SHShiyu Hu

Shiyu Hu

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Shiyu Hu has a total of 6 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Adaptive Online Discrete Degradation State Distribution Prediction Methodology for Cloud Computing Optical Fiber Ports

    by Yujie Mou, Dong Wang, Jie Liu, Wang Shi, Shiyu Hu, Ming Lu

    With the deep penetration of optical fiber communication technology in cloud computing environments, optical fiber switches and other devices have been widely deployed. A gradual decline in power of optical fiber ports will lead to signal transmission attenuation and even complete communication failure at a critical degradation point. If the failure time of optical fiber ports can be predicted in advance, preventive maintenance can be performed. Therefore, accurate remaining useful life (RUL) prediction of optical fiber ports becomes valuable. Based on the unique discrete reciprocal hopping characteristics of optical fiber port power data, this paper reports an adaptive on-line discrete degradation state distribution prediction methodology, which fills the gaps in modeling discrete degradation processes and RUL predictions. Firstly, optical fiber port degradation process with state discretization is probabilistically modeled. Then, the intractable problem of solving a discrete degradation state first hitting time distribution is transformed to the problem of solving a cumulative degradation data distribution through the first hitting time theorem. Following this, analytical expressions for the discrete degradation state first hitting time distribution are respectively derived by the sum of multinomial distributions and Central Limit Theorem. To realize RUL prediction of an on-line individual optical fiber port, a dynamic update mechanism is proposed to adaptively update distribution parameters by on-line individual optical fiber port degradation data. Finally, our proposed optical fiber RUL prediction methodology is applied to analyze industrial optical fiber port degradation data from Lenovo data centers. Industrial experiments show that our proposed methodology achieves a 95% prediction accuracy, which is significantly better than currently popular RUL prediction methods.

  • Weight-Based Large Margin Hyperdisks for Explainable Performance Degradation Modeling

    by Tongtong Yan, Dong Wang, Wang Shi, Shiyu Hu, Ming Lu

    Machine performance degradation modeling aims to construct a reliable health indicator (HI) for machine health evaluation. Deep learning based HIs are not reliable due to their inexplicability and strong volatility in assessing machine degradation. Recently, it has been an emerging topic for machine fault diagnosis based on hyperdisk modeling. Nevertheless, hyperdisks based performance degradation modeling for HI construction needs further exploration. In this paper, weight-based large margin hyperdisks for explainable machine degradation modeling are proposed to construct an explicit HI. The data preprocessing of the proposed methodology only requires simple transforms including Hilbert transform and Fourier transform. Then, based on squared envelope spectra, two hyperdisks are respectively developed to fully characterize spectral distributions under different machine health conditions. Further, a weight-based large margin hyperdisk optimization model is established for explainable machine degradation modeling, where an explicit and reliable HI can be well developed by fusing life cycle envelope spectra and optimized weight coefficients. Optimized weight coefficients that can reveal underlying informative frequencies provide an insightful perspective for the interpretability of machine performance degradation and its relevant HI. Two bearing run-to-failure datasets are studied to show the effectiveness of the proposed model and its advantages over other existing methods.