Shortcut controls for bearing-wise fault-diagnosis benchmarks: a five-dataset case study with an order-resampled envelope-spectrum network

Bearing-wise data partitioning has become the accepted remedy for the leakage that inflates fault-diagnosis results, but a bearing-wise split alone does not guarantee that a model solves the task it is credited with. This study turns that caveat into a set of controls and applies them to five public datasets (a Paderborn bearing set used for development, CWRU, UORED-VAFCLS, XJTU-SY and a Paderborn run-to-failure set) under one bearing-wise protocol. The controls are label permutation, life-stage and metadata-only references, restricted negative classes, bearing-level aggregation, a stage-free fault-type setting and a test-invariance check; most analyses were declared in writing before the runs, and the others are labeled post hoc. Two main cues remain after bearing-wise splitting: the life stage in run-to-failure labeling, and the bearing model together with the recording offset in one naturally damaged dataset; weaker metadata hints (recording offset, speed or operating condition) appear on the development set, on the Paderborn run-to-failure set and, for one label, on XJTU-SY. The same protocol is applied to a small convolutional network on order-resampled envelope spectra, with and without a per-window row z-score and dropout. The variant with both changes has a positive point estimate against the base model in every original task and in 16 of the 17 settings examined (original tasks and controls), with intervals that exclude zero in 10 of them; an ablation attributes the effect to the row z-score, since dropout alone changes the result by less than 0.4 points. Against a re-implementation of a published convolutional baseline it is ahead on CWRU and in the stage-free setting, where the gain is concentrated in one label per dataset and rests on a few bearings with mixed defects, whereas a random forest on handcrafted features is competitive on several tasks. Protocols, predictions and code are available at https://github.com/hmz-bit/bearing-shortcut-controls.