Lightweight Neural Fault Diagnosis of Analog Electronic Circuits: The Value of a Second Measurement Node and the Limits of Transfer under Severity and Operating-Condition Shift

A classifier can identify large component deviations in a simulated analog circuit and still fail when the deviation becomes smaller or the measurement conditions change. This study evaluates a compact neural diagnostic method with that distinction made explicit, and reports the negative results as carefully as the positive ones. Complex transfer estimates at eight excitation frequencies are normalized by the nominal circuit response and supplied to a multilayer perceptron with 32 and 16 hidden units. Two observation nodes are used, and an output-only variant isolates the value of the additional measurement. The numerical benchmark covers a Sallen–Key low-pass filter and a cascade of two active first-order filters. Circuit realizations are generated independently for training, validation, calibration and six test domains, and the test domains enter none of the model-selection decisions. Across five training seeds, the two-node network reaches 100.00 % matched-domain accuracy on both circuits and 98.25 % and 99.96 % accuracy at a waveform signal-to-noise ratio of 20 dB. Accuracy falls to 79.85 % and 83.56 % for unseen 12–24 % component deviations, and to 56.77 % and 76.41 % under combined stress. A support vector machine performs better under combined stress, by 4.41 and 4.33 percentage points, whereas the neural model improves on the output-only network by 7.84 and 11.49 points. The network has 1,737 trainable parameters. A local sensitivity analysis explains both the benefit and the limit of the second observation node, and an exact parameter-scaling invariance of the Sallen–Key topology is shown to bound what any classifier can recover from these measurements. Confidence analysis shows that nominal-domain temperature scaling does not ensure reliability after a severity shift: expected calibration error increases after calibration in every shifted domain, and a fixed rejection threshold accepts fewer than 10 % of mild-fault cases. The results support a small diagnostic model inside a specified operating envelope, identify the measurements and experiments required before deployment, and should not be read as a claim of robustness outside that envelope.