A new machine condition monitoring method based on likelihood change of a stochastic model

In industry, a machine condition monitoring system has become more important with ever-increasing requirements on productivity and cost saving. Although researches have been very active, many currently available intelligent monitoring methods have common drawbacks, which are the requirement of defect model for every interested defect type and inaccurate diagnostic performance. To overcome those drawbacks, authors propose a new machine condition monitoring method based on likelihood change of a stochastic model using only normal operation data. Hidden Markov model (HMM) has been selected as a stochastic model based on its accurate and robust diagnostic performance. By observing the likelihood change of a pre-trained normal HMM on incoming data in unknown condition, defect can be precisely detected from sudden drop of likelihood value. Therefore, though the types of defect cannot be identified, defects can be precisely detected with only normal model. Defect models can also be used when defect data are available. And in this case, not only the precise detection of defect but also the correct identification of defect type is possible. In this paper, the proposed monitoring method based on likelihood change of normal continuous HMM have been successfully applied to monitoring of the machine condition and weld condition, proving its great potential with accurate and robust diagnostic performance results.