Mining managers routinely must make critical operations decisions—while meeting day-to-day obligations such as safety and production targets. Such decisions typically involve complex trade-offs (such as between operating cost versus yield in processing plants). Yet many managers make sub-optimal decisions, relying on simple spreadsheets or rules of thumb. Taking advantage of machine learning can help them avoid this pitfall. But first, they must familiarise themselves with the technology.
Machine learning: A quick primer
Machine learning, a subfield of artificial intelligence, comprises algorithms that aim to understand relationships in complex data sets, and that use that understanding to make predictions. These predictions can be simple, such as what temperature an equipment component will reach under certain conditions, or complex - for instance, whether and when a crusher will fail.
The algorithms ‘learn' by discovering the relationships in the data, without a person having to program specific rules, relationships or equations. Through such learning, an algorithm can detect complex patterns emerging from thousands of variables, even in complex operating environments such as mines.
In fact, machine learning today routinely tackles problems that were impossible to solve as recently as five years ago, owing to the wealth of data available and the rise of cheap sensors and smart devices. New analytical tools, algorithms and data transmission technologies let managers analyse data they previously could not. Cheap and easy storage, along with massive increases in computing power (including scalable cloud computing), have made data analysis more affordable than ever. Forward-thinking miners are taking advantage of these developments to explore the benefits of machine learning for their operations.
Machine learning in action in mining
Machine learning can help miners uncover valuable insights, drawing on the real-time, high-volume and unstructured data typically seen in mining. Here are five proven examples, below, that miners could consider to get started.
How do you measure the impact of a high-energy explosive at a mine site when every blast varies, and you can't repeat a blast? One mining company called in a team of data scientists to answer this question. The scientists applied machine learning techniques to understand the relationship between drill hole patterns, blast design, explosive type, geology and observed rock-fragmentation from roughly 80 blast events over 6 months. Once their analytics model had learned the interrelationships among the data, it could make reliable predictions for what fragmentation would have happened if a different explosive type had been used. The mining company could then use this information to select the right explosives to minimise cost for the desired rock-fragmentation outcome.
When it comes to haulage, driver behaviour powerfully influences fuel consumption in opencut mines. With this in mind, a Congolese copper mine—with multiple pits and stockpiles, a complex road network, and a largely unskilled workforce—set out to quantify the impact of poor operator behaviour on fuel consumption. It brought in data scientists who used low-cost spatial trackers and drones to capture real-time information about trucks' location, time, speed and vibrations. The scientists then used specialised software and a statistical toolbox that leveraged neural network techniques to analyse truck dispatching and track vehicle movements. The result was direct feedback that showed operators how they were driving their trucks. Armed with this knowledge, they were able to limit peak speeds, reduce short stops and restarts, and avoid abrupt braking and strong accelerations. In just eight weeks, fuel consumption dropped by 7%.
Mining equipment often exhibits a signature ‘fingerprint' (pressure or temperature spikes, electrical signals, oil leaks, noise, vibrations) before it breaks down. With low-cost sensors, more data than ever is available on equipment status—but this can lead to information overload for engineers and managers.
Algorithms can learn to identify the unique (and often complex) signatures of a failure by modelling the relationship between observed failures and data on factors influencing equipment status, such as operator behaviour, historical maintenance and weather. In some cases, the algorithms can detect impending failures days in advance. Such analysis can help miners efficiently schedule maintenance. The result? Increased equipment uptime, and a higher proportion of (lower-cost and safer) planned maintenance.
At a mid-sized copper smelter struggling with thin margins, engineers had tried for years to optimise yield, using their knowledge of chemistry and physics. But the complex data relationships and multiple, ever-changing variables stymied them. The company brought in data scientists to build an artificial neural network that analysed years' worth of data. The model showed that declining yields stemmed from the chemical recipe, not temperature as suspected. The smelter implemented a new set of operating rules that required no capital expenditure—and achieved a 2% yield improvement.
Ballast fouling is a major problem for rail networks used in mining. Rail ballast - the track bed on which sleepers or railroad ties are laid - becomes dirty and dusty over time, leading to dangerous track deformations. Predicting when and how this will happen isn't easy. Typically, rail operators collect extensive data but have difficulty anticipating which ballast areas will need cleaning when, and what impact cleaning will have. Consequently, they spend as much as 20% of their annual maintenance and sustaining capital budgets on cleaning ballast.
One rail operator asked data scientists to conduct a five-week machine-learning proof of concept. They integrated data sets from ground-penetrating radar, maintenance and weather; built a model to predict ballast fouling; and designed a customized optimisation tool to help managers identify the best sections of track to remediate. Managers discovered that they could reduce ballast-cleaning costs by up to 13% by eliminating unnecessary maintenance.
While machine learning can help miners transform their operations, leveraging it takes some work. It takes time to decide which data to use, clean the data and trial analytics models. Plus, there's no one ‘right' machine-learning algorithm for all business problems. Each company must select and tailor algorithms that best reflect its own challenges and circumstances.
That being said, miners can take a few steps to begin exploiting machine learning—even if they don't yet have sophisticated IT systems and tech-savvy experts at hand. Most important, they can identify business decisions currently facing them that are:
• Difficult to make, owing to their complexity
• Being made using gut instinct or unsophisticated analytics tools (such as spreadsheets)
• High stakes, because the difference between a ‘good' decision and the ‘best' decision has a material impact on value
For such decisions, machine learning can be a powerful tool. ‘Sprints' can help miners generate tangible value within weeks. These pilot projects use machine learning to address a particular problem that meets these criteria. They're implemented in stages by teams comprising data scientists and in-house decision makers. Successful sprints can inspire confidence and build momentum for additional testing and learning and a broader digital-mine programme. But to get the most from sprints, miners should adopt a value-focused approach: Identify the most promising machine-learning projects using input from diverse parts of the organization, and bring in the right data science capabilities to support the projects.
Machine learning has begun helping miners sharpen their competitive edge. But today's applications are just a starting point. The future will bring even greater opportunities for miners who lay the groundwork now to take advantage of this powerful technology.
*Dale Schilling is an associate director in BCG's Sydney office, and BCG's global topic lead for mining operations. Julian King is an expert project leader in BCG's Sydney office. He leads the machine learning practitioners' network at BCG. Rohin Wood is an expert principal in BCG's Sydney office. He leads the optimisation topic at BCG. Tom Vogt is an associate director in BCG's Chicago office, and BCG's global topic lead for digital in mining. Author emails:email@example.com, firstname.lastname@example.org, email@example.com and firstname.lastname@example.org
See also, Petra reads the songs