Fault diagnosis of rotating machinery components using a deep kernel extreme learning machine under different working conditions

Deep-learning-based algorithms have shown great achievements in fault diagnosis of rotating machinery components in recent years. However, training deep networks is cumbersome and time-consuming. Furthermore, when working conditions change, their diagnostic performance degrades significantly. To address these problems, a novel deep kernel extreme learning machine (DK-ELM) is presented in this paper. First, we propose a discriminative manifold ELM auto-encoder (DM-ELM-AE), which exploits both the geometry of the training sample's marginal distribution and the label information. Then, multiple DM-ELM-AEs are stacked to form a deep feature extractor to extract discriminative and robust high-level features from vibration measurements automatically. Based on the extracted features, final fault pattern classification is carried out using a kernel ELM instead of a conventional ELM classifier. In this way, good diagnostic performance can be attained, regardless of working conditions. The experimental results show that the DK-ELM yields more promising results than other state-of-the-art algorithms in terms of diagnosis accuracy and adaptation ability to working conditions. It achieved 2% improvement in testing accuracy compared with the kernel ELM, and nearly 5% improvement compared with classical deep learning algorithms under different working conditions. In addition, it inherits the superb training efficiency from the ELM and is much easier to implement than deep learning algorithms.