Evaluating and Optimizing Conventional Training Circuits for Analog Fault Diagnosis via Transfer Learning

Abstract

The soft-fault diagnosis in analog circuits increasingly relies on neural networks; yet, diagnostic accuracy is fundamentally constrained by the quality of the training data supplied to those networks. Most studies still generate that data with a few long-standing “benchmark” training circuits (e.g., Sallen–Key bandpass filter and four-op-amp biquad high-pass filter) and tacitly assume that the circuit choice is inconsequential. Our experiments reveal the opposite: component symmetries in these classic topologies create fault-response overlap, producing ambiguous feature sets that hamper classification and generalization. To overcome this limitation, we structurally modify the original four-op-amp biquad filter to break the component symmetries responsible for fault confusion, thereby yielding more diverse and separable fault responses. Three training circuits are evaluated—Sallen–Key, original four-op-amp, and the optimized design—under consistent fault category count, excitation signal, and sample length. A fixed 1-D residual network (ResNet) is trained on each dataset; transfer learning then tests generalization on a more complex leapfrog circuit. Besides overall accuracy, we visualize embeddings with t-distributed stochastic neighbor embedding (t-SNE) and quantify separability through interclass centroid distance and intraclass variance. Across 100 independent trials, the optimized circuit achieves higher classification accuracy, the largest average interclass distance, and the lowest intraclass variance. Moreover, it consistently shows smaller or comparable standard deviations (SDs) across all metrics, indicating more stable and reliable diagnostic performance compared with the conventional circuits.