From time series measurements to rules of causality

Abstract

Data analysis procedure is described which can be used to extract events and rules of causality from time series measurements of industrial processes. The proposed procedure incorporates well-known data analysis methods, that have not been widely used in condition monitoring systems. Here, it will be demonstrated how the algorithms can be utilised in mobile work machine condition monitoring. The analysis process starts with selection of representative measurements which describe the operation of machine reasonable well. Here, variables are selected using a clustering method, in order to find groups of variables. One measurement from each group is selected for later analysis. Next, data streams are segmented to find areas on which the operation of machine continues without any abrupt changes. The segments or combinations of them are analysed to be able to extract sequences of operational states or change points. Event sequences are further analysed to extract association rules for the events. The extracted rules contain information about occurrence probability of certain event sequence. This facilitates, e.g., identification of fault precursors. Probabilities computed from measurements using this procedure can be used to adjust expert knowledge based fault-trees.