Industrial systems are often subjected to abnormal conditions due to faulty operations or external disturbances. Faults can easily propagate via the process components through material or information flows, thereby deteriorate the process performance and product quality and increase the operational costs. Therefore, it is of major importance to detect a fault, locate its root cause and reveal how it had propagated within the system. Capturing the causality of a system has a key role in fault diagnosis due to its ability to identify the root cause of a fault and retrace its propagation path. Thus far, several data-based methods have been proposed in order to capture causality from time series corresponding to process variables. However, the majority of the data-based methods suffer from several limitations and deficiencies which compromise their ability to provide adequate results, particularly when investigating complex systems with a high level of connectivity. This thesis proposes a hybrid causal analysis which automatically incorporates the information on the process connectivity into data-based analysis using a specialized search algorithm. The analysis aims to enhance the results accuracy, minimize the computational effort and to successfully tackle multivariate complex systems. This thesis entails four methodologies for a hybrid causal analysis based on the following causality estimators: Granger causality, transfer entropy, nearest neighbors and non-parametric multiplicative regression causality estimator. The hybrid causal analysis is successfully demonstrated on an industrial board machine using each of the causality estimators. The analysis aims to detect the propagation of an oscillatory disturbance due to valve stiction within the control loops of the drying section of the machine. Finally, the results of each causality estimator are evaluated and the methods are compared. The obtained results show that the hybrid causal analysis produced an enhanced causal model which depicts the oscillation propagation path and its root cause. Taken together, the findings of this study suggest that the connectivity information is essential for obtaining an adequate causal model when investigating complex systems. A natural progression of this work could be to implement the proposed hybrid analysis using other case studies with different types of faults.
|Translated title of the contribution||Data-based causality analysis by exploiting process connectivity information|
|Publication status||Published - 2019|
|MoE publication type||G5 Doctoral dissertation (article)|
- process connectivity
- fault propagation
- control loops
- board machine