LTC-AE: Liquid Time Constant Autoencoders for Time Series Anomaly Detection

This research introduces a novel class of autoencoders, termed Liquid Time-Constant Autoencoder (LTC-AEs), for anomaly detection in time series data. Anomaly detection in real-time data streams is an extremely crucial task for various domains like healthcare, finance, and cyber security but it comes with a few challenges including missing values, class imbalances and noise. Recent research has proved optimal in addressing these challenges using sequential Autoencoders. Our proposed algorithm having the ability to capture temporal dynamics, learn from irregular sequences, improve interpretability and facilitate scalability, propose to help resolve some problems that delimits the potential of existing Autoencoder models. The experiments conducted in the study are utilized to emphasize the potential of our proposed model in handling multifaceted anomalies. The workflow follows semi-supervised learning to leverage the strengths of both supervised and unsupervised approaches. The utility of LTC-AEs are evaluated on various datasets and performance comparison is done with respect to architectures like Long Short-Term Memory (LSTM-AE), Gated Recurrent Unit (GRU-AE) and Temporal Convolutional Network (TCN-AE) Autoencoders.