.png){kind=icon}
.png){kind=icon}
Return
Data-driven tool wear prediction in milling, based on a process-integrated single-sensor approach
.png){kind=icon}
PRE
AI
E
C
DOI：10.1007/s10845-026-02939-8.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Accurate tool wear prediction is essential for maintaining productivity and minimizing costs in machining. Currently, this remains challenging due to the complex nature of milling tool wear. Traditional approaches often rely on extensive data and multi-sensor setups, which are impractical in industrial environments and restrict generalization because their invasive, machine-specific hardware cannot be easily transferred. Addressing these limitations, this study proposes a cost-effective, transferable framework using a single acceleration sensor for easy industrial integration, evaluating model architectures using minimal training data. We evaluate classical algorithms (Random Forest, Support Vector Machine, Gradient Boosting) and deep learning architectures (EfficientNetV2, LSTMs, ConvNeXt). These are trained using manually engineered feature vectors and short-time Fourier transform to perform regression (flank wear) and classification (tool state). Importantly, we evaluate transferability and robustness on two different machines with varying training data. Despite the compound challenge of a single-sensor setup with limited training data, results indicate that deep learning architectures, particularly utilizing short-time Fourier transform inputs, offer excellent performance in source domain evaluations. ConvNeXt achieved 99.1% classification accuracy and a 0.95 coefficient of determination for regression, even when trained on a dataset of just four tool life cycles. Among classical algorithms, the Random Forest Classifier showed strong results, reaching 99.3% accuracy. However, the inherent variability of milling data from a completely different machine proved challenging for all models, resulting in lower performance. Overall, the presented findings establish a solid basis for data-efficient predictive maintenance strategies that can be rapidly adapted to new industrial environments through minimal retraining.
Keywords:
Tool wear prediction
Machine learning
Milling
Industry-ready models
Data-driven approach
Accelerometer
Journal
IF:
7.4
Papers:
3.4K
Citations:
1.1W
Organization
Cited Papers