Mining Causal Hypotheses in Categorical Time Series by Iterating on Binary Correlations

Abstract

Machine errors can propagate in a production field and reduce the efficiency of smart manufacturing execution systems. Since every highly automated machine can have many possible status reports, their causalities can only be detected by means of statistical analyses. We present a highly automatable methodology for iteratively analysing machine state time series and for detecting machine error causality hypotheses. First, the categorical status time series of all machines are analysed for binary correlations in two iteration steps using pairwise cross-correlation. Out of all correlations, significantly high correlations are then combined and can be validated for causalities by means of plausibility and semantic criteria. Our experimental results are presented on anonymised real production state time series and a simple representational concept for further causal interpretation is introduced.