This　paper　presents　a　fault　diagnosis　method　of　rotating　machinery　based　on　a　new　clustering　algorithm　using　a　compensation　distance　evaluation　technique　(CDET).　A　two-stage　feature　selection　and　weighting　technique　is　adopted　in　this　algorithm.　Feature　weights　are　computed　via　CDET　according　to　the　sensitivity　of　features　and　assigned　to　the　corresponding　features　to　indicate　their　different　importance　in　clustering.　Feature　weighting　highlights　the　importance　of　sensitive　features　and　simultaneously　weakens　the　interference　of　insensitive　features.　The　new　clustering　algorithm　is　described　and　applied　to　incipient　fault　and　compound　fault　diagnosis　of　locomotive　roller　bearings.　The　diagnosis　result　shows　the　algorithm　is　able　to　reliably　recognise　not　only　different　fault　categories　and　severities　but　also　the　compound　faults,　and　demonstrates　the　superior　effectiveness　and　practicability　of　the　algorithm.　Therefore,　it　is　a　promising　approach　to　fault　diagnosis　of　rotating　machinery.