AG Kommunikationstheorie


Statistical Learning for Classification of Noisy Pressure Sensor Measurements Applied to Mechanical Assembly


In the field of car manufacturing certain tasks are still hard to automate, where robots are either too slow or not flexible enough to replace human work force. To avoid faults and ensure that quality is still at a constantly high level, assistance systems are introduced to an increasing extent in the frame of Industry 4.0. Using different sensors, the activity of working in a production environment can be tracked and analyzed. In this work, time-series measurements are collected from pressure sensors attached to the factory worker's thumb at BMW's production line. Using these measurements, different machine learning techniques are applied for detecting when a mechanical piece is being assembled, predicting whether it was correctly assembled or not, and visualize our results. Algorithms such as, support vector machines, k-nearest neighbor, random forests, Gaussian mixture models and k-means clustering, are trained on laboratory generated data, then tested on real production line measurements. The best performing methods achieve less than 5% error on the detection task, less than 2% error on the classification task, and are visualized using multi-dimensional scaling with different dissimilarity metrics.

zurück zur Terminübersicht