Using Machine Learning to Mitigate Risk in Nuclear Power Plants

Steve Hess+ Jonathan Hodges

Share this post

Mar 1, 2021

Machine learning has been gaining a lot of attention in recent years. From Hollywood to university labs around the world, it’s no secret that machine learning — a subset of the artificial intelligence (AI) discipline — is here to stay . Before we take a deeper look at the application of machine learning, let’s examine the difference between machine learning and AI.

  • Artificial intelligence is a branch of computer science that refers to the simulation of human intelligence in machines.
  • Machine learning is an application of artificial intelligence that focuses on algorithms which allow machines to automatically learn and improve from experience.

Machine learning has many applications, and we regularly interact with it through our social media platforms, online stores, smartphones and internet search engines. What is less widely known, however, is how machine learning can be applied in the industrial and power generation world. In the nuclear power industry, machine learning is changing the way we address potential risks and hazards today.

Artificial Intelligence in Nuclear Power Plants to Identify Failures

Each year, a single nuclear power plant can generate anywhere from 10,000-20,000 condition reports which are documentation of events, equipment and personnel incidents near the plant. The information covers a wide range of possible situations where the safety of the plant may have been impacted, including valve failures or leaking lube pumps. Currently, an engineer must review each condition report to confirm if a functional failure with high safety significance has occurred. These functional failures could be as obvious as a loss of power to a critical reactor component or more subtle such as a part not operating at full capacity. Because a functional failure is not always readily apparent, the review process is extremely resource intensive and costly for the facility. These equipment failures have direct impacts on regulatory costs, maintenance costs, and if not appropriately addressed, can impact the continued safe operation of a plant.

Machine learning is being used to automate this process. It uses a computer to classify whether a condition report ("CR") corresponds to a functional failure based on the CR text. A neural network, or artificial brain, is developed and operates like a collection of strings on a guitar that need to be “tuned” to provide a clear sound. The strings are the network itself, and the tuning pegs are the individual components of the network that can be adjusted to provide a different pitch or outcome.

Previous CRs can be used as a reference point, much like a tuner that emits the correct pitch, to train the network. The network conducts a feedback loop after each adjustment, continually trying different combinations until it is perfectly tuned, or in this case, can correctly classify a CR as a functional failure or not. By analyzing the output from the CR, the neural network can predict future functional failures on an independent set of CRs with zero missed high safety significance CRs and less than 15% false positives to be reviewed by the plant.

While machine learning has certainly provided us with convenient shortcuts in our everyday lives, it can also be used in behind-the-scenes applications — particularly power generation — to realize huge cost and time savings, not to mention protect lives. Learn more about machine learning and other safety software applications for power plants.