Power transmission line bird damage risk assessment and early warning method and system thereof

Abstract

The invention provides a power transmission line bird damage risk assessment and early warning method and system, and relates to the technical field of intelligent power grid environment perception and risk early warning, and the method comprises the steps: data collection and multi-modal fusion: collecting image data, voiceprint data, meteorological parameters and ecological environment information along a power transmission line to form multi-modal data, the multi-modal data is uniformly coded by synchronizing timestamps and geographic coordinates, an original bird activity sample set is generated, and for the problems that the monitoring means is single and the recognition precision is insufficient, image, voiceprint, weather and ecological information are fused through a multi-modal data acquisition module, so that the recognition precision is improved. The acousto-optic image fusion algorithm and the two-channel convolutional neural network are adopted, all-weather bird type and behavior accurate recognition is achieved, the recognition accuracy is improved, key behaviors such as nesting and dung pulling can be stably captured even at night or in a foggy environment, and potential risk missing detection is remarkably reduced.

Description

Technical Field

[0001] This invention relates to the field of smart grid environmental sensing and risk early warning technology, specifically to a method and system for assessing and warning of bird damage risks to transmission lines. Background Technology

[0002] With the rapid development of power systems towards ultra-high voltage, intelligentization, and greening, the safe and stable operation of transmission lines faces increasingly severe threats from birds. Especially in densely populated areas, ecological reserves, and migratory bird corridors, bird nesting, defecation, resting, and flight crossings have become major non-meteorological factors causing line flashovers, contamination, and short circuits. For example, magpies nesting on insulator strings can cause partial discharges, crows defecating can create conductive channels, and large birds crossing conductor gaps can cause phase-to-phase short circuits. Currently, manual inspections combined with infrared cameras, sound and light bird deterrents, and insulating coatings are widely used for bird damage control on transmission lines.

[0003] However, existing bird control technologies face the following key technical challenges in achieving accurate early warning and intelligent control: Monitoring methods are limited and lack sufficient accuracy. Traditional methods rely mainly on manual inspections or fixed cameras, making it difficult to capture bird activity at night, in foggy conditions, or at long distances. For example, infrared cameras have a resolution reduced to 480P in low-light environments, resulting in less than 70% accuracy in bird species identification. Behavioral patterns (such as early nesting stages) are difficult to determine in real time, leading to the omission of numerous potential risks. Risk assessment lacks spatiotemporal dynamic modeling. Existing systems are mostly based on static empirical thresholds or single-point monitoring data, failing to reflect the spatial spread and temporal periodicity of bird activity. For example, they do not consider the weight of migration peaks or the risk diffusion from adjacent towers, causing risk assessment deviations exceeding 30%, resulting in delayed early warnings or frequent false alarms. Control strategies are passive and lack adaptive optimization. Bird deterrent devices mostly use timed or manual triggering modes, lacking intelligent linkage and closed-loop learning mechanisms with risk levels. For example, the fixed operation of sound and light bird deterrents leads to rapid adaptation and failure by birds. Manual intervention relies on experience-based judgment, making it difficult to dynamically adjust model parameters based on actual bird damage feedback, resulting in decreased control efficiency after long-term system operation.

[0004] Therefore, a method and system for risk assessment and early warning of bird damage to power transmission lines are needed to solve the above problems. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method and system for assessing and warning of bird hazards to power transmission lines, thus solving the problems of existing technologies.

[0006] To achieve the above objectives, the present invention provides a method for risk assessment and early warning of bird hazards to power transmission lines, comprising the following steps: Sp1. Data Acquisition and Multimodal Fusion: Image data, voiceprint data, meteorological parameters and ecological environment information along the transmission line are collected to form multimodal data. The multimodal data is uniformly encoded by synchronizing timestamps and geographic coordinates, and an original bird activity sample set is generated. Sp2. Bird identification and behavioral feature extraction: The original bird activity sample set is processed using an audio-visual image fusion algorithm to identify bird species and their behavioral patterns. The behavioral patterns include resting, nest building, flight crossing, group gathering and defecation behavior, and output behavioral feature parameters. Sp3, Spatiotemporal weighted gridded risk modeling: The transmission line area is divided into multi-level spatial grid units. A spatiotemporal coupled risk field model is established based on the behavioral characteristic parameters, activity frequency and time distribution, and the basic bird damage risk value of each grid is calculated. Sp4, Meteorological-Ecological-Behavioral Three-Dimensional Correction: Real-time meteorological elements and ecological factors are introduced into the spatiotemporal coupled risk field model, and the risk field is dynamically corrected through multiple correction algorithms to form a comprehensive risk distribution map; Sp5, Risk Classification and Early Warning Output: Based on the comprehensive risk value, set the early warning level threshold, output graded risk early warning information, and link with the transmission line operation and maintenance system to execute corresponding protection strategies. Sp6, Self-learning closed-loop optimization: Based on the feedback results of actual bird damage events, the model parameters are corrected in reverse, and the spatiotemporal weights and threat coefficients are continuously updated through learning to achieve adaptive optimization of the model.

[0007] Preferably, the spatiotemporal weighted rasterized risk modeling in Sp3 includes the following sub-steps: Sp31. The transmission line area is divided into grid units of fixed size according to geographical coordinates, and the size of the grid unit is dynamically adjusted according to the line voltage level and terrain complexity. Sp32. Using bird activity frequency as the main variable and activity time period as the time weight, a spatiotemporal distribution matrix is ​​constructed. The time weight is calculated by using a time-segmented statistical method to weight day and night, season and migration cycle. Sp33 uses a Gaussian spatial weighting function to calculate the risk propagation coefficient between adjacent grids, and achieves continuous spatial propagation of risk through exponential decay, generating a smooth risk field distribution.

[0008] Preferably, the bird identification and behavioral features in Sp2 are extracted using a dual-channel convolutional neural network of voiceprint and image, wherein the voiceprint recognition channel is used to identify bird species and the image recognition channel is used to extract behavioral and posture features. These features are then fused through a multimodal feature matching module to improve the identification accuracy in nighttime and foggy environments.

[0009] Preferably, the three-dimensional correction model of meteorology-ecology-behavior in Sp4 is constructed based on weighted nonlinear regression, specifically including: The meteorological factor correction weights are jointly determined by wind speed and humidity. The wind speed weight increases non-linearly with wind force level, while the humidity weight enhances the bird activity induction effect in high humidity environments. The ecological factor correction weights are determined by the distance to the water area and the vegetation type. The distance to the water area is decayed using an inverse proportional function, and the vegetation type is assigned a graded weight based on food source abundance and habitat suitability. The behavioral factor adjustment weight is determined by the duration of the behavior and the group size, where the duration is mapped using a logarithmic function and the group size is amplified by a power function to amplify the risk of high-density clustering. By normalizing the weights of each factor, a comprehensive risk correction coefficient is formed, enabling dynamic coupling correction of multiple factors including meteorology, ecology, and behavior.

[0010] Preferably, the self-learning closed-loop optimization includes the following sub-steps: Sp61. Real-time comparison of the risk level predicted by the model with the risk level of actual bird damage events, and establishment of a deviation assessment index system; Sp62. Calculate the risk prediction error based on the deviation assessment results and generate a multi-dimensional parameter correction matrix, covering spatiotemporal weights, threat coefficients and correction factors; Sp63 introduces a transfer learning algorithm to transfer model parameters from validated regions to new regions. Through few-sample fine-tuning, feature weights are adaptively adjusted, thereby improving the cross-regional adaptability and long-term self-evolution capability of the risk model.

[0011] Preferably, the prediction method is based on the risk trend curve to achieve predictive early warning. When short-term meteorological or ecological changes cause the risk gradient to rise, a dynamic early warning is issued in advance and the pre-operation strategy of protective equipment is triggered.

[0012] Preferably, a bird behavior-equipment impact matrix is ​​established, which is used to quantify the impact coefficients of various bird behaviors on line components, and dynamically adjusts the risk calculation weights based on the threat coefficient to achieve behavior-driven risk assessment.

[0013] Preferably, a bird risk knowledge graph is constructed, which includes bird species nodes, ecological distribution nodes, behavioral characteristic nodes, and risk level nodes. Potential high-risk areas are identified through semantic reasoning algorithms, and interpretability support is provided for the risk assessment model.

[0014] Preferably, the system includes: The multimodal data acquisition module is used to collect image, acoustic, meteorological, and ecological data along the power transmission line. The bird identification and behavior analysis module is used to identify bird species and behavioral characteristics and output behavioral parameters; The spatiotemporal risk modeling module is used to construct a spatiotemporal weighted rasterized risk model and calculate the basic risk value. The 3D correction and risk fusion module is used to integrate meteorological, ecological and behavioral factors to dynamically correct the risk model. The early warning decision and strategy linkage module is used to output hierarchical early warning information and link with the power transmission operation and maintenance system to execute protection strategies. The model self-learning and knowledge graph module is used to automatically correct model parameters and update the bird risk knowledge base based on event feedback. A low-power distributed monitoring terminal network is used to achieve multi-point real-time monitoring and remote data reporting.

[0015] The present invention has the following beneficial effects: To address the issues of limited monitoring methods and insufficient identification accuracy, this invention integrates image, voiceprint, meteorological, and ecological information through a multimodal data acquisition module. It employs an audio-visual image fusion algorithm and a dual-channel convolutional neural network to achieve accurate identification of bird species and behaviors in all weather conditions, improving the identification accuracy. Even at night or in foggy conditions, it can reliably capture key behaviors such as nest building and defecation, significantly reducing potential risks of missed detection.

[0016] To address the lack of spatiotemporal dynamic modeling in risk assessment, this invention constructs a spatiotemporal weighted rasterized risk model. Combining Gaussian spatial diffusion and a three-dimensional correction mechanism, it dynamically integrates activity frequency, time weight, and meteorological and ecological factors to achieve a precise mapping between the spatial continuity and temporal periodicity of the risk field. This reduces risk assessment bias and effectively eliminates early warning lag and false alarms.

[0017] To address the problem of passive prevention and control strategies lacking adaptive optimization, this invention enables protective equipment to be intelligently activated on demand through a risk-level early warning and strategy linkage module. Combined with a self-learning closed-loop optimization mechanism, it automatically iterates model parameters and threat coefficients based on feedback from actual bird-related incidents, thereby improving the adaptability of bird-repelling equipment and significantly reducing the frequency of manual intervention and the rate of bird-related flashover accidents. Attached Figure Description

[0018] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a system framework diagram of the present invention; Figure 3 This is a comprehensive risk heat map for the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Specific Implementation Example 1

[0020] like Figures 1 to 3 As shown, this invention discloses a method and system for assessing and warning of bird damage risks to transmission lines. Through multimodal data perception, spatiotemporal coupled modeling, three-dimensional correction of meteorology, ecology and behavior, closed-loop self-learning and predictive early warning, it can achieve accurate assessment, hierarchical early warning and intelligent prevention and control of bird damage risks to transmission lines. The system is based on low-power distributed monitoring network deployment and integrates edge computing and cloud analysis to achieve closed-loop intelligent management of the entire chain.

[0021] The following is a detailed explanation: 1. Detailed execution process of the method: Sp1. Data Acquisition and Multimodal Fusion: The system deploys multifunctional monitoring nodes every 50–200 meters along the transmission line corridor to simultaneously collect image, acoustic, meteorological, and ecological environment data. Image data is captured by high-definition infrared / visible dual-mode cameras (resolution ≥1080P, frame rate ≥25fps), supporting night vision and wide-angle dynamic capture; acoustic data is collected by a directional microphone array (sensitivity ≥-38dB, frequency response 20Hz–20kHz) to capture bird calls, wing flapping, and group acoustic signals; meteorological parameters are monitored in real time through an integrated micro weather station (wind speed accuracy ±0.3m / s, humidity ±3%RH, temperature ±0.5℃, rainfall ±0.1mm); ecological environment information is extracted by combining high-resolution satellite imagery (≤1m) and field surveys to obtain vegetation type (forest / farmland / wetland), water distance (≤500m marked as high risk), and food source density (fruit tree / crop coverage).

[0022] All data are appended with a unified timestamp (GPS / BeiDou time synchronization, accuracy ≤10ms) and geographic coordinates (WGS84, accuracy ≤1m) to construct an original bird activity sample set. Each sample includes an image frame sequence, a voiceprint waveform segment, a meteorological quadruple (wind speed, humidity, temperature, rainfall), an ecological triple (vegetation type, distance to water, food source density), a timestamp, and geographic coordinates, stored in a structured format for easy subsequent multimodal processing.

[0023] Sp2, Bird Recognition and Behavioral Feature Extraction: A dual-channel convolutional neural network (VNN) for voiceprint and image processing is used to process the original sample set. The voiceprint recognition channel takes the voiceprint spectral features as input and uses a deep residual network combined with a temporal attention mechanism to output bird species recognition results (supporting ≥200 common power line birds, such as magpies, crows, and egrets). The image recognition channel takes a continuous multi-frame image sequence as input and uses a dual-path temporal network (slow path extracts static pose, fast path captures dynamic trajectory) to output the location of behavioral key points and motion trajectory.

[0024] A cross-attention fusion mechanism aligns voiceprint and image features, dynamically increasing the voiceprint channel weight to 0.7 and decreasing the image channel weight to 0.3 in nighttime or foggy conditions to ensure an all-weather recognition accuracy of ≥92%. Behavioral pattern classifications include: resting (continuous stay ≥3 seconds), nest building (carrying nesting material and repeatedly traveling back and forth), flight crossing (horizontal crossing of gaps in wires), group gathering (≥5 individuals in the same frame with a distance ≤2 meters), and defecation behavior (posture matching combined with trajectory detection). Output behavioral feature parameters include: species identifier, behavior type, duration, group size, activity trajectory coordinate sequence, and threat index (0–1).

[0025] Sp3, Spatiotemporal Weighted Rasterized Risk Modeling: SP31. Spatial Grid Division: The transmission line corridor (100 meters wide) is divided into fixed-size grid units according to geographical coordinates. The basic size is 50 meters × 50 meters (applicable to 110kV lines), reduced to 30 meters × 30 meters for 220kV lines, and densified to 20 meters × 20 meters for mountainous terrain. Each grid is assigned a unique identifier, which is stored in the geographical grid index.

[0026] Sp32. Constructing a spatiotemporal distribution matrix: Using bird activity frequency as the primary variable and activity time period as the time weight. Time weights are stratified and weighted: daytime has a standard weight of 1.0, decreasing to 0.6 at night; increasing to 1.5 during spring migration and decreasing to 0.8 in winter; peak migration days are further weighted to 1.8 based on historical data. A three-dimensional spatiotemporal distribution matrix is ​​constructed, integrating activity frequency, time weight, and behavioral threat level.

[0027] Sp33, Gaussian Spatial Risk Diffusion: The risk propagation coefficient between adjacent grids is calculated using Gaussian smoothing, ensuring continuous spatial diffusion of risk. Risks in adjacent grids influence each other, with closer grids having a greater impact, generating a smoothed basic risk field distribution map (heatmap format, risk value range 0–100).

[0028] Sp4, Meteorological-Ecological-Behavioral Three-Dimensional Correction: A three-dimensional correction model is constructed based on weighted nonlinear regression to dynamically generate comprehensive risk correction coefficients (range 0.5 to 2.0).

[0029] Meteorological factor correction: The weight of wind speed increases nonlinearly with wind force level (significantly increases with >5m / s), and the weight of humidity increases in high humidity environment (>70%RH) to reflect the increased risk of defecation.

[0030] Ecological factor correction: the closer the water area, the higher the weight (inversely proportional to the decrease), and the vegetation type is weighted according to the food source and habitat suitability (wetland is the highest, forest is the second highest, and farmland is the lowest).

[0031] Behavioral factor correction: the longer the duration, the higher the weight (logarithmic growth), and the larger the group size, the greater the weight (highlighting the risk of clustering).

[0032] The weights of the three types of factors are normalized and fused to form a comprehensive correction coefficient, which is applied to the basic risk field to generate a comprehensive risk distribution map that is updated in real time.

[0033] Sp5. Risk Classification and Early Warning Output: The risk classification threshold is set into four levels based on historical bird damage statistics and expert experience: Risk value 0-30 is Level I (green, normal), indicating that bird activity poses no substantial threat to the line and only requires daily recording; Risk value 31-60 is Level II (yellow, caution), indicating potential bird damage hazards, such as minor stopovers or low-frequency crossings, which may cause slight pollution; Risk value 61-80 is Level III (orange, warning), indicating a significant bird damage risk, with possible nesting or flocking, requiring immediate intervention; Risk value 81-100 is Level IV (red, emergency), indicating an extremely high bird damage risk, which may lead to flashover or short circuits, requiring emergency handling.

[0034] The early warning information includes the line number, pole number, grid location, risk level, dominant behavior, and recommended measures, and is pushed through multiple channels such as the maintenance APP, SMS, voice broadcast, and drone dispatch system. The protection strategy is automatically linked: Level II activates the sound and light bird deterrent device (intermittent operation), Level III activates laser scanning and high-pressure air jet spray, and Level IV dispatches drones to deliver bird deterrents and arranges manual inspections.

[0035] SP6, Self-learning closed-loop optimization: Sp61, Bias Assessment: Maintenance personnel upload photos, videos, or fault records of bird-related incidents via a dedicated mobile app. The system automatically extracts the event time, location, and actual risk level (manually labeled as Level I-IV or no event). The system performs a one-to-one match and comparison between the uploaded event and the model's predicted risk level for the same grid at the same time. For example, if the model predicts Level III but the actual incident is nest building causing flashover (labeled as Level IV), it is recorded as a prediction bias. Based on the cumulative comparison sample (at least 50 valid events within 30 days), an evaluation index system is established: Accuracy is calculated by dividing the number of correctly predicted levels by the total number of events; Recall is calculated by dividing the number of correctly identified actual high-risk events (Level III and above) by the total number of actual high-risk events; the F1 score is used as the harmonic mean of accuracy and recall to comprehensively evaluate model performance.

[0036] Sp62, Parameter Correction Matrix Generation: Calculates the error based on the deviation between prediction and actual values, for example, using the absolute value of the difference between the predicted risk value and the actual labeled level as the basic error. The system generates multi-dimensional parameter correction instructions, covering spatiotemporal weights (adjusting day / night or seasonal weights), behavioral threat coefficients (reducing the weight of misjudged behaviors), and three-dimensional correction function parameters (fine-tuning the wind speed or humidity impact curves), progressively optimizing the model.

[0037] Sp63, Transfer Learning Adaptation: Introduces a transfer learning mechanism to transfer the parameters of mature models from validated lines to newly built lines or lines with large environmental differences. Only the top-level classifier is fine-tuned with a small number of samples (≤100 samples), achieving rapid cross-regional adaptation and long-term self-evolution, with automatic iteration and updates every quarter.

[0038] 2. Predictive Early Warning Mechanism: Based on the comprehensive risk change trend over the past 7 days, the system uses a Long Short-Term Memory (LSTM) network to predict the risk curve for the next 24 hours. When the risk escalation rate exceeds the threshold and the risk is predicted to reach the warning level 6 hours later, the system issues a dynamic warning 6 hours in advance, automatically triggering the bird deterrent equipment to switch from low-power standby to full-power operation, thus achieving preventive intervention.

[0039] 3. Auxiliary function module: Bird Behavior-Equipment Impact Matrix: This matrix quantifies the impact of bird behavior on line components. Nesting has the greatest impact on insulators and fittings; defecation poses a high threat to insulators but a lower threat to conductors; and resting has a moderate impact on conductors and fittings. Matrix coefficients are loaded into the risk calculation process in real time, enabling precise risk assessment driven by behavior.

[0040] Bird Risk Knowledge Graph: A knowledge graph containing bird species, ecological distribution, behavioral characteristics, and risk levels is constructed and stored in a graph database. Relationships are analyzed through a graph reasoning engine; for example, inputting "egret + wetland + spring" can automatically identify potentially high-risk areas, providing interpretability support and decision-making assistance for risk models.

[0041] 4. System Hardware and Deployment Architecture: The system is deployed on a low-power distributed monitoring network. Each node integrates the following modules: image acquisition uses a high-definition camera + infrared illumination, with an average power consumption of 80mW; soundprint acquisition uses a quad microphone array, with a power consumption of 50mW; meteorological monitoring uses a multi-parameter sensor, with a power consumption of 30mW; the main control and AI acceleration use a low-power microcontroller + edge AI chip, with a standby power consumption of <1mW; communication uses an NB-IoT or LoRa module, supporting event-triggered upload.

[0042] The node only uploads data when bird activity or sudden weather changes are detected (≤5 times per day on average), and is powered by a combination of a solar panel (5W) and a lithium battery (5000mAh), with a battery life of ≥2 years.

[0043] 5. Execution Process of Bird Hazard Risk Assessment and Early Warning System for Transmission Lines: The system comprises seven functional modules that execute collaboratively in a closed-loop process: The multimodal data acquisition module collects and uploads multi-source data in real time; the bird identification and behavior analysis module completes identification and parameter extraction at the edge; the spatiotemporal risk modeling module constructs a rasterized risk field in the cloud; the 3D correction and risk fusion module fuses multi-dimensional factors in real time to generate a comprehensive risk map; the early warning decision and strategy linkage module outputs tiered early warnings and automatically executes protection measures; the model self-learning and knowledge graph module continuously optimizes parameters and updates the knowledge base based on event feedback; and the low-power distributed monitoring terminal network achieves a network of hundreds of nodes, covering real-time monitoring and data reporting across the entire line.

[0044] The entire process forms an intelligent closed loop: perception → identification → modeling → correction → early warning → execution → feedback → optimization.

[0045] 6. Data Utilization and Transmission Methods: The system data flow adopts a layered architecture of edge preprocessing + cloud-based deep analysis + two-way closed-loop feedback to ensure low power consumption, high real-time performance, and continuous model evolution.

[0046] Raw data acquisition and local storage (node ​​end): The monitoring node continuously runs sensors, performing routine sampling every 5 minutes (image snapshots, voiceprint fragments, instantaneous meteorological values). When bird activity is detected (image motion difference > 30% or voiceprint energy > 40dB) or a sudden weather change (wind speed change > 3m / s, rainfall initiation), an event acquisition mode is triggered, recording complete multimodal data for 5 seconds before and after the event. All raw data is appended with a unified timestamp and geographic coordinates, compressed into a structured package (image H.264 encoding, voiceprint WAV, meteorological / ecological JSON), and preferentially stored on a local 128GB TF card (with 30-day cyclic overwrite) as offline backup and fault tracing source.

[0047] Edge-end preliminary identification and simplified upload (node ​​→ cloud): The node's main control chip runs a lightweight dual-channel neural network to identify event data in real time and output structured behavioral parameters (type, behavior, duration, etc.). Only a simplified parameter package (<50KB) and key evidence (1-second thumbnail video + key voiceprint fragment) are uploaded to the cloud via NB-IoT / LoRa in an event-triggered manner (average ≤5 uploads per day, single power consumption <10mJ). The original complete data is only transmitted back on demand when requested by the cloud, avoiding continuous communication power consumption.

[0048] Cloud-based multi-source fusion and risk modeling (cloud-based core processing): The cloud receives parameter packages uploaded by all nodes, aggregates them according to geographic grids and time series, and forms the spatiotemporal distribution of bird activity across the entire region. Combining satellite ecological data (updated daily) and historical behavior databases, it performs full-process modeling from Sp3 to Sp4, generating a comprehensive risk heatmap updated every minute. Risk calculation results are stored in a distributed database (supporting TB-level historical traceability) and simultaneously pushed to the operation and maintenance scheduling center and mobile APP.

[0049] Warning issuance and device linkage (cloud → node / device): Once the risk reaches Level II or above, the cloud immediately generates standardized warning instructions (JSON format, including target tower, protection level, and activation sequence), and sends them to the corresponding node or independent bird deterrent device via the MQTT protocol. Upon receiving the instructions, the node executes them locally (e.g., activating an audible and visual bird deterrent device) and sends back execution confirmation (including device status and power consumption logs). Drone scheduling instructions are simultaneously pushed to the ground station, achieving a response time within seconds.

[0050] Event Feedback and Model Closed-Loop Optimization (Operations Personnel → Cloud → Entire System): Operations personnel use the app to take photos / videos of the scene, label the actual bird damage type and severity, and upload them to the cloud. The system automatically matches the event's spatiotemporal coordinates, compares them with historical prediction records, and calculates deviation metrics (accuracy, recall, F1 score). At least 50 valid feedback samples are collected monthly, triggering model retraining: adjusting weight parameters, updating the knowledge graph, and optimizing the prediction curve. The new model version is pushed to edge nodes via OTA (Over-The-Air) updates, achieving synchronized evolution across the entire network.

[0051] Data security and redundancy assurance: All transmitted data is encrypted using AES-128, and two-way authentication is established between nodes and the cloud. Data for critical risk events is backed up three times (local TF card, cloud master database, and off-site disaster recovery) to ensure traceability. The system supports offline operation: In the event of a network interruption, nodes continue to identify and store data locally, and batch retransmission occurs upon recovery. Specific Implementation Example 2

[0052] like Figures 1 to 3 As shown, the following is a detailed description of the core hardware in this solution: This technical solution operates in all-weather outdoor environments along transmission lines, suitable for 110kV to 1000kV overhead transmission line corridors, covering diverse terrains including plains, hills, mountains, wetlands, and farmland. The system adopts an edge-cloud collaborative architecture, with edge monitoring nodes deployed on poles or dedicated supports. The operating temperature range is -40℃ to +85℃, with an IP67 protection rating, supporting extreme weather conditions such as level 12 gales, heavy rain, snow, and lightning. Nodes are powered by a hybrid of solar panels and lithium batteries, with an average daily power consumption of ≤5Wh, and can operate independently for more than 7 consecutive days of cloudy or rainy weather. Communication is based on the operator's NB-IoT public network or a self-built LoRa private network, with a coverage rate of ≥99% and a single node communication distance of ≥5km (open ground). The cloud is deployed in a highly available data center, supporting elastic expansion. A single 100km line can process ≤500MB of data per day, with an annual storage capacity of ≤200GB.

[0053] Detailed hardware composition and description of each module: The multimodal data acquisition module consists of an image acquisition unit, a soundprint acquisition unit, a micro-meteorological monitoring unit, an ecological environment sensing unit, and a time synchronization unit. The image acquisition unit uses a high-definition day / night dual-mode camera with an infrared fill light and a wide-angle lens, supporting 1080P resolution, 25fps frame rate, color imaging during the day, and switching to infrared mode (wavelength 850nm) at night. The fill light distance is ≥30m, the lens field of view is ≥90°, and it features autofocus and rain / snow enhancement functions, used to capture dynamic images, postures, and nest material details of birds. The soundprint acquisition unit uses a four-element directional MEMS microphone array with a wind noise suppression cover and a preamplifier circuit. The microphone sensitivity is ≥-38dB, the frequency response range is 20Hz~20kHz, and the array spacing is 20cm to form a beam pointing towards the circuit, suppressing environmental noise ≥20dB. It is used to collect bird calls, wing flapping, flock calls, and flight airflow sounds, achieving a sound source localization accuracy of ≤1m. The micro-meteorological monitoring unit employs an integrated micro-weather station, comprising a three-cup anemometer (starting wind speed 0.3 m / s, accuracy ±0.3 m / s), a digital temperature and humidity composite probe (humidity accuracy ±3%RH, temperature accuracy ±0.5℃), a tipping bucket rain gauge (resolution 0.1 mm), and a digital barometric pressure sensor (accuracy ±1 hPa). This unit is used to perceive local meteorological changes in real time to assess the impact of wind, rain, and humidity on bird behavior and pollution risks. The ecological environment sensing unit consists of a node-embedded high-precision GNSS receiver (supporting GPS / BeiDou / GLONASS, positioning accuracy ≤1 m) and a cloud-based satellite imagery interface. The cloud periodically acquires satellite imagery with a resolution ≤1 m, combining this with the node's location to extract vegetation type, water distance, and food source density. The time synchronization unit uses a high-stability crystal oscillator in conjunction with a BeiDou / GPS timing module, achieving a timing accuracy ≤10 ms. This provides a unified timestamp for all sensor data, ensuring strict synchronization of multimodal data.

[0054] The bird recognition and behavior analysis module consists of a main control and AI acceleration unit and a local storage unit. The main control and AI acceleration unit utilizes a low-power embedded SoC with an integrated neural network accelerator (NPU). The main control unit supports a real-time operating system, has a 32-bit RISC core operating at a frequency ≥500MHz, and ≥512MB of memory. The NPU supports INT8 quantization inference, with a computing power ≥2 TOPS, and is used to run a lightweight dual-channel convolutional neural network (voiceprint + image). A single recognition time is ≤100ms, and power consumption is ≤1W. The local storage unit uses an industrial-grade 128GB TF card, supports loop recording, and stores the most recent 30 days of raw multimodal data and recognition results for offline analysis and fault tracing.

[0055] The spatiotemporal risk modeling module consists of a computing server cluster and a geographic information processing unit. The computing server cluster employs high-performance CPU servers paired with GPU accelerator cards. The CPUs support multi-core parallel processing, with a single node having ≥32 cores, used for raster modeling and Gaussian diffusion calculations. The GPUs are used for deep learning model training and LSTM risk prediction, with ≥16GB of GPU memory. The cluster supports containerized deployment and automatic scaling. The geographic information processing unit uses a GIS engine server, integrating an open-source GIS engine, supporting the WGS84 coordinate system, and generating raster indexes and risk heatmaps in real time, outputting in GeoJSON format.

[0056] The 3D correction and risk fusion module adopts a real-time data processing unit, which consists of streaming computing engine nodes. Based on a distributed streaming processing framework, it receives data from all network nodes every 10 minutes, performs weighted nonlinear regression, and dynamically outputs the comprehensive risk correction coefficient.

[0057] The early warning decision-making and strategy linkage module consists of an early warning push server and a protection equipment execution unit. The early warning push server uses a message middleware cluster, supports MQTT / WebSocket protocols, and pushes early warnings to the operation and maintenance APP, SMS gateway, and voice broadcast terminal within seconds. The protection equipment execution unit includes an audible and visual bird deterrent device (high-decibel speaker ≥120dB + strobe light, supporting intermittent / continuous modes), a laser bird deterrent device (green laser 532nm, scanning range ±60°, safe power <5mW), and a high-pressure air jet injector (high-pressure gas cylinder + solenoid valve, spray distance ≥10m). The equipment supports remote command control and transmits execution status.

[0058] The model self-learning and knowledge graph module consists of a model training server and a knowledge graph database. The model training server uses a GPU deep learning cluster, supports transfer learning and online fine-tuning, and automatically iterates the model monthly based on feedback data. The knowledge graph database uses a graph database server, stores four-ary relationships between birds, ecology, behavior, and risk, supports semantic reasoning, and has a response time of ≤200ms.

[0059] The low-power distributed monitoring terminal network consists of a communication module, a power management unit, and a structural and protection unit. The communication module uses either an NB-IoT communication module or a LoRa wireless module. NB-IoT leverages the wide coverage of operator 4G / 5G networks, while LoRa operates in the 868MHz frequency band with strong anti-interference capabilities. Both support event-triggered uploads and have a sleep power consumption of <10μA. The power management unit uses a 5W solar panel paired with a 5000mAh lithium iron phosphate battery and an MPPT charging controller, supporting a wide voltage input (6V~24V), and features overcharge / over-discharge / reverse connection protection, with a battery life of ≥2 years. The structural and protection unit uses an aluminum alloy shell with waterproof connectors and a surge protection module. The shell has an IP67 protection rating and a built-in surge protector (8 / 20μs, 10kA), supporting pole and tower clamp installation.

[0060] Hardware Integration Description: Each monitoring node integrates image, soundprint, meteorological, main control, communication, and power supply units into a single chassis (approximately 300mm × 200mm × 150mm, weighing ≤3kg), and is fixed to the tower crossarm 1.5m below, facing the power line corridor, using standard clamps. The node has self-diagnostic capabilities, periodically reporting statuses such as heart rate, voltage, temperature, and storage space. Cloud and edge communication utilizes HTTPS+MQTT dual protocols, with all data transmissions encrypted using AES-128, and supports OTA firmware upgrades. This hardware system achieves all-weather sensing, low-power operation, edge intelligence, and cloud-edge collaboration, meeting the long-term stable operation requirements of transmission lines. Specific Implementation Example 3

[0061] like Figures 1 to 3 As shown, the specific calculation method for each step in this scheme is as follows: 1. Voiceprint-Image Dual-Channel Convolutional Neural Network (Bird Recognition and Behavioral Feature Extraction) Input data: 5-second image and video clips (1080P, 25 frames / second, 125 frames in total) collected from edge nodes and corresponding 5-second audio clips (48kHz sampling, 16-bit depth), with timestamps and geographic coordinates.

[0062] Output results: bird species identifier (e.g., "egret"), behavior type (resting, nesting, flying across, gathering in groups, defecating), behavior parameters (duration in seconds, group size and number of birds, activity trajectory coordinate sequence, threat index 0-1).

[0063] The system transforms raw multimodal data into structured behavioral parameters in real time at edge nodes, reducing the amount of data uploaded and providing accurate input for risk modeling in the cloud.

[0064] Calculation steps: First, process the image channels: extract 1 key image per second from the video, for a total of 5 frames, scale them to 224×224 pixels and normalize the brightness and contrast; the slow path network extracts static posture features of each frame, such as wing opening angle and beak orientation; the fast path network analyzes inter-frame motion and generates bird trajectory and speed; merge the slow and fast path features and output a heatmap of 17 key point positions (beak, wings, tail, claws, etc.).

[0065] Next, the voiceprint channel is processed: the 5-second audio is divided into overlapping 25-millisecond windows with a step size of 10 milliseconds, and converted into a 128×128 Mel spectrogram; the input is a deep residual network to extract voiceprint semantics, such as call frequency and rhythm pattern; a temporal attention mechanism is added to highlight high-energy call segments and suppress background noise.

[0066] Then, multimodal fusion is performed: image key point features and voiceprint semantic features are concatenated into a 512-dimensional joint vector, and the time dimension is aligned through a cross-attention mechanism (such as the synchronization of bird calls and wing flapping); when the image signal-to-noise ratio is lower than 30dB (such as at night or in foggy weather), the system automatically increases the weight of the voiceprint features to 0.7 and decreases the weight of the image to 0.3.

[0067] Final classification and parameter output: Joint features are input into a multi-label classifier, which outputs the probabilities of 5 behaviors; specific behaviors are determined according to rules: resting is when the key point remains in the hardware area for more than 3 seconds without displacement, nest building is when the beak is detected holding a slender object and going back and forth more than 3 times, flight crossing is when the trajectory passes horizontally through the gap of the wire less than 50cm, group gathering is when more than 5 birds are detected in the same frame with an average distance of less than 2 meters, and defecation is when the bird squats down and its tail presses down, leaving a white trail; a structured parameter package is generated and uploaded to the cloud.

[0068] 2. Spatiotemporal weighted rasterized risk modeling: Input data: Behavior parameter packages uploaded by all nodes (including behavior type, frequency, time, and grid ID) and grid division rules (50m×50m, 30m×30m, 20m×20m).

[0069] Output: A full-line basic risk heat map updated hourly (risk value 0-100).

[0070] Construct a spatially continuous risk base map as the basis for three-dimensional correction and early warning.

[0071] Calculation steps: First, perform spatial grid division: The system reads the GIS path of the line and divides the 100-meter-wide corridor into rectangular grids according to WGS84 coordinates. 110kV lines use 50m×50m grids, 220kV lines use 30m×30m grids, and mountainous terrain uses 20m×20m grids. Each grid generates a unique ID (such as G114523_36789) and stores it in the cloud grid index.

[0072] Next, a spatiotemporal distribution matrix is ​​constructed: the total number of bird activities in each grid per hour is counted as the frequency; time weights are added to each activity record—1.0 for daytime (6:00 to 18:00), 0.6 for nighttime, 1.5 for spring migration (March to May), 0.8 for winter, and 1.8 for historical peak migration days; threat scores are added to each behavior—0.9 for nesting, 0.8 for defecating, 0.7 for gathering, 0.4 for resting, and 0.3 for crossing; the frequency is multiplied by the time weight and then by the threat score, and the sum is obtained to obtain the original risk contribution value of the grid.

[0073] Then, Gaussian spatial risk diffusion is performed: the risk is diffused to the 8 adjacent grids centered on the current grid; the diffusion intensity decreases with distance - 60% risk is retained at a distance of 50 meters between adjacent grids, 30% risk is retained at a diagonal distance of about 70 meters, and risks beyond that are ignored; the risks diffused from all adjacent grids are superimposed on the original contribution value of this grid to form a smooth base risk value (0~100); finally, a full-line heatmap is output, with the color gradually changing from green to red.

[0074] 3. Three-dimensional correction based on meteorology, ecology, and behavior: Input data: basic risk heat map, real-time meteorological data (wind speed, humidity, temperature, rainfall), ecological data (vegetation type, water distance, food source density), and behavioral parameters (duration, population size).

[0075] Output: A comprehensive risk distribution map updated every 10 minutes.

[0076] Dynamically calibrate basic risks to accurately reflect the amplification or inhibition of bird damage threats by the environment and behavior.

[0077] Calculation steps: First, calculate the meteorological factor correction: the risk is adjusted upwards for every 5 m / s increase in wind speed, and the adjustment rate gradually increases for every 1 m / s increase in wind speed (e.g., 5% increase for 6 m / s, 20% increase for 10 m / s); the adjustment is also adjusted upwards when humidity exceeds 70% RH, with an increase of 5% for every 10% increase in humidity (reflecting the risk of electrical conductivity from damp feces); when combined, wind speed accounts for 60% of the influence and humidity accounts for 40%, resulting in the meteorological adjustment value.

[0078] Next, ecological factor corrections are calculated: water bodies less than 200 meters away are adjusted by 50%, water bodies between 200 and 500 meters away are adjusted by 20%, and water bodies more than 500 meters away are not adjusted; vegetation types are adjusted by 50% for wetlands, 30% for forests, and 10% for farmland; the ecological adjustment value is obtained by averaging the water body and vegetation adjustment values.

[0079] Then, the behavioral factor correction is calculated: the correction is increased by 10% for every 10-second increase in duration, but the growth rate slows down as time goes on (fast in the early stage and slow in the later stage); the correction is increased by 30% for 5 groups, 80% for 10 groups, and 200% for 20 groups (the clustering effect is significantly amplified); the duration accounts for 55% of the influence, and the group size accounts for 45%, which are combined to obtain the behavioral adjustment value.

[0080] Finally, a comprehensive correction coefficient is generated: the weighted average of meteorological adjustment value (40%), ecological value (30%), and behavioral value (30%) is taken, and the total coefficient is limited to between 0.5 and 2.0; the basic risk value is multiplied by this coefficient to generate a real-time updated comprehensive risk distribution map.

[0081] 4. Risk classification and early warning output: Input data: Comprehensive risk distribution map (one risk value per grid).

[0082] Output results: Tiered early warning information (Level I to IV) and protection strategy instructions.

[0083] Drive the automatic protection equipment to start and require manual intervention.

[0084] Calculation steps: The system iterates through the comprehensive risk value of each grid and classifies it: risk values ​​of 0 to 30 are Level I (green), only logs are recorded and no action is triggered; 31 to 60 are Level II (yellow), the sound and light bird deterrent is activated and runs in intermittent mode (calling for 30 seconds every 5 minutes); 61 to 80 are Level III (orange), the laser bird deterrent (±60° scanning) and high-pressure air jet are activated at the same time; 81 to 100 are Level IV (red), the drone is immediately dispatched to deliver bird deterrent and manual inspection is notified.

[0085] The generated early warning message includes the line name, tower number, grid location, risk level, dominant behavior type, and recommended measures. It is pushed out in seconds through multiple channels, including the operation and maintenance APP, SMS, voice broadcast, and drone dispatch system. After receiving the instruction, the protection equipment executes it locally and sends back a status confirmation.

[0086] 5. Predictive early warning (based on risk trend curves): Input data: The composite risk value series for each hour over the past 7 days (one time series per raster).

[0087] Output: Risk prediction curve for the next 24 hours and dynamic early warning 6 hours in advance.

[0088] Implement preventative interventions to avoid responding only when the risk reaches a critical point.

[0089] Calculation steps: The system collects the risk values ​​of each grid for the past 7 days (168 hours) to form a time series; inputs it into a Long Short-Term Memory (LSTM) network, which has been pre-trained on historical bird damage event data to learn the evolution of risk with weather, season, and behavior; and predicts the risk curve for the next 24 hours every 10 minutes.

[0090] Monitoring and prediction curve: When the risk rises at a rate exceeding 15 minutes per hour, and the risk value is predicted to reach 60 or above (Level III or above) after 6 hours, the system issues a dynamic warning 6 hours in advance; the warning triggers the bird deterrent device to switch from low-power standby (0.5W) to full-power operation (5W), and notifies the drone to fly to the nearby standby area in advance.

[0091] 6. Self-learning closed-loop optimization: Input data: Photos / videos / fault records of bird-related incidents uploaded by maintenance personnel (including manually marked actual risk levels I to IV or no incidents) and historical prediction records of the system.

[0092] Output results: Updated model parameters (spatiotemporal weights, threat coefficient, correction function), new version of knowledge graph, and transfer learning fine-tuned model.

[0093] System application: Continuously improve model accuracy and cross-regional adaptability.

[0094] Calculation Steps: Sp61 Bias Assessment: Operations personnel upload on-site evidence via a dedicated APP. The system automatically extracts the event time, location, and manually labeled level. The event is matched one-to-one with the model prediction level of the same grid at the same time. If a Level III event is predicted but a nested flashover (labeled as Level IV) actually occurs, it is recorded as a prediction bias. After accumulating at least 50 valid events within 30 days, the assessment metrics are calculated: Accuracy equals the number of correctly predicted events divided by the total number of events, Recall equals the number of correctly identified Level III and above events divided by the total number of actual high-risk events, and the F1 score is the harmonic mean of the two.

[0095] Sp62 parameter correction matrix generation: The absolute value of the difference between the predicted value and the actual labeled level is calculated as the error for each deviation event; the system generates multi-dimensional correction instructions—adjusting the day / night / seasonal time weight (e.g., reducing the night weight if there are more misjudgments at night), modifying the behavior threat score (e.g., reducing the threat score if there are high misjudgments of defecation), and fine-tuning the three-dimensional correction function curve (e.g., smoothing out its upward trend if the wind speed influence is too strong); the correction instructions are applied gradually in small steps (learning rate 0.01) to avoid oscillations.

[0096] Sp63 Transfer Learning Adaptive: When deployed to a new line, the system loads the mature model parameters of the validated line; fine-tunes the top-level classifier with only a small number of samples (≤100) on the new line, and freezes the bottom-level feature extraction network; after fine-tuning, it quickly achieves an accuracy of over 90% in the new region; and automatically triggers a global model iteration every quarter based on feedback data from the entire network, which is pushed to all edge nodes via OTA to achieve synchronous evolution of the entire system.

[0097] 7. Audio-visual image fusion algorithm: Input data: Heatmap of 17 key points extracted from the image channel + semantic vector of Mel spectrum output from the voiceprint channel.

[0098] Output: Aligned multimodal joint feature vector (512 dimensions).

[0099] System application: Solves the problem of single-modal failure at night / in foggy weather, and improves the all-weather recognition accuracy to over 92%.

[0100] Calculation steps: The system expands the image key point heatmap into a pose sequence of 5 time steps along the time axis; the voiceprint semantic vector is also divided into 5 time steps of the same 5 seconds; a cross-attention matrix is ​​constructed, and the correlation score between each image time step and each voiceprint time step is calculated (based on dot product similarity); high correlation pairs (such as flapping sound and wing movement) are given high weights, and low correlation pairs (such as background wind noise) are given lower weights; the weighted image and voiceprint features are concatenated step by step, and then compressed into a 512-dimensional joint vector through a fully connected layer; at night or in foggy weather, if the average brightness of the detected image is <50 or the contrast is <30, the voiceprint weight is automatically increased from 0.5 to 0.7 to ensure that the voiceprint dominates the recognition.

[0101] 8. Multimodal Feature Matching Module: Input data: Image key point sequence (5 frames × 17 points) + voiceprint semantic sequence (5 segments × 128 dimensions).

[0102] Output: Time-aligned multimodal behavior event labels (e.g., "chirp + takeoff").

[0103] To establish a causal relationship between images and voiceprints, and avoid misjudgments (such as mistaking wind sounds for wing flapping).

[0104] Calculation steps: Align the image and the voiceprint sequence to the millisecond level by timestamp; calculate the time difference between the rate of change of key points in the image (such as wingtip displacement velocity) and the peak value of the voiceprint energy; if the time difference is <200ms and the direction of change is consistent (such as high voiceprint energy when the wingspan is large), mark it as a strong matching event; count the number of strong matching events ≥3 within 5 seconds, then it is considered a real behavior; weak matching or no matching events trigger weight reduction processing; finally output the matching confidence (0~1), which is used for weighted voting for behavior classification.

[0105] 9. Gaussian space weighting function (risk diffusion explained separately): Input data: The original risk value of the current grid + the center distance of the 8 adjacent grids.

[0106] Output: Smoothed risk contribution value after diffusion.

[0107] Simulate the continuous spatial spread of bird activity to avoid abrupt changes in grid boundaries.

[0108] Calculation steps: Using the current grid as the center, calculate the Euclidean distance to 8 adjacent grids; when the distance is ≤50 meters (side length), retain 60% of the risk; when the distance is about 70 meters (diagonal), retain 30% of the risk; when the distance is >100 meters, retain 0%; multiply the risk value of each adjacent grid by the corresponding retention ratio and add it to the current grid; at the same time, the current grid also diffuses its own risk outward, forming a two-way smoothing; finally, all grid risks are renormalized to the range of 0 to 100, generating a continuous heatmap.

[0109] 10. Weighted Nonlinear Regression (Three-Dimensional Corrected Model): Input data: real-time values ​​of meteorological (wind speed, humidity), ecological (distance from water, vegetation), and behavioral (duration, group size).

[0110] Output: Comprehensive risk correction coefficient (0.5~2.0).

[0111] Dynamically integrate multiple source factors to avoid distortion of linear models.

[0112] Calculation steps: Establish independent regression curves for each input variable: wind speed is gradually increased from 0 to 15 m / s, with a steeper slope after 5 m / s; humidity is smoothly increased from 50% to 100%, accelerating after 70%; water distance decreases inversely from 0 to 1000 meters; duration is logarithmically mapped; population size is amplified using a power function; the outputs of the 6 variables are weighted and averaged (meteorology 40%, ecology 30%, behavior 30%); if the sum is < 0.5, it is adjusted up to 0.5, and if it is > 2.0, it is adjusted down to 2.0; output the final correction coefficient, multiplied by the aforementioned basic risk.

[0113] 11. Semantic Reasoning Algorithm (Bird Risk Knowledge Graph): Input data: User queries (such as "egrets + wetlands + spring") or real-time events (birds + location + time).

[0114] Output: List of potentially high-risk areas (grid ID + risk warning).

[0115] Provide interpretable decision support for operations and maintenance personnel to assist in manual inspections.

[0116] Calculation steps: The knowledge graph contains four types of nodes: bird species, ecological distribution, behavioral characteristics, and risk level, with edges representing the strength of association; upon receiving input, the corresponding node is activated (e.g., the "egret" node); the activation value is propagated through a graph neural network, decaying along the edges (e.g., egret → wetland edge weight 1.5); the activation value is accumulated to the risk level node, and if it exceeds the threshold of 0.7, it is marked as high risk; the activation path is output (e.g., egret → wetland → nesting → level IV), and mapped to a geographic raster to generate a highlighted area.

[0117] 12. Bird Behavior-Equipment Influence Matrix (Dynamically Loaded): Input data: Real-time identified behavior type + location of line components (insulators, conductors, fittings).

[0118] Output: Threat weighting of the behavior to specific components.

[0119] Achieve accurate risk assessment.

[0120] Calculation steps: Preset static matrix: Nesting for insulator 0.9, hardware 0.8, conductor 0.7; Manure removal for insulator 0.8, conductor 0.3, hardware 0.5; Rest for conductor 0.6, hardware 0.5, insulator 0.4; The system detects the location of the behavior. If nesting occurs within 1 meter of the insulator, a coefficient of 0.9 is applied. If multiple behaviors occur simultaneously, the largest coefficient is used. The coefficient is injected into the behavior factors of the three-dimensional correction model in real time and participates in the comprehensive risk calculation. Specific Implementation Example 4

[0121] like Figures 1 to 3 As shown, the process for staff to implement protective measures through the system's predictive early warning is as follows: Based on the risk trend over the past 7 days, the system uses a Long Short-Term Memory (LSTM) network to predict the risk curve for the next 24 hours every 10 minutes. When it is predicted that the risk of a certain grid cell is increasing at a rate exceeding 15 minutes per hour, and will reach Level III (61-80) or Level IV (81-100) in 6 hours, a dynamic early warning is automatically issued 6 hours in advance. The early warning information is simultaneously pushed through a dedicated APP for maintenance personnel, SMS, and the dispatch center's large screen. The content includes the line name, tower number, grid cell location, predicted risk level, dominant behavior type, expected occurrence time, and system-recommended protective measures.

[0122] Upon receiving the alert, maintenance personnel open the app and click on the alert message. The system immediately displays a real-time heat map of the area, video clips of bird activity over the past 3 hours, a risk prediction curve for the next 6 hours, and comparisons with similar historical events. No manual analysis of the risk's causes is required; the system automatically generates a standardized action list based on the prediction level: if the prediction is Level III, the recommended measure is to remotely activate a laser bird deterrent device for ±60° scanning, combined with high-pressure airflow spray; if the prediction is Level IV, the recommended measure is to dispatch a drone to deliver bird deterrents and arrange for manual inspections.

[0123] Staff simply click the "Confirm Execution" button, and the system immediately issues the command: the bird deterrent equipment receives the MQTT command, starts locally, and transmits its operational status; the drone ground station receives the dispatch task, automatically plans its flight path, and takes off; the electronic work order is simultaneously pushed to the mobile device of the nearest inspection team, including navigation and positioning, pole photos, and task deadlines. Upon arrival at the site, inspection personnel use the app to take photos and upload comparison images before and after bird deterrence, selecting the actual handling result option, such as "nesting materials removed," "birds dispersed," or "no abnormalities." The system automatically matches the on-site feedback with the predicted results. If they match, the prediction is marked as successful, and the model weights are strengthened; if there is a deviation, it is recorded as an optimization sample and enters the next round of self-learning.

[0124] The following are actual use cases for this solution: One spring day, a 220kV transmission line passed through a plain farmland area, surrounded by fishponds and rice paddies. A monitoring node below the crossarm of a tower captured a group of egrets. Images showed multiple egrets circling near the insulators, some carrying reed twigs and flying back and forth. Voiceprint recordings revealed high-frequency social calls synchronized with wing flapping sounds. Weather station measurements indicated high humidity and moderate wind speed. Cloud data showed nearby water bodies and vegetation of wetland farmland type. All data was stamped with a unified timestamp and geographic coordinates, stored locally, and then uploaded.

[0125] The edge node immediately runs a dual-channel neural network. The image channel identifies the egret species and extracts the object carried in its beak and its back-and-forth trajectory, while the vocalization channel confirms the call characteristics. After fusion, it is determined to be nest-building behavior, with a group size of 8 egrets, lasting for more than ten seconds, and a threat index exceeding 0.8. The structured parameter package is quickly uploaded to the cloud.

[0126] The cloud-based system divides the region into 30-meter square grids based on voltage levels. It then counts the frequency of nesting activity within each grid, combining this with the time weighting of the spring migration season and nesting threat scores to calculate a baseline risk value. Risk is then smoothly transferred between adjacent grids using Gaussian diffusion, creating a continuous baseline risk heatmap.

[0127] The system incorporates real-time meteorological and ecological data for three-dimensional correction: high humidity increases the risk of fecal conductivity, thus raising the correction coefficient; near-water and wetland vegetation further amplifies the ecological impact; and nesting duration and population size significantly increase behavioral factors. After weighted fusion of these three types of corrections, a comprehensive correction coefficient is generated. The base risk, multiplied by this coefficient, is then raised to the orange alert level.

[0128] Early warning information is immediately pushed to the maintenance personnel's APP and the dispatch center, including the line segment, tower location, risk level, and dominant behavior. The system automatically activates protective equipment: laser bird deterrents begin scanning the tower area, high-pressure air jets are periodically activated, and electronic work orders are simultaneously dispatched to the nearest inspection team.

[0129] The forecasting module analyzed the risk trends over the past week and, considering the upcoming light rain, determined that the risk would escalate to the red level within hours, thus issuing a dynamic escalation warning in advance. The drone, receiving instructions, flew to the top of the pole carrying bird deterrent agent and hovered in wait, while the bird deterrent equipment switched to full power operation.

[0130] Upon arrival, inspectors observed that the egrets had been driven away and temporarily flown off, with a small amount of nesting material remaining on the insulator surface. After climbing the tower and removing the nesting material, they uploaded before-and-after photos via an app and marked the process as complete. The system received feedback, the risk level quickly dropped to a normal level, and the protective equipment automatically returned to low-power standby.

[0131] The cloud platform compares the predicted data with the actual response results, assesses model bias, fine-tunes the ecological and behavioral correction weights, and updates the high-risk rules for "egrets + wetlands + spring" in the knowledge graph. The new model version is wirelessly pushed to all nodes along the line, completing a closed-loop optimization. The entire process, from bird appearance to nest damage removal, is automatically sensed, intelligently judged, and precisely intervened, successfully avoiding flashover tripping. Human intervention is limited to confirmation and on-site handling, significantly improving the efficiency and reliability of bird damage control on transmission lines. Specific Implementation Example 5

[0132] like Figures 1 to 3 As shown, the following is an explanation of the appendix. Figure 3 Detailed explanation: The horizontal axis (X-axis) runs from left to right, and the values ​​represent the width. The vertical axis (Y-axis) runs from top to bottom on the left, and the values ​​represent the height. The entire graph represents one region.

[0133] Colors represent risk levels (see the color bars on the right): Dark blue: 0-20 → Low impact; Light blue / green: 20-60 → Generally require precautions; Yellow: 60-80 → Danger! Call the police! The yellow circle in the picture can be interpreted as the place where birds build nests, which is the highest risk (around 80 points), and the system will automatically issue an alarm.

[0134] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising a reference structure" does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0135] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for risk assessment and early warning of bird damage to power transmission lines, characterized in that, Includes the following steps: Sp1. Data Acquisition and Multimodal Fusion: Image data, voiceprint data, meteorological parameters and ecological environment information along the transmission line are collected to form multimodal data. The multimodal data is uniformly encoded by synchronizing timestamps and geographic coordinates, and an original bird activity sample set is generated. Sp2. Bird identification and behavioral feature extraction: The original bird activity sample set is processed using an audio-visual image fusion algorithm to identify bird species and their behavioral patterns. The behavioral patterns include resting, nest building, flight crossing, group gathering and defecation behavior, and output behavioral feature parameters. Sp3, Spatiotemporal weighted gridded risk modeling: The transmission line area is divided into multi-level spatial grid units. A spatiotemporal coupled risk field model is established based on the behavioral characteristic parameters, activity frequency and time distribution, and the basic bird damage risk value of each grid is calculated. Sp4, Meteorological-Ecological-Behavioral Three-Dimensional Correction: Real-time meteorological elements and ecological factors are introduced into the spatiotemporal coupled risk field model, and the risk field is dynamically corrected through multiple correction algorithms to form a comprehensive risk distribution map; Sp5, Risk Classification and Early Warning Output: Based on the comprehensive risk value, set the early warning level threshold, output graded risk early warning information, and link with the transmission line operation and maintenance system to execute corresponding protection strategies. Sp6, Self-learning closed-loop optimization: Based on the feedback results of actual bird damage events, the model parameters are corrected in reverse, and the spatiotemporal weights and threat coefficients are continuously updated through learning to achieve adaptive optimization of the model.

2. The method for risk assessment and early warning of bird damage to transmission lines according to claim 1, characterized in that, The spatiotemporal weighted rasterized risk modeling in Sp3 includes the following sub-steps: Sp31. The transmission line area is divided into grid units of fixed size according to geographical coordinates, and the size of the grid unit is dynamically adjusted according to the line voltage level and terrain complexity. Sp32. Using bird activity frequency as the main variable and activity time period as the time weight, a spatiotemporal distribution matrix is ​​constructed. The time weight is calculated by using a time-segmented statistical method to weight day and night, season and migration cycle. Sp33 uses a Gaussian spatial weighting function to calculate the risk propagation coefficient between adjacent grids, and achieves continuous spatial propagation of risk through exponential decay, generating a smooth risk field distribution.

3. The method for risk assessment and early warning of bird damage to transmission lines according to claim 1, characterized in that, The bird identification and behavioral features in Sp2 are extracted using a dual-channel convolutional neural network of voiceprint and image. The voiceprint recognition channel is used to identify bird species, and the image recognition channel is used to extract behavioral and posture features. These features are then fused through a multimodal feature matching module to improve the identification accuracy in nighttime and foggy environments.

4. The method for risk assessment and early warning of bird damage to transmission lines according to claim 1, characterized in that, The Sp4 three-dimensional correction model of meteorology, ecology, and behavior is constructed based on weighted nonlinear regression, specifically including: The meteorological factor correction weights are jointly determined by wind speed and humidity. The wind speed weight increases non-linearly with wind force level, while the humidity weight enhances the bird activity induction effect in high humidity environments. The ecological factor correction weights are determined by the distance to the water area and the vegetation type. The distance to the water area is decayed using an inverse proportional function, and the vegetation type is assigned a graded weight based on food source abundance and habitat suitability. The behavioral factor adjustment weight is determined by the duration of the behavior and the group size, where the duration is mapped using a logarithmic function and the group size is amplified by a power function to amplify the risk of high-density clustering. By normalizing the weights of each factor, a comprehensive risk correction coefficient is formed, enabling dynamic coupling correction of multiple factors including meteorology, ecology, and behavior.

5. The method for risk assessment and early warning of bird damage to transmission lines according to claim 1, characterized in that, The self-learning closed-loop optimization includes the following sub-steps: Sp61. Real-time comparison of the risk level predicted by the model with the risk level of actual bird damage events, and establishment of a deviation assessment index system; Sp62. Calculate the risk prediction error based on the deviation assessment results and generate a multi-dimensional parameter correction matrix, covering spatiotemporal weights, threat coefficients and correction factors; Sp63 introduces a transfer learning algorithm to transfer model parameters from validated regions to new regions. Through few-sample fine-tuning, feature weights are adaptively adjusted, thereby improving the cross-regional adaptability and long-term self-evolution capability of the risk model.

6. The method for risk assessment and early warning of bird damage to transmission lines according to claim 1, characterized in that, The prediction method is based on the risk trend curve to achieve predictive early warning. When short-term meteorological or ecological changes cause the risk gradient to rise, a dynamic early warning is issued in advance and the pre-operation strategy of protective equipment is triggered.

7. The method for risk assessment and early warning of bird damage to transmission lines according to claim 1, characterized in that, A bird behavior-equipment impact matrix is ​​established, which is used to quantify the impact coefficients of various bird behaviors on line components, and dynamically adjusts the risk calculation weights based on the threat coefficient to achieve behavior-driven risk assessment.

8. The method for risk assessment and early warning of bird damage to transmission lines according to claim 1, characterized in that, A bird risk knowledge graph is constructed, which includes bird species nodes, ecological distribution nodes, behavioral characteristic nodes, and risk level nodes. Potential high-risk areas are identified through semantic reasoning algorithms, and interpretability support is provided for risk assessment models.

9. A system based on the transmission line bird hazard risk assessment and early warning method according to any one of claims 1-8, characterized in that, The system includes: The multimodal data acquisition module is used to collect image, acoustic, meteorological, and ecological data along the power transmission line. The bird identification and behavior analysis module is used to identify bird species and behavioral characteristics and output behavioral parameters; The spatiotemporal risk modeling module is used to construct a spatiotemporal weighted rasterized risk model and calculate the basic risk value. The 3D correction and risk fusion module is used to integrate meteorological, ecological and behavioral factors to dynamically correct the risk model. The early warning decision and strategy linkage module is used to output hierarchical early warning information and link with the power transmission operation and maintenance system to execute protection strategies. The model self-learning and knowledge graph module is used to automatically correct model parameters and update the bird risk knowledge base based on event feedback. A low-power distributed monitoring terminal network is used to achieve multi-point real-time monitoring and remote data reporting.

Images