Condition Monitoring and Fault Diagnosis of Aircraft Electrical Power Systems Using Data-Driven Techniques”

Abstract

Aircraft electrical power systems (EPS) are becoming larger, more electrified and more safety-critical, yet maintenance of their generators, feeders and power-conversion units is still largely scheduled or reactive. This paper presents a data-driven framework for condition monitoring and fault diagnosis of a 115/200 V, 400 Hz aircraft EPS. Three-phase voltage and current windows are acquired, reduced to 21 time-domain, frequency-domain and power-quality features, and classified into six health states: normal, overvoltage, undervoltage, open phase, feeder short circuit and rectifier (TRU) fault. Six supervised learners (logistic regression, k-nearest neighbours, support vector machine, random forest, gradient boosting and a multilayer perceptron) are trained on 2,100 windows and evaluated on 900 held-out windows. The three best models reach 97.1 to 97.8% accuracy, with the random forest achieving 97.7% accuracy and a macro-F1 score of 97.7%; 18 of its 21 test errors occur at the boundary between incipient faults and the healthy state. A noise-robustness study shows that all models remain above 90% accuracy down to 20 dB signal-to-noise ratio, while performance diverges sharply at 10 dB, where logistic regression retains 88.5% and kernel and tree-based models collapse. Feature-importance analysis identifies phase RMS voltages and current imbalance as the dominant indicators. The study uses a physics-informed synthetic dataset rather than flight or test-rig data, so the reported figures demonstrate the methodology and relative model behaviour, not certified in-service performance.

Author Affiliations & Contributions