Industrial maintenance has shifted from reactive repairs and calendar-based servicing toward data-driven predictive strategies. This paper presents a non-intrusive, low-cost IoT hardware platform for sustainable predictive maintenance of rotating machinery. The system integrates an ESP32-S3 sensor node that captures vibration (100 kHz) and temperature data, performs local logging, and communicates wirelessly. An automated spectral band segmentation framework is introduced, comparing equal-energy, linear-width, nonlinear, clustering, and peak–valley partitioning methods, followed by a weighted feature scheme that emphasizes high-value bands. Three unsupervised one-class classifiers—transformer autoencoders, GANomaly, and Isolation Forest—are evaluated on these weighted spectral features. Experiments conducted on a custom pump test bench with controlled anomaly severities demonstrate strong anomaly classification performance across multiple configurations, supported by detailed threshold-characterization metrics. Among 150 model–segmentation configurations, 25 achieved perfect classification (100% precision, recall, and F1 score) with ROC-AUC = 1.0, 43 configurations achieved ≥90% accuracy, and the lowest-performing setup maintained 81.8% accuracy. The proposed end-to-end solution reduces the downtime, lowers maintenance costs, and extends the asset life, offering a scalable, predictive maintenance approach for diverse industrial settings.