Text-sensor fusion for industrial process data.

Most modern processing plants have large numbers of sensors giving information on multiple features of parts of each piece of equipment at every part of the plant. This results in very large sets of data updated very frequently. One would like to use this data to identify and/or predict "interesting" events in the plants, such faults or periods of inefficient performance. Unfortunately, there are no labels to tell what is going at any particular time from the data, so conventional supervised machine learning cannot be used. Even if there were experts who could label the data, they would be plant engineers whose time would be too valuable to deploy on a data labeling activity.
Fortunately, shift engineers keep log files in which events and observations of the shift are recorded, often with a time stamp. The innovation of this project is to try to use the log files to generate labels for the sensor data. This requires two stages. First, key activities must be identified in the log files. Second, an activity from the log file must be associated with some features in the sensor data at a particular interval of time. If this can be done, the generated labels could be used in supervised machine learning to identify properties of the current state of the plant, or predict the future state.