Semi-Supervised Few-Shot Fault Diagnosis of Wind Turbine Pitch Bearings with Vibration-Acoustic Signal Fusion

Abstract

: Wind turbine pitch bearings operate under low-speed, heavy-load, and time-varying conditions, where fault samples are difficult to collect and label. These constraints limit the effectiveness of conventional supervised diagnostic models, especially when fault classes are imbalanced and only a few labeled samples are available. This paper proposes a semi-supervised few-shot fault diagnosis framework for wind turbine pitch bearings with vibration-acoustic signal fusion. The framework uses synchronized vibration and acoustic signals as complementary inputs and employs a hybrid convolution-attention backbone to capture local temporal features and global context within each input segment. To improve multimodal representation, a channel-exchange mechanism is introduced to replace low-importance feature channels between vibration and acoustic branches while preserving modality-specific information. To exploit unlabeled monitoring data, a pseudo-loss-based reliability estimation strategy is used to distinguish reliable pseudo-labeled samples from noisy ones. A four-component semi-supervised Gaussian mixture model fits the pseudo-loss distribution and selects high-confidence pseudo-labeled samples for iterative self-training. Dynamic class weighting is further used to reduce bias caused by uneven pseudo-label selection. Experiments on a public wind turbine pitch bearing vibration-acoustic dataset are conducted under six operating tasks with 5-way 1-shot and 5-shot settings. The proposed method achieves the highest mean accuracy across all reported tasks when compared with inductive, transductive, and semi-supervised few-shot baselines. Ablation results show that the channel-exchange mechanism improves multimodal feature representation, while the class weighting strategy provides additional gains under scarce-label settings. These results indicate that the proposed framework can improve pitch bearing fault diagnosis when labeled data are limited and unlabeled vibration and acoustic data are available.