End point detection of weld seam removal in robotic abrasive belt grinding process based on computer vision and deep learning

• Main Author: Murugan, Pushparaja
• Other Authors: Tegoeh Tjahjowidodo
• Format: Theses and Dissertations
• Language: English
• Published: 2018
• Subjects: DRNTU::Engineering::Mechanical engineering
• Online Access: http://hdl.handle.net/10356/76103
• Institution: Nanyang Technological University

Over the past decades, the demand of high quality and most consistent products with lowers cost is highly increased. In oder to tackle the demands, large number of manufacturing industries stated to adopt integrated manufacturing new products in parallel with the incorporation of automated robots. One of the attractive solution for the addressed limitation is, adaptation of industrial automations which highly facilitates the increments in productivity, quality and reduction of manufacturing cost. Automation is a process that elevate the mechanization of a particular machinery operations, especially in robotic abrasive belt grinding process. Also, automation eliminate the human involvements by using the logical programming commands. However, the determination of end point detection in abrasive belt grinding process has been a non-trivial process in the implementation of robots. Determination of end point is the lowest level of robotic abrasive belt grinding process hierarchy that includes the manual grinding process, data acquisition devices, sensors, predictive models, arti cial intelligence systems. Also it includes the controlling the robots arm with the proper logic commands and the process parameters. Due to lack of technologies, the manual grinding process is currently followed in many industries though the removal of weld seam is carried out by the robots. The manual grinding process is the involvement of human laborious practical experience to determine the end point of weld seam removal. This is highly unreliable, inconsistent and time consuming since it depends on the human thinking and experience. The process requires highly skilled labors to get the desired nishing surface. Due to the high non-linear process of the grinding process, it is highly complicated to develop an predictive model based on the chemical and mechanical parameters. Parameter independent model such as machine learning frameworks highly depends on the manually selected features. Moreover, machine learning frameworks are highly suffer from bias-variance trade off. Lower amount of features leads the algorithm to predict the undesirable results while the higher level of features leads the algorithm to suffer from curse of dimensionality. The recent developments in deep learning frameworks shows the potential of performing complex task which can be attractive solutions to the limitations mentioned. Hence, it is necessary to address the limitation of these mentioned classical method and develop a new method to determine the end point of abrasive belt grinding process. This thesis is intended to develop a parameter independent framework based on Convolutional Neural Network to determine the end point of abrasive belt grinding process. We propose a new methodology for determine the end point of weld seam removal by locating and classifying the various stages of weld seam. Weld seam is fabricated on mild steel square plate and four stages of weld seam is considered for the study. Images are acquired by using digital camera and are used for training the developed framework. Various optimization and regularization techniques are implemented for making the framework for better generalization. Multi-component loss function with ADAM optimizers are used for optimizing the learning parameters of the network. Though the hyper-parameters are selected by trial and error methods, proposed network has the high precision and recall values in detection and classi cation of the weld seam stages. Since the architecture depends on the global statistical information of weld seam images, the same network can be implemented in classi cation of any materials.