LOBSTR - Learning On-Board Signals for Timely Reaction

Important results from the project

The overall question that the project LOBSTR is trying to answer is whether it is possible to apply anomaly detection methods to temporal multivariate signals for fault detection on the vehicle´s control unit. The project´s objective was to contribute in: - adaptation of existing anomaly detection methods to work well in distributed systems (IoT) with learning - evaluate and compare performance between different anomaly detection methods for IoT systems

Expected long term effects

The project results were very positive. The big goals that were set were achieved. Two distributed models were developed, where both can detect anomalies of runs with different injected faults as expected. The distributed models gave similar results for a vehicle and when distributing the data to different vehicles. The models had some false positives, so there are a lot of room for improvement. The project was so successful and there is much more to investigate that there are already plans to continue research and development within this area.

Approach and implementation

Several methods were investigated where we proceeded with two of them - the compression-based anomaly detection with LSTM autoencoder and the mixture-based anomaly detection. The methods were evaluated on data collected from driving on vehicles with purposely injected faults. The mixture-based method was implemented on existing vehicle hardware with communication via the cloud. It listens to CAN traffic to detect anomalies real time.