Unsupervised wear detection for abrasive tools using audio features and dual-masked graph autoencoder

Abrasive tool wear detection is a critical technology for advancing Industry 4.0, ensuring high-quality and efficient machining in industries such as aerospace and optoelectronics. However, developing a high-accuracy, non-contact detection method that performs reliably under complex operating condit...

Celý popis

Uložené v:
Podrobná bibliografia
Vydané v:Engineering applications of artificial intelligence Ročník 160; s. 111931
Hlavní autori: Shi, Shuangjin, Tang, Lili, Tian, Hui, Huang, Hui, Chang, Ching-Chun, Chang, Chin-Chen
Médium: Journal Article
Jazyk:English
Vydavateľské údaje: Elsevier Ltd 23.11.2025
Predmet:
Abrasive tool
Audio signals
Graph autoencoder
Unsupervised wear detection
Unsupervised wear detection
Graph autoencoder
Audio signals
Abrasive tool
ISSN:0952-1976
On-line prístup:Získať plný text
Tagy: Pridať tag
Žiadne tagy, Buďte prvý, kto otaguje tento záznam!
Popis
Shrnutí:Abrasive tool wear detection is a critical technology for advancing Industry 4.0, ensuring high-quality and efficient machining in industries such as aerospace and optoelectronics. However, developing a high-accuracy, non-contact detection method that performs reliably under complex operating conditions remains a significant challenge. To address this problem, this paper presents a novel unsupervised approach for abrasive tool wear detection, leveraging audio-cepstral features and a dual-masked graph autoencoder, marking the first application of graph-based structures in audio-driven tool wear detection. The proposed method models audio signals with a cross-graph structure, fully integrating both numerical and structural characteristics of mel-frequency cepstral coefficients. A dual-masked graph autoencoder is introduced to learn wear-specific features, enhancing robustness against by masking the nodes again in the decoding. Additionally, a novel loss function for wear detection is designed by combining mean squared error and cosine similarity with focal and weighting factors, enabling a comprehensive reconstruction of both value differences and directional relationships in non-Euclidean graph structures. Finally, a lightweight wear detection network uses the learned features for accurate classification of wear states. Experimental results demonstrate robust and accurate detection performance under eight complex operating conditions (including various operating conditions and noise types). The proposed method achieves an area under the curve of 0.9461 in mixed conditions, outperforming state-of-the-art methods by 1.48–21.59%. Ablation studies on backbone network, masking schemes, graph transformations, and loss functions further validate the effectiveness of the proposed method.
ISSN:0952-1976
DOI:10.1016/j.engappai.2025.111931