Loading [a11y]/accessibility-menu.js
A Novel Interpretable Adversarial NMF Network for Compound Fault Decoupling in Planetary Transmission | IEEE Journals & Magazine | IEEE Xplore

A Novel Interpretable Adversarial NMF Network for Compound Fault Decoupling in Planetary Transmission


Abstract:

As the planetary transmission of heavy-duty vehicles endures heavy load impacts during service process, the key components suffer high-fault rates, which primarily compri...Show More

Abstract:

As the planetary transmission of heavy-duty vehicles endures heavy load impacts during service process, the key components suffer high-fault rates, which primarily comprise compound faults (CFs). This results in increased maintenance costs and higher risks of downtime. Although existing intelligent fault diagnosis methods based on deep learning (DL) provide solutions for CF diagnosis (CFD), the DL-based methods still have the following significant limitations: they require numerous labeled CF samples; they are ineffective at identifying single fault characteristics in complex mechanical structures; and the “closed-box” nature of DL models complicates their interpretation. This article proposes a novel interpretable adversarial nonnegative matrix factorization (ANMF) CF decoupling network to improve equipment maintenance efficiency. Representing an improvement within the framework of the generative adversarial network, this ANMF reconstructs the decoupler and discriminator using deep nonnegative matrix factorization (NMF) and convolutional sparse encoding layers. The objective function of the deep-NMF embedding model is redesigned by introducing regularization terms to guide the model’s decoupling direction, enabling accurate decoupling of CFs. The proposed method is validated through experiments involving the injection of CFs into a planetary transmission. The results demonstrate that the ANMF achieves precise decoupling of CFs at different speeds without requiring labeled CF data, with an accuracy drop of no more than 1% compared with real samples, and this represents an accuracy improvement of nearly 14.4% over state-of-the-art methods. Furthermore, both the model inference process and results are physically interpretable.
Published in: IEEE Internet of Things Journal ( Volume: 12, Issue: 2, 15 January 2025)
Page(s): 2079 - 2089
Date of Publication: 24 September 2024

ISSN Information:

Funding Agency:


References

References is not available for this document.