Bayesian online change point detection method for process monitoring

Aiming at the problem of a large amount of unlabeled observations collected in the industrial processes, an unsupervised Bayesian online change point detection method is adopted for fault detection. Firstly, a prior probability of fault occurrence is set based on the significance level. Secondly, the predictive distribution is calculated using the exponential family likelihoods as a new observation arrives. Finally, based on the observed data, a recursive message-passing algorithm is applied for calculating the fault occurrence probability at the current sampling point. The power of the Bayesian method for fault detection is tested in a numerical simulation and the Tennessee-Eastman (TE) process.