Wan, J., Yar, K. P., Hean Low, M. Y., & Du, C. (2025). Comparing Traditional Machine Learning and Deep Learning Methods for Imbalanced Vibration Signal-Based Tool Condition Monitoring. In (Editor), 2025 7th International Conference on Industrial Artificial Intelligence (IAI). https://doi.org/10.1109/iai68403.2025.11276906
Abstract:
In-process tool condition monitoring
(TCM) in the milling process demands precise
interpretation of high-frequency vibration signals under
pronounced class imbalance and industrial noise. This
study investigates the performance of eighteen models (4
classical ML and 14 DL architectures) tasked with
classifying tool wear conditions from a large-scale
vibration dataset. A feature-engineered ML pipeline
achieves an F1-score of 0.91 with k-Nearest Neighbors
(kNN) but requires ~55,000 s (≈15.3 hours) of training
(including feature extraction/selection) and 2.24 s per
inference. In contrast, a hybrid CNN+LSTM feature
extractor with an XGBoost classifier achieved balanced
accuracy 0.98 and F1-score 0.96, with an ≈11× faster
runtime. The hybrid model outperforms standalone CNN,
RNN, transformer, and autoencoder models, especially on
overlapped chunked input. It is more robust to noise and
class imbalance (even outperforming models that used
SMOTE on raw signals) and offers the best performance–
efficiency trade-off, making it well-suited for real-time
TCM.
License type:
Publisher Copyright
Funding Info:
This research / project is supported by the A*STAR - RIE2025 Manufacturing, Trade and Connectivity (MTC) Industry Alignment Fund- Pre-Positioning (IAF-PP)
Grant Reference no. : M23L4a0001