Evolutionary computation based nonlinear transformations to low dimensional spaces for sensor data fusion and Visual Data Mining

Data fusion approaches are nowadays needed and also a challenge in many areas, like sensor systems monitoring complex processes. This paper explores evolutionary computation approaches to sensor fusion based on unsupervised nonlinear transformations between the original sensor space (possibly highly-dimensional) and lower dimensional spaces. Domain-independent implicit and explicit transformations for Visual Data Mining using Differential Evolution and Genetic Programming aiming at preserving the similarity structure of the observed multivariate data are applied and compared with classical deterministic methods. These approaches are illustrated with a real world complex problem: Failure conditions in Auxiliary Power Units in aircrafts. The results indicate that the evolutionary approaches used were useful and effective at reducing dimensionality while preserving the similarity structure of the original data. Moreover the explicit models obtained with Genetic Programming simultaneously covered both feature selection and generation. The evolutionary techniques used compared very well with their classical counterparts, having additional advantages. The transformed spaces also help in visualizing and understanding the properties of the sensor data.