Statistical approaches to virtual metrology and predictive maintenance: experience from the IMPROVE project