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

Abstract

Disclose a kind of intelligent signal detection identification method based on time-frequency feature fusion and anchor-free detection head, comprising the following steps: wireless signal broadband acquisition;Time-frequency transform preprocessing;Signal detection identification based on neural network;Time-frequency information conversion.This method proposes a network suitable for signal detection identification, which is divided into backbone network, neck network and detection head network three parts.A backbone network composed of component-convolution block architecture is proposed, which can flexibly adjust the complexity of the network to effectively extract features.A multi-scale time-frequency feature aggregation module is inserted in the neck network to efficiently perform feature fusion.The detection head network is changed to an anchor-free mechanism to improve output efficiency.The final output of the network is converted to time-frequency information to obtain parameters such as signal bandwidth and duration.The method can provide support for fast and automatic signal detection and identification, effectively solving the problems of high human involvement and signal detection and identification rate in military communication intelligence reconnaissance.

Description

TECHNICAL FIELD

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

[0002] Communication signal detection and recognition is a key technology for cognitive radio, interference detection and other applications. Its task is to detect signals from received wideband signals with or without prior knowledge, while estimating signal categories, frequencies, bandwidths, durations and other parameters. With the massive growth of radio frequency data in wireless environments, communication signal detection and recognition technology has become increasingly important and is attracting attention in both civilian and military fields.

[0003] Traditional communication signal detection and recognition techniques 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 designing effective signal features relies on expert knowledge and is difficult to effectively generalize to other radio frequency data.

[0004] In recent years, deep learning-based methods have been widely used in radio frequency signal processing and have shown excellent performance in modulation recognition, signal detection and channel estimation. Most of these deep learning-based signal detection and recognition algorithms convert signals into time-frequency graphs through time-frequency transformation, which is an image representation, and then signal detection in time-frequency graphs becomes similar to target detection in computer vision. Therefore, previous methods use some classic target detection algorithms, including Faster-RCNN, SSD, YOLO series, etc. However, these methods still maintain the original network architecture of target detection algorithms and do not make targeted improvements for signal time-frequency graph datasets, making it difficult to achieve optimal detection performance in specific tasks. Specifically, these methods are anchor-based algorithms and will encounter anchor mismatch problems when dealing with signals with variable bandwidth and duration, which will lead to performance degradation. In addition, these methods do not focus on the neural network design of feature fusion, which is not conducive to achieving optimal performance. SUMMARY

[0005] To address the shortcomings of existing intelligent detection and recognition algorithms in the field of wireless signal detection and recognition, the present application proposes an intelligent signal detection and recognition method based on time-frequency feature fusion and anchor-free detection head, including the following steps:

[0006] Step 1: Wideband capture of wireless signals;

[0007] Set the center frequency and receive bandwidth according to the task to capture the complex baseband IQ signal stream containing the target signal; other signals and in-band noise interference are also present in the received signal;

[0008] Step 2: Time-frequency transform preprocessing

[0009] Time-frequency transform preprocessing converts the time-domain signal into the time-frequency domain using time-frequency transform such as short-time Fourier transform (STFT) and wavelet analysis, and generates a signal time-frequency graph, which is then used as a data set for deep learning. If x(t) is a complex time series after sampling, the STFT result of x(t) is represented as:

[0010]

[0011] where S(n, m) is the time-frequency matrix obtained by short-time Fourier transform, n and m are the frequency and time indices respectively, w(k) is an analysis window with length N, and O is the number of overlapping points between adjacent windows. The modulus or logarithmic modulus of S(n, m) is saved as a time-frequency graph. After 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.

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

[0013] The transformed time-frequency graph in step 2 is input into the trained neural network, which automatically outputs parameters such as the category of the signal and the regression coordinates.

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

[0015] During the training phase, an 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 fuses the features, the detection head network outputs the category, positioning regression, and target. The neural network output is further processed by non-maximum suppression and time-frequency information conversion to obtain the category, start and end time, center frequency, and bandwidth of the signal.

[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 in series. A CBR component includes a combination of a convolutional layer, a batch normalization layer, and a Relu activation function.

[0018] The convolution operation in the backbone network CBR component continuously improves the receptive field of the network, and as the network deepens, the number of output channels of the convolution layer in the CBR component is also increased, so that the time-frequency features of the signal are rich; at the same time, the convolution step 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 different scale feature maps 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 a multi-scale time-frequency feature context aggregation and a gated attention component;

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

[0022] The different scale feature maps are selected to be directly added or connected along the channel dimension to obtain the fused feature map; the gated attention component is added after the multi-scale time-frequency feature context aggregation;

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

[0024]

[0025] where 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 different scale feature maps taken from the backbone network; σ is the Sigmoid activation function, and is the element multiplication, which performs pixel-level fusion on the feature map through attention gating; the weight W i As the network converges, the gating weight gradually suppresses useless information and fuses information in the multi-scale feature map that is beneficial to the backend detection;

[0026] (3) Detection head network

[0027] The detection head network adopts an anchor-free decoding manner and a decoupling head structure, and the decoupling head refers to that the classification branch and the regression branch are not shared;

[0028] The fused features Y of the neck network are decoupled after passing through the CBR component, and then three output branches are obtained through a convolution layer and a Sigmoid activation function; according to the functions of the different output branches, they are named as a category branch, a positioning branch and a target branch; the category branch is used for outputting the confidence of the signal category; the positioning branch is used for outputting the coordinates of the signal prediction box; the target branch is used for outputting a quality score of a prediction box to judge the quality of the prediction box, and then suppressing the low-quality prediction box far away from the center target during label assignment;

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

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

[0031] Wherein, L det is the total cost function, L cls and L obj are the category loss and the target loss respectively, and the focal loss function and the binary cross entropy function are used in the training; L reg is the positioning regression function, and the L1 loss function is used in the training; λ and μ are hyperparameters for balancing the weights;

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

[0033] 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 box; 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 into: fs*(x2-x1), and the duration of the signal is converted into Td*(y2-y1).

[0034] The neural network architecture designed in the application adopts a backbone network with flexible complexity for extracting signal features; a neck network is inserted to effectively fuse features by using multi-scale time-frequency feature context aggregation and a gated attention component; meanwhile, the anchor detection head in the previous method is replaced by an anchor-free structure, which avoids calculating the prior anchor box and is conducive to reducing the model parameters and network inference. The method has obviously better detection performance on the constructed voice signal dataset, and the model parameters and inference time are better than other deep learning-based signal detection and recognition methods.

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

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

[0037] Figure 2 An example of a voice signal time-frequency pattern is shown;

[0038] Figure 3 The overall architecture of the neural network designed by the application is shown;

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

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

[0041] Figure 6 The detection and recognition visualization result of the actual collected voice signal by the application is shown. DETAILED DESCRIPTION

[0042] The flow chart of the intelligent signal detection and recognition method based on time-frequency feature fusion and anchor-free detection head is shown in Figure 1 , which includes four steps of wireless signal wideband capture, time-frequency transformation preprocessing, neural network-based signal detection and recognition, and time-frequency information conversion, and is specifically as follows.

[0043] Step 1: wireless signal wideband capture;

[0044] According to the task setting center frequency and receiving bandwidth, the complex baseband IQ signal stream containing the target signal is captured. Since the application adopts a wideband receiver, there are also other signals and in-band noise interference in the received signal.

[0045] Step 2: time-frequency transformation preprocessing;

[0046] Time-frequency transformation preprocessing refers to converting the time-domain signal to the time-frequency domain and generating a signal time-frequency pattern by using time-frequency transformation such as short-time Fourier transform (STFT) and wavelet analysis, and then using it as a deep learning dataset. If x(t) is a sampled complex time sequence, the STFT result of x(t) can be expressed as:

[0047]

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

[0049] Figure 2 is an example of converting a voice signal from a time sequence to a time-frequency graph, which is generated using the short-time Fourier transform described above, the light background represents the background noise interference, and the dark target surrounded by the frame is the representation form of the signal in the time-frequency graph, and the horizontal and vertical coordinates of the frame 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] The signal detection and recognition based on neural network means that the transformed time-frequency graph in step 2 is input into the trained neural network, and the neural network automatically outputs the category, regression coordinates and other parameters of the signal.

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

[0053] The neural network is refined into three parts: backbone network, neck network and detection head network, and the three network parts are improved according to the characteristics of the actual signal detection task: first, the signal detection in the time-frequency graph data set has lower complexity than the general target detection task in natural images, so the present application designs a backbone network with a component-convolution block architecture, which can adjust the complexity of the network by flexibly configuring the number of components and convolution blocks to better adapt to the signal detection task, as described below. In addition, the neck network for feature fusion in the classic target detection algorithm is designed according to the characteristics of the specific benchmark general target data set, and cannot well adapt to the distribution of the signal time-frequency graph data set. Therefore, the present application designs a new neck network, which realizes better time-frequency feature fusion by using multi-scale time-frequency feature context aggregation and gated attention components, and the 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 the obvious variation of the length-width ratio of the signal target frame, which causes the anchor frame mismatch problem of the detection neural network based on anchor frame, and further leads to performance degradation, so the anchor detection head network is replaced by the anchor-free mechanism, which is beneficial to model parameter reduction and network inference while avoiding the calculation of prior anchor frames, as described below.

[0054] In the training stage, first input the RGB three-channel time-frequency image with a batch size of B, extract the signal time-frequency features through the backbone network, and then obtain the class, positioning regression, and target three parts of the output after the feature fusion of the neck network and the detection head network. After the neural network output continues to go through post-processing such as non-maximum suppression and time-frequency information conversion, the class, start and end time, center frequency, and bandwidth of the signal time-frequency information can be obtained.

[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 has a simpler background and less color change within the signal. At the same time, there is only one type of target in the task data set. Therefore, directly using the complex backbone network of the existing algorithm may not be able to efficiently extract signal features. Therefore, designing a task-oriented network width and depth is not only beneficial to the feature fusion of the neck network and the classification regression of the detection head network, but also can reduce the network inference time and meet the real-time demand of signal detection.

[0057] As shown in Figure 4 , the overall architecture of the designed backbone network is divided into multiple convolution blocks, and each convolution block is composed of multiple basic CBR components in series. The CBR component includes a combination of a convolution layer (Convolution), a batch normalization layer (Batchnormalization), and a Relu activation function. In the actual configuration of the backbone network, the network depth and the richness of the feature map obtained 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, and the convolution channel number of the convolution layer in the component. This way can flexibly adjust the network complexity, and also facilitates code reuse of convolution blocks and components.

[0058] The backbone network proposed in the present application can flexibly adjust the number of components. In the case of a simple task, the number of components and convolution stages can be reduced to achieve the purpose of simplifying the network.

[0059] The convolution operation in the CRB component of the backbone network continuously improves the receptive field (which is well known to those skilled in the art), and as the network deepens, the output channel number of the convolution layer in the CBR component is also increased, thereby obtaining rich signal time-frequency features. At the same time, the convolution step 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, and therefore different scale feature maps can be obtained from different convolution blocks.

[0060] (2) Neck network

[0061] The neck network mainly realizes efficient feature fusion. For this purpose, the application designs a multi-scale time-frequency feature aggregation module as shown in Figure 5 The module is composed of a multi-scale time-frequency feature context aggregation and a gated attention component.

[0062] First, the different scale feature maps output by different convolution blocks in the backbone network are taken out, respectively passing 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 on 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 operation and nearest neighbor interpolation, which is well known to those skilled in the art and will not be repeated.

[0063] At this time, one can choose to directly add or connect the feature maps of different scales along the channel dimension to obtain the fused feature map, as is well known to those skilled in the art. However, the feature maps from different convolution blocks have semantic gaps, and simple fusion may bring poor results. Therefore, after the multi-scale time-frequency feature context aggregation as shown in Figure 5 The gated attention component is added.

[0064] As shown in Figure 5 Each feature map X i will 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 unchanged, and the other passes through the CBR component, convolution layer and Sigmoid activation function in turn to convert into the gating weight W i (CBR component, convolution layer, Sigmoid activation function are well known to those skilled in the art), and then the two paths are multiplied to obtain the weighted output of the branch. The outputs of different gated attention component branches are weighted and summed to obtain the final fused feature map Y, as shown in equation (2):

[0065]

[0066] Where Y, CBR(.), and 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 extracted from the backbone network. Figure 5 In the example, it's 3. σ is the Sigmoid activation function, ⊙ represents element-wise multiplication, and attention gating allows for pixel-level fusion of feature maps. Furthermore, the weight W... i It is obtained automatically through training. As the network converges, the gating weights gradually suppress useless information and fuse information that is beneficial to backend detection from the multi-scale feature maps.

[0067] (3) Detection head network

[0068] Object detection algorithms can be divided into anchored and anchorless types. Anchor boxes are a set of prior rectangular boxes. When anchor boxes are similar to the ground truth target boxes in size and aspect ratio, it is beneficial for target box regression. Therefore, the setting of anchor boxes affects the performance of detection algorithms to some extent. Previous methods obtained more suitable anchor boxes by improving the K-means algorithm, but this increased the algorithm latency. Therefore, this invention replaces the anchored mechanism with an anchorless decoding method with a decoupling head structure, where the decoupling head means that the classification and regression branches do not share the same structure.

[0069] like Figure 3 As shown, the features Y fused by the neck network are decoupled after passing through the CBR component, and then three output branches are obtained through convolutional layers and the sigmoid activation function. Based on their different roles, these branches are named the category branch, localization branch, and target branch. The category branch outputs the confidence score of the signal category. The localization branch outputs the coordinates of the predicted bounding box. The target branch causes the network to output an additional quality score for the predicted bounding box to evaluate its quality, thus suppressing low-quality predicted boxes far from the central target during label assignment.

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

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

[0072] Among them, L det Let L be the total cost function. cls and L obj These are the class loss and the target loss, respectively, using the focus loss function and the binary cross-entropy function during training; L reg λ is the localization regression function, and L1 loss is used during 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 signal category, start and end x-axis coordinates (x1, x2) and 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 detection and recognition visualization results of the voice signal collected by the present application are given. The dark detection frame in the figure is drawn according to the coordinates output by the neural network, and the numerical value above the detection frame is the signal category confidence output by the neural network. It can be seen that the present application can not only accurately locate the signal in the time-frequency diagram, but also has excellent recognition confidence.

[0076] The present application 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 the category, bandwidth and duration of the signal are obtained through time-frequency information conversion.

Claims

1. A smart signal detection and recognition method based on time-frequency feature fusion and anchorless detection head, characterized in that, Includes the following steps: Step 1: Broadband wireless signal acquisition; The center frequency and receiving bandwidth are set according to the mission to capture the complex baseband IQ signal stream containing the target signal; other signals and in-band noise interference also exist in the received signal. Step 2: Time-frequency transformation preprocessing; The time-frequency transformation preprocessing uses Short-Time Fourier Transform (STFT) and wavelet analysis to transform the time-domain signal to the time-frequency domain and generate a time-frequency graph of the signal, which is then used as a dataset for deep learning. If x(t) is a sampled complex time series, then the STFT result of x(t) is expressed as: Where S(n,m) is the time-frequency matrix obtained by short-time Fourier transform, and n and m are the frequency and time indices, respectively; w(k) is an analysis window of length N, and O is the number of overlapping points between adjacent windows; the modulus value of S(n,m) or the modulus value after logarithmic transformation is saved as a time-frequency image; after time-frequency transformation preprocessing, the signal representation changes from a one-dimensional time series to the foreground signal in a two-dimensional time-frequency image; Step 3: Signal detection and recognition based on neural networks; The transformed time-frequency graph from step 2 is fed into the trained neural network, which automatically outputs parameters such as the signal category and regression coordinates. A neural network suitable for signal detection and recognition is proposed, which is refined into three parts: a backbone network, a neck network, and a detection head network. The backbone network utilizes multi-scale time-frequency feature context aggregation and a gated attention component to achieve time-frequency feature fusion. The anchored detection head network is replaced with an anchorless mechanism. During the training phase, RGB three-channel time-frequency images with a batch size of B are input. After the backbone network extracts the time-frequency features of the signal and the neck network fuses the features, the head network outputs three parts: category, localization 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 time-frequency information such as the signal category, start and end time, center frequency, and bandwidth. (1) Backbone Network The backbone network architecture is divided into multiple convolutional blocks, each of which is composed of multiple basic CBR components connected in series. Each CBR component includes a combination of a convolutional layer, a batch normalization layer, and a ReLU activation function. The convolutional operations in the CRB component of the backbone network continuously increase the receptive field of the network. As the network deepens, the number of output channels of the convolutional layers in the CBR component also increases, resulting in rich time-frequency features. At the same time, the convolution stride is set to 2 in the first convolutional operation of each convolutional block, so that the feature map after convolution becomes half the length and width of the original feature map, thus obtaining feature maps of different scales output by different convolutional 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 a gated attention component. The feature maps of different scales output by different convolutional blocks in the backbone network are extracted and processed by the corresponding spatial pyramid pooling to obtain rich contextual feature information at different scales. Feature maps of different scales are unified to the same resolution through convolution operations and nearest neighbor interpolation. The fused feature map is obtained by directly summing or connecting feature maps of different scales along the channel dimension; a gated attention component is added 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 split into two paths, one of which retains the original feature map X. i The other path remains unchanged, passing through a CBR component, a convolutional layer, and a Sigmoid activation function before being converted into gated weights W. i Then, the two paths are multiplied element by element to obtain the weighted output of the branch; the outputs of different gated attention component branches are weighted and summed to obtain the final fused feature map Y, as shown in Equation (2): Where Y, CBR(.), and 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 extracted from the backbone network; σ is the Sigmoid activation function, ⊙ represents element-wise multiplication, and pixel-level fusion of feature maps is performed through attention gating; the weight W i As the network converges, the gating weights gradually suppress useless information and fuse information from multi-scale feature maps that is beneficial to backend detection. (3) Detection head network The detection head network adopts an anchorless decoding method with a decoupling head structure. The decoupling head means that the classification and regression branches do not share the same structure. After the neck network fusion, the features Y are decoupled by the CBR component and then passed through convolutional layers and the Sigmoid activation function to obtain three output branches. According to the role of different output branches, they are named class branch, localization branch and target branch respectively. The class branch is used to output the confidence of the signal class; the localization branch is used to output the coordinates of the signal prediction box; the target branch makes the network output an additional quality score of the prediction box to judge the quality of the prediction box, thereby suppressing low-quality prediction boxes far from the central target during label assignment. The loss function during network training consists of three parts, as shown in equation (3): L det =L cls +λL reg +μL obj (3) Among them, L det Let L be the total cost function. cls and L obj The loss functions are classified and target, respectively, and the focus loss function and the binary cross-entropy function are used during training. Lreg λ is the localization regression function, which uses the L1 loss function during training; λ and μ are hyperparameters used to balance the weights. Step 4: Time-frequency information conversion; 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 box; assuming the sampling rate of the signal is fs and the observation time of the time-frequency plot 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