<p>Reliable electronic devices play a pivotal role in empowering various industries and people’s lives. However, the circuit failure analysis still heavily relies on traditional manual inspection, which is time-consuming, labor-intensive and prone to human errors. The main challenges are the scarcity of defect samples, expensive manual labeling, random types, shapes, locations, and multiscales of defects. Limited by these factors, prior works in this domain leveraging supervised learning are too expensive and infeasible for real deployment and suffer from significant performance degradation on multi-scale circuit defects. To address these challenges, we propose a Failure-Analysis-Inspired Multi-Scale Defect Detection Framework, <i>FA-MSDDet</i>, a weakly supervised framework that emulates the practical failure-analysis process by comparing the structural features of defective and normal circuits. By constructing a normal reference repository from a scale-invariant feature learning with proposed central distance positional encoding for circuit layout and inferring abnormalities through distance-based comparison, FA-MSDDet achieves precise and scalable defect detection and localization without pixel-level labels. Extensive experiments on two challenging circuit-defect datasets demonstrate that FA-MSDDet achieves state-of-the-art performance in both one-shot and multi-scale scenarios. Furthermore, to verify its cross-domain applicability, FA-MSDDet is evaluated on the <b>mobile phone surface defect dataset</b>, achieving competitive state-of-the-art results and confirming its generalization to real-world industrial inspection tasks.</p>