Canned motor pumps serve as critical heat circulation equipment in the water cooling system of the China Spallation Neutron Source (CSNS) target station, operating continuously with high-radiation, high-reliability, and strong electromagnetic interference constraints. Conventional single parameter monitoring and offline periodic inspection fail to capture incipient electromechanical degradation due to the coupled multi physics nature of pump failures and the inaccessibility of the radiation field during beam operation. To address these limitations, this study proposes an intelligent health monitoring system based on multi-parameter fusion, integrating vibration, current, and temperature sensing modules to synchronously acquire mechanical dynamic responses and electrical operating characteristics. A layered decoupling architecture is established with lightweight edge sensing and centralized intelligent computation, where the hardware layer undertakes signal transduction and analog-to-digital conversion while the software layer performs multi-domain feature extraction and statistical decision-making. The system implements a dual-branch parallel processing framework: the real-time branch extracts time-domain statistical indicators (mean, RMS, kurtosis, skewness) and frequency-domain spectral features via Fourier transform, combined with an adaptive 3σ dynamic threshold discrimination scheme and a triple consecutive over-threshold confirmation mechanism to suppress radiation-induced noise; the archiving branch persistently stores raw data and feature sequences, enabling incremental closed-loop self-learning for baseline threshold updating. A unified anomaly detection operator A(Ψ) is formulated to encapsulate the complete monitoring pipeline from raw signal acquisition to temporal consistency verification. The integrated platform has been successfully deployed at the CSNS target station, demonstrating stable long-term unattended operation, effective incipient anomaly identification, and robust false-alarm suppression under harsh radiation conditions. This work provides a rigorous theoretical foundation and practical engineering paradigm for transitioning from scheduled maintenance to proactive predictive maintenance of high-reliability electromechanical equipment in extreme environments.

canned motor pump; multi-parameter fusion; predictive maintenance; adaptive threshold; statistical process control; high-radiation environment; closed-loop iteration