It's somewhat analogous to a z-score, but for a multivariate situation. Each data point will have a "distance" from some multi-dimensional central point, or centroid.

It's used in multivariate procedures (such as MANOVA, or multiple regression, etc.) to identify outliers.