Patrice Rabate

Feature selection for complex systems monitoring: An application using data fusion

Emergence of automated and flexible production means leads to the need of robust monitoring systems. Such systems are aimed at the estimation of the production process state by deriving it as a function of critical variables, called features, that characterize the process condition. The problem of feature selection, which consists, given an original set of features, in finding a subset such the estimation accuracy of the monitoring system is the highest possible, is therefore of major importance for sensor-based monitoring applications. Considering real-world applications, feature selection can be tricky due to imperfection on available data collections: depending on the data acquisition conditions and the monitored process operating conditions, they can be heterogeneous, incomplete, imprecise, contradictory, or erroneous. Classical feature selection techniques lack of solutions to deal with uncertain data coming from different collections. Data fusion provides solutions to process these data collections altogether in order to achieve coherent feature selection, even in difficult cases involving imperfect data. In this work, condition monitoring of the tool in industrial drilling systems will serve as a basis to demonstrate how data fusion techniques can be used to perform feature selection in such difficult cases.