A Dual‐Channel Approach Integrating Multisensory Data for Remaining Useful Life Prediction of Induction Motors

ABSTRACT The accurate prediction of remaining useful life (RUL) for induction motors, particularly in environments characterized by complex and demanding operational conditions that induce gradual degradation, is pivotal for optimizing predictive maintenance strategies and ensuring system reliability. This paper introduces a novel dual‐channel deep learning architecture, that is, Transformer and convolutional neural parallel network (TCPN), specifically designed to address the challenges inherent in RUL estimation for induction motors. The TCPN framework integrates the synergistic capabilities of a Transformer and a depth‐wise separable three‐dimensional convolutional neural network (DS‐3DCNN) within a parallel architecture. In this configuration, the DS‐3DCNN effectively captures the spatiotemporal dependencies inherent in multisensory data by exploiting time‐frequency domain features extracted via continuous wavelet transform (CWT), while the Transformer model processes one‐dimensional handcrafted features from both time and frequency domains, thereby capturing long‐term temporal dependencies. The independent feature representations from each channel are subsequently fused to deliver highly accurate RUL predictions. Extensive validation on operational data from eight induction motors demonstrates that the TCPN model significantly outperforms recent deep learning models commonly used for RUL prediction, delivering superior accuracy and enhancing predictive maintenance through improved reliability and system safety.