Machine learning explained, part one: supervised learning

Very different branches of data science are often grouped together under the same banner, and machine learning is quite often the forgotten child in the discussion. When people think of AI, they most commonly think of chatbots. Those a little deeper into the topic may think of agentic AI, where systems can plan steps, use tools and carry out tasks.

But there is another branch of data science that has been quietly transforming business decision-making for years, and that is machine learning.

At its simplest, machine learning is about using data to recognise patterns and make useful predictions, groupings or recommendations. It’s not magic, and it’s not a person sitting inside a cabinet. I’m looking at you, Mechanical Turk.

What is actually happening is a little less mysterious: machine learning uses statistics and mathematical models to learn from historical data, then apply those patterns to new situations.

That makes machine learning especially useful when a business has a lot of data and wants to answer questions like:

• What is likely to happen next?
• Which records look unusual?
• Which items are similar?
• Which action should we prioritise?
• What recommendation would be most useful here?
• Where should we focus attention?

For mining businesses, this is an important idea. Across operations, maintenance, health, safety, environment and risk, there are often years of valuable data sitting across incident systems, risk registers, inspection records, environmental monitoring systems, maintenance systems, fleet systems, IoT sensors, audit findings, corrective actions, training systems and health records.

The challenge is not always a lack of data. Often, the challenge is knowing how to turn that data into something useful, which is precisely where machine learning can help.

Clearing up the AI family tree

Before getting into the different types of machine learning, it is worth clearing up one of the most common points of confusion.

Machine learning is not the same thing as a chatbot.

A chatbot is something you talk to. You ask it a question, and it responds in natural language. Some chatbots are simple and follow rules. Others use large language models to understand and generate text.

Agentic AI goes a step further. It refers to AI systems that can take a goal, plan steps, use tools and complete tasks with more independence. For example, an agentic AI workflow might review a set of reports, find the key issues, draft a summary and send it to the right person for review.

Machine learning is different again.

A machine learning model usually has a more specific job. It might predict something, classify something, recommend something, or find groups in a dataset.

For example, in a safety context:

• A chatbot might answer the question, “How many vehicle interaction incidents did we have last quarter?”
• An agentic AI workflow might pull the data, analyse the trend, prepare a summary and send it to the HSE manager.
• A machine learning model might identify work areas that are showing early signs of elevated vehicle interaction risk.

These technologies can work together, and increasingly they will. But they are not doing the same job.

The two main types of machine learning

Most machine learning techniques sit under two broad categories: supervised learning and unsupervised learning. The easiest way to think about the difference is this:

• Supervised learning is used when we have examples from the past where we already know the answer.
• Unsupervised learning is used when we have a lot of data, but we do not yet know what patterns or groups exist inside it.

Both are useful, but they are useful for different kinds of problems.

Supervised learning

Supervised learning is probably the easiest type of machine learning to understand. The model is trained using historical examples where the outcome is already known.

Imagine you had a dataset of past safety incidents. Each incident might include details like the site, department, activity, time of day, equipment involved, worker role, incident type, contributing factors and final severity.

Because the final severity is already known for those past incidents, the model can study the relationship between the incident details and the outcome. Once it has learned from those examples, it can then be used on new records where the outcome is not yet known.

In plain English, supervised learning is a bit like learning from worked examples. The model is shown enough examples of “this situation led to that outcome” until it can start making useful predictions on its own.

The useful way to think about this is that the model is providing another signal. One based on patterns in the data, which can then support human judgement.

And that last part matters. In a business context, especially in areas like health, safety, environment and risk, machine learning should not be seen as replacing decision-makers. It is better thought of as a tool that helps people make better decisions, earlier.

Classification models

Classification models are used when the answer is a category. That category might be:

• High risk or low risk
• Serious or minor
• Compliant or non-compliant
• Normal or abnormal
• Likely or unlikely
• Requires review or does not require review

A familiar example is email spam detection. Your email system looks at the message, sender, links and wording, then decides whether the email is likely to be spam or legitimate.

The same idea can be applied in a business setting. In a mining HSE environment, a classification model could help assess incoming incident reports. It might look at the description, activity type, location, people involved, equipment and past trends, then flag records that appear more likely to be high potential. That doesn’t mean the model decides the severity. It simply helps bring attention to records that may need a closer look.

This can be valuable because HSE teams often deal with a high volume of reports. Some are straightforward. Others contain early warning signs that are easy to miss when people are busy. A classification model can act as a second set of eyes, helping teams prioritise review effort.

Classification can also be used outside incident management. Environmental readings could be classified as normal, abnormal or requiring investigation. Audit findings could be classified by likely risk level. Corrective actions could be grouped by whether they are likely to be completed on time or become overdue.

The important point is that the model is not replacing the person. It is helping the person find the right starting point.

Regression models

Regression models are used when the answer is a number. Instead of asking, “Which category does this belong to?”, regression asks, “What value should we expect?”

You already see this in everyday life. Property websites estimate house prices. Insurance companies estimate premiums. Banks estimate credit risk. Retailers estimate demand.

In mining, a regression model might be used to estimate things like:

• Likely downtime from a particular equipment issue
• Expected duration of an injury case
• Likely number of overdue actions
• Forecast monitoring workload
• Estimated cost of a claim or event
• Remaining useful life of a component

This is where machine learning can start to connect operational data with business planning.

For example, a maintenance team may want to estimate how many operating hours remain before a major component is likely to require replacement. The model might use previous maintenance history, utilisation, temperature, vibration, pressure readings and other IoT data to produce an estimate.

The value is not just avoiding failure. It can also work the other way. If a component is scheduled to be replaced every set number of hours, a model may show that under certain operating conditions, that component still has useful life remaining. That can help the business avoid replacing parts too early, while still managing the risk of running them too long.

That is an important distinction. Machine learning is not only about preventing unplanned downtime. It can also help businesses make better use of expensive assets.

If a model suggests a component is moving closer to failure, maintenance can be planned before the issue causes unplanned downtime. If the model suggests the component is still performing normally, it may support a decision to extend beyond a fixed maintenance interval, provided the right controls and engineering judgement are in place.

Regression models are useful because they help businesses move from fixed assumptions to evidence-based estimates.

Time series forecasting

Some data only really makes sense when you look at it over time. Incident volumes change month by month. Environmental readings move throughout the day. Fatigue risk can rise and fall across a roster cycle. Maintenance issues can build up slowly before they become obvious.

Time series forecasting is used when the order of events matters. A forecasting model looks at historical patterns and tries to predict what is likely to happen next. It may pick up trends, seasonality, spikes, cycles or gradual changes that are not obvious from a simple report.

Weather forecasting is the everyday example most people understand. A forecast is not just looking at one number in isolation. It is looking at how conditions have been changing over time and what usually follows those patterns.

Seasonality is another simple example. If you have ever wondered why you sell more gumboots in the wet season, or more sunscreen in summer, you are already thinking about time-based patterns. The same idea applies in business data. Some patterns are not random; they follow the calendar, the weather, the roster, the shutdown schedule, or the production cycle.

In mining, time series forecasting can be useful in both HSE and operational settings. An environment team might use it to predict dust, water or noise levels based on historical monitoring data. A health team might use it to understand how fatigue indicators change across different rosters. A safety team might use it to understand how reporting patterns change around shutdown periods, seasonal weather or high-activity months.

On the operational side, forecasting can also be used with sensor data from fixed plant, mobile equipment or processing assets. Temperature, vibration, flow rates, pressure and utilisation can all tell a story over time. A single reading may not mean much on its own, but the pattern across days, weeks or months may indicate that something is starting to drift from normal behaviour.

This is one of the main reasons forecasting is so useful. It helps teams prepare, not just report. An analytics dashboard will tell you what happened last month, a forecasting model might suggest what is likely to need attention next month.

Recommendation models

Most people have interacted with recommendation models without even knowing it. Netflix recommends shows. Spotify recommends music. Online stores recommend products. News apps recommend articles. Social platforms recommend posts.

The basic idea is simple. Based on what the system knows about you, and what it has learned from similar people or similar items, it suggests something that may be relevant.

In a business setting, recommendation models can be used to reduce the effort required to find useful information. Imagine a supervisor is reviewing a hazard report about working at heights. A recommendation model could suggest relevant controls, similar previous incidents, useful procedures, inspection checklists or lessons learned.

That could save time, but more importantly, it could improve the quality of the response. People don’t always know what information already exists in the business. They may not remember a similar incident from another site. They may not know that a procedure was updated recently. They may not be aware that a particular control has been used successfully elsewhere.

A recommendation model can help surface that information at the point it is needed. In HSE and risk, this could apply to:

• Suggested controls for a hazard
• Suggested corrective actions based on similar events
• Recommended training for a worker or team
• Recommended audit focus areas
• Recommended documents or procedures
• Recommended risk treatments

This is one area where machine learning can feel very practical. It is not necessarily making a big decision. It is simply helping people find useful information faster.

Optimisation models

Optimisation is about finding the best option when there are limits and trade-offs. That makes it slightly different from prediction. Prediction asks, “What is likely to happen?” Optimisation asks, “What should we do, given the constraints?”

You see optimisation in route planning apps. Google Maps does not simply list every possible road, it recommends a route based on travel time, distance, traffic and road conditions. It is trying to find the best path under the circumstances.

Mining has no shortage of constraints. There are people, crews, equipment, shifts, locations, production schedules, regulatory requirements, shutdown windows, maintenance requirements, equipment availability and limited time. This makes optimisation especially useful.

In operations, optimisation might be used to improve haulage routes, schedule maintenance windows, allocate equipment, reduce queuing time, or balance production requirements against asset availability.

It can also be used to test which operating settings produce the best mining outcome against a short-term plan. For example, a fleet management system may have a range of settings that influence truck allocation, digger utilisation, queuing, production rates and plan compliance. An optimisation model can help search through those settings and identify which combination produces the best result for the plan.

This is a good example of where optimisation earns its keep. The business is not just sorting items into high and low priority. It is trying to find the best combination of settings across a system where changing one thing can affect many others.

That distinction is important. If the problem is simply “which corrective actions are most serious?”, that is more likely to be a classification, scoring or prioritisation problem. A model could grade actions into priority groups, and people could start with the highest-risk items.

Optimisation becomes more relevant when there are constraints and trade-offs to solve across the whole system. For example:

• Which jobs should be scheduled into which maintenance windows?
• Which inspections should be assigned to which people, across which sites, on which days?
• Which equipment should be allocated to which tasks to meet the short-term plan?
• Which fleet settings produce the best outcome without creating bottlenecks elsewhere?
• Which combination of controls gives the greatest risk reduction within a fixed budget?

The model is not just producing a number. It is helping weigh up competing choices. That is why optimisation can be so valuable. In complex systems, the best answer is not always obvious.

Anomaly detection

Anomaly detection is about spotting things that look unusual. Banks use this for fraud detection. If your credit card is suddenly used in a way that does not fit your normal pattern, the bank may flag or block the transaction.

Businesses can use the same principle to identify unusual behaviour in operational data. In a mining environment, anomalies might include:

• Environmental readings outside the usual pattern
• Equipment behaving differently from normal
• A contractor group with an unusual trend
• A sudden change in reporting from a site
• A cluster of similar hazards appearing in a short period
• Leave patterns that look unusual for a particular person or team

This is another area where IoT data can be valuable. A pump, conveyor, truck or processing asset may produce a constant stream of readings. Most of the time, those readings follow a normal pattern. But when vibration, pressure, temperature or load starts behaving differently, it may be an early sign that something has changed.

The same idea can apply to people-related data, but it needs to be handled carefully. For example, a person’s leave pattern might change noticeably. They may begin taking more leave than usual, or taking leave in a pattern that is unusual for them. On its own, that change does not explain what is happening. It may be completely reasonable, or it may point to something that deserves a closer look.

Used responsibly, anomaly detection could help a manager or HR team identify that someone may need support. Used poorly, it could feel intrusive or lead to unfair assumptions. This is why people-related machine learning needs strong governance, privacy controls and human judgement.

One important point is that unusual does not always mean bad. The model can flag the pattern, but people still need to interpret it. A sudden increase in hazard reports, for instance, might mean risk has increased, or it might mean workers are more engaged and reporting more openly. A drop in incidents might mean performance has improved, or it might mean under-reporting. This is often where anomaly detection is most useful. It does not give the business a final answer, but it can help point people toward the right questions.