Data Driven Prognostics With Lack of Training Data Sets

Abstract

Data-driven prognostics typically requires sufficient offline training data sets for accurate remaining useful life (RUL) prediction of engineering products. This paper investigates performances of typical data-driven methodologies when the amount of training data sets is insufficient. The purpose is to better understand these methodologies especially when offline training datasets are insufficient. The neural network, similarity-based approach, and copula-based sampling approach were investigated when only three run-to-failure training units were available. The example of lithium-ion (Li-ion) battery capacity degradation was employed for the demonstration.

Review on Health Management System for Lithium-Ion Batteries of Electric Vehicles

“…The data-driven prognostics typically require sufficient offline training datasets for accurate remaining useful life for engineering products [84]. Data-based approaches of battery modeling use the battery's SOH data, which can be measured through advanced sensor technology to extract effective feature information, and construct the degradation model to predict RUL [74].…”