Feature Extraction for Uncertainty Quantification and Reduction of Noise in Condition Monitoring Using Deep Learning

Abstract

The condition monitoring of structural and mechanical systems in an industrial facility is an important component of digital maintenance platforms. In a nuclear power plant, the structures, systems, and components are continuously monitored as a part of current maintenance procedures. Real-time sensor data collected from systems such as piping-equipment is fed into online condition monitoring programs. Deep learning algorithms such as neural networks can be used as powerful data-based engines to predict damage or aging in pipingequipment systems of nuclear power plants. Factors such as the quality and quantity of data available for training the deep learning algorithm can affect the performance of the condition monitoring framework. Typically, the computational efficiency of deep learning approaches is enhanced by utilizing signal processing and feature extraction techniques on data obtained from sensors installed on the system. This paper presentes a novel feature extraction methodology to quantify uncertainty in the acquired sensor data for the condition monitoring of nuclear piping-equipment systems. Uncertainty in the severity of degradation is classified as minor, moderate, and severe. Signal processing and feature extraction are implemented to calculate a vector of degradation-sensitive elements of interest and reduce the effect of noise in the data. Artificial neural networks are designed and trained using the extracted vector from the sensor response. The condition monitoring framework is tested on various nuclear piping-equipment configurations as well as load combinations.