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

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 mu m 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.