A hybrid LSTM-KLD approach to condition monitoring of operational wind turbines

• Published in: Renewable energy 2022-01, Vol.181, p.554-566
• Main Authors: Wu, Yueqi, Ma, Xiandong
• Format: Article
• Language: English
• Subjects: Condition monitoring; Kullback-Leibler divergence; Long-short term memory; Wind turbine
• ISSN: 0960-1481; 1879-0682
• DOI: 10.1016/j.renene.2021.09.067

With the increasing installation of the wind turbines both onshore and offshore, condition monitoring technologies and systems have become increasingly important in order to reduce the downtime and operations and maintenance (O&M) cost, thus maximising economic benefits. This paper presents a novel machine learning model-based data-driven approach to accurately evaluate the performance of the turbines and diagnose the faults. The approach is based on Long-short term memory (LSTM) incorporating a statistical tool named Kullback-Leibler divergence (KLD). The hybrid LSTM-KLD method has been applied to two faulty wind turbines with gearbox bearing fault and generator winding fault respectively for fault detection and identification. The proposed method is then compared with three other well-established machine-learning algorithms to investigate its superiority. The results show that the proposed method can produce a more effective detection with accuracy reaching 94% and 92% for the turbines, respectively. Furthermore, the proposed method can effectively distinguish the alarms from the faults, from which the distinguished alarms can be considered as an early warning of the fault occurrence. Wind speed and active power output are selected as the model inputs, while temperature and pressure variables, which reflect the operation condition of the subsystem as a whole, are selected as the target output. The LSTM model is built based on those variables. The probability density distribution of prediction data and original data are then calculated respectively. The discrimination between the prediction and the real data is produced by calculating their KLD values. The normal, alarm and fault conditions are distinguished by introducing a two-level threshold strategy of the KLD values, which are defined as fault-free condition (H0) and fault condition (H1). [Display omitted] •A novel data-driven model-based CM method based on LSTM with KLD is proposed.•LSTM is adopted to capture relation features in temporal dependencies among monitoring data.•KLD is used as fault indicator to measure the severity of the fault by comparing probability distributions.•Optimised thresholds are determined to distinguish the normal, alarm and fault conditions.•The distinguished alarms are cross-referenced with SCADA alarm logs to provoke early warning of the fault.