An improved OS-ELM based Real-time prognostic method towards singularity perturbation phenomenon

Abstract

Reliable and comprehensive historical data are fundamental for data-driven based reliability prognostic. In practice, the degradation process of electronic products is usually accompanied with singularities caused by intermittent or transient interference, in the past, singularities data often need to be detected in advance or even be regarded as noise being eliminated or ignored, before estimation process. Such traditional ways would lead to waste of time and resources, decrease the accuracy of reliability prognostic as well. Real-time models are being introduced to capture complete degradation trend on time, overcome the limitations of early detection and increase forecasting performance, recently. In this study, a novel real-time fusion approach based on spline function and optimized online sequential extreme learning machine (OS-ELM) is developed to address the above singularity perturbation issue. As the approximation model, the cubic non - polynomial spline function could track the singularities’ feature by window-slide dataset and trigonometric polynomial, form the initial prediction cell frames. ELM is introduced as an effective estimation model with simple structure to save computing resources. To address the uncertainty issue caused by its inherent randomness of ELM model. Quantum behaved particle swarm optimization (QPSO) and optimally pruning method were applied to optimize the parameters of ELM hidden network, and redundant network neurons, respectively. In the verification stage, different experimental results show that the proposed fusion model could capture singularities’ characteristics, achieve better prognostic performance in terms of accuracy.