Challenging the Dynamics of Time: Generate and Evaluate Real-World Time Series to estimate NOx Emissions in a Turbo Machine

Estimating accurate NOx emissions is essential to monitor pollution and health condition of a turbo machine. We use a virtual sensor to correctly estimate the particle pollution value in real-time, leveraging modern machine learning approaches. It is well known that machine learning heavily relies on data, but real-world applications encounter datarelated issues. In this work, we address the cited business use case scenario where limited data hamper an optimal regression quality. We investigate the application of time series generative models, with a particular emphasis on evaluating their performance using a comprehensive set of quantitative and qualitative metrics. We also conduct a critical analysis of the evaluation metrics commonly employed in the literature for validating and assessing the effectiveness of generative models. The analysis highlights the limitations of these metrics, as they do not take into account the temporal dependencies present in time series data and rely heavily on the specific implementation of the evaluation model. Finally, task-specific metrics are proposed to assess the effectiveness of generated data in supporting an industrial application. By delving into those pillars, this work aims to contribute to the advancement of knowledge on temporal synthetic data generation, showing how it can impact environmental care.