Data Augmentation Using Spectral Failure Deltas to Diagnose Bearing Failure

Labeled training data are challenging to obtain in a manufacturing environment during production due to the time and cost constraints of the labelling process. Of the labeled training data that is collected, failure data comprises a small proportion or is non-existent in production datasets for condition monitoring. The small proportion can be related to failures occuring uxpectedly and parts are replaced quickly, meaning the failure state is rare and makes up a small portion of the run life and number of samples collected. The lack of labeled data and failure data leads to challenges in creating effective predictive systems, such as Digital Twins, to accurately determine equipment health state and remaining useful life. This work investigates training predictive algorithms using an augmented failure data set derived from laboratory systems with knowledge of real-world failures. Data are collected under different failure progressions and operating conditions to create variability for the variety of different production applications to apply these data augmentation methodologies. These same data are transformed by adding the variability measured through purposefully damaging the mechanical system to create the degraded and failed state data. This variability is extracted using a spectral augmentation technique on the surrogate system’s failure data under an artificial fatigue case. The fatigue case is created by incrementally damaging the bearing raceway and measuring the damaged surface area with respect to the total bearing raceway. The measured difference between these pre- and post-lab damage states is used as the damage state data set transformation function. The augmented and “true” data are then compared using class probability analysis and diagnosing particular failure instances. For future research, relatability analysis will be investigated to see how the effects change between bearings of different sizes.