Diagnostic decision algorithm based on spectral analysis and model validity

Spectral analysis based diagnosis proposes to use a prototype for the spectrum that constitutes a reference for further analyses. This prototype is used to decide that the actual signal has the same behavior as the reference one. Data used to build the prototype mast be very reliable, which is not always the case when acquisition is made in an industrial context The objective of this paper is the evaluation of the prototype validity and its use in the diagnostic decision. The study of the features of one or several reference signals, considered as prototypes of a specific behavior, is used to build a spectrum pattern representative of the specific behavior of the process (normal or specific fault). The validity of this pattern is evaluated a priori. During online diagnosis, the similarity between the actual data spectrum and the pattern is calculated thanks to the computation of a relevant distance. The pattern validity may be modified, to obtain a posteriori pattern validity. Both the similarity and the pattern validity are used to make the diagnostic decision. The proposed methodology is illustrated using an online short time Fourier transform for analyzing industrial data issued from a plate mill in metal industry. Note that plate mill and rougher of a hot rolling mill present very similar torque behavior