Adversarial variational autoencoder network for wind turbine condition monitoring

Wind turbine condition monitoring (WTCM) is crucial for wind turbine (WT) safe operation. However, WT data has class imbalances and complex underlying data distributions. Previous research fundamentally improved performance by expanding or deepening the network, which increases arithmetic resources. Therefore, we propose a model named adversarial variational autoencoder network (AVAEN), which achieves accurate WTCM while maintaining model simplicity and stability. AVAEN employs a two-stage training strategy: In the first stage, self-conditioning attention is employed to extract temporal dependencies from fine-grained data. Concurrently, the downsampling mechanism allows AVAEN to use improved variational autoencoder (VAE) to process sequences of different time scales during the encoding process. The first stage effectively achieves the extraction of fine-grained fluctuations and coarse-grained trends. In the second stage, reconstruction errors from the first stage are fed into the self-conditioning attention, enabling AVAEN to focus on subsequences with greater deviations. Furthermore, adversarial training is employed, which improves the quality of data reconstruction while avoiding the inherent limitations of GAN. The two-stage training effectively captures potential temporal correlations. Subsequently, experiments are performed on two WT datasets. AVAEN improves the F1 by 0.1416 compared to baseline methods, achieving good WTCM performance with low FLOPs and parameters.