Industrial Process Set-Point Optimization Using Autoencoder Regularization

The patent introduces autoencoders as intermediary components that bridge the gap between simple optimization algorithms and complex nonlinear process relationships. The autoencoders learn latent representations of process data and provide regularized predictions that capture nonlinearities without requiring the optimization algorithm itself to be complex. This mediator enables accurate decision-making while keeping the optimization framework relatively simple.

The patent transforms the optimization problem by changing parameters from direct process variables to latent space representations learned by autoencoders. By encoding process data into compressed latent features and then decoding them with regularized predictions, the system can handle nonlinear relationships through parameter transformations rather than through complex optimization algorithms.

The patent segments the system-wide optimization problem into individual process-level optimization problems. Each process has its own autoencoder that learns from its historical data, and these localized models are then coordinated through a master optimization framework. This segmentation allows system-wide optimization to benefit from merged historical data while avoiding the computational complexity of a fully integrated monolithic optimization problem.

The patent moves the optimization problem to another dimension by using latent space representations. Instead of optimizing directly in the high-dimensional space of all process variables, the autoencoders project data into a lower-dimensional latent space where optimization can be performed more efficiently. This dimensional transformation reduces complexity while preserving system-wide optimization capabilities.

The patent applies preliminary action by training autoencoders with regularizers on historical data before the actual optimization process. The regularized autoencoders learn robust patterns and relationships from past data, capturing nonlinearities and uncertainties in advance. This preliminary learning phase reduces the computational burden during real-time optimization, as the heavy lifting of handling model inaccuracies has already been done during training.

The patent uses copying by creating latent space representations and regularized predictions that replicate the complex relationships in the data without requiring the full computational complexity during optimization. The autoencoders create simplified copies of the process dynamics in latent space, which can be used for fast optimization while maintaining prediction accuracy.