Machine learning based online monitoring system aided by adaptive data acquisition with mobile robots in nuclear power plants

Online monitoring systems powered by a number of sensors installed over the facilities have been an essential component in securing safety and efficiency in nuclear power plants (NPPs) for decades. As more powerful machine learning techniques emerge, numerous studies have explored automated data driven diagnosis with online monitoring systems. However, due to limited information from finite number of sensors, human workers are often dispatched to the vicinity of the facilities for further inspections. To fully automate such processes with robots, this research presents a new monitoring system based on machine learning with adaptive data acquisition from mobile robots. Along with the data constantly streamed from fixed sensors, the proposed system actively utilizes data from robots for improved diagnosis. The proposed system is supported by of three major components: an anomaly detection and localization method to identify malfunctioning subsystem in a large scale NPPs which requires further inspection by robots, a measurement selection algorithm based on Bayesian inference for efficient diagnosis improvement, and diagnosis with non-synchronous data from mobile robots to compensate the time difference between measurements. The research also illustrates some key simulation techniques not only to improve dataset’s diversity with various environmental conditions and malfunctioning scenarios, but also to evaluate the developed machine learning models under different levels of faults. The proposed research will stimulate utilization of robots in many industrial facilities including but not limited to NPPs, which improves the efficiency and reliability while reducing labor work in harsh environment.