Learning diverse causally emergent representations from time series data