Managing the uncertainty of conformity assessment in environmental testing by machine learning

Abstract

A machine learning approach is described, with reference to the conformity assessment of pin solder joints for electronic devices after tests based on cyclic thermal stresses. Metrological concepts, in particular expanded uncertainty, confidence level and conformity assessment, are used to reinterpret expert judgements, with the aim of transferring as much as possible the expert judgement know-how into a semi-automated evaluation process of X-ray images. This also allows us to reduce to an acceptable level the percentage of errors of the method, with respect to the identification of faulted specimens. A tailored procedure is set, which is able to reach a satisfactory level of correct acknowledgment of the status of pieces, giving also indication of cases where the level of confidence is unsatisfactory. The obtained results show that in this way the occurrence of mistakes strongly decreases. The paper also analyses the effect of algorithms and of the most relevant data processing settings on the ambiguity percentage.