Assessing Signal and Operating-Condition Realism in Public Bearing Vibration Datasets: A Comparative Study with Real Operational Data

Publicly available bearing vibration datasets are widely used as benchmarks for developing condition monitoring and prognostic algorithms, yet the extent to which they represent the signal characteristics of real operational machinery has not been systematically investigated. This study presents a systematic comparison of five public benchmark datasets with vibration data acquired from industrial and maritime machinery operating under real service conditions. A unified signal-processing framework was applied across all datasets, including standardised segmentation, normalisation, and feature extraction in the time, frequency, and time–frequency domains. Degradation behaviour was further characterised using geometric descriptors of trajectory shape capturing trajectory regularity and smoothness. The analysis revealed consistent differences between laboratory-generated and operational vibration data, with public benchmark datasets generally exhibiting lower operating-condition variability and smoother degradation trajectories. Importantly, dataset realism emerged as a continuous characteristic rather than a binary property, with individual datasets occupying different positions along a realism continuum. An unexpected finding was that non-stationarity during nominal operation discriminated strongly between the two groups but in the direction opposite to that hypothesised, an effect attributed to latent degradation drift during accelerated laboratory testing. To quantify these aspects of realism along this continuum, a composite Realism Index (RI) was developed by combining two physically motivated and empirically complementary signal dimensions: operating-condition variability and degradation irregularity. Across 41 laboratory and 10 operational bearing runs, the RI separated the two groups with a large effect size (Cliff’s δ = 0.61, 95% CI [0.43, 1.00]), providing a quantitative framework for comparatively assessing the signal and operating-condition representativeness of benchmark datasets, with potential relevance to benchmark selection and evaluation practice in condition monitoring research.