Deep convolutional neural network architecture with wireless domain-specific FIR layers for modulation classification

Deep learning techniques have been shown to outperform traditional techniques for classifying radio signals. However, most deep learning techniques for wireless applications are inspired from image classification methods, which do not directly leverage domain knowledge and unique characteristics of over-the-air signal propagation. This thesis proposes a new operational layer for deep learning architectures, which applies a learn-able complex digital finite impulse response filter (FIR) before the traditional convolutional operations. This custom-layer is implemented in previously demonstrated architectures for the use-case of radio signal classifications and tested extensively on a dataset of 24 modulation classes collected for wide range of signal-to-noise ratios (SNR). Results show that replacing a traditional convolutional block with the proposed FIR operation increases classification accuracy on the dataset to above 95%. This paper also uses quantization techniques to perform training and classification with low precision integers with the goal of implementing the model on an FPGA. Training the model on the same dataset as the full-precision model shows a small reduction in accuracy (2%) but still greatly outperforms the baseline model at high SNR. This model is then transformed into a representational model of and emulated within FPGA. Resource usage on the FPGA and processing latency is calculated for the model. Results reveal that our approach uses 80% of available resources in Xilinx's ZU9EG FPGA while achieving an inference rate of 581.86 frames/s on 1024 sample complex frames at 45W of power. Our implementation shows a 4x reduction in power and has the advantage of deterministic latency over the GPU implantation.--Author's abstract