In　this　study,　a　novel　method　for　predicting　hardness　of　ferromagnetic　alloy　based　on　the　magnetic　Barkhausen　noise　(MBN)　is　proposed.　A　set　of　new　frequency　features　of　MBN　and　a　new　hardness　prediction　method　are　proposed.　The　new　features　are　derived　from　the　first　and　second　derivative　of　the　auto-regressive　spectrum　of　MBN　signal.　The　new　automatic　hardness　prediction　method　include　Bag-of-Words,　principal　component　analysis　and　back　propagate　neural　network　optimized　by　ensemble　learning.　The　experimental　results　of　the　hardness　classification　show　that　the　new　features　are　superior　to　the　previous　features-the　misclassification　rate　using　the　new　features　is　less　than　0.67%,　while　the　misclassification　rate　using　the　previous　features　is　about　2%.　The　efficiency　of　the　new　method　is　also　proved　by　hardness　classification　experiment.　Compared　with　the　traditional　time-domain　method　and　the　previous　frequency　domain　method,　the　misclassification　rate　of　the　new　method　decreased　significantly　from　25%　to　less　than　1%.　In　addition,　the　new　method　is　highly　automatic,　so　it　is　more　versatile　than　manual　algorithms.　The　above　characteristics　make　the　proposed　new　method　suitable　for　predicting　the　hardness　of　ferromagnetic　alloys　in　practice.