A fault diagnosis method for offshore wind turbine bearing based on adaptive deep echo state network and bidirectional long short term memory network in noisy environment

• Published in: Ocean engineering 2024-11, Vol.312, p.119101, Article 119101
• Main Authors: Du, Yuanhao, Geng, Xiuli, Zhou, Qingchao, Cheng, Sheng
• Format: Article
• Language: English
• Subjects: Adaptive deepESN; BiLSTM; CEEMDAN; Fault diagnosis; PSR
• ISSN: 0029-8018
• DOI: 10.1016/j.oceaneng.2024.119101

Accurately diagnosing offshore wind turbine bearing faults in noisy environment has long been a challenge for engineers. This necessitates enhancing diagnostic accuracy while effectively reducing noise. Accordingly, this study proposes an offshore wind turbine bearing fault diagnosis (FD) method in noisy environment based on complete ensemble empirical mode decomposition with adaptive noise-phase space reconstruction (CEEMDAN-PSR) and a hybrid neural network, namely CPAEBL. We conduct research from two main aspects: noise reduction and optimization of the FD model. In terms of noise reduction, an enhanced CEEMDAN method that integrates the correlation coefficient and skewness indicators for selecting Intrinsic Mode Functions (IMFs) is designed. Furthermore, by leveraging the capabilities of PSR to better reveal signal evolution and characteristics, these two approaches are innovatively combined to perform noise reduction and restore dynamic signal features. In terms of model optimization, a deep Echo State Network (deepESN) with adaptive constraints is established to extract features from denoised vectors. Subsequently, Bidirectional Long Short-Term Memory (BiLSTM) is effectively employed to capture temporal information within the feature vectors, ultimately enhancing the accuracy of wind turbine bearing FD models. Finally, the CPAEBL model is applied to the Case Western Reserve University (CWRU) dataset, the Paderborn University (PU) dataset, and the actual collected bearing dataset of offshore wind turbines. The ablation experiments and comparison experiments prove that the CPAEBL model has better anti-noise performance, generalization performance, and robustness. •A novel method is proposed for diagnosing offshore wind turbine bearing faults under noisy environments.•We designed a multi-indicator CEEMDAN to screen out IMFs containing valid information.•The innovative combination of CEEMDAN and PSR is used to reduce noise and restore the dynamic characteristics of signals.•An adaptive-constrained deepESN is established for feature extraction.•The feature matrix extracted from the reconstructed data is utilized as the input of the diagnostic model.