Gitlab:
Anamoly Detection of Road Sensors
Back in 2017, road sensors were used to count the volume of traffic and measure the traveling speed. There were two types of road sensors, VD(Vehicle Detector) and eTag. VD is a road sensor hidden under the road pavement. Whenever there is a vehicle passing by, the pressure caused by the vehicle measures the volume of the traffic flow. Scooter is counted as 0.5 of vehicles and huge trucks as 2 vehicles, for example. VD measures the point of speed while a vehicle passes through it. The measure of traveling speed by VD is not more accurate than eTag because eTag measures the speed of space by calculating how much time a vehicle takes from point A to point B.
To make sure we have an accurate measure of traffic volume and traveling speed, we need to ensure the quality of road sensors. However, it is not effective and reasonable to check hundreds of thousands of road sensors manually. Is there any way for us to automatically detect sensor malfunction? This is then where the project value I contribute to the advancement of the Smart City in Taiwan.
The idea of running the anomaly detection is to apply a machine learning algorithm to detect the abnormal patterns of sensor data. The methodology behind this is the clustering algorithm designed based on the domain knowledge defining the abnormal patterns.
- Double Peak Hours
- Morning Peak Hours
- Evening Peak Hours
- Abnormal Data Patterns
First, I create several sets of data patterns labeled based on domain knowledge. For example, inbound city roads always have morning peak hours while evening peak hours happen in the outbound city roads. Next, I run the clustering process by mixing the labeled samples with the unlabelled samples. The number of clusters ends when the labeled data are separated to a certain degree. This means samples of morning peak hours are almost in the same cluster. Those captured by different abnormal data patterns are then reported to the traffic control center for further investigation of their functionality.