Containers in A Network
(duration: 01/2006 - 07/2007 /// funding: Subsidie Electronische Communicatie of Senter Novum)
In the Containers in a Network (CAN) project, Almende, Chess and Salland Electronics researched how sensors can be applied in bulk container management in a Klok Containers garbage processing plant. Self-learning ad-hoc communication networks predict when containers have to be emptied.
CAN proposed several new techniques to aid an operator at Klok Containers in container management. There are several garbage containers at Klok Containers that are each filled at different rates with different materials. The current - not so elaborate - way to know when a container has to be emptied, is by kicking to its exterior and inferring its fill grade by the sound.
The project's solution exists out of sensor nodes that measure the amount of dirt in the containers, a wireless sensor network to store that information temporarily in a redundant distributed way and a PC as network drain and graphical user interface to the Klok operator. The Klok operator operator receives a signal when a container is full, and he can subsequently inform the drivers of the shovels which container to empty.
Salland Electronics developed an ultrasonic sensor to measure the amount of dirt in the garbage containers at Klok Containers. In cooperation with other parties like the VU in Amsterdam, Chess developed a wireless network mote with an energy-saving information dissemination (gossip) protocol. Those motes are called MyriaNed nodes.
Almende developed higher level software on the nodes and on the operator's computer. This level of software adapts to and learns from the environment and is implemented in an agent-oriented way.
There are two locations in which agents play a role in the CAN software. One location is on the central PC. By representing each container as an agent it is easier to let them decide matters on a mutual level of understanding, before alerting the operator. This representation is implemented in the Emerge framework (CHAP).
The other location is on the sensor nodes where the agents can predict the dirt levels and act accordingly for purpose of energy housekeeping for example. When the container is almost empty it will take a long time before an operator has to be signaled, and the signal frequency can be decreased.
The system needs to be able to learn and adapt. In the first place, it has to relieve the system installer in fine tuning parameters at installation. Those are parameters like calibration levels when a container is full, and when it is empty. The ultrasonic sensor generates a value like 350 when the container is empty (it measures a distance of 350 cm) and say 100 when full. But this is not defined beforehand. The node learns this by inspecting minimum and maximum levels and dismissing extremes.
Its second purpose is to implement the forecasting functionality needed on an application level. Therefore it has to learn how large the delay is between a generated EmptyMe! event and the actual empty event. And certain parameters have to be predefined, for example the preferred margin between an 'almost' full and a 'really' full container.