Optimal classifier design based on pairwise statistical separability maximisation of time–frequency features

This paper presents a novel classification algorithm based on the time–frequency features extracted from multiple-sensor signals. Multiple-sensor signals are difficult to handle for classification purpose since each signal may have a different separability measure between classes and, hence, it may be difficult to pick a set of best sensors for classification. This paper provides a new separability measure, the so-called miss-classification probability, in order to overcome such a difficulty. A mathematical representation of the statistical aspect of the time–frequency features is introduced for efficient calculation of the miss-classification probability. Yet, another difficulty may be encountered in extracting a set of time–frequency features, which may best represent the difference among classes. This paper also proposes a pairwise statistical separability maximisation scheme to overcome this difficulty. The resultant classification algorithm based on these new developments is validated through seeded-fault tests with rotary compressors.