The aim is to develop algorithms able to take into account the uncertainty in the learning database of neural networks. The project fits into the context of the dynamic state estimation of liquid-liquid extraction and benefits of its knowledge-based simulator as well as industrial data. Indeed, the status of an industrial chemical process is accessible through operating parameters and available monitoring measures. However, the measures being inherently associated with uncertainty, it is necessary to make the data consistent with process knowledge. Therefore, the goal is to find the best data set of operational parameters (input of the knowledge-based simulator) to provide the model to estimate the real process state known through monitoring measures (output of the knowledge-based simulator). A convolutional neural network (CNN) is being developed in another postdoctoral project to solve the inverse problem to find the best input thanks to the measured output. A consistent set of operating parameters is going to be obtained and state of the process is going to be known during the dynamic regime of the liquid-liquid extraction process. This first step is to evaluate the impact of the uncertainty of operational parameters on the outputs of the knowledge-based model. This step will need to connect the knowledge-based model to URANIE, internal platform developed by CEA ISAS. This knowledge must be taken into account in the second part of the project. The uncertainty observed on the outputs should be taken into account in the learning loop to improve the estimation of the operational parameters by the CNN. The impact of these uncertainties on the CNN computed results must be assesed in order to trust the ability of the CNN to estimate the state of the process.
Through this project, we are at the heart of the thematic of digital simulation for the best control of complex systems.