Statistical Methods for Monitoring Data Analysis

The analysis of monitoring data is often aimed at deciding if some characteristics of the monitored structure have deteriorated or are going to worsen. To do this engineers use monitoring systems which are often highly instrumented but data are confounded by measurement errors and secondary effects. Therfore it is important first to separate the main signals which identify the health of the monitored structure and, second, to give a synthetic statistical description of the possibly many measurements obtained, which may be numerous. Along these lines, the paper proposes an integrated approach, coupling the monitoring system with a set of techniques for statistical surveillance and diagnostics, which give appropriate data understanding and rules for signalling that the monitored system is experiencing a drift. The first step is statistical data analysis and modelling at the instrument level, this gives instrumental field uncertainty assessment and compensation for secondary local effects. The second step is based on modelling the monitoring system as a whole. This gives a precise assessment of the correlations among different system areas and instruments. Moreover, this is the starting point for the third step, namely a model based hierarchical multivariate detection method, which is capable of signalling deviations from overall normality while controlling false positive rates.