Unmanned aerial vehicle signal identification method based on neural network

Abstract

The invention discloses an unmanned aerial vehicle signal identification method based on a neural network. The unmanned aerial vehicle signal is identified by combining an adaptive block compressed sensing (ABCS) technology and a convolutional neural network method, and the method comprises the following steps: firstly, carrying out adaptive block compressed sampling (ABCS) on a radio frequency signal of an unmanned aerial vehicle so as to retain key features while reducing the sampling data volume; secondly, carrying out normalization preprocessing on the collected signals; thirdly, radio frequency features are extracted and classified through a convolutional neural network, a T-Max-Avg self-defined pooling layer is integrated in the network, and therefore the classification accuracy and robustness of the model are effectively improved; and finally evaluating the model performance through a cross validation method. According to the method, high-precision unmanned aerial vehicle signal identification can be realized under the conditions of low sampling rate and limited computing resources, and the real-time detection capability is effectively improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of drone detection and identification, and in particular relates to a drone signal recognition method based on a neural network. Background Art

[0002] With the widespread use of drones in both civilian and military fields, the effective detection and identification of drones, particularly in complex electromagnetic environments, has become an unresolved challenge. Existing drone identification methods include those based on acoustics, vision, radar, and radio frequency signals, but these methods have limitations in practical applications. For example, acoustic methods are susceptible to interference from ambient noise, visual methods are ineffective in low-visibility conditions, and radar methods are susceptible to obstructions when detecting small drones. Therefore, drone detection methods using neural networks are attracting increasing attention due to their stability and effectiveness. Summary of the Invention

[0003] To address the shortcomings and deficiencies of existing technologies, this paper provides a neural network-based drone signal recognition method. By combining adaptive block compressed sensing (ABCS) technology with convolutional neural networks (CNNs), this paper aims to achieve efficient and accurate drone RF signal detection and classification.

[0004] The present invention combines the adaptive block compressed sensing (ABCS) technology with the convolutional neural network method. First, the RF signal of the drone is subjected to adaptive block compressed sampling (ABCS) to reduce the amount of sampled data while retaining key features. The collected signal is then normalized and feature extracted. The RF features are then extracted and classified using a convolutional neural network. The network integrates a custom T-Max-Avg pooling layer based on the threshold parameter T, thereby effectively improving the classification accuracy and robustness of the model. Finally, the model performance is evaluated by a cross-validation method. This method can achieve high-precision drone signal recognition under conditions of low sampling rate and limited computing resources, effectively improving real-time detection capabilities. The purpose of the present invention can be achieved through the following technical solutions.

[0005] A method for identifying drone signals based on a neural network comprises the following steps:

[0006] (1) Preprocessing of sampling and normalization of the UAV’s RF signal;

[0007] (2) A convolutional neural network is used to extract and classify radio frequency features; the convolutional neural network includes a convolution layer, a pooling layer, and a fully connected layer; wherein:

[0008] The convolution layer extracts features from the RF signal; the pooling layer is a T-Max-Avg custom pooling layer based on the threshold parameter T. It selects K pixels with the highest interaction value in each pooling area and decides whether to use the maximum value or weighted average of these pixels based on the threshold parameter T. The specific pooling calculation formula is:

[0009]

[0010] Among them, Y i represents the i-th largest pixel value in the input data, Y is the parameter that controls the output, and K is the number of high-value items;

[0011] The pooled features are input to the fully connected layer, and the category probability of the drone is output through the activation function.

[0012] In the present invention, in step (1), adaptive block compressed sensing sampling is performed on the RF signal of the drone; specifically as follows:

[0013] Step 1.1: Perform fast wavelet transform on the UAV’s RF signal s(t) to obtain a signal x(t) with high sparsity and low redundancy.

[0014] Step 1.2: Divide the entire signal into several small blocks, each of which is B×B in size, and vectorize it to form a one-dimensional signal block x i ,The process is expressed as:

[0015]

[0016] That is, the small block x i,j Each row is concatenated in sequence to form a column of length B 2 A one-dimensional vector of

[0017] Step 1.3: For each signal block, construct the adaptive sampling matrix φ i , whose size is n i ×B 2 , where n i Represents the number of adaptive measurements for each block; in order to adapt to the information content of different signal blocks, the signal sampling weight is dynamically adjusted by constructing a weighting matrix W. The weighting matrix W is a diagonal matrix, in which the diagonal elements are weighting coefficients w fc :

[0018]

[0019] Weighting coefficient w fc It is calculated using the statistical properties of the signal, including the mean and variance of all frequency components in the signal block:

[0020]

[0021] Among them, μ fc and σ fc Respectively represent the mean and variance of the corresponding frequency components in the signal block;

[0022] Step 1.4: Determine the number of measurements n for each block based on the energy characteristics of the signal block i The energy E(i) of a signal block is calculated by summing the squares of the amplitudes |C(i,j)| of all its frequency components:

[0023]

[0024] The number of measurements assigned to this signal block is n i for:

[0025]

[0026] Where N is the total number of measurements;

[0027] Step 1.5: In the reconstruction phase, the sampled data is processed using the inverse weighting matrix to restore the original amplitude relationship of the signal:

[0028]

[0029] Among them, α i is the inverse of the weighting coefficient.

[0030] In the present invention, in step (2), the convolution layer uses a one-dimensional convolutional neural network technology, using asymmetric convolution kernels of different sizes to extract local features of the signal at multiple scales. The convolution operation of each convolution layer can be expressed as:

[0031]

[0032] Among them, w i is the weight of the convolution kernel, u i is the input signal, and b is the bias term. Through convolution operations of different scales, local features in the signal can be effectively extracted.

[0033] In the present invention, in step (2), the convolution layer uses a convolution operation to process the two input channels corresponding to the low-frequency and high-frequency signals, and uses a convolution kernel of 64 output channels; the output of the convolution layer is processed by the activation function ELU.

[0034] In the present invention, in step (2), the pooling layer combines the maximum value and the average value pooling method to improve the flexibility and accuracy of feature selection.

[0035] In the present invention, in step (2), the calculation formula of the activation function is:

[0036]

[0037] Among them, s j represents the output score of the fully connected layer, k is the number of categories, and p j is the probability of the jth class;

[0038] The activation function can be used to convert the output of the model into a probability distribution of each category to make the final classification decision.

[0039] Furthermore, the present invention adopts a five-fold cross-validation method to randomly divide the data set into five subsets, using one subset as the validation set each time and the other subsets as the training set, and repeats this cycle five times to finally calculate the average classification accuracy of the model.

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] (1) Efficient signal sampling and processing: Through the adaptive block compressed sensing method, the amount of signal sampling data is effectively reduced while retaining the main characteristic information of the signal, reducing the computational burden of subsequent data processing.

[0042] (2) By combining convolutional neural networks with the T-Max-Avg custom pooling layer, accurate feature extraction and aggregation are achieved, effectively improving the classification performance and robustness of the model.

[0043] (3) The recognition method of the present invention can achieve high-precision detection and classification of UAV RF signals under low sampling rate and limited computing resources, and is suitable for complex electromagnetic environments and multiple interference conditions.

[0044] (4) The method of the present invention can achieve high-precision drone signal recognition under the conditions of low sampling rate and limited computing resources, effectively improving the real-time detection capability. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a flow chart of the identification method of the present invention.

[0046] Figure 2 : a) Illustration of the output during the five-fold cross-validation experiment; b) and the output during the model training process.

[0047] Figure 3 1 is a diagram of a four-classification iterative process in an embodiment of the present invention.

[0048] Figure 4 is the confusion matrix of the classification process in this invention. DETAILED DESCRIPTION

[0049] The present invention will be further described below with reference to the accompanying drawings and embodiments, but the scope of protection of the present invention should not be limited thereto.

[0050] In order to achieve the above-mentioned purpose, the present invention provides a method for identifying drone signals based on a neural network. Figure 1 As shown, it includes the following steps:

[0051] Step 1: Adaptive Compressed Sensing Sampling

[0052] Adaptive block-level compressed sensing sampling is performed on the RF signal of the drone. The present invention adopts an improved compressed sensing method, called block-level adaptive sampling technology.

[0053] In the present invention, in step 1, the specific method of the block-level adaptive sampling technology is:

[0054] Step 1.1: Perform fast wavelet transform on the UAV’s RF signal s(t) to obtain a signal x(t) with high sparsity and low redundancy.

[0055] Step 1.2: Divide the entire signal into several small blocks, each of which is B×B in size, and vectorize it to form a one-dimensional signal block x i ,The process can be expressed as:

[0056]

[0057] That is, the small block x i,j Each row is concatenated in sequence to form a column of length B 2 A one-dimensional vector of .

[0058] Step 1.3: For each signal block, construct the adaptive sampling matrix φ i , whose size is n i ×B 2 , where n i Represents the number of adaptive measurements for each block. To adapt to the information content of different signal blocks, the present invention implements dynamic adjustment of signal sampling weights by constructing a weighting matrix W. The weighting matrix W is a diagonal matrix, where the diagonal elements are weighting coefficients:

[0059]

[0060] The weighting coefficients are calculated based on the statistical properties of the signal, including the mean and variance of all frequency components in the signal block:

[0061]

[0062] Among them, μ fc and σ fc They represent the mean and variance of the corresponding frequency components in the signal block respectively.

[0063] Step 1.4: The present invention determines the number of measurements n for each block based on the energy characteristics of the signal block.i The energy E(i) of a signal block is calculated by summing the squares of the amplitudes of all its frequency components:

[0064]

[0065] Then the number of samples n allocated to this signal block is i for:

[0066]

[0067] Where N is the total number of measurements.

[0068] Step 1.5: In the reconstruction phase, the sampled data is processed using the inverse weighting matrix to restore the original amplitude relationship of the signal:

[0069]

[0070] Among them, α i is the inverse of the weighting coefficient.

[0071] Step 2: Signal normalization

[0072] The collected signals are normalized to standardize the signal values within a fixed range for subsequent processing.

[0073] Step 3: Convolutional Neural Network Processing

[0074] This paper adopts the technology of one-dimensional convolutional neural network and uses asymmetric convolution kernels of different sizes to extract local features of the signal at multiple scales. The convolution operation of each convolution layer can be expressed as:

[0075]

[0076] Among them, w i is the weight of the convolution kernel, u i is the input signal, and b is the bias term. Through convolution operations of different scales, local features in the signal can be effectively extracted.

[0077] Step 4: Pooling layer optimization and classification

[0078] The present invention adopts a T-Max-Avg pooling layer based on a threshold parameter T, combining maximum and average pooling methods to improve the flexibility and accuracy of feature selection.

[0079] In the present invention, in step 4, the specific method of using the pooling layer for optimization and classification is:

[0080] Step 4.1: Select K pixels with the highest interaction value in each pooling area, and decide whether to use the maximum value or weighted average of these pixels based on the threshold parameter T. The specific pooling calculation formula is:

[0081]

[0082] Among them, Y i It represents the i-th largest pixel value in the input data, T is the parameter that controls the output, and K is the number of high-value items.

[0083] Step 4.2: The pooled features are input to the fully connected layer, and the activation function is used to output the drone category probability. The activation function is calculated as:

[0084]

[0085] Among them, s j represents the output score of the fully connected layer, k is the number of categories, and p j is the probability of the jth class. Through the activation function, the output of the model can be converted into the probability distribution of each class to make the final classification decision.

[0086] Step 5: Model Evaluation

[0087] A five-fold cross-validation method was used to randomly divide the dataset into five subsets. One of the subsets was used as the validation set each time, and the other subsets were used as the training set. This cycle was repeated five times, and the average classification accuracy of the model was finally calculated.

[0088] The technical solution of the present invention is described below with reference to specific embodiments.

[0089] Example 1

[0090] Step 1: Load the raw data from the DroneRF dataset. The data is loaded into three different array formats, representing the feature array, label array, and target label array.

[0091] Step 2: According to steps 1 and 2, the collected signal is preprocessed and normalized using the adaptive block compressed sensing method. The specific steps are as follows:

[0092] (1) Fast wavelet transform is used to denoise the collected 25ms UAV flight data to remove noise and interference in the RF signal and improve the sparsity of the signal.

[0093] (2) Adaptive block compressed sensing is used to sample and process the signal. By dynamically selecting the size and sampling rate of the sampling block, sampling is performed based on the importance of the signal to ensure that key features are effectively retained.

[0094] (3) Normalize the signal to ensure that the signal value is distributed within a certain range to facilitate subsequent processing and analysis.

[0095] Step 3: Further process the signal using convolutional neural networks and pooling layer optimization methods according to the methods in steps 3 and 4. In this stage, this embodiment will use multiple convolutional layers, pooling layers, and a custom T-Max-AvgPool1D pooling layer to help the network better learn the temporal characteristics of the input signal. The specific steps are as follows:

[0096] (1) First, a convolutional layer extracts low-level features from the signal. In this example, a convolution layer is used to process two input channels (low-frequency and high-frequency signals) and a convolution kernel with 64 output channels is used, meaning that 64 different feature maps are output after the convolution operation. The output of the convolution layer is processed by the activation function ELU to improve nonlinear modeling capabilities.

[0097] (2) In order to achieve the purpose of further optimizing signal feature extraction in the present invention, a customized T-Max-AvgPool1D pooling layer is used in this example. This layer combines the characteristics of maximum pooling and average pooling, and performs pooling operation by selecting the first K values. Specifically, in this example, the maximum value and average value within each pooling window are calculated, and the output is determined by the threshold T. The pooled result will further compress the feature dimension and retain key information. During the forward propagation of the network, in this example, features are first extracted through the convolutional layer, and then the features are downsampled through the pooling layer.

[0098] (3) Furthermore, in this example, four grouped convolutions are used to process deeper features in the signal. The grouped convolution layer divides the channels of the input signal into eight groups, each of which is processed with its own sub-convolution kernel to extract different signal patterns.

[0099] (4) In this example, the multi-level and multi-scale pooling outputs are concatenated to form a complete feature vector. The final classification is performed through the fully connected layer.

[0100] Step 4: In this example, initialize the network model and set hyperparameters such as learning rate, batch size, and number of training rounds. Configure the training process using the optimizer and loss function. Use 5-fold cross-validation to train the model and evaluate the performance of the model on data from different folds.

[0101] During model training, we employed 5-fold cross-validation to fully evaluate model performance, and used the cross-entropy loss function as the primary optimization metric. Adam was selected as the model optimizer, with an initial learning rate of 0.001. By gradually adjusting the learning rate, we improved the model's convergence speed and performance. Figure 2 a) and Figure 2b) are the output graphs during the five-fold cross-validation experiment and model training process, respectively.

[0102] In this specific embodiment, the public DroneRF dataset was used. The DroneRF dataset is divided into four main categories: None Drone, Bebop Drone, AR Drone, and Phantom Drone. The specific RF signals corresponding to each category are labeled as non-drone, drone 1, drone 2, and drone 3. For the drone presence problem, i.e., identifying whether a signal is a drone, the present invention achieved an accuracy rate of 96%. For specific drone models, the present invention achieved an accuracy rate of 95.15%, a significant improvement. Figure 3 is the training curve of the model. The training in each fold shows a clear convergence trend. During the training process, the initial high loss value is about 1.37, which gradually decreases as the training rounds proceed, and finally drops to about 0.017 at the tenth training cycle, showing good convergence of the model. The gradual decrease and final stabilization of the training loss indicate that the model has no obvious overfitting or underfitting phenomenon during the training stage. The model scoring table is shown in Table 1, which has a good level in terms of F1 score and running time. The confusion matrix of the classification process is shown in Figure 4 As shown, it can be seen that during the data testing process, the accuracy of identifying the four categories of non-UAV, UAV 1, UAV 2 and UAV 3 is relatively stable.

[0103] Table 1

[0104]

[0105] Although the present invention has been described in detail through the above preferred embodiments, it should be understood that the above description is not intended to limit the present invention. After reading the above description, various modifications and substitutions of the present invention will become apparent to those skilled in the art. Therefore, the scope of protection of the present invention should be defined by the appended claims.

Claims

1. A method for identifying drone signals based on neural networks, characterized in that: The following steps are involved: (1) Preprocessing of sampling and normalization of the UAV’s RF signal; (2) A convolutional neural network is used to extract and classify radio frequency features; the convolutional neural network includes a convolution layer, a pooling layer, and a fully connected layer; wherein: The convolution layer extracts features from the RF signal; the pooling layer is a T-Max-Avg custom pooling layer based on the threshold parameter T. It selects K pixels with the highest interaction value in each pooling area and decides whether to use the maximum value or weighted average of these pixels based on the threshold parameter T. The specific pooling calculation formula is: Among them, Y i represents the i-th largest pixel value in the input data, T is the parameter that controls the output, and K is the number of high-value items; The pooled features are input to the fully connected layer, and the category probability of the drone is output through the activation function.

2. The drone signal recognition method according to claim 1, characterized in that: In step (1), adaptive block compressed sensing sampling is performed on the RF signal of the drone; the details are as follows: Step 1.1: Perform fast wavelet transform on the UAV’s RF signal s(t) to obtain a signal x(t) with high sparsity and low redundancy. Step 1.2: Divide the entire signal into several small blocks, each of which is B×B in size, and vectorize it to form a one-dimensional signal block x i ,The process is expressed as: That is, the small block x i,j Each row is concatenated in sequence to form a column of length B 2 A one-dimensional vector of Step 1.3: For each signal block, construct the adaptive sampling matrix φ i , whose size is n i ×B 2 , where n i Represents the number of adaptive measurements for each block; in order to adapt to the information content of different signal blocks, the signal sampling weight is dynamically adjusted by constructing a weighting matrix W. The weighting matrix W is a diagonal matrix, in which the diagonal elements are weighting coefficients w fc : W=diag(w1,w2,...,w B2 ) Weighting coefficient w fc It is calculated using the statistical properties of the signal, including the mean and variance of all frequency components in the signal block: Among them, μ fc and σ fc Respectively represent the mean and variance of the corresponding frequency components in the signal block; Step 1.4: Determine the number of measurements n for each block based on the energy characteristics of the signal block i The energy E(i) of a signal block is calculated by summing the squares of the amplitudes |C(i,j)| of all its frequency components: The number of measurements assigned to this signal block is n i for: Where N is the total number of measurements; Step 1.5: In the reconstruction phase, the sampled data is processed using the inverse weighting matrix to restore the original amplitude relationship of the signal: Among them, α i is the inverse of the weighting coefficient.

3. The drone signal recognition method according to claim 1, characterized in that: In step (2), the convolution layer uses a convolution operation to process the two input channels corresponding to the low-frequency and high-frequency signals, and uses a convolution kernel with 64 output channels; the output of the convolution layer is processed by the activation function ELU.

4. The drone signal recognition method according to claim 1, wherein: In step (2), the activation function is calculated as: Among them, s j represents the output score of the fully connected layer, k is the number of categories, and p j is the probability of the jth class.

Images