An effective ANN-based hybrid fiber diameter control approach with gated recurrent units and selective weight optimization

Abstract

Reducing fiber diameter inconsistencies has been the focus of much research in the fiber manufacturing industry. However, the structure of the fiber drawing machine (FDM) leads to time delays in traditional diameter control methods, thereby making it difficult to achieve satisfactory results. Moreover, traditional methods apply only a single adjustment to control the diameter precision, i.e., drawing velocity, and ignore the crucial impact of furnace temperature on diameter. To address these issues and improve the control precision, we propose a novel hybrid diameter control model that uses an artificial neural network (ANN) as the main framework and integrates bidirectional gated recurrent units (BiGRUs) and a new mechanism known as selective weight optimization (SWO). Specifically, due to the irreplaceable role of diameter prediction in the control task, we adopt BiGRU to predict fiber wire diameter and utilize the results as a supplemental data input to the control policy network. Due to the high complexity of the input data, the ANN-based network is introduced as the policy network to analyze the inherent relevance of the data and provide two separate control strategies, i.e., drawing velocity adjustment and furnace temperature adjustment. Additionally, a selective weight optimization (SWO) mechanism that combines these two strategies and selectively optimizes the weights is proposed to further improve diameter control precision. The effectiveness of the BiGRU-ANN-SWO is extensively evaluated using four various real-world experiments. Results indicate that the BiGRU-ANN-SWO achieves considerably more precise and robust control of fiber diameter compared with existing approaches.