Accurate estimation of battery State of Health (SOH) is crucial in the battery monitoring and management process. Several methods have been proposed to model and estimate battery aging dynamics, either formally, model-based or data-driven. One key challenge in SOH modeling is the generality of the SOH modeling approach, which requires consideration of inherent dependencies among the various multidisciplinary stress factors involved. In this paper, we present an end-to-end self-supervised approach based on Convolutional Auto-Encoders (CAEs) for learning informative intermediate features from battery measurable properties such as voltage, current and temperature. We then employ the learned features to estimate the change in battery SOH by a light-weight feed-forward neural network. The learned features represent essential information in battery dynamics and surpass the human-engineered features in terms of correlation with the target SOH characteristic. Utilizing these representative features, our SOH estimation model yields 58.7% and 45.0% average performance improvement on two large battery datasets compared to the state-of-the-art machine learning methods.