Efficient fusion of spiking neural networks and FET-type gas sensors for a fast and reliable artificial olfactory system

• Published in: Sensors and actuators. B, Chemical Chemical, 2021-10, Vol.345, p.130419, Article 130419
• Main Authors: Kwon, Dongseok, Jung, Gyuweon, Shin, Wonjun, Jeong, Yujeong, Hong, Seongbin, Oh, Seongbin, Kim, Jaehyeon, Bae, Jong-Ho, Park, Byung-Gook, Lee, Jong-Ho
• Format: Article
• Language: English
• Subjects: Algorithms; Artificial neural networks; Back propagation networks; Bias; Datasets; Fast detection; Field effect transistors; Field-effect-transistor (FET)-type gas sensor; Gas sensing system; Gas sensors; Gases; Hardware; Hydrogen sulfide; Indium oxides; Learning theory; Machine learning; Neural networks; Neuromorphic system; Nitrogen dioxide; Pattern recognition; Semiconductor devices; Sensors; Smell; Spiking; Spiking neural network (SNN); Synaptic device; Transient current
• ISSN: 0925-4005; 1873-3077
• DOI: 10.1016/j.snb.2021.130419

•A FET-type gas sensor with In2O3 film was investigated with the micro-heater bias, pre-bias, and gas concentration.•A gas sensor data set was prepared with the gas reaction within 4.8 s.•A hardware-based spiking neural network (SNN) is designed with the bio-plausible integrate and fire neuron.•The SNN shows superior accuracy for gas prediction although a significant read fluctuation of the sensor is considered. A new artificial olfactory system based on a spiking neural network (SNN) and field-effect transistor (FET)-type gas sensors is proposed for quickly and reliably detecting toxic gases. A FET-type gas sensor was fabricated with a micro-heater, and an In2O3 film was used as a sensing material for detecting NO2 and H2S gases. The sensor was investigated with the micro-heater bias, pre-bias, and gas concentration, and an efficient data set to be used for training a neural network was prepared using the measured transient currents of the sensors within 4.8 s. Then, an artificial neural network (ANN) using the backpropagation algorithm, which is the most popular algorithm in pattern recognition, was applied to train the data set. The weights trained in the ANNs were transferred into the conductance of synaptic devices in the hardware-based SNN. The SNN using only 12 sensors shows a low error rate (∼3 %) in predicting the concentrations of NO2 and H2S. In addition, since the neuron in the SNN directly converts the sensor current into the voltage spike rate, the SNN predicts the gas concentration in real-time (within ∼5 s). Finally, considering the effect of the read fluctuation of the sensors, it turns out that the hardware-based SNN outperforms conventional machine learning algorithms.