T. Michno¹, R. Holom², S. Schmalzer², P. Meyer-Heye¹, G. Scampone^1,3, E.Riegler^1,3, M. Hartmann^1,3, U. Repanšek⁴, N. Košir⁴, P. Šifrer⁴ and K. Poczęta⁵, ¹Austrian Institute of Technology GmbH, Austria, ² RISC Software GmbH, Austria, ³LKR Light Metals Technologies, Austria, ⁴LTH Castings d.o.o., Slovenia, ⁵Kielce University of Technology, Poland

Producing a defect-free, lightweight, high-performance and complex geometry metal components is a highly challenging task. In this paper, we focused on High Pressure Die Casting (HPDC), proposing a hybrid AI model for non-destructive, in-line, and non-process-interrupting defect prediction, using thermal images. For that, a deep neural network model is used to extract features, which are then classified by a Fuzzy Cognitive Map (FCM). Experimental results show that the method improves prediction performance. The main contributions of this research include: (i) a novel hybrid model architecture for processing thermal images, (ii) a feature extractor for a FCM-based classifier, (iii) extension of FCM via three clustering techniques to enhance classification accuracy, (iv) a modular design, allowing easy addition of other data sources and classes without retraining, (v) a thorough evaluation through model comparisons and an ablation study, and (vi) to the best of our knowledge, first usage of FCM for this problem.

HPDC, defect detection, Fuzzy Cognitive Maps, Thermal imaging, Hybrid AI, Industry 4.0.