(614d) Hidden Markov Model Based Fault Diagnoser Using Binary Alarm Signals with Estimated Confidence Levels

Alarm systems play an important role in ensuring safe operation by alerting the operators when process deviations are encountered. Operators respond to alarms based on guidelines developed through the alarm rationalization process, which is the documentation of the possible causes for each of the alarms configured [1]. During an abnormal event, a large number of alarms may be triggered simultaneously making it difficult for the operators to troubleshoot and respond. In the past, several incident investigation reports have cited inadequate alarm system design as a major contributing cause for chemical incidents. Industries typically rely on qualitative methods for alarm systems design [1, 2], which are not sufficient to capture the process dynamics due to high interaction among the variables in most chemical plants arising from the complex nature of the process. Motivated by these concerns, a diagnoser is developed to extract information from the sequence of alarms generated to identify the cause of the abnormal event using a statistical approach. We have also proposed a method to determine the confidence level of the diagnoser.

A statistical model, in particular, the hidden Markov model (HMM) based approach that uses only the binary alarm signals as input information for fault diagnosis is proposed. The probabilistic framework of HMM allows to capture the stochastic elements such as sensor noises, process disturbances, model uncertainties and fault magnitudes that are prevalent in chemical systems. Individual HMMs are modeled for each fault identified through the safety review process and are trained using the Baum-Welch algorithm. The measurement alarms are modeled as the outputs of the HMMs. The probability of emitting the given alarm sequence from each of the HMMs is used to build the likelihood ratios based diagnoser. The theory of distinguishability of HMMs presented in [3-5] is used to calculate the confidence level of this diagnoser.

The proposed diagnoser was tested on the industrial case study: Tennessee Eastman Process [6]. Since the performance of the data driven diagnoser will depend heavily on the quality and amount of data available, extensive simulations were performed to obtain possible alarm sequences for the identified faults with the help of a closed-loop simulator. While collecting the data, a wide range of fault magnitudes and measurement noises were considered for each of the five identified faults. The data consisted of 300 alarm sequences for training and 100 sequences for testing the HMMs. The diagnoser was able to identify ~96% of the test sequences accurately.

Our results demonstrate that the simple HMM based diagnoser can effectively diagnose faults using only binary alarm signals information. In addition to this, the theoretical estimation of the confidence level in fault diagnosis provides insights into the limitations of the alarm system design and thereby helps improve process safety. In the future, we aim to compare the performance of this diagnoser against other similar methods that uses only binary alarm data for fault diagnosis.

References

1. Hollifield, B.R., E. Habibi, and J. Pinto, Alarm Management: A Comprehensive Guide. Research Traingle Park, NC: ISA, 2011.
2. Takeda, K., et al., A design method of a plant alarm system for first alarm alternative signals using a modularized CE model. Process Safety and Environmental Protection, 2014. 92(5): p. 406-411.
3. Doyen, L., T.A. Henzinger, and J.-F. Raskin, Equivalence of labeled Markov chains. International journal of foundations of computer science, 2008. 19(03): p. 549-563.
4. Kiefer, S. and A.P. Sistla. Distinguishing hidden Markov chains. in Proceedings of the 31st Annual ACM/IEEE Symposium on Logic in Computer Science. 2016.
5. Chen, T. and S. Kiefer. On the total variation distance of labelled Markov chains. in Proceedings of the Joint Meeting of the Twenty-Third EACSL Annual Conference on Computer Science Logic (CSL) and the Twenty-Ninth Annual ACM/IEEE Symposium on Logic in Computer Science (LICS). 2014.
6. Downs, J.J. and E.F. Vogel, A plant-wide industrial process control problem. Computers & chemical engineering, 1993. 17(3): p. 245-255.