Small sample wireless signal semi-supervised intelligent accurate identification method

Abstract

The invention discloses a semi-supervised intelligent accurate identification method for small sample wireless signals. The method comprises the following steps: collecting wireless signal data; according to the characteristics of communication signals and different task types, constructing corresponding category mark vectors; designing and constructing a general identification model suitable for communication signals; carrying out mixed training by using limited marked data and combining unmarked data; judging whether the network training is finished or not, if so, executing the next step, and if not, adding one to the number of training iterations and continuing to train the network; inputting the test set data into the network; and outputting a classification result. According to the method, a general identification framework is constructed for communication signal features, and large-scale training under a small number of marked samples is realized in combination with a large number of unmarked samples in a communication environment; the recognition accuracy under the small sample condition is obviously improved; the over-fitting risk of the network is reduced; and the end-to-end training speed of the model is improved.

Description

Technical Field

[0001] The present invention relates to the field of communication technology, and in particular to a method for semi-supervised intelligent precise recognition of small sample wireless signals. Background Art

[0002] Wireless signal recognition is a key technology in wireless intelligent communication. It recognizes different modulation types and access methods by learning the unique characteristics of the signal. With the deepening of 6G research, intelligent communication technology, as an important part of the 6G network, urgently needs further research. Among them, wireless signal recognition technologies such as modulation mode recognition and access mode recognition are particularly important. Wireless signal recognition has been widely used in military and civilian fields. In the military, it helps to obtain enemy information and generate interference; in civilian use, it can ensure the correct selection of demodulation methods. Existing wireless signal recognition methods can be divided into two categories: model-driven and data-driven. Model-driven methods, such as those based on likelihood functions, judge signal types by comparing statistics, but the calculation is complex and the recognition effect is general, especially at low signal-to-noise ratios. Data-driven methods, such as deep learning, extract features from big data through learning, and have good classification performance, but require a large number of labeled samples. In general, as the communication environment becomes increasingly complex, data-driven methods are suitable for handling dynamic and complex scenarios due to their strong learning ability and high classification accuracy, and are the development trend of wireless signal recognition.

[0003] Data-driven wireless signal recognition methods have attracted widespread attention due to their powerful feature learning capabilities. Tim O'Shea et al. proposed a residual neural network in their paper "Over the Air Deep Learning Based Radio Signal Classification" (IEEE J. Sel. Topics Signal Process., 2018). Through the residual learning structure, the network can alleviate the overfitting problem when training deep networks. However, the model is heavily dependent on a large amount of labeled data and requires a lot of computing resources. Lu Yuan et al. proposed a multiscale network-based intelligent access method recognition network model in their paper "A multiscale CNN framework for wireless technique classification in Internet of Things" (IEEE Internet of Things Journal 9 (12), 10366-10367). By capturing higher abstract features through multi-scale modules, better classification performance and higher convergence speed can be achieved. The patent application filed by the University of Electronic Science and Technology of China, "An Improved Deep Modulation Recognition Network Model" (application number CN201910650582.0, application publication number CN110503185B), discloses an improved deep modulation recognition network model. By horizontally stacking multiple ResNet module structures, the improved residual model ResNet is applied to deep modulation recognition, which can greatly improve the recognition performance under the same training complexity.

[0004] The above data-driven wireless signal recognition methods require a large number of labeled samples for training, but it is difficult to obtain sufficient labeled data in actual complex communication systems. In addition, using a small number of labeled samples to train complex neural network models is prone to overfitting risks. Therefore, it is urgent to develop wireless signal intelligent recognition methods under small sample conditions.

[0005] The invention with publication number CN116521863A discloses a label-noise-resistant text classification method based on semi-supervised learning, which processes a small number of labels in combination with a hybrid matching algorithm. However, due to the particularity of communication signals, signal recognition under small sample conditions is easily affected by environmental noise, making signal recognition more difficult. Therefore, it is urgent to build a universal signal recognition model to extract easily distinguishable signal features; at the same time, a large number of unlabeled samples in the wireless environment are used as an auxiliary to solve the overfitting problem in the case of small samples. Summary of the invention

[0006] The purpose of the present invention is to propose a semi-supervised intelligent and precise recognition method for small sample wireless signals. According to the characteristics of wireless communication signals, the soft threshold function is modified, and some useful wireless signal features can be effectively retained during denoising. On this basis, semi-supervised learning is performed through a small number of labeled samples, which significantly improves the recognition accuracy of communication signals in a small sample environment, reduces the risk of overfitting, and reduces the dependence on a large number of labeled samples. It is particularly suitable for communication environments with scarce samples.

[0007] In order to achieve the above technical objectives, the technical solution adopted by the present invention is:

[0008] A semi-supervised intelligent and accurate recognition method for a small sample wireless signal, the intelligent and accurate recognition method comprising the following steps:

[0009] S1, collect wireless signal data and regard the wireless signal identification problem as a type C hypothesis testing problem;

[0010] S2, according to the characteristics of communication signals and different task types, uses coding to represent the signal category, and constructs different category label vectors for the two tasks of modulation mode identification and access mode identification;

[0011] S3, based on the deep residual network, uses the modified soft threshold function as a shrinkage module added to the residual network, so that when the absolute value of the signal feature is less than the threshold, some useful wireless signal features are retained, and a general recognition model for communication signals is designed and constructed;

[0012] S4, semi-supervised hybrid training using limited labeled data combined with unlabeled data;

[0013] S5, determining whether the current training period has reached the maximum number of training periods, if so, executing step S6, if not, increasing the number of training iterations by one and returning to step S4 to continue training the general recognition model;

[0014] S6, input the test set data into the general recognition model and output the classification result.

[0015] In step S1, regarding the wireless signal identification problem as a type C hypothesis testing problem means:

[0016] The received signal under the c-th modulation assumption is expressed as x c (n) = s c (n)+ω c (n), where s c (n) represents the nth sampling point of the transmitted signal of the cth modulation mode, x c (n) represents the nth sampling point of the received signal of the cth modulation mode,

[0017] ωc (n) means the mean is 0 and the variance is σ 2 The received signal is represented as a vector consisting of I / Q components, x c =I c +Q c , where x c Represents x c (n) in vector form, I c and Q c Represent the in-phase and quadrature components of the signal respectively; make a decision in the Class C modulation method based on the I / Q signal samples.

[0018] Step S2 further comprises:

[0019] For modulation mode identification, AM-SSB, CPFSK, QPSK, GFSK, PAM4, QAM16, WBFM, 8PSK, QAM64, AM-DSB, BPSK are selected as identification objects; for access mode identification, LTE, Wi-Fi and DVB-T are selected as identification objects;

[0020] Different category label vectors are constructed for the two tasks of modulation mode identification and access mode identification respectively; N categories are represented by N-bit binary numbers, each category is represented by one bit, the corresponding bit is 1, and the rest are 0; for modulation mode identification, 11-bit coding is used; for access mode identification, 3-bit coding is used.

[0021] In step S3, the general recognition model includes three residual stacking modules, a fully connected layer and a classifier;

[0022] The residual stacking module includes a convolutional layer with a 1×1 kernel, two residual shrinkage units and a maximum pooling layer connected in sequence, the convolutional layer with a 1×1 kernel is used to change the number of channels and aggregate information from a previous layer, the two residual shrinkage units process the aggregated information from the previous layer to extract features, and the maximum pooling layer is used to reduce the width of the extracted feature map;

[0023] The residual shrinkage unit includes a correction shrinkage module, which combines the correction threshold operation at the channel level to perform an absolute value operation on the input feature map x followed by a global average pooling, so that the feature map x is flattened into a one-dimensional vector. The generated one-dimensional vector is passed to a fully connected layer, which contains multiple neurons, and the number is equal to the number of channels of the feature map x.

[0024] In step S3, the modified shrinkage module uses the modified soft threshold function to perform denoising on the feature; the modified soft threshold function is:

[0025]

[0026]

[0027] In the formula, τ is the threshold and x represents the input feature.

[0028] Step S4 further comprises:

[0029] Assume there is a batch of labeled examples x and an unlabeled batch u of the same size, use the hybrid matching algorithm to generate a batch of processed enhanced labeled examples and a batch of augmented unlabeled examples with guessed labels and They are used to calculate the loss terms for labeling and unlabeling respectively:

[0030]

[0031] Among them, T, K, α are hyper parameters;

[0032] Perform data augmentation on both labeled and unlabeled samples: For each labeled sample x b ∈x, by applying a random augmentation function Generate augmented data Generate K unlabeled augmented samples through the amplification function For each unlabeled sample u in u, generate a pseudo label for it; pseudo label generation includes:

[0033] Calculate the unlabeled sample u b The average class distribution predicted by all K augmented models is:

[0034]

[0035] In the formula, p model is the distribution predicted by the model, θ is the model parameter, and y is the predicted result;

[0036] By introducing the mark consistency operation, we can calculate:

[0037]

[0038] in is the guessed pseudo-label, p(y) is the true label, The tag predicted by the model;

[0039] A sharpening function is applied to reduce the distribution of labels under the entropy average prediction. The sharpening function is:

[0040]

[0041] Where p is the input classification distribution, T is the output of a hyperparameter, L is the number of samples, and i and j are the sample numbers. When the temperature approaches zero, Sharpen(p,T) will approach the one-hot encoding distribution. For the augmented data u b Make predictions;

[0042] Hybrid Algorithm For a pair of two examples with corresponding labeled distributions (x1, p1), (x2, p2), (x', p') is calculated as follows:

[0043]

[0044] Where λ is the value drawn from the Beta distribution, α is the corresponding hyperparameter, max is the maximum value function, (x1, p1), (x2, p2) are both sample and label pairs, and (x', p') is the mixed sample and label pair;

[0045] Represent all the labeled samples and all the unlabeled samples with guessed labels in all the enhanced labeled samples as a set:

[0046]

[0047]

[0048] Mixed set generation As a data source for hybrid matching algorithms; calculate hybrid algorithms and add the result to pass Will become The loss function is expressed as:

[0049]

[0050]

[0051]

[0052] Among them, H is the cross entropy loss function, and Represent the loss functions of labeled samples and unlabeled samples respectively, is the weight of the unlabeled sample, x, u and y are the labeled sample, unlabeled sample and the corresponding label, p and q are the probability distribution of the real label and the pseudo label, p model is the prediction model, and θ is the model parameter.

[0053] Compared with the prior art, the present invention has the following beneficial effects:

[0054] First, the small sample wireless signal semi-supervised intelligent and precise recognition method of the present invention introduces a modular semi-supervised signal recognition method, which combines small sample labeled samples with a large number of unlabeled samples for mixed training through wireless signal data amplification, pseudo-label guessing and mixing technologies. Compared with the traditional deep learning signal recognition framework, it reduces the dependence on the amount of labeled training samples.

[0055] Second, the semi-supervised intelligent and precise recognition method for small sample wireless signals of the present invention designs a universal signal recognition model with lower model complexity and fewer model computable parameters. While ensuring performance, it improves the model training speed and provides a guarantee for the real-time requirements in actual communication scenarios.

[0056] Third, the small sample wireless signal semi-supervised intelligent and precise recognition method of the present invention, the network structure of the designed general signal recognition model can learn discriminative deep signal features, help the model reduce the impact of environmental noise on signal recognition accuracy, and significantly improve the signal recognition accuracy under low signal-to-noise ratio conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 It is a flow chart of the semi-supervised intelligent precise recognition method of small sample wireless signals of the present invention;

[0058] Figure 2 is a framework diagram of a general wireless signal recognition model of the present invention, wherein (a) is the overall framework of the proposed model, (b) is a residual stacking module, (c) is a residual shrinkage unit, and (d) is a shrinkage module;

[0059] Figure 3 It is a semi-supervised learning framework diagram of the present invention;

[0060] Figure 4 It is a comparison chart of modulation mode recognition and classification accuracy under different signal-to-noise ratio conditions using the present invention and other existing technologies;

[0061] Figure 5 This is a comparison chart of access mode identification and classification accuracy under different signal-to-noise ratio conditions using the present invention and other existing technologies;

[0062] Figure 6 This is a comparison chart of the training convergence speed using the present invention and other existing technologies;

[0063] Figure 7 This is a comparison diagram of confusion matrices using the present invention and other existing technologies. DETAILED DESCRIPTION

[0064] The embodiments of the present invention are further described in detail below in conjunction with the accompanying drawings.

[0065] The present invention discloses a semi-supervised intelligent and accurate recognition method for a small sample wireless signal, the intelligent and accurate recognition method comprising the following steps:

[0066] S1, collect wireless signal data and regard the wireless signal identification problem as a type C hypothesis testing problem;

[0067] S2, according to the characteristics of communication signals and different task types, uses coding to represent the signal category, and constructs different category label vectors for the two tasks of modulation mode identification and access mode identification;

[0068] S3, based on the deep residual network, uses the modified soft threshold function as a shrinkage module added to the residual network, so that when the absolute value of the signal feature is less than the threshold, some useful wireless signal features are retained, and a general recognition model for communication signals is designed and constructed;

[0069] S4, semi-supervised hybrid training using limited labeled data combined with unlabeled data;

[0070] S5, determining whether the current training period has reached the maximum number of training periods, if so, executing step S6, if not, increasing the number of training iterations by one and returning to step S4 to continue training the general recognition model;

[0071] S6, input the test set data into the general recognition model and output the classification result.

[0072] Combined with Figure 1 The specific steps of the method of the present invention are described as follows.

[0073] Step 1: Wireless signal data collection.

[0074] The wireless signal recognition problem can be regarded as a multi-class hypothesis testing problem. Under the assumption of the kth signal class, the nth sampling point of the received signal can be expressed as x k (n) = s k (n)+ω k (n).

[0075] The wireless signal recognition problem can be regarded as a C-type hypothesis testing problem. The received signal under the c-th modulation hypothesis can be expressed as x c (n) = s c (n)+ω c (n). Among them, s c (n) represents the nth sampling point of the transmitted signal of the cth modulation mode, x c (n) represents the nth sampling point of the received signal of the cth modulation mode, ω c (n) means the mean is 0 and the variance is σ 2The received signal is represented as a vector consisting of I / Q components, x c =I c +Q c , where x c Represents x c (n) in vector form, I c and Q c Represents the in-phase and quadrature components of the signal respectively. Based on the I / Q signal samples, a decision is made in the Class C modulation scheme.

[0076] Step 2: Construct the corresponding category label vector according to the characteristics of the communication signal and different task types.

[0077] In the present invention, AM-SSB, CPFSK, QPSK, GFSK, PAM4, QAM16, WBFM, 8PSK, QAM64, AM-DSB, BPSK and other eleven modulation modes are used as identification objects for modulation mode identification in signal identification problems; LTE, Wi-Fi and DVB-T are used as identification objects for access mode identification. Therefore, different category label vectors are constructed for the two tasks of modulation mode identification and access mode identification respectively. The category of the signal is represented by coding, that is, N-bit binary numbers are used to represent N categories, each category is represented by one bit, the corresponding bit is 1, and the rest are 0. For modulation mode identification, 11-bit coding is used. When the category is AM-SSB, the first bit of the unique hot coding is 1, and the rest are 0, and so on; for access mode identification, 3-bit coding is used. When the category is LTE, the first bit of the coding is 1, and the rest are 0, and so on.

[0078] Step 3, design and construct a general recognition model suitable for communication signals;

[0079] The universal wireless signal recognition model constructed by the present invention is obtained based on the deep residual network. The model consists of three residual stacking modules, a fully connected layer and a classifier. The residual stacking module starts with a convolutional layer with a 1×1 kernel, and its main function is to change the number of channels and aggregate information from the previous layer. Then, two residual shrinkage units are used together to extract features. Finally, the maximum pooling operation is used to reduce the width of the feature map to reduce the amount of calculation of the subsequent layer. The residual unit consists of two units including a convolutional layer, a normalization layer, and a linear correction unit as an activation function and a residual connection. The present invention replaces the last convolutional layer of the residual unit with a correction shrinkage module to form a correction shrinkage unit. The correction shrinkage unit combines the correction threshold operation at the channel level, inputs the feature map x, and is first flattened into a one-dimensional vector by absolute value operation followed by global average pooling. Then this vector is passed into a two-layer fully connected network. The second fully connected layer contains multiple neurons, the number of which is equal to the number of channels in x.

[0080] Soft thresholding is widely used in the field of traditional signal denoising. It sets the coefficients less than the threshold to 0 (filtering noise) to achieve the denoising effect. The formula is as follows, where x represents the input feature and τ is the threshold:

[0081]

[0082]

[0083] Adding soft threshold as a shrinkage module to the residual network can achieve relatively good signal recognition. However, directly returning the features below the threshold to zero is not effective for wireless signal recognition, especially for signals with low signal-to-noise ratio. The soft threshold operation may cause the loss of some subtle features in the wireless signal, which may be important for classification and recognition, resulting in a decrease in the accuracy of wireless signal recognition. Therefore, the soft threshold operation needs to be modified to take into account the special features of wireless signals.

[0084] In order to solve this problem, the present invention proposes a modified threshold operation. Assuming that the threshold is τ, the modified threshold function is still

[0085] An odd function can be defined as y = f(x):

[0086]

[0087]

[0088] The modified threshold function does not completely return the feature to zero when the absolute value is less than the threshold, and still retains some useful wireless signal features. Although the noise information is greatly suppressed, it is not completely eliminated, which also increases the richness of the wireless signal to a certain extent and prevents overfitting. The derivative is always between 0 and 1, which can prevent the gradient vanishing and explosion problems during training.

[0089] When training the general wireless signal recognition model, first, the random initialization method is used to initialize the model weights, the initialization network training epoch number is equal to 1, the maximum epoch number is 50, the learning rate is 0.001, and the stochastic gradient descent (SGD) optimization algorithm is used as the network training optimizer. Therefore, the cross entropy loss function is selected to calculate the gap between the network output and the true category label.

[0090] Step 4: Perform semi-supervised hybrid training using a small amount of labeled data.

[0091] Suppose there is a batch of labeled examples and an unlabeled batch of the same size Generate a batch of processed enhanced labeled examples using a hybrid matching algorithm and a batch of augmented unlabeled examples with “guessed” labels Then, and Used to calculate the loss terms for labeled and unlabeled. In general, the comprehensive loss L for semi-supervised learning is defined as

[0092]

[0093] Among them, T, K, α are hyperparameters. The algorithm attached below shows the implementation process of the hybrid matching algorithm. The algorithm includes data augmentation, pseudo label guessing, and mixing.

[0094]

[0095]

[0096] Signal amplification: Data augmentation is used for both labeled and unlabeled samples. For each labeled sample By applying a random augmentation function (Line 2 of the algorithm) Generate augmented data Similarly, K unlabeled augmented samples are generated by the amplification function

[0097] Pseudo-label guessing: For For each unlabeled sample u in , the hybrid matching algorithm uses the prediction function of the model to generate a pseudo label for it. bThe average of the class distributions predicted by all K augmented models is calculated as follows (line 6 of the algorithm), p model is the distribution predicted by the model, θ is the model parameter, and y is the predicted result:

[0098]

[0099] In order to solve the label inconsistency problem in the pseudo-label guessing process caused by the particularity of the communication signal, a label consistency operation is introduced, in which is the guessed pseudo-label, p(y) is the true label, The tags predicted by the model:

[0100]

[0101] A sharpening function is applied to reduce the distribution of labels under the entropy mean prediction. In practice, the “temperature” distribution of the classification is modified, called a sharpening operation, defined as:

[0102]

[0103] where p is the input categorical distribution (in the mixture matching algorithm, p is the average class prediction for the boost (line 7 of the algorithm)), and T is an output hyperparameter. As the temperature approaches zero, Sharpen(p,T) approaches a Dirac (“one-hot encoding”) distribution. is the model for the augmented data u b Therefore, lowering the temperature helps the model generate low entropy predictions.

[0104] Mixing: A data augmentation technique for achieving semi-supervised learning. For a pair of two examples with corresponding labeled distributions (x1, p1), (x2, p2), (x', p') is calculated as follows:

[0105] λ~Beta(α,α)

[0106] λ′=max(λ,,1-α)

[0107] x′=λx1+(1-λ′)x2

[0108] p′=λp1+(1-λ′)p2

[0109] First, λ is drawn from the Beta distribution, where α is a hyperparameter, and the hybrid algorithm sets λ = λ'. Labeled and unlabeled training instances are mixed in the same mini-batch, and the batch order for calculating a single loss term should be properly preserved, which ensures that x is closer to x1 than to x2. To implement the hybrid algorithm, all augmented labeled samples are labeled and all unlabeled samples with guessed labels are represented as a set:

[0110]

[0111]

[0112] These sets are then combined, mixed and generated As the data source of the hybrid algorithm (line 11 of the algorithm). calculate and add the result to (Line 12 of the algorithm). is used The remainder of The hybrid matching algorithm will Transfer to , which is a set of labeled samples that have been amplified and mixed. become This is a collection of multi-augmented unlabeled samples with their corresponding prediction labels. L2 loss is often used for unlabeled data to measure the uncertainty of the prediction due to its finiteness and reduced sensitivity to wrong predictions. The loss function can be expressed as:

[0113]

[0114]

[0115]

[0116] Among them, H is the cross entropy loss function, and Represent the loss functions of labeled samples and unlabeled samples respectively, is the weight of the unlabeled samples.

[0117] For the above hybrid matching algorithm, the corresponding hyperparameters are set to T = 0.5, K = 2, α = 0.75, λ = 75.

[0118] Step 5: Determine whether the network training is complete.

[0119] Determine whether the current training period has reached the maximum number of training periods. If so, proceed to step 6. If not, continue to train the network in step 4.

[0120] Step 6: Input the test set data into the network.

[0121] Step 7: Output the classification results.

[0122] The effect of the present invention is further described below in conjunction with simulation experiments.

[0123] 1. Simulation conditions and parameter settings:

[0124] The simulation experiment of the present invention is carried out on a container simulation platform with operating systems of Ubuntu 18.04, Python 3.11, and Pytorch 1.13.1. The computer CPU model is i7, and it is equipped with an independent graphics card of NVIDIA Geforce RTX 3090. The data set uses the public data sets RadioML 2016.10A and RadioML 2016.04C for modulation mode identification and TechRec for access mode identification. The data set contains signal samples under different signal-to-noise ratio conditions from -12dB to 12dB, and each signal has 128 sample points. The present invention selects four modulation modes: binary phase shift keying, quaternary phase shift keying, hexadecimal quadrature amplitude modulation, and hexadecimal quadrature amplitude modulation. The signal-to-noise ratio range is -12dB to 12dB, and the step size is 2dB.

[0125] 2. Simulation results:

[0126] Attached Figure 4 The figure is a comparison chart of the modulation mode recognition and classification accuracy under different signal-to-noise ratio conditions using the present invention and the prior art. Figure 4 The horizontal axis represents different signal-to-noise ratios (dB), and the vertical axis represents the recognition accuracy. The broken line marked with a five-pointed star represents the classification accuracy curve of the method of the present invention, and the dotted line marked with a large dot, the dotted line marked with a small dot, the dotted line without a mark, the solid line marked with an oblique cross, and the dotted line marked with an oblique cross represent the classification accuracy curves of the multi-scale neural network, the residual neural network, the visual group network (VGG), the convolutional neural network, and the long short-term memory network under different signal-to-noise ratios. By comparison, it can be seen that the classification accuracy of the present invention is significantly higher than other existing methods. When the signal-to-noise ratio is greater than -1dB, the accuracy of the method of the present invention reaches more than 90%, which is 1dB more sensitive than the multi-scale neural network model. In addition, when the signal-to-noise ratio reaches 2dB, the accuracy of the method of the present invention exceeds 95%, which is about 3% higher than the residual neural network and 12% higher than VGG.

[0127] Attached Figure 5 This is a comparison chart of access mode identification and classification accuracy under different signal-to-noise ratio conditions using the present invention and the prior art. Figure 5The horizontal axis represents different signal-to-noise ratios (dB), and the vertical axis represents the recognition accuracy. The broken line marked with a five-pointed star represents the classification accuracy curve of the method of the present invention, and the dotted line marked with a large dot, the dotted line marked with a small dot, the dotted line without a mark, the solid line marked with an oblique cross, and the dotted line marked with an oblique cross represent the classification accuracy curves of the multi-scale neural network, the residual neural network, the visual group network (VGG), the convolutional neural network, and the long short-term memory network under different signal-to-noise ratios. By comparison, it can be seen that the classification accuracy of the present invention is significantly higher than other existing methods. When the signal-to-noise ratio is greater than 0dB, the present invention can achieve a recognition accuracy of more than 80%, which is 1db more sensitive than the advanced multi-scale neural network and residual neural network. In addition, when the signal-to-noise ratio reaches 4%db, the present invention can achieve an accuracy of more than 90%, which is 2db more sensitive than VGG and convolutional neural networks. In terms of overall classification accuracy, the present invention reaches 82.66% in all three categories over a wider range of signal-to-noise ratios, which is 1.5% higher than the multi-scale neural network and residual neural network, and about 5% higher than VGG and convolutional neural networks.

[0128] Attached Figure 6 This is a comparison chart of the modulation mode recognition classification accuracy using the present invention and the prior art under small sample conditions, where (a), (b), and (c) respectively represent the classification accuracy when 10%, 5%, and 3% of labeled samples are used. Figure 6 The horizontal axis represents different signal-to-noise ratios (dB), and the vertical axis represents recognition accuracy. The broken line marked with a five-pointed star represents the classification accuracy curve of the method of the present invention, and the dashed line marked with a large dot, the dashed line marked with a small dot, the dashed line without a mark, the solid line marked with an oblique cross, the dashed line marked with an oblique cross, and the dashed line marked with a cross represent the classification accuracy curves of the capsule network, the convolutional neural network-long short-term memory network, the convolutional neural network, the gated recurrent unit, the long short-term memory network, and the residual neural network under different signal-to-noise ratios. As shown in the attached figure Figure 6 As shown in (a), when a 10% labeled data set is used for training, the classification accuracy of the present invention approaches 95% as the signal-to-noise ratio increases. When a 5% labeled data set is used for training, the classification accuracy of the convolutional neural network fluctuates around 80%, the capsule network can achieve 85% accuracy when the signal-to-noise ratio is greater than 0dB, and the classification accuracy of the present invention is close to 95%. When 3% labeled training data is used, the classification accuracy of most network structures for modulated signals fluctuates between 60% and 70%, while the capsule network can achieve 80% accuracy. In contrast, the classification accuracy of the present invention fluctuates around 92%, and when the signal-to-noise ratio is 12dB, the classification accuracy is close to 95%.

[0129] Attached Figure 73 is a comparison chart of access mode identification and classification accuracy using the present invention and the prior art under small sample conditions, wherein (a), (b), (c) and (d) respectively represent the classification accuracy when 10%, 5%, 3% and 1% of labeled samples are used. Figure 7 The horizontal axis in the figure represents different signal-to-noise ratios (dB), and the vertical axis represents the recognition accuracy. The broken line marked with a five-pointed star represents the classification accuracy curve of the method of the present invention, and the dashed line marked with a large dot, the dashed line marked with a small dot, the dashed line without a mark, and the solid line marked with an oblique cross represent the classification accuracy curves of the multi-scale neural network, the residual neural network, the visual set group network, and the convolutional neural network under different signal-to-noise ratios. As shown in the attached figure Figure 7 As shown in (a), when using 1% of the labeled training data set, the classification accuracy of the present invention is close to 90% as the signal-to-noise ratio increases, while the accuracy of other advanced deep learning-based methods is less than 85%. When trained with 5% of the labeled data set, the classification accuracy of the multi-scale network fluctuates around 83%, and when the signal-to-noise ratio is greater than 4dB, the classification accuracy is close to 85%. In addition, we train the model under very few sample conditions, using only 1% of the labeled data (about 1,286 samples), as shown in Figure 7 As shown in (d), other models suffer from severe overfitting and their performance deteriorates when the signal-to-noise ratio increases; in contrast, when the signal-to-noise ratio is greater than 6 dB, the present invention can achieve a classification accuracy of over 80%.

[0130] Based on the above simulation results and analysis, the semi-supervised intelligent and precise recognition method of small sample wireless signals proposed in the present invention can achieve higher classification accuracy than existing methods by constructing a general signal recognition model combined with a large number of unlabeled samples, especially it can achieve higher recognition performance under small sample conditions.

[0131] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of complete hardware embodiments, complete software embodiments, or embodiments in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code. The scheme in the embodiments of the present application can be implemented in various computer languages, for example, object-oriented programming language Java and literal scripting language JavaScript, etc.

[0132] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0133] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0134] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0135] Although the preferred embodiments of the present application have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.

[0136] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

1. A small sample wireless signal semi-supervised intelligent accurate recognition method, characterized in that: The intelligent accurate identification method comprises the following steps: S1, collect wireless signal data and regard the wireless signal identification problem as a type C hypothesis testing problem; S2, according to the characteristics of communication signals and different task types, uses coding to represent the signal category, and constructs different category label vectors for the two tasks of modulation mode identification and access mode identification; S3, based on the deep residual network, uses the modified soft threshold function as a shrinkage module added to the residual network, so that when the absolute value of the signal feature is less than the threshold, some useful wireless signal features are retained, and a general recognition model for communication signals is designed and constructed; S4, semi-supervised hybrid training using limited labeled data combined with unlabeled data; S5, determining whether the current training period has reached the maximum number of training periods, if so, executing step S6, if not, increasing the number of training iterations by one and returning to step S4 to continue training the general recognition model; S6, input the test set data into the general recognition model and output the classification result.

2. The method for semi-supervised intelligent and accurate identification of small sample wireless signals according to claim 1 is characterized in that: In step S1, regarding the wireless signal identification problem as a type C hypothesis testing problem means: The received signal under the c-th modulation assumption is expressed as x c (n) = s c (n)+ω c (n), where s c (n) represents the nth sampling point of the transmitted signal of the cth modulation mode, x c (n) represents the nth sampling point of the received signal of the cth modulation mode, ω c (n) means the mean is 0 and the variance is σ 2 The received signal is represented as a vector consisting of I / Q components, x c =I c +Q c , where x c Represents x c (n) in vector form, I c and Q c Represent the in-phase and quadrature components of the signal respectively; make a decision in the Class C modulation method based on the I / Q signal samples.

3. The method for semi-supervised intelligent and accurate identification of small sample wireless signals according to claim 1 is characterized in that: Step S2 further comprises: For modulation mode identification, AM-SSB, CPFSK, QPSK, GFSK, PAM4, QAM16, WBFM, 8PSK, QAM64, AM-DSB, BPSK are selected as identification objects; for access mode identification, LTE, Wi-Fi and DVB-T are selected as identification objects; Different category label vectors are constructed for the two tasks of modulation mode identification and access mode identification respectively; N categories are represented by N-bit binary numbers, each category is represented by one bit, the corresponding bit is 1, and the rest are 0; for modulation mode identification, 11-bit coding is used; for access mode identification, 3-bit coding is used.

4. The method for semi-supervised intelligent and accurate identification of small sample wireless signals according to claim 1 is characterized in that: In step S3, the general recognition model includes three residual stacking modules, a fully connected layer and a classifier; The residual stacking module includes a convolutional layer with a 1×1 kernel, two residual shrinkage units and a maximum pooling layer connected in sequence, the convolutional layer with a 1×1 kernel is used to change the number of channels and aggregate information from a previous layer, the two residual shrinkage units process the aggregated information from the previous layer to extract features, and the maximum pooling layer is used to reduce the width of the extracted feature map; The residual shrinkage unit includes a correction shrinkage module, which combines the correction threshold operation at the channel level to perform an absolute value operation on the input feature map x followed by a global average pooling, so that the feature map x is flattened into a one-dimensional vector. The generated one-dimensional vector is passed to a fully connected layer, which contains multiple neurons, and the number is equal to the number of channels of the feature map x.

5. The method for semi-supervised intelligent and accurate identification of small sample wireless signals according to claim 4 is characterized in that: In step S3, the modified shrinkage module uses the modified soft threshold function to perform denoising on the feature; the modified soft threshold function is: In the formula, τ is the threshold and x represents the input feature.

6. The method for semi-supervised intelligent and accurate identification of small sample wireless signals according to claim 4 is characterized in that: Step S4 further comprises: Suppose there is a batch of labeled examples and an unlabeled batch of the same size Generate a batch of processed enhanced labeled examples using a hybrid matching algorithm and a batch of augmented unlabeled examples with guessed labels and They are used to calculate the loss terms for labeling and unlabeling respectively: Among them, T, K, α are hyper parameters; Perform data augmentation on both labeled and unlabeled samples: For each labeled sample By applying a random augmentation function Generate augmented data Generate K unlabeled augmented samples through the amplification function for For each unlabeled sample u in , generate a pseudo label for it; pseudo label generation includes: Calculate the unlabeled sample u b The average class distribution predicted by all K augmented models is: In the formula, p model is the distribution predicted by the model, θ is the model parameter, and y is the predicted result; By introducing the mark consistency operation, we can calculate: in is the guessed pseudo-label, p(y) is the true label, The tag predicted by the model; A sharpening function is applied to reduce the distribution of labels under the entropy average prediction. The sharpening function is: Where p is the input classification distribution, T is the output of a hyperparameter, L is the number of samples, and i and j are the sample numbers. When the temperature approaches zero, Sharpen(p,T) will approach the one-hot encoding distribution. For the augmented data u b Make predictions; Hybrid Algorithm For a pair of two examples with corresponding labeled distributions (x1, p1), (x2, p2), (x', p') is calculated as follows: Where λ is the value drawn from the Beta distribution, α is the corresponding hyperparameter, max is the maximum value function, (x1, p1), (x2, p2) are both sample and label pairs, and (x', p') is the mixed sample and label pair; Represent all the labeled samples and all the unlabeled samples with guessed labels in all the enhanced labeled samples as a set: Mixed set generation As a data source for hybrid matching algorithms; calculate hybrid algorithms and add the result to pass Will become The loss function is expressed as: Among them, H is the cross entropy loss function, and Denote the loss functions of labeled samples and unlabeled samples respectively, λ u is the weight of the unlabeled sample, x, u and y are the labeled sample, unlabeled sample and the corresponding label, p and q are the probability distribution of the real label and the pseudo label, p model is the prediction model, and θ is the model parameter.

Images