Cutting Condition Monitoring in Milling within Enclosed Environments Using a Microphone Array and Random Forest

Technical Summary: Cutting Condition Monitoring in Milling within Enclosed Environments Using a Microphone Array and Random Forest

Problem Statement
Real-time process monitoring is critical for preventing defects and reducing costs in modern manufacturing. While acoustic emission is an attractive, non-intrusive source for monitoring machining processes, its application in enclosed machine tools is significantly challenged by environmental factors. Specifically, the interior of machine enclosures creates a highly reverberant acoustic environment due to reflective metallic surfaces. This reverberation, combined with background industrial noise, can homogenize the acoustic field, effectively masking the specific signatures of the cutting process. Furthermore, the coexistence of direct and reverberant sounds alters spectral characteristics, potentially rendering conventional broadband signal descriptors insensitive to variations in machining parameters. There is a lack of literature specifically addressing the influence of such diffuse, reverberant fields on the reliability of microphone-based monitoring and feature extraction.

Methodology
The study evaluated the feasibility of monitoring end-milling conditions using a four-microphone MEMS array positioned inside a Romi D600 vertical machining center. The experimental setup involved:

• Experimental Design: A $2^{4-1}$ fractional factorial design was employed to generate eight distinct cutting condition classes by varying cutting speed ($v_c$), feed per tooth ($f_z$), axial depth of cut ($a_p$), and radial depth of cut ($a_e$).
• Acoustic Characterization: The enclosure was characterized as a diffuse field environment. Reverberation time ($RT_{60}$) was measured using an impulsive firecracker source, and spatial coherence was analyzed to validate the diffuse field hypothesis. The microphone array was positioned at the critical distance, resulting in a hybrid field where reverberant energy dominates.
• Feature Extraction Strategies: Two strategies were compared:
  1. Strategy 1 (Conventional): Extraction of standard time- and frequency-domain descriptors (e.g., RMS, spectral centroid, kurtosis) from the broadband signal.
  2. Strategy 2 (Proposed): Decomposition of the signal into octave bands (125 Hz to 8 kHz), followed by the application of a Wiener filter to each band to mitigate noise and reverberation. Root-mean-square (RMS) descriptors were then extracted from these filtered sub-bands.
• Classification: A Random Forest classifier was trained on data from one experimental trial (Trial 1) and validated against five independent trials (Trials 2–6) to assess generalization and robustness.

Key Results

• Environmental Characterization: The machine enclosure exhibited a reverberation time of approximately 0.64 s and a Schroeder frequency of ~752 Hz. Above this frequency, the acoustic field behaved as a statistically diffuse field. Spatial coherence analysis confirmed a high agreement (similarity index ~0.91) with the Rayleigh diffuse field model.
• Feature Sensitivity:
  • Strategy 1: Conventional broadband descriptors showed high susceptibility to the homogenizing effects of reverberation. The spectral centroid, for instance, showed almost no variation across different cutting parameters, indicating that global spectral balance remained unchanged despite energy variations.
  • Strategy 2: The octave-band RMS descriptors preserved significant sensitivity to machining parameters. The proposed method successfully separated the influence of different parameters (e.g., feed per tooth dominating lower frequencies, cutting speed dominating higher frequencies) which were obscured in the broadband analysis.
• Classification Performance:
  • Strategy 1: Achieved an average classification accuracy of 64.85% with a high standard deviation of 26.07%, indicating poor generalization across different experimental trials.
  • Strategy 2: Achieved an average classification accuracy of 94.77% with a standard deviation of only 7.92%. The proposed method maintained high accuracy (>90%) across most validation trials, demonstrating superior robustness against environmental fluctuations and reverberation.
• Feature Importance: The Random Forest model relying on Strategy 2 primarily utilized mean RMS values from specific octave bands (particularly 4 kHz, 8 kHz, 2 kHz, and 250 Hz), which overlapped with the structural resonance frequencies of the tool-holder system.

Significance and Claims
The paper claims that while the enclosed environment creates a diffuse sound field that significantly alters acoustic signals, it does not completely mask the cutting process signature. The primary contribution is the demonstration that frequency-selective analysis (octave-band decomposition combined with Wiener filtering) is essential for effective monitoring in such environments.

The authors assert that conventional broadband descriptors fail in highly reverberant settings because they average out process-related information across the spectrum. In contrast, the proposed methodology retains localized process information within specific frequency bands, providing a substantially more robust representation of the machining state. The study concludes that acoustic monitoring of milling operations inside enclosed machine tools is feasible using low-cost MEMS arrays, provided that signal processing strategies account for the specific acoustic properties of the enclosure. The findings suggest that preserving frequency-selective information is critical for overcoming the limitations imposed by industrial reverberation and background noise.