Canary of the coal mines

A new study has investigated the role artificial intelligence can play in preventing coal mine disasters.

Alongside industry and stakeholder demand for less energy-intensive operations are calls for measures to protect workers and create safer worksites.

As such, understanding and working to mitigate industry-related risks, especially those relating to health and safety, continues to be a major focus among industry leaders.

Artificial Intelligence (AI) has been applied in many areas and across a range of industries, including to assist users and decision-makers to address changes and make smarter decisions.

In the mining industry, the technology’s uptake in various applications is progressing in leaps and bounds, including analysing geological data, boosting autonomous operations and predicting equipment maintenance needs.

Intelligent hazard identification

One of the key characteristics of AI algorithms is their ability to identify anomalies and predict unknown future parameters and potential outcomes. This capability has made AI an efficient tool that can be applied in the mining industry for hazard prediction.

A recent study investigating how AI can reduce the risk of disasters was conducted by the Australian Catholic University, University of Technology Sydney, Charles Darwin University and Central Queensland University.

The study was conducted in coal mines in China and compared ten machine learning algorithms to ascertain which AI method could not only make predictions about methane gas level changes 30 minutes in advance, but also notify users of anomalies.

The study found that AI can, in fact, forecast gas-related incidents in coal mines within half an hour.

Methane is created as a byproduct of coal development, with the methane that is absorbed in coal released during mining operations. Methane is more readily combustible than coal dust and when a buildup of methane gas encounters a heat source it can result in an explosion.

As such, the careful monitoring of methane levels is a priority in keeping workers safe, and predicting methane gas levels plays a significant role in preventing accidents in coal mines, with around 60 per cent of the mining accidents[1] caused by methane gas.

Understanding this, the team of researchers, using AI algorithms, were able to predict the emission of methane gas levels with a high level of accuracy.

The process

The study was conducted in partnership with Shanxi Fenxi Mining ZhongXing Coal Industry Co (ZhongXing) which is owned by Shanxi Coking Coal Group and provided the critical data used for the study.

The lead researchers of the study, Associate Professor – Computational Intelligence, Peter Faber Business School, Australian Catholic University and Adjunct Associate Professor, Faculty of Science and Technology, Charles Darwin University, Niusha Shafiabady, who was the lead AI expert in the team, and business intelligence expert Dr Robert Wu, applied a series of machine learning algorithms, predicting the methane gas level in different scenarios.

The different scenarios incorporated data that was collected on different days to ensure the reliability of the study’s outcomes. ZhongXing provided readings that were collected every 20 seconds from temperature, methane gas and wind sensors at the coal mines.

Using the data, the AI algorithms were then asked to predict the methane gas level within the next 30 minutes. This is known as time-series prediction and is considered to be short-term forecasting.

There are four types of forecasting:

Very short-term forecasting – a few seconds to 30 minutes ahead
Short-term forecasting – 30 minutes to six hours ahead
Medium-term forecasting – six hours to one day ahead
Long-term forecasting – one day to one week ahead

The team aimed to explore more efficient machine learning algorithms with better performance for short-term forecasting. For this study, the researchers applied ten different machine learning algorithms and then compared the models’ outcomes against each other.

The ten models were selected in accordance with their previously reported capabilities in similar studies that outlined short-term prediction of hazards.

Different types of accuracies were measured for the ten applied algorithms. Additionally, the researchers calculated the time that was required for training the models for each scenario.

Tuning the algorithms

One of the critical goals for the study was tuning the models in the shortest possible time while maintaining high accuracy for the models consistently. Tuning refers to taking a pre-trained machine learning model and adapting it to new data or tasks. The process of tuning the machine learning models is referred to as training the models.

AI systems need to be trained and tuned properly for optimal capabilities. The researchers in the study trained the machine learning algorithms using the data from different days to ensure the AI models learned the hidden patterns and instead of just memorising the data.

AI systems have the capability to imitate the behaviour of real systems. This study involved ten machine learning models applied to imitate the behavioural pattern of methane gas emission in the coal mines through reading the measurements from the temperature, gas and wind sensors.

In the tuning of the algorithms, the data ZhongXing sent through was divided into two portions – the data that was provided to the models for training during the training process was called ‘train data’, the remainder of the data was not shown to the models and was referred to as ‘test data.’ Test data is kept from the models in order for researchers to see how they would perform in a real setting.

To ensure the predictive capability of the models in real settings, their performance to the unseen data had to be measured, which is why different test accuracies were recorded and compared with each other for the ten proposed models.

The accuracies reported in this study were the test accuracies of the models, with the performance of the models on the unseen data reported. The measures used for testing the accuracies of the models were commonly used errors such as mean squared error (MSE) which measures the average squared difference between actual and predicted values in a test dataset. This enabled the researchers to foresee the estimated performance of the tuned models in a real-life setting.

Results

When it came to assessing the performance of the ten applied machine learning algorithms, researchers compared their required training times and their accuracies against each other.

The steps in the process the researchers used to find the most efficient machine learning algorithm with better prediction assessments for short-term forecasting were:

Data collection and preparation
Prediction error assessment/performance assessment
Validation tests
Comparative analysis

Random Forest (RF) demonstrated better performance overall compared to the other models. In all the scenarios related to the different days’ data, the average squared difference between actual and predicted values (MSE) for RF was between 0.000025 to 0.000376.

Optimal algorithm

In terms of tuning, the training time required for the different algorithms except for Long Short-Term Memory (LSTM) Neural Networks and Recurrent Neural Networks (RNN) was less than seven seconds. These two particular neural networks have a more complex structure so training takes longer.

RF, which was the algorithm with the overall better performance than the others considering the training time and accuracy, is an ensemble learning method, meaning that more than one decision-making unit searches for the solution to the problem the expert system is designed to solve.

RF is a combination of several decision trees – a tree-like model that explores the solutions by splitting and making branches from the top of a tree. Each branch would lead to an outcome and the cost and performance of that outcome would be calculated. After the training is completed, the best outcome would be selected as the solution to the problem. RF is a combination of decision trees which means that at the end of the training phase, the best decision tree’s outcome will be selected as the solution to the problem.

In predicting the coal mine’s methane gas level, RF had a better performance in comparison with the other nine machine learning models used in the study.

Even with the move towards a greener future, in 2021-22 Australia produced 422 million tonnes[1] of coal. Considering the reports that 60 per cent of the mining accidents are caused by methane gas, and the critical importance of health and safety in the mining industry, leveraging AI and technology to provide different insights is of great importance.

Something that may seem minor – like adding even just one minute to the prediction time span of upcoming hazards – could save lives. Studies like this one, which predict upcoming risk factors, could play a significant role in mitigating risks.

The outcomes of this study can be used to inform research and provide similar solutions for the Australian coal industry to monitor the methane gas levels and leverage AI to predict gas levels to mitigate risks due to gas emission in advance.

A similar approach can be used to provide solutions to the mining industry for predicting different hazards, including rising methane gas levels, around the world.

Footnotes: