A model-free deep learning-based health prognosis methodology with epistemic and aleatoric uncertainties

Aleatoric uncertainty inherent in degradation data and epistemic uncertainty introduced by a research method directly affect the trustworthiness of health prognosis, specifically in terms of the state of health (SOH) and remaining useful life (RUL). In this paper, a deep learning methodology integrating Bayesian deep learning (BDL) and model-free approaches is proposed to eliminate the influence of uncertainties and improve the accuracy of the prediction results. First, an advanced dropout approach is introduced to quantify the epistemic uncertainty due to the neural network model parameters in a model-free manner. Then, a data augmentation approach and an arbitrary polynomial chaos expansion method are used in a model-free manner to quantify the aleatoric uncertainty inherent in the data distribution. Next, a model-free BDL model is constructed for health probability prognosis that incorporates two types of uncertainty. In addition, an incremental learning approach based on an asynchronous calibration strategy is presented to adaptively update the prediction model. To verify the effectiveness of the proposed method, lithium-ion battery SOH prognosis and turbofan engine RUL prediction experiments are conducted using degradation datasets. The results show that the proposed method can provide more accurate point estimations and more trustworthy interval estimations.