Abstract:Reducing inconsistencies in fiber diameter has always been a focus of research on optical fiber drawing machines (OFDM). Based on an established mathematical model for OFDM, we adopted a long short-term memory (LSTM)-based neural network to predict fiber wire diameter and utilized these results to feedback and control the drawing speed, thereby achieving high precision control of the wire diameter. We trained the proposed network using relevant data from two different fiber drawing processes, allowing it to learn the long-term and short-term correlations in the data. Additionally, we evaluated the performance of the model using root mean square error (RMSE) and R2 as indicators. The experimental results show that network complexity has a considerable impact on the precision of fiber wire diameter prediction. We also established that two LSTM hidden layers with 64 neural units in each layer are the most suitable for the proposed network structure. The network with such a structure has the smallest RMSE value in the validation set and the largest R2 in the testing set. Subsequently, the optimal network structure was used to conduct a multi-timespan wire diameter prediction experiment. According to the experimental results, the LSTM-based wire diameter prediction model still has a satisfactory predictive effect when the time is as long as 50 s, suggesting that the precision and robustness of this network are high. When this method is applied to the actual wire drawing process, the wire diameter error can be controlled to within ± 1.8 μm when the target wire diameter is 0.66 mm, indicating that the wire diameter precision is greatly enhanced. Results indicate that using LSTM-based prediction and control effectively avoids the interference of other factors in the process of wire drawing. Therefore, we remedied the time delay issue of conventional wire diameter control methods. Also, compared with conventional methods, we achieved more precise and robust control of wire diameter.

Machine Learning Framework Using Complex Network Features to Predict Wire-length

Machine learning enables accurate wire loop profile prediction for advanced microelectronics packaging

Worst-Case Dynamic Power Distribution Network Noise Prediction Using Convolutional Neural Network

Worst-case Power Integrity Prediction Using Convolutional Neural Network.

A Novel Learning Scheme for Inversely-Designing Bonding Wire With Multi-fidelity Data Fusion

WCPNet: Jointly Predicting Wirelength, Congestion and Power for FPGA Using Multi-task Learning

Neural network based pre-placement wirelength estimation

Hierarchical Attention-Based Machine Learning Model for Radiation Prediction of WB-BGA Package

Effective Model Integration Algorithm for Improving Link and Sign Prediction in Complex Networks

Comparison of Machine Learning Techniques for Predictive Modeling of High-Speed Links

A Novel Machine Learning-Based Short-Circuit Current Prediction Method for Active Distribution Networks

Machine Learning for Complex Emi Prediction, Optimization and Localization

New-Generation Design-Technology Co-Optimization (DTCO): Machine-Learning Assisted Modeling Framework

Learning a Provably Good Wirelength Model for Analytical Placement

Graph Theory Based Machine Learning for Analog Circuit Design

Application of machine learning in wire damage detection for safety procedure

Accurate prediction of dielectric properties and bandgaps in materials with a machine learning approach

An LSTM-based adaptive prediction control model for the wire diameter control of high-precision optical fiber drawing machines

CircuitNet: an Open-Source Dataset for Machine Learning in VLSI CAD Applications with Improved Domain-Specific Evaluation Metric and Learning Strategies.

An Effective Framework for Predicting Performance of Solid-Solution Copper Alloys Using a Feature Engineering Technique in Machine Learning

Device Modeling Based on Cost-Sensitive Densely Connected Deep Neural Networks