Dual-temperature MEMS sensors for progressive thermal-runaway-related gas prediction via a CNN–transformer network

Abstract Early detection of lithium-ion battery (LIB) thermal runaway is essential to prevent disastrous failures. While gas sensor arrays detect characteristic off-gases (H 2 , CO, CH 4 ), a large amount of sensing elements may increase heating power and hardware complexity. We propose a dual-temperature sensing system comprising only two MoS 2 @ZIF-67 microelectromechanical systems (MEMS) gas sensors with a dual-history convolutional neural network (CNN)-Transformer for progressive gas-concentration estimation. The sensors operate simultaneously at constant heater voltages of 2.5 and 3.0 V, providing two resistance channels. Independent CNN–Transformer encoders process short and longer signal histories to estimate H 2 , CO and CH 4 concentrations, followed by causal smoothing using an exponentially weighted moving average (EWMA). Exploratory evaluation on separately acquired 100 and 200 ppm records yielded a total-concentration mean absolute error (MAE) of 31.25 ppm after EWMA smoothing across gas-exposure and clean-air windows. These results establish a proof of concept for progressive three-component concentration estimation in prescribed thermal-runaway-related mixtures using a compact two-sensor configuration with a combined nominal heater power of approximately 48 mW.