January 29, 2021 — Industry Trends
Autoencoders: The next leap forward in fault detection?
The number of sensors in wind turbines keeps growing and will open up new opportunities for monitoring and optimization in the years ahead. In this article, we look at the role that autoencoders could play in detecting faults that operators might otherwise miss.
Head of Research, Greenbyte
The numbers of sensors in wind turbines keep growing. This could mean major headaches for operators that need to identify faults in hundreds of machines across their portfolios.
There are solutions, of course. In recent years, we’ve seen many operators take huge steps to use digital systems to better understand their turbines. These systems are helping them to make the right decisions so that they can boost power output and project profitability.
There has been great progress.
Even so, we must be aware that there will continue to be new approaches and technologies that help unravel the mysteries of a wind turbine. In this piece, we look at the next potential leap forward: autoencoders. We reported in the journal Renewable Energy last year on how autoencoders could help to further improve returns from wind turbines in the year ahead.
We’re going to give you a short introduction here. If you’d like further details, you can read a more detailed explanation in our ebook here.
How can we improve current processes?
Due to the large scale of modern wind farms, manual inspection of all turbines is a labour-intensive process. It is also complex and, with the time pressures on technicians, it may be challenging to do a thorough job. That’s where automated systems have been helping.
Over the last decade, companies including Greenbyte have made huge strides in building systems that detect anomalous behavior in turbine components that could lead to costly failures. Broadly, these anomaly detection systems are developed in one of three ways.
First, those producing the system do so based on their physical understanding of how each of the parts of the turbine connect and relate to each other.
Second, they produce the system based not on their physical understanding of the turbines, but on statistical methods that mimic those relationships.
Or third, they use ‘artificial neural networks’, which use nodes at key points in the turbines that act as simplified brains and are intended to replicate the ways humans learn. These are generally better at identifying changes in a turbine that are outside seasonal variations.
There are no rights and wrongs here. However, the problems with these models is that they each reconstruct a series of single signals between the various components, and the system only sees a fault when it affects one of those signals. Turbines are more complex than that.
What are autoencoders?
An ‘autoencoder’ is a type of artificial neural network that models and interprets data, and is programmed to ignore unwanted fluctuations in the signals.
In more detailed terms, this autoencoder is a neural network that tries to map the current performance of a turbine onto a model of how that individual turbine should theoretically be performing, and then flag when something is wrong.
The key word there is ‘individual’. Every turbine performs a little differently, even if it is the same model and from the same factory. The autoencoder means that each turbine has its own notional model about how it should be performing, while ignoring the signal ‘noise’.
Since autoencoders reconstruct all of their input signals, they are also capable of detecting many faults in a wind turbine at once. This enables us to go into far more detail about how the individual machines are performing and could be improved.
How could autoencoders help?
Our research so far has been very positive for operators. We have seen that autoencoders are successful at detecting issues with turbine blades and sensors, including:
- Yaw encoder malfunction
- Rotor sensor malfunction
- High generator temperature due to a malfunctioning ventilation duct
- Preventive maintenance that change the hydraulic oil temperature
- High VCP temperatures
- Grid curtailment
We should also say that research is still at an early stage, but we are now fairly confident that a fault that leaves traces in SCADA data would be detected by an autoencoder too.
Essentially, an autoencoder would be able to do what SCADA does and much more besides. In an era where operators want to squeeze more electrons and profits from their turbines, we believe this could lead to further advances in turbine monitoring in the years ahead?
Do you want to find out more? Download our ebook to find out more about our research.