Development of an interpretability approach for machine learning in the context of non-destructive material classification based on magnetic Barkhausen noise.

In the context of fine blanking, magnetic Barkhausen noise is to be considered a valuable source of data, for the purpose of determining mechanical properties of metals. Magnetic Barkhausen noise is a non-destructive testing procedure that has already be proven to have relationships with mechanical properties of ferromagnetic materials. Works in this field seem to indicate that machine learning is promising for estimating mechanical properties with magnetic Barkhausen noise as an input data. However, machine learning techniques, despite being able to model non-linear relationships, lack in transparency since their inner working remains opaque. It is possible, therefore, that an erroneous model, which performs well based by exploiting biases in the dataset, will be undetected. This thesis has therefore two main objectives. Firstly, to implement a machine learning model to estimate the hardness of materials based on magnetic Barkhausen noise. Secondly, to implement an interpretability approach, in order to understand if the predictions made by the model are compatible with current engineering knowledge in the field. To address the first main objective, a convolutional neural network-based approach has been implemented. To address the second main objective, an interpretability approach known as grad-CAM has been implemented. Furthermore, a self-explaining convolutional neural network has also been implemented, in order to achieve a more reliable interpretation of the results.