Spectrum classification for early fault diagnosis of the LP gas pressure regulator based on the Kullback-Leibler kernel

The present paper describes a frequency spectrum classification method for fault diagnosis of the LP gas pressure regulator using Support Vector Machines. Conventional diagnosis methods are not efficient because of problems such as significant noise and nonlinearity of the detection mechanism. In order to solve these problems, a machine learning method with the Kullback-Leibler (KL) kernel based on the KL divergence is introduced into spectrum classification. We use the normalized frequency spectrum directly as input with the KL kernel. The proposed method demonstrates a higher accuracy than popular kernels, such as polynomial or Gaussian kernels, or the conventional fault diagnosis method and Gaussian Mixture Model with the KL kernel for the examined problem. The high classification performance is achieved by using an inexpensive sensor system and the machine learning method. This method is widely applicable to other spectrum classification applications without limitation on the generality if the spectrums are normalized.