A state-classification approach for light-weight robotic drilling using model-based data augmentation and multi-level deep learning. (15th March 2022)

• Record Type: Journal Article
• Authors: Lu, Hao; Zhao, Xingwei; Tao, Bo; Ding, Han
• Is Part Of: Mechanical systems and signal processing. Volume 167:Part A(2022)
• Publication Date: 2022-03-15
• Subjects: Machining signal processing -- Robotic drilling -- Data augmentation -- Vibration model -- Auto encoder -- Deep residual network; Structural dynamics -- Periodicals; Vibration -- Periodicals; Constructions -- Dynamique -- Périodiques; Vibration -- Périodiques; Structural dynamics; Vibration; Periodicals; 621
• DOI: 10.1016/j.ymssp.2021.108480
• Languages: English
• ISSNs: 0888-3270

Abstract

Highlights: A method to classify robot's vertical drilling and inclined drilling is proposed. A vibration model is established to simulate vibration signals for data augmentation. Stable drilling signals are intercepted from original signals based on auto encoder. A Resnet classifier is trained by combining the simulated signals and actual signals. Abstract: In the aircraft assembly industry, a large number of holes need to be drilled so that a convenient and fast drilling quality detection method is necessary. During the robotic drilling process, the verticality of the drill directly affects the quality of the hole. In this paper, a vibration-based classification approach is proposed to find out the unqualified holes caused by inclined drilling. Firstly, robotic drilling's vibration model is established to simulate vibration signals during vertical and inclined drilling, which to prove the model information of the drill system. Then, based on auto encoder, the actual drilling signals are classified into drilling section and non-drilling section, and the stable drilling section signals are intercepted from the drilling section to be further used to improve classification performance. Finally, combining the simulated signals and a small number of actual signals as the training set, the data augmentation is realized and a deep residual network (Resnet) classifier is trained. Inputting the stable drilling section signals into the classifier, the drilling state is classified into vertical drilling and inclined drilling. In order to verify the proposed method, sixteen groups of experiments under different process parameters are carried out in this paper, and the results have proved the effectiveness and potential of this method.