Method and Device for Classifying Flight Track Targets of Birds and UAVs Based on External Radar

Abstract

The present invention provides a method and device for classifying flight tracks of birds and unmanned aerial vehicles based on an external radar. The method includes: Step 1 to Step 4. The present invention has a relatively small amount of calculation, and its advantages are particularly prominent when dealing with small targets with short flight tracks at low altitudes, which can improve the target classification and recognition performance of the radar system and is very meaningful for the practical application of external radiation source radars.

Description

Technical Field

[0001] Embodiments of the present invention relate to the technical fields of passive radar and communication, and particularly to a method and device for classifying flight track targets of birds and unmanned aerial vehicles based on an external source radar. Background Art

[0002] External radiation source radars (i.e., external source radars) have been widely studied in the radar field due to their advantages such as good concealment and non-occupation of spectrum. The terrestrial digital television signal based on the DTMB standard has been popularized at present. Using it as an external radiation source for low-altitude target detection has advantages such as low cost, good concealment, and strong survivability. Unmanned aerial vehicles and birds are detection targets that often appear in low-altitude airspace. Due to their relatively low flight altitude, there is a high possibility of collision with the fuselage during the takeoff and landing phases of an aircraft, causing danger. For the detection and identification of these two types of targets, namely birds and unmanned aerial vehicles, the previous traditional approach was to rely on the naked-eye observation and detection by ground staff. However, due to the limitations of the human eye's vision observation range, as well as restrictive factors such as weather and environmental brightness, the observation results obtained by this traditional approach are often inaccurate. Currently, some radars can achieve the detection and identification of unmanned aerial vehicle and bird targets, but often the identification path is single, the method is fixed, and there is a high dependence on the signal amplitude stability of the detected signals of unmanned aerial vehicle and bird targets, resulting in a relatively low actual identification accuracy. Due to the variable and unpredictable characteristics of the flight trajectories of unmanned aerial vehicle and bird detection targets, it has become difficult and inaccurate to detect target trajectories using traditional methods. Therefore, developing a method and device for classifying flight track targets of birds and unmanned aerial vehicles based on an external source radar, which can effectively overcome the defects in the above-mentioned related technologies, has become a technical problem urgently to be solved in the industry. Summary of the Invention

[0003] In view of the above problems existing in the prior art, embodiments of the present invention provide a method and device for classifying flight track targets of birds and unmanned aerial vehicles based on an external source radar.

[0004] First aspect, an embodiment of the present invention provides a method for classifying flight tracks of birds and drones based on an external radar, including: Step 1, obtain the flight track data of drones and birds and perform preprocessing. Use the position, speed, and acceleration information of the targets as the feature input vector, calculate the average value of the duration sequence data of all flight track files, determine four standard lengths of the files, truncate the longer sequences, and fill the sequences shorter than the average length to make the lengths of all flight track files uniform. Perform preprocessing on the duration sequence data of the flight track files, remove local outliers, and perform Chebyshev denoising; Step 2, establish a binary classification model for the flight tracks of drones and birds, and construct classification metrics for quantitative evaluation of the performance of the binary classification model. The classification metrics are accuracy, precision, recall, F1-score, and ROC curve respectively, where the F1-score is an index generated by comprehensively considering precision and recall; Step 3, construct a deep learning network for flight tracks composed of improved LSTM units and a convolutional neural network CNN; Step 4, divide the dataset for training. The division ratio of the training set, validation set, and test set is 5:1:4. Use the four standard lengths selected in Step 1 and three different parameters of the average function for comparative training of the model, and perform result analysis.

[0005] Based on the content of the above method embodiment, in the method for classifying flight tracks of birds and drones based on an external radar provided in the embodiment of the present invention, the flight track data of drones and birds in Step 1 is specifically: the position, speed, and acceleration information of the targets; the feature input vectors in Step 1 are target.csv and Datasetgroup.csv. The target.csv file is used to store the classification information of the two types of targets, drones and birds, and the category labels are marked with -1 and 1. -1 represents the drone category, and 1 represents the bird category; the Datasetgroup.csv file is used to store the division information of the training set, test set, and validation set; the division ratio can be changed according to actual needs; the method for making the lengths of the flight track files uniform in Step 1 is: calculate the average value of the duration sequences of all flight track files, truncate the longer sequences, and fill the sequences shorter than the average length. The average value is the average level of the overall data length, and perform interval estimation on the data to ensure that more than 80% of the data is within the interval; for the standard data length, select the data median as the standard length.

[0006] Based on the content of the above method embodiment, in the method for classifying flight tracks of birds and drones based on an external radar provided in the embodiment of the present invention, the preprocessing of the duration sequence data of the flight track files in Step 1 uses the removal of singular values based on the regional singular factor, and the regional singular factor is:

[0007]

[0008] Among them, k is the adjacent node of point p, K(p) is the set of point p, and L k (p) is the regional density of point p, and L k (a) is the regional density of point a. The regional singularity coefficient Lof(p) of point p is the ratio of the regional average density of point p to the density of point p itself. If Lof(p) is larger, the probability that point p is a singular value relative to the regional adjacent points is greater; conversely, it is smaller.

[0009] Based on the content of the above method embodiments, in the embodiments of the present invention, for the method for classifying flight tracks of birds and unmanned aerial vehicles based on an external radar, the accuracy rate, precision rate, recall rate, and F1-score in step 2 are as follows:

[0010] The accuracy rate refers to the proportion of the number of samples predicted correctly to the total number of samples, and the calculation formula is:

[0011]

[0012] Among them, Accuracy is the accuracy rate, TP is the number of samples that are actually classified as positive and the predicted category is also positive, TN is the number of samples that are actually classified as positive but the predicted category is negative, FP is the number of samples that are actually classified as negative and the predicted category is also negative, and FN is the number of samples that are actually classified as negative but the predicted category is positive;

[0013] The precision rate refers to the proportion of the number of samples whose predicted category is positive and the prediction is correct to the total number of samples whose predicted category is positive, and the calculation formula is:

[0014]

[0015] Among them, Precision is the precision rate;

[0016] The recall rate refers to the proportion of the number of samples whose predicted category is positive and the prediction is correct to the total number of samples whose actual category is positive, and the calculation formula is:

[0017]

[0018] Among them, Recall is the recall rate;

[0019] The F1-score is an index that comprehensively considers the precision rate and the recall rate, and the calculation formula is:

[0020]

[0021] When both Precision and recall are 1, the F1-score reaches the maximum value of 1. As an index for comprehensively considering the precision rate and the recall rate to evaluate the classification performance, if either the precision rate or the recall rate is small, the corresponding F1-score value will be small, indicating that the classification performance of the model is not good;

[0022] The ROC curve is a coordinate graphical analysis tool used for binary classification models. The horizontal axis of the ROC curve is the false positive rate, and the vertical axis is the true positive rate.

[0023] FPR refers to the ratio of samples in the negative class that are mistakenly judged as positive. The calculation formula is:

[0024]

[0025] TPR refers to the ratio of samples that are correctly judged as positive in the positive class. The calculation formula is:

[0026]

[0027] If a binary classification model and a threshold are given, a coordinate point is generated on the ROC curve. As the threshold changes, the entire ROC curve is drawn. Points above the ROC curve represent better classification results, and points below the ROC curve represent worse classification results. The diagonal line represents random classification, and the probability of predicting the correct category is 0.5. The larger the area enclosed by the ROC curve and the horizontal axis, and the closer the curve is to the upper left corner, the better the classification performance of the model.

[0028] Based on the content of the above method embodiment, the bird and drone track target classification method based on exogenous radar provided in the embodiment of the present invention, step 3 specifically includes: improving the long short-term memory network LSTM, and making different identification requirements for the track characteristics according to the different flight characteristics of birds and drones. Birds are more maneuverable during flight, while drones are more stable. An adjustable weighter is added to the LSTM input end, that is, the internal memory state vector of the neural network. When the maneuverability is strong, the weight is increased to amplify the track characteristics for easy track feature recognition. After using LSTM to learn the complex time dynamic characteristics of the target track, a convolutional neural network is continued to be used to process the input data to extract features. The feedback loop is used to optimize the processing parameters to obtain the optimal solution, and a hidden layer located in the intermediate processing process is used to extract and store key features in the input data.

[0029] Based on the content of the above method embodiment, the bird and drone track target classification method based on exogenous radar provided in the embodiment of the present invention has the following improved LSTM parameters: the memory unit is 256Mb, the number of input data vectors is 6, the activation function used in the fully connected layer is the Sigmoid function, the optimizer is the Adam optimizer, the learning rate is 0.001, and the number of training cycles is 200.

[0030] Based on the content of the above method embodiments, in the method for classifying flight tracks of birds and unmanned aerial vehicles (UAVs) based on an external radar provided in the embodiments of the present invention, in step 4, the basis for the division ratio of the data set is that the data sets of UAVs and birds are not balanced, and the amount of UAV data is less than that of bird data. Therefore, when actually dividing the data set ratio, the proportion of the test set is increased.

[0031] In a second aspect, an embodiment of the present invention provides a device for classifying flight tracks of birds and UAVs based on an external radar, including: a first main module for implementing step 1, acquiring flight track data of UAVs and birds and performing preprocessing, taking the position, speed, and acceleration information of the targets as feature input vectors, calculating the average value of the duration sequence data of all flight track files, determining four standard lengths of the files, truncating the longer sequences, and padding the sequences shorter than the average length so that the lengths of all flight track files are unified, performing preprocessing on the duration sequence data of the flight track files, removing local outliers and performing Chebyshev denoising; a second main module for implementing step 2, establishing a binary classification model for the flight tracks of UAVs and birds, and constructing classification metrics for quantitative evaluation of the performance of the binary classification model. The classification metrics are respectively accuracy, precision, recall, F1-score, and ROC curve, where the F1-score is an index generated by comprehensively considering precision and recall; a third main module for implementing step 3, constructing a deep learning network for flight tracks composed of improved LSTM units and a convolutional neural network (CNN); a fourth main module for implementing step 4, dividing the data set for training, with the division ratio of the training set, validation set, and test set being 5:1:4, performing comparative training on the model using the four standard lengths selected in step 1 and three different parameters of the average function, and performing result analysis.

[0032] In a third aspect, an embodiment of the present invention provides an electronic device, including:

[0033] At least one processor; and

[0034] At least one memory communicatively connected to the processor, where:

[0035] The memory stores program instructions executable by the processor, and the processor can execute the method for classifying flight tracks of birds and UAVs based on an external radar provided by any one of the various implementation manners in the first aspect by invoking the program instructions.

[0036] In a fourth aspect, an embodiment of the present invention provides a non-transitory computer-readable storage medium, and the non-transitory computer-readable storage medium stores computer instructions, and the computer instructions cause the computer to execute the method for classifying flight tracks of birds and UAVs based on an external radar provided by any one of the various implementation manners in the first aspect.

[0037] The method and device for classifying bird and drone track targets based on exogenous radar provided in the embodiments of the present invention have low computational complexity and are particularly advantageous when detecting small targets with short tracks at low altitudes. This improves the target classification and recognition performance of the radar system and is of great significance for the practical application of exogenous radar. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0039] Figure 1 Flowchart of a method for classifying bird and drone track targets based on external radar provided by an embodiment of the present invention;

[0040] Figure 2 A schematic diagram of the structure of a bird and drone track target classification device based on external radar provided by an embodiment of the present invention;

[0041] Figure 3 A schematic diagram of the physical structure of an electronic device provided in an embodiment of the present invention;

[0042] Figure 4 A schematic diagram of the format of a track file to be processed provided by an embodiment of the present invention;

[0043] Figure 5 A schematic diagram of the processed track file format provided by an embodiment of the present invention;

[0044] Figure 6 Schematic diagram of the LSTM-CNN classification network structure provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0045] 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 protection scope of the present invention. Additionally, the technical features in each embodiment or individual embodiment provided by the present invention can be combined with each other arbitrarily to form a feasible technical solution. Such combination is not restricted by the order of steps and / or the structural composition mode, but must be based on the ability of those of ordinary skill in the art to implement. When the combination of technical solutions results in contradictions or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the protection scope required by the present invention.

[0046] Using deep learning to achieve the classification of detection targets for the non-linear movements of unmanned aerial vehicles (UAVs) and birds, a large number of high-dimensional features of detection targets can be input into the corresponding recognition network to analyze the trajectory data of UAV and bird detection targets, thereby obtaining highly accurate recognition results. Based on the tracks of birds and UAVs obtained by a digital television passive radar, a target classification method is proposed. By using the track data obtained by the passive radar to detect UAVs and birds, the flight coordinates, speeds, and acceleration information of the two types of targets are used as features for input, and a deep learning network combining a long short-term memory network and a convolutional neural network is constructed to achieve the classification and recognition of birds and UAVs. After the network training is successful, the computational load during actual application is small, and its advantages are particularly prominent when multipath and interference are severe, which is very meaningful for the application of digital television passive radar in detecting low-altitude targets at airports. Based on this idea, the embodiments of the present invention provide a method for classifying the tracks of birds and UAVs based on an external source radar. See Figure 1, the method includes: Step 1, obtain the track data of the drone and the bird and perform preprocessing. Use the position, speed, and acceleration information of the target as the feature input vector. Calculate the average value of the duration sequence data of all track files, determine four standard lengths of the files, truncate the longer sequences, and fill in the sequences shorter than the average length to make the lengths of all track files uniform. Perform preprocessing on the duration sequence data of the track files, remove local outliers, and perform Chebyshev denoising; Step 2, establish a binary classification model for the drone and bird tracks, and construct classification metrics for quantitative evaluation of the performance of the binary classification model. The classification metrics are accuracy, precision, recall, F1-score, and ROC curve respectively, where the F1-score is an index generated by comprehensively considering precision and recall; Step 3, construct a deep learning network for flight tracks composed of improved LSTM units and convolutional neural network CNN; Step 4, divide the dataset for training. The division ratio of the training set, validation set, and test set is 5:1:4. Use the four standard lengths selected in Step 1 and three different parameters of the average function for comparative training of the model, and perform result analysis.

[0047] Based on the content of the above method embodiment, as an optional embodiment, in the method for classifying bird and drone track targets based on an external radar provided in the embodiment of the present invention, the drone and bird track data in Step 1 are specifically: the position, speed, and acceleration information of the target; the feature input vectors in Step 1 are target.csv and Datasetgroup.csv. The target.csv file is used to store the classification information of the two types of targets, namely drones and birds, and the two labels -1 and 1 are used for category annotation, where -1 represents the drone category and 1 represents the bird category; the Datasetgroup.csv file is used to store the division information of the training set, test set, and validation set; the division ratio can be changed according to actual needs; the method for making the lengths of the track files in Step 1 uniform is: calculate the average value of the duration sequences of all track files, truncate the longer sequences, and fill in the sequences shorter than the average length. The average value is the average level of the overall data length, and interval estimation is performed on the data to ensure that more than 80% of the data is within the interval; for the standard data length, the median of the data is selected as the standard length.

[0048] Based on the content of the above method embodiment, as an optional embodiment, in the method for classifying bird and drone track targets based on an external radar provided in the embodiment of the present invention, the preprocessing of the duration sequence data of the track files in Step 1 is to remove outliers based on the regional singular factor, and the regional singular factor is:

[0049]

[0050] Among them, k is the adjacent node of point p, K(p) is the set of point p, and L k (p) is the regional density of point p, and L k (a) is the regional density of point a. The regional singularity coefficient Lof(p) of point p is the ratio of the regional average density of point p to the density of point p itself. If Lof(p) is larger, the probability that point p is a singular value relative to the regional adjacent points is greater; otherwise, it is smaller.

[0051] Based on the content of the above method embodiments, as an optional embodiment, in the embodiment of the present invention, for the method for classifying flight tracks of birds and unmanned aerial vehicles based on an external radar, the accuracy rate, precision rate, recall rate, and F1-score in step 2 are as follows:

[0052] The accuracy rate refers to the proportion of the number of correctly predicted samples to the total number of samples, and the calculation formula is:

[0053]

[0054] Among them, Accuracy is the accuracy rate, TP is the number of samples that are actually classified as positive and the predicted category is also positive, TN is the number of samples that are actually classified as positive but the predicted category is negative, FP is the number of samples that are actually classified as negative and the predicted category is also negative, and FN is the number of samples that are actually classified as negative but the predicted category is positive;

[0055] The precision rate refers to the proportion of the number of samples that are predicted as positive and predicted correctly to the total number of samples predicted as positive, and the calculation formula is:

[0056]

[0057] Among them, Precision is the precision rate;

[0058] The recall rate refers to the proportion of the number of samples that are predicted as positive and predicted correctly to the total number of samples that are actually positive, and the calculation formula is:

[0059]

[0060] Among them, Recall is the recall rate;

[0061] The F1-score is an index that comprehensively considers the precision rate and the recall rate, and the calculation formula is:

[0062]

[0063] When both Precision and Recall are 1, F1-score reaches its maximum value of 1. It is an indicator that comprehensively considers precision and recall to evaluate classification performance. If either precision or recall is small, the F1-score value will be small, indicating that the model classification performance is poor.

[0064] The ROC curve is a coordinate graphical analysis tool used for binary classification models. The horizontal axis of the ROC curve is the false positive rate, and the vertical axis is the true positive rate.

[0065] FPR refers to the ratio of samples in the negative class that are mistakenly judged as positive. The calculation formula is:

[0066]

[0067] TPR refers to the ratio of samples that are correctly judged as positive in the positive class. The calculation formula is:

[0068]

[0069] If a binary classification model and a threshold are given, a coordinate point is generated on the ROC curve. As the threshold changes, the entire ROC curve is drawn. Points above the ROC curve represent better classification results, and points below the ROC curve represent worse classification results. The diagonal line represents random classification, and the probability of predicting the correct category is 0.5. The larger the area enclosed by the ROC curve and the horizontal axis, and the closer the curve is to the upper left corner, the better the classification performance of the model.

[0070] Based on the content of the above method embodiment, as an optional embodiment, the bird and drone track target classification method based on exogenous radar provided in the embodiment of the present invention, step 3 specifically includes: improving the long short-term memory network LSTM, and making different identification requirements for track characteristics according to the different flight characteristics of birds and drones. Birds are more maneuverable during flight, while drones are more stable. An adjustable weighter is added to the LSTM input end, that is, the internal memory state vector of the neural network. When the maneuverability is strong, the weight is increased to amplify the track characteristics and facilitate track feature identification. After using LSTM to learn the complex time dynamic characteristics of the target track, a convolutional neural network is continued to be used to process the input data to extract features. The feedback loop is used to optimize the processing parameters to obtain the optimal solution, and a hidden layer located in the intermediate processing process is used to extract and store key features in the input data.

[0071] Based on the content of the above method embodiments, as an alternative embodiment, in the method for classifying flight tracks of birds and unmanned aerial vehicles (UAVs) based on an external source radar provided in the embodiments of the present invention, the improved LSTM parameters are as follows: the memory cell is 256 Mb, the number of input data vectors is 6, the activation function used in the fully connected layer is the Sigmoid function, the optimizer is the Adam optimizer, the learning rate is 0.001, and the number of training epochs is 200.

[0072] Based on the content of the above method embodiments, as an alternative embodiment, in the method for classifying flight tracks of birds and UAVs based on an external source radar provided in the embodiments of the present invention, in step 4, the basis for the division ratio of the data set is that the data sets of UAVs and birds are not balanced, and the amount of UAV data is less than that of bird data. Therefore, when actually dividing the data set ratio, the proportion of the test set is increased.

[0073] The method for classifying flight tracks of birds and UAVs based on an external source radar provided in the embodiments of the present invention has a small amount of computation, and its advantages are particularly prominent when dealing with short flight tracks of small targets at low altitude. It can improve the target classification and recognition performance of the radar system, which is very meaningful for the practical application of external radiation source radars. In another embodiment, the method for classifying flight tracks of birds and UAVs based on an external source radar provided in the embodiments of the present invention includes: Step 1: The data sets of UAVs and birds used in the present invention are a series of txt files in units of flight tracks. Among them, the total number of flight track files of birds and UAVs is 73. Each flight track file contains 6 attributes: x, y, vx, vy, ax, and az, which represent the position, speed, and acceleration information of the target respectively. At the same time, the length of each flight track file, that is, the duration, is not fixed. To facilitate the introduction of the data form to be processed, a txt file is read from the data set, and the format is as Figure 4 shown. When considering the selection of feature parameters input into the network, since the time stamps and group numbers in the first and second columns of the file have little influence on the classification result and the data processing is relatively complex, they are not used as model input parameters. Therefore, the actually processed data set form is as Figure 5As shown. The dataset also includes two csv files: target.csv and Datasetgroup.csv. The target.csv file is used to store the classification information of two types of targets, drones and birds, and the categories are labeled with two tags, -1 and 1. -1 represents the drone category and 1 represents the bird category. The Datasetgroup.csv file is used to store the division information of the training set, test set, and validation set. The division ratio can be changed according to the experimental needs. Before building the model, it is necessary to process the data types. The input data is all the track files containing two categories of drones and birds. Each track file is a time series collected at 1-second intervals. The time series containing 6-column attribute information is processed into 1 array for input. Traverse all the track files and store them in a list.

[0074] In addition to processing the track dataset, the target.csv file storing the target classification results is also imported and stored as an array with only two results, 1 and -1; the Datasetgroup.csv storing the dataset division ratio is imported and stored as an array with only three results, 1, 2, and 3 (1 represents the training set, 2 represents the validation set, and 3 represents the test set). Since the lengths of each track file are different and the scales of each element in the list storing the data are not uniform, the existing dataset cannot be directly used to build the model. Therefore, we need to determine a suitable length to unify all the track files. The possible solutions are as follows: (1) Fill all sequences with a length less than the maximum length with the number 0 until they are equal to the maximum length. However, this will cause a large error in the data provided to the model. Fill to the maximum length with the data of the last row of each track file. (3) Determine the minimum length of the time series in the dataset and truncate all other sequences to this length. The disadvantage of this approach is that it will cause a large amount of data loss. (4) Calculate the average value of the time series of all track files, truncate the longer sequences, and fill the sequences shorter than the average length. Therefore, the present invention selects the fourth solution to unify the data length, and the maximum value, minimum value, and average value of the dataset file length are listed in Table 1 below:

[0075] Table 1 Track file length statistics

[0076]

[0077] The average value reflects the average level of the overall data length; the standard deviation reflects the degree of dispersion between the data and the average value; the median reflects the true characteristics of the middle level of the data length, making up for the deficiency of the average value in the skewed distribution and is used to provide a statistic that relatively respects the main situation of the sample. Next, interval estimation is performed on the data to ensure that most of the data is within the interval range.

[0078] Analyze the table. The median of the data length is 168, which is in the middle of the overall length. However, considering the data volume required by deep learning, truncating the data to the median will obviously result in the loss of a large amount of real data, and at the same time, a lot of unrealistic data will be filled, which will greatly reduce the authenticity of the data set. Taking the average value as an example, the average value plus or minus the standard deviation forms a confidence interval, and most of the data is concentrated within this interval. If the average value is selected as the standard length, more than a quarter of the long data files will lose a large amount of data. Therefore, considering minimizing the loss of long file data volume as much as possible, for the selection of the standard data length, the present invention mainly adopts the following four ideas:

[0079] (1) Select the sum of the average value and the standard deviation of all file lengths, which is 303, as the standard length.

[0080] (2) Select the larger quartile 224 as the standard length.

[0081] (3) Select the data average value 190 as the standard length.

[0082] (4) Select the data median 168 as the standard length.

[0083] The impact of the difference in the selected standard length on the model classification accuracy will be compared in step 4.

[0084] Step 2: The drone and bird flight track classification model built in this paper is a binary classification model. Therefore, the following classification metrics are considered for quantitative evaluation of the model performance, namely accuracy, precision, recall, and F1-score.

[0085] To intuitively understand the model classification results, the classification matrix TP, FN, FP, TN is used to describe the classification results, and their meanings are as follows:

[0086] TP (True positive): The number of samples that are actually classified as positive and the predicted class is also positive.

[0087] FN (False positive): The number of samples that are actually classified as negative but the predicted class is positive.

[0088] FP (False negative): The number of samples that are actually classified as negative and the predicted class is also negative.

[0089] TN (True negative): The number of samples that are actually classified as positive but the predicted class is negative.

[0090] For an ideal binary classifier, it is required that both FN and FP are 0. However, it is obvious that this ideal situation cannot be achieved in the actual scenario. Therefore, classification metrics are used to evaluate the performance of the classification model based on the number of these four types of samples. The smaller the FN and FP values, the better the performance of the classification model.

[0091] Step 3: In Steps 1 and 2, variable-length sequences are processed, training sets, validation sets, and test sets are created, the data for the LSTM classification model is prepared, and the adopted classification metrics are proposed to evaluate the model. In this Step 3, the model is built.

[0092] The LSTM neural network is improved based on the recursive neural network (RNN). In the RNN network, the hidden layer is established to ensure the persistence of information. However, the RNN network can only remember a short period of time. Once a long period of information is input, due to the problem of gradient disappearance or gradient explosion that occurs during the training stage of the RNN, this information will be lost.

[0093] Therefore, for long time series problems, the improved RNN network, namely the long short-term memory network LSTM, is used to solve them. The LSTM network determines whether to store the corresponding information through specific gates, and this specific architecture is very effective in predicting time series.

[0094] The LSTM network has a structure similar to the RNN network, but there is a specific structure outside the repeating module. Figure 1 Represents the compositional structure of an LSTM unit. This unit has two basic parts: the state unit and the gates. A single-layer LSTM network generally uses three gates: the input gate, the forget gate, and the output gate. The forget gate can be represented by the following formula 2.1:

[0095] f t =σ(W f (h t-1 ,x t )+b f ) (2.1)

[0096] Where σ represents the sigmoid activation function, W f and b f represent the weights and biases of this layer respectively. The forget gate compares the output h t-1 of the previous unit and the input information x t at the current moment, and then outputs a value between 0 and 1 to the state unit C t-1 . 0 means the forget gate hopes the state unit completely forgets this information, and 1 means the forget gate wants to remember the whole piece of information.

[0097] Next, the input gate determines which new information needs to be stored based on past sequence information. The input gate works in a similar way to the forget gate, as shown in Equation 2.2. It connects a tanh layer to create a candidate vector represented by Equation 2.3.

[0098] i t = σ(W i (h t-1 , x t ) + b i ) (2.2)

[0099]

[0100] where W c and b c are the weights and biases of this layer respectively.

[0101] As described above, the state unit and the gate are connected, and the gate can decide to add or delete information according to the needs of the state unit. The next operation is to update C t-1 to obtain the new cell state C t . The mathematical implementation method is shown in Equation 2.4.

[0102]

[0103] Finally, the information output is completed through the output gate. In this process, the sigmoid function is used to define the output, as shown in Equation 2.5, and then the tanh function multiplies the cell state by the output of the sigmoid gate to obtain the required output information, as shown in Equation 2.6.

[0104] o t = σ(W o (h t-1 , x t ) + b o ) (2.5)

[0105] h t = o t * tanh(C t ) (2.6)

[0106] The specific settings of the LSTM classification model constructed by the present invention are listed in Table 2 as follows:

[0107] Table 2 Classification Model Parameter Settings

[0108]

[0109] After adopting LTSM to learn the complex temporal dynamic characteristics of the target trajectory, it is necessary to further extract features and classify them. Therefore, a convolutional neural network (CNN) is continued to be used to complete this task. In the process of the CNN processing the input data to extract features, the processing parameters are optimized through a feedback loop to obtain the optimal solution, and the hidden layer in the middle processing process is used to extract and store the key features in the input data. The CNN network is mainly composed of a convolutional layer and a pooling layer. The convolutional layer contains multiple neurons, and the number of its neurons is related to the features to be extracted. Each neuron only receives data from the adjacent local receptive field, multiplies it with the shared weight matrix and accumulates it. The accumulated result is input into the nonlinear sigmoid function operation to increase the nonlinear segmentation ability, and the convolutional kernel is used to scan the initial data to learn the input signal features. The pooling layer is connected after the convolutional layer to prevent the network from overfitting. The pooling layer performs secondary feature extraction according to local similarity, filters out invalid parameters, and reduces the spatial dimension of the parameters. The LSTM-CNN classification model constructed by the present invention is as Figure 6 shown.

[0110] Step 4: This step mainly conducts training and testing. First, the issue to be solved is the division ratio of the dataset. Considering that the datasets of drones and birds are not balanced, the amount of drone data is much less than that of bird data, and the overall dataset is relatively small. Therefore, when actually dividing the dataset ratio, the proportion of the test set is intentionally increased. The specific division ratios of the training set, validation set, and test set are 5:1:4.

[0111] On this basis, the model is contrastively trained using the four standard lengths selected in Step 1 and three different parameters of the average function, and the results are analyzed.

[0112] For the parameter selection of the average function, in fact, it means that when different parameters are selected, the calculation methods will be different when calculating accuracy, recall rate, and F1-score. The main parameters of the average function discussed here are the following three: micro-average, macro-average, and weighted average.

[0113] (1) Micro-average means that when calculating the accuracy and recall rate indicators, single categories are not considered, but all categories are considered together. The disadvantage of doing this is that each individual of the two categories in the sample is equally considered, ignoring the proportional difference in the number of samples of the two categories.

[0114] (2) Macro-average means that when performing formula calculations, the accuracy or recall rate of the two categories is first calculated separately, and then the sum of the two is calculated and divided by 2 to obtain the final accuracy or recall rate.

[0115] (3) Compared with the macro-average, the weighted average takes into account the differences in the number of samples in different categories. When calculating the final accuracy or recall rate, each category is multiplied by its corresponding weight. The advantage of this approach is that it does not cause the weights of categories with fewer samples to be lost, making the result closer to the true classification result.

[0116] The comparative analysis of the model for the above four standard data lengths and the three calculation methods of micro-average, macro-average, and weighted average is as follows:

[0117] Under the selection of four different standard lengths, calculate the accuracy rate, precision rate, recall rate, and F1-score values of the model for each length under the three different average function parameters of micro-average, macro-average, and weighted average. The model comparison results are shown in Tables 3 to 6.

[0118] Table 3 Comparison of training results of different calculation methods when the data standard length is 303

[0119]

[0120] Table 4 Comparison of training results of different calculation methods when the data standard length is 224

[0121]

[0122] Table 5 Comparison of training results of different calculation methods when the data standard length is 190

[0123]

[0124] Table 6 Comparison of training results of different calculation methods when the data standard length is 168

[0125]

[0126] Analyzing Tables 3 to 6, when the data standard lengths are selected as 303 and 224, the accuracy rate, precision rate, recall rate, and F1-score of the model classification are relatively high, all around 83%. This may be related to the fact that the dataset retains more real data of long files, resulting in a relatively high accuracy rate of the classification result.

[0127] When the average value of the data length is 190, compared with the previous two, the model performance slightly decreases, and the accuracy rate, precision rate, recall rate, and F1-score of the classification are around 79%. When the median value of the data length is 160, it can be seen that the model performance significantly decreases, and the accuracy rate, precision rate, recall rate, and F1-score of the model classification are relatively low, around 50%. This is related to the fact that too much real data is truncated and too much non-real data is filled during data processing.

[0128] Therefore, in terms of the selection of the standard length, we can choose two lengths, 303 or 224. The reason is that the performance gaps of these two lengths in the four classification metrics are very small, and the model performances are both good.

[0129] In addition, by analyzing the selection of the average function parameters, although the differences among the three parameters are not significant, it can still be seen that when using the macro-average method for calculation, the model accuracy, recall rate, and F1-score values are all lower than the other two.

[0130] When using the micro-average method and the weighted-average method, the model performance gaps are relatively small. Compared with the two, when using the weighted-average method, the model accuracy, recall rate, and F1-score are slightly higher, at about 83%. This may be because the datasets of drones and birds are not balanced, and the weighted-average method can better ensure the weights of the categories with less data volume.

[0131] (2) Plot the change trends of the accuracy rate and loss function of the model under four different standard lengths on the training set and the validation set.

[0132] When the data lengths are respectively selected as 303, 224, 190, and 168, the change curves of the accuracy rate and loss function of the model on the training set and the validation set. The abscissa of the picture is the training cycle, and the ordinates are the accuracy rate and the loss decline respectively.

[0133] When the data length is 303, the accuracy rates of the model on the training set and the validation set are both relatively high. The loss function on the training set drops normally and smoothly, and the loss function on the validation set drops slightly fluctuating, but generally still maintains a downward trend, and the model classification performance is good.

[0134] When the data lengths are 224, 190, and 168, the number of cycles when the accuracy rate on the training set reaches the optimal value is relatively small, and the accuracy rate curve stops growing within 100 training cycles. Similarly, by analyzing the loss curve, when the data lengths are 224 and 190, the inflection points of the loss curve on the validation set appear earlier, and the loss curve rises and fluctuates violently after the inflection point. When the data length is 160, the loss curve starts to fluctuate violently almost from the beginning, and the accuracy rate on the validation set drops significantly.

[0135] The reason for this situation may be that when the selected data length is short, the number of node weights generated by the model training becomes smaller, enabling the model to complete the training quickly. Due to the large training cycle, the loss curve shows overfitting, and the generalization ability of the model from the training set to the validation set is poor. In addition, due to the small model data volume, the small proportion of the validation set, and the decline in data authenticity caused by the short data length, there are large errors in the classification results on the validation set, resulting in poor stability of the model classification performance.

[0136] (3) Plot the changing trends of the ROC curves of the models under four different standard lengths on the test set.

[0137] The horizontal axis is the false positive rate, and the vertical axis is the true positive rate. The ROC coordinate points of an ideal classifier are located at (0, 1). When the area enclosed by the curve and the horizontal axis is larger and the curve is closer to the upper left corner, the classification performance of the model is better. When the data length is 303, the area enclosed by the ROC curve and the horizontal axis is 0.79, and the classification performance of the model is good; when the data lengths are 224 or 190, the areas enclosed by the ROC curves and the horizontal axis are 0.69 and 0.66 respectively, and the classification performance of the model slightly decreases; when the data length is 168, the classification performance of the model significantly decreases, and the area enclosed by the ROC curve and the horizontal axis is only 0.53.

[0138] Therefore, based on the above analysis, considering that when the data length is 303 and the calculation method is weighted average, the performance of the model is more stable on the five evaluation indicators, the area enclosed by the ROC curve and the horizontal axis is larger, and more real data is retained. The optimal model finally determined in this paper is the model when the data standard length is 303 and the average parameter is weighted average (weighted). The classification accuracy rate of the model is 82.76%, the precision rate is 83.56%, the recall rate is 82.76%, the F1-score value is 0.8306, and the area enclosed by the ROC curve and the horizontal axis is 0.79, and the classification performance of the model is good.

[0139] The implementation basis of each embodiment of the present invention is achieved through programmed processing by a device with a processor function. Therefore, in engineering practice, the technical solutions and functions of each embodiment of the present invention can be encapsulated into various modules. Based on this actual situation, on the basis of the above embodiments, the embodiment of the present invention provides a device for classifying the flight tracks of birds and unmanned aerial vehicles based on an external radar, and this device is used to execute the method for classifying the flight tracks of birds and unmanned aerial vehicles based on an external radar in the above method embodiment. See Figure 2The device includes: a first main module for implementing step 1, obtaining the drone and bird track data and preprocessing it, taking the target's position, speed, and acceleration information as feature input vectors, calculating the average value of the time series data of all track files, determining the four standard lengths of the files, truncating the longer sequences, and filling them with sequences shorter than the average length to make the length of all track files uniform, preprocessing the time series data of the track files, removing local outliers and performing Chebyshev denoising; a second main module for implementing step 2, establishing a two-classification model for drone and bird tracks, and constructing classification indicators to determine the performance of the two-classification model. The classification indicators are accuracy, precision, recall, F1-score and ROC curve, among which F1-score is an indicator generated by comprehensively considering precision and recall; the third main module is used to implement step 3 and construct a flight track deep learning network composed of an improved LSTM unit and a convolutional neural network CNN; the fourth main module is used to implement step 4 and divide the data set for training. The ratio of training set, validation set and test set is 5:1:4. The four standard lengths selected in step 1 and three different parameters of the average function are used to perform comparative training of the model and analyze the results.

[0140] The embodiment of the present invention provides a device for classifying bird and drone track targets based on external radar, which adopts Figure 2 Several modules in it have small computational complexity and their advantages are particularly prominent when dealing with low-altitude small targets and short tracks. They can improve the target classification and recognition performance of the radar system and are very meaningful for the practical application of external radiation source radar.

[0141] It should be noted that the device in the device embodiment provided by the present invention can be used to implement the method in the above-mentioned method embodiment as well as the method in other method embodiments provided by the present invention. The only difference is that the corresponding functional modules are set. The principle is basically the same as the principle of the above-mentioned device embodiment provided by the present invention. As long as those skilled in the art refer to the specific technical solutions in other method embodiments on the basis of the above-mentioned device embodiment, obtain the corresponding technical means and the technical solutions composed of these technical means by combining technical features, and ensure the practicality of the technical solutions, they can improve the device in the above-mentioned device embodiment to obtain the corresponding device class embodiment, thereby obtaining the corresponding device class embodiment for implementing the methods in other method class embodiments. For example:

[0142] Based on the content of the above device embodiments, as an alternative embodiment, the device for classifying flight paths of birds and drones based on an external radar provided in the embodiments of the present invention further includes: a first sub-module, which is used to implement the drone and bird flight path data in step 1, specifically: the position, speed, and acceleration information of the target; the feature input vectors in step 1 are target.csv and Datasetgroup.csv. The target.csv file is used to store the classification information of the two types of targets, namely drones and birds, and the category labels are -1 and 1. -1 represents the drone category, and 1 represents the bird category; the Datasetgroup.csv file is used to store the division information of the training set, test set, and validation set; the division ratio can be changed according to actual needs; the method for unifying the lengths of the flight path files in step 1 is: calculate the average value of the duration sequences of all flight path files, truncate the longer sequences, and fill in the sequences shorter than the average length. The average value is the average level of the overall data length, and interval estimation is performed on the data to ensure that more than 80% of the data is within the interval range; for the standard data length, the median of the data is selected as the standard length.

[0143] Based on the content of the above device embodiments, as an alternative embodiment, the device for classifying flight paths of birds and drones based on an external radar provided in the embodiments of the present invention further includes: a second sub-module, which is used to preprocess the duration sequence data of the flight path files in step 1 by removing outliers based on the regional singular factor. The regional singular factor is:

[0144]

[0145] where k is the adjacent node of point p, K(p) is the set of point p, L k (p) is the regional density of point p, L k (a) is the regional density of point a. The regional singular coefficient Lof(p) of point p is the ratio of the regional average density of point p to the density of point p itself. If Lof(p) is larger, the probability that point p is an outlier relative to the regional adjacent points is greater, and vice versa.

[0146] Based on the content of the above device embodiments, as an alternative embodiment, the device for classifying flight paths of birds and drones based on an external radar provided in the embodiments of the present invention further includes: a third sub-module, which is used to implement the accuracy rate, precision rate, recall rate, and F1-score in step 2 as follows:

[0147] The accuracy rate refers to the proportion of the number of correctly predicted samples to the total number of samples, and the calculation formula is:

[0148]

[0149] Among them, Accuracy is the correct rate, TP is the number of samples that are actually classified as positive and the predicted category is also positive, TN is the number of samples that are actually classified as positive but the predicted category is negative, FP is the number of samples that are actually classified as negative and the predicted category is also negative, and FN is the number of samples that are actually classified as negative but the predicted category is positive;

[0150] Precision refers to the proportion of the number of samples whose predicted category is positive and the prediction is correct among all samples whose predicted category is positive. The calculation formula is:

[0151]

[0152] Among them, Precision is the precision;

[0153] Recall refers to the proportion of the number of samples whose predicted category is positive and the prediction is correct among all samples whose actual category is positive. The calculation formula is:

[0154]

[0155] Among them, Recall is the recall;

[0156] F1-score is an indicator that comprehensively considers precision and recall. The calculation formula is:

[0157]

[0158] When both Precision and recall are 1, F1-score reaches the maximum value of 1. As an indicator for evaluating classification performance by comprehensively considering precision and recall, if either precision or recall is small, it will correspondingly lead to a small F1-score, indicating that the classification performance of the model is not good;

[0159] The ROC curve is a coordinate graph analysis tool used for binary classification models. The horizontal axis of the ROC curve is the false positive rate, and the vertical axis is the true positive rate;

[0160] FPR refers to the ratio of samples that are wrongly judged as positive among samples of the negative class. The calculation formula is:

[0161]

[0162] TPR refers to the ratio of samples that are correctly judged as positive among samples of the positive class. The calculation formula is:

[0163]

[0164] Given a binary classification model and a threshold, a coordinate point is generated on the ROC curve. As the threshold changes, the entire ROC curve is plotted. Points above the ROC curve represent better classification results, points below the ROC curve represent worse classification results, and the diagonal line represents the case of random classification, where the probability of correctly predicting the class is 0.5. When the area enclosed by the ROC curve and the horizontal axis is larger and the curve is closer to the upper left corner, the classification performance of the model is better.

[0165] Based on the content of the above device embodiment, as an alternative embodiment, the device for classifying flight track targets of birds and unmanned aerial vehicles based on an external radar provided in the embodiments of the present invention further includes: a fourth sub-module, and the steps for implementation in step 3 specifically include: improving the long short-term memory network (LSTM). According to the different flight characteristics of birds and unmanned aerial vehicles, and the recognition requirement that the flight track features are also different. Birds are more maneuverable during flight, while unmanned aerial vehicles are more stable. At the input end of the LSTM, that is, inside the neural network memory state vector, an adjustable weight device is added. When the maneuverability is strong, the weight is increased to amplify the flight track features for facilitating the recognition of flight track features. After using the LSTM to learn the complex time dynamic characteristics of the target flight track, in the process of continuing to use a convolutional neural network to process the input data to extract features, the processing parameters are optimized through a feedback loop to obtain the optimal solution, and the hidden layer located in the middle processing process is used to extract and store the key features in the input data.

[0166] Based on the content of the above device embodiment, as an alternative embodiment, the device for classifying flight track targets of birds and unmanned aerial vehicles based on an external radar provided in the embodiments of the present invention further includes: a fifth sub-module, and the improved LSTM parameters for implementation are: the memory cell is 256 Mb, the number of input data vectors is 6, the activation function used in the fully connected layer is the Sigmoid function, the optimizer is the Adam optimizer, the learning rate is 0.001, and the number of training cycles is 200.

[0167] Based on the content of the above device embodiment, as an alternative embodiment, the device for classifying flight track targets of birds and unmanned aerial vehicles based on an external radar provided in the embodiments of the present invention further includes: a sixth sub-module, and in step 4 for implementation, the basis for the division ratio of the data set is that the data sets of unmanned aerial vehicles and birds are not balanced, and the amount of unmanned aerial vehicle data is less than that of bird data. Therefore, when actually dividing the data set ratio, the proportion of the test set is increased.

[0168] The method of the embodiments of the present invention is implemented relying on an electronic device. Therefore, it is necessary to introduce the relevant electronic device. For this purpose, the embodiments of the present invention provide an electronic device, such as Figure 3As shown in the figure, the electronic device includes: at least one processor, a communications interface, at least one memory, and a communication bus. Among them, the at least one processor, the communications interface, and the at least one memory communicate with each other through the communication bus. The at least one processor can call the logical instructions in the at least one memory to execute all or part of the steps of the methods provided in the foregoing method embodiments.

[0169] In addition, when the logical instructions in the foregoing at least one memory are implemented in the form of a software functional unit and sold or used as an independent product, they can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the method embodiments of the present invention. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs that can store program codes.

[0170] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown 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 modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. A person of ordinary skill in the art can understand and implement it without creative labor.

[0171] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on such an understanding, the technical solution, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disc, etc., and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0172] The flowcharts and block diagrams in the accompanying drawings show the possible architectures, functions and operations of the systems, methods and computer program products according to multiple embodiments of the present invention. Based on this understanding, each box in the flowchart or block diagram can represent a module, program segment or part of the code, and the module, program segment or part of the code contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in an order different from that marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, or sometimes in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flowchart, and the combination of boxes in the block diagram and / or flowchart, can be implemented with a dedicated hardware-based system that performs the specified function or action, or can be implemented with a combination of dedicated hardware and computer instructions.

[0173] It should be noted that the terms "comprise," "include," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or apparatus comprising a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus. In the absence of further limitations, the elements defined by the phrase "comprise..." do not preclude the presence of additional identical elements in the process, method, article, or apparatus comprising the elements.

[0174] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for classifying flight track targets of birds and unmanned aerial vehicles based on external radar, characterized in that, Including: Step 1: Obtain the flight track data of the drone and the bird and perform preprocessing. Use the position, speed, and acceleration information of the target as the feature input vector. Calculate the statistics of the data lengths of all flight track files, determine the four standard lengths of the files. During processing, according to the selected single standard length, truncate the longer sequences and pad the sequences shorter than the standard length to make the lengths of all flight track files uniform. Preprocess the time series data of the flight track files, remove local singular values and perform Chebyshev denoising. Among them, the statistics include the mean, standard deviation, upper quartile, and median. Step 2: Establish a binary classification model for the drone and bird flight tracks, and construct classification metrics for quantitative evaluation of the performance of the binary classification model. The classification metrics are accuracy, precision, recall, F1-score, and ROC curve respectively. Among them, the F1-score is an index generated by comprehensively considering precision and recall. Step 3: Construct a deep learning network for flight tracks composed of improved LSTM units and convolutional neural network CNN. Step 4: Divide the dataset for training. The division ratio of the training set, validation set, and test set is 5:1:

4. Use the four standard lengths selected in Step 1 and three different parameters of the average function for comparative training of the model, and conduct result analysis. The specific content of Step 3 includes: According to the different flight characteristics of the bird and the drone, resulting in different requirements for identifying flight track characteristics. The bird is more maneuverable during flight, while the drone is more stable. At the input end of the LSTM, that is, inside the neural network memory state vector, add an adjustable weight device. When the maneuverability is strong, increase the weight to amplify the flight track characteristics for easy identification of flight track characteristics. After using LSTM to learn the complex time dynamic characteristics of the target flight track, continue to use the convolutional neural network to process the input data to extract features. In the process, optimize the processing parameters through a feedback loop to obtain the optimal solution, and use the hidden layer in the middle processing process to extract and store the key features in the input data.

2. The method for classifying flight path targets of birds and unmanned aerial vehicles based on an external radar according to claim 1, characterized in that, The flight track data of the drone and the bird in Step 1 are specifically: the position, speed, and acceleration information of the target. The feature input vector in Step 1 comes from target.csv and Datasetgroup.csv. The target.csv file is used to store the classification information of the two types of targets, the drone and the bird, and is labeled with two labels, -1 and 1. -1 represents the drone category, and 1 represents the bird category. The Datasetgroup.csv file is used to store the division information of the training set, test set, and validation set. The division ratio can be changed according to actual needs.

3. The method for classifying flight path targets of birds and unmanned aerial vehicles based on an external radar according to claim 2, characterized in that, The preprocessing of the time series data of the flight track files in Step 1 uses the removal of singular values based on the regional singular factor. The regional singular factor is: where k is the adjacent node of point p, and K(p) is the set of point p. is the regional density of point p. is the regional density of point a, and the regional singularity coefficient of point p is the ratio of the regional average density of point p to the density of point p itself. If is larger, the probability that point p is a singular value relative to the regional adjacent points is greater, and vice versa.

4. The method for classifying flight path targets of birds and unmanned aerial vehicles based on an external radar according to claim 3, wherein The accuracy, precision, recall, and F1-score in Step 2 are: Accuracy refers to the proportion of the number of correctly predicted samples to the total number of samples. The calculation formula is: Among them, Accuracy is the correct rate, TP is the number of samples that are actually classified as positive and the predicted category is also positive, TN is the number of samples that are actually classified as positive but the predicted category is negative, FP is the number of samples that are actually classified as negative and the predicted category is also negative, and FN is the number of samples that are actually classified as negative but the predicted category is positive; Precision refers to the ratio of the number of samples whose predicted category is positive and the prediction is correct to the total number of samples whose predicted category is positive. The calculation formula is: Among them, Precision is the precision; Recall refers to the ratio of the number of samples whose predicted category is positive and the prediction is correct to the total number of samples whose actual category is positive. The calculation formula is: Among them, Recall is the recall; The F1-score is an indicator that comprehensively considers precision and recall, and its calculation formula is: When both Precision and recall are 1, the F1-score reaches the maximum value of 1. As an indicator for comprehensively considering precision and recall to evaluate classification performance, if either precision or recall is small, it will correspondingly lead to a small F1-score, indicating that the classification performance of the model is poor; The ROC curve is a coordinate graph analysis tool used for binary classification models. The horizontal axis of the ROC curve is the false positive rate, and the vertical axis is the true positive rate; FPR refers to the ratio of samples in the negative class that are misjudged as the positive class. The calculation formula is: TPR refers to the ratio of samples in the positive class that are correctly judged as the positive class. The calculation formula is: If a binary classification model and a threshold are given, a coordinate point is generated on the ROC curve. As the threshold changes, the entire ROC curve is drawn. Points above the ROC curve represent better classification results, points below the ROC curve represent worse classification results, and the diagonal line represents the case of random classification, where the probability of correctly predicting the category is 0.

5. When the area enclosed by the ROC curve and the horizontal axis is larger and the curve is closer to the upper left corner, the classification performance of the model is better.

5. The method for classifying flight track targets of birds and unmanned aerial vehicles based on an external radar according to claim 1, characterized in that, The parameters of the improved LSTM are: the memory cell is 256 Mb, the number of input data vectors is 6, the activation function used in the fully connected layer is the Sigmoid function, the optimizer is the Adam optimizer, the learning rate is 0.001, and the number of training epochs is 200.

6. The method for classifying flight track targets of birds and unmanned aerial vehicles based on an external radar according to claim 5, wherein, In step 4, the basis for the division ratio of the dataset is that the datasets of drones and birds are not balanced, and the amount of drone data is less than that of bird data. Therefore, when actually dividing the dataset ratio, the proportion of the test set is increased.

7. An apparatus for classifying flight path targets of birds and unmanned aerial vehicles based on an external radar, characterized in that, Including: The first main module is used to implement Step 1: obtain the flight track data of the drone and the bird and perform preprocessing, use the position, speed, and acceleration information of the target as the feature input vector, calculate the statistics of the data lengths of all flight track files, determine the four standard lengths of the files, and during processing, according to the selected single standard length, truncate the longer sequences and pad the sequences shorter than the standard length to make the lengths of all flight track files uniform, perform preprocessing on the time series data of the flight track file duration, remove local singular values and perform Chebyshev denoising. Among them, the statistics include the average value, standard deviation, larger quartile, and median. The second main module is used to implement Step 2: establish a binary classification model for the drone and bird flight tracks, and construct classification metrics for quantitative evaluation of the performance of the binary classification model. The classification metrics are accuracy, precision, recall, F1-score, and ROC curve respectively. Among them, the F1-score is an index generated by comprehensively considering precision and recall. The third main module is used to implement Step 3: construct a deep learning network for flight tracks composed of improved LSTM units and convolutional neural network CNN. The fourth main module is used to implement Step 4: divide the data set for training, and the division ratio of the training set, validation set, and test set is 5:1:

4. Use the four standard lengths selected in Step 1 and three different parameters of the average function for comparative training of the model, and perform result analysis. The specific implementation of Step 3 includes: according to the recognition requirements that the flight track characteristics are different due to the different flight characteristics of the bird and the drone. The bird is more maneuverable during flight, while the drone is more stable. At the input end of the LSTM, that is, inside the neural network memory state vector, an adjustable weight device is added. When the maneuverability is strong, the weight is increased to amplify the flight track characteristics for easy recognition of flight track characteristics. After using LSTM to learn the complex time dynamic characteristics of the target flight track, in the process of continuing to use the convolutional neural network to process the input data to extract features, the processing parameters are optimized through a feedback loop to obtain the optimal solution, and the hidden layer located in the middle processing process is used to extract and store the key features in the input data.

8. An electronic device, characterized in that, Including: At least one processor, at least one memory, and a communication interface; where The processor, memory, and communication interface communicate with each other; The memory stores program instructions executable by the processor, and the processor calls the program instructions to execute the method according to any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium stores computer instructions, and the computer instructions cause the computer to execute the method according to any one of claims 1 to 6.

Images