Cross-timestep Fault Prediction with Imbalanced Data for Optical Modules in Internet Data Centers
Zuxu Pei, Tian Song, Chao Qun Wu, Shuye Yue, Yan Li, Xiangtao Hu · 2024
Optical module faults are among the most serious threats to Internet Data Centers (IDCs), which are crucial to a company’s data processing and information storage operations. Consequently, enterprises typically aim to precisely identify faulty optical modules within an extended preparation time, a process known as cross-timestep fault prediction. However, achieving this goal encounters several challenges, including insufficient effective data, long-distance dependency issues, and the problem of data imbalance. In this paper, we propose CTFP, a novel cross-timestep fault prediction method for optical modules in IDCs, which can accurately predict optical module faults twenty-four time steps in advance based on historical data. CTFP not only leverages the Digital Diagnostic Monitoring (DDM) data of optical modules as input, but also considers port-related data that may be affected by optical module faults. In addition, we have incorporated the attention mechanism into CTFP to capture the long-term trends of historical data, thereby mitigating the issue of long-distance dependency. Moreover, we designed an improved loss function that addresses the issue of data imbalance. Finally, we evaluate CTFP on industrial datasets collected from real-world internet data centers. The experimental results demonstrate that, compared to the state-of-the-art methods Bi-GRU and LSTM, CTFP increases the recall by at least 11% and 9% under various sample proportions. Notably, CTFP maintained a remarkably low maximum false positive rate of only 0.26%. In real-world conditions, where the number of optical modules often reaches the hundreds of thousands, maintaining a low false positive rate is imperative.