, Senthil K. Chandrasegaran, Xiaoyu Zhang, Kwan-Liu Ma
Maintenance and error logs for machines in manufacturing organizations are typically written as informal notes by operators or technicians working on the machines. These logs are written using a combination of common language and internally-used abbreviations and jargon. Due to inconsistencies in the terminology used during error logging and in identifying root causes of issues, the data needs to be cleaned before automated analyses can be effectively used. This can require a human to go through and clean/tag the data, disambiguate multiple terms, and sometimes assign additional tags to the data objects to aid automated classification. With some organizations storing over a million records of legacy maintenance report data, this is not entirely feasible. We introduce a visual analytic approach to help analysts sift through such heterogeneous datasets so that the inconsistent data can be tagged and categorized with minimum manual effort. Though such data typically includes metadata such as date, time, severity, machine IDs, etc., in this paper we focus on the manually-entered text descriptions. We use metrics such as word occurrence frequency and information-theoretic metrics to visually highlight common and uncommon issues and fixes that occur in the maintenance logs. We illustrate our approach with data from industry and discuss future research directions to address scalability, metadata, and other approaches for grouping similar logs.
Proceedings of the 11th Model-Based Enterprise Summit
March 30-April 3, 2020
Model-Based Enterprise Summit
maintenance, manufacturing, Natural Language Processing, visulization