Machine Learning Techniques to Model and Predict Airflow Requirements in Underground Mining ^†

Abstract

This paper analyzes the airflow requirements of underground operations and the accurate assessment of future conditions so as to effectively adjust ventilation parameters. More particularly, ML techniques are utilized to capture patterns or prevailing conditions and to be able to generalize/predict future conditions managed via the ventilation system. The case examined is about underground bauxite mining operations, the ventilation characteristics and requirements of which have been firstly developed and modelled into a validated digital twin. With this twin model, several scenarios are developed and evaluated and more importantly data are gathered, allowing for the training of the ML algorithms used to assess and predict the required ventilation airflow, taking into account air quality data, the number of workers, and machine fleet.

1. Introduction

Underground mining operations are some of the most challenging working environments, where controlling and maintaining air quality is of fundamental importance. The main goal of a ventilation system is to supply fresh air and successfully disperse gaseous pollutants and airborne particulates. Analyzing conditions and taking into account local or overall requirements while at the same time adjusting and directing the airflow has, up to now, in the majority of cases heavily relied on the manual control of its parameters (man-in-the-loop).

In recent years, challenges such as reduction of the environmental footprint and energy consumption and social concerns about H&S issues have led the mining industry to adopt new processes and technologies in order to address these issues and ensure the long-term sustainability of the operations. In order to achieve the above requirements, researchers have attempted to combine the ventilation on demand (VoD) method with ML techniques and especially with artificial neural networks (ANNs) [1,2,3,4].

This paper analyzes the airflow requirements of underground bauxite operations and the accurate assessment of future conditions so as to effectively adjust ventilation parameters to satisfy set requirements. In this way, the design of the primary ventilation system, assisted with an ANN model, can be automatically adjusted to meet the existing conditions and requirements. Thus, with the development of a ventilation on demand strategy coupled with autonomous capabilities, both significant improvements in work safety and significant energy savings in ventilation energy consumption—and a reduced overall environmental footprint of the operation—can be achieved.

2. Description of Vargiani Mine’s Ventilation System

The bauxite mine under examination belongs to Delfoi Distomo S.A and is one of the most productive mines of the company, located in the northwestern foothills of Mount Parnassus and east of Mount Giona in Phocis, Central Greece. The local area has a strong mining character with extensive underground operations. The mine has taken its name from the village of Vargiani, which is located 320 m northeast of the mining area.

The whole mine ventilation system consists of 8 km of main tunnel while the total horsepower of the machine fleet working underground is calculated at 2170 hp with 20 miners (at full operation for both shifts). A total airflow volume of approximately 90 m³/s is needed to meet the ventilation requirements and is supplied by the use of two main fans—working in parallel—at the entrance of access tunnel [5]. The requirement airflow calculated is based on the stricter standard in effect in Greece [6], which defines the minimum airflow requirement for diesel equipment at 2.3 m³/min per HP and for personnel at 5.66 m³/min per person.

The digital twin was developed via the import of the network’s characteristics (length, friction characteristics, fans, etc.) into the Ventsim Design software program. The proper adjustments were made and the model validated via comparison with real data measurements that have been carried out at the mine, both in terms of airflow quantity and quality, under several available real-life scenarios. Based on this twin, several scenarios were developed and evaluated, and more importantly, data were gathered that allow for the training of the ML algorithms. The whole process of network modelling and validation has been already given in detail by Karagianni and Benardos (2021) [7].

3. Machine Learning & Artificial Neural Networks

3.1. Machine Learning

Machine learning (ML) is a branch of artificial intelligence (AI) and computer science which focuses on the use of data and algorithms to imitate the way that humans learn and gradually improve accuracy. In fact, ML as a discipline aims to design, understand, and apply computer programs that learn from experience for the purpose of modelling, prediction, or control. Supervised, unsupervised and reinforcement learning are the three main types of ML methods. In this paper, we applied supervised learning and trained the algorithm by using “labelled” training data from the dataset. This means that some input data are already tagged with the correct output. The goal is for the algorithm to approximate the mapping function so well that when there are new input data, it will be capable of predicting the output variables for these data.

3.2. Artificial Neural Networks

Artificial neural networks (ANN) are a ML technique and considered information-processing systems capable of learning and generalizing from “experience” [6]. They are characterized by flexibility, ability to generalize complex phenomena, and capability to satisfactorily encounter and address conditions where time-evolving situations are to be modelled. In this study, two of these were developed from the most basic and popular types of ANNs. The feed forward neural networks (FFNs) and the long/short term memory neural networks (LSTMs) and the algorithm were written using the Python programming language.

3.2.1. The Main Characteristics of an ANN Model

A simple ANN consists of an input layer, one or more hidden layers, and an output layer. These layers are interconnected with a specific number of neurons that transfer the signals. The way that all neurons are connected to each other is called topology (or architecture) and is one of the most important features, because the architecture determines the capabilities of the ANN. Besides the above, there are also two main characteristics for the composition of an artificial neural network: the training or learning algorithm, which is the method that establishes the values of the weights on the connections [8] and the type of activation function, which defines the output of a neuron for a given input.

3.2.2. FFN and LSTM Neural Networks

FFNs are the simplest type of artificial neural network. In this kind of network, the information moves in only one direction—forward—from the input layers, through the hidden layers and to the output layer. There is no “feedback loop” from the outputs of the neurons towards the inputs throughout the network.

On the other hand, LSTMs are also feed forward neural networks but with memory cell and recurrent connection. This means that these networks have a “feedback loop” from the outputs of the neurons towards the inputs at some point before being fed forward again for further processing and final output. Also, they are usually used to process and make predictions for sequences of data. In other words, the network expects the input in the form of a sequence of features and is useful for data such as time series. Figure 1 shows in schematic representation the above differences [9,10,11].

3.3. ANN Model Development

The aim of the models is to be trained and learn how to adjust the fan’s characteristics so to generalize solutions that will supply the mine with the required airflow, based on the number of workers and amount of diesel equipment that will be working at a given time (ventilation on demand) [12,13].

The dataset consists of total 813 records for each group data, 80% (650) used for the training set and 20% (163) for the test set. For the LSTM, the data splitting is in the form of a sequence, while for the FFN, it is random. Because of the different measurement units, the input data were normalized based on min–max normalization). The scenarios that have been made for both models—and the data gathered—have a maximum total diesel equipment horsepower of 1540 HP and maximum personnel of 8 workers (one working shift). In these scenarios, a quantity of equipment and workers enter and exit the mine for a given time frame with the model aiming to predict and adjust the fan’s characteristics in order to supply the mine with the required airflow. Also, the accuracy of the generalization of the ANNs is evaluated based on the mean absolute percentage error (MAPE) between the predicted and the actual value of required airflow (Equation (1)):

$$MAPE=\left|\frac{{Actual}_{value}-{Predicted}_{value}}{{Actual}_{value}}\right|\mathrm{\%}$$

(1)

3.4. Models Architecture

Finding the optimum ANN topology involves a set of trials and errors in order to achieve the minimum relative error. Finally, the best architecture for each model is presented in

Table 1. Optimum architecture for each ANN model.

Layers		LSTM Model (2 × 40 × 20 × 1)	FFN Model (2 × 65 × 45 × 30 × 20 × 10 × 1)
Input Layer		2 input neurons (hp of diesel equipment and number of workers)
Hidden Layers	#1	40 neurons	65 neurons
		20% dropout rule
	#2	20 neurons with a 10% dropout	45 neurons
	#3	-	30 neurons with a 10% dropout
	#4	-	20 neurons
	#5	-	10 neurons
Output Layer		1 output neuron (Fan Airflow) assessing the requirement airflow

and

Table 2. Training Requirements for each ANN model.

Training Requirement		LSTM Model	FFN Model
Activation Function (throughout the model)		Tanh	ReLU
Activation Function (output neurons)		ReLU	Linear
Optimizer		Adam 1
Stop Training	Number of completing epochs or Number of consecutive seasons for the achievement of the minimum MSE	200	2000
		50	100

¹ The Adam function is a method for stochastic gradient descent optimization.

, as given below:

The validation of the models is a two-step process. The first involves the achievement of a small MSE (preferably below 1) in training data; the second encompasses both the accomplishment of a small MAPE in the test set but also consistent general behavior of the model’s generalizations. Of course, there are estimation uncertainties in the model’s predictions besides the fact that the ANN shows a high degree of accuracy (e.g., even a light change in temperature can change the environmental conditions of the mine). One way to prevent these uncertainties is by importing a greater volume of data—and parameters examined—into the model algorithms, which could, in turn, expand its working envelope and generalization capabilities.

4. Models Results

At the end of the training, a major task is to check the general behavior of the model. The generalization represents the way the ANN response to the training. If it follows the trend of the data and doesn’t focus on specific groups of values, it means that it is capable of reliable predictions. Figure 2 and Figure 3 present the generalization behavior of the LSTM and FFN models accordingly. The orange line shows the actual values while the blue line shows the predicted values.

As is shown, the LSTM model is rather overfitting the training data, and it is not capable of yielding accurate predictions and probably might not be suitable for the current dataset. On the other hand, the FFN model has consistent behavior across almost the entire range of the test set and satisfactorily follows the trend of the actual values, which means that it is capable of rational predictions of the required airflow.

FFN Model’s Accuracy and Adjustment

The cross-plot (Figure 4) shows that the trained model is successful and the algorithm has the ability to lead to reliable predictions. More particularly, the FFN’s predicted values as compared with the actual airflow values show a high degree of accuracy. The model’s generalization has an average forecast accuracy of 95%.

The final step is to apply a correction rule in the model’s predictions, so as to always be in compliance with the legislative guidelines of minimum airflow requirements. Besides the fact that the average MAPE is under 4%, this rule actually ensures 100% compliance with the set airflow volume standards at every given time. Figure 5 shows the adjustment of the model for the first 50 predictions. The coding imported into the algorithm is given hereinafter:

$$IfPredictedAirflow<\left(\left(hpofdieselequipment\times 2.3\right)+\left(numberofworkers\times 5.66\right)\right)/60:$$

$$PredictedAirflow=\left(\right((hpofdieselequipment\times 2.3)+(numberofworkers\times 5.66)\left)\right)/60$$

5. Conclusions

Mine ventilation is crucial for the whole mining operation process. In this paper, machine learning (ML) models using artificial neural networks (ANNs) developed in order to adjust the fan operating parameters of the primary ventilation system of an underground bauxite mine based on the ventilation on demand (VoD) method. Via a validated digital twin, different operating scenarios have been modelled to gather the required input data. This synthetic dataset consists of 813 records that were used for the training and the test of the algorithms.

The results of the trained models showed that the FFN has a better generalization than the LSTM regarding the operation of ventilation system based on the size of machine fleet and number of mining workers. Thus, the FFN was able to generalize the anticipated conditions and adjust the fan requirements with a forecast accuracy of 95%. Its consistent accuracy in the whole range tested is also of great importance to potential future use in a mining environment that changes its prevailing conditions in a dynamic manner. Finally, a specific code line set to the FFN model in order to ensure that the airflow predictions and adjustments are in 100% compliance at any given time with the stricter standards in effect in Greece.

The aforementioned process, as presented in this study, can lead to a significant improvement of the SLO (social license to operate) by achieving a reduction of both the environmental footprint and ventilation energy consumption of the mines, but also to ensure an upgrade of work safety level. Of course, more data are required so as to develop more accurate models and expand the working environment of the models so as to include almost all scenarios.