Severity Path Invariance Learning for Bearing Fault Diagnosis Under Unseen Severities

Bearing fault diagnosis can fail to generalize when a fault size is absent from training, even when the fault classes remain known. We study this closed-set unseen-severity problem on six tasks from the CWRU and DZLRSB datasets. Severity Path Invariance Learning (SPIL) constrains same-class paths between observed severities and their severity-conditioned centers, while a compactness term limits residual within-class spread. The auxiliary score is used only as an internal interpolation-consistency variable; source-only diagnostics do not support interpreting it as a calibrated physical defect-size measurement. Models are selected using source validation only; held-out target labels are not used for stopping or model selection. In this target-free comparison, SPIL remains competitive on the difficult CWRU tasks, with its clearest task-level advantage on CWRU-SG-014, whereas WhiteningNet has the highest aggregate source-selected classification performance. A separate target-blind Full/Base confirmation under the same checkpoint rule does not show a stable Full-over-Base gain. The original target-inspected trajectory peak is retained only as an explicitly labeled oracle target-selected diagnostic. For condition-monitoring development, the protocol tests whether a known fault-type classifier remains usable when a future defect size is absent from the development data; it does not establish transfer to new bearings or machines.