Graph-Based Multi-Task Transfer Learning for Fault Detection and Diagnosis of Few-Shot Analog Circuits

Abstract

Building an interpretable fault detection and diagnostic model based on few-shot circuit samples and prior information about circuit structures is of significant importance. To fill these gaps, we propose a graph-based multi-task transfer learning method for fault detection and diagnosis of circuits under few-shot conditions. Firstly, in order to model the interconnections of nodes in a circuit, the sample data is organized into a graph-structure, and a semi-supervised graph-based structural feature fusion method is proposed. The proposed method can accept graph-structured data and process the data using feature fusion methods. Secondly, to improve the model performance under few-shot conditions, two transfer learning mechanisms are proposed for the topological structure characteristics of analog circuits as well as circuit signal characteristics. Finally, through parameter-shared strategy, we propose a task transfer-based fault diagnosis approach. Experimental results on three different circuits show that the proposed method has the best diagnostic accuracy compared to typical detection and diagnosis schemes.