Distributed online Data Anomaly Detection for connected vehicles

Wireless connectivity evolution increased the volume of acquired available data in different Internet of Things based industries. Data quality and processing time are the most challenging issues for successful data analytic algorithms to produce efficient business intelligence. In this article we tackle these two points and propose a distributed framework for data anomaly detection. In fact one of the major issues in systems that depend highly on data is detection of anomalies. Long Short Term Memory (LSTM) based anomaly detection in time series data has been studied in the past with promising results. In this article, we use LSTM model and apply distributed learning approach to train the model. Indeed, using a single machine centralized approach for model training and anomaly detection is not a feasible option when dealing with big amounts of data. Distributed approach improves the training and prediction time, model performance, and allows handle of bigger datasets and higher level of model complexity. We propose a distributed anomaly detection system framework for autonomous and connected cars with a novel online new data selection algorithm that guides the retraining and adjusts the model parameters accordingly. The framework includes the offline training of the LSTM model over many machines in a distributed fashion using all the available data. The trained parameters are then sent to the individual vehicles and the anomaly detection happens at the vehicle level. Finally, the proposed distributed framework is evaluated using MXnet framework, and it shows that with optimized settings we can reduce the model training time, use a more complex LSTM anomaly detection model and improve anomaly detection accuracy.