Discovering critical KPI factors from natural language in maintenance work orders

Optimizing maintenance practices is a continuous process that must take into account the evolving state of the equipment, resources, workers, and more. To help streamline this process, facilities need a concise procedure for identifying critical tasks and assets that have major impact on the performance of maintenance activities. This work provides a process for making data investigations more effective by discovering influential equipment, actions, and other environmental factors from tacit knowledge within maintenance documents and reports. Traditional application of text analysis focuses on prediction and modeling of system state directly. Variation in domain data, quality, and managerial expectations prevent the creation of a generic method to do this with real industrial data. Instead, text analysis techniques can be applied to discover key factors within a system, which function as indicators for further, in-depth analysis. These factors can point investigators where to find good or bad behaviors, but do not explicitly perform any anomaly detection. This paper details an adaptable procedure tailored to maintenance and industrial settings for determining important named entities within natural language documents. The procedure in this paper utilizes natural language processing techniques to extract these terms or concepts from maintenance work orders and measure their influence on Key Performance Indicators (KPIs) as defined by managers and decision makers. We present a case study to demonstrate the developed workflow (algorithmic procedure) to identify terms associated with concepts or systems which have strong relationships with a selected KPI, such as time or cost. This proof of concept uses the length of time a Maintenance Work Order (MWO) remains open from creation to completion as the relevant performance indicator. By identifying tasks, assets, and environments that have significant relevance to KPIs, planners and decision makers can more easily direct investigations to identify problem areas within a facility, better allocate resources, and guide more effective analysis for both monitoring and improving a facility. The output of the analysis workflow presented in this paper is not intended as a direct indicator of good or bad practices and assets, but instead is intended to be used to help direct and improve the effectiveness of investigations determining those. This workflow provides a preparatory investigation that both conditions the data,