Unexpected　bearing　failures　may　cause　unscheduled　downtime　and　economic　losses.　It　is,　therefore,　very　important　to　find　the　faults　symptoms　of　the　rolling　element　bearing　components.　Vibration　signal　of　fault　bearing　is　nonlinear　and　non-stationary　in　nature,　which　makes　the　stationary　assumed　methods　not　appropriate.　In　this　paper,　a　biphase　randomization　wavelet　bicoherence　method　is　introduced,　which　combines　benefits　of　the　wavelet　transform　and　the　bicoherence　analysis.　By　simultaneously　using　the　amplitude　of　the　continuous　wavelet　transform　and　biphase　information,　this　method　can　eliminate　the　spurious　bicoherence　coming　from　long　coherence　time　waves　and　non　phase　coupling　waves　efficiently.　Based　on　this　method,　two　quadratic　nonlinearity　features　are　proposed　for　fault　diagnosis　of　rolling　element　bearing.　At　the　same　time,　the　proposed　features　are　applied　to　the　real-world　vibration　data　collected　from　locomotive　roller　bearings　with　faults　on　inner　race,　outer　race　and　rollers,　respectively.　Experiment　results　demonstrate　that　the　performance　of　the　proposed　features　is　much　better　than　that　of　some　original　features.