Hier finden Sie wissenschaftliche Publikationen aus den Fraunhofer-Instituten.

A Spatio-Temporal Ensemble Deep Learning Architecture for Real-Time Defect Detection during Laser Welding on Low Power Embedded Computing Boards

: Knaak, Christian; Eßen, Jakob von; Kröger, Moritz; Schulze, Frédéric; Abels, Peter; Gillner, Arnold

Volltext ()

Sensors. Online journal 21 (2021), Nr.12, Art. 4205, 28 S.
ISSN: 1424-8220
ISSN: 1424-8239
ISSN: 1424-3210
Zeitschriftenaufsatz, Elektronische Publikation
Fraunhofer ILT ()
real-time process monitoring; recurrent neural network; high-speed infrared imaging; convolutional neural network; lack of fusion (false friends)

In modern production environments, advanced and intelligent process monitoring strategiesare required to enable an unambiguous diagnosis of the process situation and thus of the finalcomponent quality. In addition, the ability to recognize the current state of product quality inreal-time is an important prerequisite for autonomous and self-improving manufacturing systems.To address these needs, this study investigates a novel ensemble deep learning architecture based onconvolutional neural networks (CNN), gated recurrent units (GRU) combined with high-performanceclassification algorithms such as k-nearest neighbors (kNN) and support vector machines (SVM).The architecture uses spatio-temporal features extracted from infrared image sequences to locatecritical welding defects including lack of fusion (false friends), sagging, lack of penetration, andgeometric deviations of the weld seam. In order to evaluate the proposed architecture, this studyinvestigates a comprehensive scheme based on classical machine learning methods using manualfeature extraction and state-of-the-art deep learning algorithms. Optimal hyperparameters for eachalgorithm are determined by an extensive grid search. Additional work is conducted to investigatethe significance of various geometrical, statistical and spatio-temporal features extracted from thekeyhole and weld pool regions. The proposed method is finally validated on previously unknownwelding trials, achieving the highest detection rates and the most robust weld defect recognitionamong all classification methods investigated in this work. Ultimately, the ensemble deep neuralnetwork is implemented and optimized to operate on low-power embedded computing devices withlow latency (1.1 ms), demonstrating sufficient performance for real-time applications.