Leakage-aware multi-condition fault diagnosis of centrifugal pumps using sensor-sensitive vibration features, random forest, and multichannel 1D-CNN

Reliable centrifugal-pump fault diagnosis requires accurate classification supported by physically interpretable evidence. This study presents an integrated vibration-based framework combining analytical signal processing, sensor-sensitivity analysis, trial-level statistical validation, feature-based machine learning, and a five-channel one-dimensional convolutional neural network (1D-CNN). The University of Twente Motor-2 dataset, comprising ten health conditions and three operating levels, was analysed using five accelerometer channels. Following preprocessing and quality control, 948 independent trials were used for paired model development. Statistical analysis showed that 74.4% of fault-versus-Normal comparisons remained significant after Benjamini-Hochberg correction, while 67.4% exhibited medium or large Cliff's delta effects. Across five training repetitions, the 1D-CNN achieved mean windowand trial-level accuracies of 99.76 ± 0.26% and 99.90 ± 0.23%, respectively. On 191 independent test trials, the speed-excluded Random Forest achieved 98.95% trial accuracy, whereas the selected CNN correctly classified all trials. The proposed framework integrates interpretable vibration analysis with robust multichannel deep learning for reliable centrifugal-pump fault diagnosis.