What question did this study set out to answer?

February 21, 2026Open Access

Investigation of contact acoustic emission-based in-situ anomaly detection in wire arc additive manufacturing using deep unsupervised learning

Key Points

This research aims to improve anomaly detection in wire arc additive manufacturing using acoustic emission data and deep learning.
Proposed an acoustic emission-based monitoring strategy combined with deep learning.
Compared various anomaly detection methods, including conventional control charts and deep learning approaches.
Evaluated the influence of sensor placement on detection capability.
Introduced a semi-supervised clustering-based approach for anomaly severity classification.
Conventional control charts showed poor anomaly detection capabilities, with F2-scores at 12%.
Using frequency-domain energy improved F2-score to 49%.
Deep convolutional autoencoder enhanced performance, boosting F2-score from 62.9% to 77%.
The clustering-based approach provides effective anomaly severity stratification with only 8% anomalies in data.

Abstract

Wire Arc Additive Manufacturing (WAAM) is an arc-welding-based Additive Manufacturing (AM) process that enables the cost-effective production of large metallic components. Despite its potential, industrial deployment remains limited due to process instabilities and defect formation arising from the complex thermoelectric and metallurgical phenomena inherent to layer-by-layer deposition. Reliable online monitoring is therefore essential to ensure process consistency and part quality. This study proposes an acoustic emission (AE)-based monitoring strategy as a robust, noise-resilient alternative to microphone-based sensing, combined with an unsupervised deep learning approach to address the limitations of state-of-the-art supervised methods. The influence of sensor placement is experimentally assessed to identify an industrially viable configuration without sacrificing detection capability. Based on this analysis, a novel placement strategy is proposed. To investigate the performance of AE-based monitoring, conventional control charts and tree-based anomaly detection methods are compared with a novel approach that leverages automatic feature extraction from high-frequency acoustic emission data with a deep convolutional autoencoder (CAE). The proposed framework integrates tree-based anomaly detection with an additional clustering-based stage, enabling semi-supervised soft classification of anomaly severity. The results demonstrate that conventional control charts fail to detect subtle anomalies, whereas the use of frequency-domain energy instead of time-domain energy leads to a marked improvement in F2-score, increasing from 12% to 49%. The adoption of a convolutional autoencoder for automatic feature extraction further enhances Isolation Forest performance, raising the F2-score from 62.9% (obtained with manual feature extraction) to 77%. Moreover, it is shown that when unlabelled data are available and anomalies represent only a small fraction of the dataset, as in this case study (where they account for approximately 8%), the proposed clustering-based approach enables the stratification of anomaly severity, thereby providing a soft decision-support mechanism for prognostics.

اسأل الذكاء الاصطناعي

Bookmark

View Full Paper