Deep complex neural network learning for high-voltage insulation fault classification from complex bispectrum representation

Bispectrum representations previously achieved a successful classification of insulation fault signals in High-Voltage (HV) power plant. The magnitude information of the Bispectrum was implemented as a feature for a Deep Neural Network. This preliminary research brought interest in evaluating the performance of Bispectrum as complex input features that are implemented into a Deep Complex Valued Convolutional Neural Network (CV-CNN). This paper presents the application of this novel method to condition monitoring of High Voltage (HV) power plant equipment. Discharge signals related to HV insulation faults are measured in a real-world power plant using the Electromagnetic Interference (EMI) method and processed using third order Higher-Order Statistics (HOS) to obtain a Bispectrum representation. By mapping the time-domain signal to Bispectrum representations the problem can be approached as a complex-valued classification task. This allows for the novel combination of complex Bispectrum and CV-CNN applied to the classification of HV discharge signals. The network is trained on signals from 9 classes and achieves high classification accuracy in each category, improving upon the performance of a Real Valued CNN (RV-CNN).