Comparing Feature Engineering and Deep Learning for Class-Imbalanced Vibration Signal Classification in Tool Condition Assessment

<p dir="ltr">In-process tool condition assessment (ITCA) demands precise and efficient interpretation of high-frequency vibration signals, often under conditions of pronounced class imbalance and industrial-grade noise. This study investigates the performance of eighteen different models, comprising four traditional machine learning (ML) algorithms and fourteen deep learning (DL) architectures, tasked with classifying tool wear conditions from a large-scale industrial vibration dataset. A feature-engineered ML pipeline achieves a maximum F1-score of 0.91 using k-Nearest Neighbors (kNN) but requires over 4,000 seconds for training and inference. In contrast, A combined approach using Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) as feature extractors, followed by classification with Extreme Gradient Boosting (XGBoost), reached a balanced accuracy of 0.98 and an F1-score of 0.96, along with a speed enhancement factor of 2.89×. The proposed hybrid model demonstrates superior performance compared to standalone convolutional networks (CNNs), recurrent architecture (RNNs), transformers, and autoencoders, particularly when applied to chunked input data with 50% overlap. It shows improved robustness against noise and class imbalance, outperforming models trained with Synthetic Minority Over-sampling Technique (SMOTE) on unprocessed signals. The hybrid architecture delivers the best trade-off between performance and efficiency, making it well-suited for real-time ITCA in industrial environments.</p>