I’m trying to use previous work order descriptions, locations, and purchase information to develop failure codes for years worth of historical data. My first attempt was using common words (pump, piping, etc.) and just searching the array of work orders and labeling, then diving down a level for more specific component failures. This is very time consuming and since technicians write work orders differently it’s not the most accurate.

Second idea was to use LDA for untrained pattern classification and then choosing those patterns by their failure codes. This didn’t work for various reasons, namely that tokenizing failure data wasn’t very fruitful.

Is there some pre trained data out there that could assist in labeling these failure occurrences? Or any other idea that I’m not thinking of?