Intelligent signal detection and identification method based on time-frequency feature fusion and anchor-free detection head

Abstract

The invention discloses an intelligent signal detection and identification method based on time-frequency feature fusion and an anchor-free detection head. The method comprises the following steps: capturing a wireless signal broadband; carrying out time-frequency transformation preprocessing; signal detection and identification based on a neural network; and converting the time-frequency information. The invention provides a network suitable for signal detection and identification. The network is divided into a backbone network, a neck network and a detection head network. A backbone network composed of a component-convolution block architecture is provided, and the network complexity can be flexibly adjusted to effectively extract features; a multi-scale time-frequency feature aggregation module is inserted into the neck network to efficiently perform feature fusion; the detection head network is changed into an anchor-free mechanism to improve the output efficiency; and the final output of the network is subjected to time-frequency information conversion to obtain the bandwidth, duration and other parameters of the signal. According to the method, support can be provided for rapid and automatic signal detection and identification, and the problems of high human participation degree, high signal detection and identification rate and the like in military communication intelligence reconnaissance are effectively solved.

Description

Technical Field

[0001] The present invention relates to signal detection and recognition based on artificial intelligence technology, belongs to the technical field of communication signal processing, and specifically relates to an intelligent signal detection and recognition method based on time-frequency feature fusion and anchor-free detection head. Background Art

[0002] Communication signal detection and recognition is a key technology for applications such as cognitive radio and interference detection. Its task is to detect signals from received broadband signals with or without prior knowledge, and to estimate signal type, frequency, bandwidth, duration and other parameters. With the massive growth of radio frequency data in wireless environments, communication signal detection and recognition technology is becoming increasingly important and has attracted much attention in both civil and military fields.

[0003] Traditional communication signal detection and recognition technologies mainly include energy detection method, matched filtering method and feature-based method. However, the energy detection method has poor robustness, the matched filtering method requires prior knowledge of the signal, and the design of effective signal features relies on expert knowledge and is difficult to effectively generalize to other RF data.

[0004] In recent years, deep learning-based methods have been widely used in the field of RF signal processing, and have shown excellent performance in key technologies such as modulation recognition, signal detection, and channel estimation. Most of the signal detection and recognition algorithms based on deep learning transform the signal into an image representation such as a time-frequency graph through time-frequency transformation, and then the signal detection in the time-frequency graph becomes similar to the target detection in the field of computer vision. Therefore, previous methods use some classic target detection algorithms, including Faster-RCNN, SSD, YOLO series, etc. However, these methods still maintain the network architecture of the original target detection algorithm as a whole, and have not made targeted improvements to the signal time-frequency graph dataset, making it difficult to achieve optimal detection performance in specific tasks. Specifically, these methods are anchor frame-based algorithms, which will encounter anchor frame mismatch problems when facing signals with uncertain bandwidth and duration, resulting in performance degradation. In addition, these methods do not focus on the neural network design of the feature fusion part, which is not conducive to achieving optimal performance. Summary of the invention

[0005] In view of the shortcomings of existing intelligent detection and recognition algorithms in the field of wireless signal detection and recognition, the present invention proposes an intelligent signal detection and recognition method based on time-frequency feature fusion and anchor-free detection head, which includes the following steps:

[0006] Step 1: Wireless signal broadband capture;

[0007] The center frequency and receiving bandwidth are set according to the task to capture the complex baseband IQ signal stream containing the target signal; the received signal also contains interference from other signals and in-band noise;

[0008] Step 2: Time-frequency transform preprocessing;

[0009] Time-frequency transform preprocessing uses short-time Fourier transform STFT, wavelet analysis and other time-frequency transforms to convert the time domain signal to the time-frequency domain and generate a signal time-frequency graph, which is then used as a data set for deep learning; if x(t) is a sampled complex time series, the STFT result of x(t) is expressed as:

[0010]

[0011] Among them, S(n,m) is the time-frequency matrix obtained by short-time Fourier transform, n and m are frequency and time indexes respectively; w(k) is the analysis window of length N, O is the number of overlapping points between adjacent windows; the modulus value of S(n,m) or the logarithmic modulus value is saved as a time-frequency graph; after preprocessing by time-frequency transform, the representation of the signal changes from a one-dimensional time series to a foreground signal in a two-dimensional time-frequency image;

[0012] Step 3: Signal detection and recognition based on neural network;

[0013] The transformed time-frequency graph in step 2 is sent to the trained neural network, and the neural network automatically outputs parameters such as signal category and regression coordinates;

[0014] A neural network suitable for signal detection and recognition, which is refined into three parts: backbone network, neck network and detection head network. The backbone network uses multi-scale time-frequency feature context aggregation and gated attention components to achieve time-frequency feature fusion; the anchor detection head network is replaced by an anchor-free mechanism;

[0015] In the training phase, the RGB three-channel time-frequency image with a batch size of B is input. After the backbone network extracts the signal time-frequency features and the neck network features are fused, the detection head network obtains the three-part output of category, positioning regression, and target. The neural network output continues to undergo post-processing such as non-maximum suppression and time-frequency information conversion to obtain the signal category, start and end time, center frequency, and bandwidth.

[0016] (1) Backbone network

[0017] The overall architecture of the backbone network is divided into multiple convolutional blocks, each of which is composed of multiple basic CBR components connected in series; the CBR component includes a convolutional layer, a batch normalization layer, and a Relu activation function.

[0018] The convolution operation in the backbone network CRB component continuously improves the network's receptive field. As the network deepens, the number of output channels of the convolution layer in the CBR component is also increased to obtain signal-rich time-frequency features. At the same time, the convolution step size is set to 2 in the first convolution operation of each convolution block, so that the feature map after convolution becomes half of the original feature map in length and width, thereby obtaining feature maps of different scales output by different convolution blocks.

[0019] (2) Neck network

[0020] The neck network is a multi-scale time-frequency feature aggregation module, which consists of two parts: multi-scale time-frequency feature context aggregation and gated attention component;

[0021] The feature maps of different scales output by different convolution blocks in the backbone network are taken out and subjected to corresponding spatial pyramid pooling to obtain rich context feature information at different scales; the feature maps of different scales are unified to the same resolution through convolution operations and nearest neighbor interpolation;

[0022] Choose to directly add or connect feature maps of different scales along the channel dimension to obtain the fused feature map; add a gated attention component after multi-scale time-frequency feature context aggregation;

[0023] Each feature map X i Input different gated attention component branches; in the corresponding gated attention component branch, the input is divided into two paths, one of which keeps the original feature map X i The other path is converted into the gated weight W after passing through the CBR component, convolution layer, and Sigmoid activation function. i , and then the two paths are element-wise multiplied to obtain the weighted output of the branch; the outputs of different gated attention component branches are weighted and calculated to obtain the final fused feature map Y, as shown in formula (2):

[0024]

[0025] Among them, Y, CBR(.), Conv(.) represent the final fused feature map output by the gated attention component, the CBR component, and the convolution layer respectively; S is the number of feature maps of different scales taken from the backbone network; σ is the Sigmoid activation function, ⊙ is the element-wise multiplication, and the feature map is fused at the pixel level through attention gating; the weight W i As the network converges, the gating weights gradually suppress useless information and fuse information in multi-scale feature maps that is beneficial to back-end detection;

[0026] (3) Detection head network

[0027] The detection head network adopts the structure of anchor-free decoding plus decoupling head. The decoupling head means that the classification and regression branches are not shared.

[0028] The feature Y after the neck network fusion passes through the CBR component to decouple the features, and then obtains three output branches through the convolution layer and the Sigmoid activation function. According to the role played by different output branches, they are named category branch, positioning branch and target branch respectively; the category branch is used to output the confidence of the signal category; the positioning branch is used to output the coordinates of the signal prediction box; the target branch is to make the network output an additional prediction box quality score to judge the quality of the prediction box, and then suppress the low-quality prediction box far away from the central target when assigning labels;

[0029] The loss function during network training consists of three parts, as shown in formula (3):

[0030] L det =L cls +λL reg +μL obj (3)

[0031] Among them, L det is the total cost function, L cls and L obj are the category loss and target loss, respectively. The focus loss function and the binary cross entropy function are used in training; L reg is a positioning regression function, using L1 loss function in training; λ and μ are hyperparameters used to balance weights;

[0032] Step 4: Time-frequency information conversion;

[0033] The parameters output by the neural network include the signal category, the starting and ending x-axis coordinates (x1, x2) and the starting and ending y-axis coordinates (y1, y2) of the signal frame; assuming that the sampling rate of the signal is fs and the observation time of the time-frequency diagram is Td, the bandwidth of the signal is converted to: fs×(x2-x1), and the duration of the signal is converted to Td×(y2-y1).

[0034] The neural network architecture designed by the present invention uses a backbone network with flexible complexity adjustment to extract signal features; inserts a neck network, uses multi-scale time-frequency feature context aggregation and gated attention components to effectively fuse features; and replaces the anchor detection head in the previous method with an anchor-free structure, avoiding the calculation of the prior anchor frame while facilitating model parameter reduction and network reasoning. The method of the present invention has significantly better detection performance on the constructed voice signal dataset, and also performs better than other signal detection and recognition methods based on deep learning in terms of model parameters and reasoning time.

[0035] The method of the present invention can provide support for rapid and automatic signal detection and recognition, and effectively solve the problems of high human participation and low signal detection and recognition rate in military communication intelligence reconnaissance. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 A flow chart showing an intelligent signal detection and recognition method based on time-frequency feature fusion and anchor-free detection head of the present invention;

[0037] Figure 2 Shows a sample time-frequency diagram of a speech signal;

[0038] Figure 3 The overall structure of the neural network designed by the present invention is shown;

[0039] Figure 4 The backbone network composition of the neural network designed by the present invention is shown;

[0040] Figure 5 The composition of the multi-scale time-frequency feature aggregation module designed by the present invention is shown;

[0041] Figure 6 The visualization result of the detection and recognition of the actual collected voice signal according to the present invention is shown. DETAILED DESCRIPTION

[0042] The flowchart of the intelligent signal detection and recognition method based on time-frequency feature fusion and anchor-free detection head of the present invention is as follows: Figure 1 As shown, the method includes four steps: broadband capture of wireless signals, time-frequency transformation preprocessing, signal detection and recognition based on neural network, and time-frequency information conversion. The details are as follows.

[0043] Step 1: Wireless signal broadband capture;

[0044] The center frequency and receiving bandwidth are set according to the task to capture the complex baseband IQ signal stream containing the target signal. Since the present invention adopts a broadband receiver, there are interferences from other signals and in-band noise in the received signal.

[0045] Step 2: Time-frequency transform preprocessing;

[0046] Time-frequency transform preprocessing refers to using short-time Fourier transform (STFT), wavelet analysis and other time-frequency transforms to convert time domain signals to the time-frequency domain and generate signal time-frequency graphs, which are then used as data sets for deep learning. If x(t) is a sampled complex time series, the STFT result of x(t) can be expressed as:

[0047]

[0048] Among them, S(n,m) is the time-frequency matrix obtained by short-time Fourier transform, n and m are frequency and time indexes respectively. w(k) is the analysis window of length N, and O is the number of overlapping points between adjacent windows. Then the modulus value of S(n,m) or the logarithmic modulus value is saved as a time-frequency graph. After the time-frequency transform preprocessing, the representation of the signal changes from a one-dimensional time series to a foreground signal in a two-dimensional time-frequency image.

[0049] Figure 2 This is an example of converting a speech signal from a time series to a time-frequency diagram. The time-frequency diagram is generated using the short-time Fourier transform mentioned above. The light background represents background noise interference. The target in the dark-colored box is the representation of the signal in the time-frequency diagram. The horizontal and vertical coordinates of the box represent the start and end of the frequency and the start and end of the time, respectively.

[0050] Step 3: Signal detection and recognition based on neural network;

[0051] Signal detection and recognition based on neural network refers to sending the transformed time-frequency diagram in step 2 into the trained neural network, and the neural network automatically outputs parameters such as signal category and regression coordinates.

[0052] The adaptability of the neural network architecture determines the detection and recognition performance. The present invention designs a network suitable for signal detection and recognition, and its overall architecture is as follows: Figure 3 shown.

[0053] The neural network is refined into three parts: a backbone network, a neck network, and a detection head network. These three parts of the network are improved according to the characteristics of the actual signal detection task: First, the signal detection in the time-frequency graph dataset is less complex than the general target detection task in natural images. Therefore, the present invention designs a backbone network with a component-convolutional block architecture, which can adjust the complexity of the network by flexibly configuring the number of components and convolutional blocks to better adapt to the signal detection task, as described below. In addition, the neck network used for feature fusion in the classic target detection algorithm is designed for the characteristics of a specific benchmark general target dataset and cannot adapt well to the distribution of the signal time-frequency graph dataset. Therefore, the present invention designs a new neck network, which uses multi-scale time-frequency feature context aggregation and gated attention components to achieve better time-frequency feature fusion. Multi-scale time-frequency feature context aggregation and gated attention components together constitute the new neck network. In addition, the diversity of signal bandwidth and duration leads to significant changes in the aspect ratio of the signal target box, which makes the anchor box-based detection neural network face the problem of anchor box mismatch, which in turn leads to performance degradation. Therefore, the anchor detection head network is replaced by an anchor-free mechanism, which avoids the calculation of prior anchor boxes while facilitating model parameter reduction and network reasoning, as described below.

[0054] In the training phase, we first input the RGB three-channel time-frequency image with a batch size of B. After the backbone network extracts the signal time-frequency features and the neck network features are fused, the detection head network obtains the three-part output of category, positioning regression, and target. The output of the neural network continues to undergo post-processing such as non-maximum suppression and time-frequency information conversion to obtain the signal category, start and end time, center frequency, bandwidth and other time-frequency information.

[0055] (1) Backbone network

[0056] The backbone network of the target detection algorithm is a key component for extracting image features. Compared with general target detection images, the time-frequency image background is simpler and the color changes within the signal are not large. At the same time, there is only one type of target in the task data set. Directly using the complex backbone network of the existing algorithm may not be able to efficiently extract signal features. Therefore, designing task-oriented network width and depth is not only beneficial to the back-end neck network feature fusion and detection head network classification regression, but also can reduce network inference time and meet the real-time requirements of signal detection.

[0057] like Figure 4 As shown in the figure, the overall architecture of the designed backbone network is divided into multiple convolution blocks, each of which is composed of multiple basic CBR components in series. The CBR component includes a convolution layer, a batch normalization layer, and a Relu activation function. When actually configuring the backbone network, the network depth and the richness of the acquired feature maps can be changed by setting the number of convolution blocks. The network width and depth of each convolution block can also be adjusted more finely by setting the number of CBR components in each convolution block, the convolution kernel size of the convolution layer in the component, and the number of convolution channels. This method can flexibly adjust the network complexity and facilitate code reuse of convolution blocks and components.

[0058] The backbone network proposed in the present invention can flexibly adjust the number of components, and when the task is not complex, it can reduce the components and convolution stages to achieve the purpose of streamlining the network.

[0059] The convolution operation in the backbone network CRB component continuously improves the network's receptive field (well known to those skilled in the art). As the network deepens, the number of output channels of the convolution layer in the CBR component also increases, thereby obtaining signal-rich time-frequency features. At the same time, in the first convolution operation of each convolution block, the convolution step size is set to 2, so that the feature map after convolution becomes half the length and width of the original feature map, so that feature maps of different scales can be obtained from different convolution blocks.

[0060] (2) Neck network

[0061] The neck network mainly realizes efficient feature fusion. To this end, the present invention designs the following Figure 5 The multi-scale time-frequency feature aggregation module shown in Figure 1 is composed of two parts: multi-scale time-frequency feature context aggregation and gated attention component.

[0062] First, the feature maps of different scales output by different convolution blocks in the backbone network are taken out, and respectively passed through the corresponding spatial pyramid pooling (SPP) module (He K, Zhang X, Ren S, et al. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition [J]. IEEE Transactions on Pattern Analysis & Machine Intelligence, 2014, 37 (9): 1904-16. DOI: 10.1007 / 978-3-319-10578-9_23.) to obtain rich context feature information at different scales. Considering the scale distribution of the target signal in the above data set, the feature maps of different scales are unified to the same resolution through convolution operations and nearest neighbor interpolation. The specific operation is well known to those skilled in the art and will not be repeated.

[0063] At this time, you can choose to directly add or connect feature maps of different scales along the channel dimension to obtain a fused feature map, as is well known to those skilled in the art. However, if there are semantic differences between feature maps from different convolutional blocks, simple fusion may bring poor results. Figure 5 The multi-scale time-frequency feature context aggregation shown is also followed by a gated attention component.

[0064] like Figure 5 As shown, each feature map X i Different gated attention component branches are input. In the corresponding gated attention component branch, the input is divided into two paths, one of which keeps the original feature map X i The other path is converted into the gated weight W after passing through the CBR component, convolution layer, and Sigmoid activation function. i (CBR components, convolutional layers, and Sigmoid activation functions are well known to those skilled in the art), and then these two paths are element-wise multiplied to obtain the weighted output of the branch. The outputs of different gated attention component branches are weighted and calculated to obtain the final fused feature map Y, as shown in formula (2):

[0065]

[0066] Among them, Y, CBR(.), Conv(.) represent the final fused feature map output by the gated attention component, the CBR component, and the convolutional layer, respectively. S is the number of feature maps of different scales taken from the backbone network. Figure 5 In the example, it is 3. σ is the Sigmoid activation function, ⊙ is the element-wise multiplication, and the feature map can be fused at the pixel level through attention gating. i It is automatically obtained through training. As the network converges, the gating weights gradually suppress useless information and fuse the information in the multi-scale feature map that is beneficial to back-end detection.

[0067] (3) Detection head network

[0068] Target detection algorithms can be divided into two types: anchored and anchorless. The anchor box is a set of rectangular boxes set a priori. When the anchor box is similar to the real target box in size and aspect ratio, it will be beneficial to the regression of the target box. Therefore, the setting of the anchor box affects the performance of the detection algorithm to a certain extent. The previous method obtained a more suitable anchor box by improving the Kmeans algorithm, but this will increase the algorithm latency. Therefore, the present invention replaces the anchor mechanism with an anchorless decoding method plus a decoupling head structure, where the decoupling head means that the classification and regression branches are not shared.

[0069] like Figure 3 As shown in the figure, the feature Y after the neck network fusion passes through the CBR component to decouple the feature, and then three output branches are obtained through the convolution layer and the Sigmoid activation function. According to the role played by different output branches, they are named category branch, positioning branch and target branch. Among them, the category branch is used to output the confidence of the signal category. The positioning branch is used to output the coordinates of the signal prediction box. The target branch is to make the network output an additional prediction box quality score to judge the quality of the prediction box, so that the low-quality prediction box far away from the central target can be suppressed during label assignment.

[0070] The loss function during network training also consists of three parts, as shown in formula (3):

[0071] L det =L cls +λL reg +μL obj (3)

[0072] Among them, L det is the total cost function, L cls and L obj are the category loss and target loss, respectively. The focus loss function and the binary cross entropy function are used in training; L reg is the positioning regression function, and the L1 loss function is used in training; λ and μ are hyperparameters used to balance the weights.

[0073] Step 4: Time-frequency information conversion;

[0074] The parameters output by the neural network include the signal category, the start and end x-axis coordinates (x1, x2) and the start and end y-axis coordinates (y1, y2) of the signal frame, which will be converted in this step. Assuming that the sampling rate of the signal is fs and the observation time of the time-frequency diagram is Td, the bandwidth of the signal is converted to: fs×(x2-x1), and the duration of the signal is converted to Td×(y2-y1).

[0075] Figure 6 The present invention provides the visualization results of the detection and recognition of the actual collected voice signal. The dark detection box in the figure is drawn according to the coordinates output by the neural network, and the value above the detection box is the signal category confidence output by the neural network. It can be seen that the present invention can not only accurately locate the signal in the time-frequency diagram, but also has excellent recognition confidence.

[0076] The present invention first converts the captured time series into a time-frequency diagram through short-time Fourier transform. Then a neural network architecture suitable for signal detection and recognition is designed to perform signal detection and recognition on the time-frequency diagram. Finally, the parameters such as signal category, bandwidth and duration are obtained through time-frequency information conversion.

Claims

1. An intelligent signal detection and recognition method based on time-frequency feature fusion and anchor-free detection head, characterized in that: The following steps are involved: Step 1: Wireless signal broadband capture; The center frequency and receiving bandwidth are set according to the task to capture the complex baseband IQ signal stream containing the target signal; the received signal also contains interference from other signals and in-band noise; Step 2: Time-frequency transform preprocessing; Time-frequency transform preprocessing uses short-time Fourier transform STFT, wavelet analysis and other time-frequency transforms to convert the time domain signal to the time-frequency domain and generate a signal time-frequency graph, which is then used as a data set for deep learning; if x(t) is a sampled complex time series, the STFT result of x(t) is expressed as: Among them, S(n,m) is the time-frequency matrix obtained by short-time Fourier transform, n and m are frequency and time indexes respectively; w(k) is the analysis window of length N, O is the number of overlapping points between adjacent windows; the modulus value of S(n,m) or the logarithmic modulus value is saved as a time-frequency graph; after preprocessing by time-frequency transform, the representation of the signal changes from a one-dimensional time series to a foreground signal in a two-dimensional time-frequency image; Step 3: Signal detection and recognition based on neural network; The transformed time-frequency graph in step 2 is sent to the trained neural network, and the neural network automatically outputs parameters such as signal category and regression coordinates; A neural network suitable for signal detection and recognition, which is refined into three parts: backbone network, neck network and detection head network. The backbone network uses multi-scale time-frequency feature context aggregation and gated attention components to achieve time-frequency feature fusion; the anchor detection head network is replaced by an anchor-free mechanism; In the training phase, the RGB three-channel time-frequency image with a batch size of B is input. After the backbone network extracts the signal time-frequency features and the neck network features are fused, the detection head network obtains the three-part output of category, positioning regression, and target. The neural network output continues to undergo post-processing such as non-maximum suppression and time-frequency information conversion to obtain the signal category, start and end time, center frequency, and bandwidth. (1) Backbone network The overall architecture of the backbone network is divided into multiple convolutional blocks, each of which is composed of multiple basic CBR components connected in series; the CBR component includes a convolutional layer, a batch normalization layer, and a Relu activation function. The convolution operation in the backbone network CRB component continuously improves the network's receptive field. As the network deepens, the number of output channels of the convolution layer in the CBR component is also increased to obtain signal-rich time-frequency features. At the same time, the convolution step size is set to 2 in the first convolution operation of each convolution block, so that the feature map after convolution becomes half of the original feature map in length and width, thereby obtaining feature maps of different scales output by different convolution blocks. (2) Neck network The neck network is a multi-scale time-frequency feature aggregation module, which consists of two parts: multi-scale time-frequency feature context aggregation and gated attention component; The feature maps of different scales output by different convolutional blocks in the backbone network are taken out and subjected to corresponding spatial pyramid pooling to obtain rich contextual feature information at different scales. Unify feature maps of different scales to the same resolution through convolution operations and nearest neighbor interpolation; Choose to directly add or connect feature maps of different scales along the channel dimension to obtain the fused feature map; add a gated attention component after multi-scale time-frequency feature context aggregation; Each feature map X i Input different gated attention component branches; In the corresponding gated attention component branch, the input is divided into two paths, one of which keeps the original feature map X i The other path is converted into the gated weight W after passing through the CBR component, convolution layer, and Sigmoid activation function. i , and then the two paths are element-wise multiplied to obtain the weighted output of the branch; the outputs of different gated attention component branches are weighted and calculated to obtain the final fused feature map Y, as shown in formula (2): Among them, Y, CBR(.), Conv(.) represent the final fused feature map output by the gated attention component, the CBR component, and the convolution layer respectively; S is the number of feature maps of different scales taken from the backbone network; σ is the Sigmoid activation function, ⊙ is the element-wise multiplication, and the feature map is fused at the pixel level through attention gating; the weight W i As the network converges, the gating weights gradually suppress useless information and fuse information in multi-scale feature maps that is beneficial to back-end detection; (3) Detection head network The detection head network adopts the structure of anchor-free decoding plus decoupling head. The decoupling head means that the classification and regression branches are not shared. The feature Y after the neck network fusion passes through the CBR component to decouple the features, and then obtains three output branches through the convolution layer and the Sigmoid activation function. According to the role played by different output branches, they are named category branch, positioning branch and target branch respectively; the category branch is used to output the confidence of the signal category; the positioning branch is used to output the coordinates of the signal prediction box; the target branch is to make the network output an additional prediction box quality score to judge the quality of the prediction box, and then suppress the low-quality prediction box far away from the central target when assigning labels; The loss function during network training consists of three parts, as shown in formula (3): L det =L cls +λL reg +μL obj (3) Among them, L det is the total cost function, L cls and L obj They are category loss and target loss, respectively. In training, the focal loss function and the binary cross entropy function are used respectively; Lreg is a positioning regression function, using L1 loss function in training; λ and μ are hyperparameters used to balance weights; Step 4: Time-frequency information conversion; The parameters output by the neural network include the signal category, the starting and ending x-axis coordinates (x1, x2) and the starting and ending y-axis coordinates (y1, y2) of the signal frame; assuming that the sampling rate of the signal is fs and the observation time of the time-frequency diagram is Td, the bandwidth of the signal is converted to: fs×(x2-x1), and the duration of the signal is converted to Td×(y2-y1).

Images