Machine learning condition monitoring systems for rotating machinery fault detection

The vibration-severity alarms that dominate industrial practice indicate that a rotating machine has deteriorated, not what has gone wrong. This laboratory proof-of-concept study examines whether a machine learning pipeline can identify the mechanism and estimate the imposed severity of seeded faults from multi-axis vibration data, and whether it can operate within an Industrial Internet of Things architecture. An unloaded rig comprising a 0.37 kW induction motor driving a 20 mm shaft on 6204 ZZ bearings was instrumented with a piezoelectric accelerometer on the motor housing and a tri-axial MEMS accelerometer on the bearing housing, sampled at 25.6 kHz. Ten machine states (a healthy baseline and three severity levels each of bearing contamination, mass unbalance and shaft misalignment) were recorded on three assembly replications, giving 30 recordings, which are the independent observations, segmented into 3,570 overlapping one-second windows. Sixty descriptors were reduced to eighteen, and six classifiers were compared under nested leave-one-replication-out validation with recording-level partitioning. Gradient-boosted trees achieved a window-level accuracy of 98.9% (fold-level 95% confidence interval 97.9%–99.9%, indicative only), a macro-averaged F1 score of 0.988 and a Matthews correlation coefficient of 0.985, and classified all 30 recordings correctly by majority vote. A one-dimensional convolutional network reached a mean accuracy of 99.1%; the two models were not compared statistically. An exploratory univariate RMS-threshold rule applied to the recorded instrument output classified only 3 of 9 healthy sets correctly, because the campaign-specific healthy baselines spanned a sevenfold amplitude range; this failure does not depend on the uncalibrated scale of the bearing channel. Whether campaign-specific characteristics contribute to fault-family discrimination has not been excluded, and the highest-ranked feature lies beyond the rated bandwidth of the bearing sensor. The pipeline executed in 11.3 ms per window on a laptop-class processor and delivered results to a ThingsBoard dashboard with an approximate median end-to-end latency of 1.4 s and no loss of classification records. These results indicate laboratory feasibility of fault-type classification on a single unloaded rig at one speed; generalisation to industrial machines, other operating conditions and lower-cost acquisition hardware was not tested.