Abstract: To address the false alarm issue caused by abnormal sensor fluctuations due to cavitation in natural gas desulfurization purification unit, this study proposes a multi-sensor data fusion technology based on an expanded time-convolutional neural network (ETCN) enhanced long short-term memory (LSTM) network. The ETCN-LSTM model integrates data from multiple sensors related to cavitation to predict foaming behavior. Results demonstrate that the ETCN-LSTM model effectively fuses multi-sensor data and accurately predicts foaming levels over time, with predictions showing strong alignment with actual values. Compared to the standard LSTM, the ETCN-LSTM reduces root mean square error (RMSE) and mean absolute error (MAE) by 12.0% and 26.4%, respectively, while maintaining low computational cost and improving long-term prediction stability.

Key words: vacuole phenomenon, monitoring and early warning, ETCN, LSTM networks, multi sensor data, production efficiency