Fault Diagnosis With Bidirectional Guided Convolutional Neural Networks Under Noisy Labels

In recent years, supervised deep learning methods have proven to be feasible and advanced for fault diagnosis. They do, however, rely on large amounts of accurately labeled data. It is challenging to accurately label incipient or secondary fault signals in the presence of operational background noise, resulting in noisy labels. This study presents a bidirectional guidance method that enables deep networks to diagnose faults robustly with the ability to label noise tolerance. First, the proposed method uses discrete wavelet packet transform (DWPT) to preprocess vibration signals before feeding them into the deep convolutional network. Second, a bidirectional loss (BL) function is proposed to significantly improve the robustness of the model in the case of noisy labels. Third, a fast cosine decay (FCD) strategy of the learning rate is designed to further boost the recognition performance of the model under severe noisy labels. The proposed method is applied to testbed experiments and wind turbine fault diagnosis. The experimental results demonstrate that it outperforms other cutting-edge methods in fault diagnosis with severely noisy labels.