Smart fault detection in a nanosatellite electrical power system

Abstract This simulation-based proof of concept evaluates a unified residual framework for faults affecting the photovoltaic (PV) array, maximum-power-point-tracking (MPPT) converter, regulator converter, and battery-to-ground interface of a nanosatellite electrical power system (EPS). The MATLAB R2021a implementation preserves the original EPS model, fault scenarios, diagnostic architecture, and figure configuration while adding orbit-disjoint evaluation, reproducibility controls, classifier comparisons, noise tests, and resource analysis. Healthy and fault-specific multilayer perceptron (MLP) identifiers produce outputs; residual vectors are classified by an MLP, k-nearest neighbors (KNN), a decision tree (DT), and principal component analysis (PCA) with a nearest-centroid rule. For each class, 145 orbit simulations, each containing 69 samples, were separated into 87 training, 29 validation, and 29 sealed test orbits. All normalization statistics and model parameters were estimated exclusively from training data; validation supported model selection and early stopping. On the ideal test set, the PV, residual-only EPS, and augmented EPS MLPs each achieved 100% accuracy; the ten-seed mean was 100.000% ± 0.000%. Using state-of-charge/current inputs, KNN, DT, and PCA achieved 97.311%, 97.201%, and 79.620%, respectively. On an identical six-feature input, all four classifiers reached 100%, indicating that ideal feature separability, rather than classifier choice alone, drove performance. With additive white Gaussian noise applied to raw signals, EPS accuracy decreased from 99.920% at 40 dB to 40.410% at 10 dB; PV accuracy was 100% and 74.679%, respectively. The MLP bank requires 549 parameters (2.15 KiB, float32) and 430 multiply-accumulate operations per inference. These results demonstrate separability under controlled simulation conditions, not hardware-in-the-loop or flight-level accuracy.