Structure Guided Unsupervised Condition Monitoring of Wind Turbine Pitch System: A Deep One-Class Classification Approach

Currently, the operation and maintenance of wind turbine (WT) pitch systems have gained increasing attention and data-driven monitoring approaches have been widely studied. However, existing data-driven methods often cause false positives and alarms due to their purely data-driven nature. To further improve the monitoring performance, we proposed a deep one-class classification network model named Pitch-OC to detect early faults with the available supervisory control and data acquisition (SCADA) data. The training of the Pitch-OC model does not require data labels but only learns the WT pitch system health behavior characteristics. Pitch-OC integrates the prior knowledge of the pitch system structure. Specifically, we design a structure-guided group learning layer to combine the prior knowledge of the pitch structure and further capture the complex intervariable relationship in a group manner. Then, we map the hypersphere aggregation layer to obtain a hypersphere model. This hypersphere model is used to monitor the pitch system online. Our experiments on the SCADA data of two WTs prove that our model outperforms existing methods.