Real-Time Detection of Safety-Relevant Forklift Operating States using Acceleration Data with a Windowing Approach

In recent decades, many studies have been published on the topic of activity recognition. The research field is continuously evolving. This paper presents a framework to classify incoming acceleration data from a forklift fork to determine the current operating state of the forklift. The aim is to develop a system that reliably detects safety-relevant operating states, that can be reported to the driver as a warning signal or to a fleet management system as the basis for predictive maintenance of the fork. Thus, the fork wear can be reduced and the safety increased. For machine activity recognition, we train machine-learning models on a created data set. The algorithm focuses on fast and low-computation processing of incoming sensor data. Thus, a time-based windowing approach that aims at early classification of patterns is used. Therefore, we discuss an approach to future extraction and determine the optimal window size when classifying. With the chosen approaches, a robust realtime detection of the operating states with an accuracy up to 99% can be achieved. We provide the algorithms and data set on GitHub at https://github.com/tum-fml/dofos.