An intelligent broadband spectrum sensing method based on pre-trained basic model
By adopting a time-frequency fusion-tuned deep convolutional neural network in spectrum perception, the problem that traditional methods are difficult to achieve high-precision broadband spectrum perception under low signal-to-noise ratio is solved, and efficient and low-cost spectrum support set reconstruction is achieved. , suitable for intelligent perception in complex electromagnetic environments.
Patent Information
- Application Number
- CN202210662006.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-06-13
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2042-06-13
AI Technical Summary
The existing traditional spectrum perception method based on sub-Nyquist sampling is difficult to achieve high-precision, high reconstruction capability broadband spectrum perception under low signal-to-noise ratio conditions under complex electromagnetic environments, and the false alarm rate and calculation cost are high.
An intelligent broadband spectrum perception method based on pre-trained basic model is proposed, using a time-frequency fusion-tuned deep convolutional neural network (TFF_aDCNN), through modulation broadband converter technology, principal component analysis method and discrete Fourier transform, etc. Steps: Combining the dual input structure of the main network and the auxiliary network, it realizes signal feature extraction and time information fusion, reduces hardware costs, and improves the reconstruction accuracy and reconstruction capabilities of spectrum support sets.
In a low signal-to-noise ratio environment, the spectrum support set reconstruction accuracy and reconstruction capability are significantly improved, the hardware cost is reduced, and the intelligent perception under different environmental conditions can be quickly adapted to the intelligent perception under different environmental conditions and achieved efficient spectrum perception.
Smart Images

Figure CN115209418B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the fields of deep learning and spectrum perception, and in particular to an intelligent broadband spectrum perception method based on a pre-trained basic model. Background Art
[0002] Spectrum is an important scarce strategic resource. With the rapid development of next-generation wireless communication technology and the Internet of Things, the types and number of devices connected to wireless networks have exploded. In order to meet the spectrum needs of various broadband devices, future 6G communication technology must move into the ultra-wide spectrum range of millimeter waves and terahertz. However, the current static spectrum allocation strategy has made the contradiction between low spectrum resource utilization and shortage of spectrum resources increasingly acute. In a complex electromagnetic environment, the traditional broadband spectrum sensing method based on Cognitive Radio (CR) is difficult to meet future performance requirements. With the rapid development of artificial intelligence technology, using AI to achieve intelligent spectrum sensing in an extremely wide spectrum range will be a feasible solution.
[0003] Spectrum sensing allows secondary users (SUs) to sense the spectrum occupancy status of authorized primary users (PUs) in a complex surrounding radio environment. SUs use spectrum sensing technology to discover spectrum holes and then access to achieve spectrum sharing, which can greatly improve spectrum utilization and alleviate the scarcity of high-quality spectrum resources. Most of the existing traditional spectrum sensing methods based on sub-Nyquist sampling do not determine whether there is a PU in the spectrum, but directly reconstruct the support set, resulting in high false alarm rate and computational cost.
[0004] With the substantial increase in computing power, deep learning has shown its powerful capabilities. Deep learning methods can find the mapping model between signals and support sets in complex radio environments without the need for manual feature extraction. Convolutional neural networks (CNNs) have become the basic feature extraction network in image processing due to their excellent feature extraction capabilities. Many studies have attempted to use the advantages of deep learning for spectrum sensing.
[0005] However, current research (whether using traditional or deep learning-based spectrum sensing methods) still suffers from the following three problems:
[0006] (1) The relationship between PU’s activity habits and time periods is often overlooked.
[0007] (2) In low signal-to-noise ratio (SNR) environments, existing wideband spectrum sensing (WSS) algorithms based on modulated wideband converters (MWCs) require more parallel channels to ensure the probability of support set reconstruction.
[0008] (3) In low signal-to-noise ratio environments, the reconstruction capability of existing schemes is also greatly limited. Summary of the invention
[0009] The purpose of the present invention is to provide an intelligent broadband spectrum sensing method based on a pre-trained basic model. The method of the present invention aims at the problem that it is difficult to achieve high-precision and high-reconstruction capability broadband spectrum sensing in an extremely wide frequency range at an extremely low signal-to-noise ratio. A time-frequency fusion adjustable deep convolutional neural network (TFF_aDCNN) and a novel sensing framework based on TFF_aDCNN are proposed. The method can significantly reduce the hardware cost and improve the reconstruction accuracy and reconstruction capability of the spectrum support set in a complex environment with a low signal-to-noise ratio.
[0010] The technical solution of the present invention is to provide an intelligent broadband spectrum sensing method based on a pre-trained basic model, comprising the following steps:
[0011] Step 1, obtaining sampling data using modulation broadband converter technology;
[0012] Step 2: preprocessing the sampled data;
[0013] Step 2.1, using the sampling result and the observation matrix to obtain the estimated original signal with the noise and aliasing signal superimposed;
[0014] Step 2.2: Use principal component analysis to perform denoising and dimensionality reduction;
[0015] Step 2.3, performing discrete Fourier transform on the denoised and dimension-reduced data to obtain a spectrum estimation signal;
[0016] Step 2.4: Separate the real and imaginary parts of the spectrum estimation signal as the input of the network.
[0017] Step 3: Input the data obtained in step 2 into the time-frequency fusion adjustable deep convolutional neural network (TFF_aDCNN) model to obtain the support set perception result.
[0018] Furthermore, the method described in step 2 also includes: the TFF_aDCNN network is a dual-input neural network. The network is divided into two modules, namely the main network and the auxiliary network. The input of the main network is the preprocessed data, and the input of the auxiliary network is a one-hot encoding containing time information. The input of the auxiliary network is a one-hot encoding, which is upgraded through a fully connected layer and then passes through a convolution module to obtain the output of the auxiliary network. The main network first passes through a layer of convolution to extract certain signal features, and then merges with the output of the auxiliary network in the channel direction. After fusion, the convolution module (three-layer convolution) continues to extract features and finally outputs through the fully connected layer.
[0019] The beneficial effects of the present invention are:
[0020] (1) The present invention innovatively proposes an adjustable deep convolutional neural network with time-frequency fusion (TFF_aDCNN). The network consists of a main network and an adjustment network. The former is mainly used to learn complex and abstract signal features, while the latter mainly assists the main network in learning different data distribution characteristics during training and is used to adjust the focus of the main network detection during detection. The use of this network model can significantly reduce hardware costs, improve reconstruction accuracy and reconstruction capabilities, and can perform well even under very low signal-to-noise ratios.
[0021] (2) The present invention can obtain a basic model through training under a simple distribution and then perform transfer learning in a complex environment to directly obtain a new broadband spectrum sensing model (fine-tuning model) under different perception environments. Compared with traditional deep learning methods, the present invention can quickly obtain a spectrum perception model and realize intelligent perception under various environmental conditions. BRIEF DESCRIPTION OF THE DRAWINGS
[0022] Figure 1 It is a summary figure of an intelligent broadband spectrum sensing method based on a pre-trained basic model of the present invention;
[0023] Figure 2 is a schematic diagram of a cognitive radio network model in which the occupied frequency bands in different time periods vary with the usage probability of the devices in an embodiment of the present invention;
[0024] Figure 3 is a schematic diagram of a network structure in an embodiment of the present invention;
[0025] Figure 4 1 is a schematic diagram of comparing the recovery probabilities of support sets under different channel conditions in an embodiment of the present invention;
[0026] Figure 5 is a schematic diagram of the comparison of the probability of support set recovery of data (m=20) with and without PCA denoising in an embodiment of the present invention;
[0027] Figure 6 is a schematic diagram of comparison of the recovery probability of the support set between DLWSS and TFF_aDCNN using data (m=20 and N=6) that have not been processed by PCA in an embodiment of the present invention;
[0028] Figure 7 is a schematic diagram of the reconstruction probability of the TFF_aDCNN support set under different numbers of channels in an embodiment of the present invention;
[0029] Figure 8 is a schematic diagram of the performance of a network under different signal-to-noise ratios and different signal frequency bands in an embodiment of the present invention;
[0030] Fig. 91 is a schematic diagram of the recall and precision performance of the TFFaDCNN model in the embodiment of the present invention when m=20 and m=25;
[0031] Fig.10 is a ROC curve reflecting the classification ability of the model in an embodiment of the present invention (SNR=-10dB, σ1, m=20 and N=6);
[0032] Fig.11 is the confusion matrix of TFFaDCNN in the embodiment of the present invention (SNR=-10dB, m=20 and σ1). DETAILED DESCRIPTION
[0033] The following will be combined with the embodiments and Figure 1-11 The concept, specific steps and technical effects of the present invention are clearly and completely described to fully understand the purpose, features and effects of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. The equivalent transformations or substitutions in functions, methods or structures made by ordinary technicians in the field according to these implementations are all within the protection scope of the present invention.
[0034] In the description of the present invention, "several" means one or more, "more" means more than two, "greater than", "less than", "exceed" etc. are understood as not including the number itself, and "above", "below", "within" etc. are understood as including the number itself. If there is a description of "first" or "second", it is only used for the purpose of distinguishing the technical features, and cannot be understood as indicating or implying the relative importance or implicitly indicating the number of the indicated technical features or implicitly indicating the order of the indicated technical features.
[0035] In the description of the present invention, unless otherwise clearly defined, terms such as setting, installing, connecting, etc. should be understood in a broad sense, and technicians in the relevant technical field can reasonably determine the specific meanings of the above terms in the present invention based on the specific content of the technical solution.
[0036] like Figure 1 As shown in the figure, this example provides an intelligent broadband spectrum sensing method based on a pre-trained basic model, including the following features:
[0037] Step 1, obtaining sampling data using modulation broadband converter technology;
[0038] like Figure 2 As shown, consider the broadband spectrum sensing link in an area, where there are several primary users PU and one secondary user SU. According to the actual situation, the probability of the primary user using the device in different time periods is different, so the spectrum occupancy in different time periods in the area is regular.
[0039] Assume the width is B wThe extremely wide frequency band is divided into L continuous narrow frequency bands with a width of B, and the frequency band index sequence is {1, 2, 3...L}. All primary users in the system are allowed to use a total of N continuous non-overlapping narrow frequency bands with a width of B at most simultaneously. The present invention divides a day into z time periods, and the time period T j ∈{T1,T2...T z The state of the kth frequency band in} is recorded as S k ∈{0,1},S k =1 means the kth frequency band is being used, S k = 0 indicates that the kth frequency band is idle. The primary user does not occupy all available frequency bands when using it. At this time, the secondary user can sense B w Status of each frequency band in the range B state ={S1,S2...S L}, the present invention will belong to T j The set of occupied frequency band numbers M in the time period is called the support set of the signal and is defined as:
[0040] Λ={M1,M2,......M N} (1)
[0041] Sparse multi-band signals are a type of signal often encountered in cognitive radio communications. In the embodiment of the present invention, it is assumed that the received signal x(t) is a sparse bandpass analog signal with a spectrum distributed in B w =[-f nyq / 2,f nyq / 2], f nyq is the Nyquist sampling rate of the signal. The received signal is as follows:
[0042] x(t)=p(t)+w(t) (2)
[0043] w(t) is an additive white noise that follows a Gaussian distribution, p(t) is the superposition of the signals sent by active primary users in the region at time t, and the maximum bandwidth of the signal sent by the primary user is B max , the signal energy of the used frequency band is E=[E1,...E N ].
[0044] Taking into account the actual situation, the activity of different primary users is closely related to the time period they are in, which is reflected in the spectrum that the occupied frequency bands have different distributions depending on the time period during observation. For example, around 8 o'clock in the morning, the occupation of the navigation band surges, and around 9 o'clock the frequency band used by operators is very active. After 12 o'clock in the evening, there are no particularly active frequency bands, but there are still primary users randomly occupying the spectrum. In each time period during the day, in addition to certain specific active frequency bands, other frequency bands will also be randomly occupied by primary users. This feature is very similar to the characteristics of a normal distribution. Therefore, in the embodiment of the present invention, it is assumed that the carrier frequency f in the daytime signal p(t) c (T j ) is subject to the mean μ(T j )={μ(T1),...μ(T z )}, the standard deviation is σ(T j )={σ(T1),...σ(T z )} is:
[0045] f c (T j )~N(μ(T j ),σ 2 (T j )) (3)
[0046] After midnight, the spectrum occupancy is random, and the carrier frequency of p(t) follows a uniform distribution:
[0047] f c (T j )~U(-f nyq / 2,f nyq / twenty four)
[0048] The MWC algorithm is a sub-Nyquist sampling method that uses parallel channels to achieve sub-sampling at low speed. Assuming that the number of parallel sampling channels is m, on a certain channel g, the signal x(t) is multiplied by a set of period T. p =1 / f p ±1 random alternating waveform C g (t),g∈{1,2...m}, realize the shift of the signal spectrum X(f), and then pass the cutoff frequency of 1 / 2T s The low-pass filter has a frequency range of F s =[-f s / 2,f s / 2], let f p =f s , so the signal sampled by the low-pass filter contains all the characteristics of the received signal x(t). According to the Fourier transform and the relevant knowledge of the signal, the spectrum of the signal obtained on channel g is easily obtained:
[0049]
[0050] c gl Signal C g (t) The coefficient after Fourier series expansion is C g (t). l is the coefficient of the original spectrum X(f) shifted.
[0051] The spectrum of a broadband sparse signal is finite, so the number of times X(f) is shifted l is not infinite, and assuming f p =f s , so the number of translations L0 can be determined:
[0052]
[0053] The relationship between the number of band slices L and L0 in the system model is L = 2L0 + 1. Equation (5) can be rewritten as:
[0054]
[0055] The spectrum of the compressed sampling result Y(n) of m parallel channels is expressed as Y(f), which can be written in matrix form:
[0056] Y(f)=ΦZ(f) (8)
[0057]
[0058] The observation matrix Φ consists of the row vector c of the weight coefficients of the m sampling channels, and its size is (m, L). Z(f) is a matrix of signals on L frequency bands, and its size is (L, d), where d is the number of samples after undersampling of each channel.
[0059] Step 2: preprocessing the sampled data;
[0060] Step 2.1, using the sampling result and the observation matrix to obtain the estimated original signal with the noise and aliasing signal superimposed;
[0061] From formula (8), we can know that the compressed sampling result Y(n) can get the estimated original signal The spectrum of is expressed as:
[0062]
[0063] It is the pseudo-inverse of the observation matrix. Through formula (10), we can get the estimated original signal with the superposition of noise and aliasing signal.
[0064] Step 2.2: Use principal component analysis to perform denoising and dimensionality reduction;
[0065] Principal component analysis (PCA) is a commonly used data dimensionality reduction method. The work of PCA is to find the direction with large variance based on the data, project the high-dimensional features to the direction with large variance to retain most of the information and complete the dimensionality reduction. When only the features of the main components are retained, the noise with small variance will be filtered out.
[0066] Perform two consecutive MWC undersamplings, convert the estimated signal obtained according to equation (10) into a size of (L×d,1) and expand it into a (L×d,2) matrix D in the column direction:
[0067]
[0068] Find the covariance matrix C of D:
[0069]
[0070] Then, the eigenvalues and eigenvectors of the covariance matrix C are obtained. The eigenvalues are arranged from large to small, and the dimension that best represents the data characteristics is selected, that is, the direction of the eigenvector corresponding to the largest eigenvalue. After the original signal is projected onto the new dimension, the data after PCA dimensionality reduction is obtained.
[0071] The horizontal axis of D represents features, and the vertical axis represents samples. The results of two samplings are regarded as two dimensions of a sample. Both dimensions represent estimated signals, but both dimensions are affected by Gaussian white noise. PCA is used to reduce the two dimensions to one dimension, and a feature that best represents the original signal can be obtained. Since the direction of the new coordinate axis found by PCA is the direction with the largest variance and best represents the original signal, the other dimension that is discarded contains most of the Gaussian white noise. PCA removes part of the Gaussian white noise by reducing it to one dimension.
[0072] Step 2.3, performing discrete Fourier transform on the denoised and dimension-reduced data to obtain a spectrum estimation signal;
[0073] Data after PCA denoising It is necessary to use discrete Fourier transform to obtain the spectrum estimation signal
[0074]
[0075] Step 2.4: Separate the real and imaginary parts of the spectrum estimation signal as the input of the network.
[0076] Will After the real part and the imaginary part are separated, they are stacked as two signal features in the third dimension to become the input signal I of size (L, d, 2).
[0077] The traditional compressed sensing method uses the result of MWC undersampling, selects the most relevant element of the signal or residual through the greedy algorithm in each iteration, and runs multiple iterations to recover the support set:
[0078]
[0079] where j k is the residual vector r in the dictionary matrix V k-1 The most relevant column index, Λ k is the support set, which is updated through iteration.
[0080] Step 3: Input the data obtained in step 2 into the time-frequency fusion adjustable deep convolutional neural network (TFF_aDCNN) model to obtain the support set perception result.
[0081] The deep learning method also uses the results of MWC to reconstruct the support set. Considering that in the model proposed by the present invention, the frequency band occupied by the PU is related to the time period, the present invention adds information of the time dimension. Therefore, the present invention proposes a Figure 3 The system framework shown in Figure 1 maps the signal to a spectrum state. One-hot encoding uses c state representation bits to represent N time periods, each bit represents a time period, and the bit value is 0 or 1, where 1 represents that the current state is valid. z The data obtained by artificially dividing the time period in the time dimension is a one-hot encoding It is sent to the TFF_aDCNN network, and the TFF_aDCNN model is used to map the estimated signal to the frequency band state B state :
[0082]
[0083] Among them, θ is the parameter of the network. At the same time, deep learning transforms the original support set reconstruction problem into a multi-label classification problem. It only needs to judge the status of all frequency bands to get the final frequency band status. The multi-label problem is further simplified into multiple single-label binary classification problems. Finally, the Sigmoid activation function is used to solve the binary classification problem of judging L spectrum states.
[0084] The present invention proposes a dual-input convolutional neural network. The two inputs of the network represent the signal dimension and the time dimension respectively. For the input of the time dimension, when a certain H is selected, a specific state bit is 1, and other state bits are 0. The output of the fully connected layer in the adjustment network is only affected by the current state bit being 1, while the state bit being 0 cannot affect the output of the fully connected network. During the detection process, the main network adjusts the spectrum sensing of a specific data distribution by inputting a specific one-hot encoding.
[0085] Then the training set of the network can be defined as:
[0086] χ={[(I (1) ,H (1) ),y (1) ],…,[(I (w) ,H (w) ),y (w) ]} (16)
[0087] Among them, w represents the number of samples in the training set, and the input of each sample is composed of the input signal I, And the unique hot encoding H of the time period of the input signal, Assume that the training network parameter is θ and the label is y. The function of network learning is ψ. The present invention adopts supervised training, and finally activates the Sigmoid function of the fully connected layer to limit the output value range to (0,1). The cross entropy loss function l is:
[0088]
[0089] where y i is the value of the i-th spectrum slice in the training label, the occupied band is 1, the idle band is 0, ψ i (I|H,θ) n Represents a mapping of the network to the state of the i-th spectrum slice of the n-th training sample, representing the probability of the frequency band being occupied in the form of 0 to 1, and N is the number of training samples. The training of the TFF-cDCNN network is mainly driven by data. During training, the TFF-cDCNN network uses the Adam (Adaptive moment estimation) optimizer to learn based on the unique hot encoding of the current time period input by the control network and the data of the main network input that conforms to the distribution of the current time period.
[0090] When the one-hot encoding representing a certain time period is input into the control network, the main network inputs data with the distribution characteristics of this time period. After learning a large amount of data with the distribution characteristics of the current time period, the network reduces the loss of the current input combination (the one-hot encoding of the current time period and the data that conforms to the distribution characteristics of the current time period). Under the current one-hot encoding input, the main network passively learns the distribution characteristics of the input data. A large amount of training makes TFF-cDCNN more sensitive to data that obeys the distribution of the current time period when the one-hot encoding of the current time period is input into the control network.
[0091] After the training is completed, the embodiment of the present invention actively controls the focus of the TFF-cDCNN main network detection by inputting the unique hot encoding representing different time periods into the control network, so that under low signal-to-noise ratio, the network can still perform spectrum perception focusing on the current data distribution by inputting the unique hot encoding of the control network, thereby achieving performance improvement. The training of the network can be regarded as an optimization problem:
[0092]
[0093] Whenever different conditional input H is selected for learning, the parameters of the control network are learned according to the current input H, so that the control network can learn the corresponding parameters of H representing each time period through back propagation, thereby assisting the main network to learn the distribution characteristics of the data under different H inputs.
[0094] like Figure 2 As shown in the figure, the proposed network model consists of two parts: the main network and the auxiliary network. The main network consists of a convolutional part and a fully connected part. The convolutional part includes four convolutional units, each of which consists of a convolutional layer with a convolution step of 1, a batch normalization layer, and an activation layer using the ReLU function. The fully connected part consists of a fully connected network with a layer of L neurons, and uses the Sigmoid activation function to limit the output range to (0,1) to represent the band state.
[0095] Connect the input data I of the main network to the convolution part. After the data passes through the first convolution unit with a convolution kernel K1 of 1x21 and 16 channels, the effective information is extracted and the first feature map is obtained:
[0096]
[0097] Since the convolution kernel is set to 16 channels, the output is arrive 16 matrices, each matrix is as follows:
[0098]
[0099] The values in the matrix are obtained by convolving the input matrix with the convolution kernel. The value at each position is obtained by the following formula:
[0100]
[0101] Where ξ() represents the ReLU activation function, and a and b are the values of row a and column b in a channel of the feature map. Because the size of the input data I is (L, d, 2), L is the number of spectrum slices, so the first dimension of the input data I has the structure and position information of the spectrum slices. In order to ensure that the data information is not lost, the first dimension of the convolution kernel is set to 1. In the output feature map of each subsequent convolution unit, the first dimension is L unchanged.
[0102] The auxiliary network consists of a fully connected part and a convolutional part. Considering that the auxiliary network and the main network need a 16-channel feature map with the same dimension and shape for feature concatenation, in order to match the number of parameters, the fully connected part consists of a fully connected layer with a number of neurons of L×(d-20). The parameter setting not only meets the subsequent feature concatenation, but also maps the low-dimensional time information to a higher dimension, increases the learnable parameters of the network, and improves the learning ability of the network. The convolutional part consists of two convolutional units, each of which consists of a convolutional layer with a convolution kernel of 1x21, 16 channels, and same padding, a batch normalization layer, and an activation layer using the ReLU activation function. The unique hot encoding representing the time dimension information is connected to the fully connected part in the auxiliary network. The data after the data dimension increase cannot be directly input into the convolutional part, so the fully connected output data is reshaped into a two-dimensional matrix and then connected to the convolutional layer. The convolution layer has a powerful ability to mine information features and can mine the features of time information. Finally, the convolution layer outputs a feature map R2 with time information. The method of the present invention concatenates the feature map containing input data information with the feature map containing time information output by the auxiliary network to form a new feature map with 32 channels:
[0103]
[0104] The concatenated feature map R containing time dimension features and signal features concatenateThe second convolution unit with an input convolution kernel of 1x21 and a number of channels of 32. The second convolution unit fuses the information of the time dimension with the information of the input data to further extract the relationship and features between time and data. The convolution kernels of the third and fourth convolution units are set to a size of 1x31, and the number of channels is set to 32, which further extracts the fused features and finally obtains the output of the convolution part. Since the method of the present invention uses a fully connected network for final classification, the feature map output by the convolution part is connected to the fully connected part after passing through a flattening layer. The fully connected part obtains the scores of L spectrum slices through a sigmoid activation function, where 1 represents that the frequency band is occupied and 0 represents that the frequency band is idle. In order not to lose information, the pooling layer is not used in the convolution unit in the embodiment of the present invention, and in order to reduce the parameters between the result of the flattening layer and the fully connected layer, the number of parameters is reduced as much as possible by setting the convolution kernel and the number of channels, and the last layer of convolution outputs a feature map of size (L, 1, 32).
[0105] According to the literature on spectrum measurement (V aclav Valenta, Roman Marˇs alek, Genevi`eve Baudoin, Martine Villegas, Martha Suarez, and Fabien Robert. Survey on spectrum utilization in europe: Measurements, analyses and observations. In 2010 Proceedings of the fifth international conference on cognitive radiooriented wireless networks and communications, pages 1–5. IEEE, 2010.) (Roger B Bacchus, Antoni J Fertner, Cynthia S Hood, and Dennis A Roberson. Long-term, wide-band spectral monitoring in support of dynamic spectrum access networks at the iit spectrum observatory. In 2008 3rd IEEE Symposium on New Frontiers in Dynamic SpectrumAccess Networks, pages 1–10. IEEE, 2008.) (Tanim M Taher, Roger B Bacchus, Kenneth J Zdunek, and Dennis A Roberson. Long-term spectral occupancy findings in chicago. In 2011 IEEE international symposium on dynamic spectrum access networks (DySPAN), pages 100–107. IEEE, 2011.) (Adnan Ahmad Cheema and Sana Salous. Spectrum occupancy measurements and analysis in 2.4ghz wlan. Electronics, 8(9):1011, 2019.) (Yunfei Chen and Hee-Seok Oh.A survey of measurement-based spectrum occupancy modeling for cognitive radios. IEEE Communications Surveys & Tutorials, 18(1): 848–859, 2014.) The spectrum usage patterns in the same city are different, and even the same frequency band has different activity patterns in different regions. Therefore, it is difficult to have a general model that can learn the regularity of spectrum usage in different places.
[0106] The model obtained by training TFF_aDCNN is to explore the effectiveness of the proposed spectrum sensing framework and use it as a pre-trained basic model. Obviously, this basic model cannot be used as a general and specific deep learning model, and it cannot cope with broadband spectrum sensing tasks in all electromagnetic environments.
[0107] In practical applications, for a specific task, it is necessary to first conduct spectrum measurements on site for at least one day, and find the distribution pattern of the authorized spectrum based on the results of the spectrum measurements. Then, the distribution is modeled, and a considerable amount of training set data is generated through simulation. Next, the training set is used for transfer learning based on the pre-trained basic model. In the process of transfer learning, the method of the present invention freezes the main network in the pre-trained basic model and only activates the adjustable auxiliary sub-network, so that a practical new model suitable for new scenarios can be quickly trained, that is, it can quickly adapt to the local spectrum usage pattern.
[0108] Assume that the signal sent by the primary user is:
[0109]
[0110] Among them B max =50MHz, the Nyquist sampling frequency of the signal is f nyq=10GHz, the extremely wide frequency band Bw is divided into 195 continuous narrow frequency bands, allowing the main user to use up to N = 6 frequency bands including the symmetric frequency band at the same time, Ei = [1,2,3]. In order to simplify the experiment, a day is divided into z = 5 time periods in the embodiment of the present invention. The first four time periods are assumed to be the daytime radio environment, which is relatively regular. The carrier frequency obeys a normal distribution with a mean of μ = [0.2, 0.4, 0.6, 0.8], σ1 = [0.1, 0.1, 0.1, 0.1], and a normal distribution random number is used to specify the carrier frequency. The last time period is assumed to be at night, and the carrier frequency obeys a uniform distribution. A uniform distribution random number is used to specify the carrier frequency. The MWC sampling parameters are set to the number of channels m = 25, and 101 data are sampled for each channel to finally obtain a Y(f) dimension of 25x101. The dimension of the observation matrix Φ is 25x195, and a signal with a dimension of 195x101x2 is obtained after data preprocessing. The bandwidth of each narrowband frequency band is f nyq / L=51.28MHz, w(t) represents Gaussian white noise.
[0111] The software environment of the simulation experiment is as follows: 64-bit Win10 operating system, MATLAB 2018a, TensorFlow1.15.2, Cuda10.0 / cudnn7. The hardware environment is: Inter(R)Core(TM)i5-8300HCPU@2.30GHz, RAM 16GB, GPU is NVIDIAGeForce GTX 1080Ti.
[0112] Table 1. Settings of simulation experiment parameters
[0113]
[0114] Experiment 1: Support set reconstruction performance under different signal-to-noise ratios
[0115] The performance of the models with different signal-to-noise ratios is compared. First, the data is denoised by PCA. The signal-to-noise ratio of the received signal {x(t)} is SNR = {-10, -8, -6, -4, -2, 0} dB, and the other conditions are the same as the simulation settings. Under each SNR condition, the training set has 8,500 data and the test set has 1,500 data. Each set of data is divided into 5 groups, and each set of data follows a specific distribution. The experiments use the proposed TFF_aDCNN network, the TFF_aDCNN network without adjustment, SwSOMP (ZhuhuaHu, YongBai, Yaochi Zhao, and Yiran Zhang. Support recovery for multiband spectrum sensing based on modulated wideband converter with swsompalgorithm. In International Conference on 5G for Future Wireless Networks, pages 146–159. Springer, 2017.), SOMP (Joel A Tropp, Anna C Gilbert, and Martin JStrauss. Algorithms for simultaneous sparse approximation. Part I: Greedy pursuit. Signal processing, 86(3): 572–588, 2006.) and DWLSS (
[29] Shivam Chandhok, Himani Joshi, AV Subramanyam, and Sumit J. Darak. Novel deep learning framework for wideband spectrum characterization at sub-nyquist rate. Wireless Netw, 27(7):4727–4746, aug 2021.)
[0116] The ablation experiment results are as follows Figure 4As shown. When m=25, N=6, SNR={-10,-8,-6,-4,-2,0}dB, the average reconstruction probability of the proposed model is 3.5% higher than that of the TFF_aDCNN model without adjusting the network. The performance of DLWSS is similar to that of TFF_aDCNN without adjusting the network. Compared with traditional methods such as SOMP and SwSOMP, the average improvement rates of the proposed model are 65% and 69%, respectively. The results also prove that TFF_aDCNN can learn the data characteristics of each time period, so under low signal-to-noise ratio, the model can also use a single encoding of the input to obtain a higher support set reconstruction probability. At the same time, in the embodiment of the present invention, the standard deviation condition of the data distribution is set to σ2=[0.04, 0.04, 0.04, 0.04], and the support set reconstruction probability experiment is carried out when m=20 and N=6. From Figure 4 From the above, we can see that when the data standard deviation is set to σ2, the network performance is better, which means that the more obvious the usage pattern of the main users in the current time period is, the better the network can discover the pattern and learn the data distribution characteristics.
[0117] PCA denoising has greatly improved the performance of the TFF_aDCNN network. Figure 5 As shown in Figure 2, the performance of network support set reconstruction probability is compared with and without PCA processing data when m = 20, N = 6, and σ1. Figure 5 It can be seen that the results of the ablation experiment show that when the signal-to-noise ratio is very low, PCA cannot effectively remove noise. However, as the signal-to-noise ratio increases, the TFF_aDCNN network performs better on the data after PCA denoising. Compared with using data without PCA processing, denoising can improve the performance of network support set reconstruction, especially when SNR = -2dB, the support set reconstruction rate is improved by 24.66%.
[0118] Then, the estimated signal without PCA denoising is input into the TFF_aDCNN network and the DLWSS network to compare the support set reconstruction performance. The data without PCA denoising contains more Gaussian white noise, which can better show the performance of the model itself under low signal-to-noise ratio conditions. Since DLWSS is also designed for specific MWC parameters, the convolution filter size in the DLWSS network is adapted to the data set in the embodiment of the present invention, and the network is reproduced according to the method in the original paper. When N=6, m=20, σ1 and SNR={-10,-8,-6,-4,-2,0}dB, the experimental results are shown in Figure 6. Experiments have shown that, without PCA denoising, the TFF_aDCNN network support set reconstruction probability proposed by the method of the present invention is 9.73% higher than DLWSS on average when SNR = {-10, -8, -6, -4, -2, 0} dB, only in terms of the performance of the network itself in a low signal-to-noise ratio environment. The data after PCA denoising improves the SNR, while the data without PCA processing can better reflect the performance of the network itself, which strongly proves that the TFF_aDCNN network successfully integrates the input time. This also shows that TFF_aDCNN is able to learn the data distribution characteristics of each H input, and use the learned data distribution to assist the network in spectrum perception under low signal-to-noise ratio.
[0119] Experiment 2: Performance under different numbers of parallel channels
[0120] The more channels the MWC has, the better the model performs. In the embodiment of the present invention, the PCA denoised data is used to study the performance of the TFF_aDCNN network with N=6, σ1, SNR=-6dB and SNR=-8dB when the number of channels is 2 in the interval [15,35].
[0121] After comparison, it is found that under the conditions of low channel number and low signal-to-noise ratio, the TFF_aDCNN network can obtain a higher support set reconstruction probability. The reconstruction accuracy of the traditional method SwSOMP is less than 20%. Figure 7 From the above, we can see that under the conditions of SNR = -6dB and N = 35, the reconstruction probability of TFF_aDCNN reaches more than 90%, which is a high-probability reconstruction performance. This reduces the hardware requirements for WSS, making the network more suitable for deployment, with less hardware consumption and more environmentally friendly.
[0122] Importantly, if Figure 7As shown, the TFF_aDCNN network and the DLWSS network are also compared in terms of reconstruction accuracy on the support set. When the number of parallel channels is close to the theoretical lower limit (i.e., m=15, the theoretical lower limit is 2N+1=13 (Moshe Mishalian and Yonina C Eldar. From theory to practice: Sub-nyquist sampling of sparse wideband analog signals. IEEE Journal of selected topics in signal processing, 4(2):375–391, 2010.) (HJ Landau. Necessary density conditions for sampling and interpolation of certain entire functions. Acta Mathematica, 117:37–52, 1967.), the performance improvement of TFF_aDCNN is not very obvious compared with DLWSS. However, as the number of channels m increases (i.e., m>15), the performance of TFF_aDCNN is greatly improved. When m>27 and SNR=-6dB, the performance improvement of TFF_aDCNN tends to be flat compared with DLWSS, and its average performance improvement can reach 22.1%.
[0123] Therefore, the TFF_aDCNN model has better reconstruction performance than the same type of models and traditional optimization models under low signal-to-noise ratio and low hardware complexity.
[0124] Experiment 3: Performance under different signal frequency band conditions
[0125] Then, the embodiment of the present invention studies the influence of different N on the support set reconstruction capability of the TFF_aDCN network. Considering the symmetrical frequency bands, the value of the number of frequency bands must be an even number, so a new data set is created in the embodiment of the present invention, N = 2, 4, 6, and the data set size is the same as the data set in the support set reconstruction performance experiment under different signal-to-noise ratios. PCA is used to denoise the signal, the signal-to-noise ratio value is {-10, -8, -6, -4, -2, 0} dB, the number of parallel MWC channels is m = 20, and other parameters are the same as the simulation settings. Figure 8 As shown in Figure 3, as the number of signal frequency bands increases, the network performance at low signal-to-noise ratio will be severely degraded.
[0126] When N = 2 and the number of channels m = 20, the support set reconstruction rate is still close to 90% when SNR = -10dB. When N = 4 and the number of channels m = 20, the performance degradation is more serious than when N = 2, but when SNR = 0dB, the support set reconstruction rate can still reach 90%. When N = 6 and the number of channels m = 20, the performance further degrades and it is impossible to obtain a high reconstruction probability below SNR = 0dB. The number of signal frequency bands has a great impact on the performance of TFF_aDCNN.
[0127] Experiment 4: TFFaDCNN structure discussion
[0128] In order to discuss the rationality of the network structure, two different adjustment network structures are compared in the embodiment of the present invention, and the performance of the two structures in terms of reconstruction ability is compared, using PCA denoised data, and the number of parallel channels is m=25, σ1, N=6.
[0129] The experimental results are shown in Table 2, from which it can be found that when two layers of CNN are used as the adjustment network in the TFF_aDCNN structure proposed in the present invention, it has advantages over the adjustment network using only one layer of CNN. Considering the number of parameters and the support set reconstruction probability of the frequency band, the TFF_aDCNN network proposed in the present invention has a suitable number of parameters and favorable performance.
[0130] Table 2. Performance of two adjusted network structures (σ1, m=25 and N=6).
[0131]
[0132] Experiment 5: TFFaDCNN Classification Capabilities
[0133] The classification ability of TFF_aDCNN is evaluated using the receiver operating characteristic (ROC) curve and confusion matrix in the embodiment of the present invention. The confusion matrix is a matrix that reflects the classification ability of the model. Since the binary classification method is used to determine the spectrum state in the embodiment of the present invention, a matrix of size (2,2) can be obtained, such as Fig.11 As shown in the figure, the vertical 0 and 1 represent the true labels, and the horizontal 0 and 1 represent the labels predicted by the model. Then the importance of these four areas is defined as true positive (TP), true negative (TN), false positive (FP) and false negative (FN).
[0134] Since the spectrum is sparse, the number of positive and negative samples is disproportionate. Recall and precision are needed to describe the model's ability to judge positive samples.
[0135] Recall=TP / (TP+FN) (24)
[0136] Precision=TP / (TP+FP) (25)
[0137] Among them, emph in formula (24) reflects the sensitivity of the model to positive cases, while emph in formula (25) reflects the accuracy of the model's judgment on positive cases. However, in spectrum sensing, precision is not as important as recall. When the false alarm rate of the network is high, the reconstructed support set is likely to contain the true support set, which does not affect the PU's use of the network. On the contrary, when the recall rate is low, the reconstructed support set lacks the truly occupied frequency bands, which will greatly affect the normal communication of the PU.
[0138] The data set in the support set reconstruction performance experiment under different signal-to-noise ratios is used to calculate the recall and precision, N = 6, σ 1. Fig. 9 From the recall images in , it can be found that when SNR<-4dB, the recall rates of the two models are similar, while the TFF_aDCNN network proposed in this invention can show better performance at lower SNR. This shows that the TFF_aDCNN network is more sensitive to frequency band occupancy at low SNR and can capture occupied frequency bands with a higher probability at low SNR.
[0139] According to the precision data, TFF_aDCNN can more accurately judge the state of PU when the number of channels is m=25. When m=20, the precision is poorer than that of TFF_aDCNN without adjusting the network. However, TFF_aDCNN can obtain a higher support set reconstruction probability under the condition of poor precision, which means that after reducing the number of channels, TFF_aDCNN can still use the learned relationship H between the frequency band occupancy features and the input to assist the main network in correctly judging the spectrum state. Experiments have shown that the existence of the adjustment network is necessary.
[0140] ROC can reflect the classification ability of the model. In the embodiment of the present invention, the spectrum sensing problem is regarded as a binary classification problem for each frequency band. Fig.10 As shown, the ROC curve can reflect the classification ability of the model. In the embodiment of the present invention, the data under SNR = -10dB, m = 20 and N = 6 are used for classification test. The first four time periods obey the normal distribution, and the standard deviation σ1 = [0.1, 0.1, 0.1, 0.1]. The results show that compared with TFF_aDCNN without adjusting the network, the network of the method of the present invention has better band state classification ability.
[0141] pass Fig.11From the experiment of the confusion matrix in the figure, it can be clearly found that the TFF_aDCNN network can correctly predict more positive classes at low signal-to-noise ratio, and the recall rate and precision are higher than the TFF_aDCNN without adjusting the network. The TFF_aDCNN model correctly predicted 738 positive cases than the TFF_aDCNN model without adjusting the network.
[0142] Experiment 6: Model transferability
[0143] In order to verify the transferability of the pre-trained basic model proposed in the present invention, a new data set is produced for a new electromagnetic environment in the embodiment of the present invention. The parameters are set as SNR = [-10, -8, -6, -4, -2, 0] dB, m = 20, and N = 6.
[0144] The pre-trained basic model assumes that the sub-band activity in each time period is normally distributed. By changing the distribution of the spectrum, the embodiment of the present invention also simulates an unfamiliar electromagnetic environment, that is, the single normal distribution in each time period is changed to two normal distributions in each time period. After loading the pre-trained basic model, the trained main network is frozen, and only the model parameters of the adjustable network are trained. Therefore, a broadband spectrum sensing fine-tuning model can be obtained in the embodiment of the present invention, which can adapt to the new electromagnetic distribution environment through transfer learning.
[0145] Table 3 shows that the pre-trained basic model performs well in a single distribution electromagnetic environment. For sensing tasks in complex electromagnetic environments, the pre-trained basic model is used for transfer learning in the embodiment of the present invention, during which only the adjustable sub-network is retrained. It can be seen from Table 3 that the fine-tuned model also has good performance. Compared with the model fully trained from scratch, it can be concluded that the fine-tuned model can adapt to different electromagnetic environments. At the same time, the pre-trained basic model has the ability of transfer learning.
[0146] Table 3. Performance of three models (σ1, m=20, N=6).
[0147]
[0148] Experiment 7: Time Cost
[0149] The overhead of the entire system framework is simply estimated. The entire system can be divided into the MWC sampling stage, the data preprocessing stage, and the TFF_aDCNN network stage. In the embodiment of the present invention, MWC is used to sample a signal 100 times, and the average time is calculated. Then, the time of data preprocessing is recorded. Finally, the trained model is used for prediction, and the time is recorded. The tensor of the prediction result is obtained.
[0150] The experimental results in Table 4 show that MWC has the highest time cost, followed by the TFF_aDCNN network, and finally the data preprocessing part. In order to denoise the signal during data preprocessing, two MWC samples are used in the embodiment of the present invention, which takes some time, but this time is exchanged for a higher probability of support set reconstruction. It is worth noting that the IEEE802.22 standard proposes two sensing methods for PU services, namely fast sensing and fine sensing. The spectrum uses a frequency band space that changes on a larger time scale, and the real-time requirements for the sensing cycle are not so strict.
[0151] It can be seen from Table 5 that the transfer learning method based on the pre-trained basic model can achieve a faster training speed, which also verifies the correctness of the theoretical analysis of the TFF_aDCNN model. Therefore, in practical application deployment, a grounded broadband spectrum sensing model can be obtained at a faster speed.
[0152] Table 4. Time cost analysis using m = 25 and 100 iterations
[0153]
[0154] Table 5. Time cost analysis of 50 trainings
[0155]
[0156] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and the present invention can be implemented in other specific forms without departing from the spirit or basic features of the present invention. Therefore, no matter from which point of view, the embodiments should be regarded as exemplary and non-restrictive, and the scope of the present invention is limited by the attached claims rather than the above description, so it is intended to include all changes within the meaning and scope of the equivalent elements of the claims in the present invention. Although the principle of the present invention is described in detail above in conjunction with the preferred embodiments of the present invention, it should be understood by those skilled in the art that the above embodiments are only explanations of the exemplary implementation of the present invention, and are not limitations on the scope of the present invention. The details in the embodiments do not constitute limitations on the scope of the present invention. Without departing from the spirit and scope of the present invention, any obvious changes such as equivalent transformations, simple substitutions, etc. based on the technical solution of the present invention fall within the scope of protection of the present invention.
Claims
1. An intelligent broadband spectrum sensing method based on a pre-trained basic model, characterized in that: It is characterized by comprising the following steps: Step 1, obtaining sampling data by using a modulation broadband converter; Step 2: preprocessing the sampled data; Step 2.1, using the sampled data and the observation matrix to obtain an estimated original signal with noise and aliased signals superimposed; Step 2.2, use principal component analysis to perform denoising and dimensionality reduction on the original signal; Step 2.3, performing discrete Fourier transform on the denoised and dimension-reduced data to obtain a spectrum estimation signal; Step 2.4, separate the real and imaginary parts of the spectrum estimation signal as two signal features and stack them in the third dimension to become the input of the network; Step 3: Input the data obtained after the preprocessing operation in step 2 into the main network of the time-frequency fusion adjustable deep convolutional neural network (TFF_aDCNN) model, and input the unique hot encoding containing time information into the auxiliary network of the time-frequency fusion adjustable deep convolutional neural network (TFF_aDCNN) model to obtain the support set perception result; The adjustable deep convolutional neural network (TFF_aDCNN) model of time-frequency fusion consists of two parts: the main network and the auxiliary network. The main network consists of a convolutional part and a fully connected part. The convolutional part includes four convolutional units, each of which consists of a convolutional layer with a convolution step of 1, a batch normalization layer, and an activation layer using the ReLU function. The fully connected part consists of a fully connected network with one layer of neurons L, where L represents the number of spectrum slices, and uses the Sigmoid activation function to limit the output range to (0,1) to represent the frequency band state. The auxiliary network consists of a fully connected part and a convolutional part. The fully connected part consists of a fully connected layer with L×(d-20) neurons, where L represents the number of spectrum slices, d represents the number of samples after undersampling of each channel, and channel represents the sampling channel of the MWC algorithm. The convolutional part consists of two convolutional units, each of which consists of a convolutional layer with a convolution kernel of 1x21, 16 channels, and same padding, a batch normalization layer, and an activation layer using the ReLU activation function.
2. The method according to claim 1, characterized in that The method described in step 2 also includes: the TFF_aDCNN network is a dual-input neural network, and the neural network as a whole is divided into: a main network and an auxiliary network. The input of the main network is preprocessed data, and the input of the auxiliary network is a one-hot encoding containing time information. The input of the auxiliary network is a one-hot encoding, which is upgraded through a fully connected layer and then passed through a convolution module to obtain the output of the auxiliary network. The input data of the main network extracts certain signal features, and then is fused with the output of the auxiliary network in the channel direction. After fusion, it passes through a convolution unit with a convolution kernel size of 1×21 and two convolution units with a convolution kernel size of 1×31 to continue to extract features, and finally output through a fully connected layer.
