Reliability-aware waveform measurement framework for intermittent fault localization in analog circuits

Abstract Fault localization in analog circuits denotes identifying the faulty component/location through waveform-based measurement responses. It is difficult because useful measurement evidence is sparse, randomly triggered, nonstationary, and strongly dependent on waveform segments. To address this measurement challenge, this paper proposes a reliability-aware waveform measurement framework based on fault-sensitive segment weighting and multi-source waveform evidence fusion. Rather than representing each waveform with a single global feature vector, the framework organizes it into a sequence of segment tokens, each described by time-domain statistics, normalized spectral entropy, and wavelet-packet energy features. The framework learns adaptive source importance weights within each segment, extracts local and contextual waveform evidence through multiscale segment encoding, and embeds a fault-sensitive segment gate in the forward path to emphasize informative waveform intervals before generating the final measurement decision. Held-out validation and calibration subsets are used for temperature scaling and conformal prediction, respectively, enabling the model to output the fault class, calibrated confidence, prediction set, and uncertainty score as a reliability-aware measurement decision. In the 10-seed comparison with nine baseline methods, the proposed method achieves 95.96% accuracy and 95.37% macro-F1 in Case 1, and 99.90% for both metrics in Case 2. Across the ten evaluation seeds, temperature scaling changes test ECE from 10.08% ± 1.18% to 2.10% ± 0.65% in Case 1 and from 10.02% ± 0.95% to 0.10% ± 0.09% in Case 2. The proposed framework reports a fault-location prediction together with waveform-level evidence and measurement-reliability information, which is beneficial for practical analog-circuit condition monitoring.