Propeller Fault Classification for Unmanned Aerial Vehicles and Explainable Artificial Intelligence-Based Feature–Model Matching

The spread of unmanned aerial vehicles in daily operations makes the early and reliable diagnosis of propeller faults necessary. However, the performance values reported for such systems are usually obtained with sample level splits, and it is not known which feature representation should be matched with which learner. In this study, a leakage-free feature–model matching framework is presented for propeller fault classification. Microphone and six-axis inertial measurement unit data have been collected on a test bench with 980 kV and 1400 kV motors for one healthy and eight faulty propeller conditions at 16 throttle levels, and 8490 windows of 1 s have been extracted from 1735 measurement files. Four scalar feature sets and three time-frequency representations have been matched with seven ensemble learners and three compact convolutional networks under a file atomic split, and the permutation ranking of the best model has been returned to the feature selection stage. The highest macro-F1 value of 0.8027 and an accuracy of 0.8816 have been obtained with the stacked ensemble trained on the 52 input subset ranked by explainability. It is seen that the time domain statistics and the accelerometer axes are dominant, that three inertial axes reach a macro-F1 of 0.7661, and that cepstral and envelope features stay below the Welch-based features at the sampling rate of 90.9 Hz. The cross-motor experiments have shown that the models depend strongly on the motor class, and the McNemar test has confirmed that the difference between the ensemble branch and the compact convolutional branch is not accidental. In this way, the framework can be used as an evaluation protocol for low-cost multisensor setups on low-level devices.