Skip to content
Open access

Learnable wavelet-enhanced localized Transformer for low-SNR wind turbine bearing fault diagnosis

Oct 2026 · Measurement science and technology · 0 citations

Abstract

Wind turbine bearing fault diagnosis under low signal-to-noise ratio (SNR) conditions remains challenging because vibration signals are often affected by rotational-speed fluctuations, load variations, and non-stationary background interference. These factors may weaken fault-related impulsive components, disturb time–frequency energy distributions, and reduce the discriminability of fault features. To address these issues, this paper proposes a learnable wavelet-enhanced localized Vision Transformer framework, termed LW-ViT. The proposed framework focuses on three core designs: adaptive time–frequency representation, localized token interaction, and noise-aware residual regulation. First, a learnable continuous wavelet transform (LCWT) module is developed to transform one-dimensional vibration signals into adaptive time–frequency representations. In this module, the scale and center frequency of the wavelet kernel are optimized during training, thereby reducing the dependence on manually predefined wavelet parameters. Second, a localized Transformer attention mechanism is introduced to restrict token interactions within neighboring time–frequency regions, which reduces unnecessary long-range noise propagation among time frequency tokens. Third, a noise-gated MLP block is incorporated to regulate the residual feed forward pathway and limit the propagation of noise-sensitive responses in deeper layers. Experiments are conducted on an in-house LUT wind turbine bearing dataset under different SNR levels, rotational speeds, and load conditions. The CWRU dataset is further used as a public benchmark for supplementary validation of the general applicability of the proposed framework. On the LUT dataset, LW-ViT achieves classification accuracies of 93.3% and 98.4% at −6 dB and −4 dB, respectively. Comparative experiments, ablation studies, and feature visualizations indicate that the proposed framework improves adaptive time–frequency representation, localized feature interaction, and fault classification under low-SNR conditions.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.