Fault diagnosis of wind turbine gearboxes using a temporal contrastive learning model integrated with multi-kernel Shiftwise convolution

Abstract

Sustained wind turbine gearbox operation under volatile conditions and intense background noise complicates vibration signals, rendering reliable fault label acquisition cost-prohibitive. Under scarce labeling constraints, traditional data-driven methodologies frequently fail to isolate distinctive fault signatures. To circumvent this, this paper presents a temporal contrastive framework integrating a hybrid multi-kernel Shiftwise convolutional network, optimizing feature spaces without extensive manual annotation. Crucial to this framework is a multi-kernel Shiftwise architecture; by marrying channel-shift mechanisms with lightweight parallel convolutions, it captures multi-scale local anomalies efficiently, neutralizing traditional receptive field limitations. Concurrently, a dual-level self-supervised paradigm extracts both localized multi-scale features and global contextual representations directly from unlabeled field data. Downstream gearbox state classification is subsequently achieved by fine-tuning the pre-trained encoder with sparse labeled samples. Validation on the BJTU gearbox dataset and SUFD yields diagnostic accuracies of 90.84% and 91.40% using merely 2.5% and 5% labeled data, substantially outperforming existing methods.