Bringing new technologies and approaches to the operation and control of chemical process systems

Abstract

To set the federal interest rate, for example, we do not expect to ever need to know the current vending machine sales at our corner gas station.

Some manifold learning considerations toward explicit model predictive control

“…35,36 ANNs have also been widely used in control for decades, mostly for empirical approximation of state equations, 37 but also for explicit MPC problems. 14,[16][17][18] The main disadvantage of ANNs is that they require significant computational time for training, and Finally, GP regression, as a nonparametric Bayesian modeling technique, provides the conditional distribution of an output as a function of its observed inputs. In order to deal with higher dimensional input spaces, we introduce the automatic relevance determination weight in the covariance kernel, which employs an individual length scale hyperparameter for each input dimension.…”

“…This method has been most successful using artificial neural networks (ANNs) as the interpolation functions. 14,[16][17][18] Recently, this method was refined for constrained linear MPC problems using deep neural nets combined with Dykstra's projection to ensure constraint satisfaction. 14 In this work, we present an alternative framework for explicit MPC based on finding an appropriate function between the system state and the control policy that approximates the MPC controller.…”

A cyber‐secure control‐detector architecture for nonlinear processes

“…One example usage of big data is the detection and prediction of cyber‐attacks in the plant. In recent years, machine‐learning techniques have become increasingly popular in classical engineering fields in addition to computer science and engineering . Artificial neural networks were trained to model a pilot‐scale entrained‐flow gasifier in Reference , and were also used to predict industrially relevant observable values and/or stochastic PDE model parameters for nonlinear model predictive control of multiscale thin film deposition processes in References .…”

“…In recent years, machine-learning techniques have become increasingly popular in classical engineering fields in addition to computer science and engineering. [4][5][6] Artificial neural networks were trained to model a pilot-scale entrained-flow gasifier in Reference 7, and were also used to predict industrially relevant observable values and/or stochastic PDE model parameters for nonlinear model predictive control of multiscale thin film deposition processes in References 8,9. Conventional machine-learning methods (e.g., artificial neural network, principal component analysis, support vector machines) and more advanced deep learning methods (e.g., convolutional neural networks, long short-term memory neural networks, gated recurrent units) have demonstrated success in detecting machine and plant anomalies. 10 Model-based fault diagnosis and classification in electric drive systems were carried out using a fault diagnostic neural network in Reference 11 and automated fault detection and diagnosis of HVAC subsystems using hidden Markov models was studied in Reference 12.…”

Partial Observations and Conservation Laws: Gray-Box Modeling in Biotechnology and Optogenetics

“…Now, we use these two GB-ANN models and a BB-ANN model of the system to investigate how they perform with limited training data. We use the same ANN architecture as for the previous example, changing only the number of inputs and outputs to correspond to the larger number of species governed by conservation laws in Equation13. Training data were generated using the mechanistic model from Section 3.2.1, and are plotted inFigure 3a.…”