For realtime prediction on device degradation, the existing algorithms are difficult to update the trained model and the prediction results are easy to be distorted due to the outlier and noise. To solve these problems, a novel method named as online sequential regularized correntropy criterion extreme learning machine (OSRCCELM) is proposed to generate high robust prediction model and provide dynamic updating ability. Firstly, based on the regularized correntropy criterion ELM, the updating method for the model is realized. Secondly, the outlier and noise are detected with the dynamic Mestimator that is integrated with the error codebook. Finally, the corresponding influence of the outlier and noise are removed from the current model. Experiments using simulated data and CTR of optical couples show that OSRCCELM can achieve higher prediction accuracy without the effect of outliers and noises than other methods while provide accurate prediction with high speed.