Development and application of AI-based algorithms for engine emissions virtual sensing and vehicle NVH fault detection and diagnosis

This PhD thesis, developed in collaboration with the Ferrari Powertrain Development and Testing department, aims to develop algorithms and models to manage and analyse the available data, coming from experimental testing carried out in test cells, roller benches and on-board with real vehicle prototypes. To this end, the possibilities of the AI-based algorithms are explored in two different applications: (i) pollutant emissions virtual sensing and (ii) NVH fault detection and diagnosis. The data from experimental tests on the Ferrari Purosangue are used to train and develop virtual sensors based on an enhanced version of the Light Gradient Boosting Regressor. The developed virtual sensor was used to predict both engine-out and tailpipe emissions, showing excellent results. Once the system is validated, it is deployed as a software to predict emissions during driving tests without physical sensors and in virtual environment to identify potential emissions-critical manoeuvres. The second part of NVH fault detection and diagnosis is carried out by also involving the Test Division of Siemens Digital Industries. In this activity, systems based on autoencoders and a 2D-CNN classifier are developed to automatically detect and diagnose the faults emerged during the SF90 Stradale end-of-line testing, using the in-cabin audio recordings. To address the lack of anomalous experimental recordings, an NVH Simulator developed by Siemens is employed to synthesize a large dataset of faulty acoustic signals to train the AI models. Finally, these models are tested on synthesized sounds, showing excellent fault detection and diagnosis accuracy. Fine-tuning the model based on few-shot learning allows to improve the system’s accuracy even for applications on experimental data, making it implementable on future end-of-line testing. In conclusion, two innovative AI-based tools for emissions virtual sensing and NVH fault detection and diagnosis have been successfully developed and deployed to fulfil current needs of automotive manufacturers.