Free text and hand-written reports are losing ground to digitization fast, however many hours of effort are still lost across the industry to the manual creation and analysis of these data types. Work orders in particular contain valuable information from failure rates to asset health, but at the same time present operators with such analytical difficulties and lack of structure that many are missing out on the value completely. This research challenges the current mainstream practice of manual work order analysis by presenting a methodology fit for today’s context of efficiency and digitization.
A prototype text mining software for work order analysis was developed and tested in a user-oriented approach in cooperation with industrial partners. The final prototype combines classical machine learning methods, such as hierarchical clustering, with the operator’s expert knowledge obtained via an active learning approach. A novel distance metric in this context was adapted from information-theoretical research to improve clustering performance.
Using the prototype tool in a case study with real work order data, analytical effort for certain datasets was reduced by 90% - from two working weeks to a day. In addition, the active learning framework resulted in an approach that end users described as "practical" and "intuitive" during testing. An in-depth review was also conducted regarding the uncertainty of the results – a key factor for implementation in a decision-making context.
The outcomes of this work showcase the potential of machine learning to drive the digitization of not only new installations, but also older assets, where as a result the large amount of unstructured historical data becomes an advantage rather than a hindrance. User testing results encourage a wider uptake of machine learning solutions in the industry, and particularly a shift towards more accessible in-house analytical capabilities.