Monitoring combustion systems using HMM probabilistic reasoning in dynamic flame images

In this paper, a novel method of on-line flame detection in video is proposed. Processing the data generated by an ordinary camera monitoring scene, it aims to early detect the current state of the combustion system and prevent the system from further degradation and occurrence of failure. Due to the dynamic change of the combustion system, the turbulent flame flicker produces images with different spatial and high temporal resolutions. The proposed method consists of hidden Markov model (HMM) and multiway principal component analysis (MPCA). MPCA is used to extract the cross-correlation among spatial relationships in the low dimensional space while HMM constructs the temporal behavior of the sequential observation. Although the prior process knowledge may not be available in the operation processes, the probability distribution of the normal status can be trained by the images collected from the normal operation processes. Subsequently, monitoring of a new observed image is achieved by a recursive Viterbi algorithm which can find the transition state sequence from series of observed image data. The proposed method, like the philosophy of traditional statistical process control, can generate simple probability monitoring charts to track the progress of the current transition state sequence and monitor the occurrence of the observable upsets. The advantages of the proposed method, data from the monitoring practice in the real combustion systems, are presented to help readers delve into the matter.