Fault diagnosis of gas turbine based on feature fusion cascade neural network

Gas path analysis methods impose strict requirements on the number of measurement points. The multi-operating point method somewhat addresses this by using adjacent operating data, yet it relies on the assumption of uniform degradation levels and may not converge when applied to data from before and after abrupt faults and is time-intensive. To overcome these limitations, this paper introduces a feature fusion cascade neural network to diagnose both gradual and abrupt faults. Firstly, a component-level model was used to generate fault data. The multi-operating point method is then used to support fault diagnosis and provide training and testing data for the proposed method. Lastly, a novel cascade neural network architecture was proposed and the feature fusion methods including feature analysis and feature generation were adopted and incorporated into the architecture to improve the fault diagnosis accuracy. Compared with other data-driven models, it exhibits higher prediction accuracy, notably in SEC. The proportion of diagnosis error under 2×10−4 among all diagnosed values for gradual and abrupt faults diagnosis is 52.56 % and 51.18 %, respectively. These findings support the high accuracy and efficiency of the proposed method.