Mining procedural non-compliance in enterprise asset management data: a CRISP-DM study with association rules on IBM Maximo work orders

Abstract

Enterprise asset management platforms accumulate millions of maintenance transactions, yet the way users actually populate those records is almost never audited after go-live. This paper treats procedural non-compliance as a mineable object and reports a complete CRISP-DM study on the work order module of an IBM Maximo installation run by a rail transport company. Seven relational tables were extracted from IBM DB2. After filtering to corrective (CM) and preventive (PM) work, 12,368 orders remained, 5,790 corrective and 6,578 preventive. Completeness and integrity were profiled first; child tables were then aggregated to order level and recoded into a binary transaction matrix carrying eleven items, two contextual and nine describing distinct forms of procedural failure. Apriori and FP-Growth were executed over four scenarios at support 0.10, confidence 0.70 and lift above 1.0, under a two-step ablation designed to neutralise items that are almost always present. Every reported rule was additionally submitted to a one-sided Fisher exact test, corrected for multiple testing across all 155 tested rules, and each syndrome was re-estimated in a multivariable logistic model so that the items are adjusted for one another, and both thresholds were varied over a sixteen-cell grid. The two algorithms returned numerically identical rule sets, which establishes robustness, while Apriori ran 23 to 98 times faster on this low-dimensional item space; because every absolute time lies below one second, that ordering is reported as a property of small binary matrices rather than as a general ranking of the two algorithms. Two behavioural syndromes were isolated. Corrective work fails on timing: 47.0% of CM orders close within two minutes and 98.0% of those carry an earlier execution stamp, at lift 1.350. Preventive work fails on scheduling: overlapping labour combined with backdating implies late completion at 96.8% confidence and lift 1.759. Both syndromes survive Benjamini-Hochberg correction at q < 0.001, retain adjusted odds ratios of 49.3 and 69.7 in the logistic models, and reappear at every one of the sixteen threshold combinations tested, where they are also the only rules common to all cells. The extracted rules were converted into platform validation designs, a blueprint checklist and a rule engine that states each finding in natural language.