Characterizing Chaotic Systems (Chattering Valves in Nuclear Reactors)
Blue Wave AI Labs partnered with Southern Nuclear Power (SNP) to detect chattering in Boiling Water Reactor (BWR) safety valves and their component failure, as faulty reactor components can lead to downtime and consequently millions of dollars in lost revenue.
Under certain operating conditions in a nuclear power plant’s boiling water reactor (BWR), a condition arises where the pilot stages of the three-stage safety relief valves chatter sporadically. This chattering degrades the valve’s performance and eventually leads to discharge-stage leaking. The cause of this behavior is still unknown, and it used to be impossible to determine which valve was chattering.
But technical experts are now able to use real-time sensors to gather data on the system. After collecting data and training an algorithm, our team was able to identify chattering valves in real-time with 99% accuracy. After detecting a problematic valve with our data, we worked with the plant to replace that valve and substantially decrease its chattering.
Southern Nuclear Power needed a way to detect chattering
- Using SRV temperature sensor waveforms
- In order to adjust reactor parameters to mitigate chattering
SNP provided operational plant data
- Blue Wave received 1.8 terabytes of historical plant data, spanning back to 2014
Blue Wave created ML classifiers to detect chattering
Blue Wave developed ML ‘predictors’ of SRV chattering from plant sensors
- These identify plant parameters that accurately model SRV chattering
Chattering leads to deterioration and leakage
- Increased wear on SRVs
- Degraded discharge seals
- Contaminated steam enters discharge lines & suppression pool
- Increased radiation exposure for personnel
Financial Impact
- Upwards of $3 Million in lost generating revenue when SNP has shutdown as a result.
- Continual maintenance, repair, and/or replacement costs resulting from SRV chattering
Chattering Detection
- Our classifier identifies chattering in near real-time (as tested on historical data)
Greedy Analysis
- We developed a classification model to search for plant parameters that are able to identify chattering
- Strongest predictors of chattering determined iteratively
Current Outcomes
- Blue Wave has created ML classifiers for the identification of chattering in near real-time (as tested on historical data)
- Developed ML models to identify root-cause candidates
- Wind speed & direction are strong candidates
