A deep transferable motion-adaptive fault detection method for industrial robots using a residual–convolutional neural network. (September 2022)

• Record Type: Journal Article
• Authors: Oh, Yeongtak; Kim, Yunhan; Na, Kyumin; Youn, Byeng D.
• Is Part Of: ISA transactions. Volume 128(2022)Part B
• Page Start: 521
• Page End: 534
• Publication Date: 2022-09
• Subjects: Industrial robot gearboxes -- Deep-learning-based Fault detection -- Adversarial contrastive learning -- Supervised domain adaptation -- Torque ripples; Engineering instruments -- Periodicals; Engineering instruments; Periodicals; Electronic journals; 629.805
• DOI: 10.1016/j.isatra.2021.11.019
• Languages: English
• ISSNs: 0019-0578

Abstract

Abstract: Recently, in various industrial fields, including automated manufacturing processes, industrial robots are becoming indispensable equipment; these robots perform repetitive tasks and increase the productivity of the production line with consistent precision and accuracy. Thus, fault diagnostics of industrial robots is an essential strategy to prevent the significant economic losses that can be caused by a sudden stop of a production line due to an industrial robot fault However, previous data-driven industrial robot fault diagnostics are limited because a pre-trained model built for a specific motion may not accurately or consistently detect faults in other motions, due to motion discrepancies. To overcome this difficulty, in this paper, we propose a deep transferable motion-adaptive fault detection method that uses torque ripples for fault detection of industrial robot gearboxes. The proposed method is composed of two stages: (1) a residual–convolutional neural network is used to enhance the performance of feature extraction for simple motions, after first refining raw torque signals by filtering out the motion-dependent signals (2) a binary-supervised domain adaptation is performed to detect faults adaptively on multi-axial motions through adversarial contrastive learning. The efficacy of the proposed method was validated using experimental data from unit-axis and multi-axial welding motions collected from a real industrial robot testbed. The proposed method showed superior fault detection accuracy for the motion adaptation task, as compared to existing methods. Highlights: A novel, deep, transferable, motion-adaptive fault detection method for industrial robot gearboxes is proposed. Motion-invariant fault-related features of torque ripples are learned. Experimental validations are provided, examining two real industrial welding motions. The proposed method showed superior fault detection performance, as compared to earlier methods.