TUTORIAL SESSION

Remote Monitoring of Toxic Gases in Underground Mines: building knowledge for improving miners’ health and future

Dr. Isaac Olusegun Osunmakinde
Council for Scientific and Industrial Research (CSIR), South Africa
osunmakindeio@yahoo.com

Duration

3 hours

Abstract

The steadily rising gas fatality in underground mines and the need to reduce global toxic gas concentration suspended in the air to improve miners’ health and to protect our environment are today’s economical and ecological drivers for the emerging consideration of mineral consumption in mining industry. After a steep increase of contributions, the mine safety-related research is currently entering a mature phase, in which specific solutions address mining gases challenges. This tutorial will address the challenging issue of static and mobile robot gas sensing in generating knowledge to assist in improving miners’ health and future.
Interestingly, certain algorithms rooted in computational intelligence show increasing performance in generating drivability maps for robot navigation, developing behaviours for robots, and eventually generate knowledge for assisting in reducing miners diseases resulted from toxic gases (e.g. Methane, Carbon monoxide gases, etc.). Specifically, computational algorithms based on Gaussian mixture model (GMM) and expected maximisation (EM), predictive power of k-nearest neighbour (k-NN) and Bayesian Network (BN) models will be presented for addressing the challenges above.

Objectives

Our tutorial consists of three parts including objectives which are:
(1) Introduction, Motivation, and Preliminaries (10 minutes)
Safety is very important in the mining industry. Underground mine accidents could lead to fatal injuries, sicknesses, diseases, huge economic loss and death of miners. The first part of the tutorial provides an introduction to several lives that have been lost due to toxic mine gases inhalations.
The introduction then lays a solid conceptual foundation towards the purpose of this tutorial, to identify and quantify toxic gases which are extremely prevalent in the underground mines and that cannot be easily detected by human senses.
It introduces key notions of generating knowledge for a better understanding of how a healthy underground mine environment is before entering there to work, on the basis of real-time remote monitoring of the gas levels in different underground mine areas, and the important relation to static or mobile robot sensing.
(2) Drivability Mapping of Underground Mines for Autonomous Robots (10 minutes)
The second part presents illustrations of a drivability mapping of the inherently unstructured mine terrains captured from a 3D SwissRanger camera mounted on a mine safety robot when gas devices are sensing the environment.
Principles include the SR4000 time-of-flight camera producing a stream of mine frames projected as a 2.5D, which are mainly mapped as drivable and non-drivable points using mixture of Gaussians (GMM) and expected maximization (EM) models based on a hypergraph-type model within an entropy textural feature space.
(3) Developing Behaviours for Robots to Autonomously Sense Toxic Gases (10 minutes)
The third part of the tutorial provides key strategy to manage the autonomous robot navigation with respect to collision avoidance during mobile sensing of the toxic gases in the environments.
As an important basis for the tutorial outline, this part reveals an approach of training a robot to avoid obstacles through teleoperation and thereafter use the knowledge acquired to develop behaviours for robot to autonomously navigate in various environmental sensing conditions using the computational learning and predictive capabilities of k-NN and Bayesian Network models. The chosen behaviour or navigational direction of the robot determines the control command values of translational and rotational velocities the robot uses for navigation.
Its excellent performance will suggest a wider application of the behavioural models which learn tasks and command robot successfully without collisions in an unknown environment in industry.

Timeline

Tutorial Outline:
(1) Introduction, Motivations, and Preliminaries (30 minutes)
* Origin and Relevance of Improving Miners Health and Safety
* Specific Motivations and Health Effects of Toxic Mine Gases
* Recent Gas Injuries, Illness, and Fatalities
* Toxic Mine Gases and Health Effects
* Hazards Associated with Toxic Mine Gases
* Existing Measures in Combating Toxic Mine Gases Fatalities
* Introducing a Real-time Remote Monitoring Framework
* Introducing Static and Mobile Robot Sensing
(2) Computational Algorithms to be covered are
The various computational algorithms/models that will be covered in this tutorial include:
* Gaussian Mixture Models (GMM) and Expected Maximisation (EM) Models
* k-Nearest Neighbour (k-NN) Models
* Bayesian Network (BN) Models
(2.1) GMM and Expected Maximisation (EM) Models (5 minutes)
Given a set of data points such as pixels in an image frame, the points in the frame seems to be generated in clusters but they are mixed. It is not clear if a meaningful dividing line can be drawn between them by investigating which particular cluster generates a data point. A family of distributions, P called mixtures of Gaussians or Gaussian mixture models are well-suited to modelling, say K, clusters or components of data points. The EM model is the maximization algorithm that is iteratively used to find the maximum likelihood estimate of GMM parameters until convergence, that is, when optimal values are reached.
In this tutorial, we will demonstrate that GMM and EM are quite powerful computational algorithms for detecting drivable and non-drivable regions for mobile robots in the mines.
(2.2) K-Nearest Neighbour (k-NN) Models (5 minutes)
k-NN is a nonparametric instance-based learning as it allows a hypothesis of model complexity to grow with data sizes. K-NN is based on minimum distance from a query instance to all training samples to determine the K-nearest neighbours, which span the entire input space. The Euclidean distance of n-dimensional space is commonly applied for computing the minimum distance in this step.
Prediction of the query instance is taken as majority votes of the K-nearest neighbours. The choice of parameter value k is critical but k-NN is advantageously robust to uncertainty or noisy training samples. There are many variants of k-NN, but more sophisticated versions can be proposed.
In this tutorial, we will demonstrate that k-NN is quite a powerful computational algorithm for detecting the navigational directions for autonomous robots and for predicting the future trends of gases concentrations in the mines.
(2.3) Bayesian Network Models (5 minutes)
Bayesian Network (BN) technology is very useful for encoding probabilistic knowledge as graphical structures. A Bayesian belief network is formally defined as a directed acyclic graph (DAG) represented as G = {X(G), A(G)}, where X(G) = {X1,…,Xn}, vertices (variables) of the graph G and , set of arcs of G. The network requires discrete random values such that if there exists random variables X1, . . ., Xn with each having a set of some values x1, . . ., xn then, their joint probability density distribution is well defined. The Network belief technology has been successfully used for reasoning in the areas of power transformer diagnosis, medical diagnoses, telecommunication networks, etc.
In this tutorial, we will illustrate that BN is a powerful probabilistic model for also detecting the navigational directions for autonomous robots in the mines.
(3) Drivability Mapping of Underground Mines for Autonomous Robots (30 minutes)
* Various drivability mapping approaches
* Proposed drivability mapping approach
* Introduction to hyper-graph models
* Entropy model in textural feature space
* Mapping with GMM, EM models and drivability refinement
* Scoring and evaluation scheme
(4) Developing Behaviours for Robots to Autonomously Sense Toxic Gases (30 minutes)
* Approaches for developing behaviours for robots
* Behavioural and collision avoidance modelling (CAM) for robots
* Perception of ultrasound sensor (US) data
* Bayesian learning and reasoning process to robot behaviour
* The k-NN modelling and reasoning process to robot behaviour
* Evaluation scheme
(5) Sensing Toxic Mine Gases and Predicting Future Situations (25 minutes)
* Remote Monitoring Approach and Current Situations of Different Mine Regions
* A Strategy to Determine an Appropriate Kth-value for Predicting Future Gas Concentrations
* Predicting Future Gas Levels using Forecasting Power of k-NN Models
(6) Demonstration of Remote Monitoring of Toxic Gases (20 minutes)
* Participatory sensing
* Holistic implementation framework
* Real-time gas sensing demonstrations

Background Knowledge Expected of the Participants

Intended audience for the tutorial are practitioners and scholars in the broad field of computational intelligence, sensor networks, and robotics interested in getting a better understanding about the issues related to improving underground mine safety on gas fatalities. The tutorial will be organised so as to allow a wide audience to take advantage from its content, ranging from graduate students to technologists, researchers, and practitioners willing to start working on safety, computational intelligence, and in the mining industries.

Qualifications of the Instructor(s)

Dr Isaac O. Osunmakinde is currently a Senior Research Scientist in the Modelling and Digital Sciences Department at the Council for Scientific and Industrial Research (CSIR), South Africa. By building on a first class B.Sc. (Hons) Degree in Computer Sciences, a PgDS in Applied Mathematics from the University of Stellenbosch, and a M.Sc. Degree in Computer Sciences from the University of Cape Town (UCT) South Africa in 2006, He received his Ph.D. Degree in Computer Sciences in 2009 (UCT) with a specialisation in Intelligent Systems. His research interest is in Machine Learning including Intelligent Systems, Pattern Recognition, Field Robotics, Sensor Networks, Probabilistic and Mathematical modelling. He is a member of the IEEE Computational Intelligence Society.