Performance analysis and application of the bidirectional associative memory to industrial spectral signatures

A detailed study of the performance of the bidirectional associative memory (BAM) for the classification of power spectral density (PSD) functions is presented. The BAM has error-correcting properties and is capable of generating a successful recall, even with noisy and incomplete patterns. Suitable pattern coding schemes are utilized to enhance the discriminating features in the patterns. A bit-mapping scheme is used to construct binary pattern vectors from the PSD. The numerical aspects of signature encoding and fault tolerance property of the BAM are evaluated using spectral density functions of signals from a fiber-manufacturing rotating machinery system. The results of the application of the BAM are superior to those obtained with classical pattern recognition methods.<>