Computational intelligence techniques for device testing

Nowadays, Microcontrollers are widely used in almost every electronic application, from domotic to aerospace fields. In many safety-critical applications such as Automotive Control systems, in order to meet the required high quality and safety standard, an effective performance screening has to be performed to detect under performing devices, measuring the maximum frequency at which the device can operate under challenging conditions of temperature and voltage. However, this is an extremely costly and time-consuming stage. In this work, Machine Learning techniques will be exploited in order to support the screening phase, to identify not-properly-working devices on the basis of on-board embedded ring oscillators, called Speed Monitors, that measure several physical parameters of the devices. Different regression models will be evaluated and compared to predict the maximum operating frequency of the devices, using Multi-Task Regression techniques in order to take advantage of several label measured on the devices. Since label for the learning phase of machine learning models are costly to obtain, Active Learning approach will be exploited in order to build an optimized training set, aiming to reduce the number of labelled samples needed to reach a certain amount of prediction error. Then, since in the production phase no label are available and since the model may become no more accurate due to shift in the involved industrial processes, a novel approach based on classification models is used in order to try to identify an estimation of the prediction error on new unseen data in the production stage, labelling new points with a certain "Error Zone" on the basis of "how far" they are from the training set.