AI-Based Insert Classification via Acoustic Emission Signal
(2026) MMTM05 20261Production and Materials Engineering
- Abstract
- Automated sorting and classification of scrap cemented carbide cutting inserts represent a critical technological frontier for sustainable industrial recycling. This master's thesis project successfully develops, evaluates, and methodologically deconstructs an artificial intelligence (AI)-based sensor framework utilizing Acoustic Emission (AE) signals to identify discrete material grades. The preliminary baseline classifiers operating under idealized, unconstrained data achieved an overfitted validation accuracy of 92.5%. Through a systematic suite of feature rankings via ANOVA and Shapley importance, and statistical manifolds mapped by PCA and MD, the interlocking relationships between physical confounding variables and statistical... (More)
- Automated sorting and classification of scrap cemented carbide cutting inserts represent a critical technological frontier for sustainable industrial recycling. This master's thesis project successfully develops, evaluates, and methodologically deconstructs an artificial intelligence (AI)-based sensor framework utilizing Acoustic Emission (AE) signals to identify discrete material grades. The preliminary baseline classifiers operating under idealized, unconstrained data achieved an overfitted validation accuracy of 92.5%. Through a systematic suite of feature rankings via ANOVA and Shapley importance, and statistical manifolds mapped by PCA and MD, the interlocking relationships between physical confounding variables and statistical features were successfully decoupled.
Guided by these insights, the final classifier was successfully insulated from cross-geometry feature masking by enforcing a strict geometric consistency constraint, restricting the training domain exclusively to a singular geometry (SEKN). By eliminating cross-variable entanglement, the definitive model achieved a validation accuracy of 86.7% and a robust independent blind testing score of 83.2% on randomized out-of-distribution data. However, blind cross-geometry testing on alternative shapes (LFA and SNUN) resulted in a performance collapse to 41.8%, proving that the material composition fingerprints overlapped under their geometrical conditions.
Ultimately, this thesis outlines the statistical boundaries of data-driven acoustic sensing, establishing a clear paradigm for future research to deploy advanced domain adaptation algorithms or multi-model ensemble architectures to achieve true geometric invariance in automated industrial recycling. (Less) - Popular Abstract
- Every year, millions of cutting tools containing rare and high-risk metals like tungsten and cobalt are discarded. Recycling them requires flawless sorting by chemical recipe, yet current industrial methods are either too slow or fail when tools are dirty and worn. This thesis introduces a breakthrough solution: giving recycling systems a pair of "digital ears" to identify material grades using sound. By tapping each tool and recording the Acoustic Emission (AE) waves passing through it, we capture a unique acoustic "fingerprint" to train an Artificial Intelligence (AI) classifier. While the AI initially achieved a brilliant 92.5% accuracy, we uncovered a hidden trap: the AI was taking a shortcut by judging the shape of the tools rather... (More)
- Every year, millions of cutting tools containing rare and high-risk metals like tungsten and cobalt are discarded. Recycling them requires flawless sorting by chemical recipe, yet current industrial methods are either too slow or fail when tools are dirty and worn. This thesis introduces a breakthrough solution: giving recycling systems a pair of "digital ears" to identify material grades using sound. By tapping each tool and recording the Acoustic Emission (AE) waves passing through it, we capture a unique acoustic "fingerprint" to train an Artificial Intelligence (AI) classifier. While the AI initially achieved a brilliant 92.5% accuracy, we uncovered a hidden trap: the AI was taking a shortcut by judging the shape of the tools rather than their actual material composition. By restricting the training data to a single geometry, we forced the AI to ignore this shortcut, leading to a robust 83.2% accuracy on independent test samples of that shape. However, testing the same model across different shapes caused the accuracy to collapse to 41.8%. This research serves as a critical stepping stone for green manufacturing, proving that sound-based AI sorting is highly viable, but warning future engineers that they must mathematically "decouple" object geometry from material properties before safe deployment in industrial recycling environments. (Less)
Please use this url to cite or link to this publication:
https://lup.lub.lu.se/student-papers/record/9235910
- author
- Yu, Sihong LU
- supervisor
- organization
- course
- MMTM05 20261
- year
- 2026
- type
- H2 - Master's Degree (Two Years)
- subject
- keywords
- Cemented Carbide Sorting, Acoustic Emission, AI Classification
- other publication id
- LUTMDN/(TMMV-5397)/1-119/2026
- language
- English
- id
- 9235910
- date added to LUP
- 2026-06-11 14:14:26
- date last changed
- 2026-06-12 13:17:53
@misc{9235910,
abstract = {{Automated sorting and classification of scrap cemented carbide cutting inserts represent a critical technological frontier for sustainable industrial recycling. This master's thesis project successfully develops, evaluates, and methodologically deconstructs an artificial intelligence (AI)-based sensor framework utilizing Acoustic Emission (AE) signals to identify discrete material grades. The preliminary baseline classifiers operating under idealized, unconstrained data achieved an overfitted validation accuracy of 92.5%. Through a systematic suite of feature rankings via ANOVA and Shapley importance, and statistical manifolds mapped by PCA and MD, the interlocking relationships between physical confounding variables and statistical features were successfully decoupled.
Guided by these insights, the final classifier was successfully insulated from cross-geometry feature masking by enforcing a strict geometric consistency constraint, restricting the training domain exclusively to a singular geometry (SEKN). By eliminating cross-variable entanglement, the definitive model achieved a validation accuracy of 86.7% and a robust independent blind testing score of 83.2% on randomized out-of-distribution data. However, blind cross-geometry testing on alternative shapes (LFA and SNUN) resulted in a performance collapse to 41.8%, proving that the material composition fingerprints overlapped under their geometrical conditions.
Ultimately, this thesis outlines the statistical boundaries of data-driven acoustic sensing, establishing a clear paradigm for future research to deploy advanced domain adaptation algorithms or multi-model ensemble architectures to achieve true geometric invariance in automated industrial recycling.}},
author = {{Yu, Sihong}},
language = {{eng}},
note = {{Student Paper}},
title = {{AI-Based Insert Classification via Acoustic Emission Signal}},
year = {{2026}},
}