Automatic Modulation Classification Using Deep Residual Neural Network with Masked Modeling for Wireless Communications

Abstract

Automatic modulation classification (AMC) is a signal processing technology used to identify the modulation type of unknown signals without prior information such as modulation parameters for drone communications. In recent years, deep learning (DL) has been widely used in AMC methods due to its powerful feature extraction ability. The significant performance of DL-based AMC methods is highly dependent on large amount of data. However, with the increasingly complex signal environment and the emergence of new signals, several recognition tasks have difficulty obtaining sufficient high-quality signals. To address this problem, we propose an AMC method based on a deep residual neural network with masked modeling (DRMM). Specifically, masked modeling is adopted to improve the performance of a deep neural network with limited signal samples. Both complex-valued and real-valued residual neural networks (ResNet) play an important role in extracting signal features for identification. Several typical experiments are conducted to evaluate our proposed DRMM-based AMC method on the RadioML 2016.10A dataset and a simulated dataset, and comparison experiments with existing AMC methods are also conducted. The simulation results illustrate that our proposed DRMM-based AMC method achieves better performance in the case of limited signal samples with low signal-to-noise ratio (SNR) than other existing methods.

1. Introduction

In today’s era of rapid development of communications technology for drone communications, the huge demand for communication services makes wireless resources (channels, spectrum, etc.) more and more tense. In order to meet the requirements of improving spectrum utilization in a limited channel capacity, alleviating the lack of communication resources, and transmitting information quickly, accurately, and stably, increasingly complex and diverse communication systems and modulation modes have been developed [1,2,3,4]. To obtain the information carried by the communication signal, the first step is to understand the modulation mode of the signal. Modulation recognition technology is currently being widely used in modern radio communications such as spectrum management and military communications [5,6,7,8].

In terms of spectrum management, the research on modulation identification technology is to strengthen the management of spectrum resources and increase their effective utilization. The biggest problem of radio spectrum as a limited resource is that the allocation of radio spectrum is becoming tighter and tighter due to the rapid growth of the amount of user equipment (UE). The national government needs to use modulation identification technology for the unified management of spectrum resources. On the one hand, it needs to listen to illegal UE to analyze their sources and avoid interference and pollution to the frequency spectrum occupied by legal UE. On the other hand, legal UE need to be monitored to monitor whether they are strictly operating according to their assigned operating parameters [9]. The development of modulation recognition technology provides a strong guarantee for the efficiency of communication systems.

In terms of military electronic warfare, to intercept enemy communications, it is first necessary to determine the modulation type of the intercepted enemy signal, and then perform demodulation operations to decode the corresponding signal in order to finally obtain useful information [10]. Specifically, electronic support, electronic protection, and electronic attack are the three main parts of modern military electronic warfare. The task of electronic support is to collect the frequency information transmitted by the enemy radio station, after modulation and identification, use this frequency information to demodulate the intercepted enemy signal to obtain the information transmitted by the enemy. The goal of electronic protection is to protect one’s own communication results from being affected by electronic attacks from the enemy. One of the most effective means of electronic protection is to identify the jamming signal from the enemy and avoid the signal frequency band. At the same time, the own transmitter needs to detect the modulation type of the enemy jamming signal and convert the own signal into other modulation types to avoid the enemy-side interference. The main means of electronic attack is to transmit jamming signals to the enemy. Unlike the barrage jamming, blocking enemy communication with high similarity signals is more difficult to detect. In order to achieve the purpose of jamming the enemy’s communication, the frequency band and modulation type of the jamming signal must be completely consistent with the signal transmitted by the enemy. At this time, it is necessary to apply the modulation identification technology to obtain the enemy’s modulation type information [11].

In this paper, we propose an automatic modulation classification (AMC) method based on deep residual neural network with masked modeling (DRMM) for the scenario where the large amount of training samples with high quality are difficult to obtain. The main contributions of this paper are summarized as follows:

• An AMC method based on a deep residual neural network with masked modeling (DRMM) is proposed for solving the problems of modulated signal classification under limited training samples.
• The proposed AMC method is evaluated on simulated dataset and RadioML 2016.10A, and compared with three AMC methods. Simulation results show the proposed AMC method has better classification performance.
• The factor, i.e., masked areas, that makes an effort on the classification performance of the proposed AMC method is analyzed. In addition, the effectiveness of the proposed AMC method with a semi-supervised scenario is also discussed.

2. Related Work

Before the AMC technology was proposed, the modulation classification of modulated signals mainly relied on human judgment. Specifically, after the received high-frequency signal is converted into an intermediate-frequency signal, the intermediate-frequency signal is input to the demodulators corresponding to different modulation types, and relevant personnel judge the modulation type with the help of signal observation instruments based on their own experience. The disadvantages of such a modulation classification method are obvious. It requires high manpower and time costs, and lacks automation. In addition, some human judge methods will also suffer from the human subjectivity so that the accuracy rate is not high. These reasons make the emergence of AMC technology inevitable.

In 1969, C. S. Waver and others published the paper “Automatic Modulation Classification using Pattern Recognition” [12], which is the world’s first article on AMC technology. Subsequently, many researchers conducted research on AMC. The current mainstream signal modulation classification algorithms are mainly divided into two categories: the likelihood-based (LB) classification method and the feature-based (FB) statistical pattern classification method.

2.1. LB

Z. Wei and Y. Hu [13] proposed an optimized maximum likelihood method. When the signal-to-noise ratio (SNR) exceeds 0 dB, the classification accuracy for BPSK, 4PSK, and 8PSK signals is greater than 0.995, and the recognition rate for 16PSK is greater than 0.95. However, the scope of application of this method is small, only for intra-class classification of multi-ary phase modulation. J. Zheng [14] proposed a blind modulation classification algorithm based on hybrid likelihood ratio test (HLRT), using the log likelihood detector and energy detector to identify active subcarriers, and using the maximum expectation algorithm to estimate the noise variance and channel fading coefficient.

2.2. FB

K. C. Ho et al. [15] first proposed to use Haar wavelet transform to extract the characteristic parameters of PSK and FSK. However, the classification efficiency of this method is not high and it is easily affected by noise. H. Hadinejad-Mahram et al. [16] proposed a method for classification of digitally modulated signals based on performing subspace decomposition on a positive definite matrix of higher order moments of the received signals.

In recent years, machine learning has achieved good results in various classification tasks with its outstanding deep feature extraction ability [17,18,19,20,21]. Later, due to the gradual improvement of computer hardware performance and the gradual maturity of intelligent algorithms, the modulation classification methods based on deep learning (DL) have been proposed [22,23,24,25,26,27,28,29,30]. Some very related works are surveyed as follows. O’shea et al. [31] adopts a Universal Software Radio Peripheral (USRP) to collect 24 kinds of wireless modulation signals in real space, then inputs the data into Residual Networks (ResNets) to automatically extract features for classification. The recognition accuracy can reach 0.956 when the SNR is equal to 10 dB. R. Li et al. [32] proposed a deep geometric convolutional network to extract classification features layer by layer from the Wigner–Ville distribution map of the signal, and a set of geometric filters is constructed by using the least square method support vector machine (SVM) to replenish the linear singular points in the replenishment map. This algorithm has obtained a higher classification accuracy in the simulation. N. Jafar et al. [33] proposed an AMC method by simultaneously using normality test, spectral analysis, and geometrical characteristics of an in-phase-quadrature (I/Q) constellation diagram. This method has better performance at a low SNR region. J. Huang et al. [34] proposed a novel cascaded convolutional neural network (CasCNN) in which two-block convolutional neural networks are cascaded. The two blocks are used to classify different classes of modulation formats and to identify the indices of modulations in the same PSK or QAM class, respectively. The advantage of this method is that it exhibits more robustness to frequency shifts.

3. Problem Formulation

3.1. AMC Description

AMC is applied at the receiver to identify the different modulated signals. It is widely used in military and civilian fields such as cognitive radio, interference signal identification and spectrum sensing [35,36,37]. As shown in Figure 1, the input signals are modulated into radio frequency signals at the transmitter and received by the receiver after passing through the channel. The receiver converts the received signals into baseband complex signal sequence $\mathit{X}=\left\{\mathit{x}\left(0\right),\mathit{x}\left(1\right),\cdots ,\mathit{x}\left(K-1\right)\right\}$ to preprocess, where K is the number of sampling points. Then, identify the signal modulation type according to the types in the limited modulation scheme pooling $Y=\left\{y_i,i=1,2,\cdots ,N\right\}$, where N is the number of modulated types, $y_i$ is the modulated types identified.

In DL-based AMC method, instead of relying on manually designed characteristic parameters, AMC automatically classifies modulation types from sampled signals. This method has a strong generalization ability, and as shown in Formula (1), it utilizes the maximum a posteriori (MAP) criterion to design the classifier, which makes it show good classification accuracy and strong stability even in low SNR region.

$$\begin{array}{c}\hfill y_i={\arg \max }_{y_i\in Y}P\left(y_i|\mathit{X}\right)\end{array}$$

(1)

where $P\left(y_i|\mathit{X}\right)$ is the probability of that the modulated type is $y_i$. In the experiment, the structure and parameters of neural network should be designed to maximize $P\left(y_i|\mathit{X}\right)$, so as to improve the accuracy of signal recognition.

3.2. Signal Model

As shown in Formula (2), the signal model in this work assumes the time varying carrier phase offset.

$$\begin{array}{c}\hfill \mathit{x}\left(k\right)=A{e}^{j\phi \left(k\right)}\mathit{m}\left(k\right)+\mathit{n}\left(k\right),\end{array}$$

(2)

where $\mathit{m}\left(k\right)$ denotes the baseband signal sequences transmitted, $\mathit{n}\left(k\right)$ represents additive white Gaussian noise with mean value of 0, $A\in \left[0,1\right]$ is the channel gain, and $\phi \left(k\right)\in \left(0,\frac{\pi }{16}\right)$ represents the time varying phase offset.

The received modulated signals are IQ signals, including real part ($\mathit{I}$) and imaginary part ($\mathit{Q}$). This is also the reason that a complex convolutional neural network is selected in the recognition network model, that is, to avoid dividing the IQ signals and retain the information brought by the coupling of the two parts.

$$\begin{array}{c}\hfill \mathit{I}=\left[real\left(\mathit{x}\left(0\right)\right),real\left(\mathit{x}\left(1\right)\right),\cdots ,real\left(\mathit{x}\left(K-1\right)\right)\right]\end{array}$$

(3)

$$\begin{array}{c}\hfill \mathit{Q}=\left[imag\left(\mathit{x}\left(0\right)\right),imag\left(\mathit{x}\left(1\right)\right),\cdots ,imag\left(\mathit{x}\left(K-1\right)\right)\right]\end{array}$$

(4)

4. Our Proposed DRMM-Based AMC Method

The framework of our proposed DRMM-based AMC method is shown in Figure 2. Our method learns the representation by masking part of the signal samples and predicting the original signals at the masked area. The framework consists of 3 major components:

• Autoencoder-classifier architecture: It extracts the potential features of masked signals, and then uses them to reconstruct the original signals at the masked area and predict the ground-truth labels of masked signals.
• Masked modeling: Given the input signal samples, this part designs how to select the area to be masked and how to design the objective function for enabling the autoencoder to reconstruct the masked part of signals.
• Training process: The masked signals is input to autoencoder for extracting the features, reconstructing the masked area of mask signals and classifying its ground-truth label in the forward propagation. The loss is calculated and back propagated for updating the parameters of autoencoder-classifier.

In the following sections, we introduce each components of the proposed AMC method.

4.1. Autoencoder-Classifier Architecture

In the neural network model, the depth of the network has a huge impact on the learning performance of the model. The more layers of the network, the stronger the ability to learn the characteristics. However, this will also lead to the disappearance of the gradient, over fitting and other problems. At the same time, since the input data are I/Q complex data, the encoder used in this paper is a complex-residual neural network (C-ResNet) [38], and the residual network and residual unit are shown in Figure 3.

The C-ResNet has a similar structure to real-residual neural network (R-ResNet), except for the convolution layers. In C-ResNet, firstly, the complex convolution operation is performed on the input signal. Compared with the real convolution operation, the convolution kernel of the complex convolution operation combines the real part with the imaginary part, which can avoid the loss of the feature relationship and improve the recognition accuracy. Specifically, for the input signals

$$\begin{array}{c}\hfill \mathit{x}=\left\{\begin{array}{c}\mathit{I}\\ \mathit{Q}\end{array}\right\}=\left\{\begin{array}{c}real\left(\mathit{x}\right)\\ imag\left(\mathit{x}\right)\end{array}\right\}.\end{array}$$

(5)

As illustrated in Figure 4, the convolution kernel of the neural network is $W=A+iB$, and the complex convolution operation is shown as

$$\begin{array}{c}\hfill F=\left(A*\mathit{I}-B*\mathit{Q}\right)+i\left(B*\mathit{I}+A*\mathit{Q}\right)\end{array}$$

(6)

In addition, the increase in neural network depth will also lead to an increase in feature dimensions. Therefore, this paper sets a pooling layer behind each convolution layer for feature compression. At the same time, the batch normalization operation is added to the neural network, which improves the convergence speed of the neural network to improve the training speed. As shown in Figure 3, the decoder is a one-dimensional convolution with a convolution kernel size of 1 and the classifier is a linear layer to identify sample features $\mathit{z}$ as the predicted class distribution $\widehat{y}$.

4.2. Masked Modeling

Masked modeling [39,40] is shown in Figure 5, which is composed of mask 1 and mask 2. In this paper, mask 1 randomly selects 50% of the number of sample points of a given signal $\mathit{x}$ to mask, which can be expressed as

$$\begin{array}{c}\hfill \hat{\mathit{x}}=\mathit{x}\odot \mathit{M},\end{array}$$

(7)

where $\mathit{M}=\left\{z\right(n\left)\right|n=0,2,\cdots ,K-1\}$ is a sequence consisting of one and zero. Input signal $\widehat{\mathit{x}}$ into the encoder for encoding, and then input extracted features into light decoder for decoding to obtain $\tilde{\mathit{x}}$. Then, mask 2 operates on the original signals $\mathit{x}$ and the decoded signals $\tilde{\mathit{x}}$, that is masking the uncovered part in the process of mask 1, and can be formulated as

$$\begin{array}{c}\hfill {\mathit{x}}^{\prime }=\mathit{x}\odot (\mathbf{1}-\mathit{M})\end{array}$$

(8)

and

$$\begin{array}{c}\hfill {\tilde{\mathit{x}}}^{{}^{\prime }}=\tilde{\mathit{x}}\odot (\mathbf{1}-\mathit{M})\end{array}$$

(9)

where the ${\mathit{x}}^{\prime }$ represents demasked-original signals and ${\tilde{\mathit{x}}}^{{}^{\prime }}$ is demasked-reconstructed signals. The mean square error loss function is introduced for enabling the autoencoder to have the ability to reconstruct the mask part of signals with the help of information embedded in unmasked part of signals, which can be expressed as

$$\begin{array}{c}\hfill L_{MSE}=\frac{1}{N}\sum _{i=1}^N{\left({\mathit{x}}^{\prime }-{\tilde{\mathit{x}}}^{{}^{\prime }}\right)}^2,\end{array}$$

(10)

where $i=1,2,\cdots ,N$ and N is the number of sample points of masked part of a given signal.

4.3. Training Process

In the training phase, the objective function $\mathcal{L}$ can be formulated as

$$\begin{array}{c}\hfill \mathcal{L}={\mathcal{L}}_{\mathrm{MSE}}+{\mathcal{L}}_{\mathrm{CE}},\end{array}$$

(11)

$$\begin{array}{c}\hfill {\mathcal{L}}_{\mathrm{CE}}=-y\log p_y,\end{array}$$

(12)

where the y is the ground-truth label of a given signal. $p_y$ is the predicted class distribution. ${\mathcal{L}}_{\mathrm{CE}}$ is the cross-entropy loss. The value of objective function in a mini-batch is calculated and backwardly propagated to update the parameters of encoder, decoder and classifier. The details of training process are shown in Algorithm 1.

Algorithm 1 The training process of proposed DRMM-based AMC method for drones communications.

Parameters required:     ●      $\left\{\mathit{x},y\right\}$: A batch of training signals;     ●      T: The number of training iterations;     ●      B: The number of training batch size;     ●      $\eta$: Learning rate;     ●      $f_{ae}\left(·\right)$ and $f_c\left(·\right)$: The function of autoencoder and classifier, respectively;     ●      ${\mathit{\theta }}_{\mathit{ae}}$ and ${\mathit{\theta }}_{\mathit{c}}$: The parameter of autoencoder and classifier, respectively;     ●      $\widehat{y}$: The predicted label;     Train on training signals $\left\{\mathit{x},y\right\}$:

1.      for $t=1$ to T do:

2.          for  $b=1$ to B do:

[Forward propagation]:

3.              Obtain the masked original signals $\widehat{\mathit{x}}$

4.              Obtain the demasked original signals ${\mathit{x}}^{\prime }$;

5.              Obtain the reconstructed signals $\tilde{\mathit{x}}$ and features $\mathit{z}$:

$\tilde{\mathit{x}},\mathit{z}=f_{ae}\left({{\mathit{\theta }}_{\mathit{ae}}}^t;\widehat{\mathit{x}}\right)$

6.              Obtain the demarked-reconstructed signals ${\tilde{\mathit{x}}}^{{}^{\prime }}$:

${\tilde{\mathit{x}}}^{{}^{\prime }}=\tilde{\mathit{x}}\odot (\mathbf{1}-\mathit{M})$

7.              Obtain the predicted class distribution:

$\widehat{y}=f_c\left({{\mathit{\theta }}_{\mathit{c}}}^t;\mathit{z}\right)$

8.              Calculate loss function:

$L^b=L_{MSE}^b\left({\mathit{x}}^{\prime };{\tilde{\mathit{x}}}^{{}^{\prime }}\right)+L_{CE}^b\left(\widehat{y};y\right)$

[Backward propagation]:

9.              Update parameters:

$\mathit{\theta }\leftarrow \mathrm{Adam}\left({\nabla }_{\mathit{\theta }},L^b,\eta ,\mathit{\theta }\right)$

10.        end for

11.    end for

12.    Save model parameters

After the forward propagation and backward propagation process, the weights of the encoder, decoder and classifier are saved. In the testing process, the testing signals are input into the neural network for classification.

5. Experimental Results

5.1. Simulation Parameters

Our simulations are performed on an NVIDIA GeForce GTX1080Ti platform based on pytorch 1.8.1. The maximum epoch T is 150 and the batch size is 128. The learning rate is set to 0.001. Adam is selected as the optimizer.

5.2. Dataset Description

A dataset proposed in [41] and a simulated dataset are used to evaluate the proposed AMC method. The RadioML 2016.10A dataset in [41] is an open source dataset. Five modulation signals of BPSK, 8PSK, QPSK, QAM16 and QAM64 in this dataset are used as training, validating and testing samples. Each modulated signal generates 1000 samples at each SNR, and the length of each sample is 128. The numbers of samples of each modulated type per SNR are 420, 180 and 400 for training, validating and testing. The simulated dataset is generated by MATLAB, with the same modulated types, length of each samples and number of samples as RadioML2016.10A.

5.3. Comparative Methods

5.3.1. Machine Learning-Based AMC

SVM [42] is used as the representative of classic machine learning for analysis. SVM is a machine learning algorithm proposed for regression and classification problems. The specific steps are to use nonlinear mapping to map the input samples to a high-dimensional feature space, so that the original space data has approximately a linear relationship, and then construct a linear optimization decision function in the feature space. In this paper, the higher-order cumulants (HOC) of original signals are used as the input of SVM.

5.3.2. Real-Residual Neural Network-BASED AMC

To solve the degradation problem, the authors in paper [43] proposed R-ResNet. Each stacked layer of the network fits the residual mapping instead of directly fitting the underlying mapping expected by the whole building block. The advantage of this approach is that it is easy to train, can improve accuracy simply by increasing depth, and it is easily transferable.

5.3.3. Deep Reconstruction and Classification Network-Based (DRCN) AMC

DRCN [44] is a new deep reconstruction and classification network structure, which is composed of a convolutional autoencoder (CAE) and convolutional neural network (CNN). The unlabeled samples are used for modulation signal reconstruction in CAE, and the labeled samples are sent to CNN for AMC. Knowledge is transferred from the encoder layer of CAE to the feature layer of CNN to avoid invalid feature extraction of CNN under limited label samples.

5.4. Simulation Analysis

5.4.1. Classification Accuracy: Machine Learning vs. Deep Learning

The classification performance of the traditional machine learning algorithm SVM is compared with that of R-ResNet, as shown in the Figure 6. On the RadioML 2016.10A dataset, the classification performance of the deep learning algorithm is better than that of the machine learning algorithm. On the simulated dataset, although the SVM performance is better than that of R-ResNet under high SNR, the mean value of classification accuracy is lower than R-ResNet, as shown in the

Table 1. Mean value of classification accuracy of R-ResNet and SVM under different SNR.

Method	R-ResNet	SVM with HOC
Acc (Simulated Dataset)	61.87%	44.00%
Acc (RadioML 2016.10A)	80.28%	51.10%

. The experimental results show that the deep learning method is better than machine learning method in classification performance.

5.4.2. Classification Accuracy: R-ResNet vs. C-ResNet

In order to analyze the effectiveness of R-ResNet and C-ResNet, the classification performance of C-ResNet and R-ResNet, and the experimental results are shown in Figure 7. It is shown that the classification performance of C-ResNet is similar to R-ResNet on the two datasets in

Table 2. Mean value of classification accuracy of R-ResNet and C-ResNet under different SNR.

Method	R-ResNet	C-ResNet
Acc (Simulated Dataset)	61.87%	62.17%
Acc (RadioML 2016.10A)	80.28%	80.63%

. Thus, both of them are combined with masking modeling for future research.

5.4.3. Classification Accuracy: DRMM vs. Comparative Methods

To demonstrate the effectiveness of the proposed method DRMM in a limited sample scenario, the performances of DRMM and DRCN are compared. The classification performance of C-ResNet at full samples is used as the performance benchmark, while a randomly selected subset comprising $20\%$ of the full samples are used as limited samples to train DRMM and the comparative methods, as shown in Figure 8 and Figure 9 and

Table 3. Mean value of classification accuracy of DRMM and comparative methods under different SNR.

Method	C-ResNet	C-ResNet	DRCN	DRMM (R)	DRMM (C)
Acc (Simulated Dataset)	62.17%	54.96%	58.47%	57.86%	60.72%
Acc (RadioML 2016.10A)	80.63%	69.75%	70.95%	73.31%	72.11%

. The DRMM (R) and DRMM (C) in Table 3 are based on real-valued and complex-valued ResNets, respectively.

The experimental results show that the proposed DRMM method has better classification performance than comparative methods under the limited samples of simulated dataset and RadioML 2016.10A. It is worth noticing DRMM (R) has worse performance compared to the DRCN with simulated dataset. Therefore, this paper will focus on the complex-valued DRMM in the next sections, which will be directly referred to as DRMM for simplicity.

5.4.4. Influence of Masked Area on Classification Performance of DRMM

In total, $10\%$, $30\%$, $50\%$ and $70\%$ of the sample content are masked, respectively, to test the classification performance of DRMM under different masked areas, and the experimental results are shown in Figure 10 and Figure 11 and

Table 4. Mean value of classification accuracy of DRMM under different SNR regimes.

Method	DRMM	DRMM	DRMM	DRMM
Masked area	10%	30%	50%	70%
Acc (Simulated Dataset)	57.84%	$\mathbf{60}.\mathbf{95}\%$	60.72%	58.90%
Acc (RadioML 2016.10A)	64.44%	69.92%	$\mathbf{72}.\mathbf{11}\%$	70.24%

. According to the experimental results, the 50% is chosen as the optimal masked area.

5.4.5. Classification Performance of Semi-Supervised DRMM

In this paper, the semi-supervised scenario is also considered, specifically, 20% of labeled data and 80% of unlabeled data are used for training, and the experimental results are shown in Figure 12 and Figure 13 and

Table 5. Mean value of classification accuracy of semi-supervised DRMM under different SNR regimes.

Method	C-ResNet	C-ResNet	DRCN	DRMM
Sample percentage	100%	20%	20%	20%
Acc (Simulated Dataset)	62.17%	54.96%	58.88%	$\mathbf{59}.\mathbf{68}\%$
Acc (RadioML 2016.10A)	80.63%	69.75%	70.10%	$\mathbf{71}.\mathbf{45}\%$

. It can be seen that the proposed DRMM has the best classification performance under the semi-supervised scenario. In addition, compared to only using labeled samples, the above methods generally have negative effects, which is due to the sensitivity of semi-supervised learning to the number of labeled data and unlabeled data [45].

6. Conclusions

In this paper, we propose a DRMM-based AMC method with strong modulated signal recognition capability in limited sample scenarios for network communications. Specifically, a masked autoencoder is used to reconstruct the partially masked signal as the original signal. A deep residual network is used to reduce the impact caused by network degradation and to improve the network mobility. An objective function consisting of cross-entropy loss and mean square error loss is designed to obtain a better performance of feature extraction with high resolution. We evaluate the proposed DRMM-based AMC method on an open-source RadioML 2016.10A dataset and a simulated dataset and compare it with three AMC methods. A limited sample scenario and a semi-supervised scenario are considered in this paper. The simulation results illustrate that our proposed DRMM-based AMC method achieves better performance in the case of limited signal samples with low SNR than other existing methods.