| The paper proposes a hybrid machine learning framework along with a hierarchical control module for fault diagnosis, isolation, and mitigation to develop a resilient diesel engine system. The hybrid diagnostics system combines experimental data with physics-based simulation data to improve fault diagnosis, isolation, and severity prediction. The hybrid architecture consists of a denoising autoencoder to transform the engine data to a fixed lower-dimension latent space representation. The combined data is then passed to a Twin-Deep Neural Network (DNN) framework to detect and predict fault severity. The hierarchical control module consists of control calibration maps generated offline using Bayesian optimization to maintain the desired engine torque while minimizing fuel consumption. The module also uses proportional-integral (PI) and extremum seeking (ES) controllers on top of the offline map to compensate for engine faults and modeling errors. The simulation results show the efficacy of the proposed architecture to maintain the desired performance for different fault scenarios. |
*** Title, author list and abstract as seen in the Camera-Ready version of the paper that was provided to Conference Committee. Small changes that may have occurred during processing by Springer may not appear in this window.