Research Objective/Question Can DNN model be utilized for Signal decoding at receiver and assess the quality of decoding? EECS7025

Research Objective/Question Can DNN model be utilized for Signal decoding at receiver and assess the quality of decoding?

Conventional SIC decoding has issues with the error rates and decoding when subjected to multiusers superimposed signals

Results and Evaluation

This section presents the results of the simulations performed for signal decoding at the receiver end utilizing DNN based decoding model. The results of this simulation are compared with the conventional SIC decoding algorithm.

The experiments are conducted using the following software and tools Google Collab and MATLAB. TensorFlow, an open-source machine learning framework developed by Google was employed to develop the deep neural network for decoding of a superimposed signal received. MATLAB was utilized to develop the conventional SIC decoding algorithm.

As discussed in the previous section on model buildup, a two-user SISO NOMA system in a downlink transmission is considered in this study. The channel parameters are known between the transmitter and receiver. The power allocation of the users: User1 and User2, was based on their channel conditions with the total power constraint being set to unity. User1 is farthest away from the base-station being assigned a higher power coefficient as compared to User2 which is assigned a lower power coefficient.

As part of this study, we have considered both a time-invariant (TIV) and time varying channel with complex distribution of Rayleigh fading coefficients to assess the robustness of the deep neural network.

The received superimposed signal was divided into real and complex parts and fed into the input layer of the neural network consisting of two nodes as a column vector.

The deep neural network block developed for signal detection was trained using more than 60 million samples. A mini-batch gradient descent algorithm was implemented to achieve faster convergence and help prevent overfitting. The labels used in the supervised training process were one-hot encoded to remove any varying magnitude or ordinality among the samples.

The key parameters of the neural network model are summarized in the below table:

Parameter Value

Operating System Windows 11

Framework TensorFlow

Programming Language Python, MATLAB

Channel Rayleigh Fading, AWGN

Number of users 2

Modulation BPSK, QPSK

Number of training samples 65,000,000

Number of testing samples 6,000,000

Activation Functions ReLU, Sigmoid

The results were extracted in terms of a bit-error rate Vs signal-to-noise ratio (SNR) curve. Bit error rate which indicates the error level in the decoded signal provides an indication of the strength of the forward error correction code. While SNR represents the ratio of signal power to noise ratio in dB. Thus the curve generated helps in demonstrating the performance of the proposed DNN based decoding vis a vis the conventional SIC decoding algorithm system. For the same BER lower the SNR represents better performing system.

We have presented for analysis a bit-error rate Vs signal-to-noise ratio (SNR) curve to compare the two different detection blocks. For the purposes of this research, we have considered two different modulation types: BPSK and QPSK. We observe that implementing different modulation schemes exhibit varying levels of robustness to noise and interference, leading to distinct BER Vs SNR characteristics.

BPSK Modulation Scheme Results

BPSK modulated symbols, time-invariant (TIV) channel is considered for assessing the performance of the DNN based decoding system at the receiver end.

Two user configuration has been considered as described previously. User1 has been allocated a higher power-coefficient of 0.8 and User 2 has been allocated a lower power-coefficient of 0.2.

BER vs SNR curve for SIC decoding scheme were generated using Matlab and are presented. BER vs SNR curve for our model generated using DNN model is presented.

We observe comparable performance in the two different detection blocks (more quantifiable information to be added can be obtained from code outputs). No pre-processing nor post-processing was done.

In the figures below, the following wireless parameters were selected: BPSK modulated symbols, time-varying channel. User1 was allocated a power-coefficient of 0.79 and User2 was allocated a power-coefficient of 0.19. Figure to the left is the BER Vs SNR curve attained from the DNN detection block and the figure to the right is the BER Vs SNR curve attained through conventional SIC decoding.

This is being generated by similar coding in tensorflow00This is being generated by similar coding in tensorflow

We observe similar but almost better performance using the DNN detector block to decode the QPSK modulated symbols in a time-varying channel (more quantifiable information to be added can be obtained from code outputs). No pre-processing nor post-processing was done.

In the figures below, the following wireless parameters were selected: QPSK modulated symbols, time-invariant (TIV) channel. User1 was allocated a power-coefficient of 0.879 and User2 was allocated a power-coefficient of 0.12. Figure to the left is the BER Vs SNR curve attained from the DNN detection block and the figure to the right is the BER Vs SNR curve attained through conventional SIC decoding.

-473841-747970This is generated in Tensor flow 1 for QPSK modulation

00This is generated in Tensor flow 1 for QPSK modulation

We observe comparable performance in the two different detection blocks (more quantifiable information to be added can be obtained from code outputs). No pre-processing nor post-processing was done.

For QPSK modulated signals the Tensorflow program in new version is not producing results

Typical flowchart is attached

1. Define SNR array ( np.array([0,2,4,6,8,10,12,14,16,18,20,22,24,26,28,30])
2. Define Batch size 3 Define Test size 4 Generate Data - imaginary numbers (1,0), (1,1), (0,0), (0,1)
3. For QPSK data multiply the generated data with 0.707
4. Generate h1, h2, k 7 Define Power level P1 =0.8 P2=0.2.
5. DEVELOP DNN MODEL WITH 1 INPUT /OUTPUT LAYERS AND 7 INNER LAYER9 Generate BER & SNR curves