Low-power and reliable gas sensing system based on recurrent neural networks

• Published in: Sensors and actuators. B, Chemical Chemical, 2021-08, Vol.340, p.129258, Article 129258
• Main Authors: Kwon, Dongseok, Jung, Gyuweon, Shin, Wonjun, Jeong, Yujeong, Hong, Seongbin, Oh, Seongbin, Bae, Jong-Ho, Park, Byung-Gook, Lee, Jong-Ho
• Format: Article
• Language: English
• Subjects: Errors; Field effect transistors; Field-effect-transistor (FET)-type gas sensor; Gas sensing; Gas sensing system; Gas sensors; Gases; Hydrogen sulfide; Indium oxides; Neural networks; Neuromorphic system; Nitrogen dioxide; Operating temperature; Power management; Pulse duration modulation; Recurrent neural network (RNN); Recurrent neural networks; Simulation; Synaptic device
• ISSN: 0925-4005; 1873-3077
• DOI: 10.1016/j.snb.2020.129258

•A field effect transistor (FET)-type gas sensor with In2O3 film is fabricated.•Recurrent neural networks (RNNs) and fully connected neural networks (FCNNs) are trained with the sensor data.•A hardware-based and low-power (0.412 mW) neural system for gas detection is designed.•The output of system is matched to that of high-level simulation within 0.3 % error. A new gas sensing system using recurrent neural networks (RNNs) is proposed for reliable gas identification. A field-effect-transistor (FET)-type gas sensor with an In2O3 film as a sensing material was fabricated for detecting NO2 and H2S gases. Then, the transient characteristics of the sensor were investigated depending on the operating temperature and gas concentration. Training was conducted in the Pytorch framework (high-level simulation) using RNNs and fully connected neural networks (FCNNs) with the measured sensing data, and the performance of these networks was compared. In a high-level simulation, the RNNs showed better performance (1.94 % error) than that of the FCNNs (3.02 % error) even with fewer neurons and smaller synapse arrays. Using the RNNs, a hardware-based and low-power (0.412 mW) neural system was designed for gas detection with the sensors, synaptic devices, and peripheral circuits. The weights were trained in the high-level simulation and transferred to the conductance of the synaptic devices. Input encoding in the system is performed by pulse-width modulation, and the weighted sum was calculated according to Kirchhoff laws. The system operation in the time domain was verified in the SPICE simulation, and the predicted output of the system matched that of the high-level simulation within a 0.3 % error.