A Fault Diagnosis Framework Insensitive to Noisy Labels Based on Recurrent Neural Network
Xiaoyin Nie, Gang Xie
Abstract
Deep neural network (DNN)-based fault diagnosis is one of the effective means to ensure the safe and reliable operation of wind turbines (WTs). However, in practice, the complexity of labeling health condition samples leads to the possibility of health condition dataset corruption, resulting in a reduction of diagnosis effect. Considering the different distributions of noisy labels, this paper proposes a recurrent and convolutional neural network with clean revision (CRRCNN) framework, which consists of the recurrent and convolutional neural network (RCNN) and the clean revision. First, RCNN containing multi-axis input is constructed as the baseline network. Second, the clean revision containing three variations, namely backward cross-entropy loss, clean estimation of noisy transition matrix, and clean noisy joint training strategy, is embedded in the framework smoothly for better training. Finally, the proposed framework is verified by two type distribution noisy label datasets and the experiment results show the superiority of the proposed framework. Furthermore, the inner operation of CRRCNN is explored by sensitivity analysis (SA).