Fault diagnosis of hydro-turbine via the incorporation of bayesian algorithm optimized CNN-LSTM neural network

Abstract

The hydro-turbine is the core equipment of the hydropower station, and it is essential to diagnose and identify its faults. A fault diagnosis model based on Bayesian optimization (BO), which incorporates convolutional neural network (CNN) and long short-term memory (LSTM) methods for the hydro-turbine, is proposed (BO–CNN-LSTM). CNN adaptively extracts and down-scales fault features, fed into the LSTM model for feature learning and training. The BO algorithm is employed to address the challenge of model hyperparameter selection. A hydro-turbine fault experiment bench is constructed to train and validate the model. Experimental results demonstrate the superior performance of the proposed BO-CNN-LSTM model in hydro-turbine fault diagnosis, achieving accuracies of 92.7 %, 98.4 %, and 90.4 %, respectively, surpassing CNN, LSTM, and CNN-LSTM models. The BO-CNN-LSTM model improves accuracy by 5.5 %, 6.3 %, and 9.0 %, respectively, Compared to the unoptimized CNN-LSTM model. The BO algorithm is introduced to optimize CNN-LSTM from the perspective of acoustic vibration signals, which can be a beneficial supplement to the existing hydro-turbine fault diagnosis.