Learning Robust and Consistent Time Series Representations: A Dilated Inception-Based Approach