Objective Predictive maintenance is recognized as one of the core approaches to improving the reliability and operational efficiency of centrifugal pumps. A crucial component of implementing this strategy lies in the health status recognition of these pumps. However, traditional methods for health status recognition primarily depend on machine learning techniques that require a substantial amount of labeled data. This dependency often leads to their limitations in effectively and clearly representing relationships between monitoring data and health status, thus constraining their application effect in real-world conditions, which are typically complex. Therefore, it is urgently needed to develop health status recognition methods that are more accurate, efficient, and adaptable.

Methods This paper presents a health status recognition method for electrically driven centrifugal pumps, utilizing a cascade configuration of the Self-Organizing Map (SOM) neural network and the Back Propagation (BP) neural network. First, the SOM neural network was employed to preprocess the life-cycle vibration data of centrifugal pumps, focusing on extracting various statistical characteristics and entropy features in the time, frequency, and time-frequency domains, to fully characterize their operating states. Next, the extracted bearing vibration signal features were subjected to dimensionality reduction and integration using Principal Component Analysis (PCA). This facilitates the reduction of redundant information and computational complexity, the optimization of the pattern of input parameters, and the enhancement of modeling efficiency. Finally, a health status recognition model was established for electrically driven centrifugal pumps based on cascaded SOM-BP neural networks, leveraging the advantages of both neural networks.

Results Based on the health status monitoring dataset of an electrically driven centrifugal pump, a comparison was conducted between the SOM-BP model and common machine learning techniques, such as random forest and XGBoost models, to evaluate their accuracy in health status recognition of the pump, using R², MSE, and RMSE as evaluation metrics. The results indicated that the R², MSE, and RMSE values generated by the proposed model were 0.901, 0.8×10⁻⁶ m²/s⁴, and 9.12×10⁻⁴ m/s², respectively. These results were significantly superior to those obtained from traditional machine learning methods, demonstrating enhanced robustness and adaptability.

Conclusion The proposed method has demonstrated effectiveness in enhancing the accuracy of health status recognition for centrifugal pumps. Additionally, it provides valuable data support for fault diagnosis and residual life prediction of these pumps. The results of this study present a novel perspective on health status management and diagnosis that can be applied to other rotating machinery.