Predictive maintenance anticipates and prevents component failures by analysing operational data for early signs of degradation. Traditional industry-standard models for aircraft systems are often rule-based, missing complex patterns and limiting scalability. This research develops a deep learning (DL) fault detection pipeline for the engine bleed air system of wide-body twin-engine aircraft, leveraging real-world operational sensor and maintenance data. An interpretable feature engineering framework extracts physically informed features, including dual-engine comparisons, to train a gated recurrent unit (GRU) fault detection model for robust temporal modelling of healthy and faulty conditions. Bayesian optimisation is implemented for hyperparameter tuning. However, the scarcity of representative failure data, which is a common issue in aviation, limits the achievable performance and fidelity of DL models. To address this, a Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) is employed to augment the dataset with synthetic failure data. A post-processing block labeling technique is introduced to enhance fidelity, and a novel fidelity savings metric translates model predictions into operational savings. GAN-based augmentation enhances recall, precision, and F1-score, and outperforms traditional augmentation. Further case study results show that the WGAN-GP-augmented model delivers four times the operational savings compared to the industry-standard model, 50\% more than the non-augmented GRU, and 31\% more than traditional augmentation.

The detection of anomalous behaviour is fundamental to component health analysis techniques. However, detecting anomalies is a difficult and time consuming task if their form, location, and frequency are unknown. This research introduces an innovative unsupervised predictive maintenance pipeline that requires minimal domain knowledge and time to create competitive and insightful health monitoring models. First, a Dilated Convolutional Autoencoder learns to recreate healthy sensor data. Then, a Kullback-Leibler (KL) divergence based health analysis transforms discrepancies between the reconstruction and the sensor data into a single performance metric per sensor per flight. A novel evaluation method based on the KL divergence metric allows for quantitative evaluation and hyperparameter tuning of the autoencoder. Results provide new insights and show competitive performance on analysing the fuel level measuring system. Additionally, in a generalisability study on the braking system of a different aircraft type the proposed method outperforms the currently employed health monitoring model in precision and F1 score. The main advantages of the proposed method are; the ability to rapidly create unbiased health indicators on a sensor level, the capability to generalise to other components, and a framework to quantitatively evaluate the model’s performance when no truth labels are available.