Architecture Engineering and Science

Centrifugal pumps are indispensable fluid conveying equipment in key industrial fields such as nuclear power plants and petrochemical engineering. They often operate under high-temperature, high-pressure, and high-speed conditions for long periods, making them prone to faults such as impeller fracture and bearing wear. The vibration signals of centrifugal pumps exhibit non-stationary and nonlinear characteristics and are easily affected by strong background noise. In this paper, the CEEMDAN algorithm is adopted to perform adaptive decomposition of the original vibration signals, yielding multiple intrinsic mode function (IMF) components. The first five effective components are selected based on energy ratio and kurtosis criteria. Time-domain statistical features, frequency-domain energy features, and energy entropy are extracted from the selected IMF components to construct an initial high-dimensional feature set. Principal component analysis (PCA) is then applied for linear dimensionality reduction, extracting principal components with a cumulative contribution rate greater than 95%. The reduced feature sequences are subsequently input into a channel attention mechanism-based AC-CNN model for deep feature learning and classification recognition. Vibration data under four typical operating conditions are collected using a centrifugal pump fault simulation test rig. The proposed model achieves an overall diagnostic accuracy of 98.6% on the test set, which is 4.3% higher than the traditional CEEMDAN-CNN method and 2.1% higher than the single-layer AC-CNN model.