On-Chip Implementation of Neural Network-Based Classifier Models for E-Nose With Chemometric Analysis

The present work focuses on the comprehensive implementation of gas/mixture identification employing chemometric analysis followed by on-chip realization. The chemometric analysis has been accompanied by the neural network (NN)-based models and on-chip deployment has been efficiently done with high-level synthesis (HLS)-assisted flow on field programmable gate array (FPGA) and application specific integrated circuit (ASIC) platforms. Essentially, hardware-constrained NN-based classifier models have been built using neural architecture search (NAS), optimized using quantization, and implemented using HLS. The optimized NN-based models are built for three datasets extracted from three distinct applications, viz., examination of beef quality, detection of hazardous gases, and monitoring of vehicular exhausts. The well-generalized model implemented in Python is redefined in C++ using appropriate transformation functions inside the HLS process. The HLS process implements an optimized RTL logic for the prior optimized NN-based classifier models from the respective C++ descriptions. The experiment demonstrates the efficacy of compact and low-power classifier models in achieving outstanding recognition accuracy for three datasets, without any sensor array (SA) data preprocessing overhead.