Automatic process modelling (APM) is an enabling technology for the development of intelligent manufacturing systems (IMSs). The analysis of obtained models enables the prompt detection of error-prone steps and the design of proper mitigation strategies, in all aspects of the manufacturing process, from parameter optimization to development of customized personnel training. In this work we propose a Time Delay Neural Network (TDNN) applied to low level data for the automatic recognition of different process phases in industrial collaborative tasks. We selected TDNN because they are suited for modelling time dependent processes over long sequences while maintaining computational efficiency. To experimentally evaluate the recognition performance and the generalization capability of the proposed method, we acquired two novel datasets reproducing a typical IMS setting. Datasets (including manually annotated ground-truth labels) are publicly available to enable other methods to be tested on them and they replicate typical Industry 4.0 setting. The first dataset replicates a collaborative robotic environment where a human operator interacts with a robotic manipulator in the execution of a pick and place task. The second set represents a human tele-operated robotic assisted manipulation for assembly applications. The obtained results are superior to other methods available in literature and demonstrate an improved computational performance.
Automatic process modeling with time delay neural network based on low-level data
Menegozzo, Giovanni
;Dall’Alba, Diego;Roberti, Andrea;Fiorini, Paolo
2019-01-01
Abstract
Automatic process modelling (APM) is an enabling technology for the development of intelligent manufacturing systems (IMSs). The analysis of obtained models enables the prompt detection of error-prone steps and the design of proper mitigation strategies, in all aspects of the manufacturing process, from parameter optimization to development of customized personnel training. In this work we propose a Time Delay Neural Network (TDNN) applied to low level data for the automatic recognition of different process phases in industrial collaborative tasks. We selected TDNN because they are suited for modelling time dependent processes over long sequences while maintaining computational efficiency. To experimentally evaluate the recognition performance and the generalization capability of the proposed method, we acquired two novel datasets reproducing a typical IMS setting. Datasets (including manually annotated ground-truth labels) are publicly available to enable other methods to be tested on them and they replicate typical Industry 4.0 setting. The first dataset replicates a collaborative robotic environment where a human operator interacts with a robotic manipulator in the execution of a pick and place task. The second set represents a human tele-operated robotic assisted manipulation for assembly applications. The obtained results are superior to other methods available in literature and demonstrate an improved computational performance.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.