Sparse dictionary learning based adversarial variational auto-encoders for fault identification of wind turbines

Abstract

Incipient anomaly state identification of wind turbines is beneficial to improve the reliability of wind turbines, reduce operation and maintenance costs. Due to various reasons such as random noise sampling or non-convex loss function, many deep learning algorithms show instable results on the anomaly state identification. In this paper, an anomaly state identification model for wind turbines using adversarial variational auto-encoders (AVAE) and sparse dictionary learning (SDL) is proposed, named sparse dictionary learning based adversarial variational auto-encoders (AVAE_SDL). Generative adversarial network (GAN), variational auto-encoders (VAE) and SDL are combined in the proposed method to leverage the advantage of GAN as a generative model, VAE as a posterior distribution learner and SDL is used to extract important features of latent code. Supervisory Control and Data Acquisition (SCADA) is a built-in monitoring system for wind turbine, which is used for anomaly identification in this paper. First, the pre-processed healthy SCADA data is adopted to train the model, and the anomaly state threshold is determined adaptively according to the training results. Then, all the testing data is input into the trained model. A statistical score is computed as a measurement to implement anomaly detection. Due to the different operating conditions of different units, each unit is individually trained and tested. The reconstructed error of each parameter in SCADA is computed to determine the fault location. The effectiveness and robustness of the proposed method are verified by nine on-site wind turbines. The results show that the three faulty units all exceed the threshold before the actual failure time, and return below the threshold after replacement. The correct fault location is realized in all three cases. The six healthy units are always within the normal threshold range.