A Machine Learning approach for damage detection and localisation in Wind Turbine Gearbox Bearings

Increasing demand for renewable sources requires more cost-effective solutions to mitigate the cost of maintenance and produce more energy. Preventive maintenance is the most normally adopted scheme in industry for maintenance but despite being well accepted has severe limitations. Its inability to intelligently schedule maintenance at the right time and prevent unexpected breakdowns are the main downsides of this approach and consequently leads to several problems such as unnecessary maintenances. This strategy does not justify the additional costs and thereby represents a negative aspect for renewable energy resource companies that try to generate cost-competitive energy. These challenges are progressively leading towards the predictive maintenance approach to overcome these aforementioned issues. Wind Turbine Gearbox Bearings have received a lot of attention due to the high incidence failure rates provoked by the harsh operational and environmental conditions. Current techniques only reach a level one of diagnostics commonly known as the Novelty Detection stage and normally requires the expertise of a skilled operator to interpret data and infer damage from it. A data-driven approach by using Machine Learning methods has been used to tackle the damage detection and location stage in bearing components. The damage location was performed by using non-destructive methods such as the Acoustic Emission technique — these measurements were used as features to locate damage around the bearing component once the damage was detected. The implementation of this stages also led to the exploration of damage generation due to overload defects and proposed a methodology to simulate these defects in bearings — the study of this concept was implemented in a scaled-down experiment where damage detection and localisation was performed. Due to the importance of the implementation of a damage location stage, damage in AE sensors was also explored in this work. Features extracted from impedance curves allowed to train Machine Learning methods to trigger a novelty when a bonding scenario occurred. This ultimately allowed the identification of unhealthy sensors in the network that could potentially generate spurious results in the damage predictions stage.