Multi-sensor centrifugal pump fault dataset: offline recordings, window features and logged edge features

This dataset accompanies the paper “Machine Learning-Based Fault Diagnosis for Centrifugal Pumps: From Offline Training to Real-Time Inference Validation” (TSME-ICoME 2026, Paper ID TSME-2026-0132). It contains the raw recordings from a test rig with three parallel centrifugal pumps under one normal and five induced fault conditions (180 two-minute recordings, nine sessions, 27 January – 12 February 2026), the 38 window features with the train/test split used in the paper, the recording session of every file, and the feature vectors logged on a Raspberry Pi 4 during the live test, which were replayed through the trained model. See README.md for the file structure, column definitions and known limitations (no common clock between nodes, datasheet calibration, P1_Bar not a calibrated pressure, training–serving skew in the logged vibration RMS and crest factor).