AI-DRIVEN PREDICTIVE MAINTENANCE FOR MOTOR DRIVES IN SMART MANUFACTURING A SCADA-TO-EDGE DEPLOYMENT STUDY

Abstract

This quantitative study evaluated AI-driven predictive maintenance for motor drives in smart manufacturing by comparing three deployment architectures: SCADA-only, edge-only, and hybrid SCADA-to-edge fusion. A longitudinal multi-asset dataset was analyzed from 48 motor drives monitored over 16 weeks, representing 18,720 operating hours, 12,614,380 SCADA tag records, and 3,456,000 edge analysis windows. Outcomes were defined using tiered event labels, including 37 Tier-1 confirmed failures, 64 Tier-2 verified defect findings, and 142 Tier-3 operational abnormality episodes. Time-based evaluation and motor-drive clustering controls were applied, and performance was assessed under a fixed alert-budget policy using event detection, precision, false alarm density, and lead time outcomes, alongside deployment feasibility metrics. Compared with SCADA-only, edge-only deployment improved event detection (odds ratio = 1.62, p = .014) and increased precision from 0.41 to 0.53 (p = .006), while reducing false alarms from 1.48 to 1.31 per 100 operating hours (p = .041). Hybrid fusion produced the strongest predictive outcomes, increasing event detection (odds ratio = 2.08, p < .001), raising precision to 0.58 (p < .001), and lowering false alarms to 1.27 per 100 operating hours (p = .018). Median lead time increased from 18.6 hours (SCADA-only) to 31.4 hours (edge-only) and 37.9 hours (hybrid) (p < .01). Deployment tradeoffs were quantified: inference latency increased from 18 ms per window (SCADA-only) to 122 ms (edge-only) and 141 ms (hybrid), while bandwidth use was reduced by 96.1%–96.8% in edge and hybrid configurations through feature-level reporting. Overall, SCADA-to-edge fusion yielded the most stable and effective predictive maintenance performance across operating regimes with manageable system overhead.