Fault detection in wind turbine generators using a meta-learning-based convolutional neural network

Conventional fault detection methods for wind turbine (WT) generators often grapple with inadequate warning times and poor portability. These issues contribute to heightened safety risks and an increased false positive rate (FPR) and false negative rate (FNR). This study introduces a fault detection method for WT generators utilizing a 1D convolutional neural network (1DCNN) based on meta-learning principles. We incorporate the "learning to learn" concept of model-agnostic meta-learning (MAML) into a 1DCNN, enabling effective fault detection. More specifically, the training data are transformed into numerous tasks through random sampling, and the model is trained task by task. The 1DCNN is utilized as the base learner, leveraging its superior feature extraction capability to discern task features. The first order gradient of MAML is applied to ascertain the specific initialization parameters for each task, while the second-order gradient of MAML is used to understand the similarities and differences between all tasks' initialization parameters. This approach equips the 1DCNNMAML with the ability to adapt to new tasks and converge rapidly, thereby achieving swift regression prediction. We also employ the probability distribution fitting method to analyze the distribution of prediction residuals, thus setting the detection threshold. Based on this threshold, warnings can be issued for faults in WT generators. We used supervisory control and data acquisition (SCADA) data from the Liaoning wind farm in China to validate the robustness and portability of the proposed model. Experimental outcomes indicate that, compared with Reptile, FOMAML, LSTM-MAML, 1DCNN, and LSTM, our proposed method can detect faults earliest across different wind turbines and has the lowest FPR and FNR.