Deep-Learning Analysis of Acoustical Signatures Recorded in Nuclear Reactor

The transition to condition-based, risk-informed automated maintenance will contribute to a significant reduction in operations and maintenance costs that account for the majority of nuclear power generation costs in the United States (U.S.) and around the world. This transition will, however, require the application and validation of new technologies that can handle and extract meaningful information from large amounts of data recorded at commercial nuclear power plants. These technologies must be able to automatically recognize different operating conditions and advise operators on the state of the reactor so that an operator can make informed decisions and execute appropriate actions. Modern digital instrumentation systems can collect large volumes of time-series data; however, due to the amount of data, its analysis is beyond human capabilities. The recently emerged neural network-based deep-learning paradigm has shown promise in processing large amounts of imaging data with a high recognition rate. However, there are not many publications in the literature regarding the application of these deep-learning networks to challenging time-series data recorded in industrial settings for prognostics and health management. This paper describes the application of the deep-learning paradigm and other machine learning (ML) algorithms to the pattern recognition of acoustical signals, which were collected from five accelerometers installed at different locations within the nozzle trench area of the Advanced Test Reactor. The signals used for analysis in this paper had been recorded continuously in 15-s periods for 18 days. The reactor underwent several changes in its operating conditions throughout those 18 days, so the primary goal of the project was to research capabilities of ML methods and deep-learning networks to detect different operating conditions based solely on the information provided through the acoustical sensors. The obtained results show that conventional ML methods such as support vector machines and shallow neural network perform as well as or better than the deep-learning network on analyzed patterns with recognition accuracy of 99% or better. Also, the analyzed data linear classifier demonstrated very good performance. However, the deep-learning network was the only tool that achieved 100% performance for some sensors. In addition, the paper reports the results on the consistency of different ML algorithms including deep-learning networks. The results presented here show that traditional ML methods, as well as deep-learning networks, demonstrate repeatable and consistent results regardless of small variations in training data or initial training conditions. The paper concludes with an analysis of the obtained results, as well as their limitations, recommendations, and directions for future research.