FPGA Implementation for Odor Identification with Depthwise Separable Convolutional Neural Network.

Area of Science:

• * Computer Engineering
• * Artificial Intelligence
• * Sensor Technology

Background:

• * Integrated electronic nose (e-nose) systems combine sensor arrays and recognition algorithms for diverse applications.
• * Existing e-nose systems face challenges with accuracy-performance trade-offs: simple algorithms lack accuracy, while complex ones are slow.
• * Need for efficient and accurate odor identification methods in resource-constrained hardware.

Purpose of the Study:

• * To propose and implement a deep neural network for efficient odor identification on a Field-Programmable Gate Array (FPGA).
• * To develop a lightweight neural network architecture that balances accuracy and computational speed for hardware deployment.
• * To validate the proposed method's effectiveness using a real-world dataset.

Main Methods:

• * Development of a lightweight odor identification deep neural network using depthwise separable convolutional neural networks (OI-DSCNN).
• * Quantization of the OI-DSCNN model using the saturation-flooring KL divergence (SF-KL) scheme for hardware efficiency.
• * Implementation of the quantized OI-DSCNN on a Zynq-7020 System-on-Chip (SoC) FPGA.

Main Results:

• * The proposed OI-DSCNN significantly reduces parameters and enhances hardware implementation performance compared to standard CNNs.
• * Quantization using the SF-KL scheme enables efficient deployment on the target FPGA.
• * Simulation and hardware experiments demonstrated the effectiveness of the implemented system for odor identification.

Conclusions:

• * The proposed lightweight deep neural network approach enables quick and accurate odor identification on FPGAs.
• * This method addresses the limitations of current e-nose systems, offering a viable solution for embedded odor sensing.
• * Findings pave the way for advanced, real-time odor analysis in various applications using FPGAs.