Deep learning broadband spectrum sensing method based on dual power spectrum representation input and data enhancement

Abstract

The invention relates to a deep learning broadband spectrum sensing method based on dual power spectrum representation input and data enhancement, and belongs to the field of spectrum sensing. Comprising the following steps: extracting power spectrum representation of an obtained broadband time domain signal by using a periodogram method and a multi-tap spectrum estimation method to obtain a periodogram power spectrum and a multi-tap power spectrum; constructing a sample data set based on the periodogram power spectrum, the corresponding multi-tap power spectrum and the sample label; wherein the sample labels are the states of the sub-bands, and the states comprise occupied states and idle states; constructing a broadband spectrum sensing model based on the convolutional neural network, and obtaining a trained broadband spectrum sensing model by using the sample data set; obtaining a periodogram power spectrum and a multi-tap power spectrum of a signal to be detected, and obtaining confidence coefficients of a plurality of sub-bands by using the trained broadband spectrum sensing model; and performing frequency spectrum judgment based on the judgment threshold and the confidence coefficient of each sub-band to obtain a sensing result of the to-be-detected signal. The broadband spectrum sensing detection performance is improved, and the false alarm probability is effectively controlled.

Description

Technical Field

[0001] The present invention relates to the field of spectrum sensing, and in particular to a deep learning broadband spectrum sensing method based on dual power spectrum representation input and data enhancement. Background Art

[0002] With the commercialization of fifth-generation (5G) mobile networks and the rapid development of sixth-generation (6G) mobile networks, wireless communications will face a wider range of application scenarios and higher data transmission requirements in the future, and wireless spectrum resources will become increasingly scarce. Broadband spectrum sensing is a key technology for efficient spectrum management. By monitoring spectrum occupancy in real time and identifying unused spectrum holes, it can improve overall spectrum resource utilization.

[0003] With the increasing popularity of artificial intelligence (AI) technology, wideband spectrum sensing methods based on deep learning are attracting significant attention. However, existing deep learning wideband spectrum sensing methods based on multi-label classification typically use time-domain signal information or a single power spectrum as input and rely on a large number of labeled samples for training. While these methods can detect spectrum occupancy, obtaining large amounts of labeled, real-world spectrum data is difficult in practical spectrum sensing scenarios. Furthermore, the sensing model often needs to be retrained and deployed in different spectrum environments, resulting in low sensing detection rates. Relying on large amounts of labeled data for training is inefficient. Summary of the Invention

[0004] Based on the above analysis, the present invention aims to provide a deep learning-based wideband spectrum sensing method based on a small number of samples. This method solves the technical problems of low spectrum sensing detection rate and unstable false alarm probability in existing deep learning-based wideband spectrum sensing methods in scenarios with a small number of samples.

[0005] The present invention discloses a deep learning broadband spectrum sensing method based on dual power spectrum characterization input and data enhancement, comprising the following steps:

[0006] The power spectrum representation of the acquired broadband time-domain signal is extracted using the periodogram method and the multi-tap spectrum estimation method, respectively, to obtain the periodogram power spectrum and the multi-tap power spectrum;

[0007] Constructing a sample data set based on the periodogram power spectrum, the corresponding multi-tap power spectrum, and the corresponding sample labels; wherein the sample labels are states of each subband, and the states include occupied and idle;

[0008] Building a broadband spectrum sensing model based on a convolutional neural network, and obtaining a trained broadband spectrum sensing model using the sample data set;

[0009] The periodogram power spectrum and multi-tap power spectrum of the signal to be detected are obtained, and the confidence of multiple sub-bands is obtained using the trained wideband spectrum perception model; spectrum judgment is performed based on the decision threshold and the confidence of each sub-band to obtain the perception result of the signal to be detected.

[0010] Furthermore, obtaining the periodogram power spectrum includes:

[0011] Performing discrete Fourier transform on the broadband time domain signal to obtain a complex frequency domain signal;

[0012] The first periodogram power spectrum is obtained by shifting the center of the spectrum of the complex frequency domain signal, placing the frequency zero point at the center of the spectrum, and calculating the square of the modulus;

[0013] Performing data enhancement on the first periodogram power spectrum using an inter-subband spectrum rearrangement method and / or an intra-subband spectrum rearrangement method to obtain a data-enhanced second periodogram power spectrum;

[0014] The first and second periodogram power spectra are normalized to obtain the periodogram power spectrum.

[0015] Furthermore, obtaining the multi-tap power spectrum includes:

[0016] The broadband time domain signal is passed through a multi-tap filter and then subjected to discrete Fourier transform to obtain a plurality of windowed complex frequency domain signals;

[0017] By shifting the center of the spectrum of the complex frequency domain signal, aligning the frequency zero point to the center of the spectrum, and calculating the square of the modulus of the frequency domain signal to obtain the initial power spectrum; then, according to the weight coefficient of each filter, the power spectrum is weighted and combined to obtain the first multi-tap power spectrum;

[0018] Performing data enhancement on the first multi-tap power spectrum by using an inter-subband spectrum rearrangement method and / or an intra-subband spectrum rearrangement method to obtain a data-enhanced second multi-tap power spectrum;

[0019] The first and second multi-tap power spectra are normalized to obtain the multi-tap power spectrum.

[0020] Furthermore, if a primary user exists in a sub-band in the periodogram power spectrum and the multi-tap power spectrum, the sub-band is labeled as occupied, otherwise it is labeled as idle;

[0021] The labels of the sub-bands in the periodogram power spectrum and the multi-tap power spectrum are respectively combined into label vectors as sample labels.

[0022] Furthermore, the broadband spectrum sensing model includes, in sequence, a first and a second parallel input branch, a feature fusion layer, a first convolutional layer, a first average pooling layer, second and third convolutional layers, a second average pooling layer, a fourth convolutional layer, an adaptive average pooling layer, a linear layer, and an output layer;

[0023] The first input branch and the second input branch have the same structure, and both include three layers of branch convolution layers, namely the first, second and third branch convolution layers. They are used to extract the features of the periodogram power spectrum and the multi-tap power spectrum in the sample data respectively; and obtain the features F PG and F MTM ;

[0024] The feature fusion layer is used to transform F PG and F MTM Splice and get the fusion feature F cat ;

[0025] The first convolutional layer is used to transform the fusion feature F cat Performing a convolution operation to obtain a first signal feature;

[0026] The first average pooling layer is used to downsample the first signal feature to obtain a second signal feature;

[0027] The second and third convolutional layers are used to sequentially perform convolution operations on the second signal feature to obtain a third signal feature;

[0028] The second average pooling layer is used to downsample the third signal feature to obtain a fourth signal feature;

[0029] The fourth convolutional layer is used to perform a convolution operation on the fourth signal feature to obtain a fifth signal feature;

[0030] The adaptive average pooling layer downsamples the fifth signal feature to 256 dimensions to obtain a sixth signal feature;

[0031] The linear layer is a fully connected layer, used to convert the sixth signal feature into a 64-dimensional seventh signal feature;

[0032] The output layer includes a fully connected layer and a Sigmoid activation function, which is used to map the seventh signal feature into confidence levels corresponding to multiple sub-bands.

[0033] Furthermore, the first branch convolution layer is configured with 4 convolution kernels of size 7×1; the second branch convolution layer is configured with 4 convolution kernels of size 5×1; and the third branch convolution layer is configured with 8 convolution kernels of size 3×1;

[0034] The first, second and third branch convolutional layers respectively perform convolution operations, batch normalization operations and LeakRelu activation operations on the inputs of the first, second and third branch convolutional layers.

[0035] Furthermore, the training process of the broadband spectrum sensing model is as follows:

[0036] Using a training set in the sample data set to train the broadband spectrum sensing model;

[0037] During the training process, the Adam optimizer is used to iteratively update the model parameters and minimize the loss function;

[0038] If the accuracy does not improve within p training cycles, the learning rate in the current training cycle is reduced to 0.1 times that of the previous training cycle;

[0039] If the accuracy does not improve for q consecutive training cycles, the training is stopped and the network parameters with the minimum loss function on the validation set during the training process are saved as the trained wideband spectrum sensing model; where q>p.

[0040] Furthermore, the loss function is as follows:

[0041]

[0042] in, is the confidence of each sub-band output by the model; Y is the label sequence; N is the number of sub-bands into which the signal is divided; C is the total number of samples, The confidence of the i-th subband of the c-th sample output by the model; is the true label of the i-th subband of the c-th sample; BCE represents binary cross entropy; represents the average loss of all sub-bands for all samples.

[0043] Furthermore, the decision threshold is obtained by:

[0044] Extracting the corresponding periodogram power spectra and multi-tap power spectra of Z pure noise samples, inputting the trained wideband spectrum sensing model, and obtaining Z confidence scores corresponding to each subband;

[0045] Arrange the Z confidences of each sub-band in descending order to obtain a confidence sequence;

[0046] Based on the preset target false alarm probability P f , calculate the decision threshold position index P f *Z;

[0047] The first [P f *Z] value λ i, as the decision threshold for subband i, where [.] is a rounding operation.

[0048] Furthermore, spectrum decision is performed based on the decision threshold and the confidence level of each sub-band to obtain the perception result of the signal to be detected, including:

[0049] The periodogram power spectrum and multi-tap power spectrum of the signal to be detected are input into the trained wideband spectrum sensing model to obtain the sub-band confidence of the corresponding N sub-bands to be detected on the spectrum occupancy state.

[0050] like The sensing result of the subband i in the signal to be detected is determined to be in an occupied state; otherwise, the sensing result of the subband i in the signal to be detected is determined to be in an idle state;

[0051] The perception results of the N to-be-detected sub-bands corresponding to the to-be-detected signal are combined into a vector as the perception result of the to-be-detected signal.

[0052] Compared with the prior art, the present invention can achieve at least the following beneficial effects:

[0053] 1. The present invention uses only a small number of time-domain broadband signal samples, and utilizes the periodogram method and multi-tap spectrum estimation method to extract the multi-power spectrum representation of the acquired broadband time-domain signal to train the broadband spectrum perception model, which significantly reduces the model's dependence on labeled training samples.

[0054] 2. The present invention performs data enhancement on the power spectrum features extracted from broadband time-domain signals using the periodogram method and the multi-tap spectrum estimation method. Two data enhancement strategies are used to expand the training samples, which not only improves the sample utilization efficiency but also significantly enhances the generalization ability of the broadband spectrum sensing model.

[0055] 3. The present invention integrates the feature information of two different power spectrum representations of the signal sequence and extracts the features of different power spectra through two neural network branches (first and second input branches). Then, a feature data fusion mechanism is used to integrate the extracted deep feature vectors to mine the information in the signal power spectrum from multiple angles. Compared with the single sequence method, this feature extraction and feature fusion method has more stable performance, helps to improve the detection effect of spectrum perception, and effectively controls the false alarm probability, overcoming the problem of unstable false alarm probability in existing methods.

[0056] 4. The present invention extracts power spectrum features from the broadband time-domain signal to be detected and inputs them into a trained broadband spectrum sensing model for detection. Different decision thresholds are configured for different sub-bands, thereby accurately determining the occupancy status of each sub-band and overcoming the problem of unstable false alarm probability in spectrum sensing in existing methods.

[0057] Other features and advantages of the present invention will be described in the following description, and some advantages may become apparent from the description or may be understood through implementation of the present invention. The purposes and other advantages of the present invention may be realized and obtained through the contents particularly pointed out in the description and the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] The accompanying drawings are only for the purpose of illustrating particular embodiments and are not to be considered limiting of the present invention. Like reference symbols denote like parts throughout the drawings.

[0059] Figure 1 This is a flow chart of a deep learning broadband spectrum sensing method based on dual power spectrum representation input and data enhancement in an embodiment of the present invention;

[0060] Figure 2 Schematic diagram of the broadband spectrum sensing model architecture constructed in an embodiment of the present invention;

[0061] Figure 3 This is a schematic diagram of a specific network structure of a broadband spectrum sensing model according to an embodiment of the present invention;

[0062] Figure 4 Schematic diagram of a sub-band data enhancement method in an embodiment of the present invention. DETAILED DESCRIPTION

[0063] The preferred embodiments of the present invention will be described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not used to limit the scope of the present invention.

[0064] To address these issues, it is necessary to develop efficient spectrum sensing techniques based on multiple power spectrum representations under small sample data conditions. This method maximizes the advantages of signal power spectrum representation in an environment with relatively small sample data, expanding the training sample through data augmentation techniques to improve the accuracy and generalization performance of spectrum detection while strictly controlling the false alarm probability. The method in this invention improves the spectrum detection probability and generalization performance, and effectively reduces the false alarm probability.

[0065] The original signal is a wideband time-domain signal sequence x(n) acquired by a signal acquisition device, where n = 1, 2, ..., M-1, M is the total number of sampling points of the wideband time-domain signal, i.e., the length of the wideband time-domain signal, and the signal bandwidth is B. The entire bandwidth of the wideband time-domain signal is divided into N subbands, each with a bandwidth of B / N. The original wideband time-domain signal is decomposed into multiple subband signals.

[0066] like Figure 1 As shown, the present invention provides a deep learning broadband spectrum sensing method based on dual power spectrum characterization input and data enhancement, comprising the following steps:

[0067] Step S1, respectively extracting the power spectrum representation of the acquired broadband time domain signal using a periodogram method and a multi-tap spectrum estimation method to obtain a periodogram power spectrum and a multi-tap power spectrum;

[0068] Step S2: constructing a sample data set based on the periodogram power spectrum, the corresponding multi-tap power spectrum, and the corresponding sample labels; wherein the sample labels are the states of each subband, and the states include occupied and idle;

[0069] Step S3: constructing a broadband spectrum sensing model based on a convolutional neural network, and using the sample data set to obtain a trained broadband spectrum sensing model;

[0070] Step S4: Obtain the periodogram power spectrum and multi-tap power spectrum of the signal to be detected; use the trained wideband spectrum sensing model to obtain the confidence of multiple sub-bands; perform spectrum decision based on the decision threshold and the confidence of each sub-band to obtain the perception result of the signal to be detected.

[0071] Step S1 includes steps S11-S12.

[0072] Two different power spectrum estimation methods (including periodogram method and multi-tap spectrum estimation method) are used to extract two power spectrum representations of broadband time domain signals, and the periodogram power spectrum PG is obtained respectively. norm (k) and multi-tap power spectrum MTM norm (k).

[0073] like Figure 2 As shown, exemplarily, the original broadband time domain signal in the present invention includes BPSK (Binary Phase Shift Keying), QPSK (Quadrature Phase Shift Keying), PAM (Pulse Amplitude Modulation) and QAM (Quadrature Amplitude Modulation).

[0074] BPSK is one of the simplest digital modulation methods, which transmits data by changing the phase of the carrier signal. BPSK has good noise immunity.

[0075] QPSK is an extension of BPSK, which uses four phases (0°, 90°, 180°, and 270°) to represent four different states. Each state can carry two bits of information, so the data transmission rate of QPSK is twice that of BPSK. QPSK also has good noise immunity and is more spectrally efficient than BPSK.

[0076] PAM is a digital modulation technique that represents different symbols by varying the pulse amplitude. Common forms include 4-PAM and 8-PAM. While it offers high spectral efficiency, it also offers relatively weak noise immunity.

[0077] QAM combines amplitude modulation and phase modulation to achieve high spectral efficiency data transmission by simultaneously modulating signals on two channels: the in-phase (I) component and the quadrature (Q) component.

[0078] These modulation techniques are selected for use in different communication systems based on specific requirements and channel conditions. For example, BPSK and QPSK are commonly used in satellite and mobile communications because they offer strong anti-interference capabilities in environments with low signal-to-noise ratios or channel fading, ensuring communication reliability. QAM, on the other hand, is commonly used in broadband communication systems, such as cable television and wireless local area networks, to provide higher data transmission rates.

[0079] Step S11: using a periodogram method to estimate the power spectrum of the original broadband time-domain signal to obtain a periodogram power spectrum.

[0080] Obtaining the periodogram power spectrum includes:

[0081] Performing discrete Fourier transform on the broadband time domain signal to obtain a complex frequency domain signal;

[0082] The first periodogram power spectrum is obtained by shifting the center of the spectrum of the complex frequency domain signal, placing the frequency zero point at the center of the spectrum, and calculating the square of the modulus;

[0083] Performing data enhancement on the first periodogram power spectrum using an inter-subband spectrum rearrangement method and / or an intra-subband spectrum rearrangement method to obtain a data-enhanced second periodogram power spectrum;

[0084] The first and third periodogram power spectra are normalized to obtain the periodogram power spectrum.

[0085] (1) Perform discrete Fourier transform (DFT) on the broadband time-domain signal sequence x(m) to obtain the Fourier transformed sequence X(k) as a complex frequency-domain signal, as follows:

[0086]

[0087] Where X(k) is the result of discrete Fourier transform of the broadband time domain signal, which represents the complex spectrum of the signal in the frequency domain; k is the index variable in the frequency domain, k = 0, 1, …, K-1; K is the value range of the index variable in the frequency domain.

[0088] x(m) is the discrete sampling sequence of the broadband time domain signal; the value range is m = 0, 1, ..., M-1; the length N of the broadband time domain signal determines the resolution and computational complexity of the DFT.

[0089] (2) The complex frequency domain signal is shifted by shifting the spectrum center and placing the frequency zero point in the center of the spectrum to obtain the sequence XX(k) after the spectrum center is shifted.

[0090] The default output frequency range of the discrete Fourier transform is 0 to the sampling rate f s , corresponding to k = 0 to k = K-1;

[0091] k = 0 corresponds to the DC component, frequency 0 Hz;

[0092] k=K / 2 corresponds to Nyquist frequency f s / 2;

[0093] k>K / 2 corresponds to negative frequency components.

[0094] The purpose of frequency center shifting is to move the zero frequency (DC component) to the center of the spectrum, making the spectrum symmetrical and facilitating intuitive analysis of positive and negative frequency components.

[0095] The spectrum sequence is circularly shifted by K / 2 points. The original k=0 (DC component) is moved to the center position (k=K / 2); the original k=K / 2 (highest positive frequency) is moved to k=K-1; and the original k>K / 2 (negative frequency) is moved to the left half.

[0096] Calculate the square of the modulus of XX(k) after the center of the spectrum is moved to obtain the signal power spectrum sequence, and obtain the first periodogram power spectrum, as follows:

[0097]

[0098] Among them, XX(k) is the complex spectrum sequence after the spectrum center is moved, is the power spectrum of the first periodogram.

[0099] (3) performing data enhancement on the first periodogram power spectrum using an inter-subband spectrum rearrangement method and / or an intra-subband spectrum rearrangement method to obtain a data-enhanced second periodogram power spectrum;

[0100] Data enhancement is performed on the power spectrum of the first periodogram. The data enhancement methods include inter-subband spectrum rearrangement and intra-subband spectrum rearrangement. Both methods can be used simultaneously or one of them can be selected. The details are as follows:

[0101] a) Data enhancement is performed on the first periodogram power spectrum using the inter-subband spectrum rearrangement algorithm.

[0102] First, the first periodogram power spectrum is divided into N subbands:

[0103] P(k)=[P1,P2,…,P i ,…P N ]Formula (3)

[0104] Where P(k) represents the sub-band power spectrum sequence obtained by dividing the power spectrum of the first periodogram into N sub-bands; P1, P2, ..., P i ,…P N Represents the power spectrum of each subband.

[0105] The first periodogram power spectrum representation sequence is split into a plurality of sub-band power spectrum sequences.

[0106] Then, a randomly generated index sequence of length N is generated, where the index values ​​are integers between 1 and N and are not repeated, for example [3, 1, 8, 5, ..., N].

[0107] Finally, according to the generated index sequence, the split sub-band power spectrum sequence is rearranged to form a new broadband signal power spectrum sequence P new (k). Exemplarily,

[0108] P new (k)=[P3,P1,P8,P5,...,P N ].

[0109] Simulate different frequency band occupancy combinations to enhance the adaptability of subsequent models to changes in spectrum structure.

[0110] b) Data enhancement is performed using the intra-subband spectrum rearrangement algorithm.

[0111] First, the spectrum signal sequence in each sub-band is extracted from the power spectrum of the first periodogram;

[0112] Next, the spectrum signal sequence of each sub-band is randomly divided into blocks and shuffled to ensure that the spectrum sequence of each sub-band is different from the original sequence to increase the diversity of the data;

[0113] Then, the disrupted sub-band spectrum signal sequence is inserted into the first periodogram power spectrum sequence to obtain an enhanced broadband spectrum sequence.

[0114] Specific effects such as Figure 4 As shown in FIG, it is a power spectrum distribution diagram before and after enhancement of sub-band i. Before signal enhancement processing, the signal power spectrum in sub-band i is continuous and relatively concentrated, mainly concentrated in the center of the sub-band.

[0115] After signal enhancement, the power spectrum of the signal within the subband undergoes significant changes: the originally relatively concentrated power spectrum is split into multiple spectral fragments of varying lengths according to a specific random rule. These spectral fragments are then redistributed randomly and roughly evenly across the entire subband bandwidth. In other words, the position of each spectral fragment within the subband bandwidth is randomly determined, and overall, their distribution within the subband bandwidth is relatively uniform.

[0116] The sample data set obtained by performing data enhancement on the first periodogram power spectrum by using an inter-subband spectrum rearrangement method and / or an intra-subband spectrum rearrangement method is used as the second periodogram power spectrum.

[0117] (4) Normalize the set of the first periodogram power spectrum and the second periodogram power spectrum to obtain the periodogram power spectrum.

[0118] Power normalization normalizes the power of the signal power spectrum sequence to 1. The normalized power spectrum signal segment is as follows:

[0119]

[0120] Among them, PG norm (k) is the set of the normalized first and second periodogram power spectra, which serves as the final periodogram power spectrum; PG(k) includes the set of the first and second periodogram power spectra.

[0121] The broadband time domain signal extracted based on the periodogram method and the power spectrum after data enhancement and normalization are called periodogram power spectrum sequence.

[0122] Step S12: Calculate the power spectrum of the original broadband time-domain signal using a multi-tap spectrum estimation method to obtain a multi-tap power spectrum.

[0123] The obtaining of the multi-tap power spectrum includes:

[0124] The broadband time domain signal is passed through a multi-tap filter and then subjected to discrete Fourier transform to obtain a plurality of windowed complex frequency domain signals;

[0125] By shifting the center of the spectrum of the complex frequency domain signal, aligning the frequency zero point to the center of the spectrum, and calculating the square of the modulus of the frequency domain signal to obtain the initial power spectrum; then, according to the weight coefficient of each filter, the power spectrum is weighted and combined to obtain the first multi-tap power spectrum;

[0126] Performing data enhancement on the first multi-tap power spectrum by using an inter-subband spectrum rearrangement method and / or an intra-subband spectrum rearrangement method to obtain a data-enhanced second multi-tap power spectrum;

[0127] The first and second multi-tap power spectra are normalized to obtain the multi-tap power spectrum.

[0128] (1) Calculate the value of the filter with different taps L is the number of taps, the length of each filter is M, and the discrete Fourier transform of the filtered time domain sequence x(n) obtains multiple Fourier transformed sequences and the first multi-tap power spectrum, as follows:

[0129]

[0130] Where B is the bandwidth; L is the number of filters (the number of orthogonal windows used in the multi-tap power spectrum), which is limited by the time-bandwidth product and satisfies

[0131] (2) Calculate the weight coefficient of each filter as follows:

[0132]

[0133] in, Represents the energy of the lth tap.

[0134] By moving the spectrum center of the complex frequency domain signal, the frequency zero point is aligned to the center of the spectrum, and the XX after the spectrum center is moved is obtained. l (k), and calculate the square of the frequency domain signal; then, according to the weight coefficient of each filter, a weighted combination of multiple windowed power spectra is performed as the first multi-tap power spectrum, as follows:

[0135]

[0136] (3) performing data enhancement on the first multi-tap power spectrum using an inter-subband spectrum rearrangement method and / or an intra-subband spectrum rearrangement method to obtain a data-enhanced second multi-tap power spectrum;

[0137] The process here is consistent with the process of "performing data enhancement on the first periodogram power spectrum to obtain the second periodogram power spectrum after data enhancement" in step S11, and will not be repeated here.

[0138] (4) Normalizing the set of the first and third multi-tap power spectra to obtain the multi-tap power spectrum.

[0139] Power normalization normalizes the power of the signal power spectrum sequence to 1. The normalized power spectrum signal segment is as follows:

[0140]

[0141] Among them, MTM norm (k) is a set of the normalized first multi-tap power spectrum and the second multi-tap power spectrum; MTM(k) includes the first and second multi-tap power spectra.

[0142] Periodogram method power spectrum sequence PG norm (k) reflects the global spectrum characteristics of the signal, but is sensitive to noise; multi-tap power spectrum sequence MTM norm (k) Multi-tap smoothing reduces variance and is more suitable for low signal-to-noise ratio scenarios; the combination of the two provides complementary spectral information and enhances feature expression capabilities.

[0143] It should be noted that whether using the periodogram method to extract the power spectrum representation of the original broadband time domain signal or using the multi-tap spectrum estimation method to extract the power spectrum representation of the original broadband time domain signal, both methods provide sample data for subsequent broadband spectrum perception model training.

[0144] In actual operation, it is possible to flexibly choose to execute step S11 first and then step S12, or to execute step S12 first and then step S11, based on the characteristics of the signal and processing requirements, so as to achieve the best analysis effect.

[0145] The function of step S1 is to extract two power spectrum representations of the original broadband time domain signal through two different power spectrum estimation methods, namely the periodogram method and the multi-tap spectrum estimation method, to obtain the periodogram power spectrum and the multi-tap power spectrum, thereby providing rich feature information for the subsequent broadband spectrum perception model training.

[0146] Step S2: Specifically.

[0147] Constructing a sample data set based on the periodogram power spectrum, the corresponding multi-tap power spectrum, and the corresponding sample labels; wherein the sample labels are the states of each subband in the power spectrum, and the states include occupied and idle states;

[0148] If a primary user exists in a subband in the periodogram power spectrum and the multi-tap power spectrum, the subband is labeled as occupied, otherwise it is labeled as idle;

[0149] The labels of the sub-bands in the periodogram power spectrum and the multi-tap power spectrum are respectively combined into label vectors as sample labels.

[0150] Subbands refer to the multiple narrowband frequency bands into which a wideband signal is divided after spectrum analysis. In spectrum sensing, the entire bandwidth of a wideband signal can be divided into multiple subbands, each covering a specific frequency range. This division facilitates more detailed analysis and utilization of spectrum resources, as different subbands can be allocated to different primary users and may have different spectrum characteristics and occupancy.

[0151] A primary user is a user who has the right to use a specific frequency band. These users are typically authorized communications service providers or specific communications systems. They are legitimate users conducting normal communications on a specific frequency band. Detecting the presence of primary users is crucial to avoiding interference and ensuring the proper use of spectrum resources.

[0152] Exemplarily, it may be determined whether a primary user exists in a sub-band by energy detection or feature detection.

[0153] (1) Energy detection: This is a simple method that determines whether a primary user exists by calculating the energy of the signal within a subband and comparing it with a preset threshold. If the signal energy exceeds the threshold, it is considered that a primary user exists in the subband. This method is suitable for situations where there is a large difference between the noise and signal energy.

[0154] (2) For feature detection: Analyze the characteristics of the signal within the sub-band, such as power spectrum, modulation type, signal duration, etc., to identify the communication characteristics of the primary user. This method can provide richer information than energy detection, thereby improving detection accuracy.

[0155] The extracted periodogram power spectrum sequences and multi-tap power spectrum sequences are randomly divided into two sets, which serve as the training and validation sets for deep learning training. Each sample data includes two power spectrum representations: the periodogram power spectrum sequence and the multi-tap power spectrum sequence, as well as the sample label.

[0156] A single sample data is represented as {PG norm (k),MTM norm (k), Y}, where Y is the sample label, which is determined according to the occupancy status of each subband in the signal.

[0157] Y=[y1,y2,...,y N ], where y i ∈{0,1} is the label of subband i;

[0158] If there is a primary user in subband i, the subband label is "occupied", which is 1; otherwise, it is "idle", which is 0.

[0159] Finally, the labels of each sub-band are combined into a label sequence as the sample label of the power spectrum.

[0160] In practical applications, the number of subbands into which a signal is divided is determined through experiments and simulations to achieve optimal performance and efficiency. For example, an initial number of subbands can be selected, and then its performance can be evaluated through experiments and then adjusted as needed.

[0161] The function of step S2 is to use the periodogram power spectrum and multi-tap power spectrum and the corresponding subband states as labels to construct a sample data set, and randomly divide the extracted power spectrum sequence into a training set and a validation set. Each sample data includes two power spectrum representations, the periodogram power spectrum and the multi-tap power spectrum, to provide sample data for the subsequent training and verification of the broadband spectrum perception model.

[0162] Step S3 includes steps S31-S32.

[0163] Step S31: construct a broadband spectrum sensing model based on a convolutional neural network.

[0164] like Figure 2 The broadband spectrum sensing model architecture diagram shown in the figure shows that the broadband spectrum sensing model in the present invention is constructed based on a convolutional neural network; a convolutional neural network (CNN) is a deep learning model that is good at processing high-dimensional data with spatial or temporal structures.

[0165] like Figure 3 As shown in FIG, the broadband spectrum sensing model is used to extract features of the periodogram power spectrum and the multi-tap power spectrum in parallel, and then perform feature fusion.

[0166] The broadband spectrum sensing model includes, in sequence, a first and a second parallel input branch, a feature fusion layer, a first convolutional layer, a first average pooling layer, second and third convolutional layers, a second average pooling layer, a fourth convolutional layer, an adaptive average pooling layer, a linear layer, and an output layer;

[0167] The first input branch and the second input branch have the same structure, and both include three layers of branch convolution layers, namely the first, second and third branch convolution layers. They are used to extract the features of the periodogram power spectrum and the multi-tap power spectrum in the sample data respectively; and obtain the features F PG and F MTM ;

[0168] The feature fusion layer is used to transform F PG and F MTM Splice and get the fusion feature F cat ;

[0169] The first convolutional layer is used to transform the fusion feature F cat Performing a convolution operation to obtain a first signal feature;

[0170] The first average pooling layer is used to downsample the first signal feature to obtain a second signal feature;

[0171] The second and third convolutional layers are used to sequentially perform convolution operations on the second signal feature to obtain a third signal feature;

[0172] The second average pooling layer is used to downsample the third signal feature to obtain a fourth signal feature;

[0173] The fourth convolutional layer is used to perform a convolution operation on the fourth signal feature to obtain a fifth signal feature;

[0174] The adaptive average pooling layer downsamples the fifth signal feature to 256 dimensions to obtain a sixth signal feature;

[0175] The linear layer is a fully connected layer, used to convert the sixth signal feature into a 64-dimensional seventh signal feature;

[0176] The output layer includes a fully connected layer and a Sigmoid activation function, which is used to map the seventh signal feature into confidence levels corresponding to multiple sub-bands.

[0177] The broadband spectrum sensing model adopts a multi-input multi-output structure.

[0178] The first input branch inputs the periodogram power spectrum sequence PG norm (k);

[0179] The second input branch is the multi-tap power spectrum sequence MTM norm (k).

[0180] In the feature extraction stage, the first input branch and the second input branch are sequentially subjected to three layers of convolution operations. norm (k) and MTM norm (k) Perform feature extraction. The two branches have the same network structure.

[0181] Each branch contains three convolution modules with convolution kernels of different sizes. The high-dimensional feature information at different convolution scales is obtained through convolution operations. The two branches are used to extract the features of the input sample data of the first input branch and the second input branch respectively. The feature vectors outputted by each branch are represented as F PG and F MTM .

[0182] The first branch convolution layer is configured with 4 convolution kernels of size 7×1; the second branch convolution layer is configured with 4 convolution kernels of size 5×1; the third branch convolution layer is configured with 8 convolution kernels of size 3×1;

[0183] The first, second and third one-dimensional convolutional layers respectively perform convolution operations, batch normalization operations and LeakRelu activation operations on the inputs of the first, second and third one-dimensional convolutional layers.

[0184] The first branch convolution (Conv,4,7*1) is a traditional one-dimensional convolution, which realizes the preliminary feature extraction of the input data;

[0185] The second branch convolution (Conv,4,5*1) further deepens the feature extraction;

[0186] The third branch convolution (Conv,8,3*1) continues to mine data features.

[0187] The feature fusion layer (cat layer) converts F PG and F MTM Features are spliced ​​and fused.

[0188] Two eigenvectors F PG and F MTM The feature fusion layer performs splicing and fusion, and the feature vector after splicing is as follows:

[0189]

[0190] in, Represents a splicing operation.

[0191] The first convolution layer, the first average pooling layer, the second convolution layer, the third convolution layer, the second average pooling layer, and the fourth convolution layer are used in sequence to further extract the fusion feature F. cat The deep features of the signal are extracted and the length of the signal features is reduced at the same time. Finally, the adaptive average pooling layer (AdaptiveAvgPool layer) and the linear layer are used to convert the deep features into a fixed-length information representation. The output layer outputs the occupancy confidence of each subband.

[0192] Wideband spectrum sensing decisions are based on multiple binary classification problems. To flexibly control the model's decision threshold, the final output layer of the wideband spectrum sensing model is a fully connected layer with N neurons. The sigmoid activation function is used to determine the probability of occupancy for each subband, expressed as:

[0193]

[0194] Among them, γ i is the input that has not been processed by the sigmoid activation function. is the sub-band confidence normalized by the sigmoid function, that is, the output of the output layer.

[0195] The convolutional neural network structure used by the first and second input branches and the feature fusion layer is as follows: Figure 3 As shown in , all convolutional layers are one-dimensional convolutions.

[0196] "Conv,4,7*1" means that the convolution is a traditional one-dimensional convolution, the number of convolution kernels is 4, and the convolution kernel size is 7*1. Figure 3 The "Conv" layer in corresponds to the sequence convolution layer + batch normalization layer + activation layer, where the activation function is LeakRelu;

[0197] The linear layer "F1,64(8192)" is a linear layer with an input length of 8192 and an output length of 64.

[0198] The first convolutional layer, Conv32, is a 5*1 layer that uses 32 convolution kernels of size 5×1 for convolution operations, and is batch normalized and activated with LeakRelu.

[0199] The first average pooling layer "Avg_Pool,2" is an average pooling layer with a stride of 2, which downsamples the features;

[0200] The second and third convolutional layers (Conv64,1*1 layer with 64 1×1 convolution kernels and Conv64,5*1 layer with 64 5×1 convolution kernels) perform convolution, batch normalization and LeakRelu activation operations successively to further adjust and refine the features;

[0201] The second average pooling layer Avg_Pool, 2 layers for downsampling;

[0202] The fourth convolutional layer Conv32, 1*1, uses 32 convolution kernels of size 1×1 for convolution operation, and is activated by batch normalization and LeakRelu;

[0203] The adaptive average pooling layer "AdaptiveAvgpool,256" is an adaptive pooling layer with an output of 256;

[0204] The feature output is regularized into 256 dimensions by the adaptive average pooling layer AdaptiveAvgpool,256, and then transformed into 64-dimensional output by the linear layer F1,64(8192) linear layer. Finally, the output layer F2,N(64) (N number of subbands) performs further full connection transformation to obtain the final output of the model.

[0205] Step S32: training a broadband spectrum sensing model.

[0206] The training process of the broadband spectrum sensing model is as follows:

[0207] Using a training set in the sample data set to train the broadband spectrum sensing model;

[0208] During the training process, the Adam optimizer is used to iteratively update the model parameters and minimize the loss function;

[0209] If the accuracy does not improve within p training cycles, the learning rate in the current training cycle is reduced to 0.1 times that of the previous training cycle;

[0210] If the accuracy does not improve for q consecutive training cycles, the training is stopped and the network parameters with the minimum loss function on the validation set during the training process are saved as the trained wideband spectrum sensing model; where q>p.

[0211] To prevent overfitting, an early stopping strategy is adopted, that is: if the accuracy does not improve after p (such as 3) training cycles, the learning rate is reduced to 0.1 times the original; if the accuracy does not improve during q (such as 6) training cycles, training is stopped.

[0212] The convolutional neural network is trained using the training set. After the network has been trained for a period of time, the network performance is verified using the validation set.

[0213] The training process uses binary cross entropy as the loss function

[0214] The loss function is as follows:

[0215]

[0216] in, is the confidence of each sub-band output by the model; Y is the label sequence; N is the number of sub-bands into which the signal is divided; C is the total number of samples, The confidence of the i-th subband of the c-th sample output by the model; is the true label of the i-th subband of the c-th sample; BCE represents binary cross entropy; represents the average loss of all sub-bands for all samples.

[0217] The overall loss is calculated by averaging over all sub-bands to evaluate how well the wideband spectrum sensing model learns the information of each sub-band.

[0218] The Adam optimizer is used to iteratively update the model parameters and minimize the loss function.

[0219] The purpose of step S3 is to build a broadband spectrum perception model based on a convolutional neural network, using the periodogram power spectrum and multi-tap power spectrum sample data in the sample data set, as well as the corresponding sub-band state labels, and iteratively updating the model parameters to minimize the loss function, thereby obtaining a trained broadband spectrum perception model that can accurately perceive the occupancy status of each sub-band in the broadband spectrum.

[0220] Step S4 includes steps S41-S43.

[0221] Step S41: Obtain the periodogram power spectrum and multi-tap power spectrum of the signal to be detected, and use the trained wideband spectrum sensing model to obtain confidences of multiple sub-bands.

[0222] Acquire a signal to be detected, and extract two power spectrum representations, namely, a periodogram power spectrum and a multi-tap power spectrum, from the broadband time-domain signal to be detected;

[0223] The periodogram power spectrum and the multi-tap power spectrum of the signal to be detected are used with the trained wideband spectrum sensing model to obtain confidences of multiple sub-bands.

[0224] Step S42: Obtain the decision threshold of each sub-band.

[0225] The decision threshold is obtained as follows:

[0226] Extracting the corresponding periodogram power spectra and multi-tap power spectra of Z pure noise samples, inputting the trained wideband spectrum sensing model, and obtaining Z confidence scores corresponding to each subband;

[0227] Arrange the Z confidences of each sub-band in descending order to obtain a confidence sequence;

[0228] Based on the preset target false alarm probability P f , calculate the decision threshold position index P f *Z;

[0229] The first [P f *Z] value λ i , as the decision threshold for subband i, where [.] is a rounding operation.

[0230] Firstly, the power spectrum features of Z pure noise samples are extracted using the periodogram method and multi-tap spectrum estimation method to obtain the corresponding periodogram power spectrum and multi-tap power spectrum.

[0231] Next, Z pure noise samples are input into the trained wideband spectrum sensing model to obtain Z confidence values ​​corresponding to each sub-band, and the confidence values ​​of each sub-band are sorted in descending order from large to small.

[0232] For example, the preset target false alarm probability P f Set it to 0.01 and adjust it according to specific needs in specific applications.

[0233] Target false alarm probability P f In signal detection theory, it is used to measure the probability of misidentifying noise as a signal. In practical applications, this value usually needs to be set based on factors such as the system's fault tolerance and the cost of false alarms. For example, in some applications that are very sensitive to false alarms, a smaller P may be chosen.f In some applications that are more sensitive to false negatives, a larger P value may be selected. f value to reduce false negatives.

[0234] Step S43: Perform spectrum decision based on the decision threshold and the confidence level of each sub-band to obtain a perception result of the signal to be detected.

[0235] Spectrum judgment is performed based on the decision threshold and the confidence level of each sub-band to obtain the perception results of the signal to be detected, including:

[0236] The periodogram power spectrum and multi-tap power spectrum of the signal to be detected are input into the trained wideband spectrum sensing model to obtain the sub-band confidence of the corresponding N sub-bands to be detected on the spectrum occupancy state.

[0237] like The sensing result of the subband i in the signal to be detected is determined to be in an occupied state; otherwise, the sensing result of the subband i in the signal to be detected is determined to be in an idle state;

[0238] The perception results of the N to-be-detected sub-bands corresponding to the to-be-detected signal are combined into a vector as the perception result of the to-be-detected signal.

[0239] The two power spectrum sequences are input into the trained wideband spectrum sensing model to obtain the confidence of the spectrum occupancy status of the N sub-bands to be detected. Set the decision threshold λ of each subband according to the target false alarm probability i , and finally determine the occupancy status of each sub-band. Then subband i is determined to be occupied; otherwise, it is determined to be idle.

[0240] The function of step S4 is to analyze the signal to be detected using the trained wideband spectrum sensing model, obtain the confidence level and decision threshold of the sub-band, and perform spectrum decision on each sub-band, thereby obtaining the perception result of the signal to be detected.

[0241] In summary, the deep learning broadband spectrum sensing method based on dual power spectrum representation input and data enhancement in the embodiment of the present invention has the following beneficial effects:

[0242] 1. The present invention uses only a small number of time-domain broadband signal samples, and utilizes the periodogram method and multi-tap spectrum estimation method to extract the multi-power spectrum representation of the acquired broadband time-domain signal to train the broadband spectrum perception model, which significantly reduces the model's dependence on labeled training samples.

[0243] 2. The present invention performs data enhancement on the power spectrum features extracted from broadband time-domain signals using the periodogram method and the multi-tap spectrum estimation method. Two data enhancement strategies are used to expand the training samples, which not only improves the sample utilization efficiency but also significantly enhances the generalization ability of the broadband spectrum sensing model.

[0244] 3. The present invention integrates the feature information of two different power spectrum representations of the signal sequence and extracts the features of different power spectra through two neural network branches (first and second input branches). Then, a feature data fusion mechanism is used to integrate the extracted deep feature vectors to mine the information in the signal power spectrum from multiple angles. Compared with the single sequence method, this feature extraction and feature fusion method has more stable performance, helps to improve the detection effect of spectrum perception, and effectively controls the false alarm probability, overcoming the problem of unstable false alarm probability in existing methods.

[0245] 4. The present invention extracts power spectrum features from the broadband time-domain signal to be detected and inputs them into a trained broadband spectrum sensing model for detection. Different decision thresholds are configured for different sub-bands, thereby accurately determining the occupancy status of each sub-band and overcoming the problem of unstable false alarm probability in spectrum sensing in existing methods.

[0246] Those skilled in the art will appreciate that all or part of the process steps of the above-described embodiments can be implemented by instructing related hardware through a computer program, and the program can be stored in a computer-readable storage medium, such as a magnetic disk, an optical disk, a read-only memory, or a random access memory.

[0247] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by any technician familiar with this technical field within the technical scope disclosed by the present invention should be covered by the scope of protection of the present invention.

Claims

1. A deep learning broadband spectrum sensing method based on dual power spectrum representation input and data enhancement, characterized in that: include: The power spectrum representation of the acquired broadband time-domain signal is extracted using the periodogram method and the multi-tap spectrum estimation method, respectively, to obtain the periodogram power spectrum and the multi-tap power spectrum; Constructing a sample data set based on the periodogram power spectrum, the corresponding multi-tap power spectrum, and the corresponding sample labels; wherein the sample labels are states of each subband, and the states include occupied and idle; Building a broadband spectrum sensing model based on a convolutional neural network, and obtaining a trained broadband spectrum sensing model using the sample data set; The periodogram power spectrum and multi-tap power spectrum of the signal to be detected are obtained, and the confidence of multiple sub-bands is obtained using the trained wideband spectrum perception model; spectrum judgment is performed based on the decision threshold and the confidence of each sub-band to obtain the perception result of the signal to be detected.

2. The method according to claim 1, characterized in that The obtaining of the periodogram power spectrum comprises: Performing discrete Fourier transform on the broadband time domain signal to obtain a complex frequency domain signal; The first periodogram power spectrum is obtained by shifting the center of the spectrum of the complex frequency domain signal, placing the frequency zero point at the center of the spectrum, and calculating the square of the modulus; Performing data enhancement on the first periodogram power spectrum using an inter-subband spectrum rearrangement method and / or an intra-subband spectrum rearrangement method to obtain a data-enhanced second periodogram power spectrum; The first and second periodogram power spectra are normalized to obtain the periodogram power spectrum.

3. The method according to claim 2, characterized in that The obtaining of the multi-tap power spectrum includes: The broadband time domain signal is passed through a multi-tap filter and then subjected to discrete Fourier transform to obtain a plurality of windowed complex frequency domain signals; By shifting the center of the spectrum of the complex frequency domain signal, aligning the frequency zero point to the center of the spectrum, and calculating the square of the modulus of the frequency domain signal to obtain the initial power spectrum; then, according to the weight coefficient of each filter, the power spectrum is weighted and combined to obtain the first multi-tap power spectrum; Performing data enhancement on the first multi-tap power spectrum by using an inter-subband spectrum rearrangement method and / or an intra-subband spectrum rearrangement method to obtain a data-enhanced second multi-tap power spectrum; The first and second multi-tap power spectra are normalized to obtain the multi-tap power spectrum.

4. The method according to claim 3, characterized in that If a primary user exists in a subband in the periodogram power spectrum and the multi-tap power spectrum, the subband is labeled as occupied, otherwise it is labeled as idle; The labels of the sub-bands in the periodogram power spectrum and the multi-tap power spectrum are respectively combined into label vectors as sample labels.

5. The method according to claim 1, characterized in that: The broadband spectrum sensing model includes, in sequence, a first and a second parallel input branch, a feature fusion layer, a first convolutional layer, a first average pooling layer, second and third convolutional layers, a second average pooling layer, a fourth convolutional layer, an adaptive average pooling layer, a linear layer, and an output layer; The first input branch and the second input branch have the same structure, and both include three layers of branch convolution layers, namely the first, second and third branch convolution layers; they are respectively used for feature extraction of the periodogram power spectrum and the multi-tap power spectrum in the sample data; and the feature F is obtained respectively. PG and F MTM ; The feature fusion layer is used to transform F PG and F MTM Splice and get the fusion feature F cat ; The first convolutional layer is used to transform the fusion feature F cat Performing a convolution operation to obtain a first signal feature; The first average pooling layer is used to downsample the first signal feature to obtain a second signal feature; The second and third convolutional layers are used to sequentially perform convolution operations on the second signal feature to obtain a third signal feature; The second average pooling layer is used to downsample the third signal feature to obtain a fourth signal feature; The fourth convolutional layer is used to perform a convolution operation on the fourth signal feature to obtain a fifth signal feature; The adaptive average pooling layer downsamples the fifth signal feature to 256 dimensions to obtain a sixth signal feature; The linear layer is a fully connected layer, used to convert the sixth signal feature into a 64-dimensional seventh signal feature; The output layer includes a fully connected layer and a Sigmoid activation function, which is used to map the seventh signal feature into confidence levels corresponding to multiple sub-bands.

6. The method according to claim 5, characterized in that The first branch convolution layer is configured with 4 convolution kernels of size 7×1; the second branch convolution layer is configured with 4 convolution kernels of size 5×1; the third branch convolution layer is configured with 8 convolution kernels of size 3×1; The first, second and third branch convolutional layers respectively perform convolution operations, batch normalization operations and LeakRelu activation operations on the inputs of the first, second and third branch convolutional layers.

7. The method according to claim 1, characterized in that: The training process of the broadband spectrum sensing model is as follows: Using a training set in the sample data set to train the broadband spectrum sensing model; During the training process, the Adam optimizer based on stochastic gradient descent is used to iteratively update the model parameters and minimize the loss function. If the accuracy does not improve within p training cycles, the learning rate in the current training cycle is reduced to 0.1 times that of the previous training cycle; If the accuracy does not improve for q consecutive training cycles, the training is stopped and the network parameters with the minimum loss function on the validation set during the training process are saved as the trained wideband spectrum sensing model; where q>p.

8. The method according to claim 7, characterized in that: The loss function is as follows: in, is the confidence of each sub-band output by the model; Y is the label sequence; N is the number of sub-bands into which the signal is divided; C is the total number of samples, The confidence of the i-th subband of the c-th sample output by the model; is the true label of the i-th subband of the c-th sample; BCE represents binary cross entropy; represents the average loss of all sub-bands for all samples.

9. The method according to claim 1, characterized in that: The decision threshold is obtained as follows: Extracting the corresponding periodogram power spectra and multi-tap power spectra of Z pure noise samples, inputting the trained wideband spectrum sensing model, and obtaining Z confidence scores corresponding to each subband; Arrange the Z confidences of each sub-band in descending order to obtain a confidence sequence; Based on the preset target false alarm probability P f , calculate the decision threshold position index P f *Z; The first [P f *Z] value λ i , as the decision threshold for subband i, where [.] is a rounding operation.

10. The method according to any one of claims 1 to 9, characterized in that: Spectrum judgment is performed based on the decision threshold and the confidence level of each sub-band to obtain the perception results of the signal to be detected, including: The periodogram power spectrum and multi-tap power spectrum of the signal to be detected are input into the trained wideband spectrum sensing model to obtain the sub-band confidence of the corresponding N sub-bands to be detected on the spectrum occupancy state. like The sensing result of the subband i in the signal to be detected is determined to be in an occupied state; otherwise, the sensing result of the subband i in the signal to be detected is determined to be in an idle state; The perception results of the N to-be-detected sub-bands corresponding to the to-be-detected signal are combined into a vector as the perception result of the to-be-detected signal.

Images