Autonomous condition monitoring of AutoStore robot components using AI-enabled visual inspection

Abstract

AI has increased in popularity over the last decade due to its rapid advance. While it has become popular with the man on the street for its language capabilities, companies also use it in fields such as automa- tion. AutoStore, an automation company from Norway, seeks to circum- vent the need for human intervention while inspecting their R5 robots. Specifically, the belt mechanism that controls the speed and direction of the R5 robot. This thesis proposes a solution to AutoStore’s problem in- volving a convolutional neural network coupled with a recurrent neural network, or a CNN-RNN for short. The CNN-RNN was also compared to a VGG16-RNN, which is similar, but the CNN is replaced with a pre- trained VGG16. While the VGG16-RNN did not reach the desired results, the CNN-RNN showed great results when using the collected data and supports the consensus that deep learning is a great tool for video clas- sification. It is highly suggested that the amount of data is increased in the case of further testing, as data leakage from the data collection and pre-processing can have been the source behind the high accuracy.