Process and alarm data integration under a two-stage Bayesian framework for fault diagnostics

Abstract Process and alarm data are usually available from industrial processes. It is common practice to use process data for process monitoring and diagnostics. In contrast, alarm data is typically used to determine the instantaneous health state of the process, often as part of a protection system. Alarm data is also used in alarm management systems and for alarm flood detection. Despite the fact that both data types perform similar, although not identical, functions in process monitoring, they are rarely used in combination. One of the main reasons for this is that the fusion between alarm and process data is not trivial: process data is sampled continually and is numerical, while alarm data is binary and appears at discrete times. The two data sources can contain complimentary information regarding the health state of the process, therefore their fusion is a promising direction for fault diagnostics algorithm development. A two-stage Bayesian framework is proposed to fuse alarm and process data on the decision level for fault diagnostics targeting industrial processes. Instead of the raw process data, the principal components of the process variables are used as the inputs of a naive Bayes classifier. This step reduces the correlation between the process variables and reduces the dimension of the data. The alarm history is transformed into binary alarm features and input to a second, separate naive Bayes classifier. The second stage of the method fuses the local classification results of the alarm and process data and provides the final classification result. The results show that the overall performance of the method fusing alarm and process data is superior when compared to the results of a similar single stage method using either alarm or process data.