A method for identifying abnormal bird behavior based on radar monitoring data

Abstract

The invention discloses a method for identifying abnormal bird behavior based on radar monitoring data, which belongs to the technical field of target detection and solves the problem that the single image data information of a target object collected by a data acquisition module of an existing system is incomplete, which affects the accuracy of target object recognition. The method comprises normalizing target object associated data to obtain a normalized data set, performing correlation analysis on the normalized data set by a flying bird perception model to judge whether the target object is a bird target object, generating a target object path trajectory based on a target object recognition result and a path optimization algorithm, and analyzing and calculating the target object path trajectory in combination with a weakly supervised flying bird detection algorithm to generate a path abnormality degree. In the invention, the flying bird perception model can analyze and identify the bird flight trajectory based on the target object recognition result, and judge and identify the abnormal behavior of the bird in combination with the weakly supervised flying bird detection algorithm, thereby ensuring the recognition accuracy of the target object and improving the warning capability of the abnormal behavior of the bird.

Description

Technical Field

[0001] The invention belongs to the technical field of target detection, and in particular relates to a method for identifying abnormal bird behavior based on radar monitoring data. Background Art

[0002] At present, due to the requirements of airport clearance conditions, airports are usually located in the suburbs of cities. The surrounding ecological environment is favorable and bird activities are frequent, which increases the risk of bird strikes at airports. Airport bird strikes refer to collision accidents between aircraft and birds. It is a long-standing traditional safety hazard in the civil aviation industry. Therefore, it is necessary to assess the risk of bird strikes in order to scientifically guide airport bird repellent operations.

[0003] Airport bird identification is an important part of intelligent bird repellent. Real-time monitoring of bird activities and obtaining bird data are the key points of bird repellent. Common bird detection equipment at airports include radar and optoelectronic equipment. Some airports are equipped with infrared night detection and bird sound detection equipment. Various bird information detection equipment has different functional characteristics due to their different working mechanisms. Traditional bird detection and identification methods of bird detection equipment mainly rely on manual feature extraction. This process requires a lot of time and effort and cannot handle all complex situations.

[0004] Chinese patent CN117079050A discloses an AI analysis system for bird recognition based on image, including: a data acquisition module, which is used to obtain image data in the ecological environment of birds through a camera device, and number the data according to the feature data; a data preprocessing module, which is used to preprocess the acquired images, perform feature calculation processing on the images, and set the influencing factors in the feature anomaly and speed anomaly processing process; a feature recognition module, which uses the speed anomaly coefficient and the feature anomaly coefficient, and then evaluates the time, speed and texture features, and completes the anomaly recognition; a feature extraction module, which is used to extract the recognized speed features and texture features, compare the similarity between the bird image and the bird image features in the trained database, and feature mark the image frames that appear. However, the single image data information of the target object collected by the data acquisition module of the existing system is not comprehensive, which affects the target object recognition accuracy. At the same time, it is impossible to judge and analyze the abnormal behavior of birds based on the recognition results, which reduces the warning capability of abnormal behavior of birds. In view of the above problems, we propose a method for identifying abnormal behavior of birds based on radar monitoring data. Summary of the invention

[0005] The purpose of the present invention is to address the shortcomings of the prior art and provide a method for identifying abnormal bird behavior based on radar monitoring data, which solves the problem that the single image data information of the target object collected by the data acquisition module of the existing system is incomplete, affecting the accuracy of target object recognition, and at the same time, it is impossible to judge and analyze the abnormal behavior of birds based on the recognition results, thereby reducing the warning ability of abnormal bird behavior.

[0006] The present invention is implemented as follows: a method for identifying abnormal bird behavior based on radar monitoring data, the method for identifying abnormal bird behavior based on radar monitoring data comprising:

[0007] Acquire multi-source heterogeneous target object associated data, normalize the target object associated data to obtain a normalized data set, and upload the normalized data set to the target database, wherein the target object associated data includes radar monitoring data, photoelectric detection data, infrared detection data, and environmental audio data;

[0008] The bird perception model is pre-built by combining BP neural network and random forest algorithm, and the time-series perception modeling samples are selected from the target database. The perception modeling samples are divided into training set and test set in a ratio of 7:3. The bird perception model is iteratively trained through the training set and test set to output a converged bird perception model.

[0009] The normalized data set is loaded in real time, and the bird perception model performs correlation analysis on the normalized data set to determine whether the target object is a bird target;

[0010] If the target object is a bird target object, the target object is identified based on the bird perception model, and the target object identification result is output, wherein the target object identification result includes the target object type, the target object path trajectory, and the target object marking information.

[0011] Load the target object recognition result, generate the target object path trajectory based on the target object recognition result and the path optimization algorithm, analyze and calculate the target object path trajectory by combining the weakly supervised bird detection algorithm, and generate the path abnormality degree;

[0012] Determine whether the path abnormality meets the preset abnormality threshold. If it does not meet the preset abnormality threshold, trigger the abnormal behavior instruction.

[0013] Preferably, the method for normalizing target object associated data comprises:

[0014] Load the target object associated data, perform L iterative filtering on the target object associated data based on a digital low-pass filter, filter out interference components in the target object associated data, and output the associated data set after iterative filtering;

[0015] Among them, the associated data set after iterative filtering is expressed as:

[0016] (1)

[0017] (2)

[0018] in, Indicates the input target object related data, They are the output representations of the associated data after the digital low-pass filter performs one frequency shift and L frequency shifts on the associated data of the target object, respectively. is the order of the digital low-pass filter, represents the number of associated datasets, represents the magnitude response of the digital low-pass filter, is the sampling angle difference, is the frequency phase of the target object associated data, Represents the number of sampling points of the digital low-pass filter, Indicates the sampling frequency;

[0019] (3)

[0020] in, represents the angle of argument, is the sampling interval;

[0021] Obtain the correlation data set after iterative filtering, use cubic strip interpolation polynomial to interpolate and reconstruct the multiple groups of signals in the correlation data set for normalization, and set 6 groups of interaction thresholds based on the maximum and minimum values ​​of the correlation data set;

[0022] The optimal weighting coefficient of the associated data set is determined based on the Lagrange multiplier method combined with the interaction threshold. The linear weighted fusion algorithm is used to weighted average the different types of data in the associated data set according to the optimal weighting coefficient to obtain a normalized data set.

[0023] The normalized data set calculation formula is as follows:

[0024] (4)

[0025] represents the normalized data set, is the optimal weighting coefficient for different types of data, Reconstruct the normalized input representation for different types of interpolation, is the number of data types;

[0026] When setting 6 groups of interaction thresholds based on the maximum and minimum values ​​of the associated data sets, the value range of the interaction thresholds is:

[0027] (5)

[0028] (6)

[0029] (7)

[0030] in, They are the first, second, third, fourth, fifth and sixth group interaction thresholds respectively. is the number of groups of interaction thresholds, are the maximum and minimum values ​​of the associated data set, respectively. represents the initial weighting coefficients for different types of data, To correct the compensation value.

[0031] Preferably, the method of iteratively training the bird perception model through a training set and a test set includes:

[0032] The BP neural network is used as the initial model of the bird perception model, the number of nodes of the attribute discriminator, the convolutional fusion network, and the weak supervision detection module in the initial model is set, the learning step of the initial model is determined to be 0.001, the transfer function of the initial model is selected as purelin, the iteration round of the initial model is set, and the weight parameter, loss function coefficient, and momentum value of the initial model are initialized;

[0033] Load the training set and test set, use the training set as input, execute the initial model, the initial model discriminates the attributes of the target object, and extracts, fuses, and identifies the features of the training set based on the discrimination results, and outputs the target object recognition result;

[0034] The loss function of the initial model is calculated based on the target recognition result, and the weight parameters and momentum value of the initial model are learned using random forest. The initial model is iteratively trained to the final step to complete the iterative training of the initial model.

[0035] Obtain a test set, test the converged initial model through the test set, output the test recognition result, and determine whether the test recognition result meets the preset test accuracy threshold. If it meets the preset test accuracy threshold, output the converged bird perception model.

[0036] Preferably, the initial model of the bird perception model includes a data input layer, a convolutional fusion network, and an output layer, and the initial model is built using the deep learning framework PyTorch;

[0037] Among them, an attribute discriminator is introduced between the data input layer and the convolutional fusion network, and a weak supervision detection module is introduced between the convolutional fusion network and the output layer. The attribute discriminator is based on blockchain and residual connection network. The attribute discriminator is used to identify the type attributes of the target object. The attribute discriminator defines three groups of modules, namely ResidualBlock block, Inception block and Conv1d block. The Conv1d block is connected to the ResidualBlock block, and the ResidualBlock block is connected to the Inception block. The convolutional fusion network is composed of DeepLab and channel attention module. The convolutional fusion network is calculated using the cross entropy loss function, and DeepLab is calculated using the L1 loss function. The channel attention module is composed of a maximum pooling layer, an average pooling layer, a multi-layer perceptron and three convolutional layers. The weak supervision detection module is composed of four groups of parallel sub-modules, and the weak supervision detection module is based on HRNet.

[0038] Preferably, the method for identifying a target object based on a flying bird perception model specifically includes:

[0039] Obtain a normalized data set, input the normalized data set into a data input layer, and the data input layer performs a dilated convolution process on the data in the normalized data set to remove background noise in the normalized data set;

[0040] The attribute discriminator obtains the normalized data set after the background noise is filtered out, and the normalized data set is input into the multi-channel Conv1d block. The Conv1d block captures the radar wave reflection signal, radio frequency signal, and local audio information in the normalized data set;

[0041] The ResidualBlock performs one-dimensional dilated convolution and jump connection on the radar wave reflection signal, radio frequency signal, and local audio information, and outputs the residual connection result;

[0042] Obtain the residual connection result, and the Inception block down-converts the residual connection result, and determines whether the target object is a bird target based on the DDR3SDRAM mechanism;

[0043] If the target object is a bird, the residual connection result is retained and transmitted to the convolutional fusion network;

[0044] If the target object is not a bird, trigger the warning command of other targets;

[0045] Obtain a normalized data set after background noise is filtered out, input the normalized data set into DeepLab, and DeepLab performs a convolution operation on the normalized data set to obtain the self-vectors of the nodes in the normalized data set;

[0046] Integrate at least one set of self-vectors to form a vector connection matrix, multiply the vector connection matrix by itself to obtain a multi-level connection matrix, assign different weight parameters to the multi-level connection matrix based on the principal component analysis method, and obtain the target object key point matrix;

[0047] The channel attention module performs maximum pooling and average pooling operations on the target object key point matrix, compresses the dimension of the target object key point matrix, and generates a feature map of size 1×1×3. The multi-layer perceptron extracts features from the feature map to extract the trajectory features, position features, speed features, quantity features, color features, texture features, contour features, thermal radiation features, frequency features, and voiceprint features of the target object.

[0048] The three convolutional layers in the channel attention module perform convolutional fusion on the target’s trajectory features, position features, speed features, quantity features, color features, texture features, contour features, thermal radiation features, frequency features, and voiceprint features, and output the target type, target path trajectory, and target labeling information.

[0049] Preferably, when the multi-layer perceptron extracts features from the feature map, the feature map output result is expressed as:

[0050] (8)

[0051] in, represents the output feature map, Represent the maximum pooling operation and the average pooling operation respectively, is the activation function, Represents the target object key point matrix;

[0052] (9)

[0053] in, The activation function for generating the target object key point matrix after assigning weight parameters to the multi-level connection matrix based on principal component analysis, They represent the first-level connection matrix, the second-level connection matrix, and the third-level connection matrix respectively. are the weight parameters of the first-level connection matrix, the second-level connection matrix, and the third-level connection matrix, respectively. Represents a multi-level connection matrix superposition operation.

[0054] Preferably, the method of analyzing and calculating the target path trajectory in combination with the weakly supervised bird detection algorithm to generate the path abnormality degree specifically includes:

[0055] Obtaining target object recognition results, identifying the target object type, target object path trajectory, movement step length, and initial monitoring starting point in the target object recognition results, and calculating at least one set of expected path trajectories combining the target object type, target object path trajectory, movement step length, and initial monitoring starting point based on a path optimization algorithm;

[0056] The weak supervision detection module searches for the best expected path trajectory in at least one set of expected path trajectories based on the weak supervision bird detection algorithm, outputs the best expected path trajectory, and sets the best expected path trajectory as an abnormal path trajectory;

[0057] The formula for searching the expected path trajectory based on the weakly supervised bird detection algorithm is:

[0058] (10)

[0059] in, Indicates that the target has passed The expected path trajectory probability after time, Indicates that the target object passes through the path grid point The transfer entropy of Represents the velocity vector of the target object passing through the path grid point, Represents the obstacle avoidance factor vector of the path grid point where the target passes, They are entropy heuristic factor, path transfer expectation, and obstacle factor. is the total number of path grid points;

[0060] The path abnormality is calculated based on the Bayesian theorem combined with the expected path trajectory probability, entropy heuristic factor, path transfer expectation, and obstacle factor of the abnormal path trajectory;

[0061] The path abnormality is calculated by the following formula:

[0062] (11)

[0063] in, Represents the velocity vector and maximum velocity vector of the target object passing through the path grid points, Represents the obstacle avoidance factor vector of the path grid point where the target passes, is the number of obstacle avoidance factors, They are entropy heuristic factor, path transfer expectation, and obstacle factor respectively.

[0064] Compared with the prior art, the embodiments of the present application have the following beneficial effects:

[0065] In the embodiment of the present invention, the multi-source heterogeneous target object associated data is normalized, thereby facilitating the accurate identification of the target object by the bird perception model, and the bird perception model can analyze and identify the flight trajectory of the bird based on the target object identification result, and combine the weak supervision bird detection algorithm to judge and identify the abnormal behavior of the bird, thereby ensuring the recognition accuracy of the target object, and improving the warning capability of abnormal bird behavior, reducing the flight risk of aircraft at the airport, and overcoming the problem that the single image data information of the target object collected by the data acquisition module of the existing system is incomplete, affecting the recognition accuracy of the target object, and at the same time, it is impossible to judge and analyze the abnormal behavior of the bird based on the recognition result, which reduces the warning capability of abnormal bird behavior.

[0066] In an embodiment of the present invention, a method for normalizing target object associated data is provided. During the normalization process, a digital low-pass filter is used to filter multi-source heterogeneous radar monitoring data, photoelectric detection data, infrared detection data, and environmental audio data, thereby effectively filtering out interference components in the data. At the same time, a linear weighted fusion algorithm is used to weighted average different types of data in the associated data set according to an optimal weighting coefficient, thereby achieving weighted fusion of radar monitoring data, photoelectric detection data, infrared detection data, and environmental audio data, providing a complementary information source for the bird perception model, so that the bird perception model can perform a large field of view search in a wide area, and promptly detect abnormal bird activities entering airports or other key areas.

[0067] In an embodiment of the present invention, a bird perception model is provided. The bird perception model is constructed by using a BP neural network and a random forest algorithm, and an attribute discriminator, a convolutional fusion network, and a weak supervision detection module are introduced into the bird perception model, so that the target object radar signal can be detected and identified by discriminating the target object's attributes, and then the type of the target object and abnormal behavior identification are determined. At the same time, deep learning is combined with weak supervision detection methods to give full play to their respective advantages, thereby improving the robustness and reliability of the bird perception model.

[0068] In the embodiment of the present invention, the path abnormality is calculated based on the Bayesian theorem combined with the expected path trajectory probability of the abnormal path trajectory, the entropy heuristic factor, the path transfer expectation, and the obstacle factor. The Bayesian theorem is good at processing uncertain information. After being combined with other heuristic factors, it can more effectively process a large amount of complex data and improve the efficiency of data processing. At the same time, since multiple factors such as the expected path trajectory probability, the entropy heuristic factor, the path transfer expectation, and the obstacle factor are taken into consideration, the path abnormality calculation has strong adaptability, so that the bird perception model can operate stably in different environments and conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1It is a schematic diagram of the implementation flow of the method for identifying abnormal bird behavior based on radar monitoring data provided by the present invention.

[0070] Figure 2 The figure shows a schematic diagram of the implementation process of the method for normalizing target object associated data.

[0071] Figure 3 The figure shows a schematic diagram of the implementation process of the iterative training method of the bird perception model through the training set and the test set.

[0072] Figure 4 It is a structural schematic diagram of the bird perception model provided by the present invention.

[0073] Figure 5 The figure shows a schematic diagram of the implementation process of the target object recognition method based on the bird perception model.

[0074] Figure 6 The figure shows a schematic diagram of the implementation process of the method for generating path anomaly degree by combining the weakly supervised bird detection algorithm to analyze and calculate the target path trajectory. DETAILED DESCRIPTION

[0075] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by technicians in the technical field of this application; the terms used in the specification of the application herein are only for the purpose of describing specific embodiments and are not intended to limit this application; the terms "including" and "having" and any variations thereof in the specification and claims of this application and the above-mentioned drawings are intended to cover non-exclusive inclusions. The terms "first", "second", etc. in the specification and claims of this application or the above-mentioned drawings are used to distinguish different objects, not to describe a specific order.

[0076] The single image data information of the target object collected by the data acquisition module of the existing system is incomplete, which affects the accuracy of target identification. At the same time, it is impossible to judge and analyze the abnormal behavior of birds based on the identification results, which reduces the warning ability of abnormal bird behavior. To address the above problems, we propose a method for identifying abnormal bird behavior based on radar monitoring data. When implementing the method, multi-source heterogeneous target object related data is first obtained, and the target object related data is normalized to obtain a normalized data set. Then, a bird perception model is pre-built in combination with a BP neural network and a random forest algorithm. The bird perception model first performs an association analysis on the normalized data set to determine whether the target object is a bird target. Then, the target object is identified based on the bird perception model and the target object identification result is output. The target object identification result and the path optimization algorithm generate the target object path trajectory. The target object path trajectory is analyzed and calculated in combination with a weakly supervised bird detection algorithm to generate a path abnormality degree. In the embodiment of the present invention, the multi-source heterogeneous target object associated data is normalized, thereby facilitating the accurate identification of the target object by the bird perception model, and the bird perception model can analyze and identify the flight trajectory of the bird based on the target object identification result, and combine the weak supervision bird detection algorithm to judge and identify the abnormal behavior of the bird, thereby ensuring the recognition accuracy of the target object, and improving the warning capability of abnormal bird behavior, reducing the flight risk of aircraft at the airport, and overcoming the problem that the single image data information of the target object collected by the data acquisition module of the existing system is incomplete, affecting the recognition accuracy of the target object, and at the same time, it is impossible to judge and analyze the abnormal behavior of the bird based on the recognition result, which reduces the warning capability of abnormal bird behavior.

[0077] The embodiment of the present invention provides a method for identifying abnormal bird behavior based on radar monitoring data. Figure 1 The following is a schematic diagram of the implementation process of a method for identifying abnormal bird behavior based on radar monitoring data. The method for identifying abnormal bird behavior based on radar monitoring data specifically includes:

[0078] Step S10, obtaining multi-source heterogeneous target object associated data, normalizing the target object associated data to obtain a normalized data set, and uploading the normalized data set to a target database.

[0079] It should be noted that the target object related data includes but is not limited to radar monitoring data, photoelectric detection data, infrared detection data and environmental audio data;

[0080] The radar monitoring data includes but is not limited to target trajectory, echo size, echo intensity, signal fluctuation, flight speed and flight altitude, while the infrared detection data includes but is not limited to the target's location information, quantity information and behavioral characteristics. The target database stores standard data of standard targets, including but not limited to sparrows, geese, crows, pigeons, magpies, swifts, turtledoves, hoopoes, falcons, sparrowhawks, starlings and gulls.

[0081] Step S20, pre-constructing a flying bird perception model by combining the BP neural network and the random forest algorithm, selecting temporal perception modeling samples from the target database, and dividing the perception modeling samples into a training set and a test set in a ratio of 7:3, iteratively training the flying bird perception model through the training set and the test set, and outputting a converged flying bird perception model;

[0082] Step S30, loading the normalized data set in real time, and performing correlation analysis on the normalized data set by the bird perception model;

[0083] Step S40, determining whether the target object is a bird target object;

[0084] Step S50, if the target object is a bird target object, the target object is identified based on the bird perception model, and the target object identification result is output, wherein the target object identification result includes the target object type, the target object path trajectory, and the target object labeling information.

[0085] If the target object is not a bird, trigger the warning command of other targets;

[0086] Step S60, loading the target object recognition result, generating the target object path trajectory based on the target object recognition result and the path optimization algorithm, analyzing and calculating the target object path trajectory in combination with the weakly supervised bird detection algorithm, and generating the path abnormality degree;

[0087] Step S70, determining whether the path abnormality meets a preset abnormality threshold;

[0088] Step S80: If the preset abnormal threshold is not met, an abnormal behavior instruction is triggered.

[0089] In this embodiment, the preset abnormal threshold may be 0.6-00.75.

[0090] In the embodiment of the present invention, the multi-source heterogeneous target object associated data is normalized, thereby facilitating the accurate identification of the target object by the bird perception model, and the bird perception model can analyze and identify the flight trajectory of the bird based on the target object identification result, and combine the weak supervision bird detection algorithm to judge and identify the abnormal behavior of the bird, thereby ensuring the recognition accuracy of the target object, and improving the warning capability of abnormal bird behavior, reducing the flight risk of aircraft at the airport, and overcoming the problem that the single image data information of the target object collected by the data acquisition module of the existing system is incomplete, affecting the recognition accuracy of the target object, and at the same time, it is impossible to judge and analyze the abnormal behavior of the bird based on the recognition result, which reduces the warning capability of abnormal bird behavior.

[0091] The embodiment of the present invention provides a method for normalizing target object associated data. Figure 2The schematic diagram of the implementation process of the method for normalizing the target object associated data is shown. The method for normalizing the target object associated data specifically includes:

[0092] Step S101, loading target object associated data, performing L iterative filtering processes on the target object associated data based on a digital low-pass filter, filtering out interference components in the target object associated data, and outputting an iteratively filtered associated data set;

[0093] Among them, the associated data set after iterative filtering is expressed as:

[0094] (1)

[0095] (2)

[0096] in, Indicates the input target object related data, They are the output representations of the associated data after the digital low-pass filter performs one frequency shift and L frequency shifts on the associated data of the target object, respectively. is the order of the digital low-pass filter. In this embodiment, the order of the digital low-pass filter is 3-8. represents the number of associated datasets, represents the magnitude response of the digital low-pass filter, is the sampling angle difference, is the frequency phase of the target object associated data, Represents the number of sampling points of the digital low-pass filter, Indicates the sampling frequency;

[0097] (3)

[0098] in, represents the angle of argument, is the sampling interval;

[0099] Step S102, obtaining the correlation data set after iterative filtering, using a cubic strip interpolation polynomial to perform interpolation reconstruction and normalization on multiple groups of signals in the correlation data set, and setting 6 groups of interaction thresholds based on the maximum and minimum values ​​of the correlation data set;

[0100] In the embodiment of the present invention, through cubic spline interpolation, on the one hand, a smooth and continuous curve can be generated between known data points, thereby filling in data gaps or missing parts. This method can not only maintain the continuity of the data, but also improve the overall quality and reliability of the data. On the other hand, the data after cubic spline interpolation and normalization processing can provide a more accurate and reliable input for the model, thereby improving the prediction performance and generalization ability of the model.

[0101] Step S103, based on the Lagrange multiplier method combined with the interaction threshold, the optimal weighting coefficient of the associated data set is determined, and a linear weighted fusion algorithm is used to perform weighted averaging on different types of data in the associated data set according to the optimal weighting coefficient to obtain a normalized data set.

[0102] In this embodiment, the normalized data set calculation formula is as follows:

[0103] (4)

[0104] represents the normalized data set, is the optimal weighting coefficient for different types of data, Reconstruct the normalized input representation for different types of interpolation, is the number of data types. In this embodiment, the number of types can be 4-10;

[0105] When setting 6 groups of interaction thresholds based on the maximum and minimum values ​​of the associated data sets, the value range of the interaction thresholds is:

[0106] (5)

[0107] (6)

[0108] (7)

[0109] in, They are the first, second, third, fourth, fifth and sixth group interaction thresholds respectively. is the number of interaction threshold groups. In this embodiment, the number of groups is 6. are the maximum and minimum values ​​of the associated data set, respectively. Indicates the initial weighting coefficient of different types of data. The initial weighting coefficient is 0.1-0.2. To correct the compensation value, in this embodiment, the correction compensation value may be 0.3.

[0110] In an embodiment of the present invention, a method for normalizing target object associated data is provided. During the normalization process, a digital low-pass filter is used to filter multi-source heterogeneous radar monitoring data, photoelectric detection data, infrared detection data, and environmental audio data, thereby effectively filtering out interference components in the data. At the same time, a linear weighted fusion algorithm is used to weighted average different types of data in the associated data set according to an optimal weighting coefficient, thereby achieving weighted fusion of radar monitoring data, photoelectric detection data, infrared detection data, and environmental audio data, providing a complementary information source for the bird perception model, so that the bird perception model can perform a large field of view search in a wide area, and promptly detect abnormal bird activities entering airports or other key areas.

[0111] The embodiment of the present invention provides a method for iteratively training a bird perception model through a training set and a test set. Figure 3 The following is a schematic diagram of the implementation process of the iterative training method for the bird perception model using the training set and the test set. The method of iterative training the bird perception model using the training set and the test set specifically includes:

[0112] Step S201, using the BP neural network as the initial model of the bird perception model, setting the number of nodes of the attribute discriminator, convolutional fusion network, and weak supervision detection module in the initial model, determining the learning step of the initial model to be 0.001, selecting the transfer function of the initial model to be purelin, setting the iteration rounds of the initial model, and initializing the weight parameters, loss function coefficients, and momentum values ​​of the initial model. In this embodiment, the iteration rounds of the initial model are 200-350 times;

[0113] Step S202, loading the training set and the test set, taking the training set as input, executing the initial model, the initial model discriminates the attributes of the target object, and extracts, fuses, and identifies the features of the training set based on the discrimination results, and outputs the target object recognition result;

[0114] It should be noted that the training set and the test set are temporal perceptual modeling samples selected from the target database. The perceptual modeling samples in the target database are normalized in the same way as steps S101-S103 above. The perceptual modeling samples can be obtained from the "Atlas of Chinese Birds", "Field Manual of Chinese Birds" or the Chinese Bird Atlas. According to the prediction requirements, the required perceptual modeling samples can be screened by specifying species composition, micro-Doppler characteristic parameters, radar frequency bands and other conditions.

[0115] Step S203, calculating the loss function of the initial model based on the target object recognition result, using random forest to learn the weight parameters and momentum value of the initial model, iteratively training the initial model to the final step, and completing the iterative training of the initial model;

[0116] Step S204, obtaining a test set, testing the converged initial model using the test set, and outputting a test recognition result;

[0117] Step S205, determining whether the test recognition result meets a preset test accuracy threshold;

[0118] Step S206: if the preset test accuracy threshold is met, output a converged bird perception model.

[0119] If it does not meet the preset test accuracy threshold, return to step S202 to continue training the model.

[0120] In this embodiment, the preset test accuracy threshold is set to 0.96-0.98.

[0121] Correspondingly, such as Figure 4 As shown, the initial model of the bird perception model includes a data input layer, a convolutional fusion network, and an output layer, and the initial model is built using the deep learning framework PyTorch;

[0122] Among them, an attribute discriminator is introduced between the data input layer and the convolutional fusion network, and a weak supervision detection module is introduced between the convolutional fusion network and the output layer;

[0123] It should be noted that the data input layer is connected to the attribute discriminator, the attribute discriminator is connected to the convolutional fusion network, the convolutional fusion network is connected to the weak supervision detection module, and the weak supervision detection module is connected to the output layer. At the same time, the baseline network HRNet is used to conduct an ablation experiment on the flying bird perception model in bird flock counting and recognition, so that the model converges faster.

[0124] In an embodiment of the present invention, a bird perception model is provided. The bird perception model is constructed by using a BP neural network and a random forest algorithm, and an attribute discriminator, a convolutional fusion network, and a weak supervision detection module are introduced into the bird perception model, so that the target object radar signal can be detected and identified by discriminating the target object's attributes, and then the type of the target object and abnormal behavior identification are determined. At the same time, deep learning is combined with weak supervision detection methods to give full play to their respective advantages, thereby improving the robustness and reliability of the bird perception model.

[0125] The attribute discriminator is based on blockchain and residual connection network. The attribute discriminator is used to identify the type attributes of the target object. The attribute discriminator defines three groups of modules, namely ResidualBlock, Inception and Conv1d. The Conv1d block is connected to the ResidualBlock block, and the ResidualBlock block is connected to the Inception block.

[0126] The convolutional fusion network consists of DeepLab and channel attention modules. The convolutional fusion network is calculated using the cross entropy loss function, while DeepLab is calculated using the L1 loss function. The channel attention module consists of a maximum pooling layer, an average pooling layer, a multi-layer perceptron, and three convolutional layers. The weak supervision detection module consists of four groups of parallel sub-modules. The weak supervision detection module is based on the HRNet framework.

[0127] It should be noted that in the weakly supervised detection module, in order to achieve effective fusion of multi-scale features, there is at least one set of multi-scale feature fusion layers between every two adjacent stages. The multi-scale feature fusion layer performs multiple upsampling and downsampling operations between feature maps of different scales to exchange information between high-resolution and low-resolution features.

[0128] The embodiment of the present invention provides a method for identifying a target object based on a flying bird perception model. Figure 5 The following is a schematic diagram of the implementation process of a method for identifying a target object based on a flying bird perception model. The method for identifying a target object based on a flying bird perception model specifically includes:

[0129] Step S301, obtaining a normalized data set, inputting the normalized data set into a data input layer, and performing a dilated convolution process on the data in the normalized data set to remove background noise in the normalized data set;

[0130] Step S302: The attribute discriminator obtains the normalized data set after the background noise is filtered out, and the normalized data set is input into the multi-channel Conv1d block. The Conv1d block captures the radar wave reflection signal, radio frequency signal, and local audio information in the normalized data set. The Conv1d block has good environmental adaptability and can accurately capture radio frequency signals in various environments, providing stable and reliable data support for distinguishing between drones and birds.

[0131] Step S303, the ResidualBlock performs one-dimensional dilated convolution and skip connection on the radar wave reflection signal, the radio frequency signal, and the local audio information, and outputs the residual connection result;

[0132] Step S304, obtaining the residual connection result, the Inception block down-converts the residual connection result, and determines whether the target object is a bird target object based on the DDR3SDRAM mechanism;

[0133] In the embodiment of the present invention, the Conv1d block, the ResidualBlock block, and the Inception block cooperate with each other to realize the multimodal fusion of the radar wave reflection signal, the radio frequency signal, and the local audio information in the normalized data set, which can give full play to the advantages of various information, thereby improving the accuracy and reliability of distinguishing between drones and bird targets.

[0134] Step S305: if the target object is a bird target object, retain the residual connection result, and transmit the residual connection result to the convolutional fusion network;

[0135] Step S306: If the target object is a non-bird target object, trigger a warning instruction for other targets.

[0136] Step S307, obtaining a normalized data set after background noise is filtered out, inputting the normalized data set into DeepLab, and DeepLab performs a convolution operation on the normalized data set to obtain the self-vectors of the nodes in the normalized data set;

[0137] Step S308, integrating at least one group of self-vectors to form a vector connection matrix, multiplying the vector connection matrix by itself to obtain a multi-level connection matrix, assigning different weight parameters to the multi-level connection matrix based on the principal component analysis method to obtain a target object key point matrix, and the multi-level connection matrix can be a three-level connection matrix;

[0138] In this embodiment, when different weight parameters are assigned to the multi-level connection matrix based on the principal component analysis method, the importance or contribution of each connection matrix is ​​determined according to the result of the principal component analysis method. A weight parameter is assigned to each connection matrix, and this weight reflects the importance of the matrix in the overall analysis. The multi-level connection matrix represents the position information of the graph node itself, and two of the three multi-level connection matrices are randomly selected and multiplied to obtain the target object key point matrix, where the target object key point matrix collects the close-range information of the target object node as the local critical information of the target object node.

[0139] Step S309: The channel attention module performs maximum pooling and average pooling operations on the target object key point matrix, compresses the dimension of the target object key point matrix, generates a feature map of size 1×1×3, and the multi-layer perceptron extracts features from the feature map to extract the trajectory features, position features, speed features, quantity features, color features, texture features, contour features, thermal radiation features, frequency features, and voiceprint features of the target object.

[0140] In step S3010, the three convolutional layers in the channel attention module perform convolution fusion on the trajectory features, position features, speed features, quantity features, color features, texture features, contour features, thermal radiation features, frequency features, and voiceprint features of the target object, and output the target object type, target object path trajectory, and target object labeling information.

[0141] In this embodiment, when the multi-layer perceptron extracts features from the feature map, the feature map output result is expressed as:

[0142] (8)

[0143] in, represents the output feature map, Represent the maximum pooling operation and the average pooling operation respectively, is the activation function, Represents the target object key point matrix;

[0144] (9)

[0145] in, The activation function for generating the target object key point matrix after assigning weight parameters to the multi-level connection matrix based on principal component analysis, They represent the first-level connection matrix, the second-level connection matrix, and the third-level connection matrix respectively. are the weight parameters of the first-level connection matrix, the second-level connection matrix, and the third-level connection matrix, respectively. Represents a multi-level connection matrix superposition operation.

[0146] The embodiment of the present invention provides a method for analyzing and calculating the path trajectory of a target object by combining a weakly supervised bird detection algorithm to generate a path anomaly degree. Figure 6 The following is a schematic diagram of the implementation process of a method for analyzing and calculating the path trajectory of a target object by combining a weakly supervised bird detection algorithm to generate a path abnormality degree. The method for analyzing and calculating the path trajectory of a target object by combining a weakly supervised bird detection algorithm to generate a path abnormality degree specifically includes:

[0147] Step S401, obtain the target object recognition result, identify the target object type, target object path trajectory, movement step length and initial monitoring starting point in the target object recognition result, and calculate at least one set of expected path trajectories combining the target object type, target object path trajectory, movement step length and initial monitoring starting point based on the path optimization algorithm; it should be noted that the path optimization algorithm can be a cuckoo algorithm.

[0148] Step S402, the weak supervision detection module searches for the best expected path trajectory in at least one set of expected path trajectories based on the weak supervision bird detection algorithm, outputs the best expected path trajectory, and sets the best expected path trajectory as an abnormal path trajectory;

[0149] In this embodiment, the optimal expected path trajectory is the trajectory where abnormal bird flight will have the greatest impact on the airport route and take-off and landing aircraft. Abnormal flight may include suddenly changing the flight direction, lowering the flight altitude to enter the airport clearance protection zone, or crossing the route at the critical moment of aircraft take-off and landing. These abnormal trajectories are often difficult to predict, increasing the risk of collision with aircraft. Weakly supervised bird detection algorithms are used to predict the possible flight paths of birds, especially those abnormal trajectories that may cross critical areas. By outputting the optimal expected path trajectory, the potential impact of different trajectories is quantitatively analyzed, and early warnings are issued in a timely manner.

[0150] The formula for searching the expected path trajectory based on the weakly supervised bird detection algorithm is:

[0151] (10)

[0152] in, Indicates that the target has passed The expected path trajectory probability after time, Indicates that the target object passes through the path grid point The transfer entropy of Represents the velocity vector of the target object passing through the path grid point, Represents the obstacle avoidance factor vector of the path grid point where the target passes. They are entropy heuristic factor, path transfer expectation, and obstacle factor. is the total number of path grid points;

[0153] Step S403, calculating the path abnormality based on the Bayesian theorem combined with the expected path trajectory probability of the abnormal path trajectory, the entropy heuristic factor, the path transfer expectation, and the obstacle factor;

[0154] The path abnormality is calculated by the following formula:

[0155] (11)

[0156] in, Represents the velocity vector and maximum velocity vector of the target object passing through the path grid points, Represents the obstacle avoidance factor vector of the path grid point where the target passes, is the number of obstacle avoidance factors, They are entropy heuristic factor, path transfer expectation, and obstacle factor. In this embodiment, the number of obstacle avoidance factors is 150-300, the entropy heuristic factor is 0.1-0.5, the path transfer expectation is 0.02-0.05, and the obstacle factor is 0.08-0.21.

[0157] In the embodiment of the present invention, the path abnormality is calculated based on the Bayesian theorem combined with the expected path trajectory probability of the abnormal path trajectory, the entropy heuristic factor, the path transfer expectation, and the obstacle factor. The Bayesian theorem is good at processing uncertain information. After being combined with other heuristic factors, it can more effectively process a large amount of complex data and improve the efficiency of data processing. At the same time, since multiple factors such as the expected path trajectory probability, the entropy heuristic factor, the path transfer expectation, and the obstacle factor are taken into consideration, the path abnormality calculation has strong adaptability, so that the bird perception model can operate stably in different environments and conditions.

[0158] In summary, the present invention provides a method for identifying abnormal bird behavior based on radar monitoring data. In an embodiment of the present invention, the multi-source heterogeneous target object associated data is normalized, thereby facilitating the accurate identification of the target object by the flying bird perception model, and the flying bird perception model can analyze and identify the flight trajectory of the bird based on the target object identification result, and combine the weak supervision bird detection algorithm to judge and identify the abnormal behavior of the bird, thereby ensuring the recognition accuracy of the target object, and improving the warning capability of abnormal bird behavior, reducing the flight risk of airport aircraft, and overcoming the problem that the single image data information of the target object collected by the data acquisition module of the existing system is incomplete, affecting the recognition accuracy of the target object, and at the same time, it is impossible to judge and analyze the abnormal behavior of the bird based on the recognition result, which reduces the warning capability of abnormal bird behavior.

[0159] It should be noted that, for the above-mentioned embodiments, for the sake of simplicity, they are all described as a series of action combinations, but those skilled in the art should know that the present invention is not limited by the described order of actions, because according to the present invention, some steps may be performed in other orders or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the present invention.

[0160] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the scope of protection of the invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on these embodiments, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in this field can still combine, add, delete or make other adjustments to the features in the various embodiments of the present invention according to the circumstances without conflict, without making creative work, so as to obtain different other technical solutions that do not deviate from the concept of the present invention in essence, and these technical solutions also belong to the scope of protection of the present invention.

Claims

1. A method for identifying abnormal bird behavior based on radar monitoring data, characterized in that: The method for identifying abnormal bird behavior based on radar monitoring data includes: Acquire multi-source heterogeneous target object associated data, normalize the target object associated data to obtain a normalized data set, and upload the normalized data set to the target database, wherein the target object associated data includes radar monitoring data, photoelectric detection data, infrared detection data, and environmental audio data; The bird perception model is pre-built by combining BP neural network and random forest algorithm, and the time-series perception modeling samples are selected from the target database. The perception modeling samples are divided into training set and test set in a ratio of 7:

3. The bird perception model is iteratively trained through the training set and test set to output a converged bird perception model. The normalized data set is loaded in real time, and the bird perception model performs correlation analysis on the normalized data set to determine whether the target object is a bird target; If the target object is a bird target object, the target object is identified based on the bird perception model, and the target object identification result is output, wherein the target object identification result includes the target object type, the target object path trajectory, and the target object labeling information; The method for normalizing target object associated data includes: Load the target object associated data, perform L iterative filtering on the target object associated data based on a digital low-pass filter, filter out interference components in the target object associated data, and output the associated data set after iterative filtering; Among them, the associated data set after iterative filtering is expressed as: （1） （2） in, Indicates the input target object related data, They are the output representations of the associated data after the digital low-pass filter performs one frequency shift and L frequency shifts on the associated data of the target object, respectively. is the order of the digital low-pass filter, represents the number of associated datasets, represents the magnitude response of the digital low-pass filter, is the sampling angle difference, is the frequency phase of the target object associated data, Represents the number of sampling points of the digital low-pass filter, Indicates the sampling frequency; （3） in, represents the angle of argument, is the sampling interval; Obtain the correlation data set after iterative filtering, use cubic strip interpolation polynomial to interpolate and reconstruct the multiple groups of signals in the correlation data set for normalization, and set 6 groups of interaction thresholds based on the maximum and minimum values ​​of the correlation data set; Based on the Lagrange multiplier method combined with the interaction threshold, the optimal weight coefficient of the associated data set is determined, and the linear weighted fusion algorithm is used to perform weighted averaging of different types of data in the associated data set according to the optimal weight coefficient to obtain a normalized data set. The normalized data set calculation formula is as follows: （4） represents the normalized data set, is the optimal weighting coefficient for different types of data, Reconstruct the normalized input representation for different types of interpolation, is the number of data types; When setting 6 groups of interaction thresholds based on the maximum and minimum values ​​of the associated data sets, the value range of the interaction thresholds is: （5） （6） （7） in, They are the first, second, third, fourth, fifth and sixth group interaction thresholds respectively. is the number of groups of interaction thresholds, are the maximum and minimum values ​​of the associated data set, respectively. Represents the initial weighting coefficients for different types of data, To correct the compensation value.

2. The method for identifying abnormal bird behavior based on radar monitoring data according to claim 1, characterized in that: The method for identifying abnormal bird behavior based on radar monitoring data also includes: Load the target object recognition result, generate the target object path trajectory based on the target object recognition result and the path optimization algorithm, analyze and calculate the target object path trajectory by combining the weakly supervised bird detection algorithm, and generate the path abnormality degree; Determine whether the path abnormality meets the preset abnormality threshold. If it does not meet the preset abnormality threshold, trigger the abnormal behavior instruction.

3. The method for identifying abnormal bird behavior based on radar monitoring data according to claim 1, characterized in that: The method for iteratively training the bird perception model through a training set and a test set includes: The BP neural network is used as the initial model of the bird perception model, the number of nodes of the attribute discriminator, the convolutional fusion network, and the weak supervision detection module in the initial model is set, the learning step of the initial model is determined to be 0.001, the transfer function of the initial model is selected as purelin, the iteration round of the initial model is set, and the weight parameter, loss function coefficient, and momentum value of the initial model are initialized; Load the training set and test set, use the training set as input, execute the initial model, the initial model discriminates the attributes of the target object, and extracts, fuses, and identifies the features of the training set based on the discrimination results, and outputs the target object recognition result; The loss function of the initial model is calculated based on the target recognition result, and the weight parameters and momentum value of the initial model are learned using random forest. The initial model is iteratively trained to the final step to complete the iterative training of the initial model. Obtain a test set, test the converged initial model through the test set, output the test recognition result, and determine whether the test recognition result meets the preset test accuracy threshold. If it meets the preset test accuracy threshold, output the converged bird perception model.

4. The method for identifying abnormal bird behavior based on radar monitoring data as claimed in claim 2, characterized in that: The initial model of the bird perception model includes a data input layer, a convolutional fusion network, and an output layer, and the initial model is built using the deep learning framework PyTorch; Among them, an attribute discriminator is introduced between the data input layer and the convolutional fusion network, and a weak supervision detection module is introduced between the convolutional fusion network and the output layer. The attribute discriminator is based on blockchain and residual connection network. The attribute discriminator is used to identify the type attributes of the target object. The attribute discriminator defines three groups of modules, namely ResidualBlock block, Inception block and Conv1d block. The Conv1d block is connected to the ResidualBlock block, and the ResidualBlock block is connected to the Inception block. The convolutional fusion network is composed of DeepLab and channel attention module. The convolutional fusion network is calculated using the cross entropy loss function, and DeepLab is calculated using the L1 loss function. The channel attention module is composed of a maximum pooling layer, an average pooling layer, a multi-layer perceptron and three convolutional layers. The weak supervision detection module is composed of four groups of parallel sub-modules, and the weak supervision detection module is based on HRNet.

5. The method for identifying abnormal bird behavior based on radar monitoring data as claimed in claim 4, characterized in that: The method for identifying a target object based on a flying bird perception model specifically includes: Obtain a normalized data set, input the normalized data set into a data input layer, and the data input layer performs a dilated convolution process on the data in the normalized data set to remove background noise in the normalized data set; The attribute discriminator obtains the normalized data set after the background noise is filtered out, and the normalized data set is input into the multi-channel Conv1d block. The Conv1d block captures the radar wave reflection signal, radio frequency signal, and local audio information in the normalized data set; The ResidualBlock performs one-dimensional dilated convolution and jump connection on the radar wave reflection signal, radio frequency signal, and local audio information, and outputs the residual connection result; Obtain the residual connection result, and the Inception block down-converts the residual connection result, and determines whether the target object is a bird target based on the DDR3SDRAM mechanism; If the target object is a bird, the residual connection result is retained and transmitted to the convolutional fusion network; If the target object is not a bird, a warning command for other targets is triggered.

6. The method for identifying abnormal bird behavior based on radar monitoring data according to claim 5, characterized in that: The method for identifying a target object based on a flying bird perception model specifically includes: Obtain a normalized data set after background noise is filtered out, input the normalized data set into DeepLab, and DeepLab performs a convolution operation on the normalized data set to obtain the self-vectors of the nodes in the normalized data set; Integrate at least one set of self-vectors to form a vector connection matrix, multiply the vector connection matrix by itself to obtain a multi-level connection matrix, assign different weight parameters to the multi-level connection matrix based on the principal component analysis method, and obtain the target object key point matrix; The channel attention module performs maximum pooling and average pooling operations on the target object key point matrix, compresses the dimension of the target object key point matrix, and generates a feature map of size 1×1×3. The multi-layer perceptron extracts features from the feature map to extract the trajectory features, position features, speed features, quantity features, color features, texture features, contour features, thermal radiation features, frequency features, and voiceprint features of the target object. The three convolutional layers in the channel attention module perform convolutional fusion on the target’s trajectory features, position features, speed features, quantity features, color features, texture features, contour features, thermal radiation features, frequency features, and voiceprint features, and output the target type, target path trajectory, and target labeling information.

7. The method for identifying abnormal bird behavior based on radar monitoring data according to claim 6, characterized in that: When the multi-layer perceptron extracts features from the feature map, the feature map output result is expressed as: （8） in, represents the output feature map, Represent the maximum pooling operation and the average pooling operation respectively, is the activation function, Represents the target object key point matrix; （9） in, The activation function for generating the target object key point matrix after assigning weight parameters to the multi-level connection matrix based on principal component analysis, They represent the first-level connection matrix, the second-level connection matrix, and the third-level connection matrix respectively. are the weight parameters of the first-level connection matrix, the second-level connection matrix, and the third-level connection matrix, respectively. Represents a multi-level connection matrix superposition operation.

8. The method for identifying abnormal bird behavior based on radar monitoring data according to claim 7, characterized in that: The method of analyzing and calculating the target path trajectory by combining the weakly supervised bird detection algorithm to generate the path abnormality degree specifically includes: Obtaining target object recognition results, identifying the target object type, target object path trajectory, motion step length, and initial monitoring starting point in the target object recognition results, and calculating at least one set of expected path trajectories combining the target object type, target object path trajectory, motion step length, and initial monitoring starting point based on a path optimization algorithm; The weak supervision detection module searches for the best expected path trajectory in at least one set of expected path trajectories based on the weak supervision bird detection algorithm, outputs the best expected path trajectory, and sets the best expected path trajectory as an abnormal path trajectory; The formula for searching the expected path trajectory based on the weakly supervised bird detection algorithm is: （10） in, Indicates that the target has passed The expected path trajectory probability after time, Indicates that the target object passes through the path grid point The transfer entropy of Represents the velocity vector of the target object passing through the path grid point, Represents the obstacle avoidance factor vector of the path grid point where the target passes, They are entropy heuristic factor, path transfer expectation, and obstacle factor. is the total number of path grid points; The path abnormality is calculated based on the Bayesian theorem combined with the expected path trajectory probability, entropy heuristic factor, path transfer expectation, and obstacle factor of the abnormal path trajectory; The path abnormality is calculated by the following formula: （11） in, Represents the velocity vector and maximum velocity vector of the target object passing through the path grid points, Represents the obstacle avoidance factor vector of the path grid point where the target passes, is the number of obstacle avoidance factors, They are entropy heuristic factor, path transfer expectation, and obstacle factor respectively.

Images