Cutting condition monitoring in milling within enclosed environments using a microphone array and Random Forest

In modern manufacturing, monitoring systems are essential for detecting process deviations and enabling adjustments before tool or part failure. Acoustic sensing offers a practical approach for capturing machining dynamics, but signals inside enclosed machine tools are frequently contaminated by background noise and severe reverberation. This study investigates the sensitivity of acoustic signals to variations in cutting conditions using a microphone array positioned inside an enclosure, aiming to establish a basis for future process monitoring. Experiments were carried out during end-milling operations with a four-microphone array mounted laterally to the spindle. To retain frequency-dependent information in the measured acoustic signals that may be lost when broadband statistical features are used, the proposed framework combines octave-band decomposition with spatial averaging across the microphone array. Before building the feature-extraction pipeline, the acoustic field inside the enclosure was quantitatively characterized and validated against a diffuse-field model, providing a physical basis for the frequency-selective approach. Performance was evaluated by comparing two strategies: conventional broadband time- and frequency-domain features from a single microphone versus multi-channel spatial-average descriptors from octave-band signals across all four microphones. These features represent the band-wise acoustic energy measured across the array under the combined influence of cutting excitation, structural dynamics, and sound propagation within the enclosure. While conventional broadband features showed lower sensitivity to variations in machining parameters under the tested reverberant conditions, the proposed multi-channel, octave-band approach retained substantially higher sensitivity to these variations. The Random Forest classifier was trained on a single experimental run and subsequently evaluated across five unseen test trials. The conventional single-channel strategy achieved an average accuracy of $$59.69%$$ ( ± 14.47% ), whereas the multi-channel frequency-selective strategy reached $$97.70%$$ ( ± 1.81% ). These findings demonstrate that combining octave-band decomposition with multi-microphone spatial averaging improves the discrimination of the investigated cutting conditions under the noise and reverberation present in the enclosed machining environment.