Abstract: To address the problem that most of the existing composite Web service models are of low accuracy on fault disgnosis， a novel composite Web service oriented fault diagnosis approach was proposed based on an improved hidden Markov model (IHMM). Firstly， HMM model was trained by using the processed multidimensional feature sequences. In this process， the BWbased methods were not used for parameters estimation， since inaccurate parameters would often resulted in due to the single observation. Instead， a Bayes estimation based method to gain more objective paratemeters was proposed. Finally， the probabilities of different fault types caused by the current feature sequence were computed. The one of the maximum probability was inferred as the ultimate fault type. Experimental results showed that the method was effective and efficent. Due to the high diagnostic rate and the low false rate， it was suitable for realtime fault detection in network environment.

Key words: HMM(hidden Markov model), composite Web services, fault diagnosis, Bayes estimation, feature sequences