Aimed at feature extraction in machinery fault diagnosis, this paper proposes a new fault feature extraction method based on multi-wavelet coefficients. The original vibration signals of each fault category are decomposed into time-frequency representations using multi-wavelet transform. Then the maximum, minimum, mean and standard deviation of the multi-wavelet coefficients in each subband are calculated and used as the feature vector. The support vector machine method is used for machinery fault classification. Experiments are conducted on the real vibration signal of the roller bearing with normal conditions, inner fault, ball fault and outer fault. The experimental results indicate that the proposed approach can reliably identify the different fault categories, works better than the single wavelet method, and thus has potential for machinery fault diagnosis.