Effective anomaly detection in industrial applications is challenged by the complex temporal dependencies and large data volumes from multivariate time series. Traditional deep hybrid models, utilizing neural networks for feature extraction followed by standard anomaly detection algorithms like SVM, often fall short in capturing critical features over varying time scales. Addressing these limitations, we introduce VAML-Net, a novel hybrid unsupervised anomaly detection model that integrates a Variational Autoencoder (VAE) with a Multiresolution LSTM. This model enhances feature extraction through the VAE over short sequences, while the Multiresolution LSTM captures extended temporal dependencies via a hierarchical fusion mechanism. Furthermore, VAML-Net incorporates a dynamic thresholding method for precise anomaly identification. Extensive evaluations on five benchmark datasets demonstrate VAML-Net’s superior performance, significantly outperforming existing models. This approach not only advances anomaly detection in time-sensitive industrial settings but also offers a scalable solution adaptable to diverse data-intensive environments.