Multiscale Bidirectional Temporal Convolutional Framework for Industrial Ketone Organic Solvent Detection Based on Semiconductor Gas Sensor
Hao Wang, Yong Zhao, Zhenyu Yuan, Hanyang Ji, Fanli Meng
Abstract
Acetone and butanone, key ketone organic solvents in industrial electronics manufacturing, with their flammable and explosive properties pose risks at specific concentrations upon evaporation. Accordingly, accurate detection of ketone gases using stable semiconductor gas sensors holds critical significance in electronics manufacturing. However, inaccurate ketone gas detection through contemporary intelligent recognition methods is hindered by the limited differentiation in semiconductor gas sensor signals for ketone gases, especially in mixed gas environments. To address this, we propose a novel data-driven solution—the multiscale bidirectional temporal convolutional networks (Ms-BiTCN) framework, integrated with time–frequency analysis. Dynamic temperature modulation and variational mode decomposition Hilbert–Huang transform (VMD-HHT) are employed to transform gas sensor signals into the time–frequency domain. Subsignal frequency–domain data determines gas types, while subsignal time–domain data quantifies gas concentration. Experimental validation, including acetone, butanone, and acetone–butanone mixtures, demonstrates Ms-BiTCN’s superior performance over state-of-the-art methods, achieving a ketone organic recognition precision of 96.3% and a concentration quantification root mean square error (RMSE) of 2.09 ppm. This innovative framework has the potential to greatly enhance safety and efficiency in integrated circuit lithography by offering accurate gas level tracking and analysis.