Broadband multi-signal detection method based on large-size convolution kernel

Abstract

The invention belongs to the technical field of cognitive radio, and particularly relates to a broadband multi-signal detection method based on a large-size convolution kernel. The method mainly comprises the following steps of: receiving a signal by adopting a receiving antenna through electromagnetic spectrum monitoring equipment, capturing a space electromagnetic radiation signal and converting the space electromagnetic radiation signal into a baseband complex signal, carrying out large FFT point short-time Fourier transform on signal sampling data to form a high-resolution and large-size time-frequency matrix, and carrying out time-frequency spectrum detection along a frequency axis of the time-frequency matrix; a time-frequency matrix is cut into a plurality of time-frequency matrix fragments adapted to neural network processing, and parameters such as confidence probability, center frequency, bandwidth, arrival time, duration and the like of existence of signals in each time-frequency matrix fragment are automatically calculated and generated by using a specially designed and trained deep neural network model. The method has the advantages that the operation of monitoring equipment is simplified, the signal detection speed is high, the detection accuracy is high, the detection recall rate is high, and the time-frequency occupancy parameter estimation precision is high.

Description

Technical Field

[0001] The present invention belongs to the technical field of cognitive radio, and in particular relates to a broadband multi-signal detection method based on a large-size convolution kernel. Background Art

[0002] With the continuous advancement of communication system technology, spectrum sensing has become increasingly important, particularly in areas such as integrated sensing and communication (ISAC), sixth-generation (6G) wireless systems, the Internet of Things (IoT), and cognitive radio. These systems require estimating the presence and time-frequency occupancy of electromagnetic signals within the system bandwidth to facilitate subsequent processing. Blind multi-signal detection is the first and most important step in spectrum sensing. Communication or spectrum sensing receivers operate over wide bandwidths, and multiple signals coexist within the receiving bandwidth, with varying radio frequency (RF) center frequencies, bandwidths, times of arrival (ToA), and durations. Determining the presence of signals and estimating their frequency, bandwidth, time of arrival, and duration in a wideband spectrum graph is crucial. Most currently proposed signal detection methods focus on a single, fixed communication channel, determining only the presence of a signal on that fixed channel. These methods are unable to detect multiple signals in wideband received radio data or estimate their parameters, such as center frequency, bandwidth, and duration. Some signal detection methods for wideband monitoring or spectrum sensing use the spectrum as input for multi-signal detection. However, these methods have difficulty detecting bursty signals and are less resilient to interference and noise.

[0003] A few studies have used short-time Fourier transforms (STFTs) or spectrograms as input to detection algorithms. Spectrograms are very effective methods for representing signals. However, these methods are relatively rare. One reason is that spectrograms are two-dimensional time-frequency matrices, making them difficult to process using classical signal processing methods. In recent years, deep learning methods have been introduced for image and language processing and have achieved remarkable performance in computer vision tasks such as image classification and object detection. These techniques demonstrate excellent capabilities in processing two- and three-dimensional matrix data. STFT data and these image data have a very similar structure. Several deep learning methods have been applied to signal detection. However, these methods typically directly use deep neural networks for object detection in computer vision tasks. For example, the Faster-RCNN model and the YOLOv3 model have been used in related studies. While these neural network models perform well in identifying objects in natural images, they do not consider the difference between the shape of objects in the image and the shape of the signal in the spectrogram. The "shape" of the signal in the spectrogram differs from the shape of the object in the image and often appears thinner.

[0004] In image object detection tasks, deep convolutional neural networks (CNNs) typically use uniformly sized convolution kernels, such as 3x3 kernels, and employ multi-layer pooling to generate multi-scale features that can accommodate both large and small objects. These methods utilize pre-designed anchors to accommodate objects of varying shapes, with varying length-to-width ratios. When the anchors' aspect ratios are set within a range from 5:1 to 1:5, they are effective for a wide range of natural objects. For example, in Faster-RCNN, the authors designed nine anchors for detecting objects of varying aspect ratios. However, in signal detection tasks, some signals have a relatively narrow bandwidth relative to the full bandwidth of the spectrogram, while others have long durations. These "signal objects" can have very large aspect ratios. Typically, these aspect ratios range from 100:1 to 1:1, making it difficult to design a limited number of anchors that can accommodate "signal objects" of varying aspect ratios. In contrast, many anchor-free object detection methods in computer vision struggle with objects of extremely high aspect ratios. Directly using deep convolutional neural networks (CNNs) for signal detection in image object detectors faces difficulties due to these characteristics. Directly applying neural network models from computer vision tasks to process two-dimensional time-frequency matrices for signal detection can result in low detection probabilities and inaccurate extraction of time-frequency parameters such as signal frequency, bandwidth, arrival time, and duration. Summary of the Invention

[0005] The purpose of the present invention is to provide a multi-signal detection method that can simplify the operation of monitoring equipment, has fast signal detection speed, high detection accuracy, high detection recall rate, and high time-frequency occupancy parameter estimation accuracy, in order to address the defects of existing broadband electromagnetic spectrum monitoring equipment such as low detection probability of various types of electromagnetic radiation signals and inaccurate time-frequency parameter extraction.

[0006] The technical solution of the present invention is:

[0007] A broadband multi-signal detection method based on a large-size convolution kernel, characterized by comprising the following steps:

[0008] S1. The broadband electromagnetic spectrum monitoring equipment receives and captures electromagnetic radiation signals from space through a receiving antenna and converts them into electrical signals. The electrical signals are converted into intermediate frequency signals through an analog receiver, and then converted into baseband complex signals through a digital receiver.

[0009] S2. Perform short-time Fourier transform on the obtained baseband complex signal to obtain a time-frequency matrix;

[0010] S3, cutting the time-frequency matrix into multiple time-frequency matrix segments of the same size along the frequency axis of the time-frequency matrix;

[0011] S4. Construct a signal detection neural network, including a feature extraction network and a prediction head network. The input of the feature extraction network is a single time-frequency matrix fragment. In the feature extraction network, the time-frequency matrix fragment is first transformed by 2D ResNet to obtain a feature tensor of fixed scale H×W×C. This tensor can be regarded as K c =H×W vectors, corresponding to K in the input matrix c The feature vectors of each region are of dimension C. The feature tensors are enhanced by horizontal convolution and vertical convolution respectively, and the horizontal enhanced feature tensors and vertical enhanced feature tensors with the number of channels of 1 / 2 of the input tensor are obtained respectively. The enhanced feature tensors are output to the prediction head network after merging.

[0012] The prediction head network consists of two branches. One is the confidence prediction branch, which is used to predict the probability of whether there is a signal at the corresponding time-frequency position in the time-frequency matrix segment. In the confidence prediction branch, the enhanced feature tensor first passes through a multi-channel convolutional neural network layer with a convolution kernel size of 1×1, and then passes through a convolution module with a convolution kernel size of 1×1 and an output channel number of 2 to obtain the signal existence probability tensor. The probability is normalized by the Softmax layer to obtain the confidence; the other branch is the parameter regression branch. The input of the parameter regression branch is the enhanced feature tensor and the signal existence probability tensor. Specifically, the dimension of the signal existence probability tensor is first expanded, and the expanded signal existence probability tensor is element-wise and element-wise compared with the enhanced feature tensor. The feature tensors of the enhanced vectors are added to obtain the confidence-filtered feature tensor, and the confidence-filtered feature tensor is reorganized by dimension to form a query vector tag Q; at the same time, the enhanced feature tensor is directly reorganized by dimension to form a key vector tag K and a value vector tag V respectively; the Q, K, and V feature tags are added to the spatial position code and sent to the attention mechanism layer for processing, and the updated feature tags are output through multi-layer attention layer processing; the obtained feature tags are reorganized to form a feature tensor with the same size as the input feature tensor, and then a 4-channel convolution with a kernel size of 1×1 is used to output the time-frequency occupancy parameters. The 4 channels represent the signal frequency, bandwidth, arrival time, and duration regressed by the neural network respectively;

[0013] S5. Use the methods of S1-S3 to create training data, define the number of training data samples as N, and train the neural network constructed in S4. The constraint function used in training is:

[0014]

[0015] in, , , are the parameters that control the weights of the three constraint functions, is the confidence constraint function:

[0016]

[0017]

[0018]

[0019] Among them, H0 and H1 represent the probability of the signal not existing and existing respectively. Indicates the number of regions in the input time-frequency matrix used to predict whether the signal exists. Indicates the region index number, Indicates area The predicted probability of the presence of a signal;

[0020] is the bounding box constraint function:

[0021]

[0022]

[0023] in, Indicates the grid point where the signal exists in the true value label, x, y represent the position of the signal in the time-frequency matrix, w represents the width of the signal in the time-frequency matrix, and h represents the height of the signal in the time-frequency matrix; correspondingly, the definition , ,but and Represent the time frequency position true value of the label box and the time frequency position predicted value of the prediction box respectively;

[0024] is the intersection-and-union ratio constraint function. The intersection-and-union ratio constraint function of the i-th sample is:

[0025]

[0026]

[0027] in, and Indicates a label box With the true value box The intersection area and union area of ​​;

[0028] After training, a trained signal detection neural network is obtained;

[0029] S6. After processing the received signal according to the methods of S1-S3, the received signal is input into the trained signal detection neural network, and then the detection results of each time-frequency matrix fragment output by the network are integrated to complete the detection and discovery of multiple signals in the electromagnetic space and the estimation of signal time and frequency occupancy parameters.

[0030] The present invention has the beneficial effects of simplifying the operation of electromagnetic spectrum monitoring equipment and improving the ability to simultaneously detect multiple signals. The present invention, focusing on the automated detection of magnetic radiation signals by broadband electromagnetic spectrum monitoring equipment, finds that the electromagnetic spectrum monitoring equipment uses a broadband antenna to receive signals, automatically amplifies, analog-converts, and filters the signals through a radio frequency channel, and outputs them to a digital receiving module. The digital receiving module automatically samples and digitally down-converts the intermediate frequency analog signal output by the receiving channel to form a baseband complex digital signal, which is then subjected to a high-resolution short-time Fourier transform to form a large-scale time-frequency matrix representing the broadband electromagnetic multi-signal. After the detection processing automatically cuts the large-scale time-frequency matrix into preset sizes, a deep neural network model is used to automatically process the time-frequency matrix slices and generate signal presence and signal time and frequency domain parameters. Finally, the inference results of each time-frequency matrix slice are fused to form the signal detection results in the original large-scale time-frequency matrix. The entire process does not require manual setting of detection thresholds or operation of equipment. The machine automatically extracts parameters such as the signal time occupancy range and frequency occupancy range, greatly simplifying the design of the monitoring equipment, reducing manual operation, and improving the reliability of the project implementation. The signal detection neural network model processes bandwidth multi-signal time-frequency data. All results in the time-frequency data can be detected and output in parallel in one processing, which improves the ability to detect multiple signals simultaneously.

[0031] The monitoring equipment's ability to characterize and detect narrowband signals has been improved. The present invention adopts a method of fusing the results after splitting the inference of a high-resolution time-frequency matrix, targeting the characteristics of narrow bandwidth of electromagnetic space signals and wide instantaneous monitoring bandwidth of monitoring equipment in actual application scenarios. When performing the STFT transformation of the signal, an ultra-large number of points up to 131,072 are used for FFT calculation to achieve high resolution of the signal spectrum within the wide monitoring bandwidth. Compared with the extreme FFT of about 8,192 points of traditional monitoring equipment, the resolution of narrowband signals is greatly enhanced. The large-scale time-frequency matrix is ​​cut along the frequency dimension to form time-frequency matrix fragments that are suitable for neural network reasoning. The deep neural network model can perform parallel reasoning of multiple samples on a high-performance GPU or a dedicated AI processing chip, and multiple time-frequency matrix fragments are reasoned in parallel. After the obtained results are fused, they are used as the result of the large-scale time-frequency matrix before cutting. The detection capability of narrowband signals has been greatly improved.

[0032] This improves the detection probability of various signals and the accuracy of time-frequency occupancy parameter estimation. Building on a typical residual network, specially designed large-scale one-dimensional horizontal and vertical convolution kernels are used to further enhance the perception and processing capabilities of narrowband signals in the time-frequency matrix, forming enhanced features. This improves the detection probability of multiple signals. A specially designed attention-based decoder neural network model is used as the prediction head network to regress parameters such as the center frequency, bandwidth, arrival time, and duration of each signal in the time-frequency matrix. Thanks to the global receptive field characteristics of the decoder neural network model, the detection probability of longer-duration signals and the accuracy of parameter estimation, such as arrival time and duration, are greatly improved.

[0033] There is no need for multi-scale feature processing and result merging, and the processing speed is fast. The present invention uses a feature map of one scale to realize the detection of signals of different forms such as multiple bandwidths and multiple signal durations and the estimation of time-frequency occupancy parameters. Most of the current signal detection methods based on deep neural networks use neural networks to extract features, and form feature maps of different sizes at different layers, and use the corresponding prediction head network to perform signal detection of the corresponding bandwidth and duration for each feature map. The signal results detected at multiple scales are then fused to form the final detection result of the neural network processing at that time. The present invention only needs to use the features of one scale formed by the neural network, and use specially designed large-scale one-dimensional horizontal and vertical convolution kernels to enhance the features of this scale, and then use the decoder network model with a global receptive field as the prediction network, so that the features of one scale can be used to detect signals of different forms such as multiple bandwidths and multiple signal durations and estimate the time-frequency occupancy parameters. There is no need for feature calculation, prediction, and result fusion of multiple scales, and the processing speed is fast, which improves the detection processing efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a flow chart of a broadband multi-signal detection method based on a large-size convolution kernel of the present invention.

[0035] Figure 2 This is a schematic diagram of the structural principle of a broadband multi-signal detection neural network based on a large-size convolution kernel in the present invention.

[0036] Figure 3 This is a structural principle diagram of the large-size convolution kernel feature enhancement module of the present invention.

[0037] Figure 4 This is a structural principle diagram of the decoder prediction head module based on the multi-layer attention mechanism of the present invention.

[0038] Figure 5 This is a schematic diagram showing the design principle of the time-frequency occupancy parameters (including signal center frequency, signal bandwidth, signal arrival time, and signal duration) of the present invention. DETAILED DESCRIPTION

[0039] The practicability of the present invention will be described below with reference to the accompanying drawings.

[0040] like Figure 1 As shown, in the electromagnetic spectrum sensing scenario corresponding to the present invention, the broadband electromagnetic spectrum monitoring equipment uses a receiving antenna to receive signals, captures spatial electromagnetic radiation signals and converts them into electrical signals. The signals are amplified by a low-noise amplifier and converted into intermediate frequency signals through bandpass filtering and mixing. The signals are then digitized through AD sampling, and the intermediate frequency signals are converted into baseband complex signals using digital mixing and digital filtering. The signal sampling data is then subjected to a high-resolution short-time Fourier transform (STFT), using a large FFT point count to form a high-resolution, large-scale time-frequency matrix. The time-frequency matrix is ​​then cut along the frequency axis into multiple time-frequency matrix segments suitable for neural network processing and output to the detection and processing module. Specifically, for a complex signal with an instantaneous bandwidth of 60 MHz at a 96 MHz sampling rate, the present invention uses a 131,072-point FFT. Every 1,024 FFTs form a frame of the time-frequency matrix. In the frequency domain, only the effective spectral components within the 60 MHz bandwidth are selected, resulting in an effective time-frequency matrix size of 1,024 × 81,920. A single time-frequency matrix segment of 1024×2048 is then used as the data for a single inference in the signal detection neural network model. A valid original time-frequency matrix is ​​sliced ​​into 40 time-frequency matrix segments (81920 / 2048 = 40). The signal detection module uses high-precision floating-point operators to construct a deep neural network model. The deep neural network model is trained with high-precision labeled data. After loading the trained deep neural network model, the broadband time-frequency matrix segment data is fed into the deep neural network model. After inference and post-processing of the inference results, the deep neural network model automatically generates parameters such as the confidence probability of the signal's presence in the electromagnetic space, the signal's center frequency, signal bandwidth, signal arrival time, and signal duration. The signal detection neural network model uses a residual network as the feature extraction network and employs two large one-dimensional convolution kernels in the horizontal and vertical directions to further enhance features. A convolutional neural network is used as the signal confidence prediction head network, and an attention-based decoder neural network is used as the time-frequency occupancy parameter prediction head network for the signal's center frequency, signal bandwidth, signal arrival time, and signal duration. The results of the confidence prediction head network and the parameter prediction head network are combined to form the signal detection results of the time-frequency matrix segments that are fed into the neural network model for processing. Finally, the detection results of the entire large-scale, high-resolution time-frequency matrix are fused together to output the detection results. This completes the detection of multiple signals in the electromagnetic space and the estimation of the signal time and frequency occupancy parameters.

[0041] like Figure 2As shown in the figure, the overall design of the signal detection neural network model is as follows: for each segmented time-frequency matrix data, the input size of the neural network model is 1024×2048. First, a typical 2D ResNet is used for feature transformation processing. In this network model, max pooling is used, and downsampling is performed four times, each downsampling by a factor of 2. At the same time, the feature channels are expanded to C (typically 256) channels. After 2D ResNet processing, the feature tensor size is 64×128×C, which means that the feature tensor has 64×128 grid points, each grid point is a vector with C elements. A large-scale convolution kernel feature enhancement module is then used to enhance the features, while maintaining the same size of the enhanced feature tensor. The decoder prediction head module, based on a multi-layer attention mechanism, then processes the features and generates the signal detection results. The prediction head module has two branches. The first is the confidence prediction branch. After processing features using a multi-channel convolution with a 1×1 kernel, a convolution module with a 1×1 kernel and 2 output channels is used, followed by probability normalization using a softmax layer. This is the confidence prediction branch. The convolution operation in this confidence prediction branch is calculated for each position in the feature tensor of size 64×128×C. The resulting confidence prediction probability tensor has dimensions H×W×2=64×128×2. Each 1×1×2 vector represents the probability of a signal existing at each spatial location corresponding to the feature tensor. This probability is represented by two prediction results: the probability of signal presence and absence. The second branch of the prediction module is the time-frequency occupancy parameter regression branch, which processes and calculates the four time-frequency occupancy parameters of the signal at each spatial location corresponding to the feature tensor: frequency, bandwidth, arrival time, and duration.

[0042] like Figure 3 As shown in the figure, the large-scale convolution kernel feature enhancement module is designed as follows: for an input feature tensor of size H×W×C (typically 64×128×256), it processes it using two large-scale convolution kernels: vertical convolution and horizontal convolution, each with half the number of input channels. This is followed by a LayerNorm normalization layer and a Leaky ReLU activation function layer. This generates a horizontally enhanced feature tensor and a vertically enhanced feature tensor, both of the same size as the input feature tensor and with half the number of channels. Finally, these two feature tensors are concatenated and merged into the enhanced feature tensor output. This feature tensor maintains the same size as the input feature tensor. A key feature of this module is the use of large-scale convolution kernels for both vertical and horizontal convolutions. Typically, for an input feature tensor of size 64×128, the vertical convolution kernel size is 65×1, and the horizontal convolution kernel size is 1×127, meaning the kernel size is comparable to the input tensor dimensions.

[0043] like Figure 4Figure 2 shows the design of the decoder prediction head module based on a multi-layer attention mechanism. It primarily consists of two branches: the confidence prediction branch, which predicts the probability of a signal being present at the corresponding time-frequency position in the time-frequency matrix segment. This branch employs a multi-channel convolutional neural network layer with a 1×1 kernel, followed by a convolutional module with a 1×1 kernel and 2 output channels, and finally a softmax layer for probability normalization. The second branch is the parameter regression branch. This branch uses the signal presence probability tensor (H×W×1, typically 64×128×1) from the confidence prediction branch, expands its dimensions to H×W×C, and then adds it element-wise to the input feature tensor to produce a confidence-filtered feature tensor. This feature tensor is reshaped to form a query vector token, denoted as Q. The input feature tensor is directly reshaped and copied to form a key vector token, denoted as K, and a value vector token, denoted as V. The Q, K, and V feature labels are added to the spatial position encoding and then fed into the attention mechanism layer for processing. Multiple layers of attention are processed to output updated feature labels. These feature labels are then re-dimensionalized to form a feature tensor with the same dimensions as the input feature tensor. A 4-channel convolution with a kernel size of 1×1 is then used to output time-frequency occupancy parameters. The four channels represent the signal frequency, bandwidth, arrival time, and duration regressed by the neural network. The result of the entire parameter regression branch is an H×W×4 tensor. Its size is the same as the confidence prediction branch. Each 1×1×4 vector in this feature tensor represents the four parameters indicating the presence of a signal at each spatial location corresponding to a grid point in the feature tensor, corresponding one-to-one to the two probability parameters of the confidence prediction branch.

[0044] like Figure 5 As shown, the signal time-frequency position representation of the present invention is designed as follows: the time-frequency position of the signal is represented by 4 integers, x, y, w and h. Among them, x, y represent the position of the signal in the time-frequency matrix, corresponding to the signal center frequency f c and the center time (signal arrival time t a plus the duration t l half of the signal). w represents the width of the signal in the time-frequency matrix, corresponding to the signal bandwidth B s Similarly, h represents the height of the signal within the time-frequency matrix, corresponding to the duration of the signal t l . Real signal parameter [f c ,t a ,B s ,t l The relationship between ] and its representation in the time-frequency matrix [x,y,w,h] is:

[0045] (1)

[0046] When training a deep neural network model, the training constraints of the model mainly include three parts, namely confidence constraint, bounding box constraint, and intersection over union (IoU) constraint. The final output of the neural network model for a single sample is H×W×6. The vector of each grid point in the tensor (size 1×1×6) is called the probability of whether there is a signal at the spatial position corresponding to the grid point and the time-frequency occupancy parameters of the signal. For a 1×1×6 vector, 2 elements represent the confidence probability, and 4 elements represent parameters such as signal frequency, bandwidth, arrival time, and duration. Among them, a single sample i is at the grid point s k The confidence constraint is expressed as:

[0047] (2)

[0048] and

[0049] (3)

[0050] H0 and H1 represent the probability of the signal not existing and existing respectively. The total confidence constraint is

[0051] (4)

[0052] N is the total number of training samples. The bounding box for a single sample i is:

[0053] (5)

[0054] denoting the grid points where there is a signal in the ground truth label, , They represent the time-frequency position true value of the label box and the time-frequency position predicted value of the prediction box. The total bounding box constraint is:

[0055] (6)

[0056] The Intersection over Union (IoU) constraint for a single sample i is:

[0057] (7)

[0058] and

[0059] (8)

[0060] in and Indicates a label box With the true value box The intersection area and union area of ​​. Its calculation method adopts the common calculation method in the industry. The overall constraint function is

[0061] (8)

[0062] in , , are the parameters that control the weights of the three constraint functions.

[0063] In an optional embodiment, the input channel amplitude data of the neural network is 1×1024×2048, and the input data forms a 64×128×256 feature tensor after passing through the feature extraction network. The confidence prediction branch of the prediction head network forms a 64×128×2 prediction probability tensor through convolution and Softmax processing. The parameter regression branch of the prediction head network forms a tag sequence with a tag dimension of 256 for each tag and a total of 8192 tags through labeling processing. Then, after processing using the attention layer, there are also 8192 tag sequences with a dimension of 256. Then, the dimension is reconstructed into a 64×128×256 tensor, and then a 64×128×4 prediction time-frequency occupancy parameter tensor is formed through convolution processing.

[0064] The present invention has been described in detail above with reference to the accompanying drawings. However, it should be noted that the examples described above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Those skilled in the art will appreciate that the present invention is susceptible to various modifications and variations. For example, different ResNet or other neural networks with different numbers of layers can be selected as feature extraction networks based on specific engineering projects, and the number of channels in each layer of the neural network, the size parameters of the convolution kernel, and so on can be varied. Any modifications, equivalent substitutions, improvements, and the like made within the spirit and principles of the present invention shall be included within the scope of the claims of the present invention.

Claims

1. A broadband multi-signal detection method based on a large-size convolution kernel, characterized in that: The following steps are involved: S1. The broadband electromagnetic spectrum monitoring equipment receives and captures electromagnetic radiation signals from space through a receiving antenna and converts them into electrical signals. The electrical signals are converted into intermediate frequency signals through an analog receiver, and then converted into baseband complex signals through a digital receiver. S2. Perform short-time Fourier transform on the obtained baseband complex signal to obtain a time-frequency matrix; S3, cutting the time-frequency matrix into multiple time-frequency matrix segments of the same size along the frequency axis of the time-frequency matrix; S4. Construct a signal detection neural network, including a feature extraction network and a prediction head network. The input of the feature extraction network is a single time-frequency matrix fragment. In the feature extraction network, the time-frequency matrix fragment is first transformed by 2D ResNet to obtain a feature tensor of fixed scale H×W×C. The feature tensor is regarded as K c =H×W vectors, corresponding to K in the input matrix c The feature vectors of each region are of dimension C; the feature tensors are enhanced by horizontal convolution and vertical convolution respectively, and the horizontal enhanced feature tensors and vertical enhanced feature tensors with the number of channels of 1 / 2 of the input tensor are obtained respectively. After merging, the enhanced feature tensors are output to the prediction head network; The prediction head network consists of two branches. One is the confidence prediction branch, which is used to predict the probability of whether there is a signal at the corresponding time-frequency position in the time-frequency matrix segment. In the confidence prediction branch, the enhanced feature tensor first passes through a multi-channel convolutional neural network layer with a convolution kernel size of 1×1, and then passes through a convolution module with a convolution kernel size of 1×1 and an output channel number of 2 to obtain the signal existence probability tensor. The probability is normalized by the Softmax layer to obtain the confidence; the other branch is the parameter regression branch. The input of the parameter regression branch is the enhanced feature tensor and the signal existence probability tensor. Specifically, the dimension of the signal existence probability tensor is first expanded, and the expanded signal existence probability tensor is element-wise and element-wise compared with the enhanced feature tensor. The feature tensors of the enhanced vectors are added to obtain the confidence-filtered feature tensor, and the confidence-filtered feature tensor is reorganized by dimension to form a query vector tag Q; at the same time, the enhanced feature tensor is directly reorganized by dimension to form a key vector tag K and a value vector tag V respectively; the Q, K, and V feature tags are added to the spatial position code and sent to the attention mechanism layer for processing, and the updated feature tags are output through multi-layer attention layer processing; the obtained feature tags are reorganized to form a feature tensor with the same size as the input feature tensor, and then a 4-channel convolution with a kernel size of 1×1 is used to output the time-frequency occupancy parameters. The 4 channels represent the signal frequency, bandwidth, arrival time, and duration regressed by the neural network respectively; S5. Use the methods of S1-S3 to create training data, define the number of training data samples as N, and train the neural network constructed in S4. The constraint function used in training is: ， in, , , are the parameters that control the weights of the three constraint functions, is the confidence constraint function: ， ， ， Among them, H0 and H1 represent the probability of the signal not existing and existing respectively. Indicates the number of regions in the input time-frequency matrix used to predict whether the signal exists. Indicates the region index number, Indicates area The predicted probability of the presence of a signal; is the bounding box constraint function: ， ， in, Indicates the grid point where the signal exists in the true value label, x, y represent the position of the signal in the time-frequency matrix, w represents the width of the signal in the time-frequency matrix, and h represents the height of the signal in the time-frequency matrix; correspondingly, the definition , ,but and Represent the time frequency position true value of the label box and the time frequency position predicted value of the prediction box respectively; is the intersection-and-union ratio constraint function. The intersection-and-union ratio constraint function of the i-th sample is: ， ， in, and Indicates a label box With the true value box The intersection area and union area of ​​; After training, a trained signal detection neural network is obtained; S6. After processing the received signal according to the methods of S1-S3, the received signal is input into the trained signal detection neural network, and then the detection results of each time-frequency matrix fragment output by the network are integrated to complete the detection and discovery of multiple signals in the electromagnetic space and the estimation of signal time and frequency occupancy parameters.

Images