Big Data management: A Vibration Monitoring point of view

Vibration Monitoring is a particular kind of Condition Monitoring meant to infer the state of health of a machine from accelerometric measurements. From a practical point of view, the scope is then to extract from the acceleration data some valuable diagnostic information which could be used to detect the presence of possible damages (i.e., to produce knowledge about the state of health). When the monitoring is implemented online, in a continuous way, the raw accelerometric data sets can be very large and complex to be dealt with, as usually involve multiple channels (i.e., multiple locations and directions) and high sample rates (i.e., order of ksps - 103 samples per second), but the final knowledge about the state of health can, in principle, be summarized by a single binary information (i.e., healthy - 0 vs damaged - 1). This is commonly called Damage Detection. In this work, the big data management challenge is tackled from the point of view of statistical signal processing, so as to aggregate the multivariate data and condense them into single information of distance with respect to a healthy reference condition (i.e., the Novelty). When confounding influences (such as the work condition or the environmental condition) can be disregarded, the novelty information has a direct correspondence to the health information, so that an alarm indicating the detection of damage can be triggered upon exceeding a selected threshold for the limit novelty. Many different ways of solving such a binary classification problem can be found in the literature. Starting from the simplest, some of the more effective are compared in the present analysis, to finally select a reliable procedure for the big data management in vibration monitoring.