A novel normalized recurrent neural network for fault diagnosis with noisy labels

The early fault diagnosis is a kind of important technology to ensure the normal and reliable operation of wind turbines. However, due to the potential presence of noisy labels in health condition dataset and the weakly explanation of the deep neural network decisions, the performance of fault diagnosis is severely limited. In this paper, a framework called normalized recurrent neural network (NRNN) is proposed for noisy label fault diagnosis, in which the normalized long short-term memory is used to improve the training process and the forward crossentropy loss is introduced to handle the negative effect of noisy labels. The effectiveness and superiority of the proposed framework are verified by four datasets with different noisy label proportions. Meanwhile, the layer-wise relevance propagation algorithm is applied to explore the decision of framework and by visualizing the relevances of input samples to framework decisions, the NRNN does not treat samples equally and prefers signal peaks for classification decisions.

Automated summary

The paper introduces a framework called normalized recurrent neural network (NRNN) for noisy label fault diagnosis, which improves the training process with normalized long short-term memory and handles noisy labels with forward crossentropy loss. The effectiveness and superiority of the framework are verified with four datasets of varying noisy label proportions. Additionally, the layer-wise relevance propagation algorithm is used to explore decision-making within the framework, revealing that NRNN does not treat samples equally and prefers signal peaks for classification decisions.