Neerhof, Matthijs (2026) Transfer Learning and Multi-Signal Fusion for Automated Tool Wear Classification in Precision Manufacturing. Research Project, Industrial Engineering and Management.
|
Text
TransferLearningandMultiSignalFusionforAutomatedToolWearClassificationinPrecisionManufacturingIEMResearchProjectMatthijsNeerhof.pdf Download (7MB) | Preview |
Abstract
Reliable tool condition monitoring is a prerequisite for automated precision manufacturing. In the Demcon Lens Lathing Cell, cutting tools are currently replaced based on operator experience, which introduces unnecessary production stops and risks producing out-of-specification lenses. This thesis investigates whether ImageNet-pretrained vision models, fine-tuned on Short-Time Fourier Transform spectrograms of process signals, can accurately classify tool wear in two distinct settings: the PHM2010 milling benchmark and the Demcon industrial dataset. Performance is evaluated on three metrics: window accuracy, balanced accuracy, and cumulative signal accuracy. Cumulative signal accuracy reflects practical deployment by assessing tool condition once per cutting operation rather than per window. Four pretrained backbones are evaluated as single-signal models: VGG-16, ResNet-50, ViT-B/16, and Swin-V2-B. Four multi-signal fusion architectures are additionally evaluated: RGB Channel Fusion, Early Fusion, Cross-Attention Fusion, and Late Fusion. On PHM2010, ViT-B/16 achieves a mean balanced accuracy of 0.975 on the Fx force signal, competitive with state-of-the-art methods that use more complex architectures or larger sensor arrays. On the Demcon dataset, which uses encoder tracking error signals rather than dedicated force or vibration sensors, ViT-B/16 achieves a cumulative signal accuracy of 0.951 on the Ze signal. Fusion of the Xe and Ze signals produces substantially larger gains than on PHM2010: Early Fusion and Cross-Attention Fusion both achieve a perfect cumulative signal accuracy of 1.000, correctly classifying all 102 signals across four tool condition classes. The results confirm that fine-tuning frozen pretrained backbones on STFT spectrograms is an effective strategy for tool wear classification, even with limited industrial data.
| Item Type: | Thesis (Research Project) |
|---|---|
| Supervisor name: | Cheng, L. and Liu, H. |
| Degree programme: | Industrial Engineering and Management |
| Thesis type: | Research Project |
| Language: | English |
| Date Deposited: | 09 Apr 2026 14:03 |
| Last Modified: | 09 Apr 2026 14:03 |
| URI: | https://fse.studenttheses.ub.rug.nl/id/eprint/37281 |
Actions (login required)
 |
View Item |