Análise de desempenho de técnicas de indicação de causalidade aplicadas a alarmes industriais

Abstract

Industrial alarms are inherently asynchronous in nature and are critical to maintaining the operational safety and health of complex industrial processes. However, poorly configured industrial alarm systems tend to generate excessive amounts of alarms, making them inefficient. Among the possible degrading agents of an alarm system are the causal alarms. Causal alarm situations occur when the activation of a given alarm implies the activation of one or more of the resulting alarms, generating redundant information in the alarm system. Given the relevance of the problem, this dissertation analyzes the performance of two techniques for determining causal alarms: Cross-correlation and Granger causality test. The industrial alarm data is essentially of a discrete nature, before performing both techniques, there was a need to perform a preprocessing on the alarm data, through the signal smoothing technique. To obtain the results, we used alarm data from the alarm generation simulation scenarios and the Tennessee Eastman Process Benchmark. Therefore, the results indicate that, in general aspects, the Granger causality test performed a greater efficiency than the cross-correlation in the task of indicating causal relations between industrial alarms. We also performed comparative studies of the application of the Granger causality test on process variables and alarms in the Tennessee Eastman Process Benchmark, indicating their characteristics.