Adaptive monitoring of the process operation based on symbolic episode representation and hidden Markov models with application toward an oil sand primary separation

This paper presents a novel procedure for classification of normal and abnormal operating conditions of a process when multiple noisy observation sequences are available. Continuous time signals are converted to discrete observations using the method of triangular representation. Since there is a large difference in the means and variances of the durations and magnitudes of the triangles at different operating modes, adaptive fuzzy membership functions are applied for discretization. The expectation maximization (EM) algorithm is used to obtain parameters of the different modes for the durations and magnitudes assuming that states transit to each other according to a Markov chain model. Applying Hamilton's filter, probability of each state given new duration and magnitude is calculated to weight the membership functions of each mode previously obtained from a fuzzy C-means clustering. After adaptive discretization step, having discrete observations available, the combinatorial method for training hidden Markov models (HMMs) with multiple observations is used for overall classification of the process. Application of the method is studied on both simulation and industrial case studies. The industrial case study is the detection of normal and abnormal process conditions in the primary separation vessel (PSV) of an oil sand industry. The method shows an overall good performance in detecting normal and risky operating conditions.