On the relevance of graphical causa models for failure detection for industrial machinery

H. Kosorus, T. Natschläger, B. Freudenthaler, J. Küng. On the relevance of graphical causa models for failure detection for industrial machinery. pages 85-86, 2, 2013.

  • Hilda Kosorus
  • Thomas Natschläger
  • Bernhard Freudenthaler
  • Josef Küng
  • A. Quesada-Arencibia et al.
BuchEUROCAST 2013 Computer Aided Systems Theory Extended Abstracts
TypIn Sammelband
VerlagIUCTC Universidad Las Palmas
Abstract n this paper we propose to use graphical causal models for fault detection in industrial machinery to tackle this issue. Mining of causal model structures within multivariate time series, i.e. collected condition monitoring data from an industrial machinery in our setting, has received a lot of attention in the past years [2, 3, 7] and is still considered to be a major challenge, especially when dealing with complex and noisy data. Many of the developed methodologies try to leverage temporal information in order to infer the causal structure of a model using the concept of Granger causality: it is based on the notion that causes always precede their effects and that causal variables contain unique information about the effect variable, which otherwise is not available [5]. Prominent recent approaches within this field are graphical Granger methods [1] which make use of graphical lasso methods [4]. The idea behind the graphical lasso methods is to estimate a sparse graphical model by optimizing a L1 regularized log-likelihood for a Gaussian graphical model. Hence, these methods based on such first principles generalize heuristic approaches like the one presented in [7]. Studies have shown that such graphical Granger methods are well suited to reveal the true model structure, can add extra predictive accuracy and may also help to improve the interpretability of obtained models by arriving at more succinct models [1].