Window-Overlap Leakage Inflates Few-Shot Fault Detection: Quantification, a Leakage-Free Protocol, and a Re-Evaluation of a Physics-Guided Graph Network

Few-shot fault detection from multivariate sensor signals is usually evaluated by cutting recordings into overlapping sliding windows, then splitting those windows into training and test sets. This ordering leaks the test set. Auditing our own earlier experiments, 31.5% of the permanent-magnet synchronous motor (PMSM) test windows and 24.9% of the Case Western Reserve University (CWRU) test windows were byte-identical to training windows; on CWRU they covered 99.7% of the healthy test class. We propose a protocol that partitions along the time axis before windowing, inserts guard bands of one window length at every cut, and verifies disjointness at the index level, and we reevaluate a physics-guided graph in-context network against four baselines at an equal labeled-fault budget. Conclusions reverse. On PMSM a Random Forest with the same labels attains F1 = 0.997 against the network’s 0.973, and a previously reported 0.999 does not survive; on CWRU the network reaches 1.000 but a conventional graph classifier reaches 0.9996. Sensitivity studies localize what matters. Permuting the physics prior’s edges, or substituting a random graph of matched density, changes F1 by at most 0.0001. Graph sparsity and support-set size are inert; only the spectral decomposition moves accuracy. Across ten seeds and eight held-out fault types the ranking is invariant (recall exactly 1.000) while precision at a fixed threshold spans 0.690 to 0.988, and the network is poorly calibrated (ECE 0.209, unrepaired by temperature scaling). In this literature, evaluation protocol accounts for more of the reported performance than architecture does.