Anomaly detection

The Challenge

To identify on a conveyor in a parcel sorting centre abnormal cases such as two parcels bunched together or incorrectly aligned parcels using an economical solution.

Our Solution

The classical method for solving such an issue today would entail using one or more cameras and artificial vision algorithms. For the type of conveyor in question and the budget limitations this approach was not feasible.

We thus decided to use a laser measuring device able to provide us with the surface profile of the parcels as they travelled past the sensor. We then built a custom system which allowed us to acquire synchronized data from the laser sensor and time lapsed images of the parcels as the travelled down the conveyer, letting us to collect a large quantity of data in a matter of a few days.

Thanks to the images collected we classified some of the profile data and use these to train a neural network with the goal of classifying parcels as correct or abnormal.

Having verified that the system performed correctly in the virtual platform, we then proceeded to tackle the real challenge, that is execute the neural network model on a simple and cost-effective microcontroller.

A significant optimization effort allowed us to run the model on an 8-bit microcontroller sufficient fast to solve the original issue during standard operating conditions.

The smart sensor created in this way was capable of solving the customer’s issue at a fraction of the cost of a traditional artificial vision system.