Sound Data Augmentation Using Frequency Response Superposition for Machine Tool Component Running State Recognition

Monitoring the operation and state of a machine tool in real-time is indispensable for efficient production on manufacturing shop floors. However, many legacy machines remaining in use lack modern data communication capabilities. This necessitates methods to monitor operating states of these machines indirectly like non-intrusive sensing that converts physical measurements into operational information. Before advanced digital and information technologies emerged, skilled operators used to diagnose the operational state of a machine tool (e.g., axis movement, spindle rotation, coolant state) by simply listening to the sound that it generates. In particular, they run warm-up cycles after shifts to detect any potential anomaly and ensure the readiness of a machine. Inspired by such knowledge and practice in manufacturing shop floors, this paper proposes 1) a warm-up cycle designed to evaluate essential components of a three-axis milling machine (e.g., linear axes, spindle, coolant) which enables systematic collection of sound data and 2) a Machine-Learning (ML) model trained with a specialized data augmentation technique for recognizing the operation state of individual components based on sound signals. Given that CNC machine tool components can be modeled as linear time-invariant (LTI) systems, we devised a frequency response superposition approach for sound data augmentation. Specifically, this technique combines decomposed sound signals from ambient noise and component-specific frequency responses to simulate sounds generated by multiple components operating simultaneously. We apply the sound augmentation technique for the warm-up cycle collected via a microphone installed inside the machine. The warm-up cycle dataset includes isolated sounds from the coolant system, linear axes (X, Y, and Z-axis movements), and spindle rotation, while the evaluation dataset contains variations in axis movement speed, spindle speed, and combined operations that do not present in the warm-up cycle. Using the sound dataset augmented using the superposition method, an ML model composed of 1D convolutional layers for edge computing was trained to perform multi-label classification tasks. We evaluate the performance of our lightweight ML model using a confusion matrix and evaluation metrics specific to multi-label classification. Notably, compared to other traditional data augmentation methods, our superposition-based approach leads to the superior performance of the ML model, achieving 38.65% and 53.97% better in a macro F1 score and an exact match ratio, respectively. Future extension of the presented work includes not only its applications to cutting state recognition but also fusions with a data-driven method for further performance optimization.