Integrated monitoring boundary marker system based on multimodal sensing and redundant transmission

Abstract

This application discloses an integrated monitoring boundary marker system based on multimodal sensing and redundant transmission, relating to the field of nature reserve monitoring technology. The system includes multiple boundary markers, each containing sensing and imaging equipment, a dual-power management unit, a sound source identification unit, a multi-source data fusion processing unit, and a multi-mode communication unit electrically connected to the system. This system can effectively monitor and accurately transmit monitoring data for protected areas with an average area of ​​82,000 to 83,000 hectares when only 58 to 59 boundary markers are used. It offers high monitoring reliability and reduces data mistransmission.

Description

Technical Field

[0001] This application relates to the field of nature reserve monitoring technology, and in particular to an integrated monitoring boundary marker system based on multimodal sensing and redundant transmission. Background Technology

[0002] Nature reserves are generally characterized by vast areas, complex terrain, and variable climates, which imposes many limitations on traditional monitoring methods in practical applications: physical boundary markers only serve as static markers and cannot achieve dynamic monitoring. Manual patrols are limited by factors such as human resources and geographical conditions, and generally suffer from problems such as low monitoring efficiency, high response delays, and narrow coverage.

[0003] Existing electronic monitoring equipment (such as infrared cameras and single-type sensors) has relatively limited functionality and relies heavily on single-path networks (such as 4G) for data transmission. In remote or signal-free areas, it is prone to becoming a "data island," resulting in the inability to transmit monitoring data in a timely manner. At the same time, these devices lack accurate target identification capabilities and have difficulty distinguishing between different scenarios such as human activities, animal migration, and natural disturbances. This leads to a large number of invalid alarms, which not only require frequent verification by manpower and resources but may also delay the response to real threats.

[0004] Therefore, in response to the above problems, the industry urgently needs an automated monitoring terminal that is flexible and convenient to deploy, has comprehensive sensing dimensions, stable and reliable data transmission, and accurate and intelligent identification and judgment to meet the needs of refined management of nature reserves.

[0005] The information disclosed in the background section is intended only to enhance the understanding of the overall background of the invention and should not be construed as an admission or in any way implying that such information constitutes prior art known to those skilled in the art. Summary of the Invention

[0006] This application addresses the aforementioned technical problems by providing an integrated monitoring boundary marker system based on multimodal sensing and redundant transmission. This system can effectively monitor and accurately transmit monitoring data for protected areas with an average area of ​​82,000 to 83,000 hectares when only 58 to 59 boundary markers are set. The system has high monitoring reliability and can reduce data mistransmission.

[0007] This application provides an integrated monitoring boundary marker system based on multimodal sensing and redundant transmission, comprising: multiple markers, each marker being equipped with sensing and imaging equipment, a dual power supply management unit, a sound source identification unit, a multi-source data fusion processing unit, and a multi-mode communication unit electrically connected to it; The multi-mode communication unit is a communication master control module integrating at least two heterogeneous network communication modules. It is used to transmit data acquired from each channel in real time and monitor the core quality parameters of each communication channel. It adopts a channel selection mechanism based on weighted multi-attribute decision-making, combined with a dynamic threshold adjustment strategy. According to the application scenario requirements, it selects at least one of the public network communication module, self-organizing network communication module, or satellite communication module to achieve intelligent channel selection and switching. The core quality parameters of each communication channel include: signal strength, bit error rate, and transmission delay. Dual power management unit is used to power the multi-mode communication unit, sound source identification unit, multi-source data fusion processing unit, and sensing and imaging equipment; Sensing and imaging equipment includes: an integrated SHT30 temperature and humidity sensor, an air quality sensor, and a 1080P infrared night vision camera; The sound source recognition unit is used to collect ambient sounds on site, perform noise reduction preprocessing, extract spectral features and intelligent classification, and identify four categories of preset scene results. When the recognition result is human activity sound or abnormal threat sound, and the confidence level of the obtained result is greater than 85%, a trigger signal is immediately sent to the sound source main control module to start the high-definition camera to capture images and record short videos. Simultaneously, the voiceprint recognition result + real-time device location + timestamp + image / video multimedia data are integrated, packaged into a high-priority data packet, and quickly transmitted to the monitoring center through the multi-mode communication unit. The multi-source data fusion processing unit is used to analyze the voiceprint and image data packets uploaded by the sound source identification unit through multi-source data fusion algorithms to achieve integrated fusion analysis of environmental perception data, sound source identification data, and image data. In response to the uncertainty of multi-sensor data, the DS evidence theory is used for fusion reasoning to achieve complementarity and verification of multi-dimensional monitoring information.

[0008] Preferably, the sound source identification unit is used to implement the following steps: Step S31: Sound Acquisition and Preprocessing: A combination of continuous acquisition and triggered acquisition is used to capture ambient sound from all directions. Continuous acquisition uses a high-sensitivity microphone array to collect ambient sound, while triggered acquisition uses low power consumption to collect sound at a level ≥60dB. After acquisition, the audio is preprocessed using adaptive noise reduction and sound enhancement algorithms. An improved spectral subtraction combined with a wavelet threshold noise reduction algorithm is used to filter environmental interference and improve the sound signal quality to obtain a noise-reduced preprocessed sound source. Dynamic environmental acoustic baseline adaptive filtering mechanism: In response to the regional and seasonal regular noise in nature reserves, the system designs a background sound adaptive learning method based on Gaussian mixture model to completely eliminate false triggers caused by regular natural interference. Step S32: Feature Extraction and Classification: Based on a pre-trained lightweight neural network model, spectral features are extracted from the noise-reduced preprocessed sound source signal to obtain an audio spectrogram. The obtained audio spectrogram is then intelligently classified into four preset scenarios: human activity sounds, specific animal sounds, natural interference sounds, and abnormal threat sounds. Human activity sounds include talking, footsteps, and vehicle engine sounds. Natural interference sounds include wind, rain, and thunder. Abnormal threat sounds include chainsaw sounds, gunshots, and illegal fishing sounds. Step S33: Event Triggering and Data Packaging: When the recognition result is human activity sound or abnormal threat sound, a trigger signal is immediately sent to the sound source main control module to start the high-definition camera to capture images and record short videos; the sound source main control module synchronously integrates the voiceprint recognition result + real-time device location + timestamp + image / video multimedia data, packages it into a high-priority data packet, and quickly transmits it to the monitoring center through the multi-mode communication unit.

[0009] Preferably, step S31 includes the following steps: Step S311: Noise estimation: Identify silent segments using a speech activity detection algorithm and estimate the noise power spectral density based on a statistical model; Step S312: Spectral subtraction processing: Subtract the power spectrum of the noisy signal to suppress stationary noise; Step S313: Wavelet Threshold Optimization: Perform wavelet decomposition on the spectral subtraction signal and use an adaptive threshold function. , where σ is the noise standard deviation, N is the signal length, high-frequency coefficients are processed, and effective signal components are retained; Step S314: Sound Enhancement: A loudness compensation algorithm based on the characteristics of human hearing is adopted to improve the recognition of weak target sounds, improve the accuracy and reliability of sound monitoring results, and avoid interference from background noise. Step S315: Environmental baseline self-learning: During the first 72 hours of the initial deployment of the boundary stake, the system automatically and silently collects long-sequence audio data from the surrounding area, extracts Mel frequency cepstral coefficients, and uses GMM to perform probability density modeling of the normal background noise specific to the stake location, generating a unique static environmental acoustic baseline for the boundary stake. Step S316: Dynamic Adaptive Filtering: In normal monitoring, the acoustic feature distance between the input audio features and the static environmental soundprint baseline is calculated in real time, specifically by Mahalanobis distance. When the obtained acoustic feature distance is less than a set threshold, it is judged as normal natural white noise and directly filtered out without triggering the subsequent high-power classification model. Only when the acoustic feature distance breaks the baseline balance, such as the sudden appearance of chainsaw sound or human footsteps, is it judged as a sudden sound and accurately classified.

[0010] Preferably, step S32 includes the following steps: Based on MobileNet, an attention mechanism (SE) module and a knowledge distillation method are introduced to construct a lightweight sound source classification optimization model with high accuracy and low computational consumption. Step S321: Model structure optimization: Feature extraction layer: Use depthwise separable convolution to reduce computation, insert SE module to enhance the weight allocation of key spectral features; Step S322: Knowledge distillation: Using the complex ResNet-50 model as the teacher model and MobileNet as the student model, the distillation loss function with temperature coefficient adjustment is used to make the student model maintain a lightweight design while achieving classification accuracy close to that of the teacher model. Step S323: Model Quantization: Using INT8 quantization technology, the model parameters are converted from 32-bit floating-point type to 8-bit integer type, reducing memory usage and inference latency, and adapting to edge computing chips; Step S324: Training and Optimization: After optimizing the model using a hybrid data augmentation strategy, a sound recognition model is obtained. The hybrid data augmentation strategy includes: spectral masking, time stretching, and noise superposition to improve the model's adaptability to changes in the field environment. FocalLoss is used to solve the sample imbalance problem and improve the recognition rate of minority classes, such as abnormal threatening sounds. Step S325: Multimodal fusion triggering algorithm: Combine sound signals and environmental sensor data to perform multimodal fusion decision-making to reduce the false trigger rate. Environmental data includes: temperature, humidity, wind speed, and air quality. The fusion logic of sound signals and environmental data is as follows: Let the confidence level of the sound source identification result be C_s, 0≤C_s≤1, and the environmental interference coefficient be C_e, which is calculated based on wind speed and rainfall, 0≤C_e≤1. The greater the interference, the closer C_e is to 1. Then the final trigger confidence level C=C_s×(1-0.3×C_e). When C≥0.85, the camera captures an image; when 0.7≤C<0.85, the sound acquisition time is extended by 3 seconds, and the process returns to step S324 for re-evaluation; when C<0.7, the trigger is deemed invalid, and the sound is only cached locally and logged; when C≥0.85, the sound is considered a high-confidence sound, and the probability of the target sound corresponding to human activity or abnormal threat sounds is relatively high. Step S326: When the sound source recognition unit identifies a target sound with high confidence C, it immediately sends an interrupt trigger signal to the sound source main control module. The sound source main control module then starts the high-definition camera to capture images and record short videos, and calls the multi-source data fusion algorithm to jointly judge the voiceprint, image, and environmental sensor. If C ≥ 0.85, the fused data is determined to be a real threat. The sound source main control module generates a high-priority data packet, calls the multi-mode communication unit to start the optimal channel transmission, and records the event log at the same time. If the system is in low-power mode, the sound source main control module decides whether to start the camera or only upload the voiceprint recognition result to reduce energy consumption based on the power status.

[0011] Preferably, the multi-source data fusion algorithm is used to implement the following steps: Step S41: Evidence Source Construction: The voiceprint recognition results and image recognition results obtained by the sound source recognition unit, as well as the temperature and humidity sensor data and air quality sensor data obtained by the environmental sensor, are used as independent evidence sources. The basic probability allocation (BPA) function is established for each of them to obtain the evidence value. In the illegal logging scenario, chainsaw operations and vehicle traffic will bring about local temperature increases, humidity fluctuations, and increases in dust and PM2.5 concentrations. Therefore, this method combines environmental parameters to make accurate judgments and improve monitoring accuracy.

[0012] Step S42: Evidence Conflict Handling: An improved DS synthesis rule is adopted. When the conflict coefficient K between evidence values ​​is greater than 0.5, K = Σm1(A_i) × m2(A_j), A_i ∩ A_j = ∅, where i is the data obtained in the i-th second and j is the data obtained in the j-th second, i ≠ j; a conflict weight allocation factor is introduced to avoid distortion of the synthesis result. Step S43: Weighted synthesis: Based on the reliability of each evidence source, set the reliability weight of sound source identification to 0.4, the reliability weight of image recognition to 0.35, and the reliability weight of environmental data to 0.25, perform weighted synthesis, and obtain the final event confidence score. Event confidence score = 0.4 * sound source evidence value + 0.35 * image recognition evidence value + 0.25 * environmental data evidence value; Step S44: Set decision rules: When the confidence level of the merged event is ≥0.8, it is judged as a high-confidence event and reported immediately; When 0.6 ≤ confidence level < 0.8, it is judged as a medium confidence event, and secondary verification is initiated to increase the camera recording time; When the confidence level is less than 0.6, it is judged as a low-confidence event, and the data is stored locally only. Step S45: Cross-modal false alarm prevention verification based on sound-visual-motion spatiotemporal joint: When the sound source identification unit detects a sudden sound, or the environmental sensor detects the movement of a heat source, the timestamp T0 is recorded at this moment, and the low power consumption state is switched to the warning state. Audio data within T0±5 seconds is extracted, and the obtained sudden sound is matched with the voiceprint tags of human activities or abnormal threats through a lightweight sound source classification model. If the match is unsuccessful, it is determined to be natural animal activity, automatically downgraded to log recording, and the alarm is terminated. If a match is successful, the voiceprint is classified as high-risk. The main control module wakes up the high-definition camera in milliseconds, extracts key frames, and runs a lightweight target detection algorithm to detect whether a corresponding entity exists in the image.

[0013] Preferably, the multimode communication unit is used to implement the following steps: Step S11: Channel parameter normalization processing: The collected signal strength, bit error rate and transmission delay are normalized and mapped to the [0,1] interval to eliminate the difference in units. The signal strength normalization adopts forward mapping, while the bit error rate and transmission delay adopt reverse mapping. Step S12: Normalize the signal, bit error rate, and transmission delay respectively. Signal strength normalization: Let the measured signal strength at time i be S, the minimum value obtained in the continuous measurement period be S_min, and the maximum value obtained in the continuous measurement period be S_max. The normalized value S_norm = (S - S_min) / (S_max - S_min). Bit error rate normalization: Let the measured bit error rate at time i be E, and the maximum value obtained during the continuous measurement period be E_max. The normalized value E_norm = 1 - E / E_max Transmission delay normalization: Let the measured transmission delay at time i be D, and the maximum value obtained during the continuous measurement period be D_max. The normalized value is D_norm = 1 - D / D_max. Step S13: Dynamic weight allocation: The basic weights of each parameter are determined using the analytic hierarchy process (AHP), and then dynamically adjusted in combination with data priority. The weights for high-priority data are: signal strength weight is 0.4, transmission delay weight is 0.35, and bit error rate weight is 0.25. Low-priority data is assigned weights as follows: signal strength weight is 0.3, bit error rate weight is 0.3, and transmission delay weight is 0.4. Step S14: Channel comprehensive score: Calculate the comprehensive score C of each channel after weighting each parameter, C=ω_S×S_norm+ω_E×E_norm+ω_D×D_norm, where ω_S is the weight of signal strength; ω_E is the weight of bit error rate; ω_D is the weight of transmission delay, and select the channel with the highest comprehensive score as the optimal transmission channel. Step S15: Determine the priority of the output data based on the sound source identification unit. When transmitting high-priority data, monitor the status of all channels in real time. If at least two channels have a comprehensive score ≥ 0.7, and the transmission delay of any single channel is ≤ 300ms and the bit error rate is ≤ 0.5%, then start multi-channel concurrent transmission to avoid packet loss. Use a data fragmentation redundancy coding mechanism to fragment the data packets and transmit them through different channels. The receiving end merges and decodes the fragments to ensure data integrity. If only a single channel has a comprehensive score ≥ 0.7, and the transmission delay of any single channel is ≤ 300ms and the bit error rate is ≤ 0.5%, then select the optimal channel for transmission. When transmitting low-priority data, monitor the status of the public network channel. If the public network channel is available, use the public network channel for transmission. If the public network channel is unavailable, switch to the self-organizing network channel for transmission. Step S16: Channel switching: Monitor the comprehensive score of the current transmission channel in real time. If the score is lower than 0.5 for 5 seconds or a transmission interruption occurs, immediately trigger the channel reselection process, switch to the suboptimal channel, and record the channel switching log. A transmission interruption is defined as the loss of 3 consecutive data packets.

[0014] Preferably, the dual power management unit includes: a power control module, a main power supply, a backup power supply, and an intelligent power switching and power management circuit; The main power control module and the intelligent power switching and power management circuit are electrically connected to the main power supply and the backup power supply, respectively; the main power supply and the backup power supply are electrically connected to the main control module, the multi-mode communication unit, the sound source recognition unit, the sensing and imaging equipment, and the multi-source data fusion processing module, respectively. The main power supply includes: solar photovoltaic panels installed on the boundary piles and electrically connected to the energy storage battery pack to collect light energy in order to adapt to the outdoor lighting environment. The energy storage battery pack is connected to the power supply of each unit to provide continuous power for the normal operation of the system. Intelligent power switching and power management circuit: Used to monitor the voltage, remaining power, and charging / discharging status of the energy storage battery pack in real time. When the energy storage battery's power level falls below a first preset threshold, it automatically switches to backup power supply. Simultaneously, it sends a power status alarm message to the monitoring center via a multi-mode communication unit. The alarm message includes the reason for the switchover and the remaining power level of the backup power supply, facilitating timely handling by maintenance personnel. The first preset threshold is voltage < 12V or power level < 30%. When the backup power supply level is less than the second preset threshold, a low power emergency alarm message is sent to the monitoring center and then the system enters low power mode, retaining only the core function of sound source recognition. If the backup power supply level is between the first and second preset thresholds, the backup power supply is maintained and the main power supply status is continuously monitored. When the main power supply voltage recovers, the power supply is switched back to the main power supply and monitoring continues. The main power supply voltage recovers to a level greater than 13V and the power supply level is greater than 40%. The second preset threshold is when the backup power supply level is less than 10%.

[0015] Preferably, the power supply main control module is used to integrate remaining power estimation, power supply prediction, and dynamic threshold adjustment to achieve intelligent power management and fault early warning, specifically including the following steps: Step S21: Accurate estimation of remaining power: The Kalman filter algorithm is used to optimize the SOE estimation accuracy, and the SOE is dynamically corrected by combining the charge and discharge characteristic curves of the battery and the temperature compensation coefficient. The basic model used for SOE estimation is: SOE(t)=SOE(0)+(1 / C_n)×∫(I(t)×η(t,T))dt, where C_n is the rated capacity of the battery, I(t) is the charge and discharge current, and η(t,T) is the charge and discharge efficiency at temperature T. Then, the SOE estimation result is optimized by Kalman filtering: the initial value of SOE is predicted by the state equation, the observation equation is constructed using the measured values ​​of voltage and current, and the estimation error of SOE is dynamically corrected to make the SOE estimation accuracy ≤±3%. Step S22: Main power supply capacity prediction: Based on historical sunshine data, weather forecast information and photovoltaic power generation model, predict the power generation required by the main power supply and the available power supply duration for the next 24 hours as the power supply prediction result P(t). Power generation model: P(t) = P_STC × G(t) / G_STC × [1 + α(T(t) - T_STC)], where P_STC is the photovoltaic power under standard test conditions, G(t) is the real-time irradiance, G_STC is the standard irradiance, α is the power temperature coefficient, T(t) is the real-time battery temperature, and T_STC is the standard temperature; Step S23: Dynamic threshold adjustment: Adjust the power switching threshold according to the power supply forecast results; if it is predicted that there will be sufficient sunshine in the next 24 hours and the power generation of the solar photovoltaic panel is ≥ 1.5 times the average daily power consumption of the system obtained from historical data, then the first threshold is lowered to 25%; if it is predicted that there will be cloudy or rainy weather in the next 24 hours and the power generation of the solar photovoltaic panel is ≤ 0.5 times the average daily power consumption of the system obtained from historical data, then the first threshold is raised to 35%, and the backup power supply is pre-activated in advance to ensure seamless switching. Step S24: Low power warning and energy saving control: When the backup power is lower than the second preset threshold by 10%, the graded energy saving mode is activated; Level 1 energy saving: non-core sensing devices are turned off and the camera capture frequency is reduced; Level 2 energy saving: only the sound source recognition core function and emergency communication module are retained, and other devices are put into hibernation until the main power supply is restored. Level 1 energy saving consumes 10-5% of electricity; Level 2 energy saving consumes less than 5% of electricity.

[0016] Preferably, 58 to 59 piles are evenly distributed within a protected area of ​​82,000 to 83,000 hectares.

[0017] Preferably, it includes: a self-organizing network communication unit, which is installed inside the pile body. The piles are interconnected through the self-organizing network communication unit to build a distributed monitoring network. When a single boundary pile is in an area without public network signal and cannot be directly connected to the public network, it transmits data to a gateway pile with public network communication capability through the self-organizing network link of adjacent boundary piles via multi-hop forwarding. The gateway pile then aggregates the data and uploads it to the monitoring center, achieving full coverage of the monitoring network with no data transmission blind spots. The ad hoc network communication unit is used to implement the distributed ad hoc network routing optimization algorithm, which includes the following steps: Step S51: Select nodes with remaining energy, a link communication quality score ≥ 0.6, and hop count. Use the improved AODV routing protocol. When a node broadcasts a route request, it carries its own energy status and link quality information. Intermediate nodes calculate the path cost C = α × E_cost + β × L_cost + γ × H_cost based on the received RREQ information, where E_cost is the energy cost, L_cost is the link quality cost, H_cost is the hop count cost, α = 0.4, β = 0.4, and γ = 0.2. Select the path with the lowest cost as the optimal route. Step S52: Periodically monitor the energy status and link quality of each node on the routing path. When the energy of a node is lower than 15% or the link quality score is lower than 0.5, trigger route rediscovery to avoid route failure.

[0018] The beneficial effects that this application can produce include: 1) The integrated monitoring boundary marker system based on multimodal sensing and redundant transmission provided in this application has a revolutionary improvement in transmission reliability: by combining multimodal redundant communication architecture with intelligent link decision algorithm, it solves the problem of "data island" in areas without public network signal, realizes continuous transmission and complete retention of monitoring data in complex terrain and harsh environment, and the transmission success rate and stability are significantly better than traditional single-channel transmission schemes.

[0019] 2) The integrated monitoring boundary marker system based on multimodal sensing and redundant transmission provided in this application has doubled the power supply security and continuity: the dual power supply redundancy design and intelligent power management algorithm effectively cope with the main power supply interruption caused by severe weather such as continuous rain and extreme low temperature, ensure the long-term stable operation of the equipment, completely avoid the monitoring gap caused by power outages, and ensure the continuity of monitoring work.

[0020] 3) The integrated monitoring boundary marker system based on multimodal perception and redundant transmission provided in this application significantly improves the level of monitoring intelligence and accuracy: relying on the advanced sound source identification optimization algorithm system and multimodal data fusion algorithm, it has achieved a leapfrog upgrade from the traditional "alarm upon any movement" to "alarm after accurate identification", which greatly reduces the invalid alarm rate, reduces the verification pressure of operation and maintenance personnel, and improves the response speed and handling efficiency to real threats, providing accurate decision-making basis for ecological protection and violation control. Attached Figure Description

[0021] Figure 1 A schematic diagram of an integrated monitoring boundary marker system based on multimodal sensing and redundant transmission in at least one embodiment provided in this application; Figure 2 This application provides a functional flowchart of the multi-mode communication unit; Figure 3 A functional flowchart of a dual power management unit in at least one embodiment provided in this application; Figure 4 A functional flowchart of the sound source identification unit in at least one embodiment provided in this application; Figure 5 A functional flowchart of the multi-source data fusion processing unit provided in at least one embodiment of this application; Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0023] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0024] Technical means not detailed in this application and not used to solve the technical problems of this application are all set according to common general knowledge in the field, and multiple common general knowledge setting methods can be implemented.

[0025] Example 1 See Figures 1-5 The integrated monitoring boundary marker system based on multimodal sensing and redundant transmission provided in this application includes: a main control module, a multimodal communication unit, a dual power management unit, a sound source identification unit, a sensing and imaging device, and a multi-source data fusion processing module; the main control module is electrically connected to the multimodal communication unit, the dual power management unit, the sound source identification unit, and the sensing and imaging device respectively. Main control module The multi-mode communication unit integrates at least two heterogeneous network communication modules into a main control module for real-time monitoring of core quality parameters (including signal strength, bit error rate, and transmission delay) of each communication channel. It adopts a channel selection mechanism based on weighted multi-attribute decision (WMADM) and combines it with a dynamic threshold adjustment strategy. Depending on the application scenario requirements, it selects at least one of a public network communication module, a self-organizing network communication module, or a satellite communication module to achieve intelligent channel optimization and switching. It integrates a 4G / 5G public network communication module, a LoRa self-organizing network module, and a Beidou short message communication module. The LoRa module has a communication range of 3-5 kilometers, and the Beidou short message module supports emergency data transmission in environments without a public network. The dual power management unit includes: a power control module, a main power supply, a backup power supply, and intelligent power switching and management circuitry, enabling stable power supply around the clock. The main power supply includes: solar photovoltaic panels and energy storage batteries installed on the boundary piles; the solar photovoltaic panels are electrically connected to the energy storage battery packs, and the energy storage battery packs are connected to the power supply of each unit to provide continuous power for the normal operation of the system; Backup power: High-energy-density battery packs are used, featuring large capacity, long battery life, and strong environmental adaptability; Intelligent power switching and power management circuit: Real-time monitoring of the voltage, remaining power and charging / discharging status of the energy storage battery pack. When the energy storage battery power is lower than the first preset threshold (e.g., remaining power 30%), it automatically switches to backup power supply. At the same time, it sends power status alarm information to the monitoring center through the multi-mode communication unit. The alarm information includes the reason for switching and the remaining power data of the backup power supply, which facilitates timely handling by operation and maintenance personnel.

[0026] The main power supply uses a combination of a 20W high-efficiency solar photovoltaic panel and a 12V / 100Ah lithium iron phosphate battery, while the backup power supply uses a 12V / 50Ah high-energy-density lithium battery pack. The power management circuit uses TI's BQ24725 charging management chip and STM8S103 microcontroller to realize real-time monitoring and intelligent switching of power status. The sound source recognition unit is used to collect ambient sounds on site, perform noise reduction preprocessing, extract spectral features and intelligent classification, and identify four categories of preset scene results. When the recognition result is "human activity sound" or "abnormal threat sound" and the confidence level of the result is greater than 85%, it immediately sends a trigger signal to the sound source main control module to start the high-definition camera to capture images and record short videos. Simultaneously, it integrates "voiceprint recognition result + real-time device location + timestamp + image / video multimedia data", packages it into a high-priority data packet, and quickly transmits it to the monitoring center through the multi-mode communication unit. 58-59 stakes are evenly distributed throughout the 82,000-83,000 hectare protected area. This arrangement allows for effective monitoring of the protected area.

[0027] The multi-source data fusion processing module is used to analyze the voiceprint and image data packets uploaded by the sound source recognition unit through multi-source data fusion algorithms to achieve integrated fusion analysis of environmental perception data, sound source recognition data and image data. In response to the uncertainty of multi-sensor data, the DS evidence theory is used for fusion reasoning to achieve complementarity and verification of multi-dimensional monitoring information. Link decision algorithm: Real-time acquisition of signal strength (threshold set ≥ -70dBm), bit error rate (threshold set ≤1%), and transmission delay (threshold set ≤500ms) of each communication channel. High-priority data preferentially selects satellite communication or 4G / 5G public network communication. If neither is available, multi-node self-organizing network jump transmission is started. Low-priority data preferentially selects 4G / 5G public network communication. If there is no public network signal, it switches to LoRa self-organizing network transmission. Power management algorithm: The first threshold for the main power supply is set at 30% remaining power and 12V voltage. When the voltage drops below the first threshold, the power supply switches to the backup power supply. The second threshold for the backup power supply is 10% remaining power. When the voltage drops below the second threshold, an emergency alarm is sent and a low-power mode is activated, retaining only the sound source identification and emergency communication functions.

[0028] Main control module: The STM32H7 series high-performance microcontroller is selected, and an integrated FPGA chip is used for audio signal preprocessing. It has powerful computing capabilities and data processing efficiency to meet the needs of edge computing. Sound source recognition unit: Employs a 4-array high-sensitivity microphone (sensitivity -40dBV / Pa, frequency response range 20Hz-20kHz), and an embedded AI processing unit using the NVIDIA Jetson Nano developer kit, providing efficient neural network inference capabilities. A lightweight CNN model is deployed based on the TensorFlow Lite framework. The dataset consists of 3000 real-world sound samples (covering four categories: human activity sounds, animal sounds, natural disturbance sounds, and unusual threat sounds, with 750 samples in each category). After training, the resulting model achieves an accuracy of ≥92% in recognizing the sounds of objects in the test set. The detected sound can be any of the following: human activity sounds, animal sounds, natural disturbance sounds, or unusual threat sounds. Includes a high-sensitivity microphone array and an embedded AI processing unit, providing sound acquisition, intelligent analysis, and event triggering functions.

[0029] Sensing and Imaging Equipment: Integrates SHT30 temperature and humidity sensor, MQ-135 air quality sensor, and a 1080P infrared night vision camera for high-definition camera, supporting low-light capture and short video recording.

[0030] Functions and algorithms of each module: 1. Multimode communication unit It integrates at least two heterogeneous network communication modules, and selects at least one of the following according to the application scenario requirements: public network communication module (such as 4G / 5G), self-organizing network communication module (such as BLEMesh / LoRa), or satellite communication module (such as Beidou short message).

[0031] The multi-mode communication unit includes: a communication master control module; the master control module embeds an intelligent link decision algorithm, which monitors the core quality parameters of each communication channel in real time (including: signal strength, bit error rate, transmission delay) and processes them according to the following specific algorithm steps. The algorithm principle is: adopting a channel selection mechanism based on weighted multi-attribute decision (WMADM) and combining it with a dynamic threshold adjustment strategy to realize intelligent channel selection and switching.

[0032] The core quality parameter processing steps include: Step S11: Channel parameter normalization processing: The collected signal strength (dBm), bit error rate (%), and transmission delay (ms) are normalized and mapped to the [0,1] interval to eliminate the difference in dimensions. The signal strength normalization adopts a forward mapping (the larger the value, the better the quality), while the bit error rate and transmission delay adopt a reverse mapping (the smaller the value, the better the quality). Step S12: Normalize the signal, bit error rate, and transmission delay respectively. Signal strength normalization: Let the measured signal strength at time i be S, the minimum value obtained during the continuous measurement period be S_min (e.g., -110dBm), and the maximum value obtained during the continuous measurement period be S_max (e.g., -50dBm). The normalized value S_norm = (S - S_min) / (S_max - S_min). Bit error rate normalization: Let the measured bit error rate at time i be E, and the maximum value obtained during the continuous measurement period be E_max (e.g., 5%). The normalized value E_norm = 1 - E / E_max Transmission delay normalization: Let the measured transmission delay at time i be D, and the maximum value obtained during the continuous measurement period be D_max (e.g., 2000ms). The normalized value is D_norm = 1 - D / D_max. Step S13: Dynamic weight allocation: The basic weights of each parameter are determined using the Analytic Hierarchy Process (AHP), and then dynamically adjusted in combination with data priority. The weights allocated to high-priority data are: signal strength weight 0.4, transmission delay weight 0.35, and bit error rate weight 0.25; the weights allocated to low-priority data are: signal strength weight 0.3, bit error rate weight 0.3, and transmission delay weight 0.4.

[0033] Step S14: Channel comprehensive score: Calculate the comprehensive score C of each channel after weighting each parameter, C=ω_S×S_norm+ω_E×E_norm+ω_D×D_norm (where ω_S is the weight of signal strength; ω_E is the weight of bit error rate; ω_D is the weight of transmission delay), and select the channel with the highest comprehensive score as the optimal transmission channel.

[0034] Step S15: Multi-channel concurrent triggering condition: When high-priority data transmission occurs, if the overall score of at least two channels is ≥0.7, and the transmission delay of any single channel is ≤300ms and the bit error rate is ≤0.5%, then multi-channel concurrent transmission is initiated. The data fragmentation redundancy coding (RS coding) mechanism is adopted to fragment the data packets and transmit them through different channels. The receiving end merges and decodes the fragments to ensure data integrity.

[0035] Step S16: Channel switching mechanism: Real-time monitoring of the comprehensive score of the current transmission channel. When the score is below 0.5 for 5 seconds or a transmission interruption occurs (3 consecutive data packet losses), the channel reselection process is immediately triggered to switch to the suboptimal channel and the channel switching log is recorded.

[0036] Step S17: Based on the data type, define priority classification rules (where alarm data and multimedia data are high priority, and log data and normal environment data are low priority) and the real-time channel status, automatically decide the optimal transmission channel, or start a multi-channel concurrent transmission mode to ensure that high-priority critical data is transmitted without loss and with low latency.

[0037] The link decision algorithm of this module is scheduled and executed in real time by the main control module. The main control module dynamically adjusts the channel selection strategy of this module according to the current data priority, thereby ensuring accurate and effective data transmission in the "data island" state.

[0038] 2. Dual power management unit The intelligent power management and prediction algorithm is integrated into the main power control module. The algorithm's principle is to combine Remaining Energy Estimation (SOE), power supply prediction, and dynamic threshold adjustment technologies to achieve intelligent power management and fault early warning. This includes the following steps: Step S21: Accurate estimation of remaining energy (SOE): The Kalman filter (EKF) algorithm is used to optimize the SOE estimation accuracy, and the SOE is dynamically corrected by combining the charge-discharge characteristic curve of the battery and the temperature compensation coefficient. The basic model used for SOE estimation is: SOE(t)=SOE(0)+(1 / C_n)×∫(I(t)×η(t,T))dt (where C_n is the rated capacity of the battery, I(t) is the charge-discharge current, and η(t,T) is the charge-discharge efficiency at temperature T). Then, the SOE estimation result is optimized by Kalman filtering: the initial value of SOE is predicted by the state equation, and the observation equation is constructed using the measured values ​​of voltage and current to dynamically correct the estimation error of SOE, so that the SOE estimation accuracy is ≤±3%.

[0039] Step S22: Main power supply capacity prediction: Based on historical sunshine data, weather forecast information (short-term weather forecast obtained through satellite communication module) and photovoltaic power generation model, predict the required main power generation and available power supply duration for the next 24 hours as the power supply prediction result P(t).

[0040] Power generation model: P(t) = P_STC × G(t) / G_STC × [1 + α(T(t) - T_STC)] (where P_STC is the photovoltaic power under standard test conditions, G(t) is the real-time irradiance, G_STC is the standard irradiance, α is the power temperature coefficient, T(t) is the real-time cell temperature, and T_STC is the standard temperature;) Step S23: Dynamic Threshold Adjustment: Adjust the power switching threshold based on the power supply forecast results. If sufficient sunshine is predicted for the next 24 hours (solar photovoltaic power generation ≥ 1.5 times the system's average daily power consumption based on historical data), the first threshold is lowered to 25%; if cloudy or rainy weather is predicted for the next 24 hours (solar photovoltaic power generation ≤ 0.5 times the system's average daily power consumption based on historical data), the first threshold is raised to 35%, and the backup power supply is pre-activated in advance to ensure seamless switching.

[0041] Step S24: Low power warning and energy saving control: When the backup power is lower than the second preset threshold (10%), the graded energy saving mode is activated: Level 1 energy saving (power 10~5%): non-core sensing devices are turned off and the camera capture frequency is reduced; Level 2 energy saving (power <5%): only the sound source recognition core function and emergency communication module are retained, and other devices are put into hibernation until the main power supply is restored.

[0042] The above method can effectively ensure the normal power supply to the system, improve the reliability of power supply, effectively extend the power supply time in nature reserves, and improve the power supply efficiency of dual power sources.

[0043] 3. Sound source recognition unit The specific workflow is as follows: Step S31: Sound Acquisition and Preprocessing: A combination of continuous acquisition and triggered acquisition is used to capture ambient sound from all directions. After acquisition, the audio is preprocessed by adaptive noise reduction and sound enhancement algorithms. An improved spectral subtraction combined with wavelet threshold noise reduction algorithm is used to filter environmental interference and improve the sound signal quality to obtain a noise-reduced preprocessed sound source. Dynamic environmental acoustic baseline adaptive filtering mechanism: For regionally and seasonally recurring noises (such as constant flowing water sounds and regular insect chirping) within nature reserves, the system employs a background sound adaptive learning method based on a Gaussian mixture model (GMM) to completely eliminate false triggers caused by regular natural disturbances. This method is particularly suitable for precise noise filtering in complex outdoor environments.

[0044] Step S32: Feature Extraction and Classification: Based on a pre-trained lightweight neural network model (such as MobileNet or SqueezeNet, a customized variant model optimized for audio spectrograms), spectral features are extracted and intelligently classified from the noise-reduced preprocessed sound source signal, accurately identifying it into four preset scenarios: human activity sounds (including talking, footsteps, vehicle engine sounds, etc.), specific animal sounds (such as tiger roars, bird calls, animal migration sounds, etc.), natural disturbance sounds (such as wind sounds, rain sounds, thunder sounds, etc.), and abnormal threat sounds (such as chainsaw sounds, gunshots, illegal fishing sounds, etc.). Step S33: Event Triggering and Data Packaging: When the identification result is "human activity sound" or "abnormal threat sound", a trigger signal is immediately sent to the sound source main control module to start the high-definition camera to capture images and record short videos; the sound source main control module simultaneously integrates "voiceprint recognition result + real-time device location + timestamp + image / video multimedia data", packages it to generate a high-priority data packet, and quickly transmits it to the monitoring center through the multi-mode communication unit.

[0045] The sound source identification optimization algorithm of the sound source identification unit includes the following steps: Step S311: Noise estimation: Identify silent segments using the speech activity detection (VAD) algorithm and estimate the noise power spectral density based on a statistical model; Step S312: Spectral subtraction processing: Subtract the power spectrum of the noisy signal to suppress stationary noise; Step S313: Wavelet threshold optimization: Perform wavelet decomposition on the spectral subtraction signal and use an adaptive threshold function ( (where σ is the noise standard deviation and N is the signal length) processes high-frequency coefficients and retains effective signal components; Step S314: Sound Enhancement: A loudness compensation algorithm based on the characteristics of human hearing is adopted to improve the recognition of weak target sounds, improve the accuracy and reliability of sound monitoring results, and avoid interference from background noise.

[0046] Step S315: Environmental baseline self-learning: During the first 72 hours of the initial deployment of the boundary marker, the system automatically and silently collects long-sequence audio data from the surrounding area, extracts Mel frequency cepstral coefficients (MFCC), and uses GMM to model the probability density of the normal background noise specific to the boundary marker location, generating a unique "static environmental acoustic baseline" for the boundary marker.

[0047] Step S316: Dynamic Adaptive Filtering: In normal monitoring, the acoustic feature distance between the input audio features and the "static environment soundprint baseline" is calculated in real time, specifically by Mahalanobis distance. When the obtained acoustic feature distance is less than a set threshold, it is judged as normal natural white noise and directly filtered out without triggering the subsequent high-power classification model. Only when the acoustic feature distance breaks the baseline balance, such as the sudden appearance of chainsaw sound or human footsteps, is it judged as a sudden sound and the sudden sound is accurately classified.

[0048] This mechanism significantly reduces the frequency of system wake-up due to background noise, achieving an intelligent leap from "blind data collection" to "feature-triggered" data collection. It effectively reduces system power consumption and data upload / communication volume, which helps extend the system's operational lifespan in nature reserves, improves monitoring accuracy, and reduces monitoring manpower costs.

[0049] Step S32 includes the following steps: Based on MobileNet, we introduce an attention mechanism (SE module) and a knowledge distillation method to build a lightweight sound source classification optimization class model with high accuracy and low computational consumption; Step S321: Model structure optimization: Feature extraction layer: Use depthwise separable convolution to reduce computation, insert SE module squeeze-excitation unit to enhance the weight allocation of key spectral features; Step S322: Knowledge distillation: Using the complex ResNet-50 model as the teacher model and MobileNet as the student model, the distillation loss function with a temperature coefficient adjusted (T=4) is used to make the student model maintain a lightweight design while achieving classification accuracy close to that of the teacher model. Step S324: Model Quantization: Using INT8 quantization technology, the model parameters are converted from 32-bit floating-point type to 8-bit integer type, reducing memory usage and inference latency, and adapting to edge computing chips.

[0050] Step S325: Training optimization: After optimizing the model using a hybrid data augmentation strategy (spectral masking, time stretching, and noise superposition), a sound recognition model is obtained, which improves the model's adaptability to changes in the field environment; FocalLoss is used to solve the sample imbalance problem and improve the recognition rate of minority classes (such as abnormal threatening sounds).

[0051] Step S326: Multimodal Fusion Trigger Algorithm: Combining sound signals with environmental sensor data (such as temperature, humidity, wind speed, and air quality) to perform multimodal fusion decision-making, reducing the false trigger rate. Fusion logic: Let the confidence level of the sound source identification result be C_s (0≤C_s≤1), and the environmental interference coefficient be C_e (calculated based on wind speed and rainfall, 0≤C_e≤1, the greater the interference, the closer C_e is to 1). Then, the final trigger confidence level C = C_s × (1 - 0.3 × C_e). When C ≥ 0.85, the camera captures the image; when 0.7 ≤ C < 0.85, the sound acquisition time is extended (an additional 3 seconds of acquisition) and the process returns to step S324 for re-judgment; when C < 0.7, it is determined to be an invalid trigger, and only the log is recorded.

[0052] Step S327: When C ≥ 0.85, the sound is considered a high-confidence sound. When the sound source recognition unit identifies a target sound (human activity sound or abnormal threat sound) with a high confidence C, it immediately sends an interrupt trigger signal to the sound source main control module. The sound source main control module then activates the high-definition camera to capture images and record short videos, and calls a multi-source data fusion algorithm to jointly judge the voiceprint, image, and environmental sensor data of the obtained high-confidence sound. If C ≥ 0.85, the fused data is determined to be a real threat, the sound source main control module generates a high-priority data packet, calls the multi-mode communication unit to start optimal channel transmission, and records the event log. If the system is in low-power mode, the sound source main control module decides whether to activate the camera or only upload the voiceprint recognition result to reduce energy consumption based on the power status.

[0053] 4. Multi-source data fusion processing module The multi-source data fusion algorithm is a weighted DS evidence theory fusion algorithm. The algorithm principle is as follows: Addressing the uncertainty of multi-sensor data, it employs DS evidence theory for fusion reasoning to achieve complementarity and verification of multi-dimensional monitoring information. The multi-source data fusion algorithm includes the following steps: Step S41: Evidence Source Construction: The voiceprint recognition results and image recognition results obtained by the sound source recognition unit, as well as the temperature and humidity sensor data and air quality sensor data obtained by the environmental sensors, are each taken as independent evidence sources. Basic Probability Allocation (BPA) functions are established for each to obtain the evidence value. Example: The BPA values ​​for the sound source recognition evidence source for the "illegal logging" event are m1(A)=0.8, m1(¬A)=0.1, and m1(Ω)=0.1 (Ω is an uncertain set); the BPA values ​​for the image recognition evidence source for the "illegal logging" event are m2(A)=0.75, m2(¬A)=0.15, and m2(Ω)=0.1. The BPA values ​​for the "illegal logging" incident from the temperature and humidity sensor evidence source are m3(A)=0.65, m3(¬A)=0.2, and m3(Ω)=0.15; the BPA values ​​for the "illegal logging" incident from the air quality sensor evidence source are m4(A)=0.7, m4(¬A)=0.18, and m4(Ω)=0.12.

[0054] BPA value generation rules: 1) Temperature and humidity data: Compare the current data with the baseline temperature and humidity of the forest area. If the temperature / humidity fluctuation exceeds the threshold, increase the m3(A) value; if the fluctuation is within the normal range, increase the m3(¬A) value; if the data is abnormal (such as disconnection or exceeding the limit), increase the uncertainty set m3(Ω) value. 2) Air quality data: Compare the current PM2.5 / dust concentration with the forest area baseline. If the concentration increases significantly, increase the m4(A) value; if the concentration is stable at the normal level, increase the m4(¬A) value; if the data is invalid or exceeds the equipment range, increase the uncertainty set m4(Ω) value.

[0055] The methods for calculating and generating BPA values ​​include: 1) BPA generation rules for temperature and humidity sensors Let the baseline temperature of the forest area be T0 and the humidity be H0. The current data are T and H, and the fluctuation thresholds are ΔT and ΔH. If |T-T0|>ΔT or |H-H0|>ΔH, it is judged as "suspected anomaly", and a high value (such as 0.6~0.7) is assigned to m3(A). If the data is within the baseline range: it is judged as "no anomaly", and a high value (such as 0.7~0.8) is assigned to m3(¬A); If data is missing, exceeds the device's range, or jumps occur: it is judged as "uncertain" and a high value (such as 0.3~0.5) is assigned to m3(Ω).

[0056] In the example, `m3(A)=0.65` indicates that the fluctuations in temperature and humidity data are consistent with the characteristics of local environmental changes caused by illegal logging operations; `m3(¬A)=0.2` indicates a confidence level of no anomalies; and `m3(Ω)=0.15` indicates the uncertainty inherent in the data itself.

[0057] 2) Air Quality Sensor BPA Generation Rules Let the baseline PM2.5 / dust concentration in the forest area be C0, the current concentration be C, and the threshold be ΔC: If C > C0 + ΔC: it is judged as "suspected dust from work operations", and a high value (such as 0.65~0.75) is assigned to m4(A); If C is within the normal range: it is judged as "no abnormality", and a high value (such as 0.7~0.8) is assigned to m4(¬A); If the data is invalid, exceeds the range, or the equipment is faulty: it is judged as "uncertain" and a high value (such as 0.3~0.5) is assigned to m4(Ω).

[0058] In the example, `m4(A)=0.7` indicates that the air quality data has increased significantly and is highly consistent with the dust characteristics of chainsaw and vehicle operations; `m4(¬A)=0.18` indicates a confidence level of no anomalies; and `m4(Ω)=0.12` indicates the uncertainty of the data itself.

[0059] Step S42: Evidence Conflict Handling: An improved DS synthesis rule is adopted. When the conflict coefficient K between evidence values ​​is greater than 0.5 (K=Σm1(A_i)×m2(A_j), A_i∩A_j=∅, i is the data obtained in the i-th second; j is the data obtained in the j-th second, i≠j;), a conflict weight allocation factor is introduced to avoid distortion of the synthesis result.

[0060] Step S43: Weighted synthesis: Based on the reliability of each evidence source, set the reliability weight of sound source identification to 0.4, the reliability weight of image recognition to 0.35, and the reliability weight of environmental data to 0.25, perform weighted synthesis, and obtain the final event confidence score. Event confidence score = 0.4 * sound source evidence value + 0.35 * image recognition evidence value + 0.25 * environmental data evidence value.

[0061] Step S44: Decision rule: When the confidence level of the merged event is ≥0.8, it is determined to be a high-confidence event and reported immediately; when 0.6≤confidence level<0.8, it is determined to be a medium-confidence event and secondary verification is initiated (such as increasing the camera recording time); when the confidence level<0.6, it is determined to be a low-confidence event and the data is stored locally only.

[0062] Step S45: Cross-modal false alarm verification mechanism based on "sound-visual-motion" spatiotemporal joint: When the sound source identification unit detects a sudden sound, or the environmental sensor (such as PIR human infrared) detects the movement of a heat source, the timestamp T0 at this moment is recorded, and the system switches from low power consumption to warning state. Audio data within T0±5 seconds is extracted, and the obtained sudden sound is matched with the voiceprint label of "human activity" or "abnormal threat" through a lightweight sound source classification model. If the match is unsuccessful, it is determined to be natural animal activity, automatically downgraded to log recording, and the alarm is terminated. If a match is successful, the voiceprint is classified as high-risk (e.g., suspected illegal logging). The main control module wakes up the high-definition camera in milliseconds, extracts key frames, and runs a lightweight target detection algorithm (YOLO-Fastest) to detect whether there are corresponding entities (e.g., humans, vehicles, tools) in the image. A cross-modal joint decision matrix is ​​constructed. Only when the "heard sound attributes" and "seen target attributes" form a semantic closed loop (e.g., a chainsaw sound is classified and a human or tool is identified in the image) is it finally determined as a "real threat alarm".

[0063] This multi-filtering mechanism reduces the invalid alarm rate by more than 90%, greatly relieving the verification pressure on nature reserve operation and maintenance personnel.

[0064] Specific workflow: 1. First, multiple multi-functional boundary markers are evenly distributed within the natural area. 58-59 markers are evenly set up within the 82,000-83,000 hectare protected area. After system initialization, the dual-power management unit monitors the main power status. If the main power status is normal, the sensing and imaging equipment and sound source recognition unit are activated to collect soundprint, temperature, humidity, and air quality data. The obtained sound source data is transmitted to the sound source recognition unit for sound recognition, noise reduction preprocessing, and feature extraction. This is then combined with the multi-source data fusion processing unit for identification and classification. If the main power status is abnormal, the backup power is switched and a power alarm message is sent. 2. If the voiceprint classification result belongs to human activity or abnormal sound, the main control module is triggered to start the high-definition camera to capture and record a short video. Then, the "voiceprint recognition result + device real-time location + timestamp + image / video multimedia data" are packaged into a high-priority data packet and quickly transmitted to the monitoring center through the multi-mode communication unit. If the voiceprint classification result belongs to animal sound, it is recorded locally and returned to the multi-modal perception and monitoring stage. If the voiceprint classification result belongs to natural interference sound, it is logged locally and returned to the multi-modal perception and monitoring stage. 3. When the multi-mode communication unit transmits data quickly, it selects the channel with the best communication quality through link decision and transmits data according to data priority to improve the reliability of critical data transmission. The selectable channels are at least one of the following: 4G / 5G public network communication module, LoRa self-organizing network module and Beidou short message communication module. According to data priority, multiple channels can be selected to transmit high-priority data synchronously. 4. After the monitoring center obtains the uploaded data, it parses the data and displays alarms. Then it determines whether network switching is needed. If so, in the absence of public network signal, it uses a self-organizing network to switch to the gateway marker for data transmission via a nearby boundary marker. If network switching is not needed, it transmits data through the public network, thus achieving comprehensive monitoring of large-area nature reserves.

[0065] Example 2 The difference from Example 1 is that the pile body is made of fiberglass or stainless steel, and its appearance blends in with the surrounding natural environment, avoiding interference with the ecological environment and wildlife; the pile body integrates multiple types of sensing devices such as temperature and humidity sensors and high-precision tilt sensors to build a multi-dimensional, three-dimensional, multi-modal sensing fusion system to comprehensively capture environmental and ecological data.

[0066] Each unit's main control module is equipped with a high-performance edge computing chip, which has local data processing capabilities. It can complete computational tasks such as sound source recognition and simple image analysis on the device side, and only upload the recognition results, key feature data and alarm-related multimedia data to the monitoring center, which greatly reduces network transmission traffic consumption and improves data transmission efficiency.

[0067] Multiple electronic boundary markers are interconnected through self-organizing network communication units to build a distributed monitoring network. When a single boundary marker is in an area without public network signal and cannot be directly connected to the public network, it can perform "multi-hop forwarding" through the self-organizing network link of adjacent boundary markers to transmit data to a gateway marker with public network communication capabilities. The gateway marker then aggregates the data and uploads it to the monitoring center, achieving full coverage of the monitoring network with no data transmission blind spots.

[0068] The system supports remote upgrades and customized configurations. It can remotely update the sound source recognition model, optimize the link decision algorithm, and adjust the data acquisition frequency and alarm threshold through the monitoring center to adapt to the ecological characteristics, management needs and geographical environment of different nature reserves.

[0069] The self-organizing network communication unit is used to implement the distributed self-organizing network routing optimization algorithm. Algorithm Name: Hybrid Routing Algorithm Based on Energy Awareness and Link Quality (EALQ-Routing) Algorithm principle: By combining the remaining energy of nodes with the quality of links, the optimal routing path is dynamically selected to balance network load and extend network lifespan.

[0070] The distributed ad hoc network routing optimization algorithm includes the following steps: Step S51: Routing metrics: comprehensively consider three core metrics: node remaining energy (SOE ≥ 20% is considered a valid node), link communication quality (overall score ≥ 0.6), and hop count (≤ 5 hops); Step S52: Route Discovery: The improved AODV routing protocol is adopted. When a node broadcasts a route request (RREQ), it carries its own energy status and link quality information. Step S53: Route selection: The intermediate node calculates the total path cost C=α×E_cost+β×L_cost+γ×H_cost (where E_cost is the energy cost, L_cost is the link quality cost, H_cost is the hop cost, α=0.4, β=0.4, γ=0.2) based on the received RREQ information, and selects the path with the minimum cost as the optimal route; Step S54: Route maintenance: Regularly monitor the energy status and link quality of each node on the route path. When the energy of a node is lower than 15% or the link quality score is lower than 0.5, trigger route rediscovery to avoid route failure.

[0071] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An integrated monitoring boundary marker system based on multimodal sensing and redundant transmission, characterized in that, include: Multiple piles, each of which is equipped with sensing and imaging devices, a dual power management unit, a sound source identification unit, a multi-source data fusion processing unit, and a multi-mode communication unit, which are electrically connected. The multi-mode communication unit is a communication master control module that integrates at least two heterogeneous network communication modules. It is used to transmit data acquired from each channel in real time and monitor the core quality parameters of each communication channel. It adopts a channel selection mechanism based on weighted multi-attribute decision-making, combined with a dynamic threshold adjustment strategy. It selects at least one of public network communication module, self-organizing network communication module or satellite communication module according to the application scenario requirements to achieve intelligent channel selection and switching. The core quality parameters of each communication channel include: signal strength, bit error rate, and transmission delay; Dual power management unit is used to power the multi-mode communication unit, sound source identification unit, multi-source data fusion processing unit, and sensing and imaging equipment; Sensing and imaging equipment includes: an integrated SHT30 temperature and humidity sensor, an air quality sensor, and a 1080P infrared night vision camera; The sound source recognition unit is used to collect ambient sounds on site, perform noise reduction preprocessing, extract spectral features and intelligent classification, and identify four categories of preset scene results. When the recognition result is human activity sound or abnormal threat sound, and the confidence level of the obtained result is greater than 85%, a trigger signal is immediately sent to the sound source main control module to start the high-definition camera to capture images and record short videos. Simultaneously, the voiceprint recognition result + real-time device location + timestamp + image / video multimedia data are integrated, packaged into a high-priority data packet, and quickly transmitted to the monitoring center through the multi-mode communication unit. The multi-source data fusion processing unit is used to analyze the voiceprint and image data packets uploaded by the sound source identification unit through multi-source data fusion algorithms to achieve integrated fusion analysis of environmental perception data, sound source identification data, and image data. In response to the uncertainty of multi-sensor data, the DS evidence theory is used for fusion reasoning to achieve complementarity and verification of multi-dimensional monitoring information.

2. The integrated monitoring boundary marker system based on multimodal sensing and redundant transmission according to claim 1, characterized in that, The sound source identification unit is used to implement the following steps: Step S31: Sound Acquisition and Preprocessing: A combination of continuous acquisition and triggered acquisition is used to capture ambient sound from all directions. Continuous acquisition uses a high-sensitivity microphone array to collect ambient sound, while triggered acquisition uses low power consumption to collect sound at a level ≥60dB. After the acquisition is completed, the audio is preprocessed by adaptive noise reduction and sound enhancement algorithms. Then, an improved spectral subtraction combined with wavelet threshold noise reduction algorithm is used to filter environmental interference and improve the sound signal quality to obtain a noise-reduced preprocessed sound source. Dynamic environmental acoustic baseline adaptive filtering mechanism: In response to the regional and seasonal regular noise in nature reserves, the system designs a background sound adaptive learning method based on Gaussian mixture model to completely eliminate false triggers caused by regular natural interference. Step S32: Feature extraction and classification: Based on a pre-trained lightweight neural network model, spectral features are extracted from the noise-reduced preprocessed sound source signal to obtain an audio spectrogram. The obtained audio spectrogram is then intelligently classified into four preset scenarios: human activity sounds, specific animal sounds, natural interference sounds, and abnormal threat sounds. Human activity sounds include talking, footsteps, and vehicle engine sounds. Natural noise disturbances include: wind, rain, and thunder; unusual threatening noises include: chainsaws, gunshots, and illegal fishing operations. Step S33: Event Triggering and Data Packaging: When the recognition result is human activity sound or abnormal threat sound, a trigger signal is immediately sent to the sound source main control module to start the high-definition camera to capture images and record short videos; the sound source main control module synchronously integrates the voiceprint recognition result + real-time device location + timestamp + image / video multimedia data, packages it into a high-priority data packet, and quickly transmits it to the monitoring center through the multi-mode communication unit.

3. The integrated monitoring boundary marker system based on multimodal sensing and redundant transmission according to claim 2, characterized in that, Step S31 includes the following steps: Step S311: Noise estimation: Identify silent segments using a speech activity detection algorithm and estimate the noise power spectral density based on a statistical model; Step S312: Spectral subtraction processing: Subtract the power spectrum of the noisy signal to suppress stationary noise; Step S313: Wavelet Threshold Optimization: Perform wavelet decomposition on the spectral subtraction signal and use an adaptive threshold function. , where σ is the noise standard deviation, N is the signal length, high-frequency coefficients are processed, and effective signal components are retained; Step S314: Sound Enhancement: A loudness compensation algorithm based on the characteristics of human hearing is adopted to improve the recognition of weak target sounds, improve the accuracy and reliability of sound monitoring results, and avoid interference from background noise. Step S315: Environmental baseline self-learning: During the first 72 hours of the initial deployment of the boundary stake, the system automatically and silently collects long-sequence audio data from the surrounding area, extracts Mel frequency cepstral coefficients, and uses GMM to perform probability density modeling of the normal background noise specific to the stake location, generating a unique static environmental acoustic baseline for the boundary stake. Step S316: Dynamic Adaptive Filtering: In normal monitoring, the acoustic feature distance between the input audio features and the static environmental soundprint baseline is calculated in real time, specifically by Mahalanobis distance. When the obtained acoustic feature distance is less than a set threshold, it is judged as normal natural white noise and directly filtered out without triggering the subsequent high-power classification model. Only when the acoustic feature distance breaks the baseline balance, such as the sudden appearance of chainsaw sound or human footsteps, is it judged as a sudden sound and accurately classified.

4. The integrated monitoring boundary marker system based on multimodal sensing and redundant transmission according to claim 3, characterized in that, Step S32 includes the following steps: Based on MobileNet, an attention mechanism (SE) module and a knowledge distillation method are introduced to construct a lightweight sound source classification optimization model with high accuracy and low computational consumption. Step S321: Model structure optimization: Feature extraction layer: Use depthwise separable convolution to reduce computation, insert SE module to enhance the weight allocation of key spectral features; Step S322: Knowledge distillation: Using the complex ResNet-50 model as the teacher model and MobileNet as the student model, the distillation loss function with temperature coefficient adjustment is used to make the student model maintain a lightweight design while achieving classification accuracy close to that of the teacher model. Step S323: Model Quantization: Using INT8 quantization technology, the model parameters are converted from 32-bit floating-point type to 8-bit integer type, reducing memory usage and inference latency, and adapting to edge computing chips; Step S324: Training and Optimization: After optimizing the model using a hybrid data augmentation strategy, a sound recognition model is obtained. The hybrid data augmentation strategy includes: spectral masking, time stretching, and noise superposition to improve the model's adaptability to changes in the field environment. FocalLoss is used to solve the sample imbalance problem and improve the recognition rate of minority classes, such as abnormal threatening sounds. Step S325: Multimodal fusion triggering algorithm: Combine sound signals and environmental sensor data to perform multimodal fusion decision-making to reduce the false trigger rate. Environmental data includes: temperature, humidity, wind speed, and air quality. The fusion logic of sound signals and environmental data is as follows: Let the confidence level of the sound source identification result be C_s, 0≤C_s≤1, and the environmental interference coefficient be C_e, which is calculated based on wind speed and rainfall, 0≤C_e≤1. The greater the interference, the closer C_e is to 1. Then the final trigger confidence level C=C_s×(1-0.3×C_e). When C≥0.85, the camera captures an image; when 0.7≤C<0.85, the sound acquisition time is extended by 3 seconds, and the process returns to step S324 for re-evaluation; when C<0.7, the trigger is deemed invalid, and the sound is only cached locally and logged; when C≥0.85, the sound is considered a high-confidence sound, and the probability of the target sound corresponding to human activity or abnormal threat sounds is relatively high. Step S326: When the sound source recognition unit identifies a target sound with high confidence C, it immediately sends an interrupt trigger signal to the sound source main control module. The sound source main control module then starts the high-definition camera to capture images and record short videos, and calls the multi-source data fusion algorithm to jointly judge the voiceprint, image, and environmental sensor. If C ≥ 0.85, the fused data is determined to be a real threat. The sound source main control module generates a high-priority data packet, calls the multi-mode communication unit to start the optimal channel transmission, and records the event log at the same time. If the system is in low-power mode, the sound source main control module decides whether to start the camera or only upload the voiceprint recognition result to reduce energy consumption based on the power status.

5. The integrated monitoring boundary marker system based on multimodal sensing and redundant transmission according to claim 1, characterized in that, Multi-source data fusion algorithms are used to implement the following steps: Step S41: Evidence source construction: The voiceprint recognition result and image recognition result obtained by the sound source recognition unit, as well as the temperature and humidity sensor data and air quality sensor data obtained by the environmental sensor, are used as independent evidence sources. The basic probability allocation (BPA) function is established for each of them to obtain the evidence value. Step S42: Evidence Conflict Handling: An improved DS synthesis rule is adopted. When the conflict coefficient K between evidence values ​​is greater than 0.5, K = Σm1(A_i) × m2(A_j), A_i ∩ A_j = ∅, where i is the data obtained in the i-th second and j is the data obtained in the j-th second, i ≠ j; a conflict weight allocation factor is introduced to avoid distortion of the synthesis result. Step S43: Weighted synthesis: Based on the reliability of each evidence source, set the reliability weight of sound source identification to 0.4, the reliability weight of image recognition to 0.35, and the reliability weight of environmental data to 0.25, perform weighted synthesis, and obtain the final event confidence score. Event confidence score = 0.4 * sound source evidence value + 0.35 * image recognition evidence value + 0.25 * environmental data evidence value; Step S44: Set decision rules: When the confidence level of the merged event is ≥0.8, it is judged as a high-confidence event and reported immediately; When 0.6 ≤ confidence level < 0.8, it is judged as a medium confidence event, and secondary verification is initiated to increase the camera recording time; When the confidence level is less than 0.6, it is judged as a low-confidence event, and the data is stored locally only. Step S45: Cross-modal false alarm prevention verification based on sound-visual-motion spatiotemporal joint: When the sound source identification unit detects a sudden sound, or the environmental sensor detects the movement of a heat source, the timestamp T0 is recorded at this moment, and the low power consumption state is switched to the warning state. Audio data within T0±5 seconds is extracted, and the obtained sudden sound is matched with the voiceprint tags of human activities or abnormal threats through a lightweight sound source classification model. If the match is unsuccessful, it is determined to be natural animal activity, automatically downgraded to log recording, and the alarm is terminated. If a match is successful, the voiceprint is classified as high-risk. The main control module wakes up the high-definition camera in milliseconds, extracts key frames, and runs a lightweight target detection algorithm to detect whether a corresponding entity exists in the image.

6. The integrated monitoring boundary marker system based on multimodal sensing and redundant transmission according to claim 1, characterized in that, The multimode communication unit is used to implement the following steps: Step S11: Channel parameter normalization processing: The collected signal strength, bit error rate and transmission delay are normalized and mapped to the [0,1] interval to eliminate the difference in units. The signal strength normalization adopts forward mapping, while the bit error rate and transmission delay adopt reverse mapping. Step S12: Normalize the signal, bit error rate, and transmission delay respectively. Signal strength normalization: Let the measured signal strength at time i be S, the minimum value obtained in the continuous measurement period be S_min, and the maximum value obtained in the continuous measurement period be S_max. The normalized value S_norm = (S - S_min) / (S_max - S_min). Bit error rate normalization: Let the measured bit error rate at time i be E, and the maximum value obtained during the continuous measurement period be E_max. The normalized value E_norm = 1 - E / E_max Transmission delay normalization: Let the measured transmission delay at time i be D, and the maximum value obtained during the continuous measurement period be D_max. The normalized value is D_norm = 1 - D / D_max. Step S13: Dynamic weight allocation: The basic weights of each parameter are determined using the analytic hierarchy process (AHP), and then dynamically adjusted in combination with data priority. The weights for high-priority data are: signal strength weight is 0.4, transmission delay weight is 0.35, and bit error rate weight is 0.

25. Low-priority data is assigned weights as follows: signal strength weight is 0.3, bit error rate weight is 0.3, and transmission delay weight is 0.

4. Step S14: Channel comprehensive score: Calculate the comprehensive score C of each channel after weighting each parameter, C=ω_S×S_norm+ω_E×E_norm+ω_D×D_norm, where ω_S is the weight of signal strength; ω_E is the weight of bit error rate; ω_D is the weight of transmission delay, and select the channel with the highest comprehensive score as the optimal transmission channel. Step S15: Determine the priority of the output data based on the sound source identification unit. When transmitting high-priority data, monitor the status of all channels in real time. If the comprehensive score of at least two channels is ≥0.7 and the transmission delay of any single channel is ≤300ms and the bit error rate is ≤0.5%, then start multi-channel concurrent transmission to avoid packet loss. Use a data fragmentation redundancy coding mechanism to fragment the data packets and transmit them through different channels. The receiving end merges and decodes the fragments to ensure data integrity. If the overall score of a single channel is ≥0.7, and the transmission delay of any single channel is ≤300ms and the bit error rate is ≤0.5%, then the optimal channel is selected for transmission. When transmitting low-priority data, monitor the status of the public network channel. If the public network channel is available, use the public network channel for transmission. If the public network channel is unavailable, switch to the self-organizing network channel for transmission. Step S16: Channel switching: Monitor the comprehensive score of the current transmission channel in real time. If the score is lower than 0.5 and lasts for 5 seconds, or if a transmission interruption occurs, immediately trigger the channel reselection process, switch to the suboptimal channel, and record the channel switching log. A transmission interruption is defined as the loss of 3 consecutive data packets.

7. The integrated monitoring boundary marker system based on multimodal sensing and redundant transmission according to claim 1, characterized in that, The dual power management unit includes: a power control module, a main power supply, a backup power supply, and an intelligent power switching and power management circuit. The main power control module and the intelligent power switching and power management circuit are electrically connected to the main power supply and the backup power supply, respectively; the main power supply and the backup power supply are electrically connected to the main control module, the multi-mode communication unit, the sound source recognition unit, the sensing and imaging equipment, and the multi-source data fusion processing module, respectively. The main power supply includes: solar photovoltaic panels installed on the boundary piles and electrically connected to the energy storage battery pack to collect light energy in order to adapt to the outdoor lighting environment. The energy storage battery pack is connected to the power supply of each unit to provide continuous power for the normal operation of the system. Intelligent power switching and power management circuit: Used to monitor the voltage, remaining power, and charging / discharging status of the energy storage battery pack in real time. When the energy storage battery's power level falls below a first preset threshold, it automatically switches to backup power supply. Simultaneously, it sends a power status alarm message to the monitoring center via a multi-mode communication unit. The alarm message includes the reason for the switchover and the remaining power level of the backup power supply, facilitating timely handling by maintenance personnel. The first preset threshold is voltage < 12V or power level < 30%. When the backup power supply level is less than the second preset threshold, a low power emergency alarm message is sent to the monitoring center and then the system enters low power mode, retaining only the core function of sound source recognition. If the backup power supply level is between the first and second preset thresholds, the backup power supply is maintained and the main power supply status is continuously monitored. When the main power supply voltage recovers, the power supply is switched back to the main power supply and monitoring continues. The main power supply voltage recovers to a level greater than 13V and the power supply level is greater than 40%. The second preset threshold is when the backup power supply level is less than 10%.

8. The integrated monitoring boundary marker system based on multimodal sensing and redundant transmission according to claim 7, characterized in that, The power supply main control module integrates remaining power estimation, power supply prediction, and dynamic threshold adjustment to achieve intelligent power management and fault early warning. Specifically, it includes the following steps: Step S21: Accurate estimation of remaining power: The Kalman filter algorithm is used to optimize the SOE estimation accuracy, and the SOE is dynamically corrected by combining the charge and discharge characteristic curves of the battery and the temperature compensation coefficient. The basic model used for SOE estimation is: SOE(t)=SOE(0)+(1 / C_n)×∫(I(t)×η(t,T))dt, where C_n is the rated capacity of the battery, I(t) is the charge and discharge current, and η(t,T) is the charge and discharge efficiency at temperature T. Then, the SOE estimation result is optimized by Kalman filtering: the initial value of SOE is predicted by the state equation, the observation equation is constructed using the measured values ​​of voltage and current, and the estimation error of SOE is dynamically corrected to make the SOE estimation accuracy ≤±3%. Step S22: Main power supply capacity prediction: Based on historical sunshine data, weather forecast information and photovoltaic power generation model, predict the power generation required by the main power supply and the available power supply duration for the next 24 hours as the power supply prediction result P(t). Power generation model: P(t) = P_STC × G(t) / G_STC × [1 + α(T(t) - T_STC)], where P_STC is the photovoltaic power under standard test conditions, G(t) is the real-time irradiance, G_STC is the standard irradiance, α is the power temperature coefficient, T(t) is the real-time battery temperature, and T_STC is the standard temperature; Step S23: Dynamic threshold adjustment: Adjust the power switching threshold according to the power supply forecast results; if it is predicted that there will be sufficient sunshine in the next 24 hours and the power generation of the solar photovoltaic panel is ≥ 1.5 times the average daily power consumption of the system obtained from historical data, then the first threshold is lowered to 25%; if it is predicted that there will be cloudy or rainy weather in the next 24 hours and the power generation of the solar photovoltaic panel is ≤ 0.5 times the average daily power consumption of the system obtained from historical data, then the first threshold is raised to 35%, and the backup power supply is pre-activated in advance to ensure seamless switching. Step S24: Low power warning and energy saving control: When the backup power is lower than the second preset threshold by 10%, the graded energy saving mode is activated; Level 1 energy saving: non-core sensing devices are turned off and the camera capture frequency is reduced; Level 2 energy saving: only the sound source recognition core function and emergency communication module are retained, and other devices are put into hibernation until the main power supply is restored. Level 1 energy saving consumes 10-5% of electricity; Level 2 energy saving consumes less than 5% of electricity.

9. The integrated monitoring boundary marker system based on multimodal sensing and redundant transmission according to claim 7, characterized in that, 58 to 59 piles are evenly distributed within the 82,000 to 83,000 hectares protected area.

10. The integrated monitoring boundary marker system based on multimodal sensing and redundant transmission according to claim 7, characterized in that, include: Self-organizing network communication unit, which is set in the pile body, enables interconnection between piles and builds a distributed monitoring network; When a single boundary marker is in an area without public network signal and cannot be directly connected to the public network, it transmits data to a gateway marker with public network communication capability through a self-organizing network link of adjacent boundary markers via multi-hop forwarding. The gateway marker then aggregates the data and uploads it to the monitoring center, achieving full coverage of the monitoring network with no data transmission blind spots. The ad hoc network communication unit is used to implement the distributed ad hoc network routing optimization algorithm, which includes the following steps: Step S51: Select nodes with remaining energy, a link communication quality score ≥ 0.6, and hop count. Use the improved AODV routing protocol. When a node broadcasts a route request, it carries its own energy status and link quality information. Intermediate nodes calculate the path cost C = α × E_cost + β × L_cost + γ × H_cost based on the received RREQ information, where E_cost is the energy cost, L_cost is the link quality cost, H_cost is the hop count cost, α = 0.4, β = 0.4, and γ = 0.

2. Select the path with the lowest cost as the optimal route. Step S52: Periodically monitor the energy status and link quality of each node on the routing path. When the energy of a node is lower than 15% or the link quality score is lower than 0.5, trigger route rediscovery to avoid route failure.

Images