Search

In this study, we propose a novel process control strategy by assimilating multiple neural network (MNN) and reinforcement learning (RL) to generate an MNNRL controller. The artificial neural networks (ANNs) constituting the MNN are trained using repositories of optimal control and state data to predict control actions to attain or stay at a set-point. This ability is significantly improved by developing an RL component for the MNNRL controller to provide enhanced set-point tracking, robustness against disturbances, and real-time learning to successfully adapt in the presence of persistent parameter upsets. The controller performance is examined in three different case studies utilizing the simulation of continuous processes with varying degrees of nonlinearity. Relative to both MNN and nonlinear model predictive control (NMPC), the MNNRL controller is found to provide better performance with about 32% less lower integral absolute error (IAE), and 50% reduced settling time with suppressed over-/under-shoots in controlled variables. Further, it is observed that the controller is fairly robust against disturbances, and can adapt via real-time learning to altered process dynamics caused by any persistent process upsets. The controller computation time is observed to be about an order of magnitude less in comparison to NMPC. • Multiple neural network, and reinforcement learning are combined for process control • The first component provides baseline control, which is improved by the second one • The resulting controller has superior performance in three different case studies [ABSTRACT FROM AUTHOR]

Meta-learning is a branch of machine learning which trains neural network models to synthesize a wide variety of data in order to rapidly solve new problems. In process control, many systems have similar and well-understood dynamics, which suggests it is feasible to create a generalizable controller through meta-learning. In this work, we formulate a meta reinforcement learning (meta-RL) control strategy that can be used to tune proportional–integral controllers. Our meta-RL agent has a recurrent structure that accumulates "context" to learn a system's dynamics through a hidden state variable in closed-loop. This architecture enables the agent to automatically adapt to changes in the process dynamics. In tests reported here, the meta-RL agent was trained entirely offline on first order plus time delay systems, and produced excellent results on novel systems drawn from the same distribution of process dynamics used for training. A key design element is the ability to leverage model-based information offline during training in simulated environments while maintaining a model-free policy structure for interacting with novel processes where there is uncertainty regarding the true process dynamics. Meta-learning is a promising approach for constructing sample-efficient intelligent controllers. • An expert agent recognizes and controls any process in a large class. • Offline training leads to sample-efficient online operation and adaptation. • The online agent adapts automatically as plant characteristics drift. • The agent adjusts PI parameters instead of directly actuating the plant. • meta-RL design generalizes to various plant and controller structures. [ABSTRACT FROM AUTHOR]

We present a novelty detection framework for Convolutional Neural Network (CNN) sensors that we call Sensor-Activated Feature Extraction One-Class Classification (SAFE-OCC). We show that this framework enables the safe use of computer vision sensors in process control architectures. Emergent control applications use CNN models to map visual data to a state signal that can be interpreted by the controller. Incorporating such sensors introduces a significant system operation vulnerability because CNN sensors can exhibit high prediction errors when exposed to novel (abnormal) visual data. Unfortunately, identifying such novelties in real-time is nontrivial. To address this issue, the SAFE-OCC framework leverages the convolutional blocks of the CNN to create an effective feature space to conduct novelty detection using a desired one-class classification technique. This approach engenders a feature space that directly corresponds to that used by the CNN sensor and avoids the need to derive an independent latent space. We demonstrate the effectiveness of SAFE-OCC via simulated control environments. • Present novelty detection framework for convolutional neural net (CNN) sensors. • CNN sensors enable incorporation of image data into control architectures. • Novelty framework enables robust predictions by CNN sensors. [ABSTRACT FROM AUTHOR]

Manually defined control limits remain a common strategy for quality control in manufacturing due to their ease of deployment on the shop floor compared to more advanced data analysis approaches. Despite their continued importance, there is no systematic method of defining these control limits. However, sub-optimal control limits can lead to undetected faults or cause unnecessary interruption to production. This manuscript presents an algorithm that systematizes this manual process into an efficient search task. We conceptualized the search task as a sequence of sub-problems that are based on the conventional steps taken by process experts when defining control limits. This algorithm can be integrated into an expert tool for shop floor personnel to automate the definition of control limits in annotated time series data. We demonstrate the efficacy of the control limits found by our algorithm by comparing them to those manually defined by process experts in real-world process data from the automotive industry. Furthermore, we show that our algorithm generalizes to traditional time series classification problems and achieves state-of-the-art performance on selected benchmark datasets. Our work is the first effort in automating the otherwise manual definition of control limits for fault detection. • Summary of existing time series classification methods and their practical feasibility for productive deployment. • Presentation of the algorithm including a brief introduction into the theoretical fundamentals of the underlying statistical techniques. • Experimental evaluation and validation of the algorithm on public datasets and real-world manufacturing data. • Publication of the source code to facilitate adoption by practitioners and researchers. [ABSTRACT FROM AUTHOR]

Many real-world control problems involve conflicting objectives. For different objectives, it is necessary to obtain Pareto optimal solution sets for each one. Over recent years, multi-objective reinforcement learning (MORL) has been extensively studied to solve this problem. However, the multi-objective optimization problem of complex continuous control processes still requires exploration. Soft proximal policy optimization algorithms have also been proposed to be combined with a hybrid weight-generation method for application to find the Pareto front approximation of the fed-batch fermentation process. This algorithm intends to initially find a single policy for the multi-objective reinforcement learning problem. A hybrid weight-generation method is then used to change the weights between different objectives to find a set of Pareto optimal solutions. In addition, we analyzed the mechanism of the fed-batch process, established the kinetic model, and designed the experimental environment based on the OpenAI Gym library. Experimental results showed that the proposed algorithm is effective and efficient in approaching the Pareto front of the fed-batch fermentation problem. • A multi-objective optimization problem of fed-batch fermentation process is explored. • Soft PPO algorithm combined with a hybrid weight-generation method is proposed. • Experiments show the efficiency in approaching the Pareto front of the fermentation problem. [ABSTRACT FROM AUTHOR]

This work addresses a new set of tuning rules for Sliding Mode Control (SMC) applied to second-order inverse-response plus variable dead time processes. Descriptions of the dynamics of inverse response systems and sliding-mode controllers are first provided. Then, we present the design process for the set of SMC tuning rules, using dimensional analysis, experimental design, optimization, and model reduction techniques for set-point tracking. The obtained expressions are simple, which eases the application of the proposed step-by-step systematic methodology. The performance and robustness of the proposed SMC tuning method are compared to other well-known controllers, and then applied to a Van de Vusse isothermal reactor, which is a complex nonlinear second-order inverse-response system that exhibits variable dead time. The performance achieved with the new SMC tuning method is good for set-point tracking, while exhibiting a good behavior for disturbance rejection, and can be applied considering a wide-range of parameters of the process' transfer function. [Display omitted] • New step-by-step Sliding Mode Control (SMC) tuning approach. • The method uses simple expressions to facilitate the application. • The new SMC tuning method is good for set-point tracking and disturbance rejection. • Application for a nonlinear second-order inverse-response system with variable dead time. [ABSTRACT FROM AUTHOR]

This paper aims to discuss and analyze the potentialities of Recurrent Neural Networks (RNN) in control design applications. The main families of RNN are considered, namely Neural Nonlinear AutoRegressive eXogenous, Echo State Networks, Long Short Term Memory, and Gated Recurrent Units. The goal is twofold. Firstly, to survey recent results concerning the training of RNN that enjoy Input-to-State Stability (ISS) and Incremental Input-to-State Stability (δ ISS) guarantees. Secondly, to discuss the issues that still hinder the widespread use of RNN for control, namely their robustness, verifiability, and interpretability. The former properties are related to the so-called generalization capabilities of the networks, i.e. their consistency with the underlying real plants, even in presence of unseen or perturbed input trajectories. The latter is instead related to the possibility of providing a clear formal connection between the RNN model and the plant. In this context, we illustrate how ISS and δ ISS represent a significant step towards the robustness and verifiability of the RNN models, while the requirement of interpretability paves the way to the use of physics-based networks. The design of model predictive controllers with RNN as plant's model is also briefly discussed. Lastly, some of the main topics of the paper are illustrated on a simulated chemical system. [ABSTRACT FROM AUTHOR]

A semi-autogenous grinding mill is simulated with gradient and non-gradient based extremum seeking controllers to maximize the mill performance using grind curves. Grind curves map the essential performance measures of a grinding mill to the mill load and rotational speed. The curves vary with the changes in the feed ore characteristic but show generic parabolic features with extremums. The extremum seeking controllers search along the unknown input–output map to steer the process towards an unknown optimum. In this study, a classical perturbation-based method, a time-varying parameter estimation-based method and the Nelder–Mead simplex method are employed as extremum seeking control (ESC) methods to search along the grind curves to either optimize the mill throughput or grind by means of manipulating the mill feed or rotational speed. The proposed extremum seeking controller could reduce the need for a plant operator to manually select the optimal operating conditions that maximize the performance measures of a grinding mill. Since the controller is agnostic to the process model, the grinding mill can be optimized without the need for a detailed process model. The simulated results show that the extremum seeking controllers steer the mill operating conditions toward the steady-state optimum and can be used to satisfy operational objectives. However, the slow grinding mill dynamics result in a long convergence rate when the initial conditions are far from the optimal operating conditions. • Model-free extremum seeking controllers are applied to a grinding mill. • Grinding mill performance (throughput or grind) is maximized using grind curves. • Grind curves map the essential performance measures of a grinding mill. • The need for a plant operator to manually select operating conditions is reduced. • The grinding mill can be optimized without the need for a detailed process model. [ABSTRACT FROM AUTHOR]

This work considers robust moisture content and activity control in a continuous convective drying for particulate bio-materials. The major process uncertainty is in the kinetics of the second (falling-rate) drying period. Robust feedback controllers in the multi-variable and de-central case are designed to guarantee desired product specifications. The advantages of the robust controllers are discussed in comparison with a conventional PI controller. The case study presented is the drying of baker's yeast in a fluidized bed dryer. The results show constant performances concerning the uncertainty of the drying kinetics and disturbance rejection at different operating points. • Robust multivariable controller design for convective drying. • Constant controller performance concerning various process uncertainties and operating points. • Baker's yeast in fluidized bed dryer as case study. [ABSTRACT FROM AUTHOR]

The purpose of this note is to correct an error in a previous paper and to illustrate the effect the selection of the basic distillation control structure has on the performance of additive and multiplicative feedforward control structures. [ABSTRACT FROM AUTHOR]

Feedforward control has been widely used for many years to improve the closedloop dynamic response of chemical processes to load disturbances compared to simply using feedback control. If the disturbance can be measured, a feedforward control element can immediately begin to move the appropriate manipulated variable to lessen the deviation in the controlled variable. Two types of feedforward structures have been used. Most academic studies consider a structure in which the output signals from the feedback controller and the feedforward controller are added to produce the manipulated variable in the process. Most applied studies consider a structure in which the ratio of the manipulate variable to the load disturbance is adjusted by the output signal of a feedback controller (multiplicative feedforward). The feedback controller adjusts the gain of the feedforward controller. The purpose of this paper is to present a quantitative comparison of these two approaches for two important examples: a distillation column and a chemical reactor. • Additive and multiplicative feedforward/feedback control structures are studied. • Common practice uses the more simple multiplicative ratio structure. • The additive structure is demonstrated to give more effective load rejection in two important examples (distillation and CSTR). [ABSTRACT FROM AUTHOR]

In this study, a generic framework has been developed for using exergy as a dynamic measure for energy efficiency of process control systems. The proposed framework begins by restating the exergy balance in terms of nonlinear state space; followed by partitioning of input variables. The exergy balance is then linearized, and added to the plant model to construct an extended state space model which has unknown number of non-physical states. The numerical method of sub space state space system identification is applied to determine the extended model matrices along with the system order. Laplace transform is then applied to convert the extended state space into a multi-input-single-output (MISO) formula for estimating the lost work. In the next step, an scalar index is suggested for representing the steady and transient parts of the dynamic lost work. The proposed index which is defined in the time domain is transformed into Laplace domain for consistency with the formulated MISO lost work. The suggested framework is tested for two case studies; (1) selection of the control variables for a CSTR, and (2) selection of the tray number for inferential control of a distillation column. The results confirm the generality of the proposed framework not only for exergetical optimization of control systems but also for disclosing the thermodynamically unrealizable structures. • A generic framework is developed for exergetically analysis of process control systems. • Lost work is the exergy parameter as a MISO relation. • The main contributions are input variables partitioning, state space linearization, system identification and introduction of an exergy index. • The framework can easily be applied in various fields of process systems control. [ABSTRACT FROM AUTHOR]

To inhibit the effect of disturbance in a system, feedforward control is usually combined with feedback loops to provide better control. The working environment for feedforward/feedback control system generally considers the commonly measured or unmeasured external disturbances. However, the Gaussian/non-Gaussian measurement noise from sensors during the measurement process for feedback loops will affects the performance of a control system. This paper proposes dynamic data reconciliation (DDR) filter combined with the feedforward/feedback control system to account for Gaussian/non-Gaussian measurement noises. This reduces the effect of disturbances and Gaussian/non-Gaussian measurement noises on performance. The DDR filter determines the optimal feedback signals using model prediction and measurements. Simulation examples of feedforward/feedback control systems show that the proposed DDR filter enhances performance and gives better filtering than a traditional digital filter. An application case study, which combines a DC–AC​ converter with DDR, is also tested to illustrate the effectiveness of the proposed DDR filter. • Effect of Gaussian/non-Gaussian disturbances on control performances is analyzed. • Dynamic data reconciliation (DDR) is developed to enhance control performances. • DDR corrects measurements with Gaussian/non-Gaussian disturbances and noises. • SISO, MIMO, and DC–AC converter control systems are used for verification. [ABSTRACT FROM AUTHOR]

This article addresses the problem of extremum seeking of a continuous-time dynamical system with a single input and a single output. A Super-Twisting-based optimization algorithm is proposed to compute the input that leads to the extremum value of an unknown convex objective function. Our optimization algorithm requires the input–output gradient of the system's response at steady state, which we compute throughout a Super-Twisting-based differentiator. Feasibility of the proposed extremum seeking strategy is demonstrated by two simulation examples. The first one is an example of interest in academy. The second one is a novel biohydrogen production process. [Display omitted] • An extremum seeking control strategy of a continuous-time dynamical system with a single input and a single output is proposed. • A gradient-based optimization algorithm is considered. • The input–output gradient is estimated via a differentiator. • The Super-Twisting algorithm is used both, as optimization algorithm and as differentiator. • Feasibility of the proposed extremum seeking strategy is demonstrated by two simulation examples. [ABSTRACT FROM AUTHOR]

Reinforcement learning (RL) for continuous state/action space systems has remained a challenge for nonlinear multivariate dynamical systems even at a simulation level. Implementing such schemes for real-time control is still of a difficulty and remains largely unanswered. In this study, several critical strategies for practical implementation of RL are developed, and a multivariable, multi-modal, hybrid three-tank (HTT) physical process is utilized to illustrate the proposed strategies. A successful real-time implementation of RL is reported. The first step is a meta-heuristic first principles model parameter optimization, where a custom pseudo random binary signal (PRBS) is used to obtain open-loop experimental data. This is followed by in silico asynchronous advantage actor–critic (A3C/A-A2C) based policy learning. In the second step, three different approaches (namely proximal learning, single trajectory learning, and multiple trajectory learning) are utilized to explore the state/action space. In the final step, online learning (A2C) using the best in silico policy on the real process using a socket connection is established. The extent of exploration (EoE, a measure of exploration) is proposed as a parameter for quantifying exploration of the state/action space. While the online sample efficiency of RL application is enhanced, a soft constraint based constrained learning is proposed and validated. With considerations of the proposed strategies, this work demonstrates the possibility of applying RL to solve practical control problems. • Reinforcement Learning for a nonlinear multivariate continuous state/action space. • First principles-based metaheuristic parameter tuning using plant data. • Simulation based asynchronous advantage actor–critic (A3C) policy learning. • Extent of exploration's role in improving on entropy-based policy generalization. • Online real time implementation of soft constrained policy updates discussed. [ABSTRACT FROM AUTHOR]

We focus on the Lyapunov-based output feedback control problem for a class of distributed parameter systems with spatiotemporal dynamics described by input-affine linear and semilinear dissipative partial differential equations (DPDEs). The control problem is addressed via model order reduction. Galerkin projection is applied to discretize the DPDE and derive low-dimensional reduced order models (ROMs). The empirical basis functions needed for this discretization are recursively computed using adaptive proper orthogonal decomposition (APOD). To update the basis functions during process operation, APOD needs measurements of the system state's complete profile (called snapshots) at revision times. This paper's main objective is to minimize the demand for snapshots from the spatially distributed sensors by the control structure while maintaining closed-loop stability and performance. A control Lyapunov function is defined, and its value is monitored as the system evolves. Only when the value violates a closed-loop stability threshold, snapshots are requested for a brief period by APOD after which the ROM is updated, and the controller is reconfigured. The proposed approach is applied to stabilize the Kuramoto–Sivashinsky equation. • Complex processes described by semilinear PDEs are controlled via model reduction. • Lyapunov and hybrid system theory is used to prove closed-loop stability. • The proof does not invoke the restrictive assumption of extensive sampling. • The control structure is updated just-in-time by a Lyapunov-based supervisory logic. • The requirement for on-line information is minimized using Lyapunov theory. [ABSTRACT FROM AUTHOR]

In this paper, considering the uncertainties of the controller and the model, a non-fragile H ∞ control law is robustly developed for time-delay equations with additive uncertainty. Hence, exploiting the linear matrix inequality (LMI), the robust compensator design is translated to an LMI one. In a continuous - time delayed model, the stabilization around equilibrium points and the rejection of the exogenous disturbance are discussed via the non-fragile robust H ∞ controller. Such an aim is realized by checking the feasibility of the LMI. Then, through solving an LMI, the controller's parameters would be found and updated directly. To show the superiority of the tactic in process control, the reported strategy is employed in a two-stage chemical reactor. The figurative results along with the quantitative findings consistently disclose the advantages of the addressed controller over conventional techniques. • Both the closed-loop stabilization and the disturbance rejection are considered to design a non-fragile H ∞ controller in uncertain continuous-time delayed models. • The non-fragile robust H ∞ compensation is converted to the feasibility of an LMI. Then the controller's gains would be found from the LMI. • The addressed H ∞ controller is applied in two-stage chemical reactors. The results reflect the merits of the proposed scheme over the similar one. [ABSTRACT FROM AUTHOR]

Two tuning techniques are proposed to design decentralized PID controllers for weakly coupled and general MIMO systems, respectively. Each SISO loop is designed separately, and the controller parameters are obtained as a solution of a linear programming optimization problem with constraints on the process stability margins. Despite the SISO approach, loop interactions are accounted for either by Gershgorin bands (non-iterative method) or an equivalent open-loop process (iterative method). The tuning results and performance from both methods are illustrated in four simulations of linear processes, and a laboratory-scale application in a Peltier process. Four applications contemplate closed-loop performance comparisons between the proposed techniques and techniques from the literature. One application illustrates the feasibility of the proposed iterative method, based on EOPs, in tuning decentralized PIDs for a 5 × 5 system. Moreover, an analysis of the effect of model uncertainty in the phase and gain margins of the closed-loop process is performed. • Two optimal design with robustness constraints PID tuning techniques are presented. • The proposed techniques are assessed in simulation examples and a laboratory-scale application. • One technique, utilizes constraints based on Gershgorin band to guarantee closed-loop stability. • Another recasts the process into an Equivalent Open-Loop form, resulting in an iterative algorithm. [ABSTRACT FROM AUTHOR]

A non-linear model is developed to capture the transient flow and pressure behaviour of methane-rich gas (MRG) in industrial pipelines for the use in simulation and control applications. Hyperbolic partial differential equations describe the pipe pressure and flow profiles, and composition analyses are used to develop the physical properties of the gas appearing in the pipe segment equations. The spectral element method (SEM) is used to numerically solve the spatial profiles within the pipeline, and a model verification is done on the SEM tuning parameter choices. The model is developed into a compact state–space description for improved usability. Furthermore, the developed model achieves good accuracy when validated on real process data of an industrial MRG network and can be used for simulation and model-based control applications. • This paper presents an accurate and numerically robust first-principles model which can be applied to methane rich gas and natural gas pipelines. • The model is formulated into a state–space representation which is ideal for simulation and control applications. • The model is validated using data from a real industrial MRG network. • This paper presents a thorough discussion around the modelling of the real industrial MRG network and how real-world modelling challenges were overcome. [ABSTRACT FROM AUTHOR]

In this paper, we present a computationally efficient economic NMPC formulation, where we propose to adaptively update the length of the prediction horizon in order to reduce the problem size. This is based on approximating an infinite horizon economic NMPC problem with a finite horizon optimal control problem with terminal region of attraction to the optimal equilibrium point. Using the nonlinear programming (NLP) sensitivity calculations, the minimum length of the prediction horizon required to reach this terminal region is determined. We show that the proposed adaptive horizon economic NMPC (AH-ENMPC) has comparable performance to standard economic NMPC (ENMPC). We also show that the proposed adaptive horizon economic NMPC framework is nominally stable. Two benchmark examples demonstrate that the proposed adaptive horizon economic NMPC provides similar performance as the standard economic NMPC with significantly less computation time. • Computationally efficient economic NMPC formulation with adaptive horizon length. • Nominal stability of the adaptive horizon economic NMPC is shown. • Two examples demonstrate that the proposed adaptive horizon economic NMPC provides similar performance as the standard economic NMPC with significantly less computation time. [ABSTRACT FROM AUTHOR]

Economic model predictive control (EMPC) is a model-based control scheme that integrates process control and economic optimization, which can potentially allow for time-varying operating policies to maximize economic performance. The manner in which an EMPC operates a process to optimize economics depends on the process dynamics, which are fixed by the process design. This raises the question of how process and EMPC designs interact. Works which have addressed process and control design interactions for steady-state operation have sought to simultaneously develop process designs and control law parameters to find the most profitable way to operate a process that is able to prevent process constraints from being violated and to optimize capital costs in the presence of disturbances. Because EMPC has the potential to operate a process in a transient fashion, this work first focuses on how EMPC and process design interact in the absence of disturbances. Using small-scale process examples, we seek to understand the fundamental nature of the interactions between EMPC and process design, including how these interactions can impact computational complexity of the controller and the design procedure. We subsequently utilize the insights gained to suggest controller design variables which might be considered as decision variables for a simultaneous process and control design problem when disturbances are considered. • Process design and economic model predictive control design interactions are examined. • Relationships between controller computation time and process design are discussed. • Simulation explorations using a continuous stirred tank reactor are performed. [ABSTRACT FROM AUTHOR]

The process stabilization is tackled from the perspective of a family of positive systems with multivariable positive control. This framework is consistent because each of the n state variables and m control inputs in chemical processes is in R +. The aim is to provide multivariable positive feedback control schemes which take values in the hyperbox U ∈ R + m ; while sufficient conditions are satisfied to ensure the stabilization process. Two approaches are discussed: the first is a sliding mode control based on hyperplanes intersecting a (n − m) -dimensional segment of the state space which comprises the equilibrium point; such scheme guarantees robust stability. The second is a continuous feedback stabilizer based on Lyapunov functions, which enhances convergence rate while stabilization is ensured. The process stabilization is shown with two examples: a heat exchanger and an isothermal continuous stirred tank reactor. Both examples illustrate the usefulness of the proposed framework of multivariable positive control. • Sufficient conditions are given to stabilize positive systems by positive control. • Multivariable positive sliding mode control for process stabilization. • Potential industrial applications of results are shown on heat exchangers. [ABSTRACT FROM AUTHOR]

• The system theoretical properties of the standard process model are analyzed. • The standard model is reduced to a first-order input/output model. • The unstable operating points can be stabilized by a linear control law. • The stabilization enables high process speeds and stoichiometric thin films. The aim of this paper is to present a control design method for nonlinear reactive sputter processes to increase the process speed and to facilitate the deposition of stoichiometric thin films. A physical model originating the fields of thin films science and vacuum science is analyzed to develop a control-oriented model, which describes the process with respect to a reactive gas as input and the pressure as output. The model parameters are identified by experiments. A control design method is proposed to stabilize the unstable operating points of the process. Experiments for the validation of the model and the controller are given. They show that reactive sputter processes can be modeled with sufficient accuracy by a first-order nonlinear differential equation and can be controlled by a linear feedback controller. [ABSTRACT FROM AUTHOR]

• An approximately bisimilar symbolic model is constructed for a safe changeover process. • An automatic controller is designed for safe changeover process for the first time. • The synthesis of the proposed controller in a finite-state space is very fast and flexible. • The error bounds of the proposed controller are adjustable as design parameters. • The effectiveness of the symbolic controller is investigated via numerical simulation. Many process control problems with complex qualitative specifications cannot be addressed via conventional control design methods. Examples of such specifications include logic specifications expressed in the design of start-up, shut-down, changeover, and emergency shutdown operating procedures. In recent years, it has been shown in the control systems and computer science communities that symbolic models provide convenient and powerful mechanisms to synthesize controllers enforcing such qualitative specifications. The use of symbolic models reduces the synthesis of the controllers to a fixed-point computation problem over a finite-state abstract system. In this paper, after explaining the notion of approximate bisimulation for incrementally globally asymptotically stable (δ -GAS) nonlinear control systems, the construction of approximately bisimilar symbolic models for such systems is presented. Then synthesis of time-optimal symbolic controller for this class of systems is performed based on results from the computer science literature. As a benchmark chemical process control problem, an approximately bisimilar symbolic model is constructed for a safe changeover process. Then a symbolic controller is designed and it is refined to a controller to be applied to the original process. Simulation results show the effectiveness of the symbolic controller. Although the construction of the symbolic model may be complex, the synthesis of the controller in a finite-state space is fast and most importantly the error bounds are adjustable as design parameters. [ABSTRACT FROM AUTHOR]

• The method includes deterministic knowledge into data regression. • The method predicts unburned calcium oxide in clinkering process. • The method predicts max temperature in new clinkering process based on microwaves. • The method is trained and validated using real data. Soft sensing is a monitoring technique for the indirect assessment of a target variable by means of direct measurements of others and the application of data mining on the historical log, as well as simplified models of the system. Due to technical and economic advantages respect to hardware sensing, soft sensing has been increasingly used in many scenarios, in particular within the process industry. Despite the literature being wide regarding the application of conventional regression techniques on data provided by the monitoring hardware, a systematic approach for supporting and improving data regression through the deterministic knowledge of the process is still missing. This contribution presents an innovative regression method based on first principle modeling. The method is introduced and tested under the case scenario of conventional clinker making process, for predicting the instantaneous fraction of the unburned calcium oxide. A second test case involves an innovative clinkering stage based on microwaves application. [ABSTRACT FROM AUTHOR]

This work is the second of a two-part study on oscillation detection in process industries. The first part, in short, extensively reviews and classifies the detection techniques. In this part, techniques are selected and evaluated on two industrial data sets, the first extracted from the literature and the second from a refinery. Common detection problems are pointed out aiming to drive future better-grounded researches on the topic. Based on the evaluation, features of interest of the detection methods, recommended test for new methods, and a new approach based on the verdict of multiple techniques are proposed. The work finalizes by pointing out research opportunities. [ABSTRACT FROM AUTHOR]

• Review of the works related to oscillation detection in process industry. • Discussions on the definition, types, strength, and causes of oscillation. • Evaluation of 10 oscillation detection methods on industrial data. • Discussions on features of interest, recommended tests, and a new detection approach. • Pros and cons of the methods to facilitate the choice of an appropriated one. Oscillatory control loop is a frequent problem in process industries. Its incidence reduces product uniformity and increases both energy consumption and raw material waste. These impacts strongly reduce plant profitability, making fundamental the removal of such oscillation. The first step in oscillation removal is its detection, which should preferably be performed automatically to enable the evaluation of the entire plant. A large number of these automatic techniques have been proposed over the last years. Thus, understanding and selecting a proper technique for the desired application is very time-consuming. This work is the first of a two-part study on oscillation detection in process industries. While the second part focuses on the evaluation of the detection techniques applied to industrial data, this part mainly includes an extensive review of the works on the topic and discussions on the definition, types, strength, and causes of oscillation. Also, the techniques are clustered into groups, and their pros and cons are confronted. [ABSTRACT FROM AUTHOR]

Highlights • Deep reinforcement learning is a versatile tool for control problems. • A controller is developed based on DDPG framework for a polymerization reactor. • The controller follows the MW trajectory by adjusting the monomer and initiator flow rates. • The resulting MWD from the DRL controller matches with the target distribution. Abstract Reinforcement learning is a branch of machine learning, where the machines gradually learn control behaviors via self-exploration of the environment. In this paper, we present a controller using deep reinforcement learning (DRL) with Deep Deterministic Policy Gradient (DDPG) for a non-linear semi-batch polymerization reaction. Several adaptations to apply DRL for chemical process control are addressed in this paper including the Markov state assumption, action boundaries and reward definition. This work illustrates that a DRL controller is capable of handling complicated control tasks for chemical processes with multiple inputs, non-linearity, large time delay and noise tolerance. The application of this AI-based framework, using DRL, is a promising direction in the field of chemical process control towards the goal of smart manufacturing. [ABSTRACT FROM AUTHOR]

Highlights • Show that the EMPC phenomenon of inventory creep can occur within a reaction system. • Show that EMPC performance degradation can occur abruptly with horizon size. • Develop a novel infinite horizon framework for the EMPC problem. • A framework to approximate the infinite horizon policy using dynamic programming. • Advocate ELOC as the appropriate basis to approximate infinite horizon EMPC policy. • Show that the approximate policy is insensitive to computational horizon size. Abstract In recent years, the notion of Economic Model Predictive Control (EMPC) has gained significant interest. Despite a marked improvement in economic performance, it has been shown that this performance will degrade substantially if implemented with a horizon that is not sufficiently large. In the current effort, it is shown that if applied to a particular reaction process, EMPC performance will abruptly collapse at a critical horizon size. To alleviate this issue, we develop an Infinite Horizon EMPC (IH-EMPC) formulation. While this IH-EMPC problem is computationally intractable, it does lead to an approximation of the optimal policy. The resulting Approximate IH-EMPC (AIH-EMPC) is identical to the original finite horizon EMPC, but includes a final cost term that represents the objective function from the finite horizon to infinity. With two example systems, a chemical reactor and a power system with energy storage, it is shown that the AIH-EMPC policy is virtually insensitive to its computational horizon size. [ABSTRACT FROM AUTHOR]

Big Data will revolutionize modern industry by improving process optimization, facilitating insight discovery, and improving decision-making. This big data revolution presents a multitude of possibilities and challenges in evolving from traditional batch processes to smart batch processes. This tremendous potential requires the ability to extract value from vast amounts of industrial process data. Using a new three-dimensional (3D) perspective of time, batch, and operational context, this paper explores smart batch processes with higher efficiency, greater profitability, and longer sustainability. First, we review the traditional one-dimensional (1D) perspective on batch processes and summarize the existing two-dimensional (2D) perspectives on batch processes, i.e., modeling, monitoring, control, and optimization methods. Based on those results, the spotlight will focus on how big data can be used to achieve smart batch processes using the 3D perspective. This will include detailed discussions of definitions and concepts, operational mechanisms, and the benefits and advantages of smart batch processes. For further implementation of the 3D perspective, we present several monitoring and control methodologies. Next, we analyze several challenges and issues in implementing smart batch processes in the era of Big Data. In conclusion, we provide both a novel viewpoint and encouragement for future research into batch process automation from the 3D perspective. [ABSTRACT FROM AUTHOR]

The model predictive static programming (MPSP) technique, which is extended recently to incorporate applicable state and control constraints, operates on the philosophy of nonlinear model predictive control (NMPC). However, it reduces the problem into a lower-dimensional problem of control variables alone, thereby enhancing computational efficiency significantly. Because of this, problems with larger dimensions and/or increased complexity can be solved using MPSP without changing the computational infrastructure. In this paper, the MPSP technique with applicable constraints is applied to two challenging control problems in the mineral processing industry: (i) a single-stage grinding mill circuit model, and (ii) a four-cell flotation circuit model. The results are compared with a conventional nonlinear MPC approach. Comparison studies show that constrained MPSP executes much faster than constrained MPC with similar/improved performance. Therefore, it can be considered a potential optimal control candidate for mineral processing plants. [ABSTRACT FROM AUTHOR]

Membrane distillation is a separation technology that can be powered with low-temperature solar energy and perform brine concentration treatments. However, the operations under changing radiation and salinity conditions make it necessary to use specific control and optimization strategies to maximize the MD system's performance. In this work, an online feedback optimization technique is proposed to address this issue in real-time. This strategy allows us to intrinsically manage constraints on the input and output of the process while minimizing its thermal energy consumption, which makes progress concerning previous approaches in the literature for MD systems. In addition, the online feedback optimization controller is combined with the least squares technique, which turns the approach into a model-free controller that helps to overcome the difficulty in modeling MD systems. The proposed strategy was tested in a trustworthy simulation scenario, using a validated model and real data from a facility located at Plataforma Solar de Almería (Spain). The results show its superior performance over a previous controller in the literature and non-optimal approaches, allowing us to reduce the thermal energy consumption by 0.5% and up to 6.7%, respectively, which has a direct impact on the cost of water treatment. On the whole, the proposed controller results in an attractive plug-and-play strategy with energy savings achievements for general use in commercial MD modules. • A model-free online feedback controller is proposed for commercial MD modules. • A moving window of data is used to identify time-varying gradient and sensitivity. • Least squares identification turns the approach in a model-free controller. • The model-free controller reaches almost the optimal solution in batch operations. • Constraints on the input and output of the process can be intrinsically handled. [ABSTRACT FROM AUTHOR]

A flotation circuit is controlled in simulation using an extremum seeking control (ESC) approach to keep the cells operating at the optimal operating point, as represented by peak air recovery. It is assumed that optimal performance is achieved at this operating point where the froth layer is stable, and the mineral recovery of the flotation cell is maximized. Two gradient-based ESCs, a classical perturbation-based ESC and a time-varying ESC, as well as a non-gradient-based direct search Nelder–Mead simplex ESC, are compared on the flotation circuit to steer the plant through an unknown static map towards the peak in air recovery. The three ESCs can respectively optimize the flotation circuit and find the peak air recovery operating point. The simplex ESC can converge quickly to the optimum but does not adapt to changing conditions. The gradient-based ESCs can track the time-varying peak air recovery operating point and adapt to an external disturbance. Although the three ESC methods are not dependent on a process model to optimize the plant, their convergence times are relatively slow. The ESCs are ideally suited for model-independent long-term automated optimization of a flotation circuit with a slow time-varying optimal operating point. • The recovery of a flotation circuit is optimized using extremum seeking control. • Gradient-based ESC can track the time-varying peak air recovery point. • Simplex ESC converges quickly to flotation air recovery optimum. • ESC is a model-free control approach for long-term flotation optimization. [ABSTRACT FROM AUTHOR]

Highlights • Efficient calculation of terminal regions for nonlinear MPC. • Developing an NMPC strategy with adaptive horizon lengths. • Asymptotic stability of the nominal control is shown. • Input to State Practical Stability is shown for the robust case. • Adaptive NMPC strategy is demonstrated on two challenging process example. Abstract We present a new method for adaptively updating nonlinear model predictive control (NMPC) horizon lengths online via nonlinear programming (NLP) sensitivity calculations. This approach depends on approximation of the infinite horizon problem via selection of terminal conditions, and therefore calculation of non-conservative terminal conditions is key. For this, we also present a new method for calculating terminal regions and costs based on the quasi-infinite horizon framework that extends to large-scale nonlinear systems. This is accomplished via bounds found through simulations under linear quadratic regulator (LQR) control. We show that the resulting controller is Input-to-State practically Stable (ISpS) with a stability constant that depends on the level of nonlinearity in the terminal region. Finally, we demonstrate this approach on a quad-tank system and a large-scale distillation application. Simulation results reveal that the proposed approach is able to achieve significant reduction in average computation time without much loss in the performance with reference to fixed horizon NMPC. [ABSTRACT FROM AUTHOR]

Highlights • A novel model-based rapid thermal processing (RTP) approach is proposed. • A detailed mathematical model of the heat-up process is presented. • An observer is designed to estimate the wafer's temperature profile. • A systematic procedure allows to avoid the effect of controller windup. Abstract The key issue in semiconductor rapid thermal processing (RTP), where silicon wafers are heated to desired temperatures, is to ensure uniform wafer temperature profiles. This in turn guarantees that the manufactured integrated circuits (IC), which are spread over the entire wafer, can be processed in an equal measure. In this study, a mathematical model capturing the physical behavior of the heating-up process is presented. It relies mainly on the well-known partial differential heat equation, governing the temperature distribution in a given region of a given material. Based on this, an observer as also a control law is proposed, which in combination turned out to be superior to conventional RTP strategies. The approach is validated in practice at a real world heating system with different wafer types. [ABSTRACT FROM AUTHOR]

Highlights • Generating Pareto curves to show the optimal trade-off between Robustness and Performance. • Find the optimal PI and PID controller for First-Order and Integrating Plus Delay Processes. • Investigating the difference between ideal and serial PID, and the impact of filtering. • Comprehensive comparison between optimal-, SIMC- and "improved" SIMC-tuning. Abstract The first-order plus delay process model with parameters k (gain), τ (time constant) and θ (delay) is the most used representation of process dynamics. This paper has three objectives. First, we derive optimal PI- and PID-settings for this process. Optimality is here defined as the minimum Integrated Absolute Error (IAE) to disturbances for a given robustness level. The robustness level, which is here defined as the sensitivity peak (M s), may be regarded as a tuning parameter. Second, we compare the optimal IAE-performance with the simple SIMC-rules, where the SIMC tuning parameter τ c is adjusted to get a given robustness. The "original" SIMC-rules give a PI-controller for a first-order with delay process, and we find that this SIMC PI-controller is close to the optimal PI-controller for most values of the process parameters (k , τ , θ). The only exception is for delay-dominant processes where the SIMC-rule gives a pure integrating controller. The third objective of this paper is to propose and study a very simple modification to the original SIMC-rule, which is to add a derivative time τ d = θ /3 (for the serial PID-form). This gives performance close to the IAE-optimal PID also for delay-dominant processes. We call this the "improved" SIMC-rule, but we put "improved" in quotes, because this controller requires more input usage, so in practice the original SIMC-rule, which gives a PI-controller, may be preferred. [ABSTRACT FROM AUTHOR]

Graphical abstract Highlights • Model predictive control and moving horizon estimation are combined. • The controller-observer pair computations are accelerated via Carleman approximation. • Jacobian and Hessian semianalytical expressions are provided to the search algorithm. • Advanced-step concept is used to further accelerate the output controller. Abstract This manuscript develops an algorithm that fuses Carleman moving horizon estimation (CMHE) and Carleman model predictive control (CMPC) together, to design an output feedback receding horizon controller. CMHE identifies the system states as the initial condition for CMPC to make optimal control decisions. The control decisions made by CMPC update the dynamic models used in CMHE to make more precise estimations. Modeling the nonlinear system with Carleman approximation, we estimate the system evolution for both CMHE and CMPC analytically. The Gradient vectors and Hessian matrices are then provided to facilitate the optimizations. To further reduce real-time computation, we adapt the advanced-step NMHE and advanced-step NMPC concepts to our CMHE/CMPC pair to develop an asCMHE/asCMPC pair. It pre-estimates the states and pre-designs the manipulated input sequence one step in advance with analytical models, and then it updates the estimation and control decisions almost in the real-time with pre-calculated analytical sensitivities. A nonlinear CSTR is studied as the illustration example. With CMHE/CMPC pair, the computational time is decreased to one order of magnitude smaller than standard nonlinear MHE and nonlinear MPC. With asCMHE/asCMPC pair, the real-time estimation and control decisions takes a negligible amount of wall-clock time. [ABSTRACT FROM AUTHOR]

Highlights • Continuous convective drying of particulate materials is considered. • Parametric uncertainties in the second drying period are addressed. • Feedback controller providing robust stability and performance are designed. • Application is exemplified for drying of baker's yeast pellets in fluidised beds. • Very good performance in view of kinetic uncertainty and operating point. Abstract This work considers robust control of continuous convective drying of particulate solids where the major process uncertainty lies in the kinetics of the second (falling-rate) drying period. Using the concept of normalised drying curve, for arbitrary shape of the curve, the steady state particle moisture content is derived. Applying appropriate uncertainty models, a robust feedback controller is designed to guarantee desired product moisture. Advantages of this approach are discussed in comparison to convential PI and LQ-optimal control. The approach is exemplified for the case of fluidised bed drying of baker's yeast particles. Results show very good performance with respect to parametric uncertainty and disturbance rejection even at non-nominal steady-states. [ABSTRACT FROM AUTHOR]

Manufacture of ultra large-scale integrated circuits involves accurate control of a challenging nonlinear Rapid Thermal Processing (RTP) system. Precise control of temperature profile and rapid ramp-up and ramp-down rates demanded by a RTP system cannot be achieved with conventional control strategies due to nonlinear and multi time-scale effects. In this paper the control of a RTP system is reformulated as an optimal multi-step sequential decision problem using the framework of finite horizon Markov decision processes and solved using a Reinforcement Learning (RL) algorithm. Three increasingly complex RL based control strategies are explored and compared with the existing state-of-the-art approach for controlling RTPs. Simulation results indicate that the approach proposed in this paper achieves superior control of the temperature profile and ramp-up and ramp-down rates for the RTP system. [ABSTRACT FROM AUTHOR]

One-class classification plays a key role in the detection of outliers and abnormalities. Recently, several attempts have been made to extend the application of one-class classification techniques to statistical process control problems, where many of these one-class classification-based approaches have used a support vector data description algorithm. The monitoring statistics for a support vector data description-based control chart are sufficiently defined. However, the control limits are not obvious because the procedure used to derive the control limit does not include a method for controlling the false alarm rate (i.e., Type I error rate), which clearly limits its use in process monitoring. In this study, we propose a new multivariate control chart based on a technique for optimal false alarm-controlled support vector data description, which minimizes the radius of a spherically shaped boundary so that it includes the normal data that are equal to an assigned constant value. By modifying this constant value, users can precisely control the proportion of abnormal data determined by the spherically shaped boundary, which equals the expected Type I error rate. We demonstrated the usefulness of the proposed charts in experiments with simulated data and real process data based on a thin film transistor–liquid crystal display. [ABSTRACT FROM AUTHOR]

The general opinion on Åström et al.'s modification of the Smith predictor (which allows for plants with an integrator) appears to be quite poor: "good performance, yet (far) too complicated". This opinion, however, is based on missing insight and – sadly enough – a printing error in the original publication. The present article argues the case for the Smith-Åström predictor to rehabilitate its general perception by (1) explaining its working principle (focusing more on practical than on theoretical insights) and (2) comparing it to other modifications of the Smith predictor. As in the original publication, the considered plant model is an integrator with dead time for which it is shown that some subsequent modifications yield the same disturbance-to-output behaviour as the Smith-Åström predictor, yet in contrast to some of them, the latter also provides an explicit disturbance estimate. To complete the plea, the original results are extended by addressing practical issues such as parameter tuning, actuator constraints, model uncertainties (including unmodelled dynamics), and disturbance reconstruction, all of which are illustrated by means of simulation; furthermore, the design procedure is given in a more general fashion to allow for different (and possibly unstable) plants. It is shown that the Smith-Åström predictor yields good results and is easy to use! • Detailed explanation of the working principle of the modified Smith predictor by Åström et al. • Tuning rules, actuator constraints, and model uncertainties are covered. • An explicit disturbance estimate is available with the Smith-Åström predictor. • Practical relevance of the integrator with long dead time is discussed. • Input-output equivalence of different Smith predictor modifications (including Åström et al.'s) is shown for the integrator. [ABSTRACT FROM AUTHOR]

This paper presents the design and application of fractional single-input–single-output (SISO) controllers to a grinding mill circuit, which is a multiple-input–multiple-output (MIMO) process. Two kinds of controllers are presented: fractional order proportional-integral (FOPI) controllers, and a combination of FOPI and fractional order model reference adaptive controllers (FOMRAC). The parameters of the controller are tuned using off-line particle swarm optimization. In the presence of disturbances and process noise, the SISO fractional controllers achieve similar or better performance compared to linear model predictive control (LMPC). [ABSTRACT FROM AUTHOR]

A non-linear observer model of a semi-autogenous grinding mill is developed. The observer model distinguishes between the volumetric hold-up of water, solids, and the grinding media in the mill. Solids refer to all ore small enough to discharge through the end-discharge grate, and grinding media refers to the rocks and steel balls. The rocks are all ore too large to discharge from the mill. The observer model uses the accumulation rate of solids and the mill's discharge rate as parameters. It is shown that with mill discharge flow-rate, discharge density, and volumetric hold-up measurements, the model states and parameters are linearly observable. Although instrumentation at the mill discharge is not yet included in industrial circuits because of space restrictions, this study motivates the benefits to be gained from including such instrumentation. An extended Kalman filter is applied in simulation to estimate the model states and parameters from data generated by a semi-autogenous mill simulation model from literature. Results indicate that if sufficiently accurate measurements are available, especially at the discharge of the mill, it is possible to reliably estimate grinding media, solids and water hold-ups within the mill. Such an observer can be used as part of an advanced process control strategy. [ABSTRACT FROM AUTHOR]

This paper presents a nonlinear observer-based Lyapunov control for a membrane distillation (MD) process. The control considers the inlet temperatures of the feed and the permeate solutions as inputs, transforming it to boundary control process, and seeks to maintain the temperature difference along the membrane boundaries around a sufficient level to promote water production. MD process is modeled with advection diffusion equation model in two dimensions, where the diffusion and convection heat transfer mechanisms are best described. Model analysis, effective order reduction and parameters physical interpretation, are provided. Moreover, a nonlinear observer has been designed to provide the control with estimates of the temperature evolution at each time instant. In addition, physical constraints are imposed on the control to have an acceptable range of feasible inputs, and consequently, better energy consumption. Numerical simulations for the complete process with real membrane parameter values are provided, in addition to detailed explanations for the role of the controller and the observer. [ABSTRACT FROM AUTHOR]