Duration: One-day including time for hands-on experience in the use of alarm analysis and design software.

Workshop objectives: Overview of new quantitative methods for analysis and design of industrial alarm systems as an emerging research area, targeting academic researchers and industrial practitioners involved in alarm management.

Abstract: Large industrial plants contain thousands of sensors and actuators, and hundreds of control loops. All components of a plant are susceptible to faults that can disrupt the normal operation of the control system and may result in unsatisfactory performance, instability, failure or even dangerous situations. Due to the increasing complexity of process control systems and the growing demands for quality, cost efficiency and safety, it is important that faults be promptly diagnosed and appropriate remedies be applied. Detection of a fault in the plant, usually results in an alarm being raised to inform the operator about the abnormality in the process.
Recently updated industry standards (EEMUA 191 and ANSI/ISA 18.2) suggest that an operator should not receive more than six alarms per hour during the normal operation of the plant. This is, however, rarely the case in practice. Various studies show that the number of alarms each operator receives is far more than the standard (tens, hundreds or even thousands of alarms per hour, depending on the industry and their alarm generation policy); a majority of these are false or nuisance alarms. Too many false/nuisance alarms (alarm flooding) distract the operator from operating the plant and can bury important alarms. This might reach to a point where the operator no longer trusts the alarms or even shuts down the whole alarm monitoring system. Industries reportedly loose millions of dollars every year due to alarm problems. As a result, there has recently been an increasing interest in industry to address this issue and seek remedies to reduce the number of false and nuisance alarms.
Although the area of fault detection and identification is very well established in academia, design and analysis of alarm systems, as the next logical step after fault detection, has not received much attention. Industry demand for a reliable and robust alarm system is at its peak and the lack of strong theoretical support by academia has put a lot of challenges in front of industry. We believe that knowledge and expertise of control systems experts give them a unique opportunity to address alarm systems and solve problems and challenges faced by industry.
Technical material for this workshop will be complemented with several industrial case studies. Participants will be provided with software for the analysis and design of alarm systems. All attendees are encouraged to bring their laptops to obtain “hands-on” experience in analyzing process and alarm data for alarm analysis, rationalization, and design.


Agenda: Topics to be covered in this workshop will be divided into four sessions (each lasting one and half hours). Technical presentations will be complemented with industrial case studies. Each of the following 4 sessions will also include additional time for “hands-on” use of the alarm analysis and design software developed at the U of Alberta (AMtool).
This workshop is intended to introduce the following topics to the control community in industry as well as academia. Details of topics for this workshop appear below:

• Introduction to Alarm Systems
  • Alarm systems and process monitoring; historical incidents; false and nuisance alarms; alarm standards; alarm systems life cycle; alarm rationalization; current status of alarm systems; how to deal with information resources (alarm data, process data, connectivity information).
  • Introduction to the AMtool followed by hands-on experience in generating alarm overview reports using this software.
• Alarm Data: Representation, Analysis and Visualization
  • Introduction to alarm data; graphical representation; high density alarm plots; alarm correlation color map; parallel coordinate plots; run-length distributions; merged plots of process and alarm data. Detection of chattering and oscillating alarms.
  • Demonstration and hands-on use of the AMtool to generate various visualizations of process and alarm data.
• Process Data: Optimal Alarm Design
  • Introduction to process data; Alarm design trade-offs: how to minimize false and missed alarms; ROC curves; Filtering, deadbands, delay-timers; Optimal design framework; Dealing with the tradeoff between latency (delays) and accuracy.
  • Alarm processing and trip point design in a univariate framework with examples.
  • Advantages of alarm design in a multivariate framework in terms of fewer false and missed alarms.
  • Industrials case studies to illustrate alarm design procedures for univariate and multivariate processes.
• Connectivity Information: Root Cause Analysis and Alarm Flood Management
  • Introduction to process connectivity information; representation of connectivity information; signed directed graphs; ontological models, process modeling and alarm design; root cause determination of process faults using causality analysis; root cause determination of plant-wide oscillations; introduction to alarm floods; analysis of alarm flood patterns.
  • Industrial case studies to illustrate root cause analysis and of alarm flood pattern analysis.

Duration:
This half-day workshop will consist of two parts: a concise overview of the theory of polynomial chaos, and the application of the polynomial chaos framework to several model-based estimation, fault diagnosis, and control problems for uncertain, nonlinear systems.

Workshop objectives:
The workshop is designed for researchers with a basic knowledge of model-based estimation and control who intend to gain an overview of the implications of system uncertainties for estimation, fault diagnosis, and control of complex chemical and biological systems.

Abstract:
Uncertainties are ubiquitous in complex chemical and biological systems. System uncertainties typically arise from measurement noise, parametric uncertainties, and exogenous disturbances. Systematic consideration of uncertainties in model-based estimation and control of complex systems is a particularly challenging problem, and has been the subject of extensive research in the systems and control community. Despite the significant progress made, the application of the commonly used uncertainty characterization methods (such as deterministic set-based approaches e.g. [Kieffer and Walter, 2011; Malan et al., 1997; Petersen and Tempo, 2013 ] and sampling-based methods e.g., [Vidyasagar, 2000]) suffers from two major limitations: (i) the methods often rely on restrictive assumptions pertaining to the complexity of dynamics and uncertainty descriptions, which are typically not realistic for the intrinsically complex dynamics, and/or (ii) the methods are often computationally prohibitive for real-time estimation and control of complex systems.
Polynomial chaos is a potentially promising tool for uncertainty characterization and propagation through nonlinear dynamical systems with probabilistic uncertainties that possess arbitrary probability distributions. Rooted in the pioneering work of Wiener [Wiener, 1938], polynomial chaos provides a computationally tractable spectral framework for uncertainty propagation by replacing the implicit mappings between the uncertain variables/parameters and the system states with an expansion of orthogonal polynomials [Ghanem and Spanos, 1991; Xiu and Karniadakis, 2002]. This allows efficient computation of uncertain variables’ statistics using the expansion coefficients.
This workshop is intended to present the promises and challenges of the polynomial chaos framework for estimation and control of a general class of complex systems to the process systems engineering and control community. After laying the theoretical foundation of the polynomial chaos framework, a particular emphasis will be devoted to the application of the framework to stochastic model predictive control, experiment design for model discrimination and parameter estimation in the presence of probabilistic uncertainties, and fault diagnosis and isolation in stochastic settings. The use of the polynomial chaos framework will be demonstrated using several real-world biological, chemical, and pharmaceutical processes. A recently developed polynomial chaos toolbox running under Matlab will be presented to discuss the implementation aspects of the framework.

Agenda:
This half-day workshop will consist of two parts. The first part will provide a concise overview of the theory of polynomial chaos. In addition, a Matlab-based polynomial chaos toolbox will be introduced. The toolbox will be used in the subsequent talks to demonstrate the use of the polynomial chaos framework. In the second part, the application of the polynomial chaos framework to several model-based estimation, fault diagnosis, and control problems for uncertain, nonlinear systems will be presented. The estimation and control approaches will be illustrated using real-world applications. In the different talks, an introduction to the estimation/control approach as well as details and examples for its application will be given. Open research problems will be discussed in the workshop.

• The generalized polynomial chaos framework
  • The Wiener-Askey polynomial chaos framework for uncertainty propagation will be presented. This is followed by the introduction of the Galerkin-projection and collocation methods for determining the polynomial chaos expansion coefficients. The implications of the different settings of the polynomial chaos framework will be discussed.
• Introduction to the polynomial chaos toolbox
  • The main components and commands of the Matlab-based toolbox will be presented. In the subsequent talks (2.1 – 2.4), codes and examples will be provided illustrating the application of the toolbox for different model-based estimation and control tasks.
• Input design for optimal experimental design and active fault diagnosis
  • Building reliable dynamic models is important for model-based design, optimization, and control. Optimal experiment design methods will be presented for model discrimination [Streif et al., 2014] and parameter estimation [Mesbah and Streif, 2015] in the presence of time-invariant probabilistic uncertainties and measurement noise. The model discrimination approach relies on maximizing the dissimilarity of the probability distribution of outputs of candidate model structures. On the other hand, in the parameter estimation approach the input signal is designed to maximize some metric measure of the information content of the data, while minimizing variance of the data. The optimal experiment design approaches will be demonstrated for discrimination of biochemical models, and parameter estimation for a biological signaling model.
  • Reliable fault diagnosis in stochastic setting is crucial for safety and availability of high-performance systems. A probabilistic model-based approach will be presented for the design of auxiliary input signals that enhance fault diagnosability by separation of multiple models pertaining to nominal and faulty system operations in the presence of probabilistic uncertainties [Mesbah et al., 2014]. Active fault diagnosis will be demonstrated for diagnosis of multiple faults in a three-tank system.
• Stochastic MPC with chance constraints
  • A general receding-horizon stochastic optimal control approach will be presented. The control approach enables shaping the probability distribution of the system states, while incorporating chance constraints in the control approach to ensure the satisfaction of state constraints in the presence of uncertainties. SMPC control formulations are particularly advantageous when the control cost function is asymmetric and the high-performance system operation is realized in the vicinity of constraint in a stochastic setting. The stability aspects of the stochastic MPC approach will be discussed [Mesbah and Streif, 2015]. The performance of the stochastic control approach will be demonstrated for plant-wide control of a high-dimensional continuous pharmaceutical pilot plant with nearly 8,000 state variables [Paulson et al., 2014], and stochastic nonlinear MPC of a batch reactor [Streif et al., 2014]. In addition, output feedback stochastic control of batch polymorphic crystallization will be demonstrated [Mesbah et al., 2014; Mesbah et al., 2014].
• Dual identification and MPC
  • Dual control is concerned with the question of how to best control an uncertain system when assuming that future data can resolve uncertainty. To this end, an active step is taken to learn and reduce the model uncertainty, or to discriminate between competing model hypotheses such that future control decisions are made based on better understanding of the system. An MPC approach to dual control will be presented in which a modified optimal control problem is solved while a probabilistic model similarity measure is maximized to facilitate model discrimination.

Duration:
This half-day short course presents strategies for the optimization of differential-algebraic systems aided by large-scale nonlinear programming (NLP) algorithms.

Abstract:
This workshop will place special emphasis on problems encountered in process engineering and process control. The lectures will cover the basics of NLP algorithms, extensions to large-scale problems, formulation and solution of dynamic optimization problems for off-line applications and the extension of these strategies for on-line applications including nonlinear model predictive control and real-time optimization. Each set of topics will be illustrated with practical examples that illustrate the behaviour of these methods and performance of the algorithms. Details of each area are given below.

Agenda:

• Nonlinear Programming
  • Newton-based methods
  • Optimality conditions
  • Classes of Nonlinear programming methods
  • Large-scale Extensions
  • Available Optimization Methods
  • Modelling Environments
• Dynamic Optimization
  • Concepts, Background and Approaches
  • Simultaneous Approach, Analysis and Applications
  • Off-line Case Studies
• On-line Optimization
  • Optimization Formulations for Nonlinear Model Predictive Control (NMPC) and Moving Horizon Estimation (MHE)
  • Implementation and Computational Delay
  • NLP Sensitivity
  • NMPC Stability and Robustness
  • NMPC and MHE Strategies
  • On-line Case Studies

Duration:
This full-day workshop presents the result of 5-year research in time series analysis and stochastic control theory of AuLac Technologies Inc., www.aulactechnologies.com.

Workshop objectives:
This workshop will be of interest to: control/process control engineers, students and young instructors who want to consolidate their knowledge in this area, digital signal processing engineers, and time series analysts.

Abstract:
The purpose of a control professional to take a course in control theory is to be able to design a controller for his use. There are different controllers for different purposes, and there are different design methodologies for them. Since the control is usually through time, analysis and design should have time as the pivotal variable.
The control strategy to minimize the variance of the controlled variable ARMA time series with a constraint on the variance of the control variable, of a discrete feedback control system, gives a linear quadratic Gaussian control algorithm. For H-infinity control, the least sensitive controller gives the controlled variable a noninvertible ARMA time series: a time series with a at spectrum. Model predictive control requires prediction of a future value of the controlled variable ARMA time series. The ARIMA time series holds an important position in stochastic control theory. A fundamental knowledge of the ARIMA time series and its analysis is, therefore, an essential knowledge for a control professional, engineer or instructor, to have for his or her control design.
This full-day workshop presents the result of 5-year research in time series analysis and stochastic control theory of AuLac Technologies Inc., www.aulactechnologies.com. After attending the workshop, an attendee will have the knowledge to analyze and design discrete controllers and digital filters for linear systems with formulas of moments and spectra. This knowledge will be supplemented with the knowledge of periodic and aperiodic time series, prediction and (Kalman) filtering as well as fundamental statistics. An attendee will know the industrial Dahlin, Vogel-Edgar, IMC and also the PID deterministic tracking controllers. For the stochastic regulating controllers, the workshop will discuss the one-step, N-step and infinite-step optimal controllers, model predictive as well as the H-infinity controllers.
The lecture notes will be extracted from the two textbooks: “The ARIMA and VARIMA Time Series: Their Modelings, Analyses and Applications” and “Optimal Discrete Control Theory: The Rational Function Structure Model”, which come with the workshop. Demonstrations will be running MATLAB m-files. Each attendee should bring with him or her a notebook (laptop) with MATLAB software, or its clone like OCTAVE, installed on it for the demonstrations.

Agenda:
Morning: Time series literature

• Basic background
  • The random variable, probability and statistics
  • The expectation operator, moments and cumulants
  • The discrete time series: periodic, aperiodic and ARIMA
  • The model and modeling of an ARIMA/VARIMA time series
• Analyses: Moments and spectra
  • The Autocovariance (crosscovariance) generating function and the autocovariances (crosscovariances).
  • The Autospectrum and the Cross-spectra.
• Applications: Forecast, prediction, controller design and digital filter spectra
  • Forecasts: Forecasts of an ARMA and a VARMA Time Series
  • Prediction, filtering and the Kalman filters
  • Stochastic Control Theory: The Box-Jenkins model and its controllers
  • Digital Filter: Low-pass, high-pass, band-pass and all-pass filters

Afternoon: Discrete SISO controllers design

• Discrete linear control systems and time series
  • Time series models and controller types
  • Deterministic tracking control systems
  • Stochastic regulating control systems
• Designing deterministic tracking controllers
  • The minimal prototype PID and the dead-beat controllers
  • The Dahlin, Vogel-Edgar and IMC controllers
  • The set point-model tracking controllers: 1-DOF and 2.5 DOF controllers
• Designing linear quadratic stochastic controllers
  • The minimum variance controller
  • The one-step optimal controller
  • The N-step optimal controllers
  • The infinite-step optimal controller
• Designing H-infinity controllers
  • Signal and system norms
  • The sensitivity functions
  • The small gain theorem
  • The least sensitivity controller

Duration:
One day including time for hands-on exercises with PLS_Toolbox and Solo software.

Workshop Objectives:
The goal of the workshop is to familiarize the attendees with modeling methods that are useful for investigating and describing data sets with unconventional structures. Attendees will also learn how to use modeling software to explore these data sets and make use of them for data discovery and process monitoring purposes.

Abstract:
Most multivariate analysis methods, including Principal Components Analysis (PCA) and Partial Least Squares (PLS) regression, are intended for use with 2-way data, i.e. data that fits into conventional two dimensional tables. However, many analytical and process data sets do not fit the 2-way paradigm. This includes batch and semi-batch process data, but also a wide variety of situations where three dimensional data cubes are associated with a vector or matrix of data.
This workshop considers a number of model forms that can be used on data sets that do not fit the conventional 2-way approaches, e.g. PCA or PLS, but also the multi-way approaches, such as Parallel Factor Analysis (PARAFAC) and Multi-way (unfold) PCA (MPCA). This includes instances where data is in distinctly different blocks that share a common mode, in which case the multi-block variants of PCA and PLS may be appropriate. Another case is when the data is semi-batch, such as where processes are run for periods of time and then “reset” e.g. when catalysts are regenerated or the process equipment is cleaned. In these cases tools such as Simultaneous Components Analysis (SCA) or Multi-level SCA (MLSCA) may be used to understand the difference between runs and the variation within runs. Multi-level PLS variants are also potentially useful. Finally, instances where data sets consisting of blocks with different numbers of modes must be fused are considered. For example a 3-way data set might share a mode with a 2-way data set such as when a number of batch data records must be related to multiple quality parameters. In these instances models based on coupled matrix and tensor factorizations (CMTF) may be applied.
Attendees will work on hands-on exercises using PLS_Toolbox or Solo software.

Agenda:

1. Introduction
   • Definition of multi-block, multi-set and multi-level
   • Goals of Data Fusion
2. Review of Multivariate and Multi-way Models
   • Principal Components Analysis (PCA)
   • Partial Least Squares (PLS) Regression
   • Multi-way (unfold) PCA and PLS
   • Parallel Factor Analysis (PARAFAC)
   • Multi-way PLS (N-PLS) – Introduction to PLS_Toolbox/Solo
   • Hands-on exercises
3. Multi-set and Multi-Level Models
   • Simultaneous Components Analysis (SCA)
   • ANOVA SCA (ASCA)
   • Multi-level SCA (MLSCA)
   • Multi-level PLS
   • Hands-on exercises
4. Data Fusion Models
   • Coupled Matrix-Tensor Factorizations (CMTF)
   • Approaches for identifying CMTF models
   • Alternating Least Squares (ALS)
   • Direct optimization
   • Hands-on exercises