Soft sensor design and fault detection using Bayesian network and probabilistic principal component analysis

• Published in: Journal of advanced manufacturing and processing 2019-10, Vol.1 (4), p.n/a
• Main Authors: Mohammadi, Ahad, Zarghami, Reza, Lefebvre, Dimitri, Golshan, Shahab, Mostoufi, Navid
• Format: Article
• Language: English
• Subjects: Artificial intelligence; Bayesian analysis; Bayesian network; Chemical reactions; Confidence intervals; Electric power; Engineering Sciences; Fault detection; Gas streams; gas sweetening process; Hydrogen sulfide; Principal components analysis; Probabilistic models; probabilistic principal component analysis; Quality assessment; Sensors; soft sensor; Statistical analysis; Sweet gas; Tennessee‐Eastman challenge process; Uncertainty analysis
• ISSN: 2637-403X; 2637-403X
• DOI: 10.1002/amp2.10027

Bayesian network (BN) and probabilistic principal component analysis (PPCA) are powerful tools in artificial intelligence. In this study, two intelligent soft sensors, designed based on BN and PPCA, were proposed for fault detection and data prediction in chemical processes. A gas sweetening process was considered as a case study in which the existing data for H2S concentration in sweet gas stream were incomplete. In order to detect faults during the operation, a soft sensor was designed using BN and PPCA for predicting H2S concentration in the sweet gas stream. Then, another soft sensor was developed based on the BN for fault detection, considering the Gaussian mixture model with hidden nodes. The Tennessee‐Eastman challenge process was used to assess the efficiency of the fault detection method. Results showed that the predicted values of the soft sensors are very close to real data. The fault occurring in this process was detected in early stages, demonstrating the good performance of the proposed fault detection system. It was also shown that the performance of the BN is better than the PPCA in prediction of incomplete data. Moreover, the confidence interval can be evaluated for the predicted values when using the BN. An uncertainty analysis was performed to assess the quality of predicted data, and it was observed that the error magnitude of predicted data is smaller when using the BN compared with the PPCA. In particular, the results showed that the BN model is capable of estimating H2S concentration with nearly 96.1% accuracy, whereas the accuracy of the PCA‐based method was 93.2%.