For tool wear in the cutting process, hidden Markov model (HMM) is used to recognize different tool wear states. And aiming at the local minimum defect easily trapped by Baum-Welch algorithm in HMM, an improved method of optimizing initial value of B in Baum-Welch by genetic algorithm is proposed to improve the recognition accuracy of HMM. The wavelet packet decomposition technology was used to extract the feature vectors from the output current signal of spindle motor during the cutting process. After coded by Lloyd algorithm, the feature vectors was regarded as observation sequence of HMM to recognize different tool wear states. Experimental results show that this method can effectively and accurately monitor milling cutter wears.