PHYSICS-BASED AND DATA-DRIVEN MODELLING APPROACHES FOR INTELLIGENT MANUFACTURING PROCESS

In modern industrial production, achieving consistent product quality under variable process conditions requires both mechanistic understanding of the process and real-time monitoring capabilities. These tasks imply the development of suitable models that relate process parameters and in-process sensor signals to the final performance, to guide optimization, control actions, or warnings. On the other hand, the recent rapid development of machine learning methods, driven by the availability of sensor data enabled by Industry 4.0, is pushing towards the adoption of data-driven modeling approaches. On these premises, this thesis explores the possibility of modeling a small-batch industrial process using data-driven approaches while preserving its interpretability. We start from a concrete production setting by targeting a challenging machining process: centreless grinding. Centerless grinding plays a crucial role in producing cylindrical components with high throughput and reduced setup and operating times. Due to uncertainties in key process variables and the nonlinear, time varying nature of the process, current industrial practice often relies on extensive parameter exploration and costly measurements of initial and final profiles to ensure end-product conformance. This thesis addresses both quality prediction and monitoring issues. First, the process is formulated as a discrete linear dynamical system, with each stage represented by an affine transformation that captures the linearized process dynamics and nominal material removal. An augmented state enables the model to reproduce contact-filtering effects resulting from the engagement of the grinding wheel with the workpiece. The affine transformation can be constructed from Physics-Based principles or identified from experimental data sets in a purely Data-Driven manner, enabling rapid estimation of final profile quality while providing insights into profile evolution. The resulting affine transformations achieved a relative Hybrid Error below 10% and a linear correlation of approximately 0.9, demonstrating their ability to accurately predict the final workpiece profile. Second, a pure Data-Driven monitoring framework is proposed: it predicts workpiece final quality using only signals available from the machine’s Numerical Control, without requiring measurements of the initial profile, thereby saving measurement costs for the industry. Features are extracted from the time-domain, frequency-domain, time–frequency, and nonlinear domains, normalized, and combined with the final measured roundness to form a labelled dataset. Two Bayesian sparse regression models, specifically one with Least Absolute Shrinkage and Selection Operator regularization and one with conjugate priors for Stochastic Search Variable Selection, along with Gaussian Processes and a Random Forest are then used to provide predictions of workpiece roundness. Results show a Root Mean Squared Error of around 1 µm and a relative Hybrid Error of about 10% on the test set, which is adequate to assess the quality of most standard production. Bayesian-informed regression models provided the best predictive accuracy and enabled the identification of key features, while ridge regression models with correlation-guided feature selection provided compact models suitable for real-time implementation on the machine numerical control. Physics-guided analysis is conducted to investigate the impact of various process parameters and to develop additional features, leading to simpler surrogate models. This strategy improves interpretability and robustness, facilitating practical implementation in industrial environments. Overall, this thesis presents a framework for process optimization and real-time workpiece assessment. The resulting framework provides practical solutions for high-performance, intelligent manufacturing while advancing both theoretical and applied knowledge in precision machining.