(719b) Dynamic Risk Assessment in Chemical Processes Using Sparse Identification and Deep Learning.

The increasing technological advancements in chemical and petrochemical industries require a monitoring system that maintains both product quality and process safety. Moreover, failure in implementing a robust monitoring system can significantly impact the process reliability and lead to human and economic losses. Therefore, to promote the optimal and safe operation of the process, developing reliable fault detection and diagnosis schemes has received significant attention in the last few years. Recently, several model-based techniques have been widely used such as observer-based [1], parity equations [2], and parameter estimation [3] methods. However, these methods require mathematical models that describe the process accurately. Also, most of these methods have been developed for a particular class of nonlinear systems. On the other hand, data-driven approaches such as principal component analysis (PCA) [4] and partial least squares (PLS) [5] have been successfully used for fault detection and diagnosis. One important limitation of these approaches is that the models are trained offline and are not updated on-the-fly. Therefore, they cannot respond to frequently changing process dynamics. Another limitation is that they do not systematically evaluate the potential risk caused by the fault. It is important to have a risk-based fault detection framework that considers the probable hazards and consequences that occur due to a fault. This guides in eliminating the non-hazardous faults and taking appropriate action in the case of severe faults [6]. Specifically, quantifying risk in real-time helps in monitoring and managing operational risk.

Motivated by the above considerations, we propose a robust dynamic risk assessment -based fault detection scheme that uses online adaptive sparse identification of systems (OASIS) framework [7]. The OASIS is an adaptive system identification method developed based on sparse identification of nonlinear dynamics (SINDy) [8] and deep learning. The SINDy algorithm solves a sparse regression problem to identify an interpretable and sparse model of the process using the historical data offline. But it is not feasible to directly implement SINDy for process monitoring as it is computationally expensive to solve a sparse regression problem online. Hence, a deep neural network (DNN) is trained to facilitate the applicability of SINDy for online monitoring. For offline training, we consider multiple trajectories of input-output data that represent a wide range of operating conditions and obtain multiple sparse models. Next, we train a DNN using these models identified by SINDy and their corresponding training inputs. Later, the trained DNN is used online to predict and update the process models using measurement data. At every sampling time, we estimate the process states using the model obtained from the DNN. For fault detection, we compute the residuals between model prediction and measurement values. At any time instant, if the evaluated residual exceeds the threshold, a fault is observed in the process. After fault detection, we perform risk assessment by computing the probability and severity of the detected faults. By doing so, we quantify the process risk associated at each sampling time. If the calculated risk exceeds the threshold, the fault detected is regarded to be severe. The proposed OASIS-based dynamic risk assessment method has the following advantages: 1) offering an adaptive framework for fault detection and dynamic risk assessment, 2) applicable to nonlinear systems with uncertain parameters, and 3) providing interpretable models that aid in understanding the relationship between process variables, which is useful in analyzing the propagation of faults. We demonstrate the proposed method for fault identification and risk assessment through the simulation of a floating liquefied natural gas tank and a non-isothermal continuous stirred tank reactor.

Literature cited:

[1] Isermann R. Model-based fault-detection and diagnosis–status and applications. Annual Reviews in control. 2005; 29(1):71–85.

[2] Blesa J, Jim´enez P, Rotondo D, Nejjari F, Puig V. An interval NLPV parity equations approach for fault detection and isolation of a wind farm. IEEE Transactions on Industrial Electronics. 2014; 62(6):3794–3805.

[3] Rahimi A, Kumar KD, Alighanbari H. Fault estimation of satellite reaction wheels using covariance based adaptive unscented Kalman filter. Acta Astronautica. 2017; 134:159–169.

[4] Bakdi A, Kouadri A. A new adaptive PCA based thresholding scheme for fault detection in complex systems. Chemometrics and Intelligent Laboratory Systems. 2017; 162:83–93.

[5] Madakyaru M, Harrou F, Sun Y. Monitoring distillation column systems using improved nonlinear partial least squares-based strategies. IEEE Sensors Journal. 2019; 19(23):11697–11705.

[6] Zadakbar O, Khan F, Imtiaz S. Dynamic risk assessment of a nonlinear non Gaussian system using a particle filter and detailed consequence analysis. The Canadian Journal of Chemical Engineering. 2015; 93(7):1201-1211.

[7] Bhadriraju B, Narasingam A, Kwon JSI. Machine learning-based adaptive model identification of systems: Application to a chemical process. Chemical Engineering Research and Design. 2019; 152:372–383.

[8] Brunton SL, Proctor JL, Kutz JN. Discovering governing equations from data by sparse identification of nonlinear dynamical systems. Proceedings of the National Academy of Sciences. 2016; 113(15):3932-3937.