A method for identifying and classifying drones based on deep learning

Abstract

The present invention discloses a method for classifying unmanned aerial vehicles based on deep learning, comprising the following steps: S1, acquiring one-dimensional millimeter-wave data of the unmanned aerial vehicle, converting the one-dimensional millimeter-wave data into a two-dimensional image using a data conversion method, and dividing it into a training set, a validation set, and a test set; S2, building a convolutional neural network model based on ResNet18; the attention mechanism is the scSE module, and the scSE module enhances meaningful features and suppresses useless features from both the spatial and channel aspects, thereby improving the recognition accuracy of the network; S3, training the model on the training set according to the set hyperparameters, and obtaining a network model that can accurately classify unmanned aerial vehicles after the training is completed; S4, applying the unmanned aerial vehicle classification network model for classifying unmanned aerial vehicles. Converting the one-dimensional millimeter-wave data into a two-dimensional image increases the feature differences of the data, and the recognition accuracy of the network is improved by combining the attention mechanism and the residual network.

Description

Technical Field

[0001] The present invention relates to the technical fields of artificial intelligence and unmanned aerial vehicle technology, and specifically refers to a method for identifying and classifying unmanned aerial vehicles based on deep learning. Background Art

[0002] In recent years, the rapid development of unmanned aerial vehicle technology has spawned many interesting applications, such as cargo transportation, aerial mapping, military reconnaissance, etc. Although these new applications have brought many conveniences, the abuse of unmanned aerial vehicle technology will also pose threats to public safety, privacy, etc. If unmanned aerial vehicles are used for biochemical weapon attacks, the losses will be even more serious.

[0003] Compared with manned aircraft, it has the characteristics of small volume, low cost, convenient use, low requirements for the combat environment, and strong battlefield survivability, which also brings certain difficulties to identification. Therefore, real-time early detection and identification of unmanned aerial vehicles are crucial in real-world scenarios.

[0004] Using intelligent methods to classify unmanned aerial vehicles can suppress the abuse of unmanned aerial vehicles and protect people's property, personal safety, etc. Therefore, there is an urgent need for an intelligent technology to solve the problem of unmanned aerial vehicle classification. Summary of the Invention

[0005] The main purpose of the present invention is to address the above deficiencies. The present invention discloses a method for classifying unmanned aerial vehicles based on deep learning. Convert one-dimensional millimeter-wave data into two-dimensional images to increase the feature differences of the data, and improve the recognition accuracy of the network by combining the attention mechanism and the residual network.

[0006] To solve the above technical problems, the technical solution of the present invention is as follows:

[0007] A method for identifying and classifying unmanned aerial vehicles based on deep learning, comprising the following steps:

[0008] S1. Obtain the one-dimensional millimeter-wave data of the unmanned aerial vehicle, use a data conversion method to convert the one-dimensional millimeter-wave data into a two-dimensional image, and divide it into a training set, a validation set, and a test set;

[0009] S2. Build a convolutional neural network model based on ResNet18,

[0010] S2-1. Taking ResNet18 as the backbone network, embed and fuse the attention mechanism into the ResNet18 network in a parallel manner;

[0011] S2-2. Modify the number of neuron connections in the fully connected layer of the backbone network according to the number of categories in the data set;

[0012] The attention mechanism is the scSE module. The scSE module enhances meaningful features and suppresses useless features from both spatial and channel aspects, thereby improving the recognition accuracy of the network.

[0013] S3. Train the model on the training set according to the set hyperparameters. After the training is completed, a network model that can accurately classify drones is obtained.

[0014] S4. Apply the drone classification network model for drone classification.

[0015] In the above technical solution, through step S1, the Gramian angular field can convert one-dimensional data into a two-dimensional image. The two-dimensional image has richer information and expands the differences in the main features of the data.

[0016] Preferably, in step S3, the training method of the convolutional neural network model based on ResNet18 is: set the initial values of four hyperparameters, namely the number of iterations, learning rate, batch size, and optimizer. The drone classification network model is trained under the four hyperparameters until the recognition accuracy on the training set no longer improves, and the losses of the training set and the validation set are close and tend to zero.

[0017] In the above technical solution, the model is trained on the training set according to the set hyperparameters. After the training is completed, a convolutional neural network model that can accurately identify the types of drones is obtained.

[0018] Preferably, in step S3, when training the convolutional neural network model based on ResNet18, it is trained by using the cross-entropy loss function.

[0019] 1. Preferably, in step S4, the specific method for classification is as follows:

[0020] S4-1. Image preprocessing: Adjust the input image to match the size and normalized pixel values of the model input.

[0021] S4-2. Feature extraction: Extract features through convolutional layers and average pooling layers.

[0022] S4-3. Classification by the classifier: Flatten the extracted features into a feature vector, classify through a fully connected layer, and finally use the Softmax function to convert the output into the probability distribution of each category. The expression of the Softmax function is as follows:

[0023]

[0024] where z jThe original score of the j-th element in the input vector, N represents the dimension of the input vector, i.e., the number of categories; e represents the base of the natural logarithm.

[0025] Preferably, the scSE module includes a convolutional layer, a Sigmoid layer, a ReLU layer, and an average pooling layer.

[0026] Preferably, the training set, validation set, and test set are divided in a ratio of 7:2:1.

[0027] The present invention has the following characteristics and beneficial effects:

[0028] Adopting the above technical solution, the present invention improves the original ResNet18 from two aspects, improving the network recognition accuracy without increasing the number of parameters of the network model, enabling it to well complete the recognition of UAV types, and being easily deployed to actual 5G base stations, with a more reasonable and effective design. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0030] Figure 1 is the overall flowchart of an embodiment of a UAV recognition and classification method based on deep learning according to the present invention.

[0031] Figure 2 is a schematic diagram of a convolutional neural network model based on ResNet18 in an embodiment of the present invention.

[0032] Figure 3 is a schematic diagram of the scSE attention module in an embodiment of the present invention.

[0033] Figure 4 is a schematic diagram of the embedding of the scSE attention module and the convolutional neural network model based on ResNet18 in an embodiment of the present invention.

[0034] Figure 5 is the recognition accuracy of the scSE attention module at different positions in an embodiment of the present invention.

[0035] Figure 6 is the result of the ablation experiment in an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0036] It should be noted that, without conflict, the embodiments in the present invention and the features in the embodiments can be combined with each other.

[0037] In the description of the present invention, it should be understood that the orientation or positional relationship indicated by the terms "center", "longitudinal", "transverse", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc. is based on the orientation or positional relationship shown in the drawings. It is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be construed as a limitation to the present invention. In addition, the terms "first", "second", etc. are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the quantity of the technical features indicated. Thus, the features defined with "first", "second", etc. may explicitly or implicitly include one or more of such features. In the description of the present invention, unless otherwise specified, the meaning of "a plurality" is two or more.

[0038] In the description of the present invention, it should be noted that unless otherwise clearly specified and limited, the terms "installed", "connected", "connected" should be understood in a broad sense. For example, it may be a fixed connection, a detachable connection, or an integral connection; it may be a mechanical connection or an electrical connection; it may be directly connected or indirectly connected through an intermediate medium, and it may be the communication inside two elements. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood through specific circumstances.

[0039] The present invention provides a method for identifying and classifying drones based on deep learning, as Figure 1 shown, including the following steps:

[0040] Step 1: Obtain a one-dimensional millimeter-wave data set of the drone and convert it into a two-dimensional image using a data conversion method;

[0041] The one-dimensional millimeter-wave data set of the drone comes from a public data set, and the specific article is: "Analyzing Radar Cross Section Signatures of Diverse Drone Models at mmWave Frequencies"

[0042] The data conversion method used in Step 1 is the Gram Angle Field (GASF) method, and the specific steps are as follows:

[0043] Step 1-1: Assume there is a set of vectors V = [v1, v2,..., v n . The Gram matrix is the inner product matrix of every pair of vectors from V. Each element in the matrix <v i , v j> is the vector v i and v j is the vector product therebetween.

[0044]

[0045] Step 1-2: Use a Min-Max scaler to scale the original time series V data to [-1, 1];

[0046]

[0047] Step 1-3: Perform a polar coordinate system transformation on the obtained data to obtain the angle and radius corresponding to each data point;

[0048]

[0049]

[0050] Step 1-4: Use the sum-of-angles relationship and the difference-of-angles relationship to obtain the corresponding GASF diagram:

[0051] GASF = [cos(φ i + φ j )]

[0052] Step Two: Randomly divide the obtained data set into a training set, a validation set, and a test set according to a ratio of 7:2:1;

[0053] Step Three: Build a convolutional neural network model based on ResNet18. As Figure 4 shown, embed the attention mechanism module into the original ResNet18 network structure in parallel, and modify the number of neuron connections in the fully connected layer in the backbone network according to the number of categories in the data set;

[0054] It can be understood that the number of neuron connections in the fully connected layer of the original ResNet18 is 1000.

[0055] In this embodiment, the number of categories in the data set is 9. Therefore, the modified number of neuron connections is 9. Specifically, modify the OutputSize parameter of the fully connected layer, changing it from 1000 to 9.

[0056] Specifically, the scSE module includes a convolutional layer, a Sigmoid layer, a ReLU layer, an average pooling layer, and a global connection layer is added after the average pooling layer.

[0057] A further setting of the present invention is that the convolutional neural network model based on ResNet18 is composed of a preprocessing module, a feature extraction module, and a classifier module.

[0058] Specifically, as Figure 2 and Figure 3 shown, the preprocessing module sequentially includes a convolutional layer, a batch normalization layer, a ReLU layer, and a max pooling layer, so as to adjust the input image to a size and normalized pixel values that match the input of the model.

[0059] The feature extraction module includes four stages, each stage is respectively composed of two residual modules, and the second residual module in each stage is connected with an scSE module for feature extraction.

[0060] The classifier module sequentially includes an average pooling layer, a fully connected layer, a ReLU layer, a fully connected layer, and a Softmax layer.

[0061] The extracted features are flattened into a feature vector and classified through a fully connected layer. Finally, the Softmax function is used to convert the output into a probability distribution for each category. The expression of the Softmax function is as follows:

[0062]

[0063] where z j represents the original score of the j-th element in the input vector, N represents the dimension of the input vector, that is, the number of categories; e represents the base of the natural logarithm.

[0064] Step 4: Train the model on the training set according to the set hyperparameters. After the training is completed, a network model that can accurately classify drones is obtained;

[0065] The specific steps of Step 4 are as follows:

[0066] Step 4-1: Set the initial values of the four hyperparameters of the number of iterations, learning rate, batch size, and optimizer. In this embodiment, the number of iterations is set to 50, the learning rate is set to 0.0001, the batch size is set to 64, and the optimizer is set to Adam (Adaptive Moment Estimation).

[0067] Step 4-2: The network model selects the model with the highest recognition accuracy obtained in Step 3 as the final model.

[0068] Step 5: Realize the classification of different types of drones.

[0069] The present invention improves the original ResNet18 from two aspects, as Figure 5 and Figure 6As shown, it can be seen that after adding the scSE attention module and the fully connected layer, the recognition accuracy has been significantly improved, and while improving the network recognition accuracy, the number of parameters of the network model is not increased, enabling it to complete the recognition of UAV types well and being easily deployed to actual 5G base stations, making the design more reasonable and effective.

[0070] The above has described the embodiments of the present invention in detail in conjunction with the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, without departing from the principles and spirit of the present invention, various changes, modifications, substitutions, and variations to these embodiments including components still fall within the protection scope of the present invention.

Claims

1. A method for identifying and classifying drones based on deep learning, characterized in that, It includes the following steps: S1. Obtain the one-dimensional millimeter-wave data of the drone, convert the one-dimensional millimeter-wave data into a two-dimensional image using a data conversion method, and divide it into a training set, a validation set, and a test set; Step 1-1. Suppose there is a set of vectors V = [v1, v2,..., v n . The Gram matrix is the inner product matrix of every pair of vectors from V. Each element <v i , v j > in the matrix is the vector product between vector v i and v j : Step 1-2. Use a min-max scaler to scale the original time series X data to [-1, 1]; Step 1-3. Perform a polar coordinate system transformation on the obtained data to obtain the angle and radius corresponding to each data point; Step 1-4. Use the sum-angle relationship and the difference-angle relationship to obtain the corresponding GASF diagram: GASF = [cos(φ i + φ j )] S2. Build a convolutional neural network model based on ResNet18, S2-1. Use ResNet18 as the backbone network, and fuse the attention mechanism into the ResNet18 network in a parallel manner; S2-2. Modify the number of neuron connections in the fully connected layer of the backbone network according to the number of categories in the dataset; The attention mechanism is the scSE module; The scSE module includes a convolutional layer, a Sigmoid layer, a ReLU layer, an average pooling layer, and a global connection layer is added after the average pooling layer; The convolutional neural network model based on ResNet18 consists of a preprocessing module, a feature extraction module, and a classifier module; The preprocessing module sequentially includes a convolutional layer, a batch normalization layer, a ReLU layer, and a max pooling layer, so as to adjust the input image to a size and normalized pixel values that match the input of the model; The feature extraction module includes four stages, each stage is respectively composed of two residual modules, and the second residual module in the third and fourth stages is connected with an scSE module for feature extraction; The classifier module sequentially includes an average pooling layer, a fully connected layer, a ReLU layer, a fully connected layer, and a Softmax layer; S3. Train the model on the training set according to the set hyperparameters, and after the training is completed, obtain a network model that can accurately classify drones; S4. Apply the drone classification network model to classify drones.

2. The drone recognition and classification method based on deep learning according to claim 1, wherein, In step S3, the training method of the convolutional neural network model based on ResNet18 is: set the initial values of four hyperparameters, namely the number of iterations, the learning rate, the batch size, and the optimizer. The drone classification network model is trained under the four hyperparameters until the recognition accuracy on the training set no longer improves, and the losses of the training set and the validation set are close and tend to zero.

3. The method for identifying and classifying drones based on deep learning according to claim 2, wherein, In step S3, when training the convolutional neural network model based on ResNet18, it is trained by using a cross-entropy loss function, and the expression of the loss function is: where n is the number of categories, and y i is the true value, and y i ' is the network prediction value.

4. A method for identifying and classifying drones based on deep learning according to claim 1, characterized in that, In step S4, the specific method for classification is: S4-1. Image preprocessing: Adjust the input image to a size and normalized pixel values that match the input of the model; S4-2. Feature extraction: Perform feature extraction through a convolutional layer and an average pooling layer; S4-3. Classification by the classifier: Flatten the extracted features into a feature vector, perform classification through a fully connected layer, and finally use the Softmax function to convert the output into a probability distribution of each category. The expression of the Softmax function is as follows: Among them, z j represents the original score of the j-th element in the input vector, N represents the dimension of the input vector, that is, the number of categories; e represents the base of the natural logarithm.

5. A method for identifying and classifying drones based on deep learning according to claim 1, characterized in that, The training set, validation set, and test set are divided in a ratio of 7:2:1.

Images