Semi-supervised IFA with Prior Knowledge on the Mixing Process: An Application to a Railway Device Diagnosis