Semi-supervised fault diagnosis framework for underwater propeller based on speed disentanglement strategy

Underwater propellers are typically operated under variable speed conditions. When a propeller fails, fault signature generally varies across different speeds. Current propeller fault diagnosis strategies often overlook the analysis of these time-varying speeds. To address this problem, this paper introduces a novel speed disentanglement strategy utilizing a limited labeled dataset. The proposed method integrates a speed disentanglement module and an autoencoder module with a classifier. The speed disentanglement module comprises a state extractor and a condition extractor to extract state-related and speed-related features, respectively. With the assistance of the speed disentanglement module, autoencoder extracts latent features that represent solely the health state from the acoustic signal spectrum, and the classifier is subsequently trained on a small amount of labeled data. Finally, the fault diagnosis performance of the proposed method has been validated using real data under various speed conditions. The results demonstrate that, compared to state-of-the-art speed disentanglement strategies, the proposed method mitigates the impact of speed variation and achieves accuracy improvements ranging from 2.07% to 4.68% across different tasks.