Machine Learning Methods for Fault Classification

With constant evolution of technology, new issues appear in large digital circuits. Nowadays, not only permanent faults may appear in the chip, but also intermittent faults or even faults due to external noise may be present.

The correct classification of these three types of faults is crucial for the success of any integrated circuit: permanent and intermittent faults are considered critical as they may compromise the reliability of the manufactured device. Therefore, integrated circuits showing such failure behavior have to be discarded. Transient faults (noise) are considered uncritical since, either such faults are not present during regular circuit operation, or because the circuit is able to tolerate them by design.

An incorrect classification of a critical fault can result in reliability problems, while discarding a circuit affected by noise may cause unnecessary yield loss.

The effect of a fault on the behavior of a circuit depends on its logic structure, the manufacturing technology and the parameters of the test insertion (like voltage and temperature), among other factors. Although some approaches exist already to distinguish between these different types of faults, they all follow fixed global classification rules that are expected to be valid for every circuit design, test parameters and technology.

The goal of this thesis is to develop a fault classification tool using machine learning approaches, which should allow the training of the tool to obtain a more precise and reliable classification.

Prerequisites:

• Java (mandatory)

• Octave (advantage)

• “Design and Test of Systems on a Chip”, or “Hardware Verification and Quality Assessment”

 
CONTACT  
Alejandro Cook (Email: alejandro.cook@informatik.uni-stuttgart.de)
Laura Rodríguez Gómez (Email: laura.rodriguez@informatik.uni-stuttgart.de)