Neural Network-based Online Fault Diagnosis in Wireless-NoC Systems

• Published in: Journal of electronic testing Vol. 37; no. 4; pp. 545 - 559
• Main Authors: Wang, Qi, Ouyang, Yiming, Lu, Yingchun, Liang, Huaguo, Zhu, Dakai
• Format: Journal Article
• Language: English
• Published: New York Springer US 01-08-2021; Springer Nature B.V
• Subjects: Artificial neural networks; CAE) and Design; Circuits and Systems; Computer-Aided Engineering (CAD; Electrical Engineering; Engineering; Fault detection; Fault diagnosis; Finite element method; Network latency; Neural networks; System on chip; Topology; Traffic information; Wireless networks; Fault diagnosis; Wireless network on Chip; Artificial neural network
• ISSN: 0923-8174; 1573-0727
• DOI: 10.1007/s10836-021-05966-w

The recent development of wireless Network-on-Chip (WiNoC) by introducing wireless interface in traditional wired NoC has significantly increased the performance of NoC systems with higher bandwidth and low latency on-chip communication. However, the integration of more components (e.g., antenna and transceiver) on the chip also increases system complexity and makes it more susceptible to various failures. In this paper, we propose a run-time fault diagnosis mechanism based on neural network (NN) techniques, where both fully-connected (FC) and convolutional neural networks (CNN) are considered. For NoC with 2-D mesh topology that incorporates both wired and wireless interfaces, the FC and CNN neural networks for fault diagnosis and detection are presented. The NN models can be trained offline with collected traffic data from partially failed NoC with various faulty components. Then, at run-time, the NN models can be deployed on certain tiles in the NoC to detect and locate the faulty components using the run-time traffic data. Based on simulated traffic data, we have evaluated the proposed NN-based mechanism under different fault scenarios (e.g., type, location and number of faulty components). The results show that, CNN models outperform FC neural networks with higher fault diagnosis rates. CNN can successfully identify up to 81.2% faults when there is only one faulty component on the NoC with different traffic patterns. The accuracy decreases when there are more faulty components and higher traffic loads.