Steel making industries exhibit extreme working conditions characterized by high temperature, pressure, and production speed as well as intense throughput. Due to high economic and energy investments of the overall production process, an intense and expensive preventive maintenance program is adopted to avoid breakdowns. Steel making process would greatly benefit from a predictive maintenance module able to detect incoming faults from data process analysis. However, due to intense preventive maintenance, available data recording process operations enclose only a few samples of fault events, avoiding the efficient application of classical data driven anomaly detection models. In an attempt to overcome the above mentioned limits, we report the outcome of an industrial research project on data-driven anomaly detection in a steel making production process. The study assesses a fault detection strategy for rotating machines in the hot rolling mill line: we developed an automatic two-step strategy, which combines two statistical methods over the available data set: more precisely, the combination of Re-weighted Minimum Covariance Determinant estimator and Hidden Markov Models helped identify working conditions in a drive reducer of a hot steel rolling mill line and automatically isolate signs of decreasing performance or upcoming failures. The proposed strategy has been validated on real data collected in a steel making plant placed in the South of Italy.