Physics-informed unit-normalized selective embedding with harmonic loss for fault identification using vibration and acoustic signals

Induction motors generate vibration and acoustic signals from mechanical imbalance, electromagnetic forces, and structural transmission through the machine housing. Interpreting harmonic fault signatures from these responses becomes difficult when different types of sensors are used due to bandwidth and signal scaling limitations. This work investigates how physics-informed signal representation can improve fault diagnosis using both industrial sensors and cell phone sensors. A unit-consistent normalization approach is introduced to allow comparison between vibration acceleration and acoustic pressure signals by converting both signals into equivalent vibration velocity representations. Based on this formulation, a structured signal processing framework named Unit-Normalized Selective Embedding with Harmonic Loss (UNSE-HL) is proposed, combining unit normalization, selective embedding, and a physics-informed harmonic loss function to improve the representation of modulation patterns during diagnosis. Experiments are conducted using the induction motor datasets collected under multiple operating conditions with both structurally mounted sensors and cell phones. Envelope spectral analysis shows that rotor harmonics are observable across cell phone sensors after normalization; this observability does not by itself establish classification reliability, which is evaluated separately. Time-frequency representations combined with the proposed embedding strategy improve diagnosis accuracy and classification stability relative to conventional signal representation approaches on the datasets and devices evaluated in this study. These findings highlight the importance of signal transmission physics in condition monitoring and indicate that, for the specific smartphone sensors and operating conditions tested here, appropriate preprocessing improves the usability of low-bandwidth sensors for fault diagnosis.