A Method for Predicting Steam Turbine Generator Faults Based on MI

To improve the accuracy of fault prediction for power generation steam turbines and address the challenges associated with high-dimensional, nonlinear monitoring data and cumbersome hyperparameter tuning, this study proposes an intelligent fault prediction method. Although mutual information (MI)-based feature selection and Bayesian optimization (BO) for hyperparameter tuning have each demonstrated individual success in fault diagnosis applications, existing approaches predominantly treat these two critical aspects as isolated and independent procedures. This separation limits the synergistic potential between feature quality and model configuration, leaving a gap in coordinated, fully automated fault prediction frameworks for steam turbines. To bridge this gap, the proposed method, termed BO-CNN-BiLSTM, presents an automated pipeline that sequentially integrates MI-based adaptive feature selection with Bayesian optimization for hyperparameter tuning of a CNN-BiLSTM network. Initially, MI combined with K-means clustering automatically identifies and retains key features strongly associated with fault states, effectively reducing input dimensionality. Subsequently, a BO framework is employed to autonomously search for the optimal hyperparameter configuration, achieving seamless integration from feature selection to model optimization. Validation via a self-built physical-information fusion experimental platform demonstrates that the optimized model attains a root mean square error (RMSE) of 0.324 and a coefficient of determination (R2) of 0.888 on the test set. Its predictive performance significantly surpasses that of models lacking Bayesian optimization, as well as those employing standalone CNN or BiLSTM architectures. This study thus presents a highly automated, accurate, and practical intelligent fault prediction scheme for steam turbines in power generation.