing simulation system.

6. The landslide disaster cloud-edge-end monitoring and early warning scenario simulation test platform according to claim 5, characterized in that: When the host computer preprocesses the landslide data, it sequentially performs outlier removal, sensor error correction, and filtering operations.

7. The landslide disaster cloud-edge-end monitoring and early warning scenario simulation test platform according to claim 5, characterized in that: The specific landslide hazard R is as follows: R=w1×(r s ) n +w2×(r c ) n r b =(d b ×v b 2 ) / 2+d b ×a b ×v b ,b=s、c Where w1 and w2 represent the first and second weighting coefficients, respectively; (r s ) n and (r) c ) n The respective values ​​represent the risk level r of the wire displacement. s and slope crack risk level r c The normalized value of d; b v b and a b These represent the current value, rate of change, and acceleration of the wire displacement or crack information, respectively.

8. The landslide disaster cloud-edge-end monitoring and early warning scenario simulation test platform according to claim 7, characterized in that: When the landslide hazard R exceeds the preset hazard threshold, it is determined that the simulated soil has a landslide hazard.

9. The landslide disaster cloud-edge-end monitoring and early warning scenario simulation test platform according to claim 5, characterized in that: The landslide hazard assessment and early warning model also includes a feedforward neural network, which takes into account rainfall and soil moisture content and earth pressure data obtained from various monitoring points, and outputs slope deformation-related hazard levels to provide further early warning of landslide hazards.

10. The landslide disaster cloud-edge-end monitoring and early warning scenario simulation test platform according to claim 5, characterized in that: The cloud computing simulation system includes several independent hardware computing nodes, which back up and display the data transmitted from the edge early warning and monitoring system, train and lightweight deploy the landslide hazard assessment and early warning model in the edge early warning and monitoring system, and provide remote visualization functions through the data dashboard.

Images