The paper introduces a methodology of Knowledge Discovery in Serial Measurement (KDSM) for analyzing repeated very short time series with a blocking factor in ill-structured domains. This proposal focuses on results obtained on a real application to psychiatry, where common statistical analysis (time series analysis, multivariate\dots) and artificial intelligence techniques (knowledge based methods, inductive learning) used independently are often inadequate because of the intrinsic characteristics of the domain. This work shows how the limitations of the classical approaches are overcomed by using KDSM. KDSM is built as the combination of {\it clustering based on rules}, introduced by Gibert (1994), with some Inductive Learning (AI) and clustering (Statistics) techniques.