Knowledge-based grinding burn detection: Signal segmentation, feature selection, and machine learning

Grinding burn is a critical defect that compromises the integrity of high-precision components. While analytical models rely on rigorous physical formulations, they often lack the flexibility to adapt to real variability. In contrast, standard machine learning (ML) also fails by ignoring grinding process understanding. To bridge this gap, this study transfers expert knowledge to the model by extracting signal features from the distinct phases of the grinding cycle. Signals from Computer Numerical Control (CNC), accelerometers, and acoustic emissions were segmented into approach, roughing, and spark-out stages to enable independent phase learning. Additionally, a feature selection strategy was proposed, successfully reducing the 1425 initial features to a set of just 32. Machine learning models reached 96.34% accuracy in known scenarios, far exceeding the 76.72% analytical baseline. When tested on grinding conditions different from those used during training, the staged methodology achieved 80%–85% accuracy. This performance not only outperforms the analytical results, but also notably surpasses standard ML global models, which failed with accuracies below 60%. Embedding the physical structure of the process into machine learning provides the contextual awareness necessary for reliable industrial monitoring, enabling superior generalisation and robustness compared to both traditional analytical models and standard data-driven techniques.