A multi-scale graph pyramid attention network with knowledge distillation towards edge computing robotic fault diagnosis

Chen, Chong, Wang, Tao, Mao, Dong, Liu, Ying ORCID: https://orcid.org/0000-0001-9319-5940 and Cheng, Lianglun 2025. A multi-scale graph pyramid attention network with knowledge distillation towards edge computing robotic fault diagnosis. Expert Systems with Applications 260 , 125469. 10.1016/j.eswa.2024.125469 Item availability restricted.

PDF - Accepted Post-Print Version Restricted to Repository staff only until 28 September 2025 due to copyright restrictions. Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (872kB)

Abstract

The advanced graph neural networks (GNNs) algorithms and the abundant monitoring data have greatly improved the performance of fault diagnosis of industrial robots. However, collecting massive monitoring data from industrial robots and performing diagnosis in the cloud server is hard to achieve due to the heavy transmission load and computational costs. Edge computing provides a potential solution to this challenge, while it requires the neural network deployed in the edge device to be effective and lightweight. This paper proposes a multi-scale graph pyramid attention network (MsGPAT) enhanced with knowledge distillation (KD) to balance accuracy and efficiency for edge-based fault diagnosis. The approach first integrates multi-scale graph convolutional networks (GCN) to extract spatial features at different granularities from monitoring data. Subsequently, pyramidal attention is incorporated to hierarchically summarize inter-scale dependencies and intra-scale neighbors for multi-resolution temporal modeling. Finally, KD is adopted to transfer salient knowledge from the complex MsGPAT to a tailored compressed model suited for edge resources. An experimental study based on the faulty dataset collected from the real world is conducted. The experimental results show that the MsGPAT can achieve the diagnosis accuracy of 85.3 % after KD, which model size is only 1.04 MB.

Item Type:	Article
Date Type:	Publication
Status:	Published
Schools:	Engineering
Publisher:	Elsevier
ISSN:	0957-4174
Date of First Compliant Deposit:	3 October 2024
Date of Acceptance:	27 September 2024
Last Modified:	07 Nov 2024 18:08
URI:	https://orca.cardiff.ac.uk/id/eprint/172566

Actions (repository staff only)

Edit Item