@

초록

As modern manufacturing systems increasingly employ adaptive control mechanisms, there is a need for monitoring solutions that can operate robustly across varying process conditions. In particular, the thin wire flattening process requires a control-resilient model to predict roller wear-a critical factor affecting wire quality even-as roller rotation conditions are adjusted in real-time via feedback control. To minimize roller waste and maintain product quality, precise and robust monitoring of roller wear is essential. This study proposes a dynamic conditional convolutional network (DCCN) for resilient roller wear prediction, utilizing sound data from the internal sound sensors (ISSs) and operational data from machine controllers. The DCCN dynamically adjusts its internal parameters by modulating the weights in its dynamic conditional layer based on real-time machine operational data. This mechanism enables the model to respond adaptively to changes in roller rotation conditions, allowing it to maintain high predictive accuracy across diverse operational settings. The DCCN demonstrated robust performance with 93.54% accuracy under previously unseen operational conditions, compared to only 53.42% and 73.43% accuracy achieved by conventional Convolutional Neural Network (CNN) and multimodal network models, respectively. Furthermore, layer-wise feature visualization illustrated the model's capability to provide control-invariant predictions, making it well-suited for deployment in adaptive, sustainable manufacturing systems.

키워드