Data Driven Estimation of Series-RLC Parameters

Estimating circuit parameters requires balancing accuracy, measurement requirements, and processing cost. This study compares multiple multilayer perceptrons (MLPs) with bounded nonlinear least-squares fitting. The task involves parameter estimation for an ideal series RLC circuit driven by a known voltage step. A synthetic dataset containing 10,000 circuits was constructed and divided into training, validation, and testing sets. Networks trained on clean data and noise-augmented data were evaluated under Gaussian noise conditions. Comparison protocols for observation duration and sampling density were also established. At 10% noise, relative to the clean-training channel RMS, the full-record noise-augmented network achieves average median errors of 2.745%, 1.971%, and 2.012% for resistance, inductance, and capacitance, respectively. Historical full-grid CPU timings are approximately 0.048 ms for neural network inference versus 6.926 ms for nonlinear fitting, measured in separate earlier experimental runs. These timings exclude the initial 50 ms required to acquire the signal. At 101 samples per channel, a fixed five-start diagnostic removed 52 clean cases in which at least one component had an error greater than 10%. It also reduced the incidence of errors above 50% under 10% noise from 6.73% to 3.40%, which remains higher than the augmented network's 1.81%. These results are implementation-specific and do not establish the superiority of neural networks. Hardware effects and model mismatch remain outside the scope of this study.