Distributed slope monitoring system based on edge-cloud collaborative intelligent adaptive decision-making
The distributed slope monitoring system, which uses edge-cloud collaborative intelligent adaptive decision-making, solves the problems of high energy consumption, high false alarm and false alarm rates, and slow response in existing slope monitoring systems. It achieves low-power, accurate monitoring and second-level early warning, improving the system's real-time performance and robustness.
Patent Information
- Application Number
- CN202511206626.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-27
- Publication Date
- 2026-11-10
- Estimated Expiration
- 2045-08-27
AI Technical Summary
Existing slope monitoring systems suffer from problems such as rigid fixed sampling frequencies, lack of local computing power, inability to dynamically adapt to environmental changes, resulting in high energy consumption, high false alarm and false alarm rates, slow response, and strong network dependence.
The distributed slope monitoring system adopts edge-cloud collaborative intelligent adaptive decision-making. By introducing an edge computing framework and adaptive strategy engine at the sensor node level, and combining cloud-edge collaborative model distribution, it realizes local feature extraction and lightweight model inference, dynamically adjusts the sampling rate and transmission mode, and implements multi-factor adaptive strategies.
It achieves low-power, accurate monitoring and second-level early warning, reducing system power consumption by 52.7%, improving early warning timeliness, reducing false alarm rate by 40%, enhancing network robustness, and reducing cloud latency and communication load.
Smart Images

Figure CN120812055B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the cross-technical field of monitoring and early warning of rock and soil geological disasters (especially slope / landslide monitoring) and Internet of Things (IoT) edge intelligence technology, and particularly to a distributed, intelligent, low-power slope displacement monitoring system and a distributed slope monitoring system based on edge-cloud collaborative intelligent adaptive decision-making. Background Technology
[0002] Frequent landslides cause enormous loss of life and property, creating an urgent need for automated, intelligent, and cost-effective monitoring and early warning solutions for property owners (transportation, land resources, mining, hydropower). In the field of rock and soil geological hazard monitoring and early warning (especially slope / landslide monitoring), traditional monitoring methods mainly rely on manual inspections and fixed instruments. Manual inspections are inefficient, subjective, risky, and difficult to achieve real-time and continuous monitoring. Fixed instruments (such as inclinometers, crack gauges, and GNSS reference stations) are mostly point-based or small-scale monitoring, with complex wiring, high costs, high power consumption (especially for GNSS), and difficult maintenance.
[0003] With the rise of wireless sensor network applications, sensor nodes based on wireless technologies such as LoRa, NB-IoT, and Zigbee have been deployed to achieve wireless data transmission through the "sensing-transmission" mode. However, the main function of the "sensing-transmission" mode is still data acquisition and relay, with most of the computing and processing completed in the cloud or central server. The nodes have low intelligence, limited functionality, and fixed sampling frequency.
[0004] The current main development trend of intelligent sensing and monitoring is the evolution from "sensing and transmission" to the integrated "sensing-computing-transmission," that is, performing more localized intelligent processing and analysis at sensor nodes or near-node gateways to reduce transmission burden and improve response speed. The limitations of single-sensor data are obvious; therefore, fusing multi-dimensional data such as displacement (GNSS / tiltmeter / strain gauge), environment (rainfall, temperature and humidity), and inducing factors (vibration) to construct a more comprehensive steady-state assessment model (physical model + data-driven AI model) has become the mainstream direction.
[0005] Current sensors operating in the field for extended periods face high requirements for drift prevention, lightning protection, waterproofing, corrosion resistance, and earthquake resistance. Complex environmental conditions also play a role.
[0006] GNSS and wireless communication are required. The system must remain stable and reliable under harsh conditions. Low-cost MEMS sensors suffer from high noise and limited stability (especially inclinometers). Ensuring the measurement accuracy of critical data (micro-displacements) and the accuracy assessment and compensation of multi-source data fusion are challenges, given cost constraints and harsh environments. Existing system strategy adjustments mostly rely on simple threshold rules. How to intelligently and personally adjust node behavior based on data-driven principles, environmental changes, and learning mechanisms remains a technical challenge.
[0007] The closest existing technical solution to this invention is a low-power slope monitoring sensor network system based on LoRaWAN.
[0008] A preset uniform sampling frequency (e.g., once every 15 minutes) is used, which is temporarily increased to once every 5 minutes only when the rainfall exceeds the threshold (based on simple event triggering). Raw data is directly uploaded to the cloud, and all data filtering, displacement conversion, and anomaly detection are completed in the cloud.
[0009]
[0010] Its core shortcomings lie in the lack of edge intelligence and weak policy adaptation capabilities, resulting in low energy efficiency, slow response, and strong network dependence.
[0011] In summary, the existing technical solutions have the following technical defects:
[0012] (1) A simple triggering strategy based on a fixed sampling frequency or a single event (such as rainfall exceeding a threshold) is adopted without comprehensively considering factors such as displacement rate, battery status, and data noise.
[0013] (2) The monitoring strategy is rigid and cannot dynamically adapt to changes in the environment and the status of nodes.
[0014] (3) The early warning mechanism is singular, with a high rate of false alarms and missed alarms. The early warning relies entirely on fixed displacement thresholds set in the cloud and does not integrate multi-source data (such as environment and vibration) for joint analysis.
[0015] (4) Large-scale engineering slope monitoring points are numerous and widely distributed. The market needs systems that are lower in cost (high cost-effectiveness per node), easy to deploy (wireless and cordless), and simple to maintain (long-term maintenance-free or low-maintenance) to support large-scale applications. Self-sufficiency in solar / vibration energy is a key selling point. The market needs the ability to detect small, slow but potentially risky displacement trends in advance, buying time for prevention and control.
[0016] (5) Due to the lack of local computing power and intelligent decision-making mechanism of the sensing nodes, the system must upload all raw data to the cloud for processing, resulting in high communication load and cost; at the same time, relying on fixed sampling strategies and static threshold alarms, it is impossible to dynamically optimize the monitoring behavior according to displacement change trends, environmental disturbances and power status, resulting in energy efficiency imbalance (battery life shortened by 40% to 60%), key data omission (transient feature loss >70%), response lag (>300 seconds) and high false alarm rate (>45%); and in weak network environment (packet loss rate >15%), the monitoring function is paralyzed because the model relies entirely on cloud operation. Summary of the Invention
[0017] The purpose of this invention is to provide a distributed slope monitoring system based on edge-cloud collaborative intelligent adaptive decision-making. By introducing an edge computing framework and adaptive strategy engine at the sensor node level, combined with cloud-edge collaborative model distribution, a closed loop of "perception-decision-execution" is constructed, directly overcoming the bottlenecks of existing technologies in high real-time performance, ultra-low power consumption, and weak network robustness. The technological evolution path is clear and aligns with the industry trend towards "decentralized intelligence." By endowing nodes with edge intelligent processing (local feature extraction and lightweight model inference) and a multi-factor adaptive strategy engine (dynamically adjusting sampling rate, accuracy, and transmission mode), combined with a cloud-edge collaborative mechanism, the system fundamentally overcomes the deficiencies of existing technologies in real-time performance, power consumption, weak network robustness, and early warning accuracy, achieving low-power, accurate monitoring and second-level early warning.
[0018] A first aspect of the present invention is to provide a distributed slope monitoring system based on edge-cloud collaborative intelligent adaptive decision-making, comprising:
[0019] Multiple intelligent sensing unit (ISU) nodes and data processing and intelligent analysis units are deployed at key locations on the slope.
[0020] Each Intelligent Sensor Unit (ISU) node transmits data wirelessly to an edge gateway or directly to a cloud platform via a self-organizing network for slope monitoring based on a local adaptive monitoring strategy. This local adaptive monitoring strategy involves the ISU node or edge device analyzing the collected data using a built-in lightweight decision model to determine if any anomalies have occurred or if a warning threshold has been reached. The local adaptive monitoring strategy features a cloud / edge feedback mechanism, which receives model update or strategy adjustment instructions from the edge gateway or cloud platform. The cloud then distributes the optimized model, decision rules, and configuration parameters to the edge devices and ISU nodes, enabling the monitoring system to self-evolve and iterate.
[0021] The data processing and intelligent analysis unit includes an edge intelligence module, a cloud gateway, and an edge gateway. The edge gateway, as an intermediate layer, is used for protocol conversion, data aggregation, temporary storage, and preliminary analysis. The cloud gateway provides computing and storage resources, and trains more complex AI models based on historical and real-time data to perform pattern recognition, predictive analysis, and anomaly detection tasks. The edge intelligence module includes multiple edge computing nodes and a built-in lightweight AI inference unit, located at the edge, to implement edge intelligent processing. For data that does not require immediate alarm, high-complexity processing is performed at the edge, including feature extraction, anomaly detection, and event classification.
[0022] Preferably, the lightweight decision model includes a rule engine, decision tree, or micro neural network, and runs a lightweight state machine or decision tree logic in the microcontroller module of the intelligent sensing unit (ISU) node. It switches the local adaptive monitoring strategy based on built-in rules or learning models, and the key parameter thresholds of the local adaptive monitoring strategy are continuously optimized through model learning.
[0023] Preferably, the lightweight AI inference unit runs a compressed neural network model to perform image recognition and time series analysis tasks.
[0024] Preferably, the edge intelligence module adopts a real-time stream processing framework to perform real-time data pipeline processing.
[0025] Preferably, the intelligent sensing unit (ISU) node includes:
[0026] The sensing core module includes one or more of a MEMS accelerometer, a GNSS receiver chip, and a strain gauge; wherein the MEMS accelerometer is used to measure tilt angle changes and reflect displacement trends, the GNSS receiver chip is used to acquire absolute position changes, and the strain gauge is used to monitor local deformation.
[0027] An environmental sensing module includes one or more of a temperature and humidity sensor, a rain gauge, and a vibration sensor; wherein the temperature and humidity sensor is used to compensate for sensor drift, the rain gauge is used to sense key factors that trigger monitoring, and the vibration sensor is used to sense disturbance events.
[0028] The microcontroller module, including the core processing unit (MCU), performs data acquisition, preprocessing, and local computation based on computing power;
[0029] The wireless communication module uses a low-power wide area network (LPWAN) or self-organizing network protocol for communication.
[0030] The energy supply module includes a short-term power supply unit, a long-term power supply unit, and a power management unit (PMU). The short-term power supply unit includes a battery; the long-term power supply unit includes an energy harvesting module for long-term operation; and the power management unit (PMU) is used to implement intelligent power consumption management of the monitoring system.
[0031] Small-capacity memory, configured locally, is used to cache critical data and adaptive strategy parameters.
[0032] Preferably, the low-power wide area network (LPWAN) is LoRaWAN, NB-IoT, Zigbee, or BLEMesh LPWAN.
[0033] Preferably, the energy harvesting module includes a solar panel or a vibration energy harvesting unit.
[0034] Preferably, the edge intelligence module includes:
[0035] The feature extraction unit is used to perform data filtering, noise reduction, displacement calculation, and anomaly detection on the multiple intelligent sensing unit (ISU) nodes or edge gateways.
[0036] Lightweight model inference unit, used to deploy lightweight ML models for:
[0037] The trend prediction unit is used to predict short-term displacement trends based on historical displacement data, including: using a lightweight version of simple linear regression and RNN or LSTM to predict displacement trends over hours or days.
[0038] The state recognition unit is used to make a preliminary judgment on whether the current state of the slope is stable, slowly deforming, rapidly deforming, or showing signs of impending landslide.
[0039] Anomaly alarm unit is used to detect displacement mutations or abnormal events that significantly deviate from the normal pattern based on statistical thresholds or model inference.
[0040] Preferably, the strategy adjustment criteria for slope monitoring based on the local adaptive monitoring strategy include:
[0041] The dynamic adjustment of the displacement change rate includes: dynamically adjusting the sampling frequency based on a first adjustment strategy by calculating the displacement change rate in real time, thereby dynamically adjusting the displacement change rate. The first adjustment strategy includes low-frequency sampling when the displacement change rate is low to save power consumption and high-frequency sampling when the displacement change rate is high to capture details.
[0042] The strategy adjustment is based on environmental triggering events, including: when one or more of the following are detected by the rain gauge, the vibration sensor, and the temperature and humidity sensor, the monitoring frequency and accuracy are automatically increased based on a second adjustment strategy. The second adjustment strategy includes switching the GNSS receiver chip from single frequency to dual frequency and putting the MEMS accelerometer into a high-precision mode.
[0043] Adjusting the data fluctuation level or noise by means of strategy includes: adjusting the data fluctuation level or noise by calculating the standard deviation or statistical characteristics of the data, and adjusting the data fluctuation level or noise by means of a third adjustment strategy, wherein the third adjustment strategy includes appropriately reducing the data precision and reducing the data transmission frequency during the stable period, for example, reducing the data precision by means of bit compression; and restoring high data precision and high frequency data transmission when the data fluctuation intensifies.
[0044] The battery status is adjusted strategically, including: intelligently balancing power consumption and monitoring requirements based on the remaining power and energy harvesting efficiency through a fourth adjustment strategy, wherein the fourth adjustment strategy includes forcing the battery to enter the most power-saving mode when the battery enters a low power state.
[0045] Preferably, the cloud gateway and edge gateway include:
[0046] The model training and optimization unit is used to receive data uploaded from various nodes and train more complex models on the edge gateway or in the cloud; wherein, the data uploaded from various nodes is raw data and / or preprocessed data, and the more complex model is a cloud-based deep learning model;
[0047] A multi-source data fusion model is used to integrate displacement, environmental, and historical data to establish a more accurate slope stability evaluation model; wherein, the slope stability evaluation model is a machine learning model or a composite model of physical model + AI;
[0048] Long-term time-series forecasting models are used for medium- to long-term deformation trend prediction.
[0049] An adaptive policy generator is used to generate optimized adaptive policy update packages based on a global analysis of data from all nodes, and push them to the relevant ISU nodes.
[0050] Early warning models are used to provide comprehensive early warnings with higher confidence based on the fusion of information from multiple nodes and spatial correlations;
[0051] The model push update unit is used to push incremental or full updates of the trained model to edge nodes, cloud gateways, and edge gateways.
[0052] The data filtering and transmission optimization unit is used to utilize local analysis results to transmit only data that has changed significantly, anomaly reports, important feature values, or compressed model outputs.
[0053] The beneficial effects of the system of the present invention:
[0054] This invention directly overcomes the three core shortcomings of existing slope monitoring systems through an edge-intelligent dynamic decision-making mechanism and a cloud-based collaborative optimization architecture:
[0055] (1) To address the contradiction between energy consumption and accuracy caused by fixed sampling strategies, adaptive sampling driven by displacement change rate (e.g., frequency upscaling to 10Hz when the rate is >0.5mm / h) and sensor mode switching triggered by environmental events (e.g., enabling GNSS dual-frequency when rainfall is >30mm / 24h) significantly reduce system power consumption (measured average power consumption of nodes is reduced by 52.7%) while ensuring a disaster event capture rate of >95%, thus overcoming the monitoring cycle limitation caused by battery life limitations in traditional systems.
[0056] (2) By breaking through the high latency bottleneck of centralized cloud processing, a lightweight ML model (≤50KB TinyLSTM) is deployed on a resource-constrained MCU (such as ARM Cortex-M4) to achieve local millisecond-level response (latency <200ms) for displacement trend prediction (error <5%) and anomaly identification (recall rate >92%), which significantly improves the timeliness of the imminent slippage warning compared to cloud analysis (>10s);
[0057] (3) To address the environmental adaptability limitations of static strategies, a dynamic strategy update package generated by a cloud-based multi-source data fusion model (integrating geological history, real-time rainfall, and spatial displacement correlation) is used to continuously optimize the adaptive parameters of edge nodes (e.g., monthly optimization rate > 5%), reducing the false alarm rate of the system by 40% in complex field environments (compared to manually set threshold schemes). Simultaneously, through...
[0058] Mesh networking and dual-mode energy harvesting (solar energy + vibration energy) ensure network robustness (disconnection rate <0.5%). Attached Figure Description
[0059] To more clearly illustrate the technical solutions in the specific embodiments or related technologies of the present invention, the drawings used in the description of the specific embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.
[0060] Figure 1 This is a schematic diagram illustrating the structure and working principle of a distributed slope monitoring system based on edge-cloud collaborative intelligent adaptive decision-making, according to an embodiment of the present invention.
[0061] Figure 2 A schematic diagram illustrating the principle of slope monitoring strategy adjustment based on local adaptive monitoring strategy according to an embodiment of the present invention;
[0062] Figure 3 This is a schematic diagram illustrating the detailed working principle of edge intelligent data processing according to an embodiment of the present invention;
[0063] Figure 4 This is a schematic diagram illustrating the detailed working principle of the cloud-based global optimization and strategy distribution system provided in an embodiment of the present invention;
[0064] Figure 5 This is a structural diagram of an electronic device provided according to an embodiment of the present invention. Detailed Implementation
[0065] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0066] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation.
[0067] The invention is constructed and operated in a specific orientation and therefore should not be construed as a limitation of the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.
[0068] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.
[0069] The technical effects achieved by this invention are shown in Table 1 below.
[0070] Table 1
[0071]
[0072] The first aspect of this invention is to provide a distributed slope monitoring system based on edge-cloud collaborative intelligent adaptive decision-making, which is a distributed, intelligent, and low-power slope displacement monitoring system, comprising:
[0073] Multiple intelligent sensing unit (ISU) nodes and data processing and intelligent analysis units are deployed at key locations on the slope.
[0074] Each Intelligent Sensor Unit (ISU) node transmits data wirelessly to an edge gateway or directly to a cloud platform via a self-organizing network for slope monitoring based on a local adaptive monitoring strategy. The core mechanism is that the ISU node is not merely a data collector, but an edge computing node with preliminary intelligent decision-making capabilities. The local adaptive monitoring strategy involves the ISU node or edge device analyzing the collected data based on a built-in lightweight decision model to determine whether anomalies have occurred or warning thresholds have been reached. The lightweight decision model includes a rule engine, decision tree, or micro-neural network. The microcontroller module of the meta-ISU node runs a lightweight state machine or decision tree logic, which switches the local adaptive monitoring strategy based on built-in rules or learning models. The key parameter thresholds of the local adaptive monitoring strategy are continuously optimized through model learning. The local adaptive monitoring strategy has a cloud / edge feedback mechanism, which is used to receive model update or strategy adjustment instructions issued by the edge gateway or cloud platform (such as raising the monitoring level when the prediction model indicates that the risk in the area has increased). The cloud sends the optimized model, decision rules and configuration parameters to the edge devices and ISU nodes to complete the self-evolution and iteration of the monitoring system.
[0075] The data processing and intelligent analysis unit includes an edge intelligence module, a cloud gateway, and an edge gateway. The edge gateway acts as an intermediate layer, used for protocol conversion, data aggregation, temporary storage, and preliminary analysis. The cloud gateway provides virtually unlimited computing and storage resources, and enables the training of more complex AI models (such as CNN, LSTM, and Transformer) based on historical and real-time data for pattern recognition, predictive analysis, and anomaly detection tasks. The edge intelligence module includes multiple edge computing nodes and a built-in lightweight AI inference unit, located at the edge, for implementing edge intelligent processing. For data that does not require immediate alarm, high-complexity processing is performed at the edge, including feature extraction, anomaly detection, and event classification. The lightweight AI inference unit runs a compressed neural network model to complete image recognition and time-series analysis tasks. The edge intelligence module adopts a real-time stream processing framework for real-time data pipeline processing.
[0076] In a preferred embodiment, the intelligent sensing unit (ISU) node includes:
[0077] The sensing core module includes one or more of a MEMS accelerometer, a GNSS receiver chip, and a strain gauge; wherein, the MEMS accelerometer is used to measure the tilt angle change and reflect the displacement trend, the GNSS receiver chip is used to acquire the absolute position change, and the strain gauge is used to monitor local deformation. In this embodiment, a single sensor or multi-sensor fusion is selected according to cost and accuracy requirements.
[0078] An environmental sensing module includes one or more of a temperature and humidity sensor, a rain gauge, and a vibration sensor; wherein the temperature and humidity sensor is used to compensate for sensor drift, the rain gauge is used to sense key factors that trigger monitoring, and the vibration sensor is used to sense disturbance events.
[0079] The microcontroller module, including the core processing unit (MCU), has a certain computing power (such as the ARM Cortex-M series) and is used for data acquisition, preprocessing and local calculation.
[0080] The wireless communication module uses a low-power wide-area network (LPWAN) or self-organizing network protocol for communication; the wireless communication module in this embodiment has long-distance and low-power transmission capabilities.
[0081] The energy supply module includes a short-term power supply unit, a long-term power supply unit, and a power management unit (PMU). The short-term power supply unit includes a battery; those skilled in the art should know that any suitable battery can be used, and all are within the scope of protection of this invention. The long-term power supply unit includes an energy harvesting module for long-term operation. The power management unit (PMU) is used to implement intelligent power consumption management of the monitoring system.
[0082] Small-capacity memory, configured locally, is used to cache critical data and adaptive strategy parameters.
[0083] In a preferred embodiment, the key locations of the slope include one or more of the potential sliding surface, monitoring profile line, cracks, and the top and toe of the slope.
[0084] As a preferred embodiment, the low-power wide-area network (LPWAN) is LoRaWAN, NB-IoT, Zigbee, or...
[0085] BLEMesh Low Power Wide Area Network (LPWAN)
[0086] In a preferred embodiment, the energy harvesting module includes a solar panel or a vibration energy harvesting unit.
[0087] In a preferred embodiment, the edge intelligence module includes:
[0088] The feature extraction unit is used to perform data filtering (such as Kalman filtering), noise reduction (such as wavelet transform), displacement calculation (calculated from tilt angle or GNSS differential results) and anomaly detection on the multiple intelligent sensing unit (ISU) nodes or edge gateways.
[0089] Lightweight model inference units are used to deploy lightweight ML models (such as TinyML, ONNXRuntime implementations on edge devices) for:
[0090] The trend prediction unit is used to predict the displacement trend in the short term based on historical displacement data. In this embodiment, a lightweight version of simple linear regression and RNN / LSTM is used to predict the displacement trend over several hours / days.
[0091] The state recognition unit is used to make a preliminary judgment on whether the current state of the slope is stable, slowly deforming, rapidly deforming, or a sign of impending landslide.
[0092] Anomaly alarm unit is used to detect displacement abrupt changes or abnormal events that significantly deviate from the normal pattern (based on statistical thresholds or model inference).
[0093] like Figure 1 As shown, this is a schematic diagram of the structure and working principle of a distributed slope monitoring system based on edge-cloud collaborative intelligent adaptive decision-making, which is a preferred embodiment.
[0094] (I) Overview of a Distributed Slope Monitoring System Based on Edge-Cloud Collaborative Intelligent Adaptive Decision Making
[0095] Figure 1 The diagram illustrates the working principle of a typical Edge-AIIoT (Artificial Intelligence of Things) early warning system. The system integrates multiple technologies, including edge computing, IoT, adaptive decision-making, wireless sensor networks, cloud computing, and machine learning, to achieve a closed-loop processing flow from data perception to intelligent early warning. Its core features lie in "edge-cloud collaboration" and "local adaptive decision-making," enabling efficient and low-latency initial decision-making in resource-constrained edge environments, while leveraging the powerful computing capabilities of the cloud for global optimization and model iteration.
[0096] This system is applied to geological disaster monitoring scenarios (such as landslides and earthquakes) to achieve real-time perception, intelligent judgment, and multi-level early warning of abnormal events.
[0097] (II) Detailed Analysis of Each Step
[0098] 1. Deploy ISU nodes
[0099] ISU (Intelligent Sensing Unit) refers to the intelligent sensing unit, which is the lowest-level data sensing device in the system. Each ISU node typically includes a sensor, a microprocessor, a wireless communication module, and an independent power source (such as a battery or solar power).
[0100] Detailed explanation:
[0101] (1) Hardware components: including various sensors (such as accelerometers, humidity sensors, cameras, sound sensors, etc.), main control MCU (such as ARM Cortex-M series), LoRa / Wi-Fi / 4G / 5G communication modules, power management unit, etc.
[0102] (2) Deployment strategy: Nodes should be scientifically deployed according to the geographical characteristics, signal coverage, energy consumption requirements, etc. of the monitoring area. Common deployment methods include grid, strip or adaptive deployment.
[0103] (3) Software initialization: Each node is pre-installed with a lightweight operating system (such as FreeRTOS or Zephyr) and a basic communication protocol stack to complete self-test, network registration and parameter configuration.
[0104] 2. Data Acquisition and Environmental Sensing
[0105] ISU nodes continuously or periodically collect raw data from the environment, including various modal information such as images, vibration, temperature, humidity, and air pressure.
[0106] Detailed explanation:
[0107] (1) Multimodal sensing: Supports data fusion from heterogeneous sensors, improving the comprehensiveness and reliability of the data.
[0108] (2) Adaptive sampling frequency: The sampling rate can be dynamically adjusted according to environmental changes to balance data accuracy and energy consumption.
[0109] (3) Preprocessing: Perform preliminary filtering, noise reduction and calibration at the nodes to improve data quality.
[0110] 3. Local adaptive strategy decision-making
[0111] In this step, nodes or edge devices analyze the collected data based on built-in lightweight decision models (such as rule engines, decision trees, and micro neural networks) to determine whether anomalies have occurred or whether warning thresholds have been reached.
[0112] Detailed explanation:
[0113] (1) Lightweight model deployment: Use TinyML, decision tables or state machines to achieve local intelligence and avoid complete reliance on the cloud.
[0114] (2) Adaptive threshold: The judgment threshold is dynamically adjusted based on historical data to avoid false alarms.
[0115] (3) Contextual awareness: Make comprehensive judgments by combining time, geographical location and environmental context.
[0116] 4. Is an immediate warning required?
[0117] This is a binary decision point: if the local judgment is "yes", a local alarm is triggered; if it is "no", the process proceeds to the next step for more in-depth data processing or uploading.
[0118] Detailed explanation:
[0119] (1) Low latency requirement: For emergency events such as earthquakes and landslides, local decision-making can respond within milliseconds to seconds, which is much faster than cloud feedback.
[0120] (2) Confidence assessment: The local model outputs the confidence level. If the confidence level is low, it will be handed over to the next step for further processing.
[0121] 5. Trigger local alarm
[0122] Once an immediate warning is deemed necessary, the system will issue an alert to nearby personnel or equipment via audible and visual alarms, local broadcasts, or short-range communication (such as ZigBee or BLE).
[0123] Detailed explanation:
[0124] (1) Multimodal alarm: Supports buzzer, LED, vibration motor, voice prompt, etc.
[0125] (2) Local linkage: can trigger local actuators, such as closing valves or starting backup power.
[0126] (3) Low power consumption design: The alarm circuit is normally in a sleep state and is only woken up when triggered.
[0127] 6. Edge intelligent processing
[0128] For data that does not require immediate alarm, the system will perform more complex processing at the edge, such as feature extraction, anomaly detection, and event classification.
[0129] Detailed explanation:
[0130] (1) Edge computing nodes: These are typically provided by edge servers or gateway devices and possess strong computing capabilities (such as...).
[0131] Jetson Nano, Huawei Atlas).
[0132] (2) Lightweight AI inference: Run compressed neural network models (such as Mobi leNet, SqueezeNet) to complete tasks such as image recognition and time series analysis.
[0133] (3) Real-time stream processing: Real-time data pipeline processing is performed using frameworks such as Apache Kafka and Filter.
[0134] 7. Data filtering and compression
[0135] To reduce transmission bandwidth and energy consumption, the system filters, compresses, and aggregates data, uploading only valuable information.
[0136] Detailed explanation:
[0137] (1) Lossless / lossy compression: using differential coding, Huffman coding, JPEG / LZ4 and other algorithms.
[0138] (2) Event-driven upload: Data is uploaded only when an anomaly or change is detected, avoiding the transmission of redundant information.
[0139] (3) Metadata extraction: Upload feature vectors instead of raw data to further reduce the amount of data transmitted.
[0140] 8. Wireless networking transmission
[0141] The processed data is transmitted to the edge gateway via a wireless sensor network (WSN) or uploaded directly to the cloud. The network topology can be star, mesh, or hybrid.
[0142] Nodes can form a mesh or tree-like self-organizing network, solving the problem of communication distance limitations of a single node and enhancing network robustness and coverage.
[0143] Choose a suitable low-power communication protocol to achieve energy-efficient and high-performance data relay transmission.
[0144] Ultimately, the data converges at the edge network, where the gateway has stronger computing power and a stable network (such as 4G / 5G or wired) to upload critical information to the cloud platform.
[0145] Detailed explanation:
[0146] (1) Communication protocols: Commonly used protocols include LoRaWAN, NB-IoT, Wi-FiHaLow, Zigbee, etc., which are selected according to distance, power consumption, and speed requirements.
[0147] (2) Routing optimization: Use routing protocols such as AODV and OLSR to achieve multi-hop transmission, thereby enhancing network coverage and reliability.
[0148] (3) Fault tolerance mechanism: Supports disconnection reconnection, data caching, and retransmission mechanisms to ensure that data is not lost.
[0149] 9. Edge Gateway / Cloud
[0150] The edge gateway, acting as an intermediate layer, is responsible for protocol conversion, data aggregation, temporary storage, and preliminary analysis. The cloud, on the other hand, provides virtually unlimited computing and storage resources.
[0151] Detailed explanation:
[0152] (1) Edge gateway functions: MQTT / CoAP to HTTP conversion, data cleaning, cache management, and rule engine execution;
[0153] (2) Cloud platform access: Access platforms such as AWS IoT, Azure IoT, and Alibaba Cloud IoT, and use Kafka, Spark, etc. for big data processing;
[0154] (3) Secure transmission: Use TLS / DTLS encryption, device authentication, and access control to prevent data leakage and tampering.
[0155] 10. Global Model Training and Optimization
[0156] In the cloud, the system uses historical and real-time data to train more complex AI models (such as CNN, LSTM, and Transformer) to perform tasks such as pattern recognition, predictive analysis, and anomaly detection.
[0157] Detailed explanation:
[0158] (1) Distributed training: Model training is performed on a GPU cluster using TensorFlow and PyTorch.
[0159] (2) Federated learning support: Federated learning is supported in some scenarios to protect edge data privacy.
[0160] (3) Automated Machine Learning (AutoML): Automatic parameter tuning and model structure search to improve model performance.
[0161] 11. Comprehensive early warning decision-making
[0162] The cloud platform integrates data from multiple nodes, time periods, and modalities to make more accurate early warning decisions, generating warning levels, impact ranges, and recommended measures.
[0163] Detailed explanation:
[0164] (1) Multi-source information fusion: Decision-making is based on the combination of meteorological data, geographic information system (GIS), historical event database, etc.
[0165] (2) Warning levels: such as blue, yellow, orange and red warnings, corresponding to different response levels.
[0166] (3) Visualization and push notification: Generate early warning reports and map visualizations, and push them to relevant personnel via SMS, APP, large screen, etc.
[0167] 12. Strategy / Model Update Push
[0168] The cloud distributes the optimized model, decision rules, configuration parameters, etc. to edge devices and ISU nodes to complete the system's self-evolution and iteration.
[0169] Detailed explanation:
[0170] (1) OTA upgrade mechanism: Supports wireless remote upgrades without affecting the normal operation of the device.
[0171] (2) Differential upgrade: Only the difference is transmitted, saving bandwidth.
[0172] (3) Version management and rollback: It has version control and fault rollback mechanisms to ensure system stability.
[0173] (III) System Technical Features and Advantages
[0174]
[0175]
[0176] This process Figure 1 The described system is a highly intelligent, adaptive, and evolvable edge AIoT early warning platform. Through multi-layered data processing and decision-making mechanisms, it achieves seamless collaboration from the edge to the cloud, ensuring real-time response to critical events while leveraging cloud resources for global optimization and continuous learning. This architecture has broad applicability and significant practical value in modern IoT applications.
[0177] like Figure 2 As shown, in a preferred embodiment, the strategy adjustment basis for slope monitoring based on the local adaptive monitoring strategy includes:
[0178] The dynamic adjustment of the displacement change rate includes: dynamically adjusting the sampling frequency based on a first adjustment strategy by calculating the displacement change rate in real time, thereby dynamically adjusting the displacement change rate. The first adjustment strategy includes low-frequency sampling when the displacement change rate is low to save power consumption and high-frequency sampling when the displacement change rate is high to capture details.
[0179] The strategy is adjusted based on environmental triggering events, including: when one or more of the following are detected by the rain gauge, the vibration sensor, and the temperature and humidity sensor, the monitoring frequency and accuracy are automatically increased based on a second adjustment strategy. The second adjustment strategy includes switching the GNSS receiver chip from single frequency to dual frequency and putting the MEMS accelerometer into a high-precision mode.
[0180] Adjusting the data fluctuation level or noise by means of strategy includes: adjusting the data fluctuation level or noise by calculating the standard deviation or statistical characteristics of the data, and adjusting the data fluctuation level or noise by means of a third adjustment strategy, wherein the third adjustment strategy includes appropriately reducing the data precision and reducing the data transmission frequency during the stable period, for example, reducing the data precision by means of bit compression; and restoring high data precision and high frequency data transmission when the data fluctuation intensifies.
[0181] The battery status is adjusted strategically, including: intelligently balancing power consumption and monitoring requirements based on the remaining power and energy harvesting efficiency through a fourth adjustment strategy, wherein the fourth adjustment strategy includes forcing the battery to enter the most power-saving mode when the battery enters a low power state.
[0182] This involves overriding local policies by receiving cloud policy updates.
[0183] In a preferred embodiment, the cloud gateway and edge gateway include:
[0184] The model training and optimization unit is used to receive data (raw / preprocessed) uploaded from various nodes and train more complex models (such as cloud-based deep learning models) at the edge gateway or in the cloud.
[0185] Multi-source data fusion model is used to integrate displacement, environmental, and historical data to establish a more accurate slope stability evaluation model (machine learning or physical model + AI).
[0186] Long-term time series prediction models are used for medium- to long-term deformation trend prediction (such as LSTM, Prophet).
[0187] The adaptive policy generator is used to generate optimized adaptive policy update packages based on a global analysis of data from all nodes, and push them to the relevant ISU nodes.
[0188] Early warning models are used to provide comprehensive early warnings with higher confidence based on information from multiple nodes (such as considering spatial correlation);
[0189] The model push update unit is used to push the trained model to edge nodes / gateways for incremental or full updates.
[0190] The data filtering and transmission optimization unit is used to utilize local analysis results to transmit only data that has changed significantly, anomaly reports, important feature values, or compressed model outputs, greatly reducing the amount of wireless data transmission.
[0191] like Figure 3 The diagram shown illustrates the detailed working principle of edge intelligent data processing.
[0192] (I) General Overview
[0193] This process Figure 3 This describes an advanced data processing and decision-making pipeline deployed on edge computing nodes (such as smart gateways, embedded AI devices, or edge servers). Its technical role is to perform real-time, intelligent, on-site processing of incoming raw IoT sensor data without relying on the cloud, enabling event detection, status recognition, and intelligent decision-making regarding the data's subsequent fate—whether to immediately trigger an alarm, upload it with high priority, or process it normally before uploading to save bandwidth.
[0194] This pipeline is key to realizing "edge intelligence." It directly pushes computing power down to the source of data generation, greatly reducing the pressure on network transmission and the cloud, and significantly lowering the system's response latency to critical events. Essentially, it is an "intelligent filter" and "decision-maker" embedded in the data stream.
[0195] (II) Step-by-step instructions
[0196] 1. Input of raw data
[0197] Function: This is the starting point of the pipeline, receiving multimodal raw data streams transmitted wirelessly or via wired means from each ISU (Intelligent Sensor Unit).
[0198] Technical details:
[0199] (1) Data format diversity: The input data formats are extremely diverse, depending on the sensor type, including:
[0200] A. Time series data: readings from vibration sensors (accelerometers), inclinometers, acoustic sensors, and temperature / humidity / pressure sensors, typically one-dimensional numerical sequences with equal or unequal intervals.
[0201] B. Image / Video Data: RGB or grayscale images and video streams (H.264 / H.265 encoded bitstreams or RAW frames) from surveillance cameras.
[0202] C. Spectrum data: Frequency domain data from acoustic or radio frequency sensors.
[0203] D. Point cloud data: A set of three-dimensional spatial points from a LiDAR (Light Detection and Ranging) system.
[0204] (2) Data Stream Characteristics: Data typically flows in continuously in the form of a stream, characterized by high throughput, disorder, and the possibility of duplication or loss. Therefore, edge nodes need to have powerful IO processing capabilities and data buffer management mechanisms.
[0205] (3) Protocol parsing: The edge node first needs to parse the data packet according to the communication protocol (such as MQTT, CoAP, Modbus, etc.) to extract the valid sensor ID, timestamp, and payload data.
[0206] (4) Design considerations: The stability of this part is the foundation of the entire system. A robust communication stack and buffer design are required to cope with data surges and prevent data loss.
[0207] 2. Filtering / Noise Reduction Processing
[0208] (1) Function: Perform preliminary cleaning of raw data, remove noise and interference, improve signal-to-noise ratio (SNR), and provide a high-quality data foundation for subsequent feature extraction and model inference.
[0209] (2) Technical details:
[0210] Necessity of the design: IoT sensors are usually deployed in harsh environments, and the raw data is easily affected by environmental interference (such as wind noise and electromagnetic interference), thermal noise of the sensor itself, and bit errors during transmission.
[0211] Commonly used algorithms (selected according to data type):
[0212] A. For time series data:
[0213] Employing low-pass / high-pass / band-pass filters: Using digital filters (such as IIR and FIR filters) to filter out noise at specific frequencies. For example, in vibration analysis, a low-pass filter is used to filter out high-frequency electromagnetic noise.
[0214] Alternatively, a moving average filter can be used: it is simple and effective, and can smooth out short-term fluctuations.
[0215] Alternatively, Kalman filtering can be used: an optimal estimation algorithm, particularly suitable for real-time estimation of the state of dynamic systems with Gaussian white noise, and widely used in navigation and tracking. Its variants, Extended Kalman Filter (EKF) and Unscented Kalman Filter (UKF), can be used for nonlinear systems.
[0216] Alternatively, median filtering can be used: it is particularly effective in eliminating impulse noise (salt and pepper noise).
[0217] B. For image data:
[0218] Gaussian blur is used to smooth the image and suppress noise; median filtering is used to effectively remove salt-and-pepper noise; bilateral filtering is used to preserve edge information well while denoising; or wavelet denoising is used: it has a very good denoising effect on non-stationary signals (such as vibration signals and seismic waves) and can analyze signals in both the time and frequency domains.
[0219] C. Design Considerations: The choice of filtering algorithm needs to strike a balance between denoising effect and computational cost. Overly complex filtering may introduce excessive latency, making it unsuitable for edge scenarios with extremely high real-time requirements.
[0220] 3. Feature Extraction
[0221] (1) Function: Extract the feature vector that best represents the essence of the cleaned data and is most beneficial to subsequent classification or regression tasks. This is a process of data dimensionality reduction and information condensation.
[0222] (2) Technical details:
[0223] A. Necessity of the design: Directly feeding high-dimensional raw data (such as all pixels of an image or all points of a long-term series) into the model is extremely inefficient and prone to overfitting. Feature extraction is a crucial step in the machine learning pipeline.
[0224] B. Feature engineering method:
[0225] B1. For vibration / time-series signals:
[0226] Time-domain characteristics: mean, variance, root mean square (RMS), peak-to-peak value, skewness, kurtosis, waveform factor, peak factor, etc. Kurtosis is highly sensitive to impact signals and is a commonly used feature for equipment fault diagnosis.
[0227] Frequency domain features (obtained via FFT): centroid frequency, mean square frequency, frequency variance, spectral entropy, etc. After converting the signal to the frequency domain using Fast Fourier Transform (FFT), the amplitude and phase information of the spectrum, as well as the band energy, can be extracted (by dividing the spectrum into multiple sub-bands and calculating the energy of each sub-band).
[0228] Time-frequency domain features: Wavelet transform is used to extract features at different time and frequency scales, which is suitable for non-stationary signals.
[0229] B2. For image data:
[0230] Traditional handcrafted features: HOG (Histogram of Oriented Gradients) is used for object detection; SIFT (Scale Invariant Feature Transform) and SURF (Speed-Up Robust Feature Transform) are used for image matching; LBP (Local Binary Pattern) is used for texture analysis.
[0231] Automatic feature extraction: Although Figure 3 While this is listed separately, the convolutional layers of deep learning models (such as CNNs) are essentially performing automatic, hierarchical feature extraction. At the edge, the first few layers of a pre-trained small CNN can be used as feature extractors.
[0232] Design considerations: The quality of feature extraction directly determines the upper limit of subsequent model performance. Good features should possess discriminative power (signature values of samples from different categories differ significantly), robustness (insensitivity to noise), and stability (feature values of samples of the same category are concentrated).
[0233] 4. Lightweight model inference
[0234] (1) Function: Input the feature vector extracted in the previous step into a lightweight machine learning model pre-deployed on the edge device for real-time analysis and judgment. This is the "brain" of edge intelligence.
[0235] (2) Technical details:
[0236] A. Model selection principle: The optimal balance must be found between limited computing resources (CPU / memory), strict power consumption constraints, and certain accuracy requirements.
[0237] B. The lightweight model used in this embodiment:
[0238] B1. Machine learning models, including:
[0239] Support Vector Machine (SVM): Highly efficient on small feature sets, suitable for classification tasks;
[0240] Random Forest: An ensemble model with strong resistance to overfitting and fast training and inference speeds;
[0241] Lightweight Gradient Boosting Machine (LightGBM): High efficiency and accuracy, performing well in many tabular data tasks;
[0242] B2. Lightweight Neural Networks (TinyML):
[0243] B21. Model compression technology:
[0244] Knowledge Distillation: Using a large, high-precision "teacher model" to guide the training of a small "student model," allowing the small model to achieve performance close to that of the large model.
[0245] Pruning: Removes unimportant weights or connections from a neural network to form a sparse network, which significantly reduces computation and model size.
[0246] Quantization: Converting 32-bit floating-point weights and activation values in a model into 8-bit integers (INT8) or even lower precision. This significantly reduces model size, speeds up inference, and lowers power consumption, making it a key technology for edge deployment. TensorFlow Lite, PyTorchMobile, NVIDIA TensorRT, and others all provide powerful quantization tools.
[0247] B22. High-efficiency neural network architecture:
[0248] MobileNet series: uses depthwise separable convolutions, which greatly reduces computation and parameter requirements, and is designed for mobile and embedded vision applications.
[0249] SqueezeNet: The model is extremely small while maintaining AlexNet-level accuracy.
[0250] ShuffleNet: Uses channel shuffling operations to facilitate information flow, further reducing computational costs while maintaining accuracy.
[0251] B3. Inference Engine: The model requires a specialized inference framework to run efficiently on edge hardware, such as...
[0252] TensorFlow Lite, PyTorchMobile, ONNXRuntime, NVIDIA TensorRT, ArmNN, etc. These frameworks are highly optimized for specific hardware (CPU, GPU, NPU).
[0253] (3) Design considerations: Model selection and optimization is an iterative process. It is necessary to continuously validate and update the model with field data to prevent model drift (i.e., the model performance deteriorates over time due to environmental changes).
[0254] 5. Trend prediction / state recognition / anomaly detection
[0255] (1) Function: This is the specific task performed by the lightweight model inference. The model outputs one or more analysis results based on the input features.
[0256] (2) Technical details:
[0257] A. These three aspects are usually interconnected; a single model may perform multiple tasks simultaneously.
[0258] Status Recognition: A classification problem. The model categorizes the current system state into predefined categories.
[0259] Anomaly detection can be viewed as a special type of state recognition, that is, determining whether the current data point is "normal" or "abnormal". It is sometimes implemented through unsupervised learning (such as reconstruction-based autoencoders) or one-class classification (One-Class SVM), which is particularly suitable for scenarios where it is difficult to obtain a large number of anomalous samples.
[0260] Trend prediction: Regression problems or time series forecasting. The model not only judges the current state but also predicts the trend over a future period. For example, predicting the remaining useful life (RUL) of an equipment or predicting temperature changes over the next hour. Commonly used models include LSTM (Long Short-Term Memory) networks, GRU (Gated Recurrent Unit) networks, and their lightweight variants.
[0261] (3) Output format: The output of the model is usually a probability distribution (for classification tasks) or a specific value (for regression tasks). For example, [normal: 0.05, abnormal: 0.95] or predicted temperature: 25.7℃.
[0262] 6. Is there anything abnormal?
[0263] (1) Function: This is a decision point. The system determines whether the current data represents an abnormal event that requires high attention based on the results of model inference (such as the probability of the "abnormal" class) and a preset confidence threshold.
[0264] (2) Technical details:
[0265] A. Threshold Setting: The choice of threshold is crucial, as it directly determines the system's false positive rate (False Positive Rate) and false negative rate (False Negative Rate). Setting the threshold too high increases false negatives; setting it too low increases false positives. Thresholds are typically determined through ROC or PR curve analysis on the validation set.
[0266] B. Multi-model voting: In some complex systems, multiple models may be deployed for ensemble learning, and a more robust final decision may be made through voting or averaging mechanisms.
[0267] C. Context fusion: The decision-making process can not only be based on the model output, but also integrate other contextual information, such as time (alarm thresholds may be different at night and during the day), historical status of the device, and correlation information of multiple sensors, in order to make more accurate judgments.
[0268] 7. Branch 1: Yes -> Mark as high priority data
[0269] Function: If the packet is determined to be abnormal, the packet and its inference result will be marked as high priority and enter the emergency processing channel.
[0270] Technical details:
[0271] (1) Marking method: Set the priority flag (e.g., Priority:HIGH) in the data's metadata or message header.
[0272] (2) Follow-up processing:
[0273] Immediate Transmission: It may immediately interrupt the current transmission queue and prioritize sending to the cloud, ensuring that the monitoring center receives the alert as soon as possible.
[0274] Rich information: In addition to raw data or feature values, the transmission package also includes the model's inference results (such as "anomaly type: bearing failure, confidence level: 96%)", timestamps, device IDs, location information, etc., providing sufficient decision-making basis for the cloud.
[0275] Trigger local action: At the same time, it is very likely to trigger the "trigger local alarm" step in the first flowchart.
[0276] Design considerations: High-priority channels need to have Quality of Service (QoS) guarantees to ensure the reliable delivery of critical data even when the network is congested.
[0277] 8. Branch Two: No -> Regular Data Processing
[0278] Function: If the condition is determined to be normal, the data will enter the regular processing channel. Its main goal is to compress the data to the maximum extent, save storage and transmission costs, and at the same time retain enough information for the cloud to perform long-term trend analysis, model retraining, etc.
[0279] 9. Data compression / feature extraction
[0280] (1) Function: Compress normal data or extract secondary features to generate a simplified representation.
[0281] (2) Technical details:
[0282] A. Lossy compression: In order to save bandwidth to the extreme, lossy compression is usually used, which accepts a certain loss of information.
[0283] A1. Time series data:
[0284] Feature value upload: Instead of uploading the original data, only periodic statistical features are uploaded, such as the mean, maximum, minimum, and variance per minute. This is the most commonly used method.
[0285] Linear fitting / piecewise aggregation approximation (PAA): Approximates a time series data segment using a line segment or a series of piecewise constants, storing only the parameters of the line segment or the values of the segments.
[0286] More advanced encoding methods include Gore coding for integer data and run-length coding for data with many repeating values.
[0287] A2. Image data:
[0288] Frame difference method: For surveillance video, only the area that differs from the previous frame is transmitted, which greatly reduces the amount of data.
[0289] Reduce resolution / frame rate: For normal monitoring footage, reduce its resolution and frame rate before transmission.
[0290] JPEG / HEIC compression: Uses standard image compression algorithms.
[0291] B. "Discarding Raw Data": In many systems, raw data on edge devices is immediately deleted after compression or feature extraction to free up valuable storage space. Data is only deleted after it has been confirmed that it has been successfully transmitted or is no longer needed.
[0292] 10. Generate transmission packets
[0293] (1) Function: Encapsulate the processed data (whether it is high-priority abnormal data with detailed results or compressed / characteristic form of regular data) into data packets that conform to the communication protocol and prepare them for sending.
[0294] (2) Technical details:
[0295] A. Protocol Encapsulation: Pack the data payload, destination address, priority flag, CRC checksum, and other information into MQTT messages, CoAP messages, or TCP / UDP packets.
[0296] B. Batch processing: For regular low-priority data, it may be temporarily stored at the edge node for a period of time, then packaged into batches and sent all at once, further reducing the number of communications and saving power consumption.
[0297] C. Security Encryption: Before transmission, data packets are encrypted (e.g., using TLS / DTLS) and authenticated to ensure data security and trusted device identity.
[0298] According to the appendix Figure 3 The system has implemented the following process:
[0299] (1) Low latency response: Key anomalies are identified and processed within milliseconds to seconds, meeting the real-time requirements of slope monitoring scenarios.
[0300] (2) Bandwidth and cost optimization: It is possible to compress 99% of normal data into a very small volume before transmission, reducing communication costs and cloud storage costs by one to several orders of magnitude.
[0301] (3) Privacy and security: Sensitive raw data (such as facial images) can be analyzed locally, and only the results (such as "stranger detected") are uploaded instead of the raw data to protect user privacy.
[0302] (4) System reliability: Even in the event of a network outage, the edge side still has independent sensing, analysis and local alarm capabilities, ensuring the availability of core functions.
[0303] like Figure 4 The diagram shown illustrates the detailed working principle of the cloud-based global optimization and strategy distribution system.
[0304] (I) General Overview
[0305] Figure 4 This describes the workflow of the cloud control center or digital twin platform of the entire edge intelligence system. Located at the top of the architecture, it aggregates data from a wide range of edge nodes, uses advanced artificial intelligence and data analytics to perform global model training, situational awareness, trend prediction, and policy generation, and continuously distributes the optimized intelligent results (models, policies, configurations) to the edge, making the entire system an organic intelligent entity capable of continuous learning, evolution, and adaptation to environmental changes.
[0306] This process is related to the real-time response process at the edge, and... Figure 3 This creates a complementary system. Edge processing handles "tactical" real-time decisions, while the cloud focuses on "strategic" global optimization and long-term planning.
[0307] (II) Explanation of each part
[0308] 1. Receive data from multiple nodes
[0309] (1) Function: As the starting point of the cloud process, the cloud platform continuously receives data uploaded from all edge gateways and ISU nodes across the country and even the world through interfaces such as message queues and data buses.
[0310] (2) Technical details:
[0311] A. Massive Heterogeneous Data Access: The data sources are extremely diverse, including different models of sensors, different types of devices (such as cameras and vibration sensors), and nodes in different geographical locations. The data formats, protocols, and sampling rates may vary.
[0312] B. High-throughput data pipelines: The cloud needs to establish pipelines capable of handling massive data inflows. Common technologies include:
[0313] Message queues, such as Apache Kafka, RabbitMQ, and AWS Kinesis, decouple data production (edge processing) from consumption (cloud processing services), acting as buffers and smoothing out data spikes to ensure no data loss.
[0314] Stream processing platforms, such as Apache Flink and Spark Streaming, can consume data from these message queues in real time and perform initial processing.
[0315] Protocol Adaptation and Data Parsing: The cloud needs a protocol adaptation layer to parse the communication protocols used by various edge devices (such as MQTT, CoAP, HTTPS, LoRaWAN) and uniformly convert the data into an internal standard format (e.g.,
[0316] JSON, Avro, Parquet).
[0317] (3) Design considerations: Reliability and scalability are crucial in this stage. The system must be able to handle "data tsunamis" and easily integrate new node types.
[0318] 2. Multi-source data fusion
[0319] (1) Function: It can correlate, complement, combine and correct data from different nodes, different sensors and different spatiotemporal scales to generate a more accurate, complete and reliable information whole than any single data source.
[0320] (2) Technical details:
[0321] A. Technical Effects: The perspective of a single node is limited. Multi-source fusion can: First, eliminate uncertainty; an alarm from a single sensor may be a false alarm, but if multiple related sensors detect anomalies simultaneously, the probability of a real event is greatly increased. Second, provide context; abnormal vibration data + abnormal temperature data + abnormal audio data can more accurately determine the fault mode. Third, achieve spatial coverage: stitching together data from multiple nodes forms a global situation map (such as a pressure distribution map of the entire oil field or a displacement field map of the entire mountain).
[0322] B. The integration layers include:
[0323] Data-level fusion: The lowest level, directly fusing raw data (such as image stitching). It requires high communication bandwidth and is rarely used in the cloud.
[0324] Feature-level fusion: The most commonly used level. Features from each data source are extracted (such as the vibration kurtosis of node A and the temperature trend slope of node B), and then these feature vectors are fused into a joint feature vector for use in subsequent models.
[0325] Decision-level fusion: the highest level. Each data source first makes an initial decision locally (such as "normal / abnormal"), and then the cloud integrates all local decisions, and arrives at the final decision through methods such as voting and weighted averaging.
[0326] C. Key Algorithms and Technologies:
[0327] Kalman filtering and its variations: used to fuse noisy time-series data to optimally estimate the true state of the system.
[0328] Bayesian networks: model the probabilistic dependencies between different data sources and perform uncertain reasoning.
[0329] Deep learning fusion models: use neural networks (such as multimodal neural networks) to automatically learn complex correlations and fusion rules between different data sources.
[0330] (3) Design considerations: The core challenge of data fusion is to solve the problem of spatiotemporal alignment (i.e., the timestamps and coordinate systems of data from different sources are consistent) and the problem of confidence assessment (how to measure the credibility of different data sources).
[0331] 3. Training / Optimizing the Model
[0332] (1) Function: Utilize fused global big data to train and optimize more complex and accurate machine learning models. The performance and generalization ability of these models far exceed those of lightweight models on the edge.
[0333] (2) Technical details:
[0334] A. Model type:
[0335] Global diagnostic / identification model: High-precision models trained using full datasets (such as deep residual networks ResNet, large-scale...)
[0336] The Transformer model has a much higher accuracy rate than smaller edge models. These models can be used for secondary diagnosis of difficult edge-reported cases.
[0337] Predictive maintenance model: Train a Remaining Useful Life (RUL) prediction model to predict when equipment will fail, based on historical failure data.
[0338] Anomaly detection models: use unsupervised or semi-supervised learning (such as autoencoders, isolated forests) to discover unknown anomaly patterns hidden in global data.
[0339] B. Training Paradigm:
[0340] Centralized training: The most common approach is to centralize all data in a cloud data center for model training.
[0341] Federated learning: a privacy-preserving technology. Without exchanging raw data, each edge node trains its own model locally using its own data, while the cloud aggregates these model parameter updates to generate a stronger global model. It is particularly suitable for scenarios with high data privacy requirements.
[0342] Hyperparameter optimization: Use automated tools (such as Azure AutoML, Google Vizier, Hyperopt) to search for and optimize model hyperparameters to obtain the best performance.
[0343] Continuous learning: The system periodically or triggeredly retrains the model with new data to adapt to changes in the environment and data distribution (concept drift).
[0344] 4. Stability evaluation model and long-term prediction model
[0345] (1) Function: These two models are core tools for strategic decision-making in the cloud, and they analyze the system status from different time scales.
[0346] (2) Technical details:
[0347] A. Stability Evaluation Model: Used to comprehensively assess and quantify the health status and stability of the entire monitored system in the current and near future. It answers the question, "Is the system as a whole safe now?" Its input is usually real-time features after global fusion, such as statistical values of status indicators, pressure, temperature, vibration, etc. of all nodes; the output may be a stability score (e.g., 0-100), a risk level (low, medium, high), or a system health index; it can be a rule-based expert system or a trained classification / regression model.
[0348] B. Long-term prediction model: Used to predict the development trend of the system over the next few hours, days, or even months based on historical data and current status. It answers the question, "What will happen in the future if this trend continues?" and can therefore be used for long-term monitoring and prediction of slope displacement and settlement trends in this invention. It primarily uses time series prediction models, such as...
[0349] LSTM, GRU, Transformer (such as Informer), and classic time series algorithms (such as ARIMA, Prophet).
[0350] 6. Adaptive Policy Generator
[0351] (1) Function: This is the "decision brain" in the cloud. It integrates the results of stability evaluation and long-term prediction, and combines them with expert knowledge base, business rules and historical experience to automatically generate optimization strategies to guide the behavior of edge nodes.
[0352] (2) Technical details:
[0353] A. Inputs include: stability scores (e.g., "system stability drops to 60%"), long-term prediction results (e.g., "the bearing at node X is expected to fail in 14 days"), and business rules (e.g., "energy cost priority" or "safety priority").
[0354] B. Decision-making logic includes:
[0355] Based on a rule engine: using rules like IF...THEN... For example: "IF stability score < 70% AND predicted failure time < 30 days THEN generate a strategy of 'increasing the sampling frequency to 1Hz'".
[0356] Reinforcement learning: This is a more advanced approach. The cloud acts as a central intelligent agent, continuously testing different policies and observing edge feedback (reward signals, such as whether stability improves or costs decrease) to learn an optimal policy generation strategy.
[0357] Generate strategy types, including: parameter adjustment strategy: adjust the sampling frequency, alarm threshold, filtering parameters, etc. of edge nodes; model update strategy: determine which version of the model to distribute to which region of nodes; task scheduling strategy: instruct certain nodes to perform specific diagnostic tasks; resource control strategy: in scenarios such as microgrids, directly generate control commands (such as adjusting power output).
[0358] 7. Generate policy update package
[0359] (1) Function: Transform the generated abstract policy into a formatted differential data packet that edge devices can understand and execute.
[0360] (2) Technical details:
[0361] A. Generate the contents of the policy update package, including: a new version of the lightweight model file (usually in .tflite format after quantization and pruning); a new configuration file (such as config.json, containing new thresholds and parameters); an update script and digital signature (used to verify the legality and integrity of the update package and prevent malicious attacks);
[0362] B. Differential Update: To save bandwidth, the update package usually only contains the parts that differ from the current version, rather than the complete file.
[0363] 8. Push to edge gateway / ISU & 9. ISU update local policies
[0364] (1) Function: Through a secure and reliable downlink channel, the policy update package is accurately delivered to the target device group, and the device is seamlessly upgraded.
[0365] (2) Technical details:
[0366] Push mechanism: Typically employs OTA (Over-The-Air) upgrade technology. The cloud sends update notifications and download links to devices via message queues (such as MQTTTopic).
[0367] Grouping and targeted push: Devices can be grouped according to attributes such as device model, geographical location, and software version to achieve canary release. The update effect is first verified on a small scale, and then it is pushed to the whole system to maximize system stability.
[0368] Secure Updates: The entire transmission and installation process must be encrypted and authenticated. Devices must verify digital signatures before installing updates.
[0369] The update process includes:
[0370] After receiving the notification, the device downloads the update package from the specified address; verifies the integrity and validity of the package; applies the update package to the spare partition or memory area; verifies the validity of the new software / configuration; restarts or hot-swaps to the new version to complete the update; and reports the update status (success / failure) to the cloud.
[0371] Figure 4 This explains the principle of "cloud-edge collaboration," where the cloud is no longer a simple data warehouse but has evolved into the system's intelligent hub and decision-making engine. Through this closed loop, the system achieves:
[0372] 1. Global optimization: Break down the information silos of individual nodes and make optimal decisions from a God's-eye view.
[0373] 2. Continuous evolution: The system can continuously learn from operational data, automatically optimize models and strategies, and continuously improve.
[0374] 3. Proactive intervention: Shift from "post-event response" to "pre-event prediction" and "in-event control" to achieve true predictive maintenance and intelligent management.
[0375] 4. Standardized Management: Through unified management and policy distribution by the central node, the consistency and maintainability of the behavior of large-scale edge device clusters are ensured.
[0376] The core innovation of the present invention lies in that the sensing node not only collects data, but also has local computing capability, and can adaptively adjust its monitoring strategies (such as sampling frequency, data precision, transmission mode) according to historical data, current status and environmental factors, while using intelligent algorithms to process, analyze and learn from the raw data, so as to improve monitoring precision and prediction and early warning capabilities.
[0377] The technical innovations of the present invention are as follows:
[0378] (1) The present invention constructs a distributed slope monitoring system with edge intelligent adaptive decision-making capability, which breaks through the performance bottleneck of existing solutions through multi-level technical innovation.
[0379] A closed loop of dynamic monitoring strategy at the sensing node level is established, so that each ISU node autonomously adjusts the sampling frequency (switching between 1 and 100Hz) and data precision (such as switching between GNSS single-frequency / dual-frequency modes) based on the displacement change rate (e.g., threshold R1=0.5mm / h), environmental events (rainfall > P1=30mm / 24h or vibration value > V1=0.3g) and battery status (power level < E1=20%), achieving a dynamic optimal balance between energy consumption and precision, with the measured average power consumption reduced by 52.7% (compared with the fixed-frequency monitoring solution);
[0380] An edge-cloud collaborative intelligent optimization system is established, in which a lightweight ML model (such as TinyLSTM compressed to <50KB) is deployed on the MCU side to perform real-time state identification and abnormal warning (with latency <200ms), while multi-source data is fused on the cloud to generate a strategy update package that is dynamically transmitted back to the edge nodes, forming the continuous evolution capability of the monitoring strategy (monthly model optimization rate >5%).
[0381] (2) A node-level adaptive strategy execution mechanism, covering specific judgment logics such as dynamic adjustment of sampling frequency driven by displacement rate, high-precision mode switching of sensors triggered by environmental events, noise data compression, and power arbitration;
[0382] (3) A two-way interaction architecture for edge intelligence and cloud optimization, which protects the prediction / identification method of the lightweight model on the MCU side (such as Kalman residual anomaly detection) and the issuing process of the strategy update package generated by the cloud parameter optimizer;
[0383] (4) A low-power heterogeneous networking and energy management solution, covering the combined implementation of multi-protocol ad-hoc network communication stacks (LoRa / NB-IoT / Zigbee), dynamic energy harvesting allocation strategies and relay assistance mechanisms. Any monitoring system implemented without permission that simultaneously includes the three features of adaptive sampling adjustment (e.g., rate threshold triggering frequency change), edge model warning (non-raw data upload) and strategy closed-loop update constitutes an infringement.
[0384] The present invention also provides a memory that stores multiple instructions for implementing the method as described in Embodiment 1.
[0385] like Figure 5 As shown, the present invention also provides an electronic device, including a processor 301 and a memory 302 connected to the processor 301. The memory 302 stores multiple instructions, which can be loaded and executed by the processor to enable the processor to perform a distributed slope monitoring method.
[0386] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A distributed slope monitoring system based on edge-cloud collaborative intelligent adaptive decision-making, characterized in that, include: Multiple intelligent sensing unit (ISU) nodes and data processing and intelligent analysis units are deployed at key locations on the slope. Each Intelligent Sensor Unit (ISU) node transmits data wirelessly to an edge gateway or directly to a cloud platform via a self-organizing network for slope monitoring based on a local adaptive monitoring strategy. This local adaptive monitoring strategy involves the ISU node or edge device analyzing the collected data using a built-in lightweight decision model to determine if any anomalies have occurred or if a warning threshold has been reached. The local adaptive monitoring strategy features a cloud / edge feedback mechanism, which receives model update or strategy adjustment instructions from the edge gateway or cloud platform. The cloud then distributes the optimized model, decision rules, and configuration parameters to the edge devices and ISU nodes, enabling the monitoring system to self-evolve and iterate. The data processing and intelligent analysis unit includes an edge intelligence module, a cloud gateway, and an edge gateway. The edge gateway, as an intermediate layer, is used for protocol conversion, data aggregation, temporary storage, and preliminary analysis. The cloud gateway provides computing and storage resources, and trains more complex AI models based on historical and real-time data to perform pattern recognition, predictive analysis, and anomaly detection tasks. The edge intelligence module includes multiple edge computing nodes and a built-in lightweight AI inference unit, located at the edge, to implement edge intelligent processing. For data that does not require immediate alarm, high-complexity processing is performed at the edge, including feature extraction, anomaly detection, and event classification.
2. The distributed slope monitoring system based on edge-cloud collaborative intelligent adaptive decision-making according to claim 1, characterized in that, The lightweight decision model includes a rule engine, decision tree, or micro neural network. It runs a lightweight state machine or decision tree logic in the microcontroller module of the intelligent sensing unit (ISU) node and switches the local adaptive monitoring strategy based on built-in rules or learning models. The key parameter thresholds of the local adaptive monitoring strategy are continuously optimized through model learning.
3. The distributed slope monitoring system based on edge-cloud collaborative intelligent adaptive decision-making according to claim 2, characterized in that, The lightweight AI inference unit runs a compressed neural network model to perform image recognition and time series analysis tasks.
4. The distributed slope monitoring system based on edge-cloud collaborative intelligent adaptive decision-making according to claim 3, characterized in that, The edge intelligence module employs a real-time stream processing framework to perform real-time data pipeline processing.
5. The distributed slope monitoring system based on edge-cloud collaborative intelligent adaptive decision-making according to claim 4, characterized in that, The intelligent sensing unit (ISU) node includes: The sensing core module includes one or more of a MEMS accelerometer, a GNSS receiver chip, and a strain gauge; wherein the MEMS accelerometer is used to measure tilt angle changes and reflect displacement trends, the GNSS receiver chip is used to acquire absolute position changes, and the strain gauge is used to monitor local deformation. An environmental sensing module includes one or more of a temperature and humidity sensor, a rain gauge, and a vibration sensor; wherein the temperature and humidity sensor is used to compensate for sensor drift, the rain gauge is used to sense key factors that trigger monitoring, and the vibration sensor is used to sense disturbance events. The microcontroller module includes a core processing unit that performs data acquisition, preprocessing, and local computation based on computing power; The wireless communication module uses a low-power wide-area network (LPWAN) or self-organizing network protocol for communication. The energy supply module includes a short-term power supply unit, a long-term power supply unit, and a power management unit (PMU). The short-term power supply unit includes a battery; the long-term power supply unit includes an energy harvesting module for long-term operation; and the power management unit (PMU) is used to implement intelligent power consumption management of the monitoring system. Small-capacity memory, configured locally, is used to cache critical data and adaptive strategy parameters.
6. The distributed slope monitoring system based on edge-cloud collaborative intelligent adaptive decision-making according to claim 5, characterized in that, The low-power wide area network (LPWAN) is LoRaWAN, NB-IoT, Zigbee, or BLEMesh LPWAN.
7. The distributed slope monitoring system based on edge-cloud collaborative intelligent adaptive decision-making according to claim 6, characterized in that, The energy harvesting module includes a solar panel or a vibration energy harvesting unit.
8. The distributed slope monitoring system based on edge-cloud collaborative intelligent adaptive decision-making according to claim 7, characterized in that, The edge intelligence module includes: The feature extraction unit is used to perform data filtering, noise reduction, displacement calculation, and anomaly detection on the multiple intelligent sensing unit (ISU) nodes or edge gateways. Lightweight model inference unit, used to deploy lightweight ML models for: The trend prediction unit is used to predict short-term displacement trends based on historical displacement data, including: using a lightweight version of simple linear regression and RNN or LSTM to predict displacement trends over hours or days. The status recognition unit is used to make a preliminary judgment on whether the current status of the slope is stable, slowly deforming, rapidly deforming, or showing signs of impending landslide. Anomaly alarm unit is used to detect displacement mutations or abnormal events that significantly deviate from the normal pattern based on statistical thresholds or model inference.
9. The distributed slope monitoring system based on edge-cloud collaborative intelligent adaptive decision-making according to claim 8, characterized in that, The basis for adjusting the slope monitoring strategy based on the local adaptive monitoring strategy includes: The dynamic adjustment of the displacement change rate includes: dynamically adjusting the sampling frequency based on a first adjustment strategy by calculating the displacement change rate in real time, thereby dynamically adjusting the displacement change rate. The first adjustment strategy includes low-frequency sampling when the displacement change rate is low to save power consumption and high-frequency sampling when the displacement change rate is high to capture details. The strategy adjustment is based on environmental triggering events, including: when one or more of the following are detected by the rain gauge, the vibration sensor, and the temperature and humidity sensor, the monitoring frequency and accuracy are automatically increased based on a second adjustment strategy. The second adjustment strategy includes switching the GNSS receiver chip from single frequency to dual frequency and putting the MEMS accelerometer into a high-precision mode. Adjusting the data fluctuation level or noise by means of strategy includes: adjusting the data fluctuation level or noise by calculating the standard deviation or statistical characteristics of the data, and adjusting the data fluctuation level or noise by means of a third adjustment strategy, wherein the third adjustment strategy includes appropriately reducing the data precision and reducing the data transmission frequency during the stable period, reducing the data precision by means of bit compression; and restoring high data precision and high frequency data transmission when the data fluctuation intensifies. The battery status is adjusted strategically, including: intelligently balancing power consumption and monitoring requirements based on the remaining power and energy harvesting efficiency through a fourth adjustment strategy, wherein the fourth adjustment strategy includes forcing the battery to enter the most power-saving mode when the battery enters a low power state.
10. The distributed slope monitoring system based on edge-cloud collaborative intelligent adaptive decision-making according to claim 9, characterized in that, The cloud gateway and edge gateway include: The model training and optimization unit is used to receive data uploaded from various nodes and train more complex models on the edge gateway or in the cloud; wherein, the data uploaded from various nodes is raw data and / or preprocessed data, and the more complex model is a cloud-based deep learning model; A multi-source data fusion model is used to integrate displacement, environmental, and historical data to establish a more accurate slope stability evaluation model; wherein, the slope stability evaluation model is a machine learning model or a composite model of physical model + AI; Long-term time-series forecasting models are used for medium- to long-term deformation trend prediction. An adaptive policy generator is used to generate optimized adaptive policy update packages based on a global analysis of data from all nodes, and push them to the relevant ISU nodes. Early warning models are used to provide comprehensive early warnings with higher confidence based on the fusion of information from multiple nodes and spatial correlations; The model push update unit is used to push incremental or full updates of the trained model to edge nodes, cloud gateways, and edge gateways. The data filtering and transmission optimization unit is used to utilize local analysis results to transmit only data that has changed significantly, anomaly reports, important feature values, or compressed model outputs.
Citation Information
Patent Citations
Island area foundation pit deformation monitoring and predicting system and method
CN120293074A
Method for the execution of services in real time, notably of flight management and real time system using such a method
US20150334173A1
