A micro unmanned aerial vehicle detection method based on frequency domain and time domain feature fusion

Abstract

The application provides a micro unmanned aerial vehicle detection method based on frequency domain and time domain feature fusion, audio data of a no-fly zone is collected, time domain features and frequency domain features are extracted, the extracted time domain features and frequency domain features are converted to the same dimension, the frequency domain features are used to modulate the time domain features, and finally classification prediction is performed to determine whether an unmanned aerial vehicle appears or not. The application collects audio data of a no-fly zone, extracts time domain features and frequency domain features respectively, performs classification prediction after modulating the time domain features by the frequency domain features, and determines whether an unmanned aerial vehicle appears or not. The application combines time domain and frequency domain features (T-FCE), solves the problem of low unmanned aerial vehicle detection accuracy in a general scene, can improve unmanned aerial vehicle detection accuracy in a general scene, and ensures accurate output of unmanned aerial vehicle detection.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) detection, and more specifically to a method for detecting micro unmanned aerial vehicles (UAVs). Background Technology

[0002] Currently, drones are playing an increasingly important role in a growing number of fields, such as agriculture, security, reconnaissance, and aerial photography. As the cost of drones gradually decreases, more and more unregulated micro-drones are appearing in areas where they shouldn't be, such as airports, disrupting aircraft takeoffs and landings. In response to this phenomenon, numerous scholars have conducted research on counter-micro-drone technologies, including micro-drone reconnaissance, identification, location, and countermeasures. Micro-drone reconnaissance is a crucial foundation for counter-micro-drone technology research.

[0003] Micro drones have advantages such as low flight altitude and small size, while existing detection methods such as radar have problems such as small cross-section, resulting in poor detection effect of micro drones. Therefore, how to achieve accurate detection of micro drones is an urgent problem to be solved in the field of counter-micro drones.

[0004] Compared to other detection methods (including visual, thermal imaging, microwave sensors, etc.), audio information using microphones can be used for reconnaissance, identification, and location operations other than attack. Moreover, compared to visual information, audio monitoring requires less computing power, is less expensive, and is easier to deploy.

[0005] However, common frequency domain feature extraction methods suffer from problems such as feature redundancy and fixed resolution, while common time domain feature extraction methods struggle with accuracy. Therefore, in general scenarios, achieving accurate UAV detection is difficult. Summary of the Invention

[0006] Purpose of the invention: The purpose of this invention is to address the shortcomings of existing technologies by providing a micro-UAV detection method based on frequency domain and time domain feature fusion that can achieve high accuracy in general scenarios.

[0007] Technical solution: The present invention discloses a micro UAV detection method based on frequency domain and time domain feature fusion. It collects audio data from no-fly zones and extracts time domain and frequency domain features. After converting the extracted time domain and frequency domain features to the same dimension, the time domain features are modulated with frequency domain features. Finally, classification and prediction are performed to determine whether a UAV has appeared.

[0008] Furthermore, the audio data is obtained by reading data from a single microphone deployed in the no-fly zone via FPQA.

[0009] Furthermore, temporal features are extracted using a one-dimensional convolutional network neural model and a temporal attention mechanism.

[0010] Furthermore, extracting temporal features involves: first, normalizing the audio data; then, extracting nonlinear features and filtering them; and finally, using a temporal attention mechanism to globally model the filtered features.

[0011] Furthermore, extracting frequency domain features includes: after extracting frequency domain interval features through wavelet packet transform, multiplying them by the maximum energy value of each frequency domain interval to obtain the frequency domain features.

[0012] Furthermore, the wavelet packet transform uses the db4 wavelet function.

[0013] Furthermore, a single fully connected layer of a one-dimensional convolutional network neural model is used to complete the classification prediction of feature vectors.

[0014] Beneficial effects: This invention collects audio data from no-fly zones, extracts time-domain and frequency-domain features respectively, and performs classification and prediction by modulating the time-domain features with the frequency-domain features to determine whether a drone is present. This invention combines time-domain and frequency-domain features (T-FCE) to solve the problem of low drone detection accuracy in general scenarios, improving drone detection accuracy in general scenarios and ensuring accurate drone detection output. Attached Figure Description

[0015] Figure 1 This is a flowchart of the method of the present invention;

[0016] Figure 2 This is a schematic diagram of feature extraction in the time domain;

[0017] Figure 3 This is the effect of the method proposed in this invention on a single domain. Detailed Implementation

[0018] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings, but the scope of protection of the present invention is not limited to the embodiments described.

[0019] like Figure 1 As shown, a micro-UAV detection method based on frequency domain and time domain feature fusion is described. Figure 1 As shown, it includes the following steps:

[0020] Step 1: Use FPQA to read data from a single microphone deployed in a no-fly zone, encoding packets at fixed time intervals. Assume the collected time-series data is... ,in A schematic diagram of the feature extraction method is shown below. Figure 1 As shown.

[0021] Step 2: Extract temporal features using a one-dimensional convolutional network neural module (1D-CNN) and a temporal attention mechanism. Among these... Figure 1The 1D-CNN module includes 1D-CNNLayers, BatchNorm, etc., each layer achieving a different purpose, such as... Figure 2 As shown:

[0022] First, the input time series data... Normalization yields , will X norm One-dimensional convolutional features F are extracted through the Conv1d layer. Conv1d :

[0023]

[0024] Conv1d can be represented as:

[0025]

[0026] In the formula, Indicates the first The convolutional kernel at the _th ... Layer weights Representing the The first layer The bias of each convolution kernel Representing the Layer One data block, Indicates the first The first layer The input of a data block; here express .

[0027] F Conv1d BatchNorm() makes It is identically distributed with X, and then the ReLU function is used to transform F. Conv1d Extracting nonlinear features yields the nonlinear feature F. nolinear :

[0028]

[0029] BatchNorm(∙) can be represented as:

[0030]

[0031] In the formula, x represents the input data, γ represents the degrees of freedom, β represents the displacement parameter, and β and γ are trainable parameters. E(x) represents the mean of the mini-batch samples, and Var(x) represents the standard deviation of the mini-batch samples.

[0032] ReLU(∙) can be represented as:

[0033]

[0034] The max() function returns the maximum value in the input.

[0035] At this point, there are still some unnecessary disturbances, which are addressed through the Dropout layer and the Pooling layer. After performing simple filtering, the features of the one-dimensional convolutional module are obtained:

[0036]

[0037] MaxPooling can be represented as

[0038]

[0039] In the formula, N represents the number of samples, i represents the sample ordinal number, C represents the channel dimension, j represents the channel dimension ordinal number, stride represents the step size of the sliding window, k represents the number of moves, and m represents the padding.

[0040] Dropout layers allow neurons to stop working with a certain probability. Generally, the forward pass calculation formula for the network is as follows:

[0041]

[0042] r l It follows a Bernoulli binomial distribution with probability p, r l The generated vector consists of 0s and 1s, where y is the neuron's output value and l is the neuron's ordinal number. By setting the activation value to 0, some nodes in the network are rendered inactive.

[0043] Next, we will analyze the one-dimensional convolutional module features extracted from the 1D-CNN module. Divide into T F local This represents the local features extracted by the 1D-CNN module. Then, a temporal attention mechanism is used to apply this to all the features. Perform global modeling:

[0044]

[0045] The specific definition of Attention(∙) is as follows:

[0046]

[0047] In the formula, Z represents the input data, and the values ​​of Q (query), K (key), and V (value) are determined by... Trained with hyperparameters, Also called the self-attention score. The self-attention score is divided by... It is mainly used for normalization. This is called the normalized exponent score, and the final product is the normalized exponent score multiplied by the value. This process is as follows: Figure 1 The temporal self-attention mechanism part of the text.

[0048] After passing through the temporal attention mechanism, more useful information is preserved. At this point, the temporal feature extraction is complete. The hyperparameters of the temporal feature extraction network are generally selected based on empirical values; Table 1 shows one possible method for setting these values. The values ​​of the hyperparameters are not unique.

[0049] Table 1

[0050] Number of convolutional network layers 3 Activation function ReLU Batch size 64 round 100 kernel size 9-5-5 Learning rate 0.05 stride size 2-2-2 Learning rate strategy OneCycle Pooling method Max-Avg-Avg Optimizer Adam Drop rate 0.5-0.5-0.5 Parameter initialization He

[0051] Step 3: Perform frequency domain feature extraction. Addressing the issues of fixed resolution and information redundancy in traditional frequency domain feature extraction methods, this invention utilizes the zoom characteristics of wavelet packet transform and employs the db4 wavelet function to extract 2... k ( The wavelet coefficients of the interval segment are used to obtain the characteristic F of the wavelet packet transform. wpt :

[0052]

[0053] In the formula, f wpt (∙) represents the wavelet packet transform operation. Then, a filter is constructed to perform high-pass filtering on the transformed features to obtain the frequency domain features F. frequency :

[0054]

[0055] In the formula, A represents the maximum energy value of the frequency domain interval corresponding to each wavelet coefficient.

[0056] Step 4: First, Become with Same dimensions:

[0057]

[0058] in, It is a dimension transformation matrix, and then the sigmoid(∙) function is used to transform it. Become a mask Afterwards, use modulation Finally, a single fully connected LinearLayer layer is sufficient to complete the classification prediction:

[0059]

[0060] An output of 1 indicates the presence of a drone, while an output of 0 indicates the absence of a drone.

[0061] The microphone used in this embodiment is SPH0641LU4H-1. To demonstrate the effectiveness of the invention, comparative experiments were conducted in both the time and frequency domains. The experimental results are as follows: Figure 3 As shown. The evaluation metrics are accuracy, recall, precision, and F1 score, and the specific calculation formulas are as follows:

[0062]

[0063]

[0064]

[0065] (18)

[0066] In this context, TP (True Positive) refers to the number of positive samples correctly identified when the classifier predicts a positive result; FP (False Positive) refers to the number of negative samples falsely reported when the classifier predicts a positive result; TN (True Negative) refers to the number of negative samples correctly identified when the classifier predicts a negative result; and FN (False Negative) refers to the number of positive samples missed when the classifier predicts a negative result. Positive samples indicate the presence of drones; negative samples indicate the absence of drones.

[0067] Experimental results show that fusing time-domain and frequency-domain features is effective in increasing the accuracy of UAV detection.

[0068] To demonstrate the effectiveness of this invention, comparative experimental results are presented in a general scenario, comparing the method of this invention with five other UAV detection methods. The results are shown in Table 2.

[0069] Table 2

[0070] PANNS_CNN6 0.93 0.98 0.91 0.95 TDNN 0.94 0.98 0.92 0.95 ResNetSE 0.94 1.0 0.92 0.96 Cheng's Model 0.95 0.99 0.93 0.96 WPT&SVM 0.88 0.69 0.9 0.78 T-FCE 0.98 1.0 0.93 0.96

[0071] In the table, PANNS_CNN6 uses Mel frequency cepstral coefficients as input, TDNN and ResNetSE use Mel spectrum as input; Cheng's Model uses spectrogram as input; WPT&SVM uses wavelet packet transform to transform the input signal X and then uses support vector machine for classification.

[0072] As can be seen from the table, the method of this invention outperforms other methods in many aspects. Among them, WPT&SVM performed the worst because it extracted too few features. The method of this invention solves the shortcomings of other methods that apply short-time Fourier transform, and can find time-domain features from a large amount of fluctuating time-domain data. Therefore, it can achieve high-precision detection of UAVs in general environments.

[0073] As described above, although the invention has been shown and described with reference to specific preferred embodiments, it should not be construed as limiting the invention itself. Various changes in form and detail may be made without departing from the spirit and scope of the invention as defined in the appended claims.

Claims

1. A micro unmanned aerial vehicle (UAV) detection method based on frequency domain and time domain feature fusion, characterized in that, Collect audio data from the no-fly zone; Extracting temporal features: Local temporal features of the audio data are extracted using a one-dimensional convolutional network neural model, and the extracted local temporal features are globally modeled using a self-attention mechanism to obtain temporal features; Frequency domain feature extraction: After extracting frequency domain interval features through wavelet packet transform, the maximum energy value of each frequency interval is multiplied by the corresponding feature, and a filter is constructed to perform high-pass filtering on the transformed features to obtain the frequency domain features; After converting the extracted time-domain and frequency-domain features to the same dimension, the frequency-domain features are converted into a modulation mask, and the time-domain features are modulated using the modulation mask. Finally, classification and prediction are performed to determine whether a drone has appeared.

2. The micro unmanned aerial vehicle (UAV) detection method based on frequency domain and time domain feature fusion according to claim 1, characterized in that, The audio data is obtained by reading data from a single microphone deployed in a no-fly zone via an FPGA.

3. The micro unmanned aerial vehicle (UAV) detection method based on frequency domain and time domain feature fusion according to claim 1, characterized in that, Extracting temporal features involves: first, normalizing the audio data; then, extracting nonlinear features and filtering them; and finally, using a temporal attention mechanism to globally model the filtered features.

4. The micro unmanned aerial vehicle (UAV) detection method based on frequency domain and time domain feature fusion according to claim 1, characterized in that, The wavelet packet transform uses the db4 wavelet function.

5. The micro unmanned aerial vehicle (UAV) detection method based on frequency domain and time domain feature fusion according to claim 1, characterized in that, The classification and prediction of feature vectors are accomplished using a single fully connected layer in a one-dimensional convolutional network neural model.

Images