earticle

영어

Deep learning research to analyze industrial time-series data has been an active research topic. Recent studies have attempted to borrow the models for natural language processing(NLP) to handle time dependency issues. However, industrial data have different properties compared with NLP data: strongly dependent on a time axis. Moreover, because industrial information is continuously accumulating while the machine is running, it has a much longer sequence than other sequential data. In this study, we compare the performance of widely used natural language models, LSTM and Transformer, on such long-time series industrial data. For comparison, we performed experiments to detect an attack on a water treatment management system and to predict traffic flow on a highway. We confirmed that the Transformer using the attention mechanism showed better performance than the LSTM.