Reliability of a deep learning-based early warning system for table milling safety

We present a proof of concept for a deep learning-based early warning system that forecasts kickback events in handfed table milling from multivariate sensor time series. To address the lack of woodworking datasets for acute safety events, we generate a dedicated dataset using a custom experimental setup that reproducibly induces kickback and records synchronized active power, spindle speed, and triaxial vibration data under multiple process configurations. A hybrid 1D CNN-attention-BiLSTM model is trained on standardized sliding windows with class-imbalance handling and threshold optimization to balance detection performance and false alarms. Evaluation across milling configurations shows second-scale early warning lead times with millisecond inference latency, but also highlights limited transferability between configurations and insufficient robustness for safety-critical deployment, motivating further validation and methodological improvements.

Keywords: Table milling, machinery, anomaly detection, AI, machine learning.