Multi-source monitoring data cleaning-based hydraulic turbine deterioration trend interval prediction

Abstract The effective evaluation and prediction of the comprehensive deterioration state of the hydropower unit can guarantee its safe and stable operation and enable intelligent maintenance for the hydropower station. However, there are still two issues that require further study. (1) It is challenging to extract effective information that characterizes the unit’s status from multi-source monitoring signals based on working condition information. (2) Most neural network-based prediction methods cannot directly quantify the uncertainty information associated with the deterioration state prediction of hydropower units. To address these issues, this study proposes a multi-source monitoring data cleaning-based approach for the interval prediction of hydraulic turbine deterioration trends. First, the amplitude of the three-dimensional sample space of monitoring vibration data is calculated using the sparse spectral Gaussian process regression method. By calculating the upper <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" overflow="scroll"> <mml:mrow> <mml:mn>3</mml:mn> <mml:mi>σ</mml:mi> </mml:mrow> </mml:math> vibration surface of the lower frame in X direction and considering the comprehensive operational characteristics of the hydropower unit, the unit’s operating area is divided. Then the hierarchical density-based spatial clustering of applications with noise method is applied to adaptively cluster data of varying densities within each working region. Second, a dual-channel convolutional autoencoder model is constructed to explore the mapping relationships between working conditions and vibration signals, as well as pressure pulsation signals, respectively. The signal weight is calculated using self-adjusting analytic hierarchy process, and a comprehensive degradation index (CDI) is constructed after weighted calculation. The characteristic relationship between multiple source signals to accurately characterize the deterioration state is obtained. Third, in order to ensure the accuracy and reliability of prediction, we construct the temporal pattern attention-temporal convolutional network-Monte-Carlo dropout model and directly obtain the degradation state of hydropower units with prediction intervals through the neural network. Finally, the effectiveness experiment is conducted using the field monitoring data of a Francis turbine unit in Sichuan Province.