Signal-based diagnostic algorithms integrating model validity in the decision

Signal analysis can be used to perform diagnosis: features that define the representation space are extracted with a processing technique. Diagnosis consists in mapping the representation space into fault indicators easily interpreted by operators. This paper proposes two main ideas. Firstly, a limited number of realisations of one signal is considered offline to define the pattern in the representation space. Secondly, the confidence that can be attached to the pattern when the diagnostic decision is computed is taken into account. The diagnostic decision is based on multicriteria fuzzy decision-making: it aggregates the pattern validity and the similarity of the current signal features to this pattern. The proposed methodology is well adapted to FDI of multiple faults. It is illustrated with a STFT for feature generation and gives very encouraging results when applied to industrial data recorded from a roughing mill in the metal industry.