CLASSIFYING WIRELESS SIGNAL MODULATION SORTING USING CONVOLUTIONAL NEURAL NETWORK

Deep learning has recently been used for this issue with superior results in automatic modulation classification. Previous studies state that it is challenging to categorize a variety of modulation formats using traditional approaches; however, modulation classification is a crucial component of non-cooperative communication in wireless communication. The deep learning network was applied to solve the issue and get decent outcomes. This work uses a deep learning convolutional neural network (DLCNN) to classify three analog and eight digital modulation techniques by generating channel-impaired and synthetic waveforms as training data. The obtained DLCNN is tested by over-the-air indicators and a Software Define Radio(SDR) platform. The trained DLCNN estimates the modulation kind of each frame by taking 1024 samples of channel-impaired signals. The method includes generating several frames of 4-arry pulse amplitude modulation (PAM4) that are impaired with sampling time drift, Additive white Gaussian noise (AWGN), center frequency, and Rician multipath fading. The DLCNN predicts real inputs when receiving a signal with complex samples of baseband. Before updating the network coefficients and on all iterations, the data store transforms data from files and records it. This network takes about 50 minutes to train using in-memory data and 110 minutes to train using disk data. The eval-uation of the trained DLCNN is carried out by obtain-ing the classification accuracy for the test frames. The obtained outcome demonstrates that the developed network can achieve an accuracy of about 94.3 % in roughly 12 epochs for such types of waveforms, which elapsed about 26 minutes for training. This will increase the efficiency of spectrum usage and detect the modulation type of the wireless communication receivers


Introduction
Automatic modulation categorization (AMC) is a tool for primary signal processing techniques of the wireless physical layers that are used to increase the efficiency of spectrum usage and intend to blindly detect the modulation kind for the received signal of the wireless communication receivers. Numerous contemporary AMC methods utilizing deep networks have been proposed to address the current limitations of conventional approaches, which are motivated by deep learning convolutional neural network (DLCNN) high-influence achievements in several informatics fields, including radio signal processing for communications. In order to recognize modulation patterns, DLCNN can efficiently learn the fundamental properties of radio signals. This enhances the modulation classification performance even when there are channel defects [1]. In order to determine the kind of wireless signal modulation and recover the signal by demodulation, automatic modulation classification is crucial in many fields [2].
It is challenging to categorize a variety of modulation formats using traditional approaches; however, modulation classification is a crucial component of non-cooperative communication in wireless communication. The deep learning network was applied to solve the issue and get decent outcomes. The input data length for the (DLCNN) is fixed [3][4][5]. The network is unable to use the input signal data to increase classification accuracy since the signal length varies during communication [6]. Wireless communications can use automatic modulation categorization in a wide variety of situations. Deep learning has recently been used for this issue with superior results. DLCNN-based modulation categorization typically uses fixed-size inputs. The actual radio signal burst's length can vary, though.
In resident communications such as cognitive radios and military signal reconnaissance, automatic modulation classification (AMC) is a crucial technique. The majority of previous studies concentrated on the AMC in additional white Gaussian noise channels; however, it is more feasible and difficult to implement the AMC in time-varying wireless channels [7].
Wireless networks in the 5G and outside will be more heterogeneous and dynamic, necessitating the use of multistrand waveforms. The identification of a specific modulation type that the transmitter employs at the specified time to correctly decode the data is one of the biggest obstacles in such a dynamic network, especially in non-cooperative instances [8].
Significant interest is shown in automatic modulation classification in the context of both present and future wireless communication systems. Deep learning has become a potent method for modulation classification because it enables concomitant learning of discriminative characteristics and the categorization of signals. However, optimizing DLCNN architectures for modulation classification is a labor-intensive, manual procedure that needs extensive domain expertise [9]. Therefore, wireless communications are essential to apply the automatic modulation classification in a wide variety of conditions, and using DLCNN has recently been used for this issue with superior results.

Literature review and problem statements
The study [1] provided a fundamental concept of a variety of architectures, such as DLCNN, long short-term recall, recurrent neural networks, and neural networks as an essential environment. This study then discussed the use of Deep Learning (DL) of adaptive modulation and coding (AMC) in wireless communications. Although the study investigated advanced designs and several sophisticated structures of DLCNN in various data kinds of constellation images, spectrum photos, and sequential radio signals to deal with a variety of channel impairments, it provides a survey and was limited to what has been done before. The paper [10] also reviews a variety of DLCNN models and algorithms for the classification and modulation recognition of wireless communication patterns. However, it didn't discuss the recognition between the analog and the digital modulation kinds such as Single sideband amplitude modulation (SSB-AM), Double sideband amplitude modulation (DSB-AM), Broadcast FM (B-FM), Continuous phase frequency shift keying (CPFSK), Gaussian frequency shift keying (GFSK), 4-array pulse amplitude modulation (PAM4), 64-array quadrature amplitude modulation (64-QAM), 16-array quadrature amplitude modulation (16-QAM), 8-array phase shift keying (8-PSK), Quadrature phase shift keying (QPSK), and the Binary phase shift keying (BPSK).
The paper [2] contributed to exploring the appropriate design of the DLCNN technique in the area of communication signals recognitions. Although the study adopted the de-noising auto-encoder to provide the ability to resist finite perturbations of the input and pre-process the received data, it fails to cover the analog and digital modulations. The research [6] proposed a revolutionary DLCNN technique that makes use of a multi-stream topology. The proposed shorter network structure makes it easier to compare and recognize some digital modulation signals and avoid over-fitting issues. However, only a limited number of communication signals have been recognized. The study [11] offered a multi-stream architecture to enrich the types of signal features obtained and increase the network width, but this study also didn't compare the analog and digital modulation types. A wide expansion for signal recognition has been done by the paper [12], where full use of the complete signal burst was used to get better classification accuracies. However, the results obtained showed only three types of recognized communication signals.
The article [7] investigated a time-varying AMC guide using the DLCNN technique to get high classification accuracies. Its experimental outcome showed that the presented AMC system achieved superior classification precision in both fast and slow fading, but it was with very complicated AMC architectures.
All this allows to assert that it is expedient to conduct a study on the use of a deep learning convolutional neural network (DLCNN) to classify analog and eight digital modulation signals with generated waveforms as training data. Then, the obtained DLCNN network is tested by a platform/hardware such as over-the-air signals and software-defined radio (SDR).

The aim and objectives of research
The aim of research is to classify wireless signal modulations by sorting them according to the type using deep learning convolutional neural network DLCNN. This will make it possible to categorize various modulation schemes as modulation classification is a crucial component of non-cooperative communication in wireless communication.
To achieve this aim, the following objectives are accomplished: -to develop a DLCNN for classifying three analog and eight digital modulation techniques; -to generate channel-impaired and synthetic waveforms as training data; -to train and test the developed DLCNN.

1. Object and research hypothesis
This work develops a deep learning convolutional neural network (DLCNN) architecture to classify three analog and eight digital wireless modulation techniques by generating channel-impaired and synthetic waveforms as training data. The categorized modulation types include BPSK, QPSK, 8PSK, 16QAM, 64QAM, PAM4, GFSK, CPFSK, B-FM, DSB-AM, and SSB-AM. The obtained DLCNN is tested by over-the-air signals and software-defined radio (SDR) platform using MAT-LAB environment functions. The assumptions made in the work are described in section 4. 2, while the simplifications adopted are demonstrated in the following subsections.

2. Using deep learning to predict modulation type
This work includes recognizing the analog and digital modulation kinds as listed in Table 1.
4. The trained DLCNN gets channel-impaired 1024 samples to predict the type of modulation for every frame.
5. Creating some frames of PAM4 to be impaired with Additive white Gaussian noise (AWGN), sampling time drift, center frequency, and Rician multipath fading.
6. Using the following MATLAB functions to create synthetic signals for testing the DLCNN.
7. Using the DLCNN to estimate the frames modulation types.

3. Waveform generation for training
The proposed work generates about 10,000 frames for every modulation kind, of which 10 % is used for testing, 10 % is used for validation, and the rest 80 % is used for training. The validation and training frames are used throughout the training phase of the network. The ultimate accuracy of classification type is acquired via the testing frames. Every frame is with a sample rate of 200 kHz and is 1024 samples long. For digitally modulated types, a symbol is represented by eight samples. Instead of considering numerous consecutive frames, the network bases each judgment on a single frame. Let's assume that the middle frequencies for the digital and analog modulation kinds are 902 MHz and 100 MHz, correspondingly. Fig. 2 shows the flowchart of the algorithm assumptions.
The inaccuracy of transmitters' and receivers' internal clock sources results in clock offset. The sample rate of the digital-to-analog converter and the middle frequency, which are utilized to down-convert the signal to the baseband, deviate from the optimum values due to clock offset (Δclock). The clock offset factor C for the channel is given by C=1+Δclock/10 6 . Each channel creates an arbitrary Δclock of homogeneous distribution domain from the range [−maxΔclock; maxΔclock]. Table 1 List of the recognized analog and digital modulation types
Let's start by training the convolutional neural network (CNN) with labeled (known) data and generating synthetic data for training. Next, let's define, train, and test the network to classify the modulation. Finally, by using the platforms of software-defined radio (SDR), let's examine the CNN performance with over-the-air signals.

Fig. 1. Using deep learning to predict modulation type
For every modulation type, let's create a loop that produces channel-impaired frames and stores them in MAT files with the relevant labels. It is possible to avoid having to generate the data each time let's execute the function by storing the data in files. The data can be distributed more successfully as well. Therefore, the flowchart showing the waveform generation for training is shown in Fig. 3.
To get rid of transients and guarantee starting frames at an arbitrary location concerning the symbol limitations, let's take an arbitrary number of samples out of the beginning of each frame.

Training the Deep Learning Network
This work employs a DLCNN with one fully connected layer and six convolution layers. Every convolution layer includes a max pooling layer, activation layer (rectified linear unit (ReLU)), and batch normalization layer except the final convolution level, where an average pooling level is used in place of the maximum layer. There is softmax activation in the output layer. The screenshot for the details of the developed network layers is shown in Fig. 4.

Fig. 4. Screenshot for the details of the developed network layers
The training configuration is as follows: 1. Every nine epochs, the learning rate is decreased by a factor of 10.
3. Setting the maximum number of epochs to 12. 4. Using an SGDM solver with a mini-batch size of 256. 5. Plotting the training progress.

1. Using deep learning network to predict modula tion type
The probability of every frame containing the anticipated modulation kind is represented by the score shown in Fig. 5.
The network classifier predicts the analog decisions and successfully identifies the frames as PAM4 [7×1] categorical. A score vector of every frame can be represented by the classifier.

2. Waveform generation for training
The sample numbers against the magnitude of the imaginary and real regions for the adopted frames are shown in Fig. 6, while the plotting of the spectrograms for the adopted frames is shown in Fig. 7.
All these Fig. 6, 7 represent the signal waveforms generated for training the network.

3. Training the deep learning network
The training process has been performed over a Titan XpNVIDIA® GPU, which elapses about 54 minutes for training. The network converges to more than 95 % accuracy in roughly 12 epochs, as seen by the training progress graphic as shown in Fig. 8.
To obtain the classification accuracy for the test frames, a confusion matrix is prepared as shown in Fig. 9.
The obtained confusion matrix indicates that for this collection of waveforms, the network achieves an accuracy of roughly 94.5455 %.   Table 2 shows a comparison between the presented approach and the existing studies in terms of the number of epochs, training elapsed time, and the maximum accuracy achieved.

Discussion of the results of the proposed classifying wireless signal modulations
The developed network classifier predicted the analog decisions and successfully identified the frames as PAM4 [7×1] categorical. The probability that every frame contains the anticipated modulation kinds is represented by the score as shown in Fig. 4. The sample numbers against the amplitude of the real and imaginary regions for the adopted frames are shown in Fig. 5, while the plotting of the spectrograms for the adopted frames is shown in Fig. 6.
The training process by measuring the accuracy and loss for the developed network has been shown in Fig. 7. The confusion matrix shown in Fig. 8 indicates that the network conflates 64/QAM and 16/QAM frames, as the matrix demonstrated. Since each frame only contains 16/QAM and 128 symbols are a subset of 64/QAM, this issue is expected. Because the constellation of those modulation kinds appears related when rotated-phase caused by the frequency offset and fading channels, this network conflates 8/PSK and QPSK packets.
A comparison between the presented approach and the existing studies in terms of the number of epochs, training elapsed time, and the maximum accuracy achieved is shown in Table 2.
The disadvantage of the proposed solution is the complexity of the number of hidden layers of the network of the classifier, which is increased with the number of classes. This can be eliminated by exploring more deep learning networks such as ResNet50, and ResNet80 networks over a more modern modulation categorization in the future.
2. The presented work has successfully generated channel-impaired and synthetic waveforms as training data for the developed DLCNN. The process elapses about 54 minutes for training and the network converges to more than 95 % accuracy in roughly 12 epochs.
3. The trained DLCNN has been tested and achieved an accuracy of roughly 94.5455 %.

Conflict of interest
The authors affirm that they do not have any conflicts of interest with this research financial, authorship, personal, or otherwise that would have an impact on the findings of the study and how they are reported in this work.

Financing
The study was performed without financial support.

Data availability
Data will be made available on reasonable request.