A machine learning approach is used for analysis of pin solder joints for electronic devices after tests based on cyclic thermal stresses. Metrological concepts, in particular budget uncertainty, are used to reinterpret expert's judgements, to individuate the variability causes of errors of a semi-automated evaluation process of x-ray images, aiming to replicate the expert's judgement. The actions are also described for reduction to an acceptable level of the error percentage with respect to the faulted specimens identification. In this way a tailored approach is set, which is able to remarkably reduce the errors and to evaluate the most significant contributions to the variability of results. The suggestions deriving from this study could be useful also for different applications.