The workshop will be hands-on with participants developing their own models and control solutions. Participants must bring their own laptops, but will be provided with the necessary software.

On day one local industry and academic experts will introduce popular modern data analytics methods. Participants will then use these methods to infer a key performance measure of a simulated milling circuit.

On day two participants will be introduced to model predictive control (MPC). They will use their inferred performance measure to implement an MPC controller to stabilize and (or) optimize the milling circuit.

The workshop will be arranged such that participants can attend only day 1 or day2.

Date

26 – 27 August 2019

Pricing

SACAC member one day of workshop – R2,500

SACAC member both days of workshop – R4,200
Non SACAC member one day of workshop – R3,000
Non SACAC member both days of workshop – R5,000
Student one day of workshop – R2,200
Student both days of workshop – R3,500

Venue

The Stellenbosch Institute for Advanced Study (STIAS) is located on the Mostersdrift farm, nestled in the centre of Stellenbosch.

A “Creative Space For The Mind” has been created through sustainable architecture and beautiful, tranquil surroundings. Intellectual leaders and top researchers find ingenious and maintainable solutions to the issues that have arisen in the world, the continent of African and, in particular, South Africa.

 

Downloads

Call For Papers

MMM2019 Workshop Flyer

Program – Day One – Data Analytics for Soft Sensing

Introduction to analytics for soft sensing (Chris Aldrich)

Soft sensors, virtual sensors or state observers from a control theory perspective, are commonly used in the resource industries for combined processing of high-dimensional signals. This can be used to infer quantities from these signals that would otherwise have had to be measured physically. As a consequence, soft sensors are used in data fusion, fault diagnosis, as well as various control applications.

The proliferation of large volumes of data connected with mining systems and process plants is opening up ever-widening scope for the development of soft sensors. Even so, the soft sensing components of smart plant sensing networks are significantly impacted by the by 3Vs of Big Data, i.e. volume, velocity and veracity. The growing heterogeneity of sensors and sensing platforms can lead to problems related to the trustworthiness of sensed data and this needs to be addressed prior to model building. Anomaly detection and recommender systems to filter malfunctioning sensors or external interference in sensor readings can be used for this purpose.

Increased numbers of connected sensing devices can generate global data traffic that can lead to a networking bottleneck between sensors and the cloud. This problem can be alleviated by localised processing of the data, or by other methods to boost processing turnover in otherwise computationally intensive algorithms, such as deep learning neural networks that are designed to deal with massive volumes of data.

Data preparation best practices (Danielle Winter)

Real world data isn’t perfect – more often than not, sampled sensor data is riddled with missing values, noise and outliers. This data may also come from disparate sources with different sampling rates. Before an accurate model can be designed for soft sensing, the raw data must be cleansed, processed and aligned. This part of the workshop is a practical overview of the best practice techniques for getting data into the right format for creating soft sensor models.

Model selection and training (John McCoy)

With the increasing availability of libraries and packages which make it easy to apply a wide range of predictive models, it has become very important that the user develop an understanding of the different model types, how to choose between them, how to train them, and how to evaluate them to choose the most useful model for the problem at hand. This talk will cover: common supervised model types and hyperparameters, model training with cross validation or a validation set, using unsupervised models for feature engineering, and hyperparameter tuning.

Model maintenance best practices (Danielle Winter)

The behaviour of systems changes over time depending on environmental factors and changes in inputs. When developing soft sensing models, those models typically capture the system dynamics at a point in time. Over longer periods, the model will drift from the true response of the system, making predictions less accurate. This portion of the workshop covers the principles behind detecting system drift and calibrating and maintaining models for retaining a true representation of the system.

Program – Day Two – Model Predictive Control

An introduction to grinding circuit control (John Burchell)

By optimising the control of a grinding circuit a business unit can achieve significant energy savings, increase its throughput potential, and experience improved product grade and recoveries. This session will present a high level overview of a grinding circuit’s operation, the primary factors that influence its performance, as well as some of the most common strategies employed for its control.

Dynamic modelling and simulation (Jason Miskin)

Dynamic modelling and simulation have become increasingly popular due to both the availability of computing power as well as software that provide a framework for execution. This technology involves mimicking time-dependent real-life systems in simulation, with the purpose of understanding and inspecting interactions between complex subsystems. These insights are often used to make design decisions and improve or monitor existing systems. This talk will cover a step-by-step implementation of a system in Simulink. Practical considerations and best practices
will be discussed. A non-linear dynamic model of a milling circuit will then be exposed through the OPC protocol, which will act as a real-life system, in preparation for MPC prototyping.

Linear MPC (Kevin Brooks)

Linear dynamic models are the basis of the vast majority of industrial APC. The generation of these models, and their use in multivariable control will be discussed. Expected benefits are described.

Nonlinear MPC (Loutjie Coetzee)

Nonlinear APC is relatively new and not a widely used tool, due primarily to the complexity of defining and maintaining nonlinear models, but also the benefit for continuous processes running at a fixed operating point is negligible. There may, however, be scope to start with a linear approximation of a process and only add nonlinear behaviour where it will add value, such as time-varying nonlinear dynamics, optimisation that moves the process across nonlinear dynamics,
discontinuities by switching in or switching out process units or even nonlinear economic objectives. Nonlinear APC is well suited as an enabler for recent trends, such as time-varying economic objectives (eMPC) and plant-wide control.

 

Speakers

Prof Chris Aldrich

Professor Chris Aldrich holds a Chair in Process Systems Engineering in the Western Australian School of Mines, in Perth, Australia. His core expertise is in process automation in the mining and metallurgical industries, with a strong focus on artificial intelligence and machine learning. He was the recipient of a number of early and mid-career national awards based on his research in this area, including the President’s Award of the Foundation of Research and Development (FRD) in South Africa, as well as the British Association (S2A3) Silver Medal.
In addition, he is a Fellow of the South African Academy of Engineers and serves on the editorial boards of several academic journals.

Dr Loutjie Coetzee

Loutjie Coetzee (Mintek) is a research and development engineer at Mintek. He specialises in developing the model based control toolbox for Mintek’s StarCS control platform and have been involved in commissioning a number of APC systems for the mining industry, in the fields of milling, thickener and leaching control.

Dr Kevin Brooks

Kevin Brooks’ (BluESP) control career began with Honeywell Hi-Spec Solutions, in the area of APC and optimization. He worked on a number of processes, mainly with Sasol. As APC Engineering Leader he led projects in both South Africa and Europe, in particular Eastern Europe. Kevin joined BluESP, the Aspentech distributor for South Africa, in 2007, where he has led projects in ethylene, fuel gas network optimization and novel applications in the platinum, copper, lead and zinc industries.

Dr John McCoy

John McCoy is a data scientist at Stone Three Digital, working on projects which include sugar cane yield prediction, sugar crystallisation batch monitoring and optimisation, and process optimisation for pharmaceuticals. John got his PhD in Chemical Engineering from the University of Cape Town, worked in operations and process engineering at Sasol’s Secunda plant, and did a postdoctoral fellowship at the University of Stellenbosch on machine learning methods for the minerals and process industries.

John Burchell

John Burchell is an APC specialist for Sibanye-Stillwater’s platinum operations in South Africa. Throughout his career he has led a number of projects to optimise processes in smelting, concentrating, and refining. Before joining operations at Sibanye-Stillwater in 2013 he worked as a consultant and academic researcher to develop bespoke soft sensors for industry. Since then
he has mainly focused on optimisation and plant wide control using technologies that include model predictive control, machine learning, and evolutionary programming.

Jason Miskin

Jason Miskin is a Consultant Team Leader at Opti-Num Solutions, focusing on Advanced Process Control and Predictive Maintenance in the process industry. He is passionate about solving practical challenges with the art of programming.

Danielle Winter

Danielle is an Application Engineer at Opti-Num Solutions and is passionate about machine learning, signal and image processing and computer vision applications. She is a specialist in MathWorks tools for data analytics in engineering solutions, with major focus on data science and machine learning in process engineering and machine vision.