Vibration Severity Criteria Development for MEMS-Enabled Data-Driven Condition Monitoring of Electric Motors

The rapid proliferation of artificial intelligence in industrial diagnostics has fundamentally altered predictive maintenance for rotating machinery. However, even with all its benefits, operational reliability for applications using it remains constrained by a "cold start" dependency on prolonged training periods to establish reference baselines. This introduces a pronounced vulnerability known as baseline poisoning, wherein latent defects present during the initial learning window bias the characterization of nominal behavior, “teaching” the models that high levels of vibration are standard for the machine. Such normalization of deviance compromises diagnostic sensitivity, delaying the detection of early-stage mechanical degradations. For industrial motors, while the ISO 20816-3 standard offers normative thresholds to mitigate this initialization gap, its absolute limits trace back to actuarial machinery studies from the first half of the 20th century, scaled via a rigid mathematical geometric progression. This legacy methodology creates a significant disconnect when applied to the structural dynamics of contemporary motors and modern MEMS sensing technologies of the 21st century. Echoing the recent methodological shift seen in standards like ISO 20816-5, which transitioned from fixed multipliers to statistical probability models derived from large-scale asset databases, this study proposes an empirical framework for establishing vibration severity thresholds tailored specifically to contemporary hardware and computational capabilities. Utilizing a robust triaxial dataset from over 49,000 induction motors, the research employs a multidimensional analytical approach across nine distinct operational categories. By permuting rated power and rotational speed, the study accounts for the physical scaling laws governing electric motors.