Posters-Accepted Abstracts: Adv Robot Autom
Alarms are used in industrial plants to remind operators of abnormality so that they can take proper actions to solve the problems to remove or alleviate the symptoms. Traditionally, alarms are triggered according to comparison with thresholds configured on specific process variables. To reduce the influence of noise and instability, techniques such as filtering, dead bands, and time delays are used to avoid false and missed alarms as much as possible, whereby keeping the detection delay in an acceptable range. All the above techniques are based on the idea of univariate alarm strategy. However, due to the relationship between variables, alarms should ideally be designed in a multivariate framework, i.e., as combined indices generated by a set of variables, including both continuous process variables and binary variables. For this purpose, the similarity or extended correlation between variables needs to be defined and therefore statistical techniques such as principal components analysis can be employed for the index generation and visualization tools such as alarm similarity color map (ASCM) can be used for intuitive analysis. In addition, especially considering of dynamic factors, there may be evident causality between alarms, leading to parent-child alarms; thus various causality analysis techniques can be utilized to build topology and find root causes. These techniques include data-based analysis and formulized knowledge description and discovery. The former includes Granger causality, transfer entropy, and partial directed coherence, and the latter can be ontology models and their reasoning. Some case studies are shown to illustrate the proposed techniques.
Advances in Robotics & Automation received 1004 citations as per Google Scholar report