CN116738225A - Signal detection method based on improved YOLOv5 - Google Patents

Abstract

The invention discloses a signal detection method based on improved YOLOv5, which comprises the following steps: 1. performing short-time Fourier transform on the received signals to obtain a time-frequency diagram thereof, graying the time-frequency diagram, and constructing a signal detection time-frequency diagram data set; 2. in order to improve the feature extraction capability of the deep learning network, a CBAM module is introduced into classical YOLOv5; 3. to improve the accuracy of the final prediction frame of the algorithm, replacing the NMS algorithm in YOLOv5 with the WBF algorithm; 4. enhancing the impact of high quality predictions on training using improved Focal-EIoU loss functions; 5. training an improved YOLOv5 network model using an Adam optimizer and a signal time-frequency graph dataset; 6. and inputting the time-frequency diagram of the signal to be detected into a trained network model to obtain a signal detection result. The embodiment of the invention provides a method for detecting the target signal existing in the received broadband data by using the improved YOLOv5 network model for the first time, which is simple and practical, realizes higher signal detection performance with lower complexity, and has development significance for the application of the deep learning network in signal target detection.

Description

A signal detection method based on improved YOLOv5

Technical field

The invention belongs to the field of communication technology, and specifically relates to a signal detection method based on improved YOLOv5.

Background technique

With the continuous improvement of communication requirements due to the development of modern society, the communication environment has become extremely crowded, the electromagnetic environment is complex and changeable, and signals are in various forms and can arrive at the same time. Therefore, the primary task in non-cooperative communication is to detect narrowband signals in broadband channels, measure their related parameters and perform signal extraction.

Signal detection is to use certain technical means to judge the existence of the target signal of interest based on the received broadband data. The earliest signal detection was mainly completed by rough positioning of the signal on the video image by the human eye. The neural network algorithm in deep learning has powerful feature extraction capabilities and has achieved great success in image processing and other fields. Among them, the YOLO series is currently a popular target detection algorithm, and its detection speed and accuracy have certain advantages among algorithms of the same category. Based on this, consider using the improved YOLOv5 algorithm to detect the time-frequency diagram of the signal.

Contents of the invention

Based on the success of the YOLOv5 neural network algorithm in the field of target detection, the present invention studies the reasons why the YOLOv5 algorithm can be used for signal detection. By establishing a signal detection time-frequency diagram data set, the classic YOLOv5 algorithm is improved based on the data set, and the improved algorithm is trained using the data set to locate the signal in the time-frequency diagram, and ultimately achieve the purpose of signal detection.

The first aspect is a signal detection method based on improved YOLOv5, including:

S1: Perform short-time Fourier transform on the received signal to obtain its time-frequency diagram, and use a variety of preprocessing methods to construct a signal detection time-frequency diagram data set;

S2: The classic YOLOv5 algorithm includes Input, Backbone, neck, head and loss function. Based on the characteristics of the signal detection time-frequency graph data set, this method improves the Input, Backbone and loss function of the classic YOLOv5;

S3: The time-frequency graph in the training set is used as input, and the training set label vector is used as the output label. Use the signal detection time-frequency graph data set to train the improved YOLOv5 algorithm until convergence, and obtain the final signal detection model;

S4: Input the time-frequency diagram of the signal to be detected into the final YOLOv5 model to obtain the detection result;

Preferably, the step S1 specifically includes:

The signal to be detected is intercepted. The minimum duration of each signal is 50ms, the maximum duration is 4s, the noise base is Gaussian noise, and the SNR range is -5~10dB. The short-time Fourier transform is performed on the preprocessed signal to obtain The time-frequency diagram is converted into grayscale. Each time-frequency diagram sample contains 5 to 10 target signals, and the center frequency and duration of each signal are randomly determined.

Preferably, the step S2 specifically includes:

Step 2.1: Use the improved Mosaic data enhancement method. During the training process, randomly cut off a part of each of the four images, reduce them to a certain proportion, and then randomly arrange and splice them to obtain new images, thereby increasing the number of data samples;

Step 2.2: Combine the average pooling on the time-frequency graph space and the average pooling on the channel, introduce the CBAM (Convolutional Block Attention Module) module, and use the attention mechanism to improve the feature extraction capability of the algorithm;

Step 2.3: Replace the NMS algorithm in the classic YOLOv5 with the WBF algorithm, use the information of all prediction boxes to perform a weighted sum, and improve the accuracy of the final prediction box of the algorithm;

Step 2.4: Use the improved Focal-EIoU loss function to enhance the impact of high-quality prediction results in the training process;

Step 2.5: The Head part uses three convolutional layers as the final detector. Different convolutional layers are responsible for detecting signals with different scales;

Preferably, the step 2.1 specifically includes:

The content in the time-frequency diagram represents the information about the signal frequency domain energy changing with time. This method of random cropping and random splicing is not suitable for processing time-frequency diagrams. This method converts a time-frequency diagram from top to bottom. Divide the image into four areas evenly. During the training process, four images are randomly selected. Each of the four images corresponds to a region and a part is cropped. The four parts are then spliced into a new image according to the corresponding area. That is, the improved Mosaic data enhancement method clips and splices the time-frequency diagram in the frequency domain direction, thereby completely retaining the time-frequency information of the signal.

Preferably, the step 2.2 specifically includes:

CBAM is a deep convolutional neural network attention module that combines spatial average pooling and channel average pooling, and uses the attention mechanism to improve the performance of the model. This method introduces the CBAM module into the CSP structure of YOLOv5. Increase the model's ability to extract features from time-frequency diagrams and improve the algorithm's ability to detect signals; CBAM consists of two sub-modules: spatial attention sub-module and channel attention sub-module; the calculation process of the CBAM module is: first, for the input features In the channel attention sub-module, the feature map is passed through two parallel maximum pooling and average pooling to calculate a set of feature maps with unchanged channels and a scale of 1×1; secondly, a set of fully connected layers are used to continue Compress the number of channels of the feature map; finally, use parallel maximum pooling and average pooling to expand to the original number of channels. The two parallel results are added and activated by Sigmoid to obtain a set of channel attention/> Multiply the channel attention by the original input to generate a new feature map/> In the spatial attention sub-module, the feature map F' processed by the channel attention module is sequentially processed through maximum pooling and average pooling to obtain a feature map with a constant scale and a channel number of 1, which is used to characterize the characteristics of a specific area; Then, a convolutional layer and activation layer are used to generate a set of spatial attention feature weights on these feature maps. Finally, these feature weights are multiplied by the feature map F' to generate the final output feature map/>

Preferably, the step 2.3 specifically includes:

The WBF algorithm sets different weights for each predicted bounding box, and calculates a result through weighted fusion as the final fusion result. The result is more accurate. This method replaces the NMS algorithm in YOLOv5 as the prediction box fusion algorithm. , the method is as follows: first, create two lists B and C to store all current prediction boxes and corresponding confidences respectively; secondly, create two empty lists L and F. L is used to store suitable bounding boxes, and F only contains A fused bounding box; finally, set a threshold T, traverse list B, and find all prediction boxes whose IoU with F is greater than T. If no bounding box is found, add the bounding box to the tail of B and F. If found, If the bounding box is missing, add the bounding box to the tail of B, and update the bounding box in F according to the following formula

Among them, C is the confidence of the prediction box, and X and Y are the X and Y coordinates of the center point of the prediction box respectively.

Preferably, the step 2.4 specifically includes:

Considering that there is an imbalance of training samples in the regression of prediction frames, that is, the number of high-quality prediction frames with small regression errors in an image is much less than the low-quality samples with large errors. Poor-quality samples will generate excessive The gradient of affects the training process; therefore, refer to the practice of FocalLoss to give greater weight to high-quality prediction boxes so that they have a greater impact on the training process. The improved Focal-EIoU expression is

Among them, γ is a parameter that controls the degree of outlier suppression, IoU is the intersection ratio of the predicted box and the real box, Represents the Euclidean distance between the center of the predicted box and the center of the true box,/> and/> Represents the width and height of the minimum circumscribed rectangle of the prediction box and the true box respectively. α is a weight factor with a value less than 1. Its value is obtained based on the height-to-width ratio of all targets in the Kmeans clustering statistical data set.

Preferably, the step S3 specifically includes:

Set the maximum number of iterations to 100, and update the algorithm parameters according to the neural network backpropagation algorithm by minimizing the loss function until the loss function converges or the number of iterations is reached.

Preferably, the step S4 specifically includes:

The time-frequency diagram with multiple signals is fed into the improved YOLOv5 algorithm as input. The algorithm marks the position of the signal in the diagram in the form of a generated prediction box, and indicates parameters such as signal category, bandwidth, center frequency, and start and end time. , to achieve the purpose of signal detection.

Description of drawings

In order to explain the embodiments of the present invention or the technical solutions in the prior art more clearly, the drawings needed to be used in the description of the embodiments or the prior art are briefly introduced below. Obviously, the drawings in the following description are only These are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without exerting creative efforts.

Figure 1 is a schematic flow chart of a signal detection method based on improved YOLOv5 provided by an embodiment of the present invention;

Figure 2 is an architecture diagram of the improved YOLOv5 algorithm of the present invention;

Figure 3 is a structural diagram of the CBAM module in the present invention;

Figure 4 is a structural diagram of the channel attention module of the present invention;

Figure 5 is a structural diagram of the spatial attention module in the present invention;

Figure 6 is a diagram of the detection effect of the signal time-frequency diagram according to the present invention;

Figure 7 is a graph comparing the performance of the present invention and three other neural network algorithms;

Figure 8 is a comparison graph of the effectiveness of four improvement methods in the present invention;

Detailed ways

In order to make the purpose, 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 in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments These are some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention.

S1: Establish signal detection time-frequency graph data set.

S2: Use the signal detection time-frequency diagram data set to train the improved YOLOv5 algorithm until convergence.

S3: Use the improved YOLOv5 algorithm that has been trained to identify the signals in the test set.

Specifically, step S1 includes:

A. Create a training set

The signal to be detected is intercepted. The minimum duration of each signal is 50ms, the maximum duration is 4s, the noise base is Gaussian noise, and the SNR range is -5~10dB. The short-time Fourier transform is performed on the preprocessed signal to obtain The time-frequency diagram is converted into grayscale. Each time-frequency diagram sample contains 5 to 10 target signals, and the center frequency and duration of each signal are randomly determined.

The short-time Fourier transform is defined as:

where w(t) is the sliding window function. It can be seen that STFT is a local spectrum. When w(t)=1, the definition formula is:

The STFT at this time is the definition of Fourier transform. When the STFT of s(t) is equal to the Fourier transform, only frequency information exists.

The inverse transform of STFT is:

When the window functions h(t) and w(t) satisfy the following relationship, the signal s(t) can be reconstructed from STFT(τ,ω) through the above formula:

The discrete STFT of a sequence is defined as:

Where N is the length of the sequence s[n], and the width of the window function w[n] is controllable. For time-frequency analysis of signals, what is usually analyzed is its amplitude, and its spectrum can be defined as:

The corresponding amplitude diagram is the time-frequency diagram.

B. Create a test set

The test set is divided into a test set under Gaussian channel and a test set under Rayli fading channel, which are simulated in Gaussian white noise environment and Rayleigh channel respectively. The signal-to-noise ratio SNR values are -5dB, -4dB, ..., 5dB, Each signal uses short-time Fourier transform to generate 300 time-frequency diagram samples at each signal-to-noise ratio, and a total of 3300 samples form two test sets.

Further, step S2 includes:

The improved YOLOv5 in the present invention mainly consists of four parts: Input, Backbone, Neck and Head. Among them, the Input stage mainly performs data enhancement on the input image. Backbone is the backbone of YOLOv5, which is mainly stacked by the CBS module and the CSP module. It is responsible for feature learning of the input. In the Neck stage, the SPPF structure is first used to pass the input feature map through the maximum Each feature map after the pooling operation is feature fused, and the Head part uses three convolutional layers as the final detector.

The learning rate of the training process is set to 0.001, the batch size is set to 32, the maximum epoch is set to 100, and the Adam optimizer is used. The hardware and software configuration information of the experimental environment is as shown in the following table:

Table 1 Experimental environment configuration information

Further, the step S3 specifically includes:

Set the maximum number of iterations to 100, and update the algorithm parameters according to the neural network backpropagation algorithm by minimizing the loss function until the loss function converges or the number of iterations is reached.

Further, the step S4 specifically includes:

The time-frequency diagram with multiple signals is fed into the improved YOLOv5 algorithm as input. The algorithm marks the position of the signal in the diagram in the form of a generated prediction box, and indicates parameters such as signal category, bandwidth, center frequency, and start and end time. , to achieve the purpose of signal detection.

Claims (9)

1. A method for improved YOLOv 5-based signal detection, the method comprising:

s1: performing short-time Fourier transform on the received signals to obtain time-frequency diagrams of the received signals, and constructing a signal detection time-frequency diagram data set by using a plurality of preprocessing methods;

s2: the classical YOLOv5 algorithm comprises Input, backbone, neck, head and loss functions, and the method improves Input, backbone and loss functions of the classical YOLOv5 according to the characteristics of a signal detection time-frequency chart data set;

s3: training the improved YOLOv5 algorithm to be converged by using a signal detection time-frequency diagram data set to obtain a final signal detection model;

s4: and inputting the time-frequency diagram of the signal to be detected into a final YOLOv5 model to obtain a detection result.

2. The improved YOLOv 5-based signal detection method of claim 1, wherein step S1 specifically comprises:

intercepting signals to be detected, wherein the minimum duration of each signal is 50ms, the maximum duration is 4s, the noise substrate is Gaussian noise, the SNR range is-5-10 dB, performing short-time Fourier transform on the preprocessed signals to obtain a time-frequency diagram and gray scale, each time-frequency diagram sample contains 5-10 target signals, and the center frequency and the duration of each signal are randomly determined.

3. The improved YOLOv 5-based signal detection method of claim 1, wherein step S2 specifically comprises:

step 2.1: the improved Mosaic data enhancement method is used, four pictures are randomly cut off in a part respectively in the training process, and new images are obtained by shrinking according to a certain proportion and then randomly arranging and splicing, so that the number of samples of the data is increased;

step 2.2: combining the average pooling in the time-frequency diagram space and the average pooling in the channel, introducing a CBAM (Convolutional Block Attention Module) module, and improving the feature extraction capability of the algorithm by using an attention mechanism;

step 2.3: the NMS algorithm in the classical YOLOv5 is replaced by the WBF algorithm, the information of all the prediction frames is used for weighted summation, and the accuracy of the final prediction frame of the algorithm is improved;

step 2.4: enhancing the impact of high quality predictions on training using improved Focal-EIoU loss functions;

step 2.5: the Head section uses three convolution layers as the final detector, with different convolution layers being responsible for detecting signals of different scale sizes.

4. The method for improved YOLOv 5-based signal detection of claim 3, wherein said step 2.1 comprises:

the method comprises the steps that a time-frequency diagram is divided into four areas from top to bottom, four images are randomly selected in the training process, a part of each of the four images is cut out corresponding to one area, and then the four parts are spliced into a new image according to the corresponding area; the improved Mosaic data enhancement method cuts and splices the time-frequency diagram in the frequency domain direction, so that the time-frequency information of the signals is completely reserved.

5. The method for detecting a signal based on improved YOLOv5 of claim 3, wherein said step 2.2 comprises:

the method introduces the CBAM module into a CSP structure of YOLOv5, increases the feature extraction capacity of the model on a time-frequency chart, and improves the detection capacity of an algorithm on signals; CBAM consists of two sub-modules: a spatial attention sub-module and a channel attention sub-module; the calculating process of the CBAM module is as follows: first, for input featuresIn the channel attention submodule, the feature map is subjected to two parallel maximum pooling and average pooling, and a group of feature maps with unchanged channels and the scale of 1 multiplied by 1 are calculated; secondly, continuously compressing the channel number of the feature map by using a group of full connection layers; finally, using parallel maximum pooling and average pooling to expand to the original channel number, adding two parallel results and activating by Sigmoid to obtain a group of channel attention +.>Multiplying the channel attention by the original input to generate a new profile +.>In the space attention sub-module, the feature map F 'processed by the channel attention module sequentially passes through maximum pooling and average pooling to obtain a feature map with the unchanged scale and the number of channels of 1, and the feature map F' is used for representing the features of a specific area; then, a set of spatial attention feature weights are generated for the feature maps using a convolution layer and an activation layerFinally, these feature weights are multiplied by the feature map F' to generate the final output feature map +.>

6. The method for improved YOLOv 5-based signal detection of claim 3, wherein said step 2.3 comprises:

the WBF algorithm sets different weights for each predicted boundary box, calculates a result through weighted fusion as a final fusion result, and the obtained result has higher accuracy. Firstly, establishing two lists B and C, and respectively storing all current prediction frames and corresponding confidence coefficients; secondly, two empty lists L and F are established, wherein L is used for storing proper bounding boxes, and F only contains one fused bounding box; finally, a threshold T is set, a list B is traversed, all predicted frames which are larger than T with IoU of F are found, if no boundary frame is found, the boundary frames are added to the tail parts of B and F, if the boundary frame is found, the boundary frame is added to the tail part of B, and the boundary frame in F is updated according to the following formula

Where C is the confidence level of the predicted frame and X, Y is the X, Y coordinates of the center point of the predicted frame.

7. The method for improved YOLOv 5-based signal detection of claim 3, wherein said step 2.4 comprises:

considering the problem that training samples are unbalanced in regression of prediction frames, namely the number of high-quality prediction frames with small regression errors in one image is far less than that of low-quality samples with large errors, and the samples with poor quality can generate overlarge gradients to influence the training process; therefore, referring to Focal Loss, the high-quality prediction frame is given a larger weight, so that the influence on the training process is larger, and the improved Focal-EIoU expression is as follows

Wherein, gamma is a parameter for controlling the suppression degree of abnormal values, ioU is the intersection ratio of a prediction frame and a real frame,euclidean distance representing the prediction frame center and the true frame center,>and->The values of the alpha are weight factors with the values smaller than 1, and the values are obtained according to the aspect ratio of all targets in the Kmeans cluster statistical data set.

8. The improved YOLOv 5-based signal detection method of claim 1, wherein step S3 specifically comprises:

setting the maximum iteration number as 100, and updating algorithm parameters according to a neural network back propagation algorithm by minimizing the loss function until the loss function converges or the iteration number is reached.

9. The improved YOLOv 5-based signal detection method of claim 1, wherein step S4 specifically comprises:

the time-frequency diagram with various signals is used as input to be sent into an improved YOLOv5 algorithm, the algorithm marks the position of the signals in the diagram in the form of generating a prediction frame, and the parameters such as signal category, bandwidth, center frequency, start-stop time and the like are marked, so that the purpose of signal detection is achieved.