Feasibility Study of Transfer Learning on LSTM Recurrent Neural Networks for Fiber Manufacturing Commercialization

Abstract

This thesis explores business pathways to commercialize Device Realization Lab’s technology that uses deep reinforcement learning for optical fiber manufacturing control systems. A viable business solution is proposed based on feedback from venture capital investors. The solution comprises developing cloud-based software that can generate digital twins for fiber manufacturing companies. These digital twins can serve as anomaly detectors and suggest optimal input parameters that reduce production variation and tolerance, improving quality and decreasing scrap rate. Efforts to define a minimum viable product (MVP) for this business solution began with the creation of a long short-term memory recurrent neural network (LSTM RNN) model for a desktop fiber extrusion system that mimics the fiber extrusion process on the manufacturing floor. Transfer learning on the LSTM RNN was then implemented to explore the feasibility of reusing a well-developed machine learning (ML) model for a fiber material (e.g. glass fiber) to construct an ML model for a separate fiber material (e.g. nylon fiber) for which a relatively low amount of data is available. The study found that applying transfer learning reduced the mean squared error of the new fiber material model by over 40% compared to developing the model without transfer learning. This thesis strives to reveal the innovative applications of the technology that can benefit the fiber manufacturing field and defines an MVP that can be shared with venture capital investors as a first step toward commercializing this technology.