Health State Recognition Method for Electric-Driven Centrifugal Pumps Based on SOM-BP Cascaded Neural Network

Predictive maintenance of centrifugal pumps is essential for improving equipment reliability. Traditional machine learning methods for health condition identification often require large labeled datasets and fail to clearly link monitoring data with health states. This research proposes a health condition identification method using a Self-Organizing Map (SOM) and Back Propagation (BP) neural network cascade for electric-driven centrifugal pumps. The SOM network preprocesses vibration data, extracting features in time, frequency, and time-frequency domains. Principal Component Analysis (PCA) reduces dimensionality, optimizing input patterns and improving efficiency. A SOM-BP cascade neural network model is then developed, and its performance is compared with traditional machine learning methods (random forests, XG-boost). Evaluation metrics (R2, MSE, RMSE) demonstrate that the SOM-BP model outperforms these methods, achieving an R2 value of 0.901, MSE of 0.8×10⁻⁶ m²/s⁴, and RMSE of 9.12×10⁻⁴ m/s². This approach improves the accuracy of health condition identification and provides support for fault diagnosis and remaining useful life prediction in centrifugal pumps, with broader applicability to rotating machinery.