Envelope-Spectrum Kurtosis of Wavelet-Packet Subbands for Motor-Current Bearing Diagnosis: Recording-, Condition- and Specimen-Level Evaluation

In motor current signature analysis, bearing-related components are weak and their spectral location varies with operating conditions and fault morphology. This paper proposes KESK, the kurtosis of the Hilbert-envelope spectrum of the most impulsive wavelet-packet subband in each current window, and evaluates it as one additional feature to a 38-feature baseline using one phase current from 20 Paderborn bearings (11,173 windows). With supervised energy selection inside training folds, new recordings, unseen operating conditions and unseen bearings were evaluated separately. On the primary recording-level partition, XGBoost macro-F1 increased by 0.002 with KESK, 0.003 with a training-selected fixed node and 0.003 with a DC-free KESK variant; over ten additional partitions, the mean increments were 0.002, 0.003 and 0.007, respectively. With an operating condition held out, the fixed node reduced macro-F1 (−0.014; 95% CI −0.024 to −0.005) and the DC-free variant increased it (+0.016; 0.008 to 0.023). With bearings held out, macro-F1 was 0.48–0.49 for every representation. On nine bearings excluded from all design decisions, a pre-specified test did not confirm the DC-free advantage (−0.003; −0.012 to 0.005). Envelope-spectrum kurtosis therefore provides a small but reproducible amount of complementary information at negligible additional computational cost, although baseline composition and validation level have a substantially greater influence on estimated performance.