Polynomial Improved Convolution Kernel Graph Network For Fault Diagnosis

Diagnosis is widely performed using convolutional neural networks (CNN). There are, however, limitations to CNN in non-Euclidean space. Based on the superiority of the graph convolutional neural networks (GCN) in non-Euclidean space, the GCN was adopted in this paper. In order to optimize the convolution process, the convolution kernel is approximated by four orthogonal polynomials, Chebyshev, Legendre, Laguerre, and Hermite. Then, ChebyNet, LegenNet, LargerNet, and HermiNet models are constructed using these four polynomials. According to the experimental results, HermiNet is superior when time-domain vibration signals are used as input, and LargerNet is superior when frequency-domain vibration signals are used.