Unmanned aerial vehicle I / Q signal classification method based on pyramid frequency domain Transform and complex convolution

Abstract

The invention provides an unmanned aerial vehicle I / Q signal classification method based on pyramid frequency domain Transform and complex convolution, and the method comprises the steps: 1, carrying out the linear calculation of an inputted complex signal feature vector xa through employing learnable weight matrixes WQ, WK, and WV, obtaining the complex matrix input Q, K, and V of an attention mechanism, and introducing the Fourier transform to the Q, K, and V; step 2, respectively calculating a phase and an amplitude for a query matrix Q in the Transform, and interacting and integrating frequency characteristics of the phase and the amplitude by using complex 1 * 1 convolution; step 3, introducing a user-defined convolution kernel into the K in the Transform so as to extract signal features of different stages; and 4, calculating the attention obtained in the step 2 and the step 3 by using an attention score formula, inputting an obtained result into a classification layer for mapping, and obtaining a final classification result. By using the global feature extraction capability of Transform and the characteristic that the signal features in the frequency domain are more significant, the unmanned aerial vehicle I / Q signal features can be effectively extracted, and high-precision classification and identification of the unmanned aerial vehicle signals are realized.

Description

Technical Field

[0001] The present invention belongs to the field of drone signal classification and recognition, and in particular relates to a drone I / Q signal classification method based on pyramid frequency domain Transformer and complex convolution. Background Art

[0002] Drones were originally used mainly in the military, but now their civilian applications have penetrated into daily life, covering agriculture, disaster monitoring, material transportation, security and other fields. Although this widespread use is convenient, it also brings potential risks such as privacy violations and illegal activities, and effective detection of illegal operations is urgently needed. To this end, it is particularly important to accurately and efficiently classify and identify drone signals.

[0003] In recent years, drone detection technology is mainly divided into traditional methods and deep learning methods. Traditional methods include acoustic, radio frequency, radar and visual detection: acoustic detection matches voiceprint features, but the accuracy decreases in noisy environments; radio frequency detection matches drone communication protocols with known radio frequency features to provide effective detection capabilities, but it will be interfered by other radio frequency sources, resulting in reduced performance, and the communication protocol of drones is not open to the public, so it may cause privacy and security issues related to drones; radar detection uses the principle of the drone body reflecting the electromagnetic waves emitted by the radar system to detect drones, but it cannot detect small drones. Visual detection relies on images and video data captured by the camera, and performs real-time or offline detection and classification through computer vision and image processing algorithms. It is suitable for a variety of lighting and environmental conditions, but will be affected by weather conditions. Although these methods have their own advantages, their limitations also need to be further optimized to improve the effectiveness and safety of drone detection.

[0004] In order to break free from the functional shortcomings of traditional methods, many deep learning algorithms and feature extraction methods are used in drone classification problems, because they do not require enough information as prior knowledge, and rely on neural networks to autonomously learn deep abstract representations of drone signals to achieve feature extraction. However, traditional deep learning methods are limited to local feature representations. Considering that most drone signals are long sequences, it is not enough to use them alone to identify the overall characteristics of drone signals. This limitation will affect the classification effect of most drone signal data sets. Summary of the invention

[0005] The purpose of the present invention is to provide a method for classifying drone I / Q signals based on pyramid frequency domain Transformer and complex convolution. This method improves the traditional Transformer model and can efficiently extract the unique features of drone signals by mining the frequency domain features of phase and amplitude, thereby improving the accuracy of classification and recognition and quickly identifying the drone model corresponding to the signal. To achieve the above purpose, the following scheme is adopted:

[0006] 1. In order to capture the global characteristics of signals in long sequence signal datasets, the proposed framework contains a pyramid frequency domain Transformer layer at each stage.

[0007] 2. The pyramid frequency domain Transformer layer used in this method can effectively extract relevant information of the real and imaginary parts of the complex signal. It is suitable for processing complex signals and exhibits excellent performance.

[0008] 3. This method mines important features by introducing Fourier transform (FFT) and extending it to the frequency domain, thereby capturing the key information in the signal more intuitively.

[0009] 4. In order to observe signal features of different sizes from multiple scales, this method sets the model framework into a pyramid structure. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 This is the improved pyramid frequency domain Transformer module diagram;

[0011] Figure 2 It is the process of complex 1×1 convolution;

[0012] Figure 3 This is a network structure diagram for signal classification based on pyramid frequency domain Transformer and complex convolution; DETAILED DESCRIPTION

[0013] The following is a more detailed description of the method for classifying drone I / Q signals based on pyramid frequency domain transformers of the present invention in conjunction with a schematic diagram, which shows a preferred embodiment of the present invention. It should be understood that those skilled in the art can modify the present invention described herein and still achieve the advantageous effects of the present invention. Therefore, the following description should be understood as being widely known to those skilled in the art and not as a limitation of the present invention.

[0014] like Figures 1 to 3 ,The drone I / Q signal classification system based on pyramid frequency domain Transformer and complex convolution is based on the classic Transformer Encoder framework. The core self-attention mechanism of Transformer can establish the correlation between different positions and has the ability of global vision. The signal features analyzed in the frequency domain will be more obvious than the signal correlation features analyzed in the time domain. The present invention combines the advantages of both, can efficiently extract the overall features of drone signals, and can also bring better interpretability to the entire model.

[0015] The present invention uses a pyramid structure to observe signal features of different sizes at multiple scales. We set the entire model to four stages, each of which includes a pyramid frequency domain Transformer layer, and the size of the drone signal in each stage is half the size of the drone signal in the previous stage. The following embodiments of the present invention are described based on one stage.

[0016] The present invention first combines the original complex I / Q signal with the learnable weight matrix W Q , W K , W V Perform linear operations to obtain complex matrices Q, K, and V, and introduce Fourier transforms into Q, K, and V respectively.

[0017] Secondly, the amplitude and phase corresponding to the query vector Q matrix are calculated respectively. The frequency domain characteristics of the signal are mainly studied from the two angles of signal amplitude and phase. The complex 1×1 convolution is used to assign different weights to different frequency components of the amplitude and phase. Through the learnable weight parameters, the model can emphasize the information that needs to be focused on, and can efficiently extract the frequency domain characteristics of the signal, thereby improving the model's ability to identify key signal information.

[0018] Then the present invention introduces a custom convolution kernel to the key matrix K, and the size of the convolution kernel is obtained by debugging step by step through experiments.

[0019] Then, the attention score obtained by the attention score formula of the results of the first two parts is recorded as ATTN. Finally, ATTN is input into the classification layer for mapping to obtain the final classification result.

[0020] The methods for drone signal classification based on pyramid frequency domain Transformer include:

[0021] Step 1: Learnable weight matrix W Q , W K , W V For the input I / Q signal x a Perform linear calculations to obtain the complex matrix inputs Q, K, and V of the attention mechanism, and introduce Fourier transforms to Q, K, and V respectively.

[0022] Step 11: Learnable weight matrix W Q , W K , W V For the input I / Q signal x a Perform linear calculations.

[0023] For a general complex signal, let S(t) be used, it can be expressed as the sum of an in-phase component and an orthogonal component according to formula (1):

[0024] S(t)=A(t)cos(θ(t))+B(t)sin(θ(t)) (1)

[0025] Among them, A(t) and B(t) are functions of the amplitude of the signal changing with time, and θ(t) is a function of the phase of the signal changing with time. The in-phase component I is A(t)cos(θ(t)), and the quadrature component Q is B(t)sin(θ(t)). The phases of I and Q differ by 90°. The I / Q signal data format is a complex vector I+iQ composed of the in-phase component I and the quadrature component Q.

[0026] The complex matrices Q, K, and V can be obtained by formula (2):

[0027] Q = x a ×W Q

[0028] K=x a ×W K (2)

[0029] V = x a ×W V

[0030] Step 12: Introduce Fourier transform to Q, K, and V respectively, and extend Q, K, and V to the frequency domain. The calculation expression of Fourier transform is as shown in formula (3).

[0031]

[0032]

[0033] Where F(Q), F(K), and F(V) represent the frequency domain signals of Q, K, and V after Fourier transform, f(Q), f(K), and f(V) represent the original time domain signals, u is the length of the time series, and j is an imaginary number.

[0034] Step 2: Calculate the phase and amplitude x corresponding to the query vector Q matrix pha With x mag The calculation expressions of phase and amplitude are shown in formula (4).

[0035] x pha =arctan(x Q / x I )

[0036]

[0037] where x l is the in-phase component of the signal, x Q are the orthogonal components of the signal.

[0038] Step 21: Use complex 1×1 convolution to interact and integrate the frequency features of phase and amplitude, assign different weights to different frequency components of amplitude and phase, and through learnable weight parameters, the model can emphasize the information that needs to be focused on. The process of complex 1×1 convolution is shown in formula (5).

[0039] Y(f)=X(f)·K=(X r (f)+iX i (f))·(K r +iK i )

[0040] =X r (f)K r -X i (f)K i +i(X r (f)W i +X i (f)W r ) (5)

[0041] where K is represented by a complex 1×1 convolution kernel with Kr+iKi, where Kr is the real weight and Ki is the imaginary weight. r (f) represents the real part of the phase or amplitude, X i (f) represents the imaginary part of the phase or amplitude.

[0042] Step 3, in order to better capture the key information of the input data, the present invention promotes the learning of the self-attention mechanism by enhancing the correlation between the contextual information (i.e., the features of adjacent time series) between the key matrices. Specifically, the present invention convolves the key matrix in each stage with convolution kernels of different sizes by adopting a custom convolution kernel, introduces local features, and changes the receptive field of the key matrix in each stage to capture the contextual information between the key matrices. In this way, the key matrix expresses the relationship between adjacent features. When calculating attention, the query matrix Q can more accurately match the relevant context, thereby improving the accuracy of the attention mechanism in capturing key information. A custom convolution kernel is introduced to the key matrix K, and the result obtained by this part is recorded as k.

[0043] k=Conv k×k (k) (6)

[0044] Step 4: Calculate the results of step 2 and step 3 using the attention score formula, and the obtained attention score is recorded as ATTN.

[0045] Step 41: Aggregate the frequency features of the phase and amplitude obtained in step 2 to obtain a new query matrix NewQuery, as shown in formula (7).

[0046] NewQuery=F(x pha )+F(x mag ) (7)

[0047] The F(·) function is the process of step 2.

[0048] Step 42: Calculate the results of step 2 and step 3 using the attention score formula to generate the attention result ATTN. The attention score calculation formula is shown in formula (8).

[0049]

[0050] where F sf (·) represents the softmax activation function, represents the scaling factor, T represents the transpose operation of the matrix, and F R (·) represents the real part operation, F I (·) represents the imaginary part operation.

[0051] Step 5. Finally, ATTN is input to the classification layer. In the classification stage, the input size is flattened into one dimension and transposed with the channel dimension. The purpose of this is to make it easier to operate on different dimensions. Then, a global average pooling operation is performed on the flattened spatial dimension to integrate different spatial positions into the global information of each channel. Finally, a fully connected layer is used to map the pooled feature vector to the final classification output.

[0052] Therefore, this embodiment proposes a method for drone signal classification based on pyramid frequency domain Transformer.

[0053] The present invention firstly uses a learnable weight matrix W Q , W K , W V For the input I / Q signal x a Perform linear calculations to obtain the complex matrix inputs Q, K, and V of the attention mechanism, and introduce Fourier transforms to Q, K, and V respectively.

[0054] Secondly, the amplitude and phase corresponding to the query vector Q matrix are calculated respectively. The frequency domain characteristics of the signal are mainly studied from the two angles of signal amplitude and phase. The complex 1×1 convolution is used to assign different weights to different frequency components of the amplitude and phase. Through the learnable weight parameters, the model can emphasize the information that needs to be focused on, and can efficiently extract the frequency domain characteristics of the signal, thereby improving the model's ability to identify key signal information.

[0055] Then, a custom convolution kernel is introduced to the key matrix K to enhance the correlation of contextual information between key matrices, and the result obtained from this part is used as part of the attention score. Then, the attention score obtained by using the attention score formula of the results of the first two parts is recorded as ATTN, and finally ATTN is input into the classification layer for mapping to obtain the final classification result.

[0056] The above is only a preferred embodiment of the present invention and does not limit the present invention in any way. Any technician in the relevant technical field, without departing from the scope of the technical solution of the present invention, makes any form of equivalent replacement or modification to the technical solution and technical content disclosed in the present invention, which does not depart from the content of the technical solution of the present invention and still falls within the protection scope of the present invention.

Claims

1. A UAV I / Q signal classification method based on pyramid frequency domain Transformer and complex convolution, characterized in that: The following steps are involved: Step 1: Use a learnable weight matrix W Q , W K , W V For the input complex signal eigenvector x a Perform linear calculations to obtain the complex matrix inputs Q, K, and V of the attention mechanism, and introduce Fourier transforms to Q, K, and V respectively; Step 2: Calculate the phase and amplitude X corresponding to the query vector Q matrix respectively pha With x mag , use complex 1×1 convolution to interact and integrate the frequency characteristics of phase and amplitude for Q in Transformer, and the result after interaction and integration is x pha ' and x mag '; Step 3: Introduce a custom convolution kernel into K in Transformer to extract signal features at different stages; Step 4: Use the attention score formula to input the final result ATTN into the classification layer for mapping to obtain the final classification result.

2. The method for classifying unmanned aerial vehicle I / Q signals based on pyramid frequency domain Transformer and complex convolution according to claim 1, characterized in that: Step 1 comprises the following steps: Step 11: Convert the complex drone signal x a With the learnable weight matrix W Q , W K , W V Perform linear operations to obtain complex matrices Q, K, and V; Q=x a ×W Q K=x a ×W K V=x a ×W V Step 12, introduce Fourier transform to Q, K, and V respectively; Q′=FFT(Q) K′=FFT(K) V′=FFT(V) Where FFT is Fourier transform.

3. The method for classifying unmanned aerial vehicle I / Q signals based on pyramid frequency domain Transformer and complex convolution according to claim 1, characterized in that: Step 2 comprises the following steps: Step 21: Calculate the phase and amplitude X corresponding to the query vector Q matrix pha With x mag ; x pha =arctan(x Q / x I ) where x I is the in-phase component of the signal, x Q are the orthogonal components of the signal. Step 22: Use complex 1×1 convolution to interact and integrate the frequency characteristics of phase and amplitude; x pha′ =Cpx 1×1 (x pha ) x mag′ =Cpx 1×1 (x mag ) Where Cpx 1×1 is a complex 1×1 convolution.

4. The method for classifying unmanned aerial vehicle I / Q signals based on pyramid frequency domain Transformer and complex convolution according to claim 1, characterized in that: Step 3 comprises the following steps: Step 31: Introduce a custom convolution kernel into K in Transformer to extract signal features at different stages; k=Conv k×k (k) Among them, Conv k×k It is a K×K convolution kernel, and the convolution kernel size is customized.

5. The method for classifying UAV I / Q signals based on pyramid frequency domain Transformer and complex convolution according to claim 1, characterized in that: Step 4 comprises the following steps: Step 41, use the attention score formula to obtain the result ATTN by applying the results obtained in step 2 and step 3; NewQuery=F(x pha )+F(x mag ) The F(·) function is the process of step 2. Step 42: The final result ATTN is input into the classification layer for mapping to obtain the final classification result.

Images