Airborne Acoustic Signal Processing and Deep Ensemble Learning for Tool Condition Monitoring in Gear Hobbing (May 2026)


Torun Y., Yerlikaya R.

IEEE Access, cilt.1, sa.1, ss.1-12, 2026 (SCI-Expanded)

  • Yayın Türü: Makale / Tam Makale
  • Cilt numarası: 1 Sayı: 1
  • Basım Tarihi: 2026
  • Doi Numarası: 10.1109/access.2026.3705350
  • Dergi Adı: IEEE Access
  • Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED)
  • Sayfa Sayıları: ss.1-12
  • Sivas Cumhuriyet Üniversitesi Adresli: Evet

Özet

Noncontact tool-condition monitoring (TCM) based on airborne cutting sound offers an inexpensive and retrofit-friendly route for predictive maintenance in industrial gear hobbing; however, its performance and deployability depend strongly on the chosen signal representation. This study has investigated four hobbing-tool conditions (intact, worn, thermally damaged, and fractured) using microphone recordings acquired from an industrial hobbing machine. Audio has been segmented into 0.5 s analysis windows with 25% overlap, and two feature representations have been compared under a common segmentation and preprocessing pipeline: (i) a high-dimensional multi-resolution spectral representation (HDMRS; 11,163 features/window) that preserves fine spectral detail via linear/mel/bark/ERB band partitions, and (ii) a compact dynamics-centric representation (CDCR; 60 features/window) designed for embedded deployment using spectral statistics and MFCC-based temporal descriptors. Three complementary multilayer perceptron (MLP) architectures (Deep, Wide, and Regularized) have been trained and combined using soft voting to form a deep ensemble. Generalization has been assessed using group-stratified 10-fold cross-validation in which all windows from the same recording remained within a single fold to prevent window-level leakage. The proposed ensemble achieved high and stable accuracy across both feature sets, whereas the compact CDCR representation retained competitive performance with a substantially lower memory footprint and input dimensionality. These findings support airborne-acoustic TCM as a practical alternative for hobbing operations and provide concrete guidance on accuracy–complexity trade-offs for real-time implementation.