Deep Neural Networks for Predictive Maintenance of Wind Turbine System

Renewable energy resources (RER) have gained tremendous popularity over the last decade due to the demand for energy sustainability and lower energy generation costs. Solar and wind are the two most dominant resources; wind turbines (WT) can produce energy around the clock. However, the corrective maintenance (failure-based) of the WT becomes very challenging due to their accessibility and weather conditions, which lengthen the outage duration and increase expenditure. Besides, preventive maintenance (time-based) is sometimes unnecessary as the maintenance is scheduled before it is needed. In response, predictive maintenance (PM) offers an economical solution to overcome the mentioned challenges for WT systems. In the WT systems, gearbox bearing failure is one of the most prevailing issues that seriously affect the system's performance and stability. This article proposes an intelligent PM approach employing deep learning methods for the gearbox-bearing using the data acquired through the WT supervisory control and data acquisition (SCADA) data system. The developed model efficacy is evaluated based on the statistical performance metrics (SPM). The satisfactory values of the SPM indicate the effectiveness and strength of the developed models. A post-processing analysis is also carried out using the Shewhart control chart (SCC) to determine the deviation of predictions from the actual SCADA data measurement to identify the PM schedule. The developed models accurately predict the failure event at least 72 hours before it occurs, as reported in the wind farm maintenance logbook. This predictive maintenance approach could potentially lead to significant cost savings, offering an optimistic outlook on the economic benefits of our proposed approach.