SAFE: Statistical Analysis and Integration of Laboratory and Field Sensor Data for Gearbox Health Monitoring

Predictivemaintenanceforindustrialgearboxesrequiresmodelsthatcandetect early signs of mechanical degradation from sensor data, but deploying such models across different installations is difficult when the available laboratory and field datasets differ in sensor type, placement (internal versus external), sampling rate, and acquisition period, especially when the two acquisitions are separated by years with no temporal overlap or shared infrastructure. Even defining which sensors correspond to which requires inference from distributionalpropertiesrather thandirect alignment. A preprocessing pipeline applies Otsu’s method to the motor current dis- tribution to automatically separate operational from non-operational periods, producingstationarytimeseriesfreeofidle-statecontamination. Asensorcor- respondence mapping based on Wasserstein distance and physical reasoning then identifies four matching sensor pairs between the Bonfiglioli laboratory testbench (2022–2023) and the Tampieri field installation (2024–2025), pro- ducing a common five-sensor format with a shared temporal index. Using the aligned representation, the anomaly detection pipeline progresses from Ridge regression and conditional kernel density estimation to a conditional NeuralSplineFlowthatmodelsthejointprobabilityofvibrationobservations conditionedonlaggedsensorvalues. A head-to-head comparison on labelled accelerated-fatigue data shows that Ridge regression detects 1.0% and the conditional KDE detects 0.8% of degraded observations, while the normalizing flow detects 66.5% ±31.7% across five random seeds. The prediction-based methods fail because their laggedfeaturesshiftintandemwiththedegradation,causingthemodeltotrack thefaultratherthanflagit. Jointtrainingonlaboratoryandfielddatatogether achieves 59.1% ±12.3% detection; the mean is lower than the laboratory- only model because the field data broadens the learned normal envelope, but seed-to-seedstandarddeviation dropsbya factor of2.6.