A mountain torrent early warning monitoring device and method based on rain sound and moisture condition
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NINGBO HONGTAI WATER RESOURCES INFORMATION TECH CO LTD
- Filing Date
- 2026-03-09
- Publication Date
- 2026-07-14
Smart Images
Figure CN122392228A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of flash flood early warning technology, specifically to a flash flood early warning monitoring device and method based on rain sound and soil moisture. Background Technology
[0002] Currently, monitoring and early warning of flash floods mainly rely on various sensor devices deployed in risk areas. However, most existing flash flood monitoring devices still have many limitations in practical applications. First, traditional monitoring methods often have a single monitoring dimension, usually relying solely on rain gauges or water level gauges for independent monitoring. They lack the ability to simultaneously collect and analyze multi-source data such as soil moisture, mountain micro-vibrations, and rainfall intensity, making it difficult for early warning models to establish a correlation between "soil moisture-mountain structure stability-rainfall intensity," which easily leads to false alarms and missed alarms. Second, in terms of soil moisture monitoring, most existing devices only monitor the soil moisture content at a single depth, failing to obtain the moisture change gradient of soil layers at different depths. This makes it difficult to accurately determine whether the deep soil has reached saturation, thus failing to effectively capture precursory information of landslides. In addition, the data processing mode of existing devices also has significant defects. Most rely on uploading raw monitoring data to cloud servers for analysis and processing. In remote mountainous areas, this mode often leads to data transmission delays due to network bandwidth limitations, resulting in early warning decision response times of minutes or even longer, which cannot meet the urgent need for second-level early warning responses in the "short-term outbreak" of flash floods. Furthermore, the deployment and maintenance of existing equipment face challenges in complex terrain conditions. For example, some equipment relies on wired network transmission, resulting in high deployment costs and difficulties. Simultaneously, the lack of effective on-site positioning and debugging methods makes it difficult for maintenance personnel to quickly locate equipment in complex mountainous environments. Moreover, parameter modifications or software upgrades often require disassembling the equipment, leading to low maintenance efficiency. Additionally, existing displacement monitoring methods mostly rely on GPS positioning, which suffers from significantly reduced accuracy (to the meter level) in areas with severe signal blockage, such as mountain valleys, failing to detect minute centimeter-level displacements before landslides occur, thus hindering early warning. Therefore, overcoming these shortcomings in existing technologies and providing a flash flood early warning monitoring device capable of multi-source data fusion sensing, real-time edge processing, high-precision positioning, and convenient maintenance is a technical problem urgently needing to be solved by those skilled in the art. Summary of the Invention
[0003] To achieve multi-source data fusion sensing, real-time processing at the edge, high-precision device positioning, and convenient maintenance, this invention proposes a flash flood early warning and monitoring device based on rainfall sound and soil moisture, comprising: A pole having an underground portion buried in the ground and an above-ground portion located on the ground; A multi-layer soil moisture sensor includes at least two sensor probes fixed at intervals along the vertical direction to the underground part of the pole, for collecting soil moisture content data at different depths. The infrasound vibration sensor is rigidly connected to the inside of the pole and is used to collect infrasound signals generated by micro-vibrations of the mountain. The rain-listening cap is a conical, thin-walled shell fixed to the top of the pole; A microphone array, comprising multiple microphones evenly distributed inside the rain cap, for collecting audio signals generated by raindrops impacting the rain cap; The UWB module, located on the ground portion, is used for ranging communication with external devices; An edge computing controller is installed on the above-ground part and is electrically connected to the multi-layer soil moisture sensor, the infrasound vibration sensor, the microphone array and the UWB module respectively. It is used to receive and process the soil moisture data, the infrasound signal, the audio signal and the ranging data obtained through the UWB module, and run a preset early warning model locally to output early warning results. The wireless communication module is electrically connected to the edge computing controller and is used to send the early warning result to a remote platform.
[0004] This invention achieves efficient early warning of flash flood disasters by integrating multi-source sensors and utilizing an edge computing controller to process soil moisture, infrasound, rain sound, and ranging data locally in real time.
[0005] Furthermore, the UWB module is also configured to network and communicate with UWB modules on adjacent poles, calculate the relative displacement data between adjacent poles in real time, and send the relative displacement data to the edge computing controller.
[0006] Furthermore, the edge computing controller is configured to run machine learning-based signal processing algorithms locally, including: performing a fast Fourier transform on the infrasound signal to extract effective frequency band features, performing Kalman filtering on the soil moisture data to remove measurement noise, and performing a beamforming algorithm on the audio signal to enhance the rain sound signal and suppress environmental noise.
[0007] Furthermore, the edge computing controller is also configured to run a pruned and quantized compressed LSTM-CNN spatiotemporal fusion early warning model locally. The LSTM-CNN spatiotemporal fusion early warning model takes the soil moisture content data, the characteristic frequency band data of the infrasound signal, the rain sound intensity data of the audio signal, and the relative displacement data as inputs, and outputs the flash flood and landslide risk level.
[0008] Furthermore, the multi-layer soil moisture sensor includes three frequency domain reflectance (FDR) soil moisture sensor probes, which are respectively fixed at three preset depths in the underground part of the pole.
[0009] Furthermore, the infrasound vibration sensor is rigidly fixed inside the pole by a metal bracket, and the sensitive axis is set vertically downward.
[0010] Furthermore, the microphone array consists of four microphones, which are evenly distributed in a circular pattern and fixed to the inner wall of the rain-listening cap.
[0011] This invention also includes a flash flood early warning and monitoring method based on rainfall sound and soil moisture, comprising the following steps: S1: Real-time acquisition of multi-layer soil moisture content data through at least two soil moisture sensor probes, and real-time acquisition of infrasound signals generated by micro-vibration of the mountain through an infrasound vibration sensor. S2: The multi-channel audio signal generated by raindrops impacting the rain-listening cap is acquired in real time through a microphone array, and the edge computing controller performs beamforming algorithm on the multi-channel audio signal to output rain sound intensity characteristic data; S3: The edge computing controller performs a fast Fourier transform on the infrasound signal to extract the spectral feature data of the effective frequency band of the infrasound, and performs Kalman filtering on the multi-layer soil moisture content data to output soil moisture change data. S4: The soil moisture change data, infrasound spectrum feature data and rain sound intensity feature data are input into the LSTM-CNN spatiotemporal fusion early warning model through the edge computing controller. The LSTM layer extracts the temporal change features, the CNN layer extracts the spatial coupling features of multi-source data, and the feature fusion layer outputs the risk level of flash floods and landslides. S5: When the risk level reaches the preset threshold, the warning information containing the risk level and device location information is sent to the remote IoT platform through the wireless communication module.
[0012] Furthermore, the entire operation of the flash flood early warning and monitoring device also includes the following steps: Real-time relative displacement data between adjacent poles is obtained through periodic ranging communication between UWB modules; Spatiotemporal correlation analysis was performed on relative displacement data, soil moisture change data, and infrasound spectral characteristic data. When the relative displacement data exceeds the preset displacement threshold, the soil moisture change data shows a saturation trend, and the infrasound spectrum characteristics show a sudden drop in the dominant frequency, the output risk level will be upgraded by one level.
[0013] Furthermore, the edge computing controller also performs the following adaptive optimization steps: The early warning events generated by real-time monitoring and their corresponding multi-source sensor data are packaged into verification samples and periodically uploaded to a remote IoT platform via a wireless communication module. Receive incremental model parameters from the remote IoT platform after optimizing and updating the early warning model parameters based on verification samples; The incremental model parameters are loaded into the LSTM-CNN spatiotemporal fusion early warning model for adaptive iterative optimization.
[0014] Compared with the prior art, the present invention has at least the following beneficial effects: (1) The present invention proposes a flash flood early warning monitoring device and method based on rain sound and soil moisture. Through the FDR soil moisture sensor layered on the pole, the rigidly connected infrasound vibration sensor and the microphone array evenly distributed in the rain listening cap, it realizes the accurate perception of soil saturation gradient, mountain micro-vibration and rainfall intensity, and solves the defects of traditional monitoring with single dimension and difficulty in capturing small signals of landslide precursors. (2) The edge computing controller is used to perform FFT spectrum analysis of infrasound signals, Kalman filtering of soil moisture data and beamforming noise reduction of multi-channel audio signals locally, which significantly improves the signal-to-noise ratio and feature extraction efficiency of the original data. (3) By running the pruned and compressed LSTM-CNN spatiotemporal fusion model, multi-source feature data is coupled and analyzed in real time at the edge, and the second-level early warning results are output, thus getting rid of the dependence on cloud computing power and reducing the early warning delay. Attached Figure Description
[0015] Figure 1 This is a schematic diagram of the overall structure of the device; Figure 2 A schematic diagram of the exploded unfolding structure of the rain-sensing cap. Figure 3 This is a schematic diagram of the internal connection status of the functional section of the infrasound vibration sensor. Figure 4 A flowchart illustrating the steps of a flash flood early warning and monitoring method based on rainfall sound and soil moisture. Explanation of reference numerals in the attached diagram: 1-pole, 11-underground part, 12-above-ground part, 2-multi-layer soil moisture sensor, 3-infrasound vibration sensor, 4-rain cap, 5-microphone array. Detailed Implementation
[0016] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments to provide a better understanding of the technical solution and beneficial effects of the present invention. It should be noted that the drawings are for illustrative purposes only, and the proportions and dimensions of the components may be adjusted for clarity, and do not constitute a limitation on the scope of protection of the present invention.
[0017] This invention provides a flash flood early warning monitoring device and method based on rainfall sound and soil moisture, aiming to solve the core technical problems of existing flash flood monitoring equipment, such as single monitoring dimensions, high early warning delay, difficult deployment and maintenance, and insufficient displacement monitoring accuracy. The device integrates multi-source sensors through an integrated pole structure and combines it with an edge computing controller to achieve local intelligent early warning. Simultaneously, it utilizes UWB technology to achieve dual functions of device positioning and displacement monitoring, improving the timeliness, accuracy, and ease of operation and maintenance of flash flood disaster early warnings.
[0018] like Figures 1 to 3 As shown, the flash flood early warning monitoring device based on rain sound and soil moisture described in this invention has a core pole-type data acquisition unit. The pole 1 has a vertical rod-like structure, specifically divided into an underground part 11 buried in the ground and an above-ground part 12 exposed above ground. The length of the underground part 11 is typically set to 40cm to 60cm to ensure that key sensors can be buried in stable soil layers. The height of the above-ground part 12 is set to 1.2m to 1.5m depending on the terrain and vegetation conditions of the monitoring point to ensure that the rain-listening cap 4 at the top is unobstructed and can effectively receive raindrop impact signals. The main body of the pole 1 is made of high-strength, corrosion-resistant, and rust-resistant metal materials, such as stainless steel or hot-dip galvanized steel pipe. The interior is hollow to accommodate various sensor cables, controllers, and power supply modules. All interfaces of the underground part are strictly sealed with epoxy resin, achieving an IP68 protection level, and can work stably in humid and corrosive underground environments for a long time.
[0019] In the underground portion 11 of pole 1, a multi-layer soil moisture sensor 2 is fixedly installed using a multi-functional segmented / layered assembly. This sensor addresses the problem that traditional monitoring methods cannot obtain the vertical distribution gradient of soil moisture. In a preferred embodiment, the multi-layer soil moisture sensor 2 includes three independent sensor probes that measure moisture using the frequency domain reflection (FDR) principle. The three probes are spaced apart along the vertical direction of pole 1. Specifically, the first probe is fixed 10cm from the bottom of the pole, corresponding to a depth of 10cm underground; the second probe is fixed at 20cm; and the third probe is fixed at 40cm. Through this layered arrangement, the device can monitor the water infiltration and saturation process from the surface soil to the deep soil in real time. The FDR sensor probes are rigidly connected to pole 1, ensuring that the probe position remains relatively fixed even when the soil experiences slight creep or displacement, avoiding data distortion caused by probe offset. The sensor probe housing is made of waterproof and corrosion-resistant material, and the probe is electrically connected to the edge computing controller of the above-ground portion 12 via a shielded cable inside pole 1.
[0020] Also installed inside the underground portion 11 of pole 1 is an infrasound vibration sensor 3. This sensor is used to capture infrasound signals generated by the fracturing and friction of rock and soil before a landslide or flash flood. These signals typically have frequencies below 20Hz, imperceptible to the human ear, and are important precursors to geological disasters. To ensure the sensitivity and accuracy of signal acquisition, the infrasound vibration sensor 3 is rigidly connected to the inner wall of pole 1 via a metal bracket, ensuring its sensitive axis is vertically downward and aligned with the main direction of the mountain vibration, thereby maximizing signal acquisition efficiency. In a preferred embodiment, to improve the reliability and redundancy of monitoring data, two identical infrasound vibration sensors 3 are arranged side-by-side inside pole 1. Their data are cross-checked to effectively eliminate occasional interference and provide richer raw information for subsequent data analysis. The infrasound vibration sensor 3 is preferably a piezoelectric accelerometer with a frequency response range covering 0.5Hz to 6kHz, fully encompassing the infrasound frequency band, and a measurement range of ±800m / s. 2 It can capture various signals ranging from weak creep to violent vibration, and operates over a wide temperature range of -40℃ to +150℃, making it suitable for harsh mountainous climates. The 4-20mA standard current signal output by the sensor is transmitted to the edge computing controller via a shielded cable.
[0021] At the top of the pole 1, a rain-listening cap 4 is fixedly installed. The rain-listening cap 4 is a conical thin-walled shell structure made of stainless steel. Compared with hemispherical or flat designs, the conical design has better airflow guidance and sound wave focusing effects. When raindrops impact the cap at different angles, it can excite richer and more easily distinguishable vibrations and acoustic characteristics. The interior of the rain-listening cap 4 forms a sealed cavity, and a microphone array 5 is fixedly installed on the inner wall of the cavity. The microphone array 5 consists of multiple high-sensitivity MEMS microphones, evenly distributed in a circle, used to collect the audio signals generated by raindrops impacting the rain-listening cap 4. The core advantage of using a microphone array instead of a single microphone is that it can utilize the spatiotemporal differences in the signals received by different microphones, combined with subsequent signal processing algorithms, to accurately distinguish rain sounds from environmental noise (such as wind sounds, insect chirps, and bird calls), and determine the impact direction and energy distribution of raindrops. In a preferred embodiment, the microphone array 5 consists of four microphones evenly distributed at 90-degree intervals on the inner circumference of the rain cap 4. This configuration strikes a balance between cost, power consumption, and performance, enabling 360-degree omnidirectional sound acquisition while providing sufficient data channels for beamforming algorithms. The four-channel audio signals acquired by the four microphones are pre-amplified and converted from analog to digital before being transmitted to the edge computing controller via internal cables.
[0022] At the ground level 12 of pole 1, one or more sealed control compartments are installed, integrating the edge computing controller of this invention, along with its supporting UWB module, wireless communication module, GPS positioning module, and power management module. The edge computing controller is the "brain" of the entire device, electrically connected to the multi-layer soil moisture sensor 2, infrasound vibration sensor 3, microphone array 5, UWB module, GPS module, and wireless communication module. It is responsible for receiving raw data collected by all sensors and performing localized signal processing and intelligent early warning model inference. In a preferred embodiment, the edge computing controller employs a high-performance microcontroller based on the ARM Cortex-M55 core, such as the Renesas Electronics RA6M5 series chip. The biggest technological advantage of this chip lies in its Arm Helium vector processing technology, specifically the Arm v8.1M architecture's M-ProfileVector Extension, which provides higher performance support for digital signal processing (DSP) and machine learning (ML) applications while maintaining low power consumption. Its power consumption during AI inference is only about one-fifth that of traditional processors, making it ideal for battery-powered edge monitoring devices that require long-term operation. In addition, the controller complies with the ISO26262 functional safety standard, meeting the stringent requirements for high reliability in geological disaster monitoring equipment.
[0023] The UWB module is also integrated into the control compartment of the ground section 12. It uses ultra-wideband pulse radio technology and has high-precision ranging and positioning capabilities. In this invention, the UWB module is configured with two core operating modes, realizing functional reuse. The first is the "maintenance positioning mode," in which the UWB module establishes two-way communication with a separate handheld debugger (hereinafter referred to as the "handheld device"). The handheld device itself also integrates the UWB module, a touch screen display, and a buzzer. When maintenance personnel carry the handheld device into the UWB communication range of pole 1 (usually 100-150 meters), a dedicated APP on the handheld device screen will generate a radar-like interface, displaying the relative position and distance of pole 1 in real time. At the same time, the buzzer in the handheld device will automatically adjust the beeping frequency according to the distance; the closer the distance, the more rapid the beeping; the farther the distance, the lower the beeping frequency. This dual guidance mechanism of vision and hearing helps maintenance personnel to quickly and accurately locate the equipment that needs maintenance, even in mountainous areas with low visibility and complex terrain. The second mode of the UWB module is the "networked displacement monitoring mode." In this mode, multiple poles 1 deployed at different locations on the same hillside (such as the top, middle, and bottom of the slope) have UWB modules that periodically communicate with each other through distance measurement, automatically forming a local monitoring network. By accurately calculating the relative distance changes between each pole, the system can capture centimeter-level displacements of the mountain in real time. This represents a significant improvement over the meter-level accuracy of traditional GPS in mountainous areas with signal obstruction, providing crucial data support for early landslide warnings.
[0024] The wireless communication module is preferably a 4G communication module that supports all network types. It is electrically connected to the edge computing controller and is responsible for sending locally generated early warning information and key monitoring data to a remote IoT platform, while also receiving downlink commands from the platform. For extremely remote areas with no public network signal, the wireless communication module can also be replaced with a Beidou short message communication module to ensure reliable transmission of early warning information.
[0025] The power management module is connected to the battery compartment located on the ground level (12). This battery compartment is designed to accommodate two different types of batteries, allowing users to flexibly choose according to the nature of the monitoring task. For fixed monitoring points requiring long-term (2-3 years) unattended operation, disposable high-capacity lithium thionyl chloride batteries (lithium-thionyl chloride batteries) can be used; for emergency rescue, short-term intensive observation, and other scenarios, rechargeable lithium iron phosphate batteries can be used, which can be periodically charged or replaced on-site by maintenance personnel. This dual power supply design enhances the device's adaptability to different application scenarios.
[0026] In addition, this system includes a handheld device as a supporting tool. Besides the aforementioned UWB module for positioning, touchscreen display, and buzzer, the handheld device also integrates a Type-C interface. This Type-C interface is a multi-functional interface, serving as a charging port for the handheld device's built-in battery. Furthermore, when the handheld device is connected to an external computer via the Type-C interface, maintenance personnel can copy historical monitoring data files exported from the pole-mounted data collector, or save the latest firmware upgrade file for the pole-mounted data collector to the handheld device, and then upgrade the pole-mounted equipment wirelessly via UWB. The entire process does not require disassembling the pole, greatly improving the convenience and efficiency of on-site maintenance.
[0027] The following will provide a detailed description of the flash flood early warning and monitoring method based on rainfall sound and soil moisture provided by this invention, combined with specific application scenarios and embodiments. This method relies on the hardware device described above, and its core lies in fully leveraging the local intelligent processing capabilities of the edge computing controller. For example... Figure 4 As shown, the method mainly includes the following steps: S1: Real-time acquisition of multi-layer soil moisture content data through at least two soil moisture sensor probes, and real-time acquisition of infrasound signals generated by micro-vibration of the mountain through an infrasound vibration sensor. S2: The multi-channel audio signal generated by raindrops impacting the rain-listening cap is acquired in real time through a microphone array, and the edge computing controller performs beamforming algorithm on the multi-channel audio signal to output rain sound intensity characteristic data; S3: The edge computing controller performs a fast Fourier transform on the infrasound signal to extract the spectral feature data of the effective frequency band of the infrasound, and performs Kalman filtering on the multi-layer soil moisture content data to output soil moisture change data. S4: The soil moisture change data, infrasound spectrum feature data and rain sound intensity feature data are input into the LSTM-CNN spatiotemporal fusion early warning model through the edge computing controller. The LSTM layer extracts the temporal change features, the CNN layer extracts the spatial coupling features of multi-source data, and the feature fusion layer outputs the risk level of flash floods and landslides. S5: When the risk level reaches the preset threshold, the warning information containing the risk level and device location information is sent to the remote IoT platform through the wireless communication module.
[0028] Specifically, after the device is powered on and completes its initial self-test, the system enters a low-power real-time monitoring loop. First, step S1, the data acquisition process, is executed. This process is triggered by the edge computing controller according to a preset sampling strategy. For the multi-layer soil moisture sensor 2, the controller samples at a lower frequency (e.g., once every 5 minutes) to save power, acquiring raw data on soil volumetric water content at three depths (10cm, 20cm, and 40cm). This raw data exists in the form of voltage or frequency signals, containing information on the soil dielectric constant, but also includes interference from sensor polarization and soil particle contact noise. For the infrasound vibration sensor 3, the controller continuously samples at a higher frequency (e.g., 200 times per second) to ensure the capture of brief, sudden micro-vibration signals. The raw data is a continuous voltage waveform, representing the acceleration time history of the mountain vibration. For the microphone array 5, the controller is normally in a "listening" state, only monitoring the energy of the audio signal. Once a continuous rise in energy is detected, indicating potential rainfall, the controller immediately activates the full-band recording mode, simultaneously acquiring audio stream data from the four microphones.
[0029] Next, steps S2 and S3, namely the signal preprocessing and feature extraction stages at the edge, are performed. This stage leverages the advantages of the Cortex-M55 controller's Helium vector processing technology. The edge computing controller first executes a beamforming algorithm on the multi-channel audio signals acquired by the microphone array 5. The core of this algorithm is to utilize the signal phase difference and time delay difference generated by the different spatial positions of the four microphones, and construct a spatial filter pointing towards the surface of the rain cap 4 through matrix operations accelerated by Helium technology. This filter can enhance the raindrop impact sound coming from the direction of the rain cap 4, while significantly suppressing interference from other directions such as wind noise and environmental noise. Finally, it outputs a single-channel rain sound signal with a significantly improved signal-to-noise ratio, and extracts feature data that reflects the rainfall intensity from this signal, such as short-time energy, zero-crossing rate, and spectral centroid, to achieve accurate classification of rainfall intensity (light rain, moderate rain, heavy rain, and torrential rain).
[0030] Simultaneously, the controller performs a real-time Fast Fourier Transform (FFT) on the raw waveform data stream acquired by the infrasound vibration sensor 3. Through FFT, the time-domain vibration signal is converted to the frequency domain, allowing the controller to accurately extract the spectral characteristics of the 0.1-20Hz effective infrasound frequency band, including the dominant frequency and spectral energy distribution. Simultaneously, the controller performs a Kalman filter algorithm on the raw data of multi-layer soil moisture content. Kalman filtering removes random noise and anomalous jumps from the soil moisture sensor data, outputting smooth and stable filtered soil moisture change data, and enabling the calculation of the rate of moisture change.
[0031] After preprocessing in steps S2 and S3, the original multi-source heterogeneous data was transformed into high-quality, physically meaningful feature data: filtered soil moisture change data (including three depths), infrasound spectrum feature data (dominant frequency and energy), and rain sound intensity feature data.
[0032] Subsequently, step S4, the inference process of the edge intelligent early warning model, is executed. These feature data are input in real-time into the LSTM-CNN spatiotemporal fusion early warning model stored locally on the edge computing controller. This model undergoes targeted pruning and quantization compression, reducing its storage capacity from over 100 MB to less than 30 MB, enabling smooth operation on the Cortex-M55 core. The model's dual-branch structure begins to work in parallel: the CNN (Convolutional Neural Network) branch specifically handles static spatial features, taking into account local underlying surface data containing information such as terrain slope, elevation, and soil type distribution. It extracts key spatial patterns of landslide-prone areas through convolutional and pooling layers. The LSTM (Long Short-Term Memory) branch specifically handles dynamic time-series features, namely the real-time change sequences of soil moisture, infrasound, and rainfall intensity output from steps S2 and S3. Through the memory and forget gates of the LSTM units, it captures the trends and patterns of these elements evolving over time, such as the saturation process of soil moisture and the decreasing trend of the infrasound dominant frequency. Finally, the model's feature fusion layer weighted couples the spatial features extracted by CNN with the temporal features extracted by LSTM. Through a fully connected layer and a Softmax function, it finally outputs one or more early warning results, such as the probability of flash floods and landslides and the corresponding risk levels (usually divided into four levels: blue, yellow, orange, and red).
[0033] Finally, step S5, namely the sending and handling of the warning information, is executed. When the risk level output in step S4 reaches a preset threshold (e.g., orange or higher), the edge computing controller immediately sends a structured warning message to the remote IoT platform via the wireless communication module. This message includes the unique ID of the device that triggered the warning, the precise latitude and longitude location obtained through the GPS module, the risk level of the warning, and a snapshot of key real-time monitoring data that triggered the warning. After receiving the information, the IoT platform can automatically notify relevant responsible persons via SMS, APP push, etc., and activate the emergency plan.
[0034] To further improve the accuracy and reliability of early warning, this invention also introduces spatiotemporal correlation analysis based on UWB networking as a supplement. In scenarios involving network monitoring of multiple poles, the method further includes the following steps: The edge computing controller, through its UWB module, engages in high-frequency periodic ranging communication with UWB modules on adjacent poles to acquire high-precision relative displacement data between the current pole and its neighbors in real time. Subsequently, the controller performs spatiotemporal correlation analysis on this relative displacement data along with the soil moisture change data and infrasound spectral characteristic data obtained in step S3. This coupled analysis of multidimensional data enables the construction of a more comprehensive picture of disaster evolution.
[0035] For example, when the UWB displacement data of a certain pole is detected to be continuously increasing and exceeding the preset 5cm displacement threshold, while its soil moisture data shows a saturation trend and the infrasound spectrum characteristics show a typical landslide precursor pattern with the main frequency dropping sharply from 5Hz to 2Hz, the controller will determine that the risk of disaster has increased sharply and automatically upgrade the risk level output by the LSTM-CNN model by one level, for example, directly upgrading from a yellow warning to an orange warning, thereby providing an earlier and stronger warning.
[0036] Furthermore, to achieve continuous evolution of the monitoring model and accurate adaptation to different regional characteristics, this invention also includes an adaptive optimization step. The edge computing controller packages each early warning event generated during local real-time monitoring (regardless of whether the sending threshold is reached) and the corresponding multi-source sensor data before and after the triggering warning time into a structured verification sample, which is temporarily stored in local memory. When communication conditions permit (such as during the early morning network off-peak hours), the controller periodically uploads these verification samples as "learning results" to the model training center of the remote IoT platform via the wireless communication module. The platform's automatic training program uses these sample data from the real monitoring environment to verify, correct, and optimize the global early warning model, generating incremental model parameters for the specific monitoring area. Then, the platform sends these incremental parameters to the edge computing controller of pole 1 via downlink. After receiving the incremental parameters, the controller loads them into the locally running LSTM-CNN spatiotemporal fusion early warning model, realizing adaptive, lightweight iterative optimization of the local model. In this way, the model on each pole can continuously learn the unique environmental characteristics and disaster evolution patterns of its location.
[0037] In summary, this invention constructs a flash flood early warning system with real-time edge processing capabilities through the collaborative design of multi-source sensor fusion and edge intelligent computing.
[0038] At the data acquisition level, the FDR soil moisture sensor, rigidly connected infrasonic vibration sensor, and microphone array evenly distributed inside the rain-listening cap are used to accurately perceive the soil saturation gradient, mountain micro-vibration, and rainfall intensity, which solves the shortcomings of traditional monitoring which has a single dimension and is difficult to capture the subtle signals of landslide precursors.
[0039] At the data processing level, the edge computing controller utilizes the Helium vector processing technology of the Cortex-M55 core to perform FFT spectrum analysis of infrasound signals, Kalman filtering of soil moisture data, and beamforming noise reduction of multi-channel audio signals locally, significantly improving the signal-to-noise ratio and feature extraction efficiency of the original data.
[0040] At the early warning decision-making level, by running a pruned and compressed LSTM-CNN spatiotemporal fusion model, multi-source feature data is coupled and analyzed in real time at the edge, outputting second-level early warning results, completely eliminating the dependence on cloud computing power and shortening the early warning delay from minutes to seconds.
[0041] In addition, the dual-function design of the UWB module not only enables centimeter-level displacement monitoring in multi-pole networking, but also significantly reduces the difficulty of equipment maintenance in complex mountainous environments through radar map positioning and buzzer guidance functions of the handheld device.
[0042] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indication will also change accordingly.
[0043] Furthermore, in this invention, descriptions involving terms such as "first," "second," and "a" are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.
[0044] In this invention, unless otherwise explicitly specified and limited, the terms "connection," "fixed," etc., should be interpreted broadly. For example, "fixed" can mean a fixed connection, a detachable connection, or an integral part; it can mean a mechanical connection or an electrical connection; it can mean a direct connection or an indirect connection through an intermediate medium; it can mean the internal communication of two components or the interaction between two components, unless otherwise explicitly limited. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.
[0045] Furthermore, the technical solutions of the various embodiments of the present invention can be combined with each other, but only if they are feasible for those skilled in the art. If the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.
Claims
1. A flash flood early warning and monitoring device based on rainfall sound and soil moisture, characterized in that, include: The pole 1 has an underground portion 11 buried in the ground and an above-ground portion 12 located on the ground. The multi-layer soil moisture sensor 2 includes at least two sensor probes that are fixed vertically at intervals to the underground portion 11 of the pole 1, for collecting soil moisture content data at different depths. The infrasound vibration sensor 3 is rigidly connected to the inside of the pole 1 and is used to collect infrasound signals generated by micro-vibrations of the mountain. The rain-listening cap 4 is a conical thin-walled shell fixed to the top of the upright 1; The microphone array 5 includes multiple microphones evenly distributed inside the rain cap 4, used to collect audio signals generated by raindrops impacting the rain cap 4; The UWB module is located on the ground part 12 and is used for ranging communication with external devices; An edge computing controller is located on the above-ground part 12 and is electrically connected to the multi-layer soil moisture sensor 2, the infrasound vibration sensor 3, the microphone array 5 and the UWB module respectively. It is used to receive and process the soil moisture data, the infrasound signal, the audio signal and the ranging data obtained through the UWB module, and run a preset early warning model locally to output early warning results. The wireless communication module is electrically connected to the edge computing controller and is used to send the early warning result to a remote platform.
2. The flash flood early warning and monitoring device based on rainfall sound and soil moisture as described in claim 1, characterized in that, The UWB module is also configured to network and communicate with UWB modules on adjacent poles, calculate the relative displacement data between adjacent poles in real time, and send the relative displacement data to the edge computing controller.
3. The flash flood early warning and monitoring device based on rainfall sound and soil moisture as described in claim 1, characterized in that, The edge computing controller is configured to run machine learning-based signal processing algorithms locally, including: performing a fast Fourier transform on the infrasound signal to extract effective frequency band features, performing Kalman filtering on the soil moisture data to remove measurement noise, and performing a beamforming algorithm on the audio signal to enhance the rain sound signal and suppress environmental noise.
4. The flash flood early warning and monitoring device based on rainfall sound and soil moisture as described in claim 3, characterized in that, The edge computing controller is also configured to run a pruned and quantized compressed LSTM-CNN spatiotemporal fusion early warning model locally. The LSTM-CNN spatiotemporal fusion early warning model takes the soil moisture content data, the characteristic frequency band data of the infrasound signal, the rain sound intensity data of the audio signal, and the relative displacement data as inputs, and outputs the flash flood and landslide risk level.
5. A flash flood early warning and monitoring device based on rainfall sound and soil moisture as described in claim 1, characterized in that, The multi-layer soil moisture sensor 2 includes three frequency domain reflectance (FDR) soil moisture sensor probes, which are fixed at three preset depths in the underground part of the pole.
6. The flash flood early warning and monitoring device based on rainfall sound and soil moisture as described in claim 1, characterized in that, The infrasonic vibration sensor 3 is rigidly fixed inside the pole 1 by a metal bracket, and the sensitive axis is set vertically downward.
7. A flash flood early warning and monitoring device based on rainfall sound and soil moisture as described in claim 1, characterized in that, The microphone array 5 consists of four microphones, which are evenly distributed in a circular pattern and fixed to the inner wall of the rain-listening cap 4.
8. A flash flood early warning monitoring method based on rainfall sound and soil moisture, applied to the flash flood early warning monitoring device based on rainfall sound and soil moisture as described in any one of claims 1 to 7, characterized in that, Includes the following steps: S1: Real-time acquisition of multi-layer soil moisture content data through at least two soil moisture sensor probes, and real-time acquisition of infrasound signals generated by micro-vibration of the mountain through an infrasound vibration sensor. S2: The multi-channel audio signal generated by raindrops impacting the rain-listening cap is acquired in real time through a microphone array, and the edge computing controller performs beamforming algorithm on the multi-channel audio signal to output rain sound intensity characteristic data; S3: The edge computing controller performs a fast Fourier transform on the infrasound signal to extract the spectral feature data of the effective frequency band of the infrasound, and performs Kalman filtering on the multi-layer soil moisture content data to output soil moisture change data. S4: The soil moisture change data, infrasound spectrum feature data and rain sound intensity feature data are input into the LSTM-CNN spatiotemporal fusion early warning model through the edge computing controller. The LSTM layer extracts the temporal change features, the CNN layer extracts the spatial coupling features of multi-source data, and the feature fusion layer outputs the risk level of flash floods and landslides. S5: When the risk level reaches the preset threshold, the warning information containing the risk level and device location information is sent to the remote IoT platform through the wireless communication module.
9. A flash flood early warning and monitoring method based on rainfall sound and soil moisture as described in claim 8, characterized in that, The entire operation of the flash flood early warning and monitoring device also includes the following steps: Real-time relative displacement data between adjacent poles is obtained through periodic ranging communication between UWB modules; Spatiotemporal correlation analysis was performed on relative displacement data, soil moisture change data, and infrasound spectral characteristic data. When the relative displacement data exceeds the preset displacement threshold, the soil moisture change data shows a saturation trend, and the infrasound spectrum characteristics show a sudden drop in the dominant frequency, the output risk level will be upgraded by one level.
10. A flash flood early warning and monitoring method based on rainfall sound and soil moisture as described in claim 8, characterized in that, The edge computing controller also performs the following adaptive optimization steps: The early warning events generated by real-time monitoring and their corresponding multi-source sensor data are packaged into verification samples and periodically uploaded to a remote IoT platform via a wireless communication module. Receive incremental model parameters from the remote IoT platform after optimizing and updating the early warning model parameters based on verification samples; The incremental model parameters are loaded into the LSTM-CNN spatiotemporal fusion early warning model for adaptive iterative optimization.