Intelligent inspection and bird repelling method and system for power transmission line

Abstract

The invention discloses a power transmission line intelligent inspection and bird repelling method and system, and relates to the technical field of bird repelling, and the method comprises the steps: S1, collecting a bird activity signal and a weather condition parameter from a foggy environment around an airport through a sensor network, and obtaining a bird gathering density and a weather influence factor; classifying the data by adopting a K-means clustering method in combination with Euclidean distance measurement and cluster center initialization, and determining a bird activity mode and a weather change trend by applying data standardization processing and iterative convergence conditions; according to the intelligent routing inspection and bird repelling method and system for the power transmission line, accurate monitoring and dynamic bird repelling are achieved through multi-technology fusion, interference of bird activities on airport safety in foggy days is effectively reduced, and the aviation operation efficiency is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of bird repelling, in particular to a power transmission line intelligent inspection and bird repelling method and system. BACKGROUND

[0002] Bird repelling technology is crucial in agriculture, aviation and urban management, directly related to crop safety, flight safety and ecological balance. Effective bird repelling measures can significantly reduce bird interference with human activities, ensuring economic benefits and public safety.

[0003] Currently, traditional bird repelling devices such as sound wave bird repellers, laser equipment or physical nets have shown certain effects in practical applications, but have obvious limitations. Most of these devices rely on manual operation or fixed modes, making it difficult to adapt to the dynamic changes of bird behavior and the complexity of environmental conditions. For example, traditional devices cannot flexibly adjust according to the activity patterns of bird flocks or weather changes, resulting in unstable bird repelling effects, and even failure in specific scenarios such as night or heavy fog weather, where devices cannot accurately perceive bird activity. The core challenge lies in achieving the intelligentization and adaptive adjustment of bird repelling devices. Traditional devices lack real-time perception ability of bird activity patterns, making it impossible to dynamically adjust bird repelling strategies according to the gathering time, location or species of bird flocks. This lack of perception ability often leads to insufficient or excessive bird repelling in the face of different bird habits or environmental changes. Further, due to the lack of comprehensive analysis of weather conditions such as wind speed, rainfall or light, devices cannot optimize bird repelling frequency and intensity. For example, in strong wind weather, the propagation effect of sound wave bird repellers may be significantly weakened, but the device cannot automatically adjust the output power or switch to other bird repelling methods, resulting in bird repelling failure. SUMMARY

[0004] The purpose of the present application is to provide a power transmission line intelligent inspection and bird repelling method and system, which deeply integrates traditional bird repelling devices with Internet of Things technology, enabling the device to autonomously adjust bird repelling intensity and frequency based on real-time data of bird activity patterns and weather conditions, achieving intelligent upgrade of bird repelling devices.

[0005] To achieve the above purpose, the present application provides the following technical solution: a power transmission line intelligent inspection and bird repelling method, comprising:

[0006] S1, collecting bird activity signals and weather condition parameters from the surrounding fog environment of the airport through a sensor network to obtain bird flocking density and weather influence factors;

[0007] S2, according to the bird flocking density and weather influence factors, using the K-means clustering method combined with Euclidean distance measurement and cluster center initialization to classify and process the data, and applying data standardization processing and iterative convergence conditions to determine bird activity patterns and weather change trends;

[0008] S3, if the bird activity pattern shows a clustering trend and the cluster quality is confirmed by the contour coefficient evaluation, the bird flock movement path is predicted by a neural network model to obtain a potential interference area;

[0009] S4, for the potential interference area, state information of an Internet of Things device is obtained, a bird repelling intensity demand level is judged, and if the demand level exceeds a preset threshold, a frequency adjustment process is activated;

[0010] S5, according to the bird repelling intensity demand level, a genetic algorithm defined by a chromosome coding method and a fitness function is used to optimize the bird repelling frequency parameters, a cross probability and a mutation rate are set to control the generation of a population size determined by the selected parameter group, and an adjusted output scheme is obtained through a selection operation mechanism and a termination condition judgment, while the optimal individual is reserved as a fog day exclusive configuration;

[0011] S6, by the adjusted output scheme, control instructions are sent to the bird repelling device to determine the final sound wave or laser intensity value;

[0012] S7, if the weather change trend shows that the propagation is blocked and the signal attenuation is confirmed by the anomaly point detection, the standby bird repelling mode is switched to obtain an enhanced adaptive response, wherein the standby bird repelling mode integrates the state information of the Internet of Things device to recalculate the intensity value.

[0013] Preferably, the S1 comprises:

[0014] Bird activity signals and weather parameters in the fog environment around the airport are collected by a sensor network to generate an original signal data set and a weather data set;

[0015] The original signal data set is denoised by a signal processing technology to obtain a bird activity signal set;

[0016] According to the bird activity signal set, the bird clustering density is calculated based on the signal intensity to generate a density distribution data;

[0017] If the clustering density in the density distribution data exceeds a preset threshold, the weather data set of the corresponding area is analyzed to extract humidity data and visibility data;

[0018] The humidity data and the visibility data are fused by a weighted fusion algorithm to generate a weather influence factor;

[0019] According to the weather influence factor and the density distribution data, a decision tree algorithm is used to judge the correlation between bird activity and fog environment to generate a correlation analysis result;

[0020] Through the correlation analysis result, a regression analysis algorithm is used to predict the bird clustering density change trend to generate prediction data.

[0021] Preferably, the S2 comprises:

[0022] Collecting bird activity signals and weather parameters from the airport surrounding fog environment through the sensor network to generate a raw signal dataset and a weather dataset;

[0023] Preprocessing the raw signal dataset and the weather dataset by using a data standardization technology to generate a standardized signal dataset and a standardized weather dataset;

[0024] Classifying the standardized signal dataset by using a K-means clustering algorithm, combining with Euclidean distance measurement and cluster center initialization to generate a bird activity pattern classification result;

[0025] If the number of samples of a certain cluster in the bird activity pattern classification result exceeds a preset threshold, performing dimensionality reduction processing on the standardized weather dataset to generate a reduced dimension weather dataset;

[0026] According to the reduced dimension weather dataset, extracting main weather features by using a principal component analysis algorithm to obtain a weather influence factor;

[0027] By using a linear regression algorithm, analyzing the correlation between the weather influence factor and the bird activity pattern classification result to obtain a correlation model of bird activity and weather change;

[0028] According to the correlation model, predicting the distribution of the bird activity pattern under different weather change trends to generate prediction distribution data.

[0029] Preferably, the S3 comprises:

[0030] Obtaining bird activity signals from the sensor network to generate a raw signal dataset;

[0031] If the number of signal points in the raw signal dataset exceeds a preset threshold, using an outlier detection algorithm to remove isolated signal points to obtain a filtered signal dataset;

[0032] According to the filtered signal dataset, using a K-means clustering algorithm to generate a bird activity pattern to obtain a clustering trend classification result;

[0033] If the contour coefficient of the clustering trend classification result exceeds a preset threshold, confirming the cluster quality to generate a cluster quality confirmation result;

[0034] According to the cluster quality confirmation result, using a neural network model to train the filtered signal dataset to predict a bird flock moving path to obtain a prediction path dataset;

[0035] Through the prediction path dataset, combining with environmental interference factors, analyzing the path coverage range to determine a potential interference area;

[0036] According to the potential interference region, generate interference region boundary data set, and obtain region distribution result.

[0037] Preferably, the S4 comprises:

[0038] Obtain state information of the Internet of Things device, including device power and location coordinates, and generate device state data set;

[0039] Through the device state data set, use support vector machine algorithm to classify the device power, judge whether the power is lower than the preset threshold, and obtain the power device list whose power is lower than the preset threshold;

[0040] According to the power device list whose power is lower than the preset threshold and the location coordinates, calculate the geometric distance between the device and the potential interference region, and generate distance distribution data set;

[0041] If the distance in the distance distribution data set is lower than the preset threshold, use K nearest neighbor algorithm to analyze the proximity of the device location and the interference region, and obtain intensity demand data set;

[0042] Through the intensity demand data set, judge whether the demand level exceeds the preset threshold, and generate demand level classification result;

[0043] According to the demand level classification result, use the pre-established frequency adjustment rule to generate adjustment instruction data set;

[0044] Through the adjustment instruction data set, generate device frequency adjustment signal, and obtain frequency adjustment execution result.

[0045] Preferably, the S5 comprises:

[0046] Obtain state information of the Internet of Things device, including device power and location coordinates, and generate device state data set;

[0047] Through the device state data set, use genetic algorithm to initialize population, set chromosome coding mode, define fitness function, and obtain initial population parameter set;

[0048] According to the initial population parameter set, set the crossover probability and mutation rate, perform crossover operation and mutation operation, and generate candidate parameter group;

[0049] Through the candidate parameter group, use selection operation mechanism, calculate the fitness value of each individual according to the fitness function, and obtain optimal parameter set;

[0050] If the fitness value in the optimal parameter set meets the preset termination condition, the optimal individual is retained, and the fog special frequency configuration is generated;

[0051] According to the fog special frequency configuration, combine the location coordinates, calculate the distance distribution between the device and the target region, and generate distance distribution data set;

[0052] By the distance distribution dataset, the K-Nearest Neighbor algorithm is used to analyze the proximity between the device location and the target area, and the frequency adjustment execution result is obtained.

[0053] Preferably, the S6 comprises:

[0054] By the bird flock movement path data, the trajectory analysis algorithm is used to generate the bird flock movement trend, and the dynamic distribution of the potential interference area is determined.

[0055] According to the dynamic distribution, the center coordinates and the boundary range of the potential interference area are calculated, and the area coverage range is obtained.

[0056] By the area coverage range, the relative distance between the device and the interference area is calculated in combination with the location coordinates of the bird repelling device, and the distance distribution dataset is generated.

[0057] If the relative distance in the distance distribution dataset is less than the preset threshold value, the K-Nearest Neighbor algorithm is used to analyze the proximity between the device and the interference area, and the weight set is obtained.

[0058] By the weight set, the generation rule of the control instruction set is adjusted, and the distribution proportion of the sound wave intensity value and the laser intensity value is determined.

[0059] According to the distribution proportion, the control instruction set for covering the potential interference area is generated, and the final sound wave or laser intensity value is determined.

[0060] By the control instruction set, the execution signal is sent to the bird repelling device, and the intensity adjustment is completed.

[0061] Preferably, the S7 comprises:

[0062] By the weather change trend data obtained by the weather sensor, the attenuation effect of rainfall or fog on signal propagation is analyzed, and the signal propagation blocked state is determined.

[0063] If the signal propagation blocked state is confirmed, the anomaly point detection algorithm is used to analyze the time series data of the signal intensity, and the signal attenuation confirmation result is obtained.

[0064] According to the signal attenuation confirmation result, the device state information is collected from the Internet of Things device, and the state dataset is generated.

[0065] Preferably, the S7 further comprises:

[0066] By the state dataset, the decision tree algorithm is used to classify the applicability of the device state to the bird repelling mode, and the switching condition of the standby bird repelling mode is determined.

[0067] If the switching condition is met, the standby bird repelling mode is activated, the preset sound wave and laser intensity parameters are obtained, and the initial intensity value set is generated.

[0068] According to the initial intensity value set, combined with the weather change trend data, the sound wave or laser intensity value is recalculated to obtain an adaptive intensity adjustment set;

[0069] Through the adaptive intensity adjustment set, a control signal is sent to the bird repelling device to complete intensity reconfiguration.

[0070] A power transmission line intelligent inspection and bird repelling system is used to realize the steps of the power transmission line intelligent inspection and bird repelling method, and the system comprises:

[0071] A sensor network module collects bird activity signals and weather condition parameters from the foggy environment around the airport through a sensor network to obtain a bird gathering density and weather influence factors;

[0072] A data classification processing module classifies data by using a K-means clustering method combined with Euclidean distance measurement and cluster center initialization according to the bird gathering density and weather influence factors, and applies data standardization processing and iterative convergence conditions to determine a bird activity mode and weather change trend;

[0073] A path prediction module predicts a bird group movement path through a neural network model if the bird activity mode shows a gathering trend and the cluster quality is confirmed by contour coefficient evaluation to obtain a potential interference area;

[0074] A demand evaluation module obtains state information of an Internet of Things device for the potential interference area, judges a bird repelling intensity demand level, and activates a frequency adjustment process if the demand level exceeds a preset threshold;

[0075] A bird repelling frequency optimization module optimizes bird repelling frequency parameters by using a genetic algorithm defined by a chromosome coding mode and an adaptive function according to the bird repelling intensity demand level, sets a cross probability and a mutation rate to control the generation of a population size determined candidate parameter group, obtains an adjusted output scheme through a selection operation mechanism and a termination condition judgment, and retains an optimal individual as a foggy day exclusive configuration;

[0076] A bird repelling device control module sends a control instruction to the bird repelling device through the adjusted output scheme to determine a final sound wave or laser intensity value;

[0077] A backup bird repelling mode module switches to a backup bird repelling mode if the weather change trend indicates that propagation is blocked and signal attenuation is confirmed by anomaly point detection to obtain an enhanced adaptive response, wherein the backup bird repelling mode integrates state information of an Internet of Things device to recalculate an intensity value.

[0078] From the above technical solutions, the present application has the following beneficial effects:

[0079] The power transmission line intelligent inspection and bird repelling method and system aim at the potential threat of bird activity to aviation safety in foggy environment, fuse data analysis of bird gathering density and weather influencing factors, and solve the problem of how to accurately predict bird group moving path and optimize bird repelling strategy. The bird activity signal and weather parameter are collected through a sensor network, K-means clustering is combined with data standardization and dimension reduction processing, bird activity mode and weather change trend are identified. When high gathering trend is detected and cluster quality is verified by contour coefficient, the bird group path is predicted by using a neural network, and abnormal signal points are removed to improve accuracy. For the potential interference area, the bird repelling intensity demand is judged by using the state information of the Internet of Things device, if the threshold is exceeded, the genetic algorithm is used to optimize the sound wave or laser frequency parameter, and the fog special configuration is generated. If the signal is attenuated due to weather change, the standby mode is switched, and the intensity value is recalculated to ensure continuity. The present application realizes accurate monitoring and dynamic bird repelling through multi-technology fusion, effectively reduces the interference of bird activity in foggy weather to airport safety, and improves aviation operation efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0080] Figure 1 The method flowchart of the present application is shown in the figure.

[0081] Figure 2 The system connection diagram of the present application is shown in the figure. DETAILED DESCRIPTION

[0082] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0083] As shown in the figure, Figure 1 The present application provides a technical solution: a power transmission line intelligent inspection and bird repelling method, comprising:

[0084] S1, collecting bird activity signal and weather condition parameter from the foggy environment around the airport by a sensor network, obtaining bird gathering density and weather influencing factor;

[0085] S2, according to the bird gathering density and weather influencing factor, adopting K-means clustering method combined with Euclidean distance measurement and cluster center initialization to classify and process data, applying data standardization processing and iterative convergence condition to determine bird activity mode and weather change trend;

[0086] S3, if the bird activity mode shows gathering trend and the cluster quality is confirmed by contour coefficient evaluation, the bird group moving path is predicted by a neural network model, and the potential interference area is obtained;

[0087] S4, acquiring state information of the Internet of Things device for a potential interference area, judging a bird repelling intensity demand level, and activating a frequency adjustment process if the demand level exceeds a preset threshold;

[0088] S5, using a genetic algorithm optimized by a chromosome coding mode and a fitness function according to the bird repelling intensity demand level to optimize the bird repelling frequency parameters, setting a cross probability and a mutation rate to control the generation of a population size-determined candidate parameter group, obtaining an adjusted output scheme through a selection operation mechanism and a termination condition judgment, and retaining an optimal individual as a fog day exclusive configuration;

[0089] S6, sending a control instruction to the bird repelling device through the adjusted output scheme to determine a final sound wave or laser intensity value;

[0090] S7, if the weather change trend indicates that the signal propagation is blocked and the signal attenuation is confirmed through an anomaly point detection, switching to a backup bird repelling mode to obtain an enhanced adaptive response, wherein the backup bird repelling mode integrates the state information of the Internet of Things device to recalculate the intensity value.

[0091] The embodiment utilizes a sensor network to monitor the area where the power transmission line around the airport is located in real time, and collects bird activity signals and meteorological data under foggy conditions. First, through data preprocessing combined with the K-means clustering algorithm and the Euclidean distance measurement, the classification and identification of bird aggregation density and weather factors are realized, and a preliminary model of bird activity pattern and weather trend is formed. Subsequently, the contour coefficient is used to evaluate the clustering quality to ensure the reliability of the classification result. When the bird aggregation trend is detected to be obvious, the neural network model predicts the moving track of the bird flock according to the historical and real-time data, so as to lock the potential interference area. Then, the system calls the running state information of the Internet of Things device to evaluate the bird repelling demand intensity, and if the demand exceeds the threshold, the genetic algorithm is started to optimize the bird repelling frequency parameters. The genetic algorithm continuously iterates before the termination condition is met through chromosome coding, fitness function, cross and mutation operations, outputs the optimal parameter group, and generates a fog day exclusive configuration. The control system sends accurate control instructions of sound wave or laser intensity to the bird repelling device according to the parameter group to achieve efficient bird repelling. When it is monitored that the weather trend has an adverse effect on signal propagation and the signal attenuation is confirmed through anomaly detection, the system switches to the backup mode, re-integrates the state of the Internet of Things device and adjusts the intensity value to maintain the bird repelling effect.

[0092] The embodiment realizes intelligent inspection and precise bird repelling of the power transmission line under foggy conditions around the airport, effectively reducing the threat of bird groups to the power transmission line and flight safety. By fusing clustering analysis and neural network prediction, the bird repelling behavior can be targeted to cover the potential interference area, improving resource utilization. The genetic algorithm is used to optimize the bird repelling frequency parameters, significantly improving the propagation efficiency and effectiveness of the bird repelling signal, reducing energy consumption and prolonging the service life of the equipment. The introduction of the standby mode enhances the adaptability and stability of the system under complex weather conditions, ensuring the continuity and reliability of the bird repelling effect. The overall structure of the method is modular, with strong algorithm adaptability, high scalability and field application value.

[0093] S1 includes: collecting bird activity signals and weather parameters in the foggy environment around the airport through a sensor network, generating an original signal data set and a weather data set;

[0094] The original signal data set is denoised using signal processing technology to obtain a bird activity signal set;

[0095] According to the bird activity signal set, the bird aggregation density is calculated based on the signal strength, and a density distribution data is generated;

[0096] If the aggregation density in the density distribution data exceeds a preset threshold, the weather data set of the corresponding region is analyzed, and humidity data and visibility data are extracted;

[0097] The humidity data and the visibility data are fused by a weighted fusion algorithm to generate a weather influence factor;

[0098] According to the weather influence factor and the density distribution data, a decision tree algorithm is used to determine the correlation between bird activity and foggy environment, and a correlation analysis result is generated;

[0099] Through the correlation analysis result, a regression analysis algorithm is used to predict the change trend of the bird aggregation density, and a prediction data is generated.

[0100] First, data collection is performed. The sensor network synchronously collects bird activity signals and weather parameters in the foggy environment around the airport with a sampling interval of 1 second, and the length of a single monitoring period is 900 seconds. All bird activity signals obtained in sequence within the period are grouped into an original signal data set according to time, and humidity and visibility obtained in sequence within the same period are grouped into a weather data set.

[0101] Secondly, denoising is performed, and the denoising adopts a certain three-step method without changing the order. The first step is time window smoothing, the time window length is 5 seconds, the sliding step length is 1 second, and the arithmetic mean of all samples in the time window is used as the smoothing value of the time point and replaces the original value. The second step is fixed frequency band pass filtering, the band pass lower limit frequency is 300, the band pass upper limit frequency is 8000, and all components below the lower limit and above the upper limit are removed. The third step is energy deduction, the sequence is divided into continuous non-overlapping time windows according to the time window length of 5 seconds, and the energy value at the 20th percentile is selected as the noise reference of the window after sorting the sample energy from small to large in each time window. The energy of all samples in the window is subtracted from the noise reference, and the result less than zero is directly set to zero. The bird activity signal set is obtained after completion.

[0102] Then, the aggregation density calculation is performed, and the calculation adopts a fixed statistical time unit scheme, and the statistical time unit length is 10 seconds. In each statistical time unit, the arithmetic mean of all sample intensities in the bird activity signal set is calculated as the average intensity of the statistical time unit, and the average intensities of the continuous statistical time units are sequentially combined to form an intensity sequence. Subsequently, scale normalization processing is performed on the intensity sequence to eliminate the amplitude difference between different devices. The scale normalization processing method is to obtain the global minimum value and the global maximum value in the data of the last 30 days, subtract the global minimum value from each average intensity, divide by the difference between the global maximum value and the global minimum value, and truncate the result between 0 and 1. The truncated result is the aggregation density of the statistical time unit.

[0103] Then, threshold judgment is performed, and the preset threshold is determined by traversing the search of the labeled samples in the offline stage. Specifically, a sample set covering different seasons and different time periods is selected from historical fog monitoring, and the sample set includes real aggregation and real non-aggregation. The candidate threshold is taken between 0 and 1 with a step of 0.01, and for each candidate threshold, the comprehensive recognition accuracy is calculated when the aggregation density above the candidate threshold is judged as aggregation and the aggregation density below or equal to the candidate threshold is judged as non-aggregation. The comprehensive recognition accuracy is defined as the weighted average of the proportion of real aggregation correctly judged as aggregation and real non-aggregation correctly judged as non-aggregation, and the weight is determined according to the proportion of the two types of samples in the historical operation. The candidate threshold with the maximum comprehensive recognition accuracy is selected as the preset threshold and is fixed as the only threshold of the system. When the aggregation density of any statistical time unit is higher than the preset threshold in the online operation, it is determined that the region corresponding to the time unit exists aggregation and enters the weather extraction and fusion process.

[0104] In the weather extraction link, the time range completely consistent with the statistical time unit is located from the weather data set, and the humidity data and the visibility data are read. The humidity unit is percentage and is calibrated to zero and full scale by the weather equipment with standard samples at each power-on self-test, and the visibility unit is meter and is calibrated daily by the visibility meter with a standard reference plate. Then the scale unification processing is performed on the humidity and the visibility respectively, and the scale unification processing method is to perform linear scaling of the minimum value and the maximum value in the respective historical data range of the last 30 days, and the two types of data are respectively mapped between 0 and 1, and after completion, it enters the weighted fusion link.

[0105] The weighted fusion is used to generate the weather influence factor, and the weight is determined by grid search in the offline stage, the search range is 0 to 1, and the step value is 0.05. The humidity weight is taken as the candidate value, and the visibility weight is taken as 1 minus the humidity weight. For each weight set, the binary classification accuracy when determining whether it belongs to the high influence period of foggy day is calculated on the historical samples, and the high influence period is jointly annotated by the aviation oil station, the tower and the field inspection record. Select the weight combination with the highest binary classification accuracy as the fixed weight. In online operation, the standardized humidity is multiplied by the fixed humidity weight and added to the standardized visibility multiplied by the fixed visibility weight to obtain the weather influence factor, and the result is limited between 0 and 1.

[0106] Then the correlation determination is performed, and the decision tree algorithm is used to take the gathering density and the weather influence factor as two input features, and whether it belongs to the high activity related period of birds caused by foggy day as the output category. The decision tree is trained in the offline stage with the last 90 days of labeled data, the maximum depth is selected from 3, 5, and 7 by five-fold cross-validation to select the value with the highest verification accuracy, the minimum leaf sample number is fixed at 10, and the split criterion is fixed at information gain ratio. In online operation, the gathering density and the weather influence factor of the current statistical time unit are input into the trained decision tree to output the determination conclusion of correlation or not, as the correlation analysis result.

[0107] Finally, the trend prediction is performed, and the regression analysis algorithm is used to predict the gathering density of each statistical time unit within the next 30 minutes. The regression model selects one of the three candidate models in the offline stage, which are linear regression, quadratic regression and piecewise linear regression. The training and verification data are divided in chronological order, and the verification index is the mean absolute error. The model with the smallest mean absolute error is selected as the fixed model. In online operation, new data is used every 5 minutes to perform incremental update with a small step to adapt to slow seasonal changes without changing the determined model structure. The prediction output is the predicted value of the gathering density of each 10-second statistical time unit within the next 30 minutes, which together with the correlation analysis result forms the prediction data.

[0108] All parameters appearing in the above process and their determination methods are as follows: the sampling interval is 1 second, which is determined by the sensor hardware sampling capability and the time alignment mechanism to ensure that the timing alignment error is less than 1 second; the single monitoring period length is 900 seconds, which is determined by the airport scheduling window and the bird repelling response time limit to ensure that each batch of data is analyzed and output within 15 minutes; the time window length is 5 seconds and the sliding step length is 1 second, which are determined by the signal-to-noise ratio analysis of historical data to ensure that the smoothed phase delay is not more than 2 seconds and the amplitude loss is not more than 10 percent; the band-pass lower limit frequency is 300 and the band-pass upper limit frequency is 8000, which are determined by the joint test of the field background noise spectrum and the device sensitivity curve to remove wind noise and electromagnetic high-frequency noise stably; the noise reference is selected as the 20th percentile, which is determined by the combination minimization experiment of false positive rate and false negative rate to ensure that the false positive rate is less than 5 percent and the weak and small targets are not amplified; the statistical time unit length is 10 seconds, which is determined by the minimum response period of the bird repelling control device to keep the evaluation and control synchronized; the scale normalization processing adopts a sliding window of the last 30 days, which is determined by the seasonal and device drift characteristics to ensure numerical stability; the candidate step length of the preset threshold is 0.01, which is determined by the comprehensive trade-off of traversal time consumption and identification resolution; the fixed weights of humidity and visibility are determined by the maximum criterion of grid search binary classification accuracy, and once determined, they will not be changed in the online stage; the maximum depth and minimum leaf sample number of the decision tree and the splitting criterion are determined by the optimal results of cross-validation and are fixed; the regression model type is determined by the minimum criterion of mean absolute error and is fixed; the time interval of online incremental update is 5 minutes, which is determined by the system computing resources and real-time requirements, and the learning step length is a very small fixed value, which is determined by the stability test of preventing model drift.

[0109] S2 includes: collecting bird activity signals and weather parameters from the foggy environment around the airport through a sensor network to generate an original signal data set and a weather data set;

[0110] The original signal data set and the weather data set are preprocessed by using a data standardization technology to generate a standardized signal data set and a standardized weather data set;

[0111] The standardized signal data set is classified by using a K-means clustering algorithm, combined with Euclidean distance measurement and cluster center initialization, to generate a bird activity pattern classification result;

[0112] If the number of samples of a certain cluster in the bird activity pattern classification result exceeds a preset threshold, the standardized weather data set is subjected to dimensionality reduction processing to generate a reduced dimension weather data set;

[0113] According to the reduced dimension weather data set, main weather features are extracted by using a principal component analysis algorithm to obtain weather influence factors;

[0114] The correlation between the bird activity and the weather change is obtained by using a linear regression algorithm to analyze the correlation between the weather influencing factors and the bird activity pattern classification results.

[0115] According to the correlation model, the distribution of the bird activity pattern under different weather change trends is predicted, and prediction distribution data is generated.

[0116] In one possible implementation, data acquisition and data set generation are performed according to a determined process, and offline calibration and online verification are completed before running. The sensor network synchronously acquires bird activity signals and weather parameters in a foggy environment around the airport at a sampling interval of 1 second. The length of a single monitoring period is 900 seconds. The bird activity signals obtained in time sequence within the period form an original signal data set, and the humidity and visibility obtained in time sequence within the same period form a weather data set. The acquisition start and end time, equipment number, geographical coordinates, and time stamp are recorded and stored in a data record table. Missing data is processed according to a fixed interpolation rule. When the continuous missing time is not more than 5 seconds, the arithmetic mean of the adjacent two ends is filled. When the continuous missing time is more than 5 seconds, it is directly marked as invalid and the corresponding time slice is removed to ensure that subsequent calculation is based only on valid samples.

[0117] Data standardization processing is performed according to a determined step to generate a standardized signal data set and a standardized weather data set. First, the original signal data set is subjected to amplitude normalization to eliminate differences in sensitivity of different devices. The method is to obtain the global minimum value and the global maximum value within the statistical range of the historical signal of the last 30 days, map each sample intensity to 0 to 1 by linear scaling, and perform zero-one truncation to ensure numerical stability. Then, the humidity and visibility of the weather data set are processed respectively. The humidity is subjected to the same linear scaling as the minimum value and the maximum value of the last 30 days of history. The visibility is subjected to the same linear scaling as the minimum value and the maximum value of the last 30 days of history, and the abnormal high value of the visibility is subjected to percentile clipping. The clipping threshold is the 99th percentile to eliminate the influence of extreme values. The two standardized data sets form a standardized signal data set and a standardized weather data set respectively. The minimum value and the maximum value used for scaling are updated once a day and the last 60 days of versions are retained for tracing. If a new extreme value appears after the zero point of the day, the scaling boundary is updated and the update time is recorded when the first batch processing is performed on the day.

[0118] The classification of the standardized signal dataset adopts a K-means clustering algorithm combined with Euclidean distance measurement and a determined cluster center initialization strategy to complete the bird activity pattern classification. The specific steps are as follows: first, determine the value range of the cluster number as 2 to 6, and in the offline stage, fix the cluster number by the sample internal Euclidean distance square sum drop rate threshold rule. The drop rate threshold value is 5 percent. When the cluster number increases, if the relative drop amplitude of the total square sum is less than the threshold value for the first time, the last number of this time is taken as the fixed cluster number; in the online operation, the fixed number is used for clustering calculation, the cluster center initialization adopts an explicit sequential center selection strategy, the first center is the sample with the maximum Euclidean distance from the sample mean, and each subsequent center is the sample with the maximum minimum Euclidean distance from the nearest center in the selected center set, until the fixed number of centers is selected, and after the initialization is completed, the iteration is executed. Each iteration assigns each sample to the center with the minimum Euclidean distance, and then updates the center of each cluster to the arithmetic mean of all samples in the cluster. The iteration termination condition is that the maximum change of the center position is less than the preset convergence threshold value or the iteration number reaches the upper limit. The convergence threshold value is 0.001 of the standardized scale, and the iteration upper limit is 300 times. The output includes the cluster label of each sample, the center coordinates of each cluster, and the number of samples in each cluster. The above results are recorded as the bird activity pattern classification results and enter the threshold trigger judgment.

[0119] The threshold trigger judgment and weather data dimension reduction are executed according to the following determination process. First, check the sample number of each cluster in the bird activity pattern classification result. When the sample number of any cluster exceeds the preset threshold value, start the dimension reduction processing of the standardized weather dataset. The preset threshold value is determined in the offline stage according to the proportion of high activity period in the historical samples of the airport. Specifically, calculate the proportion of samples labeled as obvious gathering period by manual verification in the history and take the arithmetic mean of the proportion as the threshold value. The threshold value is between 0 and 1 and is expressed in percentage. In this embodiment, the threshold value is 30 percent. In the online operation, multiply the total number of samples in the current batch by the threshold value to get the sample number threshold value, and compare it with the sample number of each cluster. When the condition is met, perform the dimension reduction preprocessing on the standardized weather dataset. The preprocessing steps are as follows: first, align the time slices of the standardized weather dataset with the time slices of the bird activity pattern to ensure one-to-one correspondence. Then, for each time slice, retain humidity and visibility as the original weather feature input for dimension reduction algorithm, and retain the time index in the output to support subsequent association.

[0120] The main weather feature extraction and weather influence factor are obtained by adopting principal component analysis algorithm and determining the number of principal components and contribution threshold according to fixed rules. The calculation steps are as follows: input the aligned standardized weather data set into the principal component analysis process, first calculate the variance contribution of two features on the sample set, and then sequentially include the principal components according to the contribution from high to low, and the cumulative variance explanation rate threshold is fixed at 90%. When the cumulative explanation rate first reaches or exceeds the threshold, the number of principal components is determined and fixed. Then, the two standardized features of each time slice are linearly combined according to the determined principal components to obtain the reduced weather data set, and the score of the first principal component is linearly scaled to 0 to 1 as the weather influence factor. The scaling boundary uses the minimum and maximum values of the first principal component score of this batch and records them when output to ensure traceability. If the minimum and maximum values are equal, the weather influence factor is uniformly set to 0.5 to avoid numerical instability.

[0121] The establishment of the correlation analysis model adopts linear regression algorithm and takes the weather influence factor and the bird activity pattern classification result as the explicit corresponding relationship between input and output. Specifically, first, the bird activity pattern classification result of each time slice is converted into the proportion of each cluster. The method is to divide the number of samples belonging to a certain cluster in the time slice by the total number of samples in the time slice to obtain the cluster proportion, and the same calculation is performed for all clusters. Then, linear regression models are established for each cluster respectively. The weather influence factor is used as the independent variable input of the cluster proportion, and the cluster proportion is used as the dependent variable output. The training data is the time slice sample within the last 90 days confirmed by manual inspection. During training, the batch least error criterion is adopted and the fixed learning order is input piece by piece. The convergence criterion is that the average absolute error decreases by less than 1% within 10 consecutive iterations or the iteration times reach the upper limit. The iteration upper limit is set to 10,000 times. Finally, the linear regression coefficient and constant term corresponding to each cluster are obtained and solidified as the correlation model between bird activity and weather change. The model fitting quality is double-indexed by the average absolute error and the decision index of the validation set. The average absolute error threshold is set to 0.05, and the decision index threshold is set to 0.6. If either of them does not meet the standard, it will be rolled back to the last stable version to ensure the reliability of the model output.

[0122] The generation of the prediction distribution based on the correlation model is performed in explicit steps to output prediction distribution data that can be directly used for decision-making. First, a weather change trend sequence is generated, which is derived from the change direction of weather influencing factors in the time dimension in the reduced dimension weather data set. Three types of trends are defined: upward trend, downward trend, and stable trend. The upward trend increases by a fixed increment at each time slice, with the fixed increment being 0.5 times the standard deviation of the batch of weather influencing factors and the upper limit being truncated to 1. The downward trend decreases by the same fixed increment at each time slice, with the lower limit being truncated to 0. The stable trend remains unchanged at each time slice. The actual weather influencing factor at the current time slice is taken as the starting point for the three types of trends. Then the trend sequence is input into the correlation model piece by piece to obtain the prediction values of the cluster proportion of each cluster. Normalization processing is performed on all cluster proportions at the same time slice to eliminate numerical drift, with the sum being 1. Finally, the cluster proportions of each cluster under the three types of trends at each time slice are output together with the corresponding time stamp as prediction distribution data, along with the model version number used, the number of clusters, the threshold value, the number of principal components, the cumulative explanation rate threshold, the fixed increment value, and the truncation boundary for auditing and reproduction.

[0123] The sampling interval is 1 second, determined by the sensor hardware capability and the error evaluation of the time synchronization mechanism to ensure that the time alignment error does not exceed 1 second. The single monitoring period length is 900 seconds, determined by the control cycle of the airport operation and the bird repelling response requirement to ensure that each batch of data is processed within 15 minutes. The history window length is 30 days, determined by seasonal analysis and equipment drift analysis to ensure that the scaling boundary is representative and not excessively delayed. The visibility percentile clipping threshold is the 99th percentile, determined by extreme value analysis to avoid a small number of abnormal values dominating the results. The number of clusters is in the range of 2 to 6, determined by the upper limit of distinguishable activity patterns in actual applications and the upper limit of computing resources. The Euclidean distance is the only similarity measure, determined by algorithm requirements and implementation simplicity. The cluster center initialization uses a sequential center selection strategy, determined by convergence speed and repeatability evaluation. The convergence threshold is 0.001 and the iteration upper limit is 300 times, determined by historical convergence statistics to ensure stable results and not waste computing power. The sample number preset threshold is 30%, determined by the mean value of historical high activity proportion. The principal component cumulative explanation rate threshold is 90%, determined by the balance between information retention and dimension compression. The linear regression training iteration upper limit is 10,000 times and the convergence criterion is that the absolute error is less than 1% for 10 consecutive times, determined by stability testing. The average absolute error threshold in the validation phase is 0.05 and the decisive indicator threshold is 0.6, determined by error tolerance and business availability requirements. The weather trend fixed increment coefficient is 0.5 times the standard deviation, determined by sensitivity experiments to reflect trend changes without amplifying noise. The upper and lower boundaries are truncated to 0 and 1, determined by the value range definition.

[0124] S3 comprises: obtaining bird activity signals from the sensor network, generating an original signal dataset;

[0125] If the number of signal points in the original signal dataset exceeds a preset threshold, an outlier detection algorithm is used to remove isolated signal points, obtaining a filtered signal dataset;

[0126] According to the filtered signal dataset, a K-means clustering algorithm is used to generate a bird activity pattern, obtaining a clustering trend classification result;

[0127] If the contour coefficient of the clustering trend classification result exceeds a preset threshold, the cluster quality is confirmed, and a cluster quality confirmation result is generated;

[0128] According to the cluster quality confirmation result, a neural network model is used to train the filtered signal dataset, predict the bird group moving path, and obtain a predicted path dataset;

[0129] Through the predicted path dataset, combined with environmental interference factors, the path coverage range is analyzed, and the potential interference area is determined;

[0130] According to the potential interference area, an interference area boundary dataset is generated, and a region distribution result is obtained.

[0131] In one possible implementation, the generation of the original signal dataset is performed according to a fixed process, and all parameters are fixed to the system parameter library after offline calibration and online verification before going online. The sensor network continuously collects bird activity signals in a foggy environment around the airport with a sampling interval of 1 second. Each signal point contains time stamp, spatial position and intensity value. The duration of a single monitoring batch is 900 seconds. All signal points recorded in this period in chronological order form the original signal dataset. At the same time, the device number, collection start and end time and geographic reference coordinates are recorded for tracing. The data integrity is verified according to the determined rules. When multiple devices report repeatedly in any second, the arithmetic mean of the intensity values is taken as the only value for that second. When there is no report in any second and the continuous missing time does not exceed 5 seconds, the arithmetic mean of the adjacent seconds at both ends of the missing segment is linearly filled. When the continuous missing time exceeds 5 seconds, the corresponding time slice is directly removed, and the removal reason and time period boundary are recorded in the log to ensure that subsequent calculations are based on valid samples only.

[0132] The abnormal point detection and filtering is triggered and performed in a fixed calculation sequence according to a determined threshold. First, the total number of signal points in the original signal data set is counted and compared with a preset number threshold. If the total number exceeds the threshold, the abnormal point detection is performed. The determination method of the preset number threshold is to take the 95th percentile of the statistical results based on the historical non-aggregated period and to fix it before going online. In this embodiment, the threshold value is the specific value corresponding to the 95th percentile of the single batch signal point number distribution in the original data set. The abnormal point detection adopts a two-step parallel execution of the outlier elimination method based on neighborhood counting and the consistency check based on time neighbors, and is judged as abnormal if any condition is met. The spatial neighborhood counting method is to calculate the number of neighboring points within a certain radius for each signal point. The radius value is 30 meters, which is determined by the joint evaluation of the device positioning accuracy and the on-site channel width. The neighboring point number threshold value is 3, which is determined by the false positive rate and false negative rate balance experiment. When the number of neighboring points of a certain point within the radius is less than 3, it is determined as a spatial isolated point and marked for elimination. The time consistency check method is to count the number of repeated occurrences in the same small area within the time window of 2 seconds for each signal point. The small area side length value is 20 meters, which is determined by the raster resolution and the device coverage radius. When the number of repeated occurrences is less than 2 times, it is determined as a time isolated point and marked for elimination. After the two steps are completed, all the marked points are removed from the original signal data set, and the remaining samples form the filtered signal data set. At the same time, the elimination ratio and reason distribution are output for auditing and parameter review.

[0133] The generation of the aggregation trend classification result adopts the mean clustering algorithm and uses the geometric distance as the only similarity measure and uses the determined cluster center initialization strategy. Specifically, the fixed value of the number of clusters is determined in the offline stage. The candidate range is 2 to 6. The evaluation standard is the decreasing rate of the total intra-class geometric distance square sum with the increase of the number of clusters. When the decreasing rate is less than 5 percent for the first time after increasing by one step, the last cluster number is taken as the fixed cluster number and is fixed. It does not change dynamically after going online. In the online operation, the fixed cluster number is used to perform clustering on the filtered signal data set. The cluster center initialization adopts the farthest point first strategy. The first center is the sample point with the maximum distance from the geometric center of all samples. Each subsequent center is the sample point with the maximum value of the minimum distance from the nearest center in the selected center set. The iteration update is performed according to the standard steps. In each round, each sample is assigned to the center with the minimum geometric distance. Then, the cluster center is updated to the weighted arithmetic mean of all sample positions and intensities in the cluster, where the intensity is used as the weight to highlight the representative of high-intensity trajectory points. The iteration termination condition is that the maximum value of the position change of all centers after any round of update is less than 0.001 meters or the iteration number reaches 300. The earlier one is taken as the standard. After clustering, the cluster number of each sample, the center position of each cluster and the number of samples in the cluster are output, which is recorded as the aggregation trend classification result.

[0134] The cluster quality confirmation is performed according to the profile coefficient threshold method, and the threshold is determined by labeled data in the offline stage and is fixed before going online. Specifically, two quantities are calculated for each sample in the clustering trend classification result. One quantity is the average value of the geometric distance from the sample to all other samples in the same cluster, and the other quantity is the average value of the geometric distance from the sample to all samples in the nearest other cluster. The profile coefficient is obtained by subtracting the two values and dividing by the larger one of the two values, resulting in a value between -1 and 1. The calculation sequence is described in the embodiment, and the formula is not written in the text. The arithmetic average of the profile coefficients of all samples is taken as the overall cluster quality indicator, which is compared with the preset threshold. The preset threshold determination method is to traverse the candidate values from 0.3 to 0.8 with a step size of 0.05 in the historical sample set with artificial quality evaluation, calculate the proportion consistent with the artificial evaluation, select the candidate value with the highest consistent proportion as the threshold and fix it. The threshold value of the present embodiment is 0.5. When the overall cluster quality indicator is greater than or equal to 0.5, the cluster quality confirmation result is passed, otherwise the result is failed and the number of clusters, the iteration round and the sample balance degree in the cluster are recorded in the log for rollback processing.

[0135] The generation of the predicted path data set adopts a neural network model and uses the filtered signal data set as the only training data source. The input and output construction and the training process are performed according to fixed steps. First, the filtered signal data set is time-sliced with a sliding window length of 20 seconds and a step size of 1 second. In each slice, the data is aggregated into a time-ordered trajectory point sequence. Each trajectory point contains a position and an intensity. After all the values are normalized in the range of zero to one, they are used as the model input. The output is the predicted position and intensity of the next time point. The neural network structure is a three-layer feedforward structure. The number of hidden layer nodes is 64, 32 and 16 in sequence. The activation function is a rectified linear function to ensure stable training. The loss metric is the mean absolute error. The optimization process adopts batch training. The batch size is 256. The learning step size is 0.001. The maximum training round is 100. The early stopping rule is to stop and roll back to the best round weight when the mean absolute error on the validation set decreases by less than 1 percent in 10 consecutive rounds. The training data and the validation data are time-sliced according to the time sequence. The validation proportion is 20 percent. After the training is completed, the latest sliding window is used as the starting point in the inference stage to perform step-by-step rolling prediction. Each time, a new predicted point is added to the tail of the window and removed from the head of the window. The continuous extrapolation of 300 seconds obtains a predicted trajectory sequence in seconds. The connection of all the predicted points with time stamps constitutes the predicted path data set. The used model version, training time and training sample number are recorded to ensure traceability.

[0136] The determination of the potential interference area is based on the predicted path data set and combined with environmental interference factors, which are imported from existing airport data at system online and only read in this step. The six types of factors include terrain obstacle position and height, power transmission line corridor boundary, existing exclusion buffer width, dominant wind direction, and typical wind speed interval. No additional acquisition module is needed. The coverage analysis is completed in two steps. The first step is path buffer generation. The path buffer radius is expanded to a circular area along the predicted path, and the adjacent circles are connected to form a continuous band. The basic buffer radius is 50 meters, which is determined by the action radius of the bird repeller and the path positioning error. The second step is environmental weight correction. If the path band overlaps with the terrain obstacle, the buffer radius is expanded by 10 meters for every 10 meters of obstacle height in the overlapping section, but the total radius does not exceed 150 meters. If the path corridor is close to the power transmission line boundary, which is less than or equal to 30 meters, a directional expansion distance of 20 meters is added on the side close to the boundary to cover the line protection zone. If the angle between the dominant wind direction and the current wind speed interval and the path direction is less than or equal to 45 degrees, an expansion distance of 10 meters is added in the wind direction to reflect the propagation extension. The above expansions are added in a superimposed manner, and after each segment is calculated, the upper and lower boundaries are truncated to keep the total radius between 50 meters and 150 meters. The corrected path band of all time slices is fused and time-merged to obtain the spatial coverage range of the potential interference area.

[0137] The generation of the interference area boundary data set takes the potential interference area as input and outputs the regional distribution results according to fixed boundary extraction and simplification rules. Specifically, the merged coverage range is first analyzed for connectivity to separate non-adjacent areas into independent patches. Then, the boundary of each patch is extracted to generate an external boundary polyline. Subsequently, shape simplification is performed to remove high-frequency polyline details that do not affect the area and connectivity. The simplification tolerance is 5 meters, which is determined by the display accuracy and control instruction resolution. After simplification, the boundary polyline is closed to a polygon, and the polygon area, circumscribed rectangle, and geometric center are calculated as additional attributes. Finally, the polygon boundaries and additional attributes of all patches are indexed by time stamp to form the interference area boundary data set and output as the regional distribution results. Key parameter values and rule trigger records are also recorded to support review.

[0138] The sampling interval is 1 second, which is determined by the device sampling capability and the time alignment error requirement of not more than 1 second, the batch length is 900 seconds, which is determined by the airport operation control window, the upper limit of the missing filling is 5 seconds, which is determined by the sensitivity test of the prediction stability, the preset number threshold is the 95th percentile, which is determined by the non-aggregated distribution statistics, the spatial neighborhood radius is 30 meters, which is determined by the positioning accuracy and the sensor coverage radius, the number of adjacent points threshold is 3, which is determined by the minimum principle of false alarm rate and false alarm rate, the time window is 2 seconds, and the small area side length is 20 meters, which is determined by the sampling frequency and the spatial resolution, the clustering number candidate range is 2 to 6, and the falling rate threshold is 5%, which is determined by the offline elbow method experiment, the iteration convergence tolerance is 0.001 meters, and the iteration upper limit is 300 times, which is determined by the historical convergence statistics, the contour coefficient threshold is 0.5, which is determined by the result with the highest consistency with the artificial quality evaluation, the sliding window length is 20 seconds, and the step value is 1 second, which is determined by the motion smoothness and real-time trade-off, the hidden layer nodes are 64, 32 and 16, which are determined by the training speed and fitting ability, the batch size is 256, which is determined by the video memory capacity and gradient stability, the learning step value is 0.001, the maximum training round value is 100, and the early stopping rule is that the continuous 10 rounds of falling is less than 1%, which is determined by the overfitting prevention experiment, the extrapolation time length is 300 seconds, which is determined by the bird repelling decision preparation time, the basic buffer radius is 50 meters, and the maximum radius is 150 meters, which is determined by the device action radius and safety margin, the edge orientation expansion distance is 20 meters, and the wind expansion distance is 10 meters, which is determined by the propagation experiment, the obstacle expansion rule is that every 10 meters of height increases 10 meters, which is determined by the shielding diffraction experiment, and the simplified tolerance is 5 meters, which is determined by the map display and control accuracy.

[0139] S4 comprises: obtaining state information from the Internet of Things device, including device power and location coordinates, and generating a device state data set;

[0140] Through the device state data set, the support vector machine algorithm is used to classify the device power, to determine whether the power is lower than the preset threshold, to obtain a power device list with power lower than the preset threshold;

[0141] According to the power device list with power lower than the preset threshold and the location coordinates, the geometric distance between the device and the potential interference area is calculated, and a distance distribution data set is generated;

[0142] If the distance in the distance distribution data set is lower than the preset threshold, the K nearest neighbor algorithm is used to analyze the proximity of the device location and the interference area, and a strength demand data set is obtained;

[0143] Through the strength demand data set, it is judged whether the demand level exceeds the preset threshold, and a demand level classification result is generated;

[0144] According to the demand level classification result, a pre-established frequency adjustment rule is used to generate an adjustment instruction data set;

[0145] A device frequency adjustment signal is generated through the adjustment instruction data set, and a frequency adjustment execution result is obtained.

[0146] In a possible implementation, the generation of the device state data set is performed according to a fixed process, and all parameters are calibrated offline before being put into operation and written into a parameter library. The system pulls state information from Internet of Things devices at a sampling interval of 1 second, and the state information includes two mandatory fields of device power and location coordinates, and two traceable fields of device identification and time stamp. The duration of a single monitoring batch is 900 seconds, and the system collects 900 records for each device within the period and arranges them in chronological order to form the device state data set. When the same device repeatedly reports within the same second, the record with the latest time stamp is retained and the remaining records are discarded. When a device has a reporting gap within a range of no more than 5 seconds, the arithmetic mean of the power and location coordinates of the adjacent two seconds before and after the gap is used for linear filling. When the gap exceeds 5 seconds, all records of the device within the gap are directly excluded and marked as invalid to ensure that subsequent calculations are based on valid samples. The location coordinates are uniformly projected in meters and calibrated once when the device is connected to the network. The power is recorded in percentage and calibrated by the device at zero o'clock every day to eliminate drift.

[0147] The power classification and threshold determination are performed by using a support vector machine algorithm and in an explicit training and inference step. First, a training set is constructed in the offline phase. The training samples are derived from the time period in which the power and fault records in the historical operation data are consistent. The selection criteria for low-power samples are the time when the power drops to cause the device to continuously lose packets for 3 times or more and the data within 30 seconds before and after the time. The selection criteria for high-power samples are data in which the power is stable and no packet loss occurs within a 5-minute window. The training features include the current power, the power drop speed in the past 60 seconds, and the power fluctuation amplitude in the past 60 seconds. The labels are low power and non-low power. The model kernel function is compared among three candidates, namely, linear kernel, polynomial kernel, and radial basis kernel, in the offline phase. The comparison index is the average accuracy of five-fold cross-validation. When the average accuracy difference is less than 1 percentage point, the linear kernel with the lowest calculation cost is preferred. In this embodiment, the linear kernel is finally selected. The classification threshold is determined according to the principle of the maximum comprehensive accuracy on the validation set and is fixed as a preset threshold. In this embodiment, the threshold value is 20 percent. In online operation, the current power, the drop speed in the past 60 seconds, and the fluctuation amplitude are extracted from the device state data set second by second, and are input into the trained support vector machine. The output is a determination conclusion of low power or non-low power. The system adds the device determined as low power to the list of devices with power lower than the preset threshold, and records the device identification and determination time for subsequent linkage.

[0148] The generation of the distance distribution dataset takes the list of power equipment and the coordinates of the equipment positions as input and is executed according to a fixed geometric calculation process. The system reads the projection coordinates of each equipment in the same second from the list of power equipment obtained in the previous stage, and reads the boundary data of the potential interference area generated in the previous step, and calculates the geometric distance from each equipment to the nearest area boundary as the target quantity. The calculation rule is the minimum straight-line distance from a point to a polygon boundary. The straight-line distance is calculated in meters on the projection coordinate plane. If the equipment projection point is inside the polygon, the distance is defined as zero. The system records a distance value for each equipment in the current second and forms the distance distribution dataset, and calculates the minimum and maximum values of the distance and the quantiles for threshold comparison and statistical analysis. To avoid the influence of instantaneous jitter on the judgment, the system applies a time smoothing rule on the distance sequence. The rule takes the minimum value in a 5-second window as the representative distance, and the window step is 1 second. This rule is determined through a comparison test of false alarm rate and missed alarm rate before going online.

[0149] The proximity analysis and the generation of the intensity demand dataset are executed according to a fixed algorithm parameter with a clear threshold trigger. The system first compares the distance distribution dataset with the distance preset threshold, which is determined in the offline stage based on the safety buffer width and the equipment positioning error. In this embodiment, the value is 50 meters. When the representative distance of any equipment is less than or equal to 50 meters, the proximity analysis is entered. The proximity analysis uses the K-nearest neighbor algorithm and takes the equipment position and the representative point set on the boundary of the interference area as the object. The representative point set is obtained by sampling the boundary polyline at a fixed interval. The sampling interval is 10 meters. The distance measurement uses the straight-line distance consistent with the previous description. The value of K is 5, which is determined by the offline sensitivity experiment by weighing between the misjudgment rate and the calculation cost. The weight uses distance reciprocal weighting to emphasize the influence of the nearest representative point. The algorithm output is the proximity score of each triggered equipment. The calculation steps of the proximity score are to weight and summarize the distances of the equipment and its five nearest representative points and normalize them to between 0 and 1. The system combines the proximity score with the equipment power state to form the intensity demand indicator. The combination rule is to add a power penalty coefficient with a value of 1.2 to the proximity score when the equipment is judged to be low power to improve the demand sensitivity of the low-power neighbor equipment. When the equipment is not low power, the penalty coefficient is valued at 1.0 without adjustment. The system generates the intensity demand dataset for all triggered equipment at the same time, which contains four fields of equipment identification, proximity score, power state, and intensity demand indicator.

[0150] The demand level classification and threshold determination are performed according to the determined grading rules and output the demand level classification results. The system first classifies the intensity demand indicators in the intensity demand data set, and the grading threshold is fitted by the historical intervention effect and expert scoring corresponding relationship in the offline stage and solidified as a three-section threshold. In this embodiment, the value is that when the intensity demand indicator is greater than or equal to 0.7, it is determined as high demand, when the intensity demand indicator is greater than 0.4 and less than 0.7, it is determined as medium demand, and when the intensity demand indicator is less than or equal to 0.4, it is determined as low demand. The system then aggregates all device demand levels in the same second in the same potential interference region. If there is at least one high demand device or more than three medium demand devices in the region, it is determined that the demand level of the region exceeds the preset threshold. The preset threshold is the parallel condition of the above two rules and is solidified before going online. The system records the determination results of each region and the trigger device list as the demand level classification results.

[0151] The generation of the adjustment instruction data set adopts the pre-established frequency adjustment rules and determines the mapping with the demand level classification results as input. The frequency adjustment rules are calibrated by field experiments in the offline stage and solidified as table-driven form without introducing external modules. The rule content is the combined mapping of three-frequency strategy, two-duration strategy and one-cooling time strategy. In this embodiment, the three-frequency strategy takes the values of low frequency, medium frequency and high frequency, and the corresponding frequency values are 30%, 60% and 90% of the available frequency range of the device, respectively. The selection of medium frequency and high frequency is based on the demand level. High demand is mapped to high frequency, medium demand is mapped to medium frequency, and low demand is mapped to low frequency. The two-duration strategy takes the values of short time and long time, with short time taking 60 seconds and long time taking 180 seconds. The selection of duration is based on the number of trigger devices in the region. When the number of trigger devices is less than 3, take short time, and when it is greater than or equal to 3, take long time. The cooling time strategy is that after each execution, it must wait for at least 120 seconds before issuing a new instruction to the same region to prevent frequent switching. The system generates an adjustment instruction for each region determined to exceed the threshold according to the above mapping. The instruction includes target region identification, target device list, frequency bin, corresponding frequency value, duration, cooling time, and issue timestamp. All instructions constitute the adjustment instruction data set.

[0152] The generation of the device frequency adjustment signal and the output of the execution result are completed through a fixed delivery and confirmation process. The system parses the adjustment instruction data set by region and generates a frequency adjustment signal for each target device. The signal field includes device identification, target frequency value, execution start time, execution duration, and cooling time after execution. The system waits for an execution receipt in the device status channel after delivery. The receipt includes the actual effective time, current power, current frequency, and execution status code. The system compares the consistency of the instruction and the receipt after the receipt arrives and records the comparison result. When the comparison is consistent and completed within the planned time, the execution result of the device is marked as successful. When the receipt is overdue or inconsistent, it is marked as failed and the failure reason category and occurrence time are recorded in the log. The system summarizes the execution status of all devices in the region, calculates the success rate, and forms the frequency adjustment execution result. At the same time, the key parameters used this time and the threshold value and rule version number are written into the result record to ensure audit traceability.

[0153] The sampling interval is 1 second and the batch length is 900 seconds, which is determined by the control window and computing resources. The upper limit of the missing filling is 5 seconds, which is determined by the sensitivity test of trend stability. The position coordinates use meter-level projection, which is determined by the airport surveying and mapping reference. The power calibration is performed at zero o'clock every day, which is guaranteed by the device firmware. The support vector machine kernel function is selected by the principle of maximum average accuracy of five-fold cross-validation and prefers low-cost kernels. The classification threshold value is 20 percent, which is determined by the principle of maximum accuracy of the validation set. The distance preset threshold value is 50 meters, which is determined by the superposition of safety buffer and positioning error. The distance time smoothing window value is 5 seconds, which is determined by the false positive and false negative trade-off test. The representative point sampling interval is 10 meters, which is determined by the balance between boundary density and computing cost. The K value is 5, which is determined by the minimum time consumption experiment. The power penalty coefficient value is 1.2, which is determined by the improvement amplitude of the response demand of low-power devices in the field survey. The intensity demand classification threshold values are 0.7 and 0.4, which are determined by the fitting of historical intervention effects. The judgment rule for regions exceeding the threshold value is at least one high demand or at least three medium demands, which is determined by the field tolerance and device concurrency capability. The three frequency levels correspond to 30 percent, 60 percent, and 90 percent of the available frequency, which is determined by the device frequency response curve calibration. The duration short value is 60 seconds and the duration long value is 180 seconds, which is determined by the scene coverage demand. The cooling time value is 120 seconds, which is determined by the device thermal management and stability test.

[0154] S5 includes: obtaining state information of the Internet of Things device, including device power and position coordinates, and generating a device state data set;

[0155] Through the device state data set, a genetic algorithm is used to initialize a population, set a chromosome coding method, define a fitness function, and obtain an initial population parameter set;

[0156] According to the initial population parameter set, set the crossover probability and mutation rate, perform the crossover operation and mutation operation, and generate the alternative parameter set;

[0157] Through the alternative parameter set, adopt the selection operation mechanism, calculate the fitness value of each individual according to the fitness function, and obtain the preferred parameter set;

[0158] If the fitness value in the preferred parameter set meets the preset termination condition, the optimal individual is retained, and the fog-specific frequency configuration is generated;

[0159] According to the fog-specific frequency configuration, the distance distribution between the device and the target area is calculated in combination with the position coordinates, and the distance distribution data set is generated;

[0160] Through the distance distribution data set, the K-nearest neighbor algorithm is used to analyze the proximity of the device location and the target area, and the frequency adjustment execution result is obtained.

[0161] In one possible implementation, the generation of the device state data set is completed according to fixed collection and arrangement rules, and all parameters are fixed to the system parameter library after offline calibration and online verification before going online. The system continuously pulls state information from the Internet of Things device with a sampling interval of 1 second, and the single monitoring batch duration is 900 seconds. The state information includes two mandatory fields of device power and position coordinates, and two traceable fields of device identifier and timestamp. The position coordinates use the meter-level projection coordinate system and complete the coordinate zero offset correction through the reference point calibration when the device first enters the network. The power is recorded in percentage and calibrated to zero point and full scale by the device at zero o'clock every day. The collected data is de-duplicated and completed second by second. When multiple records of the same device are generated in the same second, only the record with the largest timestamp is retained. When any device is continuously missing for no more than 5 seconds, the arithmetic mean of the previous and next two seconds is filled in the power and position coordinates. When the continuous missing is more than 5 seconds, all records in the missing interval of the device are excluded, and the device identifier, time range and exclusion reason are written into the log. Finally, the device state data set is formed in ascending order of device and time, and the batch start time, end time and total number of devices are recorded for traceability.

[0162] The genetic algorithm initializes the initial population parameter set strictly according to a fixed process and does not introduce additional modules, the optimization target is to determine the optimal value of the device frequency parameter under foggy conditions to improve the bird repelling effect while controlling energy consumption, the chromosome coding mode uses real number coding and each chromosome only contains one gene for representing the target frequency value, the upper and lower bounds of the frequency value are read from the safe working range provided by the device manufacturer and are fixed before going online, the lower bound value in this embodiment is the value corresponding to the 10th percentile of the device allowed frequency range, and the upper bound value is the value corresponding to the 90th percentile of the device allowed frequency range to reserve a safety margin; the population size is determined by offline pressure testing by balancing the convergence speed and computing resources, the value in this embodiment is 50 individuals, the frequency value of each individual in the initial population is independently sampled according to a uniform distribution between the upper and lower bounds; the fitness function is realized by a weighted score driven by pure data and only uses the data sources allowed in the claim, the first part of the score is the effect score, the calculation method is to read the representative distance of each device to the target area at the same time and sort the distance from small to large, mark the devices in the smallest third as high-efficiency area devices and give full marks, give the devices in the middle third an average score, and give the devices in the largest third a low score, then take the arithmetic average of the device scores according to the number of devices to get the effect score; the second part of the score is the energy consumption score, the calculation method is to linearly scale the frequency value corresponding to the individual to between zero and one according to the upper and lower bounds, and then take the opposite number as the energy consumption score to express the fact that the higher the frequency, the greater the energy consumption, the effect score and the energy consumption score are combined into the fitness value according to fixed weights, the effect weight is 0.7 and the energy consumption weight is 0.3, this proportion is selected by statistical intervention effect of historical batches so that the effect index is maximized under the premise of achieving the same energy consumption.

[0163] The crossover and mutation operations are performed in a determined order and probability and are immediately corrected for out-of-bound values to maintain legality. The crossover probability is set to 0.8, determined by offline performance curve tests, and the mutation rate is set to 0.1, determined by diversity maintenance and convergence speed balancing tests. The crossover uses double-point crossover, with the specific steps being to determine two cutting positions on the frequency scale for two parent individuals and exchange the frequency segments between the two positions to generate two child individuals. When the frequency is a single gene, it degenerates to generating a child by proportional sampling between the two parent frequency values. After the crossover is completed, if the frequency of either child exceeds the upper or lower bound, it is truncated to the nearest boundary value. The mutation uses uniform mutation, with the specific steps being to independently sample each child to determine whether to mutate with the mutation rate. If mutation occurs, a new frequency is randomly sampled between the upper and lower bounds of the frequency to replace the original frequency, and an upper and lower bound check and truncation are performed on the result. After all the crossover and mutation are completed, all new individuals are collected to form a candidate parameter set. At the same time, the top individuals in the previous generation are directly copied to the candidate parameter set without modification to prevent the optimal value from being destroyed by random operations. The number of elite individuals is set to 2, determined by the results of experiments to maintain the optimal value and avoid excessive convergence.

[0164] The selection operation mechanism uses fixed-size tournament selection and uses the numerical value of the fitness function as the only criterion to generate the preferred parameter set. The tournament size is set to 3, determined by the trade-off between selection pressure and diversity. The execution steps are to randomly select 3 individuals from the candidate parameter set, compare their fitness values, and select the maximum value individual to join the new generation population. Repeat the extraction until the new generation population reaches the specified size. During the entire selection process, record the fitness and source of the selected individual for each comparison to facilitate tracing. At the same time, calculate the average fitness and optimal fitness of the new generation population to determine the termination condition. When a generation of preferred parameter sets is generated, immediately recalculate the fitness of each individual in the set to eliminate numerical errors and store it in the preferred parameter set for the next step of judgment and output.

[0165] The termination condition and the generation of the fog-specific frequency configuration follow fixed rules and do not introduce subjective judgment. The termination condition consists of two parallel constraints, and either one satisfies the condition to stop iteration. The first constraint is that the optimal fitness improvement amplitude of the last 20 generations is less than 1 percent. The second constraint is that the total number of iterations reaches 200 generations. The values of the two constraints are determined by offline stability tests to avoid overfitting and premature stopping. When the termination condition is met, locate the individual with the maximum fitness value in the current preferred parameter set as the optimal individual and extract its frequency value as the fog-specific frequency configuration. At the same time, record the key statistics for generating the configuration, including the final generation, optimal fitness value, average fitness value, elite individual retention number, crossover probability, and mutation rate. Write the configuration record together with the batch time and device list to the configuration record for review and reproduction.

[0166] The calculation and generation of the distance distribution dataset is completed in a determined geometric step with the fog-specific frequency configuration and the device position coordinates as inputs. The system first reads the projection coordinates of each device from the device state dataset and reads the boundary polyline from the target area boundary obtained in the previous step. The shortest straight-line distance from each device to the nearest boundary of the target area is taken as the representative distance of the device. If the projection point of a device is located inside the polygon of the target area, the representative distance is defined as zero. The representative distances of all devices are written into the distance distribution dataset together with the device identification and the corresponding fog-specific frequency configuration for use in proximity analysis. To reduce false positives caused by instantaneous position jitter, the system applies a determined time smoothing rule on the representative distance time series. The rule takes the minimum value within a 5-second sliding window as the final representative distance for that second, with a window step value of 1 second. This rule is determined through comparison tests on false positive rate and false negative rate in the offline stage.

[0167] The proximity degree analysis and the generation of the execution result of frequency adjustment use a fixed-parameter nearest neighbor method and take the distance distribution dataset as the only input. A set of representative points on the boundary of the target area is sampled at a fixed interval to obtain the boundary polyline. The sampling interval is determined by precision and computational load balance tests and has a value of 10 meters. The number of nearest neighbors is determined by false judgment and time consumption trade-off and has a value of 5. The distance metric is the straight-line distance on the projection plane and is measured in meters. The weight is the reciprocal of the distance to enhance the influence of the nearest neighbor points. The system calculates the weighted distance average of each device with its nearest 5 representative points and obtains the proximity score by scaling the average value to zero to one based on the maximum and minimum weighted distances of all devices in the batch. The closer the proximity score is to one, the closer the device is to the target area and the more consistent it is with the boundary distribution. The system delivers the proximity score and the corresponding fog-specific frequency configuration as execution parameters. The execution parameters include the target frequency value, the execution start time, the execution duration, and the execution area identification. The execution duration in this embodiment is determined by coverage stability tests and has a value of 180 seconds. The minimum cooling time between two consecutive deliveries for the same device is determined by device thermal management and stability tests and has a value of 120 seconds. After the execution is completed, the system collects the device receipts, which include the actual effective time, the actual frequency, the execution duration, the current power, and the execution result code. The system compares the consistency of the receipts and the instructions one by one and calculates the success rate by summarizing the number of successful devices and the total number of devices in the region to generate the frequency adjustment execution result. At the same time, the system records the nearest neighbor parameters and the distance smoothing parameters used this time, as well as the frequency configuration version number and the batch identification, to ensure that the audit is traceable.

[0168] All the parameters and threshold values involved are as follows and are fixed before going online and are updated at 0 o'clock every day in a 30-day sliding window for items dependent on historical statistics, with a sampling interval of 1 second and a batch length of 900 seconds determined by the airport operation window and system computing power, the upper limit of missing data completion is 5 seconds determined by the trend maintenance test, the upper and lower limits of frequency are 10 and 90 percentiles of the device available frequency range determined by the manufacturer's specifications and safety margin, the population size is 50 determined by the convergence speed and computing resource balance test, the effect and energy consumption weights are 0.7 and 0.3 determined by the historical intervention effect statistics fitting, the crossover probability is 0.8 and the mutation rate is 0.1 determined by the performance curve test, the number of elite individuals is 2 determined by the need to prevent optimal loss, the tournament size is 3 determined by the selection pressure and population diversity compromise, the continuous generation threshold in the termination condition is 1 percent, the continuous generation number is 20, and the maximum generation number is 200 determined by the stability test, the distance smoothing window is 5 seconds determined by the false positive and false negative trade-off, the boundary sampling interval is 10 meters and the number of neighbors is 5 determined by the accuracy and computing overhead, the execution duration is 180 seconds and the cooling time is 120 seconds determined by the scene coverage requirement and device thermal management constraints.

[0169] S6 includes: generating bird flock movement trend by using trajectory analysis algorithm through bird flock movement path data, determining dynamic distribution of potential interference area;

[0170] According to the dynamic distribution, the center coordinates and boundary range of the potential interference area are calculated to obtain the area coverage range;

[0171] Through the area coverage range, the relative distance between the device and the interference area is calculated by combining the position coordinates of the bird repelling device to generate distance distribution data set;

[0172] If the relative distance in the distance distribution data set is less than the preset threshold value, the K-nearest neighbor algorithm is used to analyze the proximity between the device and the interference area to obtain a weight set;

[0173] Through the weight set, the generation rule of the control instruction set is adjusted to determine the distribution proportion of the sound wave intensity value and the laser intensity value;

[0174] According to the distribution proportion, the control instruction set for covering the potential interference area is generated to determine the final sound wave or laser intensity value;

[0175] Through the control instruction set, an execution signal is sent to the bird repelling device to complete the intensity adjustment.

[0176] In one possible implementation, trajectory analysis first takes the flock movement path data as input, which consists of the predicted paths output by the previous step in chronological order, each time point containing planar coordinates in meters and time information in seconds. The system time-aligns the trajectory with a sampling interval of 1 second, and performs smoothing on the coordinate sequence using a moving average with a fixed time window length of 5 seconds and a sliding step of 1 second to reduce transient jitter. Then, the displacement between adjacent time points is divided by the sampling interval to obtain the second-by-second speed and direction, and the weighted average of the speed and direction is obtained within a fixed window of the last 20 seconds, with higher weights for data closer to the current time. The weight sequence is determined by the historical error minimization principle in the offline stage and is fixed. The threshold for movement significance is that the trend speed is not less than 0.5 meters per second and is continuously maintained for 10 seconds. The threshold value is determined by the balance test of prediction bias and response delay in the field test.

[0177] When the significance condition is met, the current trend direction is taken as the main axis, and the trend path is extrapolated every second within the planning time domain of the future 300 seconds, and an equal-width buffer zone is generated on both sides of the trend path with a base buffer radius of 50 meters, forming a dynamic distribution. If adjacent buffer zones are adjacent in time and have overlap in space, a union operation is performed and the earliest start and latest end time window is retained as the effective time range of the area. The value of the buffer radius is determined by the joint evaluation of the bird repelling device action radius and the positioning error, and is limited to a maximum of 150 meters to ensure a safety margin.

[0178] The coverage range is extracted from the dynamic distribution piece by piece, and the center coordinates of the coverage range are calculated according to the geometric center of the boundary polygon of the piece, and the boundary range is determined by the minimum and maximum horizontal coordinates and the minimum and maximum vertical coordinates, and is recorded in meters. At the same time, the area and perimeter of the piece are calculated as additional attributes for robustness verification in subsequent weight allocation.

[0179] The fixed position coordinates of the bird repelling device are taken as input, and the relative distance between the device and the coverage range is calculated, which is defined as the shortest straight line distance from the device projection point to the boundary of the piece, and if the device is located inside the piece, the distance is zero. The system outputs a distance record for each device in each effective time slice, forming a distance distribution data set, and takes the minimum distance as the final representative distance in a 5-second sliding window. The window length is determined by the false alarm and missed alarm trade-off test. When the representative distance is less than or equal to the distance preset threshold, the adjacent analysis is triggered, and the distance preset threshold is 90 percent of the effective action radius of the device, and is fixed as a definite value (unit: meters) before going online through multiple batches of field tests to ensure early intervention before entering the action boundary.

[0180] The proximity analysis adopts a fixed-parameter proximity method. First, a set of representative points is generated on the boundary polyline of the area at a fixed interval of 10 meters (determined by the precision and computational overhead trade-off). Then, the straight-line distance between the trigger device and all representative points is calculated, and the five smallest distances are selected as the nearest neighbors (the number of nearest neighbors is 5, determined by the false positive rate and time consumption trade-off). The reciprocals of the five distances are then linearly scaled so that their sum is 1, obtaining a weight set. The components of the weight set are strictly non-negative, the sum is equal to 1, and they monotonically decrease from near to far.

[0181] The generation rules of the control instruction set are adjusted according to the weight set to determine the allocation ratio of sound waves and lasers. When the area of the area is greater than or equal to the area threshold and the sum of the two heaviest weights in the weight set is less than or equal to 0.6, it is determined that the area needs coverage-type interference, the sound wave ratio is set to 0.7, and the laser ratio is set to 0.3. When the area of the area is less than the area threshold or the sum of the two heaviest weights in the weight set is greater than 0.6, it is determined that the area needs point-type interference, the sound wave ratio is set to 0.3, and the laser ratio is set to 0.7. The area threshold is the median of the historical area distribution and is fixed as a specific value. If there are multiple adjacent areas in the same batch and the effective time range overlaps, they are combined according to the weighted center distance of the weight set, and when the distance is less than or equal to 100 meters, they are considered as a single area and the area and boundary are recalculated before determining the ratio.

[0182] After the ratio is determined, the system maps the ratio to the intensity level available to the device. The allowed range of the device's sound waves and lasers is written into the parameter library when it is connected to the network, including two integers, the minimum level and the maximum level. The mapping rule is to multiply the corresponding maximum level by the ratio and round to get the target level, while not lower than the minimum level and not higher than the maximum level. If multiple devices in the area meet the trigger condition, fine-tune them according to the heaviest weight of their respective weight sets in the same area: if the heaviest weight is greater than or equal to 0.5, increase it by one level without exceeding the maximum level, and if the heaviest weight is less than 0.2, decrease it by one level without falling below the minimum level, to concentrate the interference resources.

[0183] Generate control instruction set, each instruction writes target device identifier, sound wave intensity level, laser intensity level, action start time, action duration, target area identifier. The action duration is 180 seconds (determined by the coverage stability test), and the cooling time between two consecutive instructions to the same device is 120 seconds (determined by the device thermal management and stability test). After the instruction is issued, the system waits for the device's feedback, which includes the actual effective time, the actual sound wave level, the actual laser level, the execution time, the current power, and the execution status code. The system compares the feedback and the instruction for each device, allowing a deviation of no more than 5 seconds in the effective time, and the levels are consistent and within the range, which is considered successful, otherwise it is considered failed and the failure reason is recorded.

[0184] The sampling interval is 1 second, the smoothing time window is 5 seconds, the sliding step is 1 second, which is determined by the track noise characteristics and real-time requirements; the moving significance threshold speed lower limit is 0.5 meters per second, and the duration is 10 seconds, which is determined by the path stability requirement of field measurement; the planning time domain is 300 seconds, which is determined by the bird driving scheduling and execution preparation time length; the buffer radius is 50 meters and the maximum is 150 meters, which is determined by the device action radius and safety margin; the center coordinates are the geometric center; the distance sliding window is 5 seconds, which is determined by the false alarm and false alarm trade-off; the distance preset threshold is 90% of the action radius, which is determined by the principle of early intervention; the boundary sampling interval is 10 meters, and the number of neighbors is 5, which is determined by the precision and calculation overhead test; the area threshold is the median of the historical area distribution, which is determined by the experience rule; the sound wave and laser ratio is 0.7 and 0.3 in the coverage type, and 0.3 and 0.7 in the fixed-point type, which is determined by the fitting of historical intervention effect; the minimum and maximum levels are limited and frozen by the manufacturer; the action duration is 180 seconds and the cooling time is 120 seconds, which are determined by the coverage stability and equipment thermal management experiment; the effective time deviation is 5 seconds, which is determined by the control link time delay statistics.

[0185] S7 includes: obtaining weather change trend data through a meteorological sensor, analyzing the attenuation effect of rainfall or fog on signal propagation, and determining the signal propagation blocked state;

[0186] If the signal propagation blocked state is confirmed, an outlier detection algorithm is used to analyze the time series data of signal strength, and a signal attenuation confirmation result is obtained;

[0187] According to the signal attenuation confirmation result, device state information is collected from the Internet of Things device to generate a state data set;

[0188] Through the state data set, a decision tree algorithm is used to classify the applicability of the device state to the bird driving mode, and the switching condition of the standby bird driving mode is determined;

[0189] If the switching condition is met, the standby bird driving mode is activated, the preset sound wave and laser intensity parameters are obtained, and an initial intensity value set is generated;

[0190] According to the initial intensity value set, combined with the weather change trend data, the sound wave or laser intensity value is recalculated to obtain an adaptive intensity adjustment set;

[0191] Through the adaptive intensity adjustment set, a control signal is sent to the bird driving device to complete the intensity reconfiguration.

[0192] In one possible implementation, the weather change trend data is continuously acquired by the weather sensor at a sampling interval of 1 second, and the rainfall intensity (in millimeters per hour), visibility (in meters), timestamp, and measurement point coordinates are recorded every second, and quality inspection and missing data completion are performed on the data in each 30-second time window. When the missing data is continuous for no more than 5 seconds, the arithmetic mean interpolation of the front and rear endpoints is used; when the missing data is continuous for more than 5 seconds, the time period is marked as invalid and excluded. Then the rainfall intensity in each time window is linearly scaled to 0 to 1 according to the minimum and maximum values of the last 30 days of history, and the visibility is also linearly scaled and then the inverse index (1 minus the scaled value) is taken to reflect the fog concentration change. The system sums the rainfall and fog attenuation values weighted by fixed weights to obtain a comprehensive attenuation coefficient, the rainfall weight is 0.4, and the fog weight is 0.6. The weight combination is determined by offline grid search and minimum prediction error. If the average of the comprehensive attenuation coefficient of any 30-second time window is greater than or equal to 0.7, it is determined that the signal propagation is blocked, and the attenuation confirmation link is entered, otherwise the current round of determination is directly ended.

[0193] When the signal propagation is blocked, the system calls the abnormal point detection algorithm to process the signal strength time sequence reported by the bird repelling device. The sampling interval is 1 second, and the detection window length is 30 seconds. The system calculates the arithmetic mean and standard deviation in the window, and marks the samples below the arithmetic mean minus 3 times the standard deviation as attenuation abnormal points. Then the proportion of abnormal points is counted, if the proportion is greater than 0.4, the signal attenuation confirmation result is true, otherwise it is false and the current round of process is terminated.

[0194] When the signal attenuation confirmation result is true, the system batch pulls all related bird repelling device state information from the Internet of Things side, and generates a state data set. The state fields include: device power (percentage), location coordinates (meter-level projection coordinates), current running mode (sound wave, laser or combination), and the latest self-check time (seconds). If the same device reports repeatedly in the same second, the timestamp with the maximum time is retained; intervals with continuous missing reports for no more than 5 seconds are completed by interpolation, and intervals with continuous missing reports for more than 5 seconds are excluded.

[0195] The state data set is input into the decision tree algorithm for applicability classification. The features include device power, running mode, self-check overdue flag (if the latest self-check is more than 7 days, it is yes, otherwise it is no), and the minimum relative distance to the potential interference area (5-second sliding window takes the minimum value). The maximum depth of the decision tree is 5, the minimum leaf sample number is 10, and the split criterion is information gain ratio. These parameters are selected offline by cross-validation to obtain the optimal parameters. The switching condition is: the comprehensive attenuation coefficient is greater than or equal to 0.7, the signal attenuation confirmation is true, and the classification result is "applicable to standby mode". When the three conditions are met at the same time, the standby bird repelling mode is triggered; otherwise, the current mode is maintained.

[0196] After the standby mode is triggered, the system reads the preset sound wave intensity level and laser intensity level of the mode from the parameter library to generate an initial intensity value set. The intensity level is an integer, ranging from the minimum level 1 to the maximum level 10, which is determined and frozen by the equipment manufacturer's safety specification and on-site verification.

[0197] The initial intensity is proportionally adjusted using the comprehensive attenuation coefficient of the latest 30-second window. When the comprehensive attenuation coefficient is greater than or equal to 0.7, the sound wave intensity is increased by 20% and the laser intensity is decreased by 10%; when the coefficient is between 0.5 and 0.7, both remain unchanged; when the coefficient is less than 0.5, the sound wave intensity is decreased by 10% and the laser intensity is increased by 20%. The adjustment result is rounded and truncated within the range of 1 to 10. The percentage adjustment value is determined and fixed by the off-line double-target optimization test (energy consumption and interference efficiency).

[0198] A control signal is generated for each target device, including device identification, sound wave intensity level, laser intensity level, execution start time, execution duration (fixed at 180 seconds, determined by coverage maintenance test), and mode identification. The cooling time of the same device issued twice in succession is fixed at 120 seconds, determined by the thermal management test. The control signal is issued through the Internet of Things link, and the device returns a receipt containing the actual effective time, the actual output intensity, the execution status code, and the current power. If the effective time deviation is not more than 5 seconds and the actual output intensity is consistent with the issued one, it is marked as successful, otherwise it is recorded as a failure and classified as link delay, device rejection, or insufficient energy.

[0199] The historical minimum and maximum values of rainfall and visibility are updated at zero o'clock every day using a 30-day sliding window; the rainfall and fog weight is 0.4 and 0.6, respectively, determined by grid search minimum prediction error; the comprehensive attenuation coefficient threshold is 0.7, determined by on-site test when the false positive rate is less than 5% and the false negative rate is less than 10%; the sampling interval is 1 second and the detection window is 30 seconds, determined by balancing detection sensitivity and timeliness; the abnormal point is determined as the mean value minus 3 times the standard deviation, verified by historical noise distribution; the abnormal proportion threshold is 0.4, determined by false positive rate control target; the power threshold is 30%, determined by low power stability test; the self-check overdue threshold is 7 days, determined by manufacturer's recommendation; the decision tree parameters are determined by cross-validation; the intensity level range is limited by device specifications; the intensity adjustment proportion (±10%, ±20%) is determined by double-target optimization test; the execution duration is 180 seconds and the cooling time is 120 seconds, determined by coverage maintenance and thermal management test; the allowed effective time deviation is 5 seconds, determined by link delay statistics.

[0200] Although embodiments of the present application have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, alternatives, and variations can be made thereto without departing from the principles and spirit of the application, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent inspection and bird repelling of power transmission lines, characterized in that, The method comprises the following steps: S1, collecting bird activity signals and weather condition parameters from the foggy environment around the airport through a sensor network to obtain bird gathering density and weather influence factors; S2, classifying the data by using the K-means clustering method combined with Euclidean distance measurement and cluster center initialization according to the bird gathering density and the weather influence factors, applying data standardization processing and iterative convergence conditions to determine the bird activity mode and the weather change trend; S3, if the bird activity mode shows a gathering trend and the cluster quality is confirmed by the contour coefficient evaluation, predicting the bird flock moving path through a neural network model to obtain a potential interference area; S4, obtaining the state information of the Internet of Things equipment for the potential interference area, judging the bird repelling intensity demand level, and activating the frequency adjustment process if the demand level exceeds the preset threshold; S5, optimizing the bird repelling frequency parameters by using the genetic algorithm defined by the chromosome coding method and the fitness function according to the bird repelling intensity demand level, setting the crossover probability and mutation rate to control the generation of the selected parameter group of the population size, and obtaining the adjusted output scheme through the selection operation mechanism and the termination condition judgment, while retaining the optimal individual as the fog-specific configuration; S6, sending control instructions to the bird repelling device through the adjusted output scheme to determine the final sound wave or laser intensity value; S7, if the weather change trend indicates that the propagation is blocked and the signal attenuation is confirmed by the anomaly point detection, switching to a backup bird repelling mode to obtain an enhanced adaptive response, wherein the backup bird repelling mode integrates the state information of the Internet of Things equipment to recalculate the intensity value.

2. The method of claim 1, wherein the method further comprises: The S1 comprises: Collecting bird activity signals and weather parameters in the foggy environment around the airport through a sensor network to generate an original signal data set and a weather data set; Using signal processing technology to denoise the original signal data set to obtain a bird activity signal set; According to the bird activity signal set, calculating the bird gathering density based on the signal intensity to generate density distribution data; If the gathering density in the density distribution data exceeds the preset threshold, analyzing the weather data set of the corresponding area to extract humidity data and visibility data; Fusing the humidity data and the visibility data by a weighted fusion algorithm to generate weather influence factors; According to the weather influence factors and the density distribution data, using a decision tree algorithm to judge the correlation between bird activity and foggy environment to generate correlation analysis results; Using a regression analysis algorithm to predict the bird gathering density change trend through the correlation analysis results to generate prediction data.

3. The method of claim 1, wherein the method further comprises: The S2 comprises: Collecting bird activity signals and weather parameters from the foggy environment around the airport through a sensor network to generate an original signal data set and a weather data set; Using data standardization technology to preprocess the original signal data set and the weather data set to generate a standardized signal data set and a standardized weather data set; Classifying the standardized signal data set by using the K-means clustering algorithm, combining Euclidean distance measurement and cluster center initialization to generate bird activity mode classification results; If the number of samples of a certain cluster in the bird activity mode classification results exceeds the preset threshold, performing dimensionality reduction processing on the standardized weather data set to generate a reduced dimension weather data set; According to the dimensionality reduction weather dataset, the principal component analysis algorithm is adopted to extract the main weather characteristics, and a weather influence factor is obtained; Through the weather influence factor and the bird activity pattern classification result, the linear regression algorithm is adopted to analyze the correlation between the two, and a correlation model of bird activity and weather change is obtained; According to the correlation model, the distribution of the bird activity pattern under different weather change trends is predicted, and a prediction distribution data is generated.

4. The method of claim 1, wherein the method further comprises: The S3 includes: Obtain bird activity signals from the sensor network to generate an original signal dataset; If the number of signal points in the original signal dataset exceeds a preset threshold, adopt an outlier detection algorithm to remove isolated signal points to obtain a filtered signal dataset; According to the filtered signal dataset, adopt a K-means clustering algorithm to generate a bird activity pattern, and obtain a clustering trend classification result; If the contour coefficient of the clustering trend classification result exceeds a preset threshold, confirm the cluster quality, and generate a cluster quality confirmation result; According to the cluster quality confirmation result, adopt a neural network model to train the filtered signal dataset to predict the bird flock moving path, and obtain a prediction path dataset; Through the prediction path dataset, combined with environmental interference factors, analyze the path coverage range, and determine the potential interference area; According to the potential interference area, generate an interference area boundary dataset, and obtain a region distribution result.

5. The method of claim 1, wherein: The S4 includes: Obtain state information from the Internet of Things device, including device power and location coordinates, and generate a device state dataset; Through the device state dataset, adopt a support vector machine algorithm to classify the device power, judge whether the power is lower than the preset threshold, and obtain a power device list whose power is lower than the preset threshold; According to the power device list whose power is lower than the preset threshold and the location coordinates, calculate the geometric distance between the device and the potential interference area, and generate a distance distribution dataset; If the distance in the distance distribution dataset is lower than a preset threshold, adopt a K-nearest neighbor algorithm to analyze the proximity of the device location and the interference area, and obtain a strength demand dataset; Through the strength demand dataset, judge whether the demand level exceeds a preset threshold, and generate a demand level classification result; According to the demand level classification result, adopt a pre-established frequency adjustment rule to generate an adjustment instruction dataset; Through the adjustment instruction dataset, generate a device frequency adjustment signal, and obtain a frequency adjustment execution result.

6. The method of claim 1, wherein: The S5 includes: Obtain state information of the Internet of Things device, including device power and location coordinates, and generate a device state dataset; Through the device state dataset, adopt a genetic algorithm to initialize a population, set a chromosome coding mode, and define a fitness function, and obtain an initial population parameter set; According to the initial population parameter set, set a crossover probability and a mutation rate, perform a crossover operation and a mutation operation, and generate a candidate parameter group; Through the candidate parameter group, adopt a selection operation mechanism, calculate the fitness value of each individual according to the fitness function, and obtain an optimal parameter set; If the fitness value in the optimal parameter set meets a preset termination condition, retain the optimal individual, and generate a fog day exclusive frequency configuration; According to the fog day exclusive frequency configuration, combined with the location coordinates, calculate the distance distribution between the device and the target area, and generate a distance distribution dataset; The K-Nearest Neighbor algorithm is used to analyze the proximity between the device location and the target area based on the distance distribution dataset, and the frequency adjustment execution result is obtained.

7. The method of claim 1, wherein the method further comprises: The S6 includes: The trajectory analysis algorithm is used to generate bird flock movement trends based on bird flock movement path data, and the dynamic distribution of potential interference areas is determined. According to the dynamic distribution, the center coordinates and boundary range of the potential interference area are calculated to obtain the area coverage range. By combining the position coordinates of the bird repelling device with the relative distance between the device and the interference area, a distance distribution dataset is generated. If the relative distance in the distance distribution dataset is less than the preset threshold, the K-Nearest Neighbor algorithm is used to analyze the proximity between the device and the interference area, and a weight set is obtained. The generation rule of the control instruction set is adjusted based on the weight set to determine the allocation ratio of the sound wave intensity value and the laser intensity value. According to the allocation ratio, a control instruction set that covers the potential interference area is generated, and the final sound wave or laser intensity value is determined. The control instruction set is sent to the bird repelling device to complete the intensity adjustment.

8. The method of claim 1, wherein: The S7 includes: The weather change trend data is obtained through the meteorological sensor, the attenuation effect of rainfall or fog on signal propagation is analyzed, and the signal propagation blocked state is determined. If the signal propagation blocked state is confirmed, the anomaly point detection algorithm is used to analyze the time series data of signal intensity to obtain a signal attenuation confirmation result. According to the signal attenuation confirmation result, device state information is collected from the Internet of Things device to generate a state dataset.

9. The method of claim 8, wherein the method further comprises: The S7 also includes: The decision tree algorithm is used to classify the applicability of device state to the bird repelling mode based on the state dataset to determine the switching condition of the standby bird repelling mode. If the switching condition is met, the standby bird repelling mode is activated, the preset sound wave and laser intensity parameters are obtained, and an initial intensity value set is generated. According to the initial intensity value set, the sound wave or laser intensity value is recalculated in combination with the weather change trend data to obtain an adaptive intensity adjustment set. The control signal is sent to the bird repelling device through the adaptive intensity adjustment set to complete the intensity reconfiguration.

10. A power line intelligent inspection and bird repelling system for implementing the steps of the power line intelligent inspection and bird repelling method of any one of claims 1-9, characterized in that, The system includes: The sensor network module collects bird activity signals and weather condition parameters from the foggy environment around the airport through the sensor network to obtain bird aggregation density and weather influence factors. The data classification processing module classifies and processes data based on bird aggregation density and weather influence factors using the K-means clustering method combined with Euclidean distance measurement and cluster center initialization, applies data standardization processing and iterative convergence conditions to determine bird activity patterns and weather change trends. The path prediction module predicts the bird flock movement path through the neural network model if the bird activity pattern shows an aggregation trend and the cluster quality is confirmed by the contour coefficient evaluation to obtain potential interference areas. The demand evaluation module obtains the state information of the Internet of Things device for the potential interference area, judges the bird repelling intensity demand level, and activates the frequency adjustment process if the demand level exceeds the preset threshold. The bird repelling frequency optimization module adopts a genetic algorithm defined by a chromosome coding mode and a fitness function to optimize the bird repelling frequency parameters according to a bird repelling intensity requirement level, sets a cross probability and a mutation rate to control the generation of a candidate parameter group with a determined population size, obtains an adjusted output scheme through a selection operation mechanism and a termination condition judgment, and retains an optimal individual as a fog day exclusive configuration; The bird repelling device control module sends a control instruction to the bird repelling device through the adjusted output scheme to determine a final sound wave or laser intensity value; The standby bird repelling mode module switches to a standby bird repelling mode to obtain an enhanced adaptive response if a weather change trend indicates that the propagation is blocked and a signal attenuation is confirmed through an abnormal point detection, wherein the standby bird repelling mode integrates state information of Internet of Things devices to recalculate the intensity value.

Images