Fault signal reconstruction for multi-sensors in gas turbine control systems based on prior knowledge from time series representation

Abstract

Improving efficiency through intelligence is the current development trend of industrial gas turbines. Among the fault statistics of gas turbines, the number of sensor fault is the highest during use. The fault signal diagnosis and reconstruction are of great significance to the efficient and safe operation of gas turbines. In order to eliminate the sensor fault signal and transmit the normal signal to the control system, a multivariate fault signal reconstruction method based on the prior knowledge of the time-series representation was proposed in this work. The proposed multivariate signal reconstruction method can reconstruct almost all fault cases with high accuracy by training only one model. Firstly, the prior knowledge is applied to improve the conventional time series data representation. Secondly, three steps are employed to utilize spatial or temporal information and obtain three intermediate data. The masks combine the third intermediate data and the original time series to obtain the final reconstruction results. Then, reconstruction data sets are built based on exhaust gas temperatures (EGTs) from real-world power plant to verify the effectiveness of the proposed method. Different evaluation metrics and visualization reveal the high accuracy results. Compared results with different reconstruction algorithms reflect both the robustness and high speed of this reconstruction method. Finally, three typical fault signal reconstruction cases reveal the generalizability of this model.