Self-labeling techniques for semi-supervised time series classification: an empirical study