In modern manufacturing, operating and environmental factors commonly induce abnormal failures and health degradation, leading to safety incidents and losses. Rolling bearings, pervasive in industrial systems, produce vibration signals that reflect equipment condition, making vibration-based anomaly detection central to fault diagnosis. Traditional signal-processing methods depend on expert priors and scale poorly, whereas deep-learning approaches often lack interpretability. To address these limitations, we propose an unsupervised method that combines a dual-channel autoencoder with a node-weighted isolation forest: one channel extracts time- and frequency-domain statistical features, and the other ingests complex time–frequency representations to capture frequency and phase dynamics; anomalies are then scored using node-level weighting within the isolation forest. Experiments on three public datasets confirm the effectiveness of the proposed approach.