Wind turbines are big and expensive machines, so keeping them running smoothly helps keeping their operational cost down. The sensor data generated by the turbine can help achieving this – by analysing it, you can spot potential failures earlier. The longer the warning period before a part fails, the better you can prepare for it.
Main parts of a wind turbine
The business case
Preventive maintenance saves you money when you have:
- Shorter downtime and less lost production
- Better planning of people and materials
- Cheaper repairs
To do that, you need to be able to anticipate failures in heavy and expensive parts like the gearbox, generator and main shaft.
For Vattenfall we, Algoritmica, have implemented a solution based on this concept that is currently monitoring 78 turbines. It provides a daily overview of the turbine status, helping Operations staff to prioritize inspections and investigate potential problems.
Shortly after going live, this solution helped identify a bearing defect which could be claimed under the manufacturer’s warranty, thereby partially paying for itself.
How it works
Wind turbines have an array of sensors that measure temperatures, pressures, voltages, currents, and blade angles. This data is available for analysis, typically as 10-minute averages of the sensor values.
The computer that controls the turbine uses these measurements for its operations. This includes error thresholds like ‘the gearbox oil temperature should be below 120 degrees Celsius’. However, by the time the threshold is exceeded it is usually too late: the damage has already been done. To catch failures earlier we should look for anomalies, e.g. measurements that are unexpected and therefore might indicate a problem – but are not yet so severe that they exceed a threshold.
A typical wind farm has about 30 turbines, with around 50 sensors each. That means about 200 thousand new measurements every day. Clearly, looking for anomalies is not something a human analyst can do manually; fortunately we can automate this.
Anomaly detection begins by defining what measurement values are expected and then calculating the difference with the actual situation. Since sensor data is delivered as a time series, we create a model that predicts the next value of a specific sensor given its previous values as well as the previous values of any other sensors that may be relevant. Based on these multiple inputs, the model then calculates its predicted value and compares it with the actual sensor reading. The difference (or residual) is now a measure of how much the turbine is deviating from its expected performance. If it is persistent or grows too large (i.e. becomes an anomaly), an analyst can investigate the cause and decide on a course of action together with the Operations staff at the wind farm.
The anomaly detection process
To create such a sensor model we apply machine learning, i.e. one or more algorithms that use a set of examples (the ‘training set’) to learn a predictive model. For a wind turbine, it is a natural fit to use a year of sensor data as the training set so that all seasonal variation is included. Since each turbine can have its own training set we can train a model for each individual turbine. This has the advantage that we automatically take into account the influence of the specific location and turbine-specific parts. Note that we have to be careful in selecting the training period: if the turbine had a defect during that time, the model will learn from bad examples and think the defective situation is normal.
Driven by data
This is a data-driven approach: the model learns the relationship between the various sensor readings purely based on the training data. This is in contrast to a so-called physical model that explicitly describes the turbine design using detailed knowledge of its physical characteristics. The main advantage of a data-driven approach is that the model can be trained by a non-turbine expert and matches the actual situation by definition, whereas a physical model has to be carefully calibrated by an expert.
As shown in the illustration, the data-driven approach is generic. This means you can apply it to other machinery like trains, cars or printing presses – anywhere it makes sense from an economic or safety perspective to know the condition of your assets.
Please contact us if you want more information.