تجزیه و تحلیل گسل های یاطاقان در توربین های بادی: یک روش داده کاوی

Bearings are an essential part of turbine generators and gearboxes. Dynamic and unpredictable stress causes the bearings to wear prematurely, leading to increased turbine maintenance costs, and could lead to sudden, expensive turbine breakdowns. Over temperature impacts the performance of turbine bearings. In this paper, data mining is applied to identify bearing faults in wind turbines. Historical wind turbine data are analyzed to develop prediction models for bearing faults. Such models are generated by neural network algorithms, using data from 24 turbines collected over a period of four months. Models predicting normal behavior are constructed. The performance of the models is validated on different wind turbines with over 97% accuracy. The model error residuals are analyzed using moving average windows to predict the occurrence of over-temperature events. Five over-temperature events are analyzed. The research reported in this paper has led to the prediction of faults 1.5 h before their occurrence.

Wind energy is one of the most viable sources of renewable energy. The growing number of wind farms has increased the importance of their operation and maintenance [1]. Due to the variability of wind, shifting loads, and fluctuating energy demands, components of wind turbines such as gearboxes and generators are susceptible to damage. The replacement cost of failed high-value components can be high. In the event of faults, the monitoring systems of wind turbines issue alarms. However, such alarms are usually signaled once damage to the component has already occurred. There is a need to find solutions for predicting faults ahead of time to avoid extensive damage to turbine components. Data-mining algorithms build fault prediction models using data collected by supervisory control and data acquisition (SCADA) systems. Such data—e.g., power output, gearbox bearing temperature, and generator speed—are usually acquired for over 100 turbine parameters. Traditional condition-monitoring systems are imperfect and can be costly. Data-mining algorithms use historical wind turbine data to predict faults. Such algorithms have been successfully applied to predict faults across various wind turbine components such as: turbine blades [2], generators [1] and [3], and gearboxes [4] and [5]. The results presented in this paper are based on data collected at twenty-four 1.5 MW wind turbines. The parameter values recorded at 10-s intervals (10 s data) over a period of four months constitute the dataset used in this current research.

In this paper, a data-mining approach was applied for the detection of generator bearing anomalies. Five neural network algorithms were optimized to capture the relationship between input parameters and generator bearing temperature during normal behavior. The neural network models were first tested on two different turbines for model validation. The error residuals were analyzed using an average moving window of size 360. The five over-temperature events were tested using the normal behavior model. The over-temperature events were predicted 1.5 h ahead of the fault occurrence. This research will be advanced in the future once more data becomes available for detection of gearbox, shaft-bearing anomalies.