Automated condition monitoring of railway switches and crossings (S&C) requires classification models whose reported accuracy reflects genuine generalization rather than evaluation artefacts. This paper presents a methodologically rigorous, leak-free machine-learning framework for vibration-based event classification, evaluated on accelerometer data from a full-scale outdoor S&C test facility. The pipeline enforces strict ordering (split, select, augment, standardize, train, evaluate) and partitions the data at the level of physical events, so that all measurements of a given event are assigned together to either the training or the test subset. A symmetric tabular autoencoder generates synthetic minorityclass samples through latent-space interpolation. Twenty-one classifiers spanning eight families are benchmarked on held-out data and by group-aware five-fold cross-validation. The strongest models reach 81.5% held-out accuracy (ROC-AUC ≈0.94) and 80.4%±2.1% under cross-validation; ensemble methods are the most stable. Feature standardization is essential: without it, neural networks collapse below chance level. Computational profiling (inference latency 0.005–0.63 ms per one-second segment; model size 0.002–2.4 MB) maps three deployment scenarios to specific algorithm recommendations. Because the minority crossing class has only six held-out samples, its per-class metrics carry wide confidence intervals and should be interpreted with caution.