Neural Network Spots Nanosatellite Power Faults in Simulation With 100 Percent Accuracy

A nanosatellite’s electrical power system is a small, unforgiving machine. Photovoltaic arrays generate the energy, converters condition it, a battery stores it, and every other subsystem on the spacecraft depends on the result. With restricted mass, volume, power, telemetry bandwidth, and redundancy, a fault that begins in a single power-conditioning element can propagate rapidly to mission-level loss of function. A new open-access study in Discover Electronics presents a simulation-based proof of concept for catching such faults early, using a software monitor that reads only the current and voltage telemetry a nanosatellite already produces, without adding a dedicated sensor suite.

The work, by Alireza Rezaee, Niloofar Nobahari, and Ayub Taheri, evaluates a unified residual framework covering the photovoltaic array, the maximum-power-point-tracking converter, the regulator converter, and the battery-to-ground interface. The diagnostic classes are healthy operation, PV open circuit, PV line-to-line fault, MPPT-switch open circuit, regulator-switch open circuit, regulator-switch short circuit, and battery-to-ground fault. Each injected fault is a severe, binary condition; incipient resistance changes, intermittent faults, and progressive degradation are explicitly outside the demonstrated scope. The architecture itself is deliberately conservative: residual-based diagnosis, multilayer perceptrons, PCA, k-nearest neighbors, and decision trees are all established methods, and the authors claim none as an algorithmic invention. The systems-level contribution is the use of one low-sensor pipeline to represent and distinguish faults distributed across four distinct parts of the power system.

The diagnostic logic works by comparison. A bank of neural models contains one healthy identifier plus one fault-specific identifier for each relevant state, all trained to reproduce the system’s response from solar irradiance and PV surface temperature alone. At each sample, the measured response is subtracted from each candidate model’s prediction, producing a residual. The model whose prediction best matches reality should yield the smallest structured mismatch, while the remaining residuals provide class-discriminating context. The whole pattern of residuals is then fed to a classifier that assigns a fault label. Notably, the framework intentionally avoids requiring input from the attitude determination and control subsystem, which keeps EPS monitoring available during an ADCS anomaly, though the authors acknowledge that attitude and solar-incidence angle are latent drivers of PV output that irradiance and temperature can substitute for only within the illumination envelope represented during model development.

A key methodological feature is the leakage-prevention protocol. Each of the seven classes contains 145 independently initialized simulated orbits with 69 samples per orbit, split into 87 training, 29 validation, and 29 sealed test orbits, equivalent to 60, 20, and 20 percent. The orbit, not the sample, is the indivisible grouping unit, and sliding windows never cross orbit boundaries. All normalization statistics and model parameters were estimated exclusively from training data; validation supported only model selection and early stopping; and the sealed test set was evaluated once, after the pipeline was finalized. Ten classifier initializations used seeds 20,260,730 through 20,260,739, with results reported as mean plus or minus sample standard deviation. The implementation runs in MATLAB R2021a and preserves the original EPS model, fault scenarios, diagnostic architecture, and figure configuration while adding the orbit-disjoint evaluation, reproducibility controls, classifier comparisons, noise tests, and resource analysis.

On the ideal sealed test set, the results were striking. The PV subsystem MLP classified all 6,003 sealed test samples correctly, achieving 100 percent accuracy, macro precision, macro recall, macro F1, and macro one-vs-rest AUC. The residual-only whole-EPS MLP classified all 10,005 sealed test samples correctly, likewise reaching 100 percent on every reported metric, and the augmented MLP that adds a causal sliding mean of load current was also perfectly diagonal, with 2,001 correct test samples in each class and a ten-seed mean accuracy of 100.000 percent plus or minus 0.000 percent. The healthy whole-EPS identifier achieved a correlation of 0.993031 and an RMSE of 0.009282 amperes on test orbits, while the healthy PV-current and PV-voltage identifiers reached correlations of 0.999957 and 0.999997, respectively.

Comparisons with other classifiers revealed something important about where the performance actually comes from. Using only state-of-charge and load-current inputs, KNN, decision tree, and PCA with a nearest-centroid rule achieved 97.311 percent, 97.201 percent, and 79.620 percent accuracy on sealed test orbits, respectively. But when all four algorithms received the identical training-standardized six-feature vector of residuals plus the causal current mean, on the same orbit split with no test-driven tuning, every one of them reached 100 percent. The authors interpret this tie as showing that ideal feature separability, rather than classifier choice alone, drove the performance, and they explicitly state it does not establish MLP superiority. They also caution that because the residual-only MLP already achieved the same result, no incremental accuracy gain can be attributed to the causal mean feature under these ideal conditions; its benefit must be reassessed under matched and overlapping load profiles.

The noise tests puncture the optimism of the perfect scores. When additive white Gaussian noise was applied to the raw measured signals and features were recomputed without retraining, the whole-EPS classifier fell from 99.920 percent accuracy at 40 dB signal-to-noise ratio to 98.251 percent at 30 dB, 75.342 percent at 20 dB, and 40.410 percent at 10 dB. The PV classifier held at 100 percent through 30 dB, then declined to 96.585 percent at 20 dB and 74.679 percent at 10 dB. The authors note that sensor bias and drift, load transients, component tolerance, fault severity, aging, and unmodeled attitude may create systematic shifts more damaging than zero-mean noise.

Onboard feasibility was assessed analytically rather than by measurement. The complete MLP bank, if both the whole-EPS and PV paths are resident, requires 549 parameters, 2,196 bytes of float32 weight memory, or 2.15 KiB, and 430 multiply-accumulate operations per complete inference, with temporary activations adding less than 0.25 KiB under sequential evaluation. Latency figures are ideal arithmetic floors assuming one MAC per clock cycle at 100 MHz, not measured execution times; activation-function libraries, memory access, the state-of-charge estimator based on an unscented Kalman filter, telemetry acquisition, scheduling, and communication can all dominate in practice. The common-feature KNN is described as much less attractive onboard because it stores all 30,015 training points. The authors also emphasize that classification confidence should pass through an unknown/reject option and temporal persistence, such as three consecutive identical high-confidence labels, before any mission-specific fault-detection, isolation, and recovery action is taken.

The limitations are stated plainly. The study is bounded by simulation-derived data, severe binary fault assumptions, and the absence of hardware-in-the-loop or flight telemetry. The group-disjoint protocol prevents trajectory overlap but does not create independent hardware physics, since training and test orbits come from the same physics-level simulation family. The PV model omits explicit incidence angle, eclipse phase, degradation, series-resistance evolution, aging, and contamination, and the broader qualification matrix covering gain and offset error, drift, component tolerance, umbra and penumbra transitions, dynamic load steps, graded and simultaneous faults, radiation-induced parameter drift, and bit upsets remains unexecuted, with no numerical claims made for those conditions. In low Earth orbit, radiation can cause parametric drift and single-event upsets, and thermal cycling or battery aging could move a healthy trajectory toward a fault class, so a classifier without an unknown state might make a confident but unsafe assignment.

The team lays out a staged path forward: expanding the physics-level simulator across independently sampled illumination, temperature, matched load schedules, and graded or intermittent faults; processor-in-the-loop profiling with frozen quantized and float32 implementations on a flight-representative microcontroller; hardware-in-the-loop validation with a PV emulator, converters, battery emulator, and programmable loads; and eventually thermal-vacuum, radiation, and shadow-mode testing on flight-like telemetry, with autonomous recovery enabled only after false alarms, missed detections, latency, and safe FDIR interfaces meet mission requirements. For now, the message is measured: a compact, auditable, low-sensor diagnostic architecture can perfectly separate severe simulated power-system faults under controlled conditions, and the same architecture’s real-world worth will be decided on representative hardware, not in simulation.

Subject of Research: Machine-learning fault detection in a simulated nanosatellite electrical power system

Article Title: Smart fault detection in a nanosatellite electrical power system

Article References: Rezaee, A., Nobahari, N., & Taheri, A. (2026). Smart fault detection in a nanosatellite electrical power system. Discover Electronics, 3(1), Article 148. https://doi.org/10.1007/s44291-026-00301-z

Image Credits: AI Generated

DOI: 10.1007/s44291-026-00301-z

Keywords: nanosatellite, electrical power system, fault detection, machine learning, multilayer perceptron, residual classification, photovoltaic, MPPT converter, battery fault, simulation, noise sensitivity, CubeSat