Fault diagnosis of gas turbine based on partly interpretable convolutional neural networks

The fault diagnosis of gas turbines plays a vital role in engine reliability and availability. The data-driven diagnostic model has been verified useful for identifying and characterizing engine degradation. But Convolution Neural Networks (CNN) is considered to perform poorly in the fault diagnosis for time-series signals. And most of the studies do not involve the interpretability of CNN, leading model hard to be optimized and integrated with physical mechanisms. For the fault diagnosis of the gas turbines, strong coupling often exists between gas path faults and sensor faults, making fault diagnosis difficult when both faults occur simultaneously. A novel method is proposed to improve the performance of typical CNN through optimizing the influence of input measurement parameter sequencing. Extreme Gradient Boosting (XGBoost) is used to make the effects of the sequencing on CNN diagnostic accuracy interpretable. In the simulation experiment, the diagnostic accuracy of CNN after optimization is 95.52%, higher than that of conventional CNN (accuracy rate 91.10%), RNN (accuracy rate 94.21%) and other methods. For the analysis of field data, the new method has shown stronger feature extraction ability and can detect typical gas path faults in advance. The new method performs well in precision, stability, and comprehensibility.