Vibration Fault Diagnosis for Steam Turbine-Generators Based on History Cases and Artificial Neural Network

In this work, automatic fault diagnosis of a steam turbine-generator system based on the vibration data is studied. In practice, the sampling rate of the data might be different. In this work, when the sampling rate is high, we show that the renowned empirical mode decomposition (EMD) works well to extract the features, which further are used to train a neural network. However, in our case only vibration data with a relatively low sampling rate is available for the two most common faults, i.e., rotor unbalance and rubbing between the rotor and cylinder. To copy with this situation, we propose a novel data preprocessing method and a novel statistic feature. The testing results show that the neural network with features extracted by our method outperforms significantly the ones with the EMD and the classical statistic features. Our method is simple to implement and has low computational complexity. It can be utilized to mitigate the disturbance caused by the different values of signal when the system is of the same state and to capture the temporal relation between different inputs.