A Convolutional Autoencoder Framework for Probabilistic Anomaly Detection on Infrastructure Systems

Data-driven autonomous anomaly detection, empowered by advanced sensing and machine learning techniques, has the potential to significantly reduce costs associated with infrastructure maintenance. In this paper, we studied the impact of probabilistic anomaly detection in the context of a proposed deep learning-empowered framework for multivariate time series anomaly detection. In this framework, two-dimensional convolutional autoencoders are utilized to learn the patterns of normal data and detect unseen anomalous data points that fail to be reconstructed by the well-trained autoencoders. Kernel density estimation was utilized to generate a probability distribution function based on raw window-based anomaly predictions. Our assessments showed promising performance in a real-world case study of Class I railroad track anomaly detection. We conducted a comparative study that proved our hypothesis that probabilistic anomaly detection outperforms the deterministic approach of anomaly detection using direct autoencoder output. Furthermore, the impact of multivariate time series data processing methods, autoencoder architecture, and anomaly score thresholds was also investigated.