Intelligent fault diagnosis for wind turbine main bearings through multi-source signal fusion and lightweight model design with knowledge distillation

Deep learning has achieved notable success in bearing fault diagnosis. However, this success is limited to single-dimension fault information and issues with larger models and complex computations, making it impractical for industries requiring high accuracy and easy deployment. Considering the need for intelligent diagnostic models in wind power operation and maintenance to balance diagnostic accuracy and practicality, this paper proposes an architecture for constructing a lightweight intelligent fault diagnosis model based on multi-source heterogeneous signal fusion and knowledge distillation (KD). The model is validated on the fault conditions of wind turbine main bearings using a dataset from Paderborn University (PU), focusing on the combination of motor current and vibration signals. The model utilises Markov transition fields (MTFs) to represent the different features extracted from these two signals, enriching the multi-dimensional description of bearing faults. Subsequently, convolutional neural networks (CNNs) extract features from the Markov transition field time-frequency diagrams of the two signals for fusion diagnosis. Finally, knowledge distillation is applied for model compression, transferring the knowledge learned by the larger and more complex teacher network, which results in the fusion of multiple input signals to a smaller and more computationally efficient student network for effective deployment in practical engineering applications. Experimental results show that the proposed lightweight student model with multi-source signal fusion significantly reduces complexity and computational resource consumption while maintaining high diagnostic accuracy. This makes it easier to achieve fault diagnosis of wind turbine main bearings and provides practical models and valuable references for intelligent operation and maintenance diagnostic engineering applications.