Hardware-Based Architecture for DNN Wireless Communication Models.

Area of Science:

• Electrical Engineering
• Computer Science
• Signal Processing

Background:

• Multiple Input Multiple Output Orthogonal Frequency Division Multiplexing (MIMO OFDM) is crucial for wireless communications.
• High peak-to-average power ratio (PAPR) in MIMO OFDM transmitters causes non-linear distortions.
• These distortions degrade signal detection and channel estimation at the receiver.

Purpose of the Study:

• To propose three deep neural network (DNN) models to mitigate non-linear distortions in MIMO OFDM systems.
• To replace conventional digital signal processing (DSP) modules with DNNs at the receiver.
• To evaluate the performance of DNN-based receivers in 2x2 and 4x4 MIMO OFDM systems.

Main Methods:

• Developed three DNN models: Type I for signal de-mapping, Type II for noise filtering, and Type III for combined de-mapping and detection.
• Implemented and tested DNN models in software for 2x2 and 4x4 MIMO OFDM systems.
• Designed and implemented hardware architectures for DNN models using quantization and pipelining on an FPGA board.

Main Results:

• Achieved robust bit error rate (BER) performance with the proposed DNN receivers.
• Demonstrated the practical feasibility of DNN models through successful hardware implementation.
• Validated the effectiveness of quantization and pipelining techniques for hardware acceleration.

Conclusions:

• DNN models effectively address non-linear distortions caused by power amplifiers in MIMO OFDM transmitters.
• The proposed DNN-based receivers offer improved signal detection and channel estimation performance.
• Hardware implementation on FPGA confirms the practical viability of DNNs for future wireless communication systems.