August 2019

The path to autonomous mineral processing operations

  • By Arthur Gooch, Director of Innovation Automation Solutions, ANDRITZ Automation Ltd

Autonomous technology is being deployed across the resources sector. A new type of technology, based on proven techniques, looks to bring the benefits of AI to processing plants.

Automating mineral processing operations can increase safety, optimise performance, improve system reliability and help to detect problems early. In mining, autonomous technology is now applied to haul trucks, trains, drill rigs, drone-based monitoring and other simple systems, but not yet to processing plants. Industry 4.0 brings the opportunity for a step change in processing automation as a new range of technologies address challenges that have stood for 50 years.

The hype surrounding Industry 4.0 can be overwhelming, and the move towards a highly digitised industry is perceived as risky to many in the mining industry, with seemingly little proven tangible benefit behind the hype. Concepts like Artificial Intelligence (AI), machine learning and digital twins are becoming ubiquitous in the media, but the results of such technologies are not always readily apparent. In light of this, ANDRITZ’s IDEASTM Intelligent Control combines three elements that are each proven technologies: 

  1. an AI controller, as used in many business and consumer applications (eg search engine and mapping applications, inventory and logistics management or stock market trade decision making)
  2. an AI learning method called ‘reinforcement learning’ – a model where an AI system is rewarded for completing the right action or penalised for completing the wrong one
  3. IDEAS dynamic process simulation models, which have been used by mining companies for more than 15 years for applications including piping and instrumentation diagram (P&ID) validation, control system check-out and simulator-based operator training. 

Each of these elements is explored in more detail below. 

1. An AI controller for industry

When the public thinks of AI, the perception is often that of an independent intelligent entity, capable of learning abstract tasks. Known academically as ‘strong AI’, this capability is still far out of reach for us – probably to the relief of many. When we talk about AI for industry, we mean ‘weak AI’, which may appear to show intelligence, but really is closely tailored to automate a particular task or system. When it comes to Industry 4.0, the successful application of weak AI means automatic solving of specific problems without completely removing humans from the loop. This type of AI is successfully employed in many consumer and business applications as mentioned above. 

Human factors – the case for automation

For all the cognitive strengths of humans, we are still bound by physiological limitations. The average person can only process about 120 bits per second of information. When it comes to short-term memory, the average person can only hold about seven objects at once. Compared to the quantity of data available in a modern control system, the amount that an operator can concentrate on at once is relatively small. 

Operators’ limited attention capacity is further compounded by fatigue. Safety is the number one priority for many mining operations, so there is strong incentive to manage operator fatigue.

At its heart, Industry 4.0 means applying new technologies so that operators and software complement each other, capitalising on strengths and patching weaknesses.

2. AI and reinforcement learning

Due to its complexity, decision-making has been the last holdout in automatic control. Tasks that require complicated action steps, or require an operator to weigh current versus future rewards, have typically been left to humans. Modern machine learning techniques offer us a way to tackle these challenges. Consider the problem of needing to reach an empty space in a large car park by driving a roundabout route. The idea of having to move away from a goal to maximise future reward is straightforward for humans, but a real problem or even an impossibility for conventional automatic control algorithms. Using a type of training for the AI called reinforcement learning, we are able to control situations like this in industry. 

Reinforcement learning explores the operating space, learning which choices in each state lead to the defined goal. A simple example would be for the AI to learn to navigate the maze in Figure 1. 

To learn the system, we assign a value of 0 to the starting point and 1 to the finish line. The AI then ‘walks’ around the maze, exploring each state methodically. By remembering which choices led to the finish (when it eventually does find the finish) it learns to associate a value with each state, as shown in Figure 2.

Once this map of values has been created, navigating through the maze is possible from any point by advancing to whichever adjacent cell has the highest value. To accomplish the same outcome with conventional control system logic would be a significant task, even for this simple example.

Figure 1. The maze for an AI to navigate.
Figure 2. The AI learns to navigate the best path from start to finish by allocating a value to each state. 

3. IDEAS dynamic process simulation models

The ANDRITZ approach to training AI is not new. Mining companies have been using IDEAS process models built as an Operator Training Simulator (OTS) to train human operators for mineral processing plants for many years. This involves using the simulation process models to create the industrial analog of a ‘flight simulator’. The controlled nature of the simulator allows the creation of specific certification scenarios that are used to prove competency before the operator runs the real plant. These scenarios can also explore non-standard operating conditions and enable training in dangerous upset conditions without placing personnel or equipment at risk. 

These IDEAS simulation models can be used to train the AI to run the plant using reinforcement learning. The simulation can also run faster than real time, so the AI gains months or years of experience in a much shorter time frame before operating real equipment. 

The plant simulation model is high fidelity and dynamic. The model is built using objects for each unit process, as well as first principle equations, thermodynamic and chemical reactions of the processes and equipment. Detailed engineering information is entered into the IDEAS model, including equipment specifications, dimensions, elevations, pump curves, valve CVs, etc. The model includes the maximum design operating allowances, such as motors operating at maximum load for start-up, conveyor capacity limits, etc. The IDEAS model runs a simulation of the plant operation that is not just accurate in steady state conditions, but dynamic as the plant is starting up, shutting down and going through other changing operating conditions. This is especially important for processes where there are recycle flows, with different separation and yield effectiveness at different ore and capacity conditions. The simulation is highly accurate (typically >95 per cent) to real plant operations.

Whether training human operators or AI, the IDEAS process simulation model allows the trainee to do much more than would be practical on a running plant. Virtual equipment can be started and shut down numerous times, allowing the trainee to experience unstable and upset conditions without consequences to plant productivity. A systematic approach drives the creation of each upset scenario. Past maintenance records and experience from other plants indicate what sort of failure scenarios should be considered for training in each area. Each possible event, such as an individual instrument failing, is used to create a test scenario for all similar devices in the area. Each of these scenarios are then added to the learning set. The AI repeats each scenario until it has learned to act optimally in each case. The AI learns to safely operate the plant to the maximum physically possible performance and capacity in the most cost-effective way. 

IDEAS Intelligent Control 

Bringing these three elements together – a proven AI decision-making agent; teaching the AI through reinforcement learning; and using the IDEAS dynamic simulation model – is the basis of ANDRITZ IDEAS Intelligent Control. The AI helps by working with the human operator; humans retain control over any tasks requiring creativity and intuition, while the AI performs procedural tasks needing attentiveness, speed and discipline. 

IDEAS Intelligent Control (previously called IDEAS Reinforcement Learning AI) won the CAD$1 million first prize in Newmont Goldcorp’s annual #DisruptMining competition in March 2019. One of the key reasons for the technology’s success is that it combines three proven technologies, resulting in a robust, reliable system that will deliver value with minimal risk. 

The proof of concept project is currently being implemented at a Newmont Goldcorp mineral processing plant in Mexico, with the aim that results will be shared in early 2020. 

ANDRITZ Automation in Australia

ANDRITZ has opened an office in Perth, which focuses on advanced automation and simulation for mining and mineral processing. The ANDRITZ Australian regional head office is located in Melbourne. ANDRITZ is coauthoring a paper for presentation at AusIMM’s World Gold conference in Perth in September with Oceana Gold. ANDRITZ is looking to collaborate with mining companies globally who are interested in implementing digitalisation technologies for tangible operational value. 

Share This Article