A multi layer kernel learning machine combining kernel principle comp onent analysis with support vector machine is presented for processing no nst ationary vibration signals of diesel cylinder head and dealing with the linear nonseparating problem of multiple faults of the valve train. The method, firs tly, extracts non1i near principle components from original features using kernel principle componen t analysis. Then an SVMbased multiclass classifier is constructed by “oneagainstrest” algorithm. Finally, multiple faults of the valve train ar e diagnosed quantitatively. The experimental results indicate that six state s of the valve train can be recognized exactly umder conditions of a small samp le size, and both precis ion of recognition and test speed are superior to the multiclass SVM method in practice.