| Abstract: |
Maintenance in discrete manufacturing is still dominated by reactive repair and calendar-based preventive replacement, a combination that produces unpredictable stoppages, inflated spare-part inventories and persistent erosion of overall equipment effectiveness (OEE). This paper reports an empirical evaluation of an artificial-intelligence-based predictive maintenance (AI-PdM) framework deployed across forty-two machining, press and robotic welding cells in a component plant over twenty-four months. Multisensor telemetry-triaxial vibration, spindle current, bearing temperature, acoustic emission and hydraulic pressure, acquired over 198 channels in twelve-minute windows-was streamed through an edge gateway into a time-series data lake. A gated ensemble coupling a bidirectional long short-term memory network for remaining-useful-life regression with a gradient-boosted classifier for short-horizon failure detection was trained on 18,470 machine-weeks of labelled history and validated on a held-out 650 machine-weeks. Against the pre-deployment baseline the framework reduced unplanned downtime by 37.7 percent, from 11.4 to 7.1 hours per machine per month; raised mean time between failures by 28.6 percent, from 168 to 216 hours; cut mean time to repair by 31.1 percent, from 4.37 to 3.01 hours; and lifted OEE from 71.3 to 84.9 percent. Detection precision, recall and F1-score were 0.90, 0.92 and 0.91, with a receiver operating characteristic area of 0.96 and a median warning lead time of 214 minutes. Paired t-tests confirmed that every OEE component improved significantly at p < 0.01. The evidence shows that AI-PdM delivers warning long enough to be acted upon, and that reliability gains convert into measurable productivity and cost benefits rather than merely relocating maintenance effort earlier. |
