On-line detection of outliers in control process data based on autoregressive hidden Markov model with order self-learning
Zhizhong Mao · Control theory & applications · 2011
For the accurate online detection and collection of massive real-time data of a control process in strong noise environment,we propose an autoregressive hidden Markov model(ARHMM) algorithm with order self-learning.This algorithm employs an AR model to fit the time series and makes use of the hidden Markov model as the basic detection tool for avoiding the deficiency in presetting the threshold in traditional detection methods.In order to update the parameters of ARHMM online,we adopt the improved traditional BDT(Brockwell-Dahlhaus-Trindade) algorithm with double iterative structures,in which the iterative calculations are performed respectively for both time and order.To reduce the influence of outlier on parameter updating in ARHMM,we adopt the strategy of detection-before-update,and select the method for updating based on the detection results.This strategy improves the robustness of the algorithm.Simulation with emulation data and practical application verify the accuracy,the robustness and the property of online detection of this algorithm.Comparison between the traditional AR-model-based algorithm and the proposed algorithm shows the superiority of the proposed algorithm in outlier detection in industrial control processes.