A novel data augmentation-based intelligent fault diagnosis method of wind turbine gearbox with limited labeled samples

Abstract

The acquisition of labeled vibration signal samples for wind turbine gearbox fault diagnosis is costly and labor-intensive in engineering practice, often leading to overfitting of deep learning models and limiting their deployment in real-world industrial scenarios. To address this challenge, this study introduces a data augmentation-based intelligent fault diagnosis (DAIFD) method. First, DAIFD incorporates a Spatiotemporal Fusion Generative Adversarial Network that synergizes bidirectional gated recurrent units and multi-scale convolutional layers to jointly extract temporal dynamics and spatial patterns in vibration signals. This enables the generation of high-fidelity synthetic fault data that emulates real-world gearbox degradation scenarios. Furthermore, a Cross-Task Coordination Diagnosis Network is proposed to establish a closed-loop framework where a shared spatiotemporal feature extractor is jointly optimized for both synthetic data generation and fault classification. This ensures that synthetic data generation actively aligns with the discriminative requirements of fault diagnosis tasks, enhancing model generalizability in data-scarce conditions. Experimental validation using test-bench data and field data from 2.5 MW wind turbines demonstrates that DAIFD achieves the best diagnosis performance. By addressing the industry’s critical pain point of labeled data scarcity, DAIFD provides an optional solution for improving gearbox reliability, offering practical value for wind farm operators and predictive maintenance platforms.