Multi-sensor signals augmented multimodal MAML-1DCNN-RBEAM deep transfer learning algorithm for wind turbine bearing fault diagnosis

In the fault diagnosis of wind turbine bearings, although there are multiple sensors, the fault signals are extremely scarce, mixed with high noise, and the working conditions are complex and variable. For the above problems, this paper proposes a novel multimodal deep transfer learning framework, the MAML-1DCNN-RBEAM algorithm, specifically designed to enhance diagnostic precision and robustness. This algorithm uniquely combines multi-sensor signals for signal augmentation, facilitating a broader and richer data representation. Unlike existing approaches that rely solely on single-modality or fixed-feature extraction, this algorithm innovatively integrates time-frequency spatial and statistical modalities, maximising feature diversity from source domain data. The global features in the time-frequency spatial modality are optimised through a Model-Agnostic Meta-Learning (MAML) scheme, enhancing the model’s adaptability to new fault conditions. Concurrently, statistical modal features are extracted via the custom-designed 1DCNN-RBEAM model, which incorporates residual blocks enhanced by attention mechanisms for heightened feature sensitivity. Fusion occurs at the decision level, resulting in a robust multimodal learning framework adaptable to varying operational scenarios. Comparative experiments validate that this approach not only surpasses baseline methods in diagnostic accuracy and generalisation but also effectively addresses the complexities of small-sample, fluctuating conditions in wind turbine environments.