Machine learning-based adaptive model identification of systems: Application to a chemical process Academic Article uri icon

abstract

  • Recently, sparse identification of nonlinear dynamics (SINDy) has delivered promising results in identifying interpretable models using data for various process systems. However, SINDy cannot completely comprehend the dynamics of an evolving complex process without relying on impractically large data sets. Another important challenge is that at any instance of plant-model mismatch or process upset, re-training the model using SINDy is computationally expensive and cannot guarantee to catch up with rapidly changing dynamics. As a solution to this, a systematic procedure capable of identifying and predicting the nonlinear dynamics on the fly promises to provide a useful representation of the process model. Motivated by this, we propose an adaptive model identification framework that relies on the methods of sparse regression and feature selection. The proposed method is a three-step procedure: (1) identifying potential functions from a candidate library using SINDy, (2) updating coefficients of the identified model using ordinary least-squares regression, (3) selecting the most important features using stepwise regression. Initially, a baseline model is identified off-line using SINDy, and as a new data becomes available, the subsequent on-line steps are triggered based on a pre-specified tolerance to further update the model. Such an adaptive identification scheme facilitates in perceiving the model structure using a less amount of data than its off-line counterpart, SINDy. Based on the previously proposed method, we further propose online adaptive sparse identification of systems (OASIS) framework to extend the capabilities of SINDy for accurate, automatic, and adaptive approximation of process models. The OASIS method combines SINDy algorithm and deep learning for system identification during online control of a process. First, we use SINDy to obtain multiple models from historical process data for varying input settings. Next, using these identified models and their training data, we build a deep neural network that approximates the functional relationship between SINDy coefficients and the state-input pairs in the training data. Once trained, the deep neural network is incorporated in a model predictive control framework for closed-loop operation. We demonstrate both the methods on a continuous stirred tank reactor

published proceedings

  • Chemical Engineering Research and Design

author list (cited authors)

  • Bhadriraju, B., Narasingam, A., & Kwon, J.

citation count

  • 28

complete list of authors

  • Bhadriraju, Bhavana||Narasingam, Abhinav||Kwon, Joseph Sang-Il

publication date

  • December 2019