LSTM RNN in support of Manufacturing Simulation-based Digital Twin; A Case from Offshore Wind Turbine Production

This paper describes how the application of the long short-term memory recurrent neural network (LSTM RNN) model can support the simulation-based digital twin (Sim.-based DT) of a manufacturing system in predicting the new outcome of production upon a changed parameter in the manufacturing setup. Simulation of the assembly production flow, as the only comprehensive ground-truth of the flow, represents the sequence and multi-factor dependencies of the assembly processes (associated with the spatial and environmental parameters) and renders a realistic realization of the production process. Manufacturing digital twin (DT) has adopted such virtualization capability of the simulation [1] to capture, identify, and analyze the real-time abnormalities during the flow and mitigate the risk alongside the process. But simulation software systems are impotent in handling heavy detailed graphics, inflexible (by hard-coded logic) for optimization scenarios, and unresponsive to real-time changing parameters. This paper proposes a dual-layer LSTM RNN model, backed by three Dense layers, to be trained by the production simulation, so to render the generative pattern of the flow behavior and the predicted production outcome when the determinant parameters are changed in real-life. The final model will be a part of a bigger proposal for a hybrid optimization model, which is under development and not covered by the scope of this paper. However, the stand-alone proposed LSTM RNN model in this paper has capability to enhance the functionality of the Sim.-based DT in response to the real-time variations in determinant parameters. With the use of data from offshore wind turbine (OWT) generator's production line in Siemens Gamesa Renewable Energy (SGRE) Nacelle factory in Denmark and the corresponding simulation, a pilot model has managed to predict the total number of operators and their allocation to the workstations when a tact time is changed. This study denotes the logic behind the concept, methodologies, and frameworks, which are reusable in any typical manufacturing system supported by simulation. The result of the training evaluation is an affirmative proof of concept for the hypothesis, although, to utilize the trained model in SGRE's real life, there are still some hurdles to overcome. The paper concludes with highlights of difficulties, risks, and limitations of the model helping other industries to predict potential challenges and realistically estimate the required budget for such implementation.