Abstract: A novel method using wavelet-domain hidden Markov tree (HMT) model for on-line detection of abnormal process trends is presented in this paper where all scale coefficients and several selected large wavelet coefficients are taken into account. The way how to select the large wavelet coe fficients is also introduced. The experiments prove that the new method can reduce the ambiguities of identification and decreases the time spending on detecting abnormal process trends.