Are Novel Deep Learning Methods Effective for Fault Diagnosis?

In recent years, deep learning has become the standard approach for fault diagnosis in mechanical equipment, with models becoming increasingly complex and large in scale. Through a thorough review and analysis of existing literature, we found that while many studies report performance improvements with the latest models, there has been no comparison of state-of-the-art (SOTA) methods. This gap is primarily due to two factors: first, the wide variation in the quality and nature of fault diagnosis data leads to significant performance fluctuations across different datasets; second, the diverse preprocessing methods employed make it challenging to compare models objectively. For instance, while deep learning has demonstrated high accuracy in bearing fault diagnosis, variations in vibration signal processing methods often skew the evaluation of model performance. To address these issues and evaluate the true performance of the latest deep learning models for fault diagnosis, this article establishes a unified data processing framework that ensures fair performance comparisons across models. Using this framework, we reproduce eight SOTA deep learning models and assess their effectiveness on three publicly available bearing datasets. Additionally, we design three benchmark models to quantify performance differences. The experimental results highlight that current deep learning-based fault diagnosis methods still face significant challenges in real-world applications. Finally, the future research directions in this domain are given.