Signal detection methods, devices, and media based on U-Net and spectrum.

Abstract

This invention discloses a signal detection method, device, and medium based on U-Net and spectrum, belonging to the interdisciplinary field of signal processing and artificial intelligence. The method includes the following steps: calculating the spectral feature map of the signal frame to be detected; inputting the spectral feature map into a U-Net spectral line segmentation network after training and convergence; using the U-Net spectral line segmentation network to output a signal spectral line probability vector of the same length as the spectral feature map, where each element in the signal spectral line probability vector represents the probability that the corresponding spectral line is a signal spectral line; and then using the signal spectral line probability vector for post-processing to output the signal detection result. This invention applies deep learning to signal detection, utilizing the feature representation capabilities of deep neural networks to improve signal detection performance, providing a novel approach for electronic warfare signal processing.

Description

Technical Field

[0001] This invention relates to the intersection of signal processing and artificial intelligence, and more specifically, to a signal detection method, device, and medium based on U-Net and spectrum. Background Technology

[0002] Signal detection, as the name suggests, is the process of extracting useful signals from the background (usually noise). Signal detection can be considered the foundation of all signal processing and is of paramount importance. James Tsui, in his classic work "Wideband Digital Receivers" [M]. Beijing: Electronic Industry Press (translated by Yang Xiaoniu, Lu Anan, and Jin Biao), introduces the most common and classic signal detection methods for digital reconnaissance receivers, mainly including time-domain and frequency-domain detection.

[0003] The basic principle of time-domain detection is to detect a signal by utilizing the difference between signal energy and noise energy. A common approach is to use a virtual window that slides across the time domain to divide the signal into detection frames, measure the amplitude of each frame, and then compare the amplitude to a fixed threshold. If the amplitude exceeds the threshold, the frame is considered to contain a signal. While time-domain detection is easy to understand, it is rarely used in modern digital reconnaissance systems. This is because time-domain detection requires a high signal-to-noise ratio (SNR). If the SNR is not ideal, it becomes difficult to determine the detection threshold; setting it too low results in numerous false alarms, while setting it too high leads to numerous missed detections.

[0004] The difference between frequency domain detection and time domain detection is that frequency domain detection does not directly use the signal amplitude within the detection frame. Instead, it first calculates the signal's spectrum and then compares it with a specific threshold to determine whether a signal exists. Compared to time domain detection, frequency domain detection is more complex, but due to its superior detection sensitivity, it is widely used in modern digital reconnaissance systems. Besides its advantage in detection sensitivity, for signal frames containing multiple frequency points, with appropriate processing, frequency domain detection can potentially output information for several frequency points simultaneously. In contrast, time domain detection typically only determines whether a signal exists within the detection frame.

[0005] With technological advancements, numerous detection methods have been developed to address various application scenarios and detection challenges. For example, correlation detection can be used to improve detection sensitivity, and dynamic thresholds can be employed to adapt to scenarios with constantly changing noise floor conditions. In recent years, with the development of deep learning technology, deep learning techniques have begun to be applied in signal detection. For instance, Y. Junfei, L. Jingwen, S. Bing, and J. Yuming, in their paper "Barrage Jamming Detection and Classification Based on Convolutional Neural Network for Synthetic Aperture Radar[J]. 38th Annual IEEE International Geoscience and Remote Sensing Symposium, IGARSS, 2018-July: 4583-4586," and Yao Peng, in his paper "Research on Deep Learning Signal Detection Method Based on Spectral Features[D]. Wuhan: Huazhong University of Science and Technology, 2019," both proposed using convolutional neural networks for signal detection. However, the signal detection performance of existing solutions needs further improvement. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a signal detection method, device and medium based on U-Net and spectrum, apply deep learning to signal detection, improve signal detection performance by utilizing the powerful feature representation capabilities of deep neural networks, and provide a new approach for electronic warfare signal processing.

[0007] The objective of this invention is achieved through the following solution:

[0008] A signal detection method based on U-Net and spectrum includes the following steps:

[0009] Calculate the frame of the signal to be detected Spectral feature map

[0010] The spectral feature map The input is passed to a U-Net spectral line segmentation network after training and convergence, and the output of the U-Net spectral line segmentation network is used to... Vectors of equal length Name it the signal spectral line probability vector; The elements in represent The probability that the corresponding spectral line is a signal spectral line;

[0011] Reuse Post-processing is performed to output the signal detection results.

[0012] Furthermore, the calculation of the signal frame to be detected Spectral feature map Including sub-steps: for After performing the Fast Fourier Transform, if If it is a real signal frame, then taking the first half and normalizing the modulus will give the result. if If it is a complex signal frame, then taking the complete spectrum and normalizing the modulus will yield the result.

[0013] Furthermore, the aforementioned It is the normalized signal spectrum, with a length of specLen, which contains specLen normalized spectral lines. The spectral lines that correspond to the actual signal frequencies are signal spectral lines, and the spectral lines that do not correspond to the actual signal frequencies are non-signal spectral lines.

[0014] For real signals, the spectrum exhibits a symmetrical structure. `specLen` represents half of the complete spectrum, and its frequency resolution `reso = fs / (specLen × 2)`. If the signal frequency is known, denoted as `fre`, the index value `idx` of the corresponding spectral line can be calculated by dividing the frequency by the frequency resolution and rounding off.

[0015]

[0016] in This indicates a round-down operation;

[0017] The signal spectral lines are those that most closely approximate the frequencies of the real signal, by... The spectral lines in the image are divided into two non-overlapping subsets, which are then used to segment the image using the U-Net spectral line segmentation network. The spectral lines within.

[0018] Furthermore, the aforementioned reuse Post-processing is performed to output the signal detection results. Including sub-steps:

[0019] It is a binary vector with values ​​of 0 or 1, where a value of 1 indicates that the corresponding spectral line is a signal spectral line, and a value of 0 indicates that it is not.

[0020]

[0021] Among them, sdr i =1 indicates that the U-Net signal detection model considers the corresponding spectral line p to be 1 i A signal was detected nearby, SDR i= 0 indicates that no signal is detected near the corresponding spectral line;

[0022] The threshold method is adopted, that is, a detection threshold thr is set. As long as the sslp in i is greater than the threshold, that is, the probability that the corresponding spectral line is a signal spectral line is greater than the threshold, it is considered that the signal is detected; that is

[0023]

[0024] The threshold thr is a probability value, and the value of thr ranges from 0 to 1.

[0025] Further, before inputting the spectral feature map into the U-Net spectral line segmentation network after training convergence, it includes the steps of: constructing a data set composed of data pairs, and using the supervised learning method to pre-train the U-Net spectral line segmentation network offline, where the label is the label corresponding to and the label contains the information of the position of the true signal frequency in the signal frame to be detected; after the U-Net spectral line segmentation network converges in training, it is used for the actual signal detection task.

[0026] Further, the U-Net spectral line segmentation network includes an input layer, a convolutional layer and an output layer; the input layer includes a one-dimensional vector with a shape of length specLen or a two-dimensional array of (specLen, 1); the convolutional layer includes using one-dimensional convolution operations; the output layer sets the number of output layer channels = 1, and the activation function selects the activation function applicable to the binary classification problem; or the output layer sets the number of output layer channels = 2, and uses softmax as the activation function.

[0027] Further, the construction of the data set composed of data pairs includes sub-steps:

[0028] is the label corresponding to and is a binary vector in which the element values are 0 or 1, and the length is equal to the length specLen of ; denoted as:

[0029]

[0030] Suppose the signal frame to be detected is composed of FN signals superimposed, and the center frequencies are denoted as f0 to f FN-1 then the element corresponding to f j in takes the value of 1, and other elements take the value of 0, 0 ≤ j < FN, that is:

[0031]

[0032] in, This indicates a floor operation, where reso is the frequency resolution of the spectral feature map.

[0033] Furthermore, the aforementioned reuse Post-processing is performed to output the signal detection results. Including sub-steps: Based on the threshold method, it includes the following steps:

[0034] For a real-sampling digital receiver with a sampling frequency of fs, the intermediate frequency bandwidth range [IF] is pre-defined. min ,IF max SDRs other than ] i Set all to 0; and / or, set a frequency spacing fd, and once a signal spectral line is detected, set other SSLPs with a spacing smaller than fd to zero;

[0035] or,

[0036] Based on the multi-signal detection capability, a maximum number of frequency points (MFN) that may be contained within a signal frame is set, and a maximum of MFN signal spectral lines are output. If the frequency exceeds this range, even if the SSLP value is very high, it will not be considered a signal spectral line.

[0037] A signal detection device based on U-Net and spectrum, comprising:

[0038] The preprocessing module is used to calculate the signal frame to be detected. Spectral feature map

[0039] The deep learning module will process the spectral feature map. The input is passed to a U-Net spectral line segmentation network after training and convergence, and the output of the U-Net spectral line segmentation network is used to... Vectors of equal length Name it the signal spectral line probability vector; The elements in represent The probability that the corresponding spectral line is a signal spectral line;

[0040] Post-processing module, reuse Post-processing is performed to output the signal detection results.

[0041] A readable storage medium storing a computer program, the computer program being loaded by a processor and executed as described in any of the preceding methods.

[0042] The beneficial effects of this invention include:

[0043] For conventional single-tone signals, the technical solution of this invention is expected to achieve results similar to those of classic frequency domain detection algorithms using the U-Net signal detection model.

[0044] The technical solutions of this invention address common problems in classic signal detection algorithms, such as signal saturation and large signals suppressing small signals. Similar phenomena also exist in the U-Net signal detection model.

[0045] If the signal frame to be detected contains a phase transition, the signal spectrum output by the U-Net signal detection model in this embodiment of the invention may have a better match with the true frequency of the signal compared to the classic frequency domain detection algorithm.

[0046] In some cases, the U-Net signal detection model can output signal frequency points that are difficult for conventional frequency domain detection algorithms to distinguish, as shown in the attached figure. Figure 4 As shown in b.

[0047] The U-Net signal detection model proposed in the technical solution of this invention realizes signal detection through the powerful segmentation capability of U-Net. Its beneficial effects will be further explained in conjunction with implementation examples in specific embodiments. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 This is a signal detection model based on U-Net and signal spectrum according to an embodiment of the present invention;

[0050] Figure 2 This is a conceptual illustration related to the U-Net signal detection model in an embodiment of the present invention;

[0051] Figure 3 The training curves of the U-Net spectral line segmentation network in this embodiment of the invention are shown.

[0052] Figure 4 The signal detection effect of the signal detection model in this embodiment of the invention is shown. Detailed Implementation

[0053] All features disclosed in all embodiments of this specification, or steps in all methods or processes implied in the disclosure, may be combined and / or extended or replaced in any way, except for mutually exclusive features and / or steps.

[0054] The technical solution of this invention involves applying the U-Net network to achieve signal detection. Deep learning has achieved unprecedented success in many fields such as image processing, speech processing, NLP, and intelligent game development. The purpose of this invention is to apply deep learning technology to signal detection, utilizing the powerful feature representation capabilities of deep neural networks to improve signal detection performance, and providing a new approach to electronic warfare signal processing.

[0055] In a further inventive concept, the input of the signal detection model based on U-Net and signal spectrum (hereinafter, the U-Net signal detection model) is the signal frame to be detected, and the output is the signal detection result, which contains information about the frequency points contained in the signal frame to be detected.

[0056] like Figure 1 As shown, the input to the U-Net signal detection model is the signal frame to be detected. (shaped like) The arrow character indicates that this is a vector (the same applies below), and it consists of three modules: data preprocessing, U-Net spectral segmentation network, and post-processing. The function of the data preprocessing module is to calculate the spectral feature map of the signal. The method is to Perform a Fast Fourier Transform, then take the first half (for real signal frames). For complex signal frames, the complete spectrum is taken, and modulus normalization is performed to obtain the result. Next, the U-Net spectral segmentation network is used for segmentation. Output a sum Vectors of equal length (Signal Spectral Line Probability), named the signal spectral line probability vector because its elements represent... The corresponding spectral line is the probability of the signal spectral line; the third step utilizes Post-processing is performed to output the final signal detection result. (Signal Detection Result) It is a binary vector with a value of 0 or 1. An element with a value of 1 indicates that the U-Net signal detection model considers the corresponding spectral line to be a signal spectral line, while an element with a value of 0 indicates that it is not.

[0057] The present invention introduces a concept called "signal spectral line" in its technical solution, which needs to be explained. This is the normalized signal spectrum, with a length of specLen. It contains specLen normalized spectral lines. Some of these lines correspond to the actual signal frequencies, while others do not. The former are called signal spectral lines, and the latter are called non-signal spectral lines. For real signals, the spectrum exhibits a symmetrical structure, so specLen can be half of the complete spectrum. Its frequency resolution is reso = fs / (specLen × 2). If the signal frequency is known, denoted as fre, the index value idx of the corresponding signal spectral line can be calculated by dividing the frequency by the frequency resolution and rounding. That is:

[0058]

[0059] here This indicates the floor function.

[0060] It must be emphasized that the so-called signal spectral lines, in simpler terms, are "spectral lines that are closest to the actual signal frequency." This is actually a classification method, used to... The spectral lines are divided into two non-overlapping subsets. This definition is for the convenience of using the U-Net spectral segmentation network for segmentation. From a signal processing perspective, the purpose of spectral lines cannot be understood as signal energy being distributed only at the frequency points corresponding to the spectral lines.

[0061] To obtain a high-performance U-Net spectral line segmentation network, it is necessary to construct a network consisting of... The dataset consists of data pairs, and a U-Net spectral line segmentation network is pre-trained offline using a supervised learning method, where the labels... Is and The corresponding labels contain information about the location of the actual signal frequency in the frame to be detected. Only after the U-Net spectral segmentation network has converged during training can it be used for actual signal detection tasks.

[0062] Figure 2This illustrates the related concepts. In this example, the signal frame to be detected is the result of superimposing two signals with frequencies of 100MHz and 200MHz, respectively. The former has a power 20dB greater than the latter. The signal frame length is 128 sampling points, and the sampling rate of the sampling system is 1000MHz. After preprocessing, the signal frame to be detected yields a spectral feature map. The frequency resolution of the spectral feature map, reso = 1000 / 128 = 7.8125MHz, easily calculated. The 13th and 26th spectral lines correspond to the signal spectral lines of 100MHz and 200MHz, respectively. Therefore, in the expected detection output, the 13th and 26th elements should have a value of 1, and all other elements should have a value of 0, which is precisely the value of the label. The goal of training the U-Net spectral segmentation network is to make the network output as close as possible to the above result, so that the same output result as the label vector can be obtained through post-processing.

[0063] signal frame Spectral feature map Signal spectral line probability vector The related concepts will be discussed in more detail in this section.

[0064] a) Data preprocessing

[0065] Let the sampling frequency of the digital reconnaissance system be fs, and let the frame of the signal to be detected be denoted as fs. It is a sequence of real sample points of a digital signal of fixed length. (Note: The original text contains some formatting errors and inconsistencies. A more accurate translation would require the full context.) The length is frameLen, and for simplicity, frameLen is always limited to a power of 2, so:

[0066]

[0067] use Calculate the spectral feature map The method is as follows:

[0068]

[0069] in,

[0070] 1) `fillzero` refers to the operation of padding the end of the signal frame to be detected with zeros, ensuring that the length of the padded signal remains a power of 2. Zero padding is a common method to improve frequency resolution in spectral analysis. Zero padding is optional for calculating spectral feature maps; whether to pad with zeros, and how many zeros to use, depends on the frequency resolution requirements of the signal spectrum for signal detection.

[0071] 2) FFT stands for Fast Fourier Transform;

[0072] 3) For real signal input, the spectrum data output by FFT is symmetrical. Half means taking the first half of the data. Let specLen be the length of the spectrum data after the half operation.

[0073] 4) abs represents the modulo operation;

[0074] 5) 20log 10 This indicates taking the logarithm by 20.

[0075] 6) `normalize` represents the normalization operation. The normalization interval can be defined according to the specific reconnaissance and receiving system. For example, if the signal detection model is designed for a signal with a saturation power of P... s The noise floor power is P n For a digital receiver, the normalization interval can be taken as [P]. n ,P s ],

[0076] Or a certain and [P] n ,P s The range that is close to the target range.

[0077] After preprocessing, record the spectral feature map. The length is specLen, and it has

[0078]

[0079] Where, p i yes The i-th spectral line is physically represented by the normalized value of the relative power of the signal frame to be detected at the i-th frequency component.

[0080] b) U-Net spectral segmentation network

[0081] The U-Net spectral segmentation network is the core processing module of a signal detection model based on U-Net and signal spectra. Its input is the spectral feature map. The output is a sum Equal-length signal spectral line probability vector

[0082]

[0083] Among them, sslp i The meaning refers to the output of the U-Net spectral segmentation network. The i-th spectral line p in i It is the probability of the signal spectral lines.

[0084] Key design considerations for the U-Net spectral line segmentation network include:

[0085] 1) The input layer is a 1D vector with a length of specLen. At the specific implementation level, the input layer may also be a two-dimensional array in the form of (specLen, 1) (this is the setting made by some deep learning development packages at the software implementation level);

[0086] 2) Since is a 1D vector, 1D convolution operations are used in all convolutional layers of the U-Net spectral line segmentation network;

[0087] 3) For each spectral line in , there are two possible output results: signal spectral line or non-signal spectral line. This can be regarded as a binary classification problem. The number of output layer channels can be set to 1, and activation functions such as sigmoid or relu, which are suitable for binary classification problems, can be used for the output layer. The problem can also be regarded as a multi-class classification problem with 2 classes. Correspondingly, the number of output layer channels can be set to 2, and softmax can be used as the activation function;

[0088] Table 1 shows the U-Net network used in the implementation example. According to specific detection requirements, the configuration of the U-Net network can be locally adjusted. For example, the upsampling and downsampling ratios and the number of layers of symmetric convolutional layers can be different from those in Table 1. However, regardless of the specific structure of the network, as long as the U-Net network is used to segment and output it is within the protection scope of this invention.

[0089] c) Dataset construction

[0090] To train the U-Net network, a training and validation dataset needs to be constructed. The dataset consists of a large number of data pairs in the form of . The method of using the signal frame to be detected to generate has been described in the above technical solution. is the label corresponding to , and it is a binary vector whose elements take values of 0 or 1, and its length is equal to the length specLen of <\ . Denote

[0091]

[0092] Suppose the signal frame to be detected is composed of FN signals superimposed, and denote the center frequencies of each as f0 ~ f FN-1 , then [[ID=X]] the element in j corresponding to f j (0 ≤ j < FN) takes the value of 1, and other elements take the value of 0. That is

[0093]

[0094] In the above formula, This indicates a floor operation, where reso is the frequency resolution of the spectral feature map.

[0095] Once the dataset is constructed, the U-Net spectral segmentation network can be trained using classic supervised learning methods. The goal of training the U-Net spectral segmentation network is to minimize the difference between the U-Net network's output and the label. Various common machine learning platforms in the industry (such as Keras, Caffe, and PyTorch) all provide support for supervised learning algorithms. This part is common to deep learning research, and related details will not be elaborated upon in this invention.

[0096] d) Post-processing

[0097] The goal of post-processing is to utilize This yields the final signal detection result. The signal detection result is visually identical to the label (ideally, they should be equal), also a binary vector where each element takes the value 0 or 1.

[0098]

[0099] Among them, sdr i =1 indicates that the U-Net signal detection model considers the corresponding spectral line p to be 1 i A signal was detected nearby, SDR i =0 indicates that no signal was detected near the corresponding spectral line.

[0100] The simplest post-processing can use a threshold method, that is, setting a detection threshold thr. As long as... sslp in i If the probability that the corresponding spectral line is a signal spectral line is greater than a threshold, then the signal is considered detected.

[0101]

[0102] Note that the threshold `thr` here is a probability value, which is significantly different from the detection threshold in classic detection algorithms (usually a value related to signal energy, such as amplitude or power). Clearly, `thr` takes a value between 0 and 1, and the specific value needs to be determined by a trade-off between the system's requirements for false alarm and false miss probabilities. Increasing `thr` decreases the false alarm rate and increases the false miss rate, and vice versa. If there is no other prior information, setting `thr = 0.5` is a reasonable choice.

[0103] In a further implementation, based on the threshold method, for a real-sampling digital receiver with a sampling frequency of fs, the receiver typically has an intermediate frequency bandwidth range. Signals outside the bandwidth range will not be processed; therefore, [IF] can be pre-programmed. min ,IF max SDRs other than ] i Set all values ​​to 0. In certain special cases, the processing bandwidth of the receiver system may not be a complete interval, but rather the union of several intervals. This can be used to limit post-processing.

[0104] In a further implementation, based on the above method, a frequency distance fd can be set. Once a signal spectral line is detected, other SSLPs with a distance less than fd from the signal spectral line are set to zero. The purpose of this is to ensure that there is a certain distance between the two or more signal spectral lines output by the signal detection model. This can prevent multiple output signal spectral lines from being located within the same main lobe (or side lobe).

[0105] In a further implementation, based on the above method, a maximum number of frequency points (MFN) that may be contained within a signal frame can be set according to the multi-signal detection capability, and a maximum of MFN signal spectral lines will be output. Lines exceeding this range, even if the SSLP value is very high, will not be considered signal spectral lines.

[0106] Other post-processing strategies may also be used in the technical solutions of this invention, and any method that is beneficial to improving detection performance is worth trying. However, adjustments made only in the post-processing stage do not necessarily fall outside the scope of protection of this invention.

[0107] The embodiments of the present invention are not limited to this specific example.

[0108] In the technical solution of this invention embodiment, the training and signal detection effects of the U-Net signal detection model are demonstrated through simulation. The simulation software is developed using the Python+keras framework.

[0109] In the implementation example, the U-Net spectral line segmentation network is first designed according to the settings in step "b) U-Net spectral line segmentation network" of the above technical solution. The specific structure is shown in Table 1.

[0110] Table 1. U-Net Spectral Line Segmentation Network Structure

[0111]

[0112]

[0113] After constructing the U-Net spectral segmentation network, the next step is to construct the dataset according to the method in step "c) Dataset Construction" of the above technical solution. The simulation parameters are set as follows when generating the dataset for simulation:

[0114] 1) Receiver sampling frequency: fs = 1000MHz;

[0115] 2) Intermediate frequency bandwidth range: [IF min ,IF max [100, 400] (MHz). The frequencies of all signal frames to be detected are randomly selected within this range.

[0116] 3) Signal saturation amplitude: satuA = 1 (dimensionless). When the signal amplitude is greater than 1, amplitude limiting is applied. The relative power of a single-tone signal is -3dB when its amplitude is exactly 1, which is used to determine the relative power of other signals. Therefore, this is equivalent to limiting the signal amplitude to 1 when the relative power of the signal is greater than -3dB.

[0117] 4) Relative signal power range: [minP, maxP] = [-44, 0] dB. The power of all signal frames to be detected is randomly selected within this range;

[0118] 5) Relative noise power: Pn = -38dB;

[0119] 6) The frame length of the signal to be detected is frameLen = 128, that is, a spectral feature map is generated each time using a 128ns signal frame;

[0120] 7) In the implementation example, when generating the spectral feature map from the signal frame, zeros of the same length were padded to the end of the signal. Therefore, the length of the spectral feature map, specLen, is 128, and the corresponding frequency resolution, reso, is fs /

[0121] (specLen×2)=3.90625MHz;

[0122] 8) There are five possible types of signal frames to be detected: pure noise, normal (no modulation), LFM, NLFM (parabolic modulation), and phase-jump signal. The modulation bandwidth of the frequency-modulated signal within a 128ns signal frame is randomly selected between 5 and 30kHz, and after proportional linear amplification, it is approximately equivalent to modulation of 0.4–2.3MHz every 10μs. The phase jump value is always set to π, and the phase jump occurs at most once per signal frame.

[0123] 9) The maximum number of frequency points (MFN) in the signal frame to be detected is 2, meaning that at most 2 frequency points are considered when generating the dataset. A signal frame containing only a single frequency point is called a single-frequency signal frame, and a signal frame containing 2 frequency points is called a dual-frequency signal frame. Accordingly, for one signal frame to be detected, the signal detection model can output at most 2 frequency points;

[0124] 10) The frequency spacing fd = 20MHz, which means that when generating a dual-frequency signal frame, the spacing between the two frequency points should be greater than 20MHz. Correspondingly, if the signal detection model outputs two frequency points, the spacing between these two frequency points should also be greater than 20MHz.

[0125] Strictly speaking, steps 9) and 10) also include parameters used in post-processing, but there's no need to distinguish them so clearly here. Once the dataset is constructed, it can be used to train the U-Net segmentation network. The main training parameter settings are as follows:

[0126] 1) A maximum of 300 epochs can be trained;

[0127] 2) Train 20 batches per epoch;

[0128] 3) When constructing the dataset, use Keras's built-in methods to generate and update the training dataset for each batch to increase the randomness of the data;

[0129] 4) Each batch of training data contains 100 samples. The data pairs are divided into single-frequency data pairs and dual-frequency data pairs, each accounting for half. The proportion of signal frames of various modulation types is shown in Figure 0.

[0130] 5) Use the Adam optimization algorithm;

[0131] 6) The loss function is the binary classification cross-entropy loss function;

[0132] 7) The initial learning rate is set to 0.001;

[0133] 8) Use accuracy as the network evaluation metric. If the training set accuracy does not improve for 16 consecutive epochs, the learning rate is halved. If there is no improvement for 40 consecutive epochs, training is terminated.

[0134] Table 2. Proportion of various signal frame types

[0135]

[0136] The final training results are shown in Figure 3 As shown, where Figure 3 In the middle (a), the accuracy curve is shown. Figure 3 Figure (b) shows the loss curve. Training was conducted for 193 epochs, with the peak accuracy of 99.79% achieved on the training set in the 153rd epoch. The corresponding U-Net network parameters were saved as the final U-Net spectral line segmentation network. After training, the test dataset was regenerated 100 times consecutively to evaluate the U-Net spectral line segmentation network, with an average segmentation accuracy of 99.76%.

[0137] The post-processing strategy in the implementation example is set as follows:

[0138] 1) Intermediate frequency bandwidth range [IF] min ,IF max ] = [100, 400] MHz. Spectral lines outside this range are not considered signal spectral lines even if their probability is greater than the detection threshold.

[0139] 2) Frequency distance fd = 20MHz, which means that the search will find... After the peak, the four spectral lines on either side are not considered signal spectral lines even if the signal spectral line probability sslp value is greater than 0.5.

[0140] 3) The maximum number of frequency points MFN = 2, the detection threshold for the first frequency point thr1 = 0.5, and the detection threshold for the second frequency point thr1 = 0.1. That is, for the signal frame to be detected, a maximum of 2 frequency points are output. If the peak probability of the signal spectrum is greater than 0.5, the first frequency point is considered to be detected, and if the second peak probability is greater than 0.1, the second frequency point is considered to be detected.

[0141] Following the post-processing strategy described above, the false alarm rate (predicted value positive, labeled negative) is calculated, specifically reflected in the SDR output of the signal detection model. i =1, corresponding to label l i =0) and false negative rate (predicted value is negative while labeled positive, specifically reflected in the SDR output of the signal detection model). i =0, and the corresponding label l i =1) They were 14.8% and 11% respectively.

[0142] The detection performance of the signal detection model in this implementation example is shown in the figure. Figure 4 As shown in the figure, solid dots mark the corresponding positions of spectral lines in the spectrum; the intensity of the dot color indicates the probability that the corresponding spectral line is a signal spectral line; the vertical line below the dot marks the elements with a value of 1 in the label, which actually marks the location of the signal frequency when the signal frame is generated; ★ and the vertical line above the dot mark the final signal detection result output.

[0143] Figure 4 Images (a) and (b) show the correct detection output. The ★ and vertical lines indicate the same spectral lines, meaning the U-Net signal detection model correctly outputs the positions of the true signal spectral lines. Classical frequency domain detection methods typically search for local maxima in the spectrum. For a typical single-tone signal, the local maxima in the spectrum will likely correspond to the true signal frequency. However, if the signal frame to be detected contains phase transitions, the dip between two local maxima in the spectrum may be closer to the true signal frequency, such as... Figure 4As shown in (a), the U-Net signal detection model in the technical solution of this embodiment of the invention can better distinguish this situation, and the matching of the output signal spectrum and signal frequency is better. Figure 4 In (b), the signal detection model outputs two signal spectral lines. The signal spectral line inside the black circle on the left is very difficult to detect using conventional frequency domain detection algorithms.

[0144] Figure 4 (c) and (f) highlight the two scenarios of error detection output. When faced with similar situations, it is also very common for the classic frequency domain detection method to output similar error results. Figure 4 (c) shows the detection output of a single-frequency signal frame. The relative power of this signal is -0.35dB. According to the simulation settings, the signal will be saturated and limited when the relative power of a single signal is greater than -3dB. This results in strong harmonics in the signal, and the signal detection model treats the harmonics at this frequency as the signal frequency. Figure 4 In (d), the frequency 1 (281.4MHz) and the frequency 2 (316.94MHz) differ by about 36MHz, which is not far apart. The power difference is 16dB, and a phenomenon occurs where the sidelobe of the strong signal suppresses the weak signal, causing the signal detection model to be unable to detect frequency 2.

[0145] Next Figure 4 Examples (c) and (f) also illustrate common false alarms or missed alarms in implementation examples: the index of the signal spectral line output by the signal detection model differs from the index of the signal spectral line in the label by 1. This situation is less an error and more an error in the detection model's output; similar errors are very common in classic signal detection algorithms. Previously, all such cases were included in the false alarm or missed alarm statistics, which is arguably too strict. If the statistical criteria are adjusted to exclude these cases where the model output differs from the label by 1 from the false alarm and missed alarm statistics, the corresponding false alarm rate and missed alarm rate decrease to 5.9% and 5.5%, respectively. Note that this reduction in the false alarm rate and missed alarm rate is not due to any improvement in model performance, but purely due to the adjustment of the statistical criteria.

[0146] It should be noted that, within the scope of protection defined in the claims of this invention, the following embodiments can be combined and / or extended or replaced in any logical manner from the above specific embodiments, such as the disclosed technical principles, disclosed technical features or implicitly disclosed technical features.

[0147] Example 1

[0148] A signal detection method based on U-Net and spectrum includes the following steps:

[0149] Calculate the frame of the signal to be detected Spectral feature map

[0150] The spectral feature map The input is passed to a U-Net spectral line segmentation network after training and convergence, and the output of the U-Net spectral line segmentation network is used to... Vectors of equal length Name it the signal spectral line probability vector; The elements in represent The probability that the corresponding spectral line is a signal spectral line;

[0151] Reuse Post-processing is performed to output the signal detection results.

[0152] Example 2

[0153] Based on Example 1, the calculation of the signal frame to be detected Spectral feature map Including sub-steps: for After performing the Fast Fourier Transform, if If it is a real signal frame, then taking the first half and normalizing the modulus will give the result. if If it is a complex signal frame, then taking the complete spectrum and normalizing the modulus will yield the result.

[0154] Example 3

[0155] Based on Example 2, the It is the normalized signal spectrum, with a length of specLen, which contains specLen normalized spectral lines. The spectral lines that correspond to the actual signal frequencies are signal spectral lines, and the spectral lines that do not correspond to the actual signal frequencies are non-signal spectral lines.

[0156] For real signals, the spectrum exhibits a symmetrical structure. `specLen` represents half of the complete spectrum, and its frequency resolution `reso = fs / (specLen × 2)`. If the signal frequency is known, denoted as `fre`, the index value `idx` of the corresponding spectral line can be calculated by dividing the frequency by the frequency resolution and rounding off.

[0157]

[0158] in This indicates a round-down operation;

[0159] The signal spectral lines are those that most closely approximate the frequencies of the real signal, by... The spectral lines in the image are divided into two non-overlapping subsets, which are then used to segment the image using the U-Net spectral line segmentation network. The spectral lines within.

[0160] Example 4

[0161] Based on Example 1, the reuse Post-processing is performed to output the signal detection results. Including sub-steps:

[0162] It is a binary vector with values ​​of 0 or 1, where a value of 1 indicates that the corresponding spectral line is a signal spectral line, and a value of 0 indicates that it is not.

[0163]

[0164] Among them, sdr i =1 indicates that the U-Net signal detection model considers the corresponding spectral line p to be 1 i A signal was detected nearby, SDR i =0 indicates that no signal was detected near the corresponding spectral line;

[0165] The threshold method involves setting a detection threshold thr, as long as... sslp in i If the probability of the corresponding spectral line being a signal spectral line is greater than the threshold, then the signal is considered detected.

[0166]

[0167] The threshold thr is a probability value, which takes a value between 0 and 1.

[0168] Example 5

[0169] Based on Example 1, the spectral feature map is then... Before inputting into the U-Net spectral line segmentation network after training and convergence, the steps include: constructing a network composed of... The dataset consists of data pairs, and a U-Net spectral line segmentation network is pre-trained offline using a supervised learning method, where the labels... Is and The corresponding labels contain information about the location of the actual signal frequency in the frame to be detected; the U-Net spectral segmentation network is only used for actual signal detection tasks after it has been trained and converged.

[0170] Example 6

[0171] Based on Example 1, the U-Net spectral segmentation network includes an input layer, a convolutional layer, and an output layer; the input layer includes a 1-dimensional vector of shape and length specLen. A two-dimensional array of (specLen, 1); the convolutional layer includes 1D convolutional operations; the output layer sets the number of output layer channels = 1, and the activation function selects an activation function suitable for binary classification problems; or the output layer sets the number of output layer channels = 2, and uses softmax as the activation function.

[0172] Example 7

[0173] Based on Example 5, the structure consists of A dataset composed of data pairs, including sub-steps:

[0174] is and The corresponding label is a binary vector with elements taking values of 0 or 1, and its length is equal to the length specLen of ; denoted as:

[0175]

[0176] Suppose the signal frame to be detected is composed of FN signals superimposed, and the center frequencies are denoted as f0 to f FN-1 , then The element in j corresponding to f takes the value of 1, and other elements take the value of 0, 0 ≤ j < FN, that is:

[0177]

[0178] Among them, represents the floor operation, and reso is the frequency resolution of the spectral feature map.

[0179] Example 8

[0180] Based on Example 4, the reuse of is used for post-processing to output the signal detection result including sub-steps: Based on the threshold method, the following steps are included:

[0181] ]>For a real-sampling digital receiver with a sampling frequency of fs, all sdr min outside the intermediate frequency bandwidth range [IF max are set to 0; and / or, a frequency interval fd is set. Once a signal spectral line is detected, other sslp with a distance less than fd from the signal spectral line are set to zero; i Or,

[0182] Or,

[0183] According to the multi-signal detection ability, a maximum number of frequency points MFN that may be included in a signal frame is set, and at most MFN signal spectral lines are output. For those exceeding this range, even if the sslp value is very high, they are not considered signal spectral lines.

[0184] Example 9

[0185] A signal detection device based on U-Net and spectrum, comprising:

[0186] The preprocessing module is used to calculate the signal frame to be detected. Spectral feature map

[0187] The deep learning module will process the spectral feature map. The input is passed to a U-Net spectral line segmentation network after training and convergence, and the output of the U-Net spectral line segmentation network is used to... Vectors of equal length Name it the signal spectral line probability vector; The elements in represent The probability that the corresponding spectral line is a signal spectral line;

[0188] Post-processing module, reuse Post-processing is performed to output the signal detection results.

[0189] Example 10

[0190] A readable storage medium storing a computer program, the computer program being loaded by a processor and executed as described in any one of Examples 1 to 8.

[0191] The units described in the embodiments of the present invention can be implemented in software or hardware, and the described units can also be located in a processor. The names of these units do not necessarily limit the specific unit itself.

[0192] According to one aspect of the present invention, a computer program product or computer program is provided, the computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and executes the computer instructions, causing the computer device to perform the methods provided in the various optional implementations described above.

[0193] In another aspect, embodiments of the present invention also provide a computer-readable medium, which may be included in the electronic device described in the above embodiments; or it may exist independently and not assembled into the electronic device. The computer-readable medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to perform the methods described in the above embodiments.

[0194] All parts not covered in this invention are the same as or can be implemented using existing technologies.

[0195] The above technical solution is only one embodiment of the present invention. For those skilled in the art, based on the application methods and principles disclosed in the present invention, it is easy to make various types of improvements or modifications, and not limited to the methods described in the above specific embodiments of the present invention. Therefore, the methods described above are only preferred and are not restrictive.

[0196] In addition to the examples above, other embodiments may be obtained by those skilled in the art based on the above disclosure or by making modifications using knowledge or technology in related fields. The features of each embodiment may be interchanged or replaced. Modifications and changes made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A signal detection method based on U-Net and spectrum, characterized in that, Includes the following steps: Calculate the frame of the signal to be detected Spectral feature map ; The spectral feature map The input is passed to a U-Net spectral line segmentation network after training and convergence, and the output of the U-Net spectral line segmentation network is used to... Vectors of equal length This is named the signal spectral line probability vector; The elements in represent The probability that the corresponding spectral line is a signal spectral line; Reuse Post-processing is performed to output the signal detection results. ; The calculation of the signal frame to be detected Spectral feature map This includes sub-steps: for After performing the Fast Fourier Transform, if If it is a real signal frame, then taking the first half and normalizing the modulus will give the result. ;if If it is a complex signal frame, then taking the complete spectrum and normalizing the modulus will yield the result. ; The It is the normalized signal spectrum, with a length of specLen, which contains specLen normalized spectral lines. The spectral lines that correspond to the actual signal frequencies are signal spectral lines, and the spectral lines that do not correspond to the actual signal frequencies are non-signal spectral lines. For real signals, the spectrum exhibits a symmetrical structure. `specLen` represents half of the complete spectrum, and its frequency resolution `reso = fs / (specLen × 2)`. If the signal frequency is known, denoted as `fre`, the corresponding spectral line index `idx` can be calculated by dividing the frequency by the frequency resolution and rounding. in This indicates a round-down operation; The signal spectral lines are those that most closely approximate the frequencies of the real signal, by... The spectral lines in the image are divided into two non-overlapping subsets, which are then used to segment the image using the U-Net spectral line segmentation network. The spectral lines within.

2. The signal detection method based on U-Net and spectrum according to claim 1, characterized in that, The reuse Post-processing is performed to output the signal detection results. This includes sub-steps: It is a binary vector with values ​​of 0 or 1, where a value of 1 indicates that the corresponding spectral line is a signal spectral line, and a value of 0 indicates that it is not. =[ sdr0, sdr1, sdr specLen-1 ] Among them, sdr i =1 indicates that the U-Net signal detection model considers the corresponding spectral line p to be... i A signal was detected nearby, SDR i =0 indicates that no signal was detected near the corresponding spectral line; The threshold method involves setting a detection threshold thr, as long as... sslp in i If the probability of the corresponding spectral line being a signal spectral line is greater than the threshold, then the signal is considered detected. The threshold thr is a probability value, which takes a value between 0 and 1.

3. The signal detection method based on U-Net and spectrum according to claim 1, characterized in that, In the spectral feature map Before inputting into the U-Net spectral segmentation network after training convergence, the steps include: constructing a network composed of ( , The dataset consists of data pairs, and a U-Net spectral line segmentation network is pre-trained offline using a supervised learning method, where the labels... Is and The corresponding labels contain information about the location of the actual signal frequency in the frame to be detected; the U-Net spectral segmentation network is only used for actual signal detection tasks after it has been trained and converged.

4. The signal detection method based on U-Net and spectrum according to claim 1, characterized in that, The U-Net spectral segmentation network includes an input layer, a convolutional layer, and an output layer; the input layer includes a 1-dimensional vector of shape and length specLen. Alternatively, it can be a two-dimensional array of (specLen, 1); the convolutional layer includes 1-dimensional convolution operations; the output layer is set to have 1 output channel and the activation function is selected to be suitable for binary classification problems; or the output layer is set to have 2 output channels and softmax is used as the activation function.

5. The signal detection method based on U-Net and spectrum according to claim 3, characterized in that, The construction is composed of ( , The dataset, consisting of data pairs, includes the following sub-steps: Is and The corresponding label is a binary vector in which each element takes the value 0 or 1, and its length and sum are given. The lengths of specLen are equal; Recorded as: =[ l0, l1,…, l specLen-1 ] T Suppose that the signal frame to be detected is composed of FN superimposed signals, and let the center frequencies of each signal be f0 to f1. FN-1 ,but neutralization f j The corresponding element has a value of 1, and all other elements have a value of 0. ,Right now: in, This indicates a floor operation, where reso is the frequency resolution of the spectral feature map.

6. The signal detection method based on U-Net and spectrum according to claim 2, characterized in that, The reuse Post-processing is performed to output the signal detection results. The process includes sub-steps: Based on the threshold method, it includes the following steps: For a real-sampling digital receiver with a sampling frequency of fs, the intermediate frequency bandwidth range [IF] is pre-defined. min ,IF max SDRs other than ] i Set all to 0; and / or, set a frequency spacing fd, and once a signal spectral line is detected, set other SSLPs with a spacing smaller than fd to zero; or, Based on the multi-signal detection capability, a maximum number of frequency points (MFN) that may be contained within a signal frame is set, and a maximum of MFN signal spectral lines are output. If the frequency exceeds this range, even if the SSLP value is very high, it will not be considered a signal spectral line.

7. A signal detection device based on U-Net and spectrum, characterized in that, include: The preprocessing module is used to calculate the signal frame to be detected. Spectral feature map ; The deep learning module will process the spectral feature map. The input is passed to a U-Net spectral line segmentation network after training and convergence, and the output of the U-Net spectral line segmentation network is used to... Vectors of equal length This is named the signal spectral line probability vector; The elements in represent The probability that the corresponding spectral line is a signal spectral line; Post-processing module, reuse Post-processing is performed to output the signal detection results. ; Furthermore, the device is used to perform the signal detection method for U-Net and spectrum as described in any one of claims 1 to 6.

8. A readable storage medium, characterized in that, A computer program is stored in a readable storage medium, the computer program being loaded by a processor and executing the method as described in any one of claims 1 to 6.

Images