A hybrid data-driven modeling method on sensor condition monitoring and fault diagnosis for power plants

• Mechanism analysis could guarantee all relevant variables to be obtained.
• The main modeling variables is selected by the method of MIV.
• The model has good generalization ability, simple structure, and fewer parameters.
• The method can solve the online real-time monitoring and diagnosis problems.
• The data of a gas turbine active power verifies the suitability of this method.

This paper proposes a new hybrid data-driven soft measurement modeling method for power plant sensor condition monitoring and fault diagnosis. The method integrates Generalized Regression Neural Network (GRNN), Mean Impact Value (MIV), Partial-Least Squares Regression (PLSR) and B-Spline transformation techniques. First, the relevant parameters are obtained from mechanism analysis and a GRNN model is built to assess the average contribution rate of each independent variable and filter out the main modeling parameters by method of MIV. Then, the main modeling parameters are modeled with a PLSR method based on the cubic B-Spline transformation, which is an effective approach to the nonlinear modeling and multicollinearity problems. The final reliable model is completed to monitor and diagnose the sensors. Taking the active power sensor of a combined cycle generator unit of Siemens V94.3A as an example, the computational result shows that this modeling approach to sensor measurement data fits well in both accuracy and generalization ability under different conditions. Through fault signs and fault diagnosis methods analysis, this model could accurately identify sensor fault types. Most importantly, only a few model parameters need to be saved, and the model has low computation cost and strong robustness. Therefore the model is more suitable in solving the online real-time monitoring and diagnosis problems.