A Deep Double-Convolutional Neural Network-Based Fault Detection

Abstract

Fault detection is crucial for saving valuable maintenance time and costs of industrial systems when faults occur. Faults are usually cataloged into sudden faults and degraded faults, which have diverse impacts on system safety. So, distinct treatment should be treated during fault detection. In this paper, an intelligent fault detection method is proposed based on a double-convolutional neural network (CNN) model architecture to detect whether faults occur and which type they belong. Firstly, a CNN model is employed to judge whether faults occur. Furthermore, the "Majority rule" is applied to effectively eliminate the influence of outliers for enhancing the robustness of fault detection model. Then, Another CNN model combined with root-mean-square (RMS) is introduced to detect fault types that belong to the sudden fault or degraded fault. Finally, an experimental study involving four types of bearing degradation scenarios is conducted to validate the effectiveness of the proposed method.