Real-time monitoring of drilling operations in Computer Numerical Control (CNC) machine tools is crucial to ensure process reliability and to minimize downtime. This is enabled by the detection of anomalies such as missing workpieces, incorrect measurements and drill bit breakages. The aforementioned anomalies can be detected by segmenting the drilling process into distinct phases and identifing unexpected transitions. In addition, the segmented data can be utilized for a fine-grained, phase-specific monitoring and support further analytical processes, enabling the development of more sophisticated algorithms for process optimization and fault diagnosis.
This paper presents a novel, computationally lightweight unsupervised machine learning algorithm—inspired by k-means clustering—that efficiently determines a binary state indicating whether the drill is engaged or not. The algorithm continously adapts to handle complex scenarios like sensitive or noisy signals. By combining the engagement state with other machine signals a second segmentation into finer phases, like Repositioning, Air-Drilling or Unexpected Torque Drop is performed.
The presented approach is specifically designed for streaming data directly from the CNC-control unit. This eliminates the need for additional sensors or costly hardware investments and keeps the complexity to a minimum. The algorithm adapts quickly to new parameters and requires minimal memory.
Given its low computational overhead and real-time capability, the algorithm is widely applicable in industrial settings and can be seamlessly integrated into existing production lines. This work therefore constitutes an important step toward more effective and robust condition monitoring strategies in diverse manufacturing environments.