Power transmission line bird damage prevention and early warning method and device and medium
By constructing a distribution condition model and selecting a bird classification expert model using attention mechanism, the bird and drone identification problem is solved, and accurate identification and early warning of birds in the transmission corridor is achieved, thereby reducing the calculation and hardware costs.
Patent Information
- Application Number
- CN202510231300.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-28
- Publication Date
- 2025-06-20
AI Technical Summary
The prior art is difficult to effectively identify and distinguish birds from drones, resulting in identification errors in bird damage prevention and control on transmission lines, affecting the early warning effect.
By constructing a distribution condition model, the relationship between bird distribution and time and environment is modeled based on statistical time, environment and bird population activity count, bird distribution encoding is provided, and an attention mechanism is used to select appropriate bird classification expert models for identification.
It significantly improves the accurate identification ability of birds in the transmission corridor, reduces identification errors, enhances early warning effects, and reduces calculation costs and hardware costs.
Smart Images
Figure CN120180217A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of bird radar signal recognition, and particularly to a method, device and medium for preventing and warning bird damage to transmission lines. Background Art
[0002] With the improvement of the ecological environment, the breeding and inhabiting areas of birds have increased, and the transmission line corridors are often utilized by birds. Behaviors such as nest building and bird droppings excretion are likely to cause tripping accidents such as line short circuits and flashovers, seriously affecting the power supply reliability and bringing huge economic losses. With the development of unmanned aerial vehicles (UAVs), UAVs with various functions have been applied to the inspection and operation and maintenance process of transmission lines. Since the moving speed, flight trajectory, RCS and other characteristics of UAVs and bird targets are highly similar, it is a difficult task to distinguish them.
[0003] Most of the research on the identification of UAVs and birds focuses on the differences in the micro-Doppler characteristics of the two types of targets. Byung Kwan Kim et al. obtained a combined Doppler image by combining micro-Doppler features and a rhythm velocity map, analyzed the Doppler information of the rotary-wing UAV in the time domain and frequency domain, and realized the classification of the rotary-wing UAV by combining with a convolutional neural network (CNN). S. Park et al. used the micro-Doppler bandwidth as the main feature and used a K-nearest neighbor classifier to classify birds and UAVs. Riccardo Palama et al. used the NeXtRAD radar system to extract features such as the maximum value of the negative half-axis of the micro-Doppler spectrum, the average value and variance of the singular values obtained by SVD, and the variance of the positive half-axis of the micro-Doppler spectrum from the micro-Doppler spectra of UAV and bird targets, and used a K-nearest neighbor classifier to realize the discrimination of bird and rotary-wing UAV targets. Molchanov et al. designed an automatic target recognition system, performed singular value decomposition (SVD) on the time-frequency diagrams of UAVs, humans, and birds to extract features such as signal energy, spectral amplitude, and target speed, and combined with SVM to realize the classification of UAVs, humans, and birds. Liu Jia et al. extracted features such as average speed, speed standard deviation, course deflection standard deviation, maneuver factor, and oscillation factor according to the different flight trajectories of birds and rotary-wing UAVs, and used a random forest model to construct a classifier to realize the classification of birds and UAVs. In order to realize the recognition of multiple bird species, the neural network for extracting radar features is often very deep and has a large number of parameters. By adding more parameters to support the recognition of more bird species, this neural network is prone to overfitting during the training process, resulting in low environmental adaptability, and a large number of parameters resulting in slow subsequent calculations, high calculation costs, poor real-time performance, and high hardware costs for deployment. Summary of the Invention
[0004] In order to solve the above technical problems or at least partially solve the above technical problems, the present invention provides a method, device and medium for preventing and warning bird damage to transmission lines.
[0005] In a first aspect, the present invention provides a method for early warning of bird damage control in power transmission lines, comprising:
[0006] Statistics on the periodic changes of bird population activity in the area around power transmission lines under different environmental conditions over the years and days;
[0007] A distribution condition model is constructed. During the training process, the distribution condition model models the relationship between bird distribution and time and environment based on the statistical time, environment and the number of bird population activities in the area around the transmission line. During the early warning process, the distribution condition model provides bird distribution coding according to the current time and environment. The bird distribution coding is converted into model selection condition weights using attention. The dimension of the model selection condition weights corresponds to the total number of bird classification expert models. The model selection condition weights are used to select a bird classification expert model with a better prediction effect on the main birds in the distribution to identify the birds according to the collected radar features. When it is identified that the birds are gathering in the transmission channel, the preset bird repellent is triggered to drive the birds away.
[0008] Furthermore, the number of active birds in the area around the transmission line is counted every day, and the number of active birds in the area around the transmission line is counted every hour of the day; the area around the transmission line is obtained by taking the maximum coverage of the bird activity time on both sides of the transmission line and merging them; the number of active birds in the area around the transmission line is counted using the sampling average method: infrared cameras are deployed at sampling points to collect bird activity data in the area around the transmission line, including: frequency of occurrence, location and type; and a complete timestamp and environmental parameters are annotated for each record, the timestamp is year-month-day-hour, and the environmental parameters include: temperature, weather, wind speed; the collected data is cleaned to remove outliers; and time series data is established according to birds, and the time series data is in hours as the smallest unit.
[0009] Furthermore, for all birds with obvious multi-scale periodicity, the average normalized result of the active data in each time series data is used as the label, and the time and environmental parameters in the time series data, combined with the bird type and the periodic characteristics of the bird, are used as input to train the distribution conditional model; during the training process, the cross entropy loss between the actual bird distribution and the predicted bird distribution is constrained to be minimized.
[0010] Furthermore, in order to enable the distribution condition model to understand time, environmental parameters, and bird types, the time, environmental parameters, and bird types are encoded; wherein the time encoding includes: annual sine and cosine encoding and daily sine and cosine encoding,
[0011] The annual sine and cosine codes are expressed as:
[0012]
[0013] The daily sine-cosine coding representation is as follows:
[0014]
[0015] The time coding is: [year_sin, year_cos, day_sin, day_cos], where Day is the day of the year and hour is the hour of the day;
[0016] In the environmental coding: the temperature and wind speed features are encoded using Z-score standardization, and the weather features are encoded for different weather types through one-hot encoding;
[0017] Birds are encoded using one-hot encoding.
[0018] Furthermore, the distribution conditional model adopts an architecture that combines a recurrent neural network and attention. First, a recurrent neural network for time series prediction is based on time series features. The hidden state output by the recurrent neural network corresponds to the prediction result of the time series. The hidden states within a certain period of the recurrent neural network are aggregated through the attention mechanism to form a distribution weight, and the hidden states are weighted to obtain the bird distribution coding.
[0019] Furthermore, the bird classification expert model includes: a feature extraction network with shared weights constructed based on CNN. The feature extraction network is composed of several residual convolutional blocks and pooling layers stacked. Each residual convolutional block contains two cascaded convolutional layers, and each convolutional layer is followed by a non-linear activation function and a normalization layer. The output of the last residual convolutional block is connected to multiple parallel classification heads;
[0020] The feature extraction network and each classification head form a bird classification expert. The radar signal features and micro-motion features extracted from the echo spectrum by the feature extraction network are input into the top K classification heads activated according to the model selection condition weights. The classification heads perform a linear mapping on the input features through a fully connected layer to obtain a vector of the total number of categories. The output vectors of the selected top K classification heads are weighted and summed according to the weights and then input into the Softmax layer to obtain the probability that the identified radar signal features belong to various types of birds.
[0021] Furthermore, during the training process of the bird classification expert model,
[0022] Within the first training round range, the bird distribution coding of the distribution conditional model is not provided, and the bird classification expert model is trained using the radar signal features of various birds with uniform proportion and the radar signal features of drones;
[0023] In the range between the first training round and the second training round, for each piece of bird data in the training data of the bird classification expert model, several similar birds are preferentially selected according to the similarity of periodic features to play a dominant role, and the remaining birds are used as supplements to form multiple groups of biased training sets corresponding to the bird classification expert model; for all the biased training sets, the selected dominant birds cover all the birds to be identified, and the coincidence rate of the dominant birds in any two biased training sets is lower than the difference rate; the features of any one biased training set are specifically assigned to a classification head for training, so that the classification head has a strong recognition effect on several birds specifically.
[0024] After the second training round, a bird distribution code is provided. The bird distribution code is converted into a model selection condition weight by using attention. Through training, an attention score between the model selection condition weight and the model index is established independently, so that the bird distribution code is converted into a model selection condition weight corresponding to the classification head index according to the attention score, and TopK classification heads are adaptively selected to participate in the prediction.
[0025] Furthermore, the bird radar signal and the UAV radar signal features include the micro-motion features of the bird radar signal and the UAV radar signal; among them, the UAV radar signal is obtained by modeling the time-domain radar signal based on the UAV structure and motion state.
[0026] In a second aspect, the present invention provides a warning device for preventing bird damage to transmission lines, including: at least one processing unit, the processing unit is connected to a storage unit and a collection unit through a bus unit, the storage unit stores a computer program, and when the computer program is executed by the processing unit, the warning method for preventing bird damage to transmission lines as described above is implemented.
[0027] In a third aspect, the present invention provides a computer-readable storage medium, the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the warning method for preventing bird damage to transmission lines as described above is implemented.
[0028] The above technical solutions provided by the embodiments of the present invention have the following advantages compared with the prior art:
[0029] This application counts the periodic changes of the number of bird population activities in the area around the transmission line over the years and days under different environmental conditions; constructs a distribution condition model, which models the relationship between bird distribution and time and environment based on the statistical time, environment, and the number of bird population activities in the area around the transmission line, so as to provide a bird distribution code for the current time environment during the early warning process; converts the bird distribution code into a model selection condition weight, the dimension of which corresponds to the total number of bird classification expert models, and uses the model selection condition weight to select a bird classification expert model with a better prediction effect on the birds that account for the main body of the distribution to identify birds based on the collected radar features. When it is identified that birds gather in the transmission channel, a preset bird repeller is triggered to drive away the birds. Since each classification head focuses on the classification of several types of birds, compared with full-type classification, it requires fewer parameters and is easier to converge during training. There is natural sparsity between different bird classification expert models, and the whole will not overfit. This application trains and independently establishes the attention score between the model selection condition weight and the model index, so as to convert the bird distribution code into the model selection condition weight corresponding to the classification head index according to the attention score, and adaptively select the top K classification heads to participate in the prediction. So that the selected top K classification heads can adapt to the current bird distribution characteristics and give more accurate predictions. This solution significantly improves the accurate identification ability of birds in the transmission corridor by deeply integrating the ecological periodic law and the innovation of the deep learning architecture, and provides reliable technical support for the bird ecological protection of the smart grid under the premise of maintaining reasonable computational overhead. BRIEF DESCRIPTION OF THE DRAWINGS
[0030] The accompanying drawings here are incorporated into the specification and form a part of this specification, showing embodiments consistent with the present invention, and are used together with the specification to explain the principles of the present invention.
[0031] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, for those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.
[0032] Figure 1 It is a flowchart of a method for preventing and warning bird damage to transmission lines provided by an embodiment of the present invention;
[0033] Figure 2 It is a schematic diagram of the overall structure of the neural network model provided by an embodiment of the present invention;
[0034] Figure 3 It is a stage flowchart of the training of the bird classification expert model provided by an embodiment of the present invention;
[0035] Figure 4 Schematic diagram of the warning device for preventing bird damage to transmission lines provided by the embodiments of the present invention. Detailed implementation manners
[0036] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Apparently, the described embodiments are some but not all of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the scope of protection of the present invention.
[0037] It should be noted that in this document, the terms "include", "comprise" or any other variant thereof are intended to cover non-exclusive inclusion, such that a process, method, article or device including a series of elements not only includes those elements but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or device. Without further limitation, an element defined by the phrase "including one..." does not exclude the existence of additional identical elements in the process, method, article or device including the said element.
[0038] Embodiment 1
[0039] As Figure 1 shown, the present invention technically realizes a method for preventing and warning bird damage to transmission lines, including:
[0040] Statistically analyzing the periodic changes of the number of bird population activities in the surrounding area of the transmission line over the years and days under different environmental conditions.
[0041] During the specific implementation process:
[0042] Statistically analyze the daily active number of birds in the surrounding area of the transmission line, and statistically analyze the hourly active number of birds in the surrounding area of the transmission line within a day; the surrounding area of the transmission line is obtained by respectively taking the maximum coverable range of the daily activity time of birds on both sides of the transmission line and combining them. The statistical analysis of the active number of birds in the surrounding area of the transmission line adopts the sampling average method: deploy infrared cameras at the sampling points to collect the bird activity data in the surrounding area of the transmission line, including: the appearance frequency, location and species; and label each record with a complete timestamp and environmental parameters, the timestamp is year-month-day-hour, and the environmental parameters include: temperature, weather, wind speed; clean the collected data to eliminate outliers; and establish time series data according to the birds, and the time series data uses hours as the minimum unit; an example of cleaning is: eliminate the outliers brought by extreme weather according to extreme weather, and the bird activity amount decreases abnormally under extreme weather, destroying the periodic law.
[0043] For various bird species, perform multi-scale periodicity verification based on time series data to determine their periodicity:
[0044] Determination of annual-scale periodicity: Count the daily bird activity numbers in the area around the transmission line within adjacent years, and calculate the average of the daily bird activity numbers within adjacent years; perform a fast Fourier transform on the average of the daily bird activity numbers within adjacent years and take the absolute value. These absolute values represent the amplitudes of different frequency components. The length of the array returned by the fast Fourier transform function is the same as the length of the input array. That is, the result of the fast Fourier transform of the 365-day bird activity numbers will also have 365 points. Then, retain the absolute values of the fast Fourier transform for the first half of the year through slicing. Use the absolute values of the fast Fourier transform retained for the first half of the year to find the index of the dominant frequency component, and then calculate the dominant period by dividing 365 by the index of the dominant frequency component. Take the annual dominant period as the annual cycle feature.
[0045] Determination of daily-scale periodicity: Count the hourly bird activity numbers in the area around the transmission line every day within adjacent days, and calculate the average of the hourly bird activity numbers within adjacent days; perform Gaussian kernel density estimation on the average of the hourly bird activity numbers within adjacent days to calculate the logarithmic probability density of the given bird activity numbers for each hour, find the peak of the logarithmic probability density. There can be multiple peaks. Determine the daily cycle situation by finding the peak moment corresponding to the index of the peak of the logarithmic probability density, and perform OneHot encoding on the peak moment within the range of 24 hours to obtain the daily cycle feature.
[0046] Construct a distribution conditional model. During the training process, the distribution conditional model models the relationship between bird distribution and time and environment based on the statistical time, environment, and the number of bird population activities in the area around the transmission line. And during the warning process, the distribution conditional model provides a bird distribution code according to the current time and environment.
[0047] To enable the distribution conditional model to understand time and environment parameters, and bird types, encode the time and environment parameters and bird types.
[0048] Among them, the time encoding includes: annual sine-cosine encoding and daily sine-cosine encoding.
[0049] The annual sine-cosine encoding is expressed as:
[0050]
[0051] The daily sine-cosine encoding is expressed as:
[0052]
[0053] The time encoding is: [year_sin, year_cos, day_sin, day_cos], where Day is the day of the year and hour is the hour of the day.
[0054] In the environmental encoding: The temperature and wind speed features are encoded using Z-score normalization, and the weather features are encoded for different weather types through one-hot encoding.
[0055] The birds are encoded using one-hot encoding.
[0056] After taking the average and normalizing the time series data, the elements are combined with the corresponding time encoding, environmental encoding, and bird encoding, and arranged in chronological order to obtain a data set for training the distribution conditional model. For all birds with obvious multi-scale periodicity, the average normalization result of the active data in their respective time series data is used as the label, and the time and environmental parameters in the time series data, combined with the bird type and the periodic characteristics of the birds, are used as the input to train the distribution conditional model; during the training process, the loss function includes the cross-entropy loss between the actual bird distribution and the predicted bird distribution, and the mean squared error loss between the predicted time series data and the real time series data, and the parameters of the distribution conditional model are optimized with the goal of minimizing the loss function.
[0057] The distribution conditional model adopts an architecture combining a recurrent neural network and attention. First, a recurrent neural network for time series prediction based on time series features, such as an LSTM or GRU model, is used. The hidden state output by the recurrent neural network corresponds to the prediction result of the time series. The hidden states within a certain period of the recurrent neural network are aggregated through the attention mechanism to form a distribution weight, and the hidden states are weighted to obtain the bird distribution encoding. During the training process, the distribution conditional model models the relationship between the bird distribution and time and environment based on the statistically collected time, environment, and the number of bird population activities in the area surrounding the transmission line. And during the warning process, the distribution conditional model provides the bird distribution encoding according to the current time and environment.
[0058] The bird distribution encoding is converted into model selection conditional weights through attention. The dimension of the model selection conditional weights corresponds to the total number of bird classification expert models. Using the model selection conditional weights, a bird classification expert model with a better prediction effect on the birds accounting for the main body of the distribution is selected to identify the birds based on the collected radar features.
[0059] To take into account bird hazard warnings day and night, this application uses radar to collect target features, and a bird classification expert model under the control of a distribution condition model identifies birds based on the collected radar signal features. Specifically, in the actual power grid operation and maintenance process, drones are used. Therefore, to avoid interference from drones in bird hazard prevention, the bird classification expert model needs to be able to distinguish drones and birds; to achieve the above purpose, a training dataset containing drone radar signals and bird radar signals is constructed for the bird classification expert model. Among them, to reduce the cost of dataset collection, for drones, their drone radar signals are obtained by modeling the time-domain radar signal model based on the drone structure and motion state:
[0060] The radar transmitted signal is:
[0061]
[0062] Among them, rect(t / T) is a rectangular window function: the mathematical meaning of rect(t / T) is that the signal exists in the time range [-T / 2, T / 2], and is zero at other times, restricting the duration of the radar pulse to T, that is, rect(t / T) corresponds to a rectangular pulse with a width of T. is the complex exponential frequency modulation term, and the starting frequency is f c , and the frequency increases according to the slope k, then the phase of the radar signal is: f c t + 0.5kt 2 .
[0063] Under the radar transmitted signal, a drone time-domain radar signal model is constructed according to the structure and motion state of the drone. In the specific implementation process, the radar reflection signal of the drone comes from the radar scattering unit of the drone body and the radar scattering unit of the rotor. The time-domain radar signal model of the drone is as follows:
[0064]
[0065] Among them, D0 is the distance between the drone body and the radar, c is the speed of light, λ is the radar wavelength, σ d is the body scattering coefficient of the drone body, σ is the point target scattering coefficient, M is the number of drone rotors, and N is the number of drone rotor blades;
[0066]
[0067] Among them, α is the angle between the line connecting the projection point of the radar and the drone in the xy plane and the line connecting the radar and the drone, θ m is the initial phase of the m-th rotor, ω m is the angular velocity of the m-th rotor, t o$T_{o}$ is the full time corresponding to the $o$-th pulse, and $\beta$ is the angle between the line connecting the projection points of the radar and the UAV in the $xy$ plane and the $x$-axis.
[0068] The characteristics of bird radar signals and UAV radar signals for analysis include echo spectra, as well as the micro-motion characteristics of bird radar signals and UAV radar signals. The micro-motion characteristics include characteristic spectral energy entropy and spectral symmetry peak pair characteristics. The calculation process of characteristic spectral energy entropy includes: the radar echo undergoes short-time Fourier transform and singular value decomposition to obtain the left singular vector, and the entropy of the left singular vector is calculated to obtain the characteristic spectral energy entropy. The calculation process of spectral symmetry peak pair characteristics includes: extracting the main peak of the radar signal echo spectrum, and taking the main peak as a reference to find symmetric peaks at the same distance from the main peak. There are obvious differences in the micro-motion characteristics of the radar signals of birds and UAVs. The spectral diagram of bird targets has no symmetry and weak periodicity, while the spectral diagram of UAV targets has significant regularity and periodicity. The micro-motion characteristics of the two targets provide a theoretical basis for the subsequent discrimination of birds and rotor UAVs.
[0069] As Figure 2 shown, the bird classification expert model includes: a feature extraction network with shared weights constructed based on CNN. The feature extraction network is composed of several residual convolutional blocks and pooling layers stacked. Each residual convolutional block contains two cascaded convolutional layers, and each convolutional layer is followed by a non-linear activation function and a normalization layer. The output of the last residual convolutional block is connected to multiple parallel classification heads. The feature extraction network and each classification head form a bird classification expert model. The radar signal features extracted from the echo spectrum by the feature extraction network combined with the micro-motion characteristics are input into the top $K$ classification heads activated according to the model selection condition weights. The classification heads perform linear mapping on the input features through a fully connected layer to obtain a vector of the size of the total number of categories. The output vectors of the selected top $K$ classification heads are weighted and summed according to the weights and then input into the Softmax layer to obtain the probability that the identified radar signal features belong to various birds. Since each classification head focuses on the classification of several types of birds, compared with full-type classification, the number of required parameters is less, and it is easier to converge during training. There is natural sparsity between different bird classification expert models, and the overall will not overfit.
[0070] To realize the combination of the distribution condition model and the bird classification expert model, during the training process of the bird classification expert model, as Figure 3 shown, it is trained in the following way:
[0071] Within the range of the first training round, the bird distribution encoding of the distribution condition model is not provided, and the bird classification expert model is trained through the characteristics of bird radar signals and UAV radar signals with uniform proportion of various birds.
[0072] In the range between the first training round and the second training round, for each bird data in the training data of the bird classification expert model, several similar birds with dominant periodic feature similarity are preferentially selected, and the remaining birds are used as supplements, forming multiple groups of biased training sets corresponding to the bird classification expert model; for all the biased training sets, the selected dominant birds cover all the birds to be recognized, and the coincidence rate of the dominant birds in any two biased training sets is lower than the difference rate; the features of any one biased training set are specifically assigned to a classification head for training, so that the classification head has a strong recognition effect on several birds. The biased training sets bring periodic priors to each bird classification expert model, and can form several bird classification experts focusing on the recognition of highly active birds in spring and autumn, and several bird classification experts focusing on diurnal or nocturnal birds; the training in the second stage uses the cross-entropy between the predicted results and the true results of the dominant birds of each bird classification expert model as the loss function.
[0073] After the second training round, the distribution condition model provides a bird distribution encoding according to time and environmental parameters, converts the bird distribution encoding into model selection conditional weights through attention, and autonomously establishes the attention scores between the model selection conditional weights and the model index through training. Then, according to the attention scores, the bird distribution encoding is converted into model selection conditional weights corresponding to the classification head index, and adaptively selects the top K classification heads to participate in the prediction. This enables the selected top K classification heads to adapt to the current bird distribution characteristics and give more accurate predictions; the subsequent training in the third stage uses the cross-entropy between the overall predicted results and the true results as the loss function.
[0074] After the training is completed, the entire network is dynamically evaluated: the validation set is divided by time period for periodic sensitivity accuracy detection. Compared with the existing baseline model, this application has significant improvements in the average accuracy over all time periods, the recognition recall rate of rare birds, and the F1 score of nocturnal recognition detection. In summary, through the deep integration of ecological periodic laws and deep learning architecture innovation, this solution significantly improves the accurate recognition ability of birds in the transmission corridor while maintaining a reasonable computational cost, providing reliable technical support for the bird ecological protection of smart grids.
[0075] Embodiment 2
[0076] See Figure 4As shown in the figure, an embodiment of the present invention provides a warning device for preventing bird damage to transmission lines, including: at least one processing unit, which is connected to a storage unit, a collection unit, and a bird repelling unit through a bus unit. The storage unit, as a computer-readable storage medium, can be used to store software programs, computer-executable programs, and modules, such as the software programs, computer-executable programs, and modules corresponding to a method for preventing bird damage to transmission lines in an embodiment of the present invention, and can be used to store radar signals collected by the collection unit. By running the software programs, computer-executable programs, and modules stored in the storage unit, the processing unit realizes the above-mentioned method for preventing bird damage to transmission lines, including:
[0077] Statistically analyze the periodic changes of the number of bird population activities in the surrounding area of the transmission line over the years and days under different environmental conditions;
[0078] Construct a distribution condition model. During the training process, the distribution condition model models the relationship between bird distribution and time and environment based on the statistically analyzed time, environment, and the number of bird population activities in the surrounding area of the transmission line. And during the warning process, the distribution condition model provides a bird distribution code according to the current time and environment; convert the bird distribution code into a model selection condition weight by using attention. The dimension of the model selection condition weight corresponds to the total number of bird classification expert models. Using the model selection condition weight, select a bird classification expert model that has a better prediction effect on the birds that account for the main body of the distribution to identify the birds according to the collected radar features. When it is identified that the birds gather in the transmission channel, trigger a preset bird repeller to repel the birds.
[0079] Certainly, the computer program stored in the storage unit of the device for implementing the method for preventing bird damage to transmission lines provided by the embodiment of the present invention is not limited to the above-mentioned method operations, and can also execute relevant operations in a method for preventing bird damage to transmission lines provided by any embodiment of the present invention.
[0080] An exemplary collection unit is a radar device installed near the transmission line, which is used to collect radar signals of airborne moving targets. An exemplary bird repelling unit uses a sonic bird repeller.
[0081] Embodiment 3
[0082] An embodiment of the present invention provides a computer-readable storage medium. The computer-readable storage medium stores a computer program, and when the computer program is executed, it realizes the above-mentioned method for preventing bird damage to transmission lines, including:
[0083] Statistically analyze the periodic changes of the number of bird population activities in the surrounding area of the transmission line over the years and days under different environmental conditions;
[0084] Build a distribution condition model. During the training process, the distribution condition model models the relationship between bird distribution and time and environment based on the statistically counted time, environment, and the number of bird population activities in the area around the transmission line. And during the early warning process, the distribution condition model provides a bird distribution code according to the current time environment; convert the bird distribution code into a model selection conditional weight by using attention. The dimension of the model selection conditional weight corresponds to the total number of bird classification expert models. Use the model selection conditional weight to select a bird classification expert model that has a better prediction effect on the birds accounting for the main body of the distribution to identify birds according to the collected radar features. When it is identified that birds gather in the transmission channel, trigger a preset bird repeller to drive away the birds.
[0085] A computer-readable storage medium provided by an embodiment of the present invention, the computer program stored therein is not limited to the method operations as described above, and can also execute related operations in a method for preventing and warning bird damage to transmission lines provided by any embodiment of the present invention.
[0086] In the embodiments provided by the present invention, it should be understood that the disclosed structure and method can be implemented in other ways. For example, the structural embodiments described above are only illustrative. For example, the division of the units is only a logical function division. In actual implementation, there may be other division methods. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point, the displayed or discussed coupling or direct coupling or communication connection between each other can be through some interfaces, and the indirect coupling or communication connection of the structure or unit can be in an electrical, mechanical or other form.
[0087] The units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or may be distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.
[0088] In addition, each functional unit in various embodiments of the present invention can be integrated in a processing unit, or each unit can exist physically alone, or two or more units can be integrated in one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of a software functional unit.
[0089] The above are only specific embodiments of the present invention, enabling those skilled in the art to understand or implement the present invention. Various modifications to these embodiments will be obvious to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to these embodiments shown herein, but rather to the broadest scope consistent with the principles and novel features claimed herein.
Claims
1. A bird damage prevention and early warning method for power transmission lines, characterized in that: include:. Statistics on the periodic changes of bird population activity in the area around power transmission lines under different environmental conditions over the years and days; Constructing a distribution condition model, during the training process, the distribution condition model models the relationship between bird distribution and time and environment based on the statistical time, environment and the number of bird population activities in the area around the power transmission line, and during the early warning process, the distribution condition model provides bird distribution coding according to the current time and environment; The bird distribution code is converted into a model selection condition weight using attention, and the dimension of the model selection condition weight corresponds to the total number of bird classification expert models; using the model selection condition weight, a bird classification expert model with a better prediction effect on the birds that occupy the main part of the distribution is selected to identify the birds according to the collected radar features; when it is identified that birds gather in the power transmission channel, a preset bird repellent is triggered to drive away the birds.
2. The method for early warning of bird damage prevention in power transmission lines according to claim 1, characterized in that: The number of active birds in the area around the transmission line is counted every day, and the number of active birds in the area around the transmission line is counted every hour of the day; the area around the transmission line is obtained by taking the maximum coverage of the bird activity time on both sides of the transmission line and merging them; the number of active birds in the area around the transmission line is counted using the sampling average method: infrared cameras are deployed at sampling points to collect bird activity data in the area around the transmission line, including: frequency of occurrence, location and type; and a complete timestamp and environmental parameters are marked for each record. The timestamp is year-month-day-hour, and the environmental parameters include: temperature, weather, wind speed; the collected data is cleaned and outliers are removed; and time series data is established according to birds, and the time series data is in hours as the smallest unit.
3. The method for early warning of bird damage prevention in power transmission lines according to claim 2, characterized in that: For all birds with obvious multi-scale periodicity, the average normalized result of the active data in each time series data is used as the label, and the time and environmental parameters in the time series data, combined with the bird type and the periodic characteristics of the bird, are used as input to train the distribution conditional model; during the training process, the cross entropy loss between the actual bird distribution and the predicted bird distribution is constrained to be minimized.
4. The method for early warning of bird damage prevention in power transmission lines according to claim 3, characterized in that: In order to enable the distribution condition model to understand time and environmental parameters, bird types, time, environmental parameters and bird types are encoded; wherein the time encoding includes: annual sine and cosine encoding and daily sine and cosine encoding, The annual sine and cosine codes are expressed as: The daily sine and cosine codes are expressed as: The time code is: [year_sin, year_cos, day_sin, day_cos], where Day is the day of the year and hour is the hour of the day; In environmental coding: temperature and wind speed features are coded using Z-score standardization, and weather features are coded using One-hot coding to encode different weather types; Birds are encoded using one-hot encoding.
5. The method for early warning of bird damage prevention in power transmission lines according to claim 3, characterized in that: The distribution conditional model adopts a recurrent neural network and attention combined architecture. First, a recurrent neural network is used to perform time series prediction based on time series features. The hidden state output by the recurrent neural network corresponds to the prediction result of the time series. The hidden states within a certain period of the recurrent neural network are aggregated through the attention mechanism, and distribution weights are formed based on the aggregation results. Each hidden state is weighted to obtain the bird distribution code.
6. The method for early warning of bird damage prevention in power transmission lines according to claim 1, characterized in that: The bird classification expert model includes: a feature extraction network with shared weights constructed based on CNN, wherein the feature extraction network is composed of a plurality of residual convolution blocks and a stack of pooling layers, each residual convolution block includes two convolution layers connected in series, each convolution layer is followed by a nonlinear activation function and a normalization layer, and the output of the last residual convolution block is connected to a plurality of parallel classification heads; The feature extraction network and each classification head form a bird classification expert. The radar signal features extracted from the echo spectrum by the feature extraction network are combined with the micro-motion features and input into the TopK classification heads that are activated according to the weights of the model selection conditions. The classification heads perform linear mapping on the input features through the fully connected layer to obtain a vector of the total number of categories. The output vectors of the selected TopK classification heads are weighted and summed according to the weights and then input into the Softmax layer to obtain the probability of identifying that the radar signal features belong to each type of bird.
7. The method for early warning of bird damage prevention in power transmission lines according to claim 6, characterized in that: The training process of the bird classification expert model includes: In the first training round, the bird distribution coding of the distribution condition model is not provided, and the bird classification expert model is trained by the bird radar signal and drone radar signal characteristics of the uniformity of the proportion of various birds; In the range between the first training round and the second training round, each bird data in the training data of the bird classification expert model is dominated by several bird stations with similar periodic feature similarity, and the remaining birds are supplemented to form multiple groups of biased training sets corresponding to the bird classification expert model; for all biased training sets, the selected dominant birds cover all birds to be identified, and the overlap rate of the dominant birds in any two biased training sets is lower than the difference rate; the features of any biased training set are targetedly assigned to a classification head for training, so that the classification head has a strong recognition effect on several birds; After the second round of training, bird distribution encoding is provided, and the bird distribution encoding is converted into model selection condition weights using attention. The attention score between the model selection condition weights and the model index is independently established, so as to adaptively select the TopK classification heads to participate in the prediction according to the attention score.
8. The method for early warning of bird damage prevention in power transmission lines according to claim 7, characterized in that: The characteristics of bird radar signals and UAV radar signals include micro-motion characteristics of bird radar signals and UAV radar signals; wherein, the UAV radar signal is obtained by modeling a time domain radar signal model based on the structure and motion state of the UAV.
9. A bird damage prevention and early warning device for power transmission lines, characterized in that: include: At least one processing unit, the processing unit is connected to a storage unit via a bus unit, the storage unit stores a computer program, and when the computer program is executed by the processing unit, the power transmission line bird damage prevention and early warning method as claimed in any one of claims 1-8 is implemented.
10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the power transmission line bird damage prevention and early warning method as claimed in any one of claims 1 to 8 is implemented.