This investigation showed that a Keras TensorFlow deep learning neural network model can be used to analyze process data created by a simulation, and to learn the relationships between process inputs and the resulting process yield. The model was able to determine which process variables had an effect on the yield, and which variables did not. It was also able to show the optimum operating conditions for the most important variables. 
If this were a real application, the next step would be to investigate what improvements in the process automation system would be needed to consistently achieve the best operating conditions. It may be simply tuning PID control loops, or more extensive hardware changes may be required. 
This process simulation was designed to introduce random noise in the data, but the level of noise was such that the neural network was still able to learn the underlying mathematical relationships. Higher levels of noise in the data would have caused the less influential process relationships to become unresolvable by the model and would leave higher levels of residual error in the predicted results. This is something which must be considered in real life applications. 
This investigation provides a demonstration of the power of new data analytics tools such as Keras TensorFlow and NumPy as applied to industrial process optimization. While these types of mathematical tools have been available for many decades, they are becoming much easier to use. As the tools are now freely available in a Python environment, they will be standard fare for the next generation of engineering professionals. The propagation of this technology is also being driven by its incorporation into cloud computing platforms like Microsoft’s Azure and Amazon Web Services which are strongly supported by training and certification programs. 
Overall, this investigation demonstrates how one specific aspect of IIOT technology, namely machine learning, is becoming a potential factor to improve financial performance in the process industries. 

copyright 2019 Powell Simulation LLC