Industrial processes provide lots of operational data on different timescales. Those data are well-structured and used now for daily control, longer-term management and forensics. We propose to pre-process that data and treat them the way the natural language processing is done - first, in order to find common ways the process is controlled. Such knowledge can then be used in prediction or early detection of faults, or necessary manufacturing shifts. Gas transmission operational data are considered here as the live example.