Improved TabM LightGBM Model for Construction Hoist Vibration Diagnosis Using Motor Current and Acceleration Data
Authors
Files
Abstract
Construction-hoist vibration diagnosis must distinguish mechanical abnormalities from electrical demand and directional motion changes caused by payload and travel conditions. This study evaluates a load-coupled TabM-LightGBM pipeline using motor current, three acceleration directions and operating metadata. Two-second windows sampled at 512 Hz yield 24 engineered signal descriptors and four context descriptors. A healthy-state regression fitted on training episodes supplies 24 condition-adjusted residuals; eight parameter-sharing TabM members transform the resulting 52 features. Their hidden- state mean and dispersion join the original descriptors in a 116-feature LightGBM input. A separate calibration partition selects probability temperature before confidence-ranked review. The controlled numerical reference contains 288 independent episodes across six load-centered groups and four diagnostic states, giving 1,728 windows. Two held-out groups supply 576 test windows. The complete pipeline attains 98.09% accuracy, macro-F1 0.9809, abnormal-state recall 97.92% and healthy-state false-alarm rate 1.39%; its median descriptor-to-probability inference time is 0.551 ms. Random forest and standalone TabM obtain higher reported macro-F1 values of 0.9965 and 0.9913, respectively, and a separately trained acceleration-only pipeline obtains 0.9879. At the selected full checkpoint, axial-feature interruption reduces macro- F1 to 0.6048, while descriptor noise and load-metadata bias yield 0.9281 and 0.9390. A 115-window confidence-priority list captures nine of eleven test errors. These results characterize source dependence, calibration and review yield in a controlled signal generator, with independent operating data needed to establish field applicability.
