Abstract:
Remaining useful life (RUL) prediction is crucial for ensuring the safety and reliability of industrial equipment. Although deep learning methods have proposed high-accur...Show MoreMetadata
Abstract:
Remaining useful life (RUL) prediction is crucial for ensuring the safety and reliability of industrial equipment. Although deep learning methods have proposed high-accuracy solutions for RUL prediction, a conflict arises between the limited edge computing resources and the significant memory requirements of deep models. This article proposes a binary temporal composite attention network (BTCAN) aiming to achieve deep model compression with acceptable performance degradation. First, a fully binary RUL prediction framework is proposed to achieve a high model compress ratio. Then, a binary composite trend attention (BCTA) module is proposed to allocate weight to multisource degraded features. Finally, a binary temporal convolution network (BTCN) is proposed to extract time-series information from degraded data. The proposed approach achieves about 28.7 times reduction in model memory footprint and 27 times improvement in inference efficiency while maintaining competitive accuracy. The BTCAN exhibits significant memory consumption advantages and competitive prediction accuracy compared to the state-of-the-art models.
Published in: IEEE Transactions on Instrumentation and Measurement ( Volume: 73)