Real-time synchronization in neural networks for multivariate time Series anomaly detection

Deep learning has gained momentum over traditional methods in recent years due to its ability to scale up to an unforeseen rise in both volumes and dimensions of data emerging from IoT. It is suitable for modeling arbitrary complex dependencies, such as those exacerbated by asynchrony in the inputs. We target real time anomaly detection in asynchronous multivariate time series of regular seasonal variations, which lack sufficient research contribution, albeit their prominence in industrial applications. We propose a mathematical formulation of neural network layers, which generate a synchronized representation from asynchronous multivariate input. The layers can be added to any network architecture and are pre-trained to learn the multivariate input periodic properties, then use synchronizing - desynchronizing filters within networks to improve learning performance and detection accuracy. For demonstration, we apply the proposed method to an Autoencoder and evaluate on labeled anomaly detection data generated at eBay during business availability monitoring.