MRPM: Multistep Robust Prediction Machine for Degradation Time Series Projection

Currently, since electric systems are developed with extremely fast speeds and operated under variable environments, existing residual useful life (RUL) estimation methods fail to meet the requirements of both time complexity and case flexibility. Although online RUL estimation has shown great potential for performing accurate predictions, these methods have many limitations regarding probability-based parameter assignment and local similarity measurement with least mean squares (LMS). To overcome these problems, we proposed a novel multistep robust prediction machine (MRPM). The MRPM combines a recurrent neural network (RNN) and an extreme learning machine (ELM) to link historical information and estimated values under the RNN structure with a simple prediction unit used by the ELM. In addition, an error codebook (ECB) is introduced with a correntropy-based similarity measurement to avoid the influence of noise and outlier interruption from measured signals. Finally, a dynamic adjustment process is developed to gain flexibility under variable device operation conditions. It is demonstrated that the MRPM is not influenced by Gaussian noise and other noise with even probability distributions. Even under arbitrary noise, the MRPM can constrain its influence to an acceptable range in most cases. Experiments show that the MRPM has better performance than state-of-art methods and can provide accurate RUL estimates of health conditions for IGBT modules.