Vibration-based diagnosis of propulsion-shaft misalignment supports condition monitoring in electric ships, but reported accuracies are often inflated by overlapping-window leakage and by untested cross-recording, cross-session, and cross-speed generalization. We evaluated leakage-free, deployment-oriented protocols (chronological, leave-one-recording-out (LORO), leave-one-session-out, and leave-one-speed-out) on a 50 kW electric-propulsion land-based test system with six accelerometer channels under 0, 2, and 4 mm offset misalignments, comparing feature-engineered gradient-boosting models, raw-signal deep models, and nominal-speed order-normalized features. Under LORO evaluation, a raw-signal one-dimensional convolutional neural network (1D-CNN) achieved 0.93–1.00 accuracy and outperformed LightGBM in all 15 folds. Under leave-one-speed-out testing, MiniRocket retained 0.775 accuracy, whereas InceptionTime dropped to 0.468. Fixed-Hz features outperformed nominal-speed order-normalized features by 0.079 (p = 0.022), and single-window 1D-CNN inference required 0.50 ms on a CPU. Recording-level and cross-session evaluation are therefore essential for reliable misalignment diagnosis, and raw-signal models provide the strongest deployable performance on the present test rig.