Fault Detection for Aircraft Fuel System with Machine Learning Techniques

The complexity of aircraft fuel systems is increasing; this complexity introduces additional interdependencies between systems and makes fault diagnosis more challenging. Detecting and locating faults are necessary to ensure aircraft safety, smooth operation, and quick service restoration. Signal processing and data-driven techniques have become increasingly popular for faults detection and classification in aerospace. In this paper, machine learning techniques, namely: Neural Network (NN), K-Nearest Neighbor (KNN), Naïve Bayes, and Decision-tree (DT), are implemented and compared to detect the clogging filter problem and classify its severity. To generate the data, we developed and built a simplified aircraft fuel system to emulate the clogging filter fault in the system. The fault is emulated by using a direct-acting proportional valve (DPV3). The fully opened DPV3 is equivalent to a healthy filter, while the partially closed DPV3 is equivalent to a clogged filter with a certain degree of severity. The proposed classifiers are trained to classify the five classes of clogging filters using the pressure difference across DPV3. The generated data is split into training, validation, and testing sets. The implemented classifiers' performance is compared based on the confusion matrix and receiver operating characteristics (ROC) curve. The experimental results demonstrated that the neural network performs better than KNN Naive Bayes and DT.