Unlicensed band wireless signal recognition method based on deep residual shrinkage network

Abstract

The application discloses a kind of unlicensed frequency band wireless signal identification method based on deep residual shrinkage network, it is related to wireless signal identification technical field, comprising: constructing multiple IQ signals based on time domain characteristic length sequence to generate signal sample set;The IQ signal is derived from the sampling signal of multiple unlicensed frequency band wireless signal categories;IQ signal in signal sample set is used to train the deep residual shrinkage network model including residual shrinkage module, obtain the trained deep residual shrinkage network model;The input IQ signal is identified using the trained deep residual shrinkage network model, obtains unlicensed frequency band wireless signal category.The application can accurately identify wireless signal under low signal-to-noise ratio condition in unlicensed frequency band.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wireless signal recognition, and particularly relates to a non-licensed frequency band wireless signal recognition method based on a deep residual shrinkage network. BACKGROUND

[0002] In order to solve the problem of wireless spectrum resource explosion caused by the increasing communication traffic, it is imminent to find a method to meet the current traffic demand. In addition, the effective utilization rate of the resource frequency band also needs to be improved. There are rich unlicensed frequency band resources in the 5GHz spectrum, and deploying long-term evolution technology (LTE) in the unlicensed frequency band can effectively alleviate the spectrum resource pressure. LTE-U technology deployed in the unlicensed frequency band is initially an extension of LTE technology in the licensed frequency band, and since deploying communication technology in the unlicensed frequency band can reduce costs, it is an effective method to effectively alleviate spectrum pressure and meet the overload wireless communication capacity. However, there are WiFi, Bluetooth and other signals in the unlicensed frequency band, and whether the signals can coexist friendly and fairly is closely related to the effective utilization rate of the spectrum, so it is necessary to find a reasonable spectrum allocation scheme to achieve the highest spectrum efficiency in the signal coexistence network. Implementing a WiFi and LTE-U friendly coexistence network can allocate the required traffic resources to users with high efficiency and provide the best quality of service for users. In this problem, whether the signal categories can be accurately distinguished is the key to fair and efficient spectrum allocation, and only by adopting different spectrum allocation schemes for different signals can the utilization rate of the overall network be improved and the overload wireless communication capacity be met.

[0003] The existing scheme automatically recognizes wireless signals through a deep learning algorithm, so as to accurately select channels / subframes and achieve fair spectrum allocation of the two signals. However, these schemes have low recognition accuracy under low signal-to-noise conditions and cannot accurately recognize different signals. SUMMARY

[0004] The present application aims to overcome the shortcomings of the prior art and proposes a non-licensed frequency band wireless signal recognition method based on a deep residual shrinkage network. The residual shrinkage module in the network is used to perform noise reduction processing on the input data, so as to minimize the influence of noise features and features close to noise on the network recognition algorithm, thereby improving the efficiency and recognition accuracy of the subsequent network in extracting useful information from the signal.

[0005] To achieve the above purpose, the technical scheme of the present application is a non-licensed frequency band wireless signal recognition method based on a deep residual shrinkage network, comprising:

[0006] A non-licensed frequency band wireless signal recognition method based on a deep residual shrinkage network comprises:

[0007] constructing a plurality of IQ signals based on time-domain feature length sequences to generate a signal sample set; the IQ signals are derived from sample signals of a plurality of unlicensed frequency band wireless signal categories;

[0008] training a deep residual shrinkage network model including a residual shrinkage module using the IQ signals in the signal sample set to obtain a trained deep residual shrinkage network model;

[0009] using the trained deep residual shrinkage network model to identify an input IQ signal to obtain an unlicensed frequency band wireless signal category.

[0010] Preferably, the IQ signal is the in-phase and quadrature two-way data extracted from the initial signal using an IQ sampling rate of 20 Msps;

[0011] For the initial signal y(n), the real part and the imaginary part are represented as real(y(n)) and imag(y(n)) respectively, and the two are combined into a 2×N matrix S IQ , as follows:

[0012]

[0013] wherein S IQ is the IQ signal composed of the real part value and the imaginary part value of the signal; after processing, S IQ is used as the input value of the deep residual shrinkage network for training.

[0014] Preferably, the initial signal is a sample signal or a signal including noise and random phase offset after data preprocessing;

[0015] The signal including noise and random phase offset after data preprocessing is represented as follows:

[0016] y(n) = De jΔθ x(n) + w(n), n = 0, 1,..., N-1

[0017] wherein y(n) represents the signal after phase offset processing; D represents the coefficient factor; w(n) represents the additive noise after the channel; N represents the number of sampling points; Δθ is the added phase offset, which is subject to a random distribution of (0, π); x(n) is the received sample signal,

[0018] Preferably, before training the deep residual shrinkage network model including the residual shrinkage module using the IQ signals in the signal sample set, it further includes:

[0019] normalizing the IQ signals in the signal sample set;

[0020] The input IQ signal is identified by using the trained deep residual shrinkage network model, specifically comprising:

[0021] The input IQ signal is normalized, and the normalized input IQ sampling signal is identified by using the trained deep residual shrinkage network model.

[0022] Preferably, the deep residual shrinkage network model comprises a convolutional layer, a plurality of residual shrinkage modules RSBU, a batch normalization layer, a linear activation function PReLU with parameter correction, a global average pooling layer, and a fully connected output layer connected in sequence; the input of the deep residual shrinkage network model is an IQ signal, and feature extraction is performed by the deep residual shrinkage network model; the input IQ signal is first processed by a convolutional layer, and the convolutional layer is used to extract shallow feature information in the IQ signal;

[0023] The data output matrix after the convolutional layer processing is sent as input into a plurality of residual shrinkage modules, and the residual shrinkage modules are used to dynamically update the threshold by an attention mechanism to eliminate noise;

[0024] The batch normalization layer BN is used to ensure the consistent rule of data distribution, avoid the influence of the previous network on the data distribution, ensure that the network extracts the original features of the data, and improve the network training speed;

[0025] The linear activation function PReLU with parameter correction can adaptively learn the parameters of the linear unit to reduce the influence of the positive and negative of the signal feature value distribution on the model;

[0026] The fully connected layer output layer containing two neurons at the end of the network predicts the results; the Softmax function is used to map the input sample to the interval (0, 1), and if a certain wireless signal has a higher probability, the signal category output by the deep residual shrinkage network model is the wireless signal.

[0027] Preferably, each residual shrinkage module comprises two batch normalization layers BN, two linear activation functions PReLU with parameter correction, two convolutional layers Conv, one identity mapping, and one threshold learning subnetwork connected in sequence.

[0028] The threshold learning subnetwork is used for adaptively setting the threshold, in which one path takes the absolute value of all features of the input feature map, and then the features are averaged through global average pooling GAP to obtain a feature, denoted as A; in another path, the feature map after global average pooling is input into a fully connected network FC, the fully connected network takes the Sigmoid function as the last layer, and the output is normalized to between 0 and 1 to obtain a coefficient, denoted as alpha, and finally the threshold alpha A is obtained.

[0029] From the above description of the present application, compared with the prior art, the present application has the following beneficial effects:

[0030] (1) A non-licensed frequency band wireless signal recognition method based on a deep residual shrinkage network, which uses a residual shrinkage module in the network to perform noise reduction processing on input data, minimizes the influence of noise features and features close to noise on the network recognition algorithm, and thereby improves the efficiency and recognition accuracy of the subsequent network for extracting useful information from the signal;

[0031] (2) The PReLU linear activation function with parameter correction in the residual shrinkage module of the present application can adaptively learn and correct the parameters of the linear unit, and can improve the accuracy with negligible additional computational cost, and through the characteristics of the PReLU adaptive learning parameters, the negative features of the signal are retained as much as possible, and the generalization ability of the model is improved;

[0032] (3) The present application introduces random phase offset to preprocess the input sampling data, and uses the preprocessed data as the training data of the deep residual shrinkage network model, and through the data set obtained by adding random phase offset to the signal, the signal can be recognized in a communication environment closer to the actual signal transmission, thereby improving the generalization ability of the trained deep residual shrinkage network model;

[0033] (4) After adding random phase offset to the received sampling signal, the present application uses IQ sampling with a sampling rate of 20Msps to extract the In-phase and Quadrature data of the initial signal as the input value of the deep residual shrinkage network model, which can ensure that more feature information of the signal can be extracted by the deep residual shrinkage network without increasing the data amount too much.

[0034] The present application will be further described in detail below in combination with the drawings and embodiments, but the non-licensed frequency band wireless signal recognition method based on a deep residual shrinkage network of the present application is not limited to the embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 The flowchart of the non-licensed frequency band wireless signal recognition method based on a deep residual shrinkage network of the present application is shown in the figure.

[0036] Figure 2 A structural schematic diagram of a deep residual shrinkage network model of an embodiment of the present application;

[0037] Figure 3 A structural schematic diagram of a residual shrinkage module of an embodiment of the present application;

[0038] Figure 4 A schematic diagram of wireless signal recognition accuracy of different network models of an embodiment of the present application under a signal-to-noise ratio;

[0039] Figure 5 A schematic diagram of a deep residual network recognition model accuracy curve of an embodiment of the present application with random phase offset. DETAILED DESCRIPTION

[0040] The technical solutions in the embodiments of the present application will be described and discussed in detail below with reference to the accompanying drawings of the present application. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0041] Referring to Figure 1 The present application is a non-licensed frequency band wireless signal recognition method based on a deep residual shrinkage network, which includes the following steps.

[0042] S101, constructing a plurality of IQ signals based on a time domain feature length sequence to generate a signal sample set; the IQ signals are derived from a plurality of non-licensed frequency band wireless signal categories of sampled signals;

[0043] S102, using the IQ signals in the signal sample set to train a deep residual shrinkage network model including a residual shrinkage module, to obtain a trained deep residual shrinkage network model;

[0044] S103, using the trained deep residual shrinkage network model to recognize the input IQ signals, to obtain the non-licensed frequency band wireless signal category.

[0045] In this embodiment, the data set for training and testing verification comes from WiFi and LTE-U signals collected by radio equipment in a laboratory environment by the international organization EWINE PROJECT. In order to ensure the universality of the experimental results, two different data forms and the introduction of random phase offset are used to preprocess the input data, as follows.

[0046] (1) Introducing random phase offset

[0047] Since the signal data used is idealized data collected in the laboratory, there is a certain difference from the real communication environment, and the influence of phase on the actual signal transmission process is not considered. Therefore, the initial signal sample is not close enough to the wireless signal sample in the actual communication environment, which reduces the generalization ability of the trained signal recognition model, and easily leads to overfitting phenomenon, and the result does not have universality and generality. Therefore, in order to simulate the actual communication environment, random phase offset is added to the received signal data, and the generalization ability of the recognition model is improved by training different data sets. Define x(n) as the received original signal, which satisfies the following formula:

[0048]

[0049] Define y(n) as the signal after phase offset processing, as shown in the following formula:

[0050] y(n)=De jΔθ x(n)+w(n),n=0,1,...N-1

[0051] Where D represents the coefficient factor, w(n) represents the additive noise through the channel, and N represents the number of sampling points. Δθ is the added phase offset, which obeys the random distribution of (0, π). The data set obtained by adding random phase offset to the signal can identify the signal in a communication environment closer to the actual signal transmission, thereby improving the generalization ability of the trained model.

[0052] (2) Use in-phase quadrature or amplitude phase data

[0053] After adding random phase offset to the received original signal by the method in (1), considering that IQ sampling can ensure that the signal has more feature information extracted by the deep neural network without increasing the data amount too much, the in-phase (In-phase) and quadrature (Quadrature) two-way data of the initial signal are extracted by IQ sampling with a sampling rate of 20Msps as the input value of the deep neural network.

[0054] For the signal y(n), the real part and the imaginary part are represented as real(y(n)) and imag(y(n)) respectively, and the two are combined into a 2xN matrix S IQ , which is represented as follows:

[0055]

[0056] S IQ is a real matrix with a dimension of 2xN, which is composed of the real part value and the imaginary part value of the signal. After processing, S IQThe input value of the deep residual shrinkage network can be trained as a deep residual shrinkage network, and sent to a shallow convolutional neural network to extract signal features.

[0057] The deep residual shrinkage network (DRSN) is a variant of the residual network (ResNet) based on soft thresholding to eliminate noise-related features. Under the premise of ensuring high accuracy, the ability of ResNet to extract useful features from high-noise vibration signals is improved.

[0058] Referring to Figure 2 The network model shown is built to solve the identification of WiFi and LTE-U signals under low signal-to-noise ratio conditions in complex heterogeneous networks. A complete deep residual shrinkage network includes a convolutional layer, several RSBU residual shrinkage modules, a batch normalization layer, a parametric rectified linear unit (PReLU) with parameter correction, a global average pooling layer, and a fully connected output layer. The input of the deep residual shrinkage network is a numerical matrix of size (2, 1024) after IQ sampling and preprocessing, which contains the feature information of the input signal and needs to be sent to the deep residual shrinkage network for feature extraction. The input signal is first processed by a convolutional layer, which extracts shallow feature information from the IQ signal.

[0059] The data output matrix after convolutional layer processing will be sent as input to several deep residual shrinkage modules. Within each basic deep residual shrinkage module, batch normalization is used to ensure consistent data distribution and avoid the impact of previous neural networks on data distribution, ensuring that the network extracts the original features of the data and improves the network training speed. In addition, to increase the nonlinearity of the model and reduce the influence of positive and negative signal feature value distribution on the model, PReLU is used instead of rectified linear unit (ReLU). PReLU can adaptively learn the parameters of the rectified linear unit and improve the accuracy with negligible additional computational cost. By using the adaptive learning parameter feature of PReLU, the negative features of the signal are preserved as much as possible, improving the generalization ability of the model. The calculation is as follows.

[0060]

[0061] The last connection of the deep residual shrinkage network contains a fully connected layer with two neurons to output the prediction result. The input sample is mapped to the interval (0, 1) using the Softmax function, and if the WiFi signal probability is higher, the neural network finally predicts the signal category output as the WiFi signal, and vice versa as the LTE-U signal.

[0062] Further, referring to Figure 3 As shown in the figure, the residual shrinkage module (Residual Shrinkage Building Unit, RSBU) is the basic module of the DRSN. Unlike the artificial setting of the threshold value, the RSBU can dynamically update the threshold value by means of the attention mechanism. Through the attention mechanism, the importance of each feature channel is evaluated, and then each feature channel is assigned a suitable weight according to its importance, combined with the soft threshold update function, to adapt to the change of data distribution. First, the weight coefficient is generated through the attention mechanism, then multiplied by the result compressed by the global average pooling layer to obtain the threshold value of each feature channel, and finally the feature values less than the threshold value in the output result of the convolution layer are set to zero to obtain the denoised feature map. The structure of the residual shrinkage module is as shown in the figure. Figure 1 As shown in the figure, the width of the input feature is 1, the height is w, and the number of channels is c. The composition structure of this module includes 2 batch normalization layers (Batch Normalization Layer, BN), 2 linear activation functions PReLU with parameter correction, 2 convolutional layers (Convolutional Layer, Conv), 1 identity mapping, and 1 threshold learning subnetwork. Figure 3 The role of the threshold learning subnetwork in the figure is to adaptively set the threshold value. The implementation mechanism is as follows: in this subnetwork, first, take the absolute value of all features of the input feature map; then, after global average pooling (Global Average Pooling, GAP), a feature is obtained, denoted as A; in another path, the feature map after global average pooling is input to a small fully connected network (Fully Connected Network, FC). The fully connected network takes the Sigmoid function as the last layer, normalizes the output to between 0 and 1, obtains a coefficient denoted as α, and finally obtains the threshold value α·A. In this way, it is ensured that the threshold value is positive, and the threshold value will not be too large. Moreover, different samples have different threshold values, so to some extent, the above process can be understood as a special soft thresholding, setting them to zero; or, it is noticed that the features related to the current task are retained. After identity mapping, the output value of each layer of the network will be added to the output of the last layer, which reduces the complexity of the network compared to nonlinear mapping.

[0063] The following will be verified by experiments.

[0064] All deep learning algorithm model training, validation and testing in this experiment use PyTorch1.9 to build a deep learning framework, and the programming language is Python3.7. All experiments are completed on a computer with an NVIDIA RTX 2060Ti GPU and 2 Intel Xeon E5 CPUs.

[0065] The algorithm recognition accuracy curve of the signal recognition model proposed in the application under each signal-to-noise ratio dataset is shown in the following figure. Figure 4 In the figure, the CNN (Convolutional Network, CNN) algorithm uses a convolutional neural network to extract features from the signal and then identifies it, and the RNN (Recurrent Network, RNN) algorithm distinguishes different signals based on a recurrent neural network; the SelfAttention algorithm first extracts shallow features through a layer of convolutional neural network and then focuses on the features that have a greater impact on the result through the self-attention mechanism, and then classifies different signals. Compared with other algorithms, the deep residual shrinkage network algorithm model performs well. It can be seen that in the case of low signal-to-noise ratio, that is, SNR is between -20dB and 0dB, the recognition accuracy of the DRSN algorithm is higher than that of the other three neural network algorithms. Especially when SNR is -20dB, the accuracy of the DRSN algorithm is about 57%, while the accuracy of the other three algorithms is only about 51%, and the recognition accuracy of DRSN is improved by 6%. When the signal-to-noise ratio is higher than 0dB, it can be seen that the recognition accuracy is more than 90%, and the performance is relatively excellent. However, compared with the recognition algorithm based on the self-attention mechanism, it is slightly insufficient. The reason is that although the deep residual shrinkage network can perform noise reduction processing on the signal under high noise conditions to improve the feature learning ability of the model, the improvement effect can be ignored in the case of good channel conditions. In addition, the self-attention mechanism pays more attention to the extraction of signal-related information, and can better fuse the signal features, so the recognition accuracy is slightly higher than that based on the deep residual shrinkage network under high signal-to-noise ratio conditions.

[0066] In order to better simulate the wireless communication environment in real life, and at the same time simulate and test the robustness of the DRSN model, this section adds random phase offset (PO) to the WiFi and LTE-U signals used in the simulation experiment to improve the universality of the research results.

[0067] From Figure 5It can be seen that the recognition model based on the deep residual shrinkage network is affected to a certain extent after introducing the random phase offset, and the overall performance is slightly decreased. However, even if there is an unknown phase offset in the input signal data, the deep residual shrinkage network recognition model can accurately distinguish the two signals when the signal-to-noise ratio is high, that is, the channel condition is good or the signal gain is high. This result is consistent with that when the random phase offset is not introduced. When the signal-to-noise ratio decreases, the influence of the random phase offset on the recognition accuracy of the algorithm will also increase to some extent, but even so, it is still higher than the recognition model precision of other network recognition models without adding the phase offset. It can be seen that the influence of introducing the random phase offset on the recognition network is not very high, but even if the random phase offset is added, the recognition model based on the deep residual shrinkage network proposed in the present application is still better than the recognition model based on other algorithms in performance under the condition of low signal-to-noise ratio.

[0068] The present application uses an improved deep residual shrinkage network to propose a model suitable for wireless signal recognition. The soft threshold module in the network can perform noise reduction processing on the signal, better capture useful features in a high-noise environment, improve the feature learning ability of the model, and greatly improve the performance of low signal-to-noise ratio signal recognition, with an increase of about 7% in recognition accuracy.

[0069] The above is only one preferred embodiment in the examples of the present application. However, the present application is not limited to the above-mentioned embodiments, and any equivalent changes and modifications made according to the present application, as long as the resulting functional effects do not exceed the scope of the present application, are within the scope of protection of the present application.

Claims

1.A method for unlicensed band wireless signal recognition based on a deep residual contraction network, characterized in that, The method comprises the following steps: Constructing a plurality of time-domain feature length sequence-based IQ signals to generate a signal sample set; The IQ signals are derived from a plurality of unlicensed frequency band wireless signal categories; Training a deep residual shrinkage network model comprising residual shrinkage modules using the IQ signals in the signal sample set to obtain a trained deep residual shrinkage network model; Using the trained deep residual shrinkage network model to identify an input IQ signal to obtain an unlicensed frequency band wireless signal category; The deep residual shrinkage network model comprises a convolutional layer, a plurality of residual shrinkage modules RSBU, a batch normalization layer, a linear activation function PReLU with parameter correction, a global average pooling layer, and a fully connected output layer connected in sequence; the input of the deep residual shrinkage network model is an IQ signal, and feature extraction is performed through the deep residual shrinkage network model; the input IQ signal is first subjected to a convolutional layer, and the convolutional layer is used to extract shallow feature information in the IQ signal; The data output matrix after the convolutional layer processing is sent as input into a plurality of residual shrinkage modules, and the residual shrinkage modules are used to dynamically update the threshold value through an attention mechanism to eliminate noise; The batch normalization layer BN is used to ensure the consistent regularity of data distribution, avoid the influence of the previous network on the data distribution, ensure that the original features of the data are extracted by the network, and improve the network training speed; The linear activation function PReLU with parameter correction can adaptively learn to correct the parameters of the linear unit, and reduce the influence of the positive and negative signal feature value distribution on the model; The fully connected layer output layer containing two neurons is connected at the end of the network to predict the result; the Softmax function is used to map the input sample to the (0, 1) interval, and if a certain wireless signal has a higher probability, the signal category output by the deep residual shrinkage network model is the wireless signal; Each residual shrinkage module comprises two batch normalization layers BN, two linear activation functions PReLU with parameter correction, two convolutional layers Conv, one identity mapping, and one threshold learning subnetwork connected in sequence; The threshold learning subnetwork is used to adaptively set the threshold value; in the subnetwork, one path takes the absolute value of all features of the input feature map; then the global average pooling GAP is performed to average, obtaining a feature, denoted as A; in the other path, the feature map after the global average pooling is input into a fully connected network FC, and the fully connected network takes the Sigmoid function as the last layer to normalize the output to between 0 and 1, obtaining a coefficient, denoted as α, and finally obtaining the threshold value α·A. 2.The unlicensed band wireless signal recognition method based on deep residual shrinkage network according to claim 1, wherein, The IQ signal is the in-phase and quadrature two-way data of the initial signal extracted by IQ sampling with a sampling rate of 20 Msps; For the initial signal y(n), the real and imaginary parts are denoted as real(y(n)) and imag(y(n)), respectively, and combined into a 2 x N matrix S IQ , as follows: Wherein, S IQ is an IQ signal composed of a real part value and an imaginary part value of the signal; after processing, S IQ is trained as an input value of a deep residual shrinkage network; N represents the number of sampling points. 3.The unlicensed band wireless signal recognition method based on deep residual shrinkage network according to claim 2, wherein, The initial signal is a sampling signal or a signal that has undergone data preprocessing and includes noise and random phase offset; The signal that has undergone data preprocessing and includes noise and random phase offset is represented as follows: y(n) = De jΔθ x(n) + w(n), n = 0, 1,... N - 1 where y(n) represents a signal after phase offset processing; D represents a coefficient factor; w(n) represents additive noise through a channel; Δθ is an added phase offset, which is subject to a random distribution of (0, π); x(n) is a received sampling signal, 4.The unlicensed band wireless signal recognition method based on deep residual shrinkage network according to claim 1, wherein, Before training the deep residual shrinkage network model comprising residual shrinkage modules using the IQ signals in the signal sample set, the following steps are further included: Normalizing the IQ signal in the signal sample set; The input IQ signal is recognized by using the trained deep residual shrinkage network model, and specifically includes: The input IQ signal is normalized, and the normalized input IQ sampling signal is recognized by using the trained deep residual shrinkage network model.

Images