1. Introduction
In underground mining, the movement of crews is realized by wheeled vehicles. This is due to the long distances between the shaft and the mining sections where the underground crews work. In Polish copper ore mines, almost all vehicles are powered by diesel engines [
1,
2]. In addition to vehicles for transporting people, the largest part of the mine’s fleet consists of trucks and loaders, which transport the material to be excavated. Underground crews face mobility problems due to long distances of up to several kilometres. One of the major ventilation problems is the transport of fresh air to distant worksites. In the deep mines, apart from an increase in air temperature along with the transport distance, air pollution also increases.
According to Dong et al. [
3], underground mines face many natural hazards that affect air quality. One of the most important hazards affecting air quality is the gas hazard. The most dangerous of these to the human body are carbon monoxide (CO), methane (CH
), hydrogen sulphide (H
S), and nitrogen oxides (NOx) [
4,
5,
6,
7]. Gases such as H
S or CH
are gases that are of natural sources. NOx and CO are gases whose presence in mine air is generated mainly by technological processes.
Diesel machinery is considered the source of the highest amount of nitrogen oxides in an underground mine [
8]. The NOx emissions from vehicles have been analysed to develop analytical tools based on ANN for their prediction. A type of Feed-Forward Neural Network, the Multi-Layer Perceptron (MLP), has been used to develop a nonlinear relationship between input and output layers for predicting NOx emissions from load–haul–dumping (LHD) vehicles in underground mines.
The obtained NOx prediction accuracy allows the use of the MLP “soft-sensor” for estimating the environmental impact and to plan the ventilation system demand depending on the number of operating LHDs in the underground mine. The measurement data were analysed from the SYNAPSA system mounted on the LHD machine. This system monitors various parameters, such as engine speed, engine acceleration, fuel consumption, oil temperature, and oil pressure, some of which also measure NOx concentrations. Based on these data, a prediction model was created using the MLP network, which is often used to forecast gas emissions [
9,
10].
To ensure the safety of underground mining crews, it is crucial to have accurate information about the emissions of harmful compounds from mining machinery. The prediction model developed in this study is a valuable tool that can be used to estimate the emissions values of mining machinery that does not have a nitrogen oxide sensor installed. Using the prediction model to estimate emissions values, mining companies can take appropriate measures to ensure the safety of their workers and the surrounding environment.
2. State of the Art
Ventilation in underground mines can be a costly and complex process, as it depends on factors such as the mining technology and the geometry of the mine. Additionally, local air quality can fluctuate, which poses a risk to miners.
Ensuring the safety of underground mining crews is of utmost importance in mining operations. This is reflected in various regulations that impose restrictions on human work in excavations where working conditions may pose a risk to workers’ health or safety [
11,
12,
13,
14,
15]. Examples of such regulations include legal acts that limit the temperature or the concentration of harmful substances in the workplace. These acts also define different types of natural hazards that may affect workers in underground mines [
16].
Adhering to these regulations and taking appropriate measures to mitigate risks is crucial for maintaining a safe work environment for underground mining crews. As mining operations move away from the main underground excavations and intake shafts, the risk of natural hazards in underground mines increases [
17,
18,
19]. Of these hazards, the atmospheric and gas hazards pose the greatest danger. Harmful gases can enter the mine atmosphere naturally, such as methane and hydrogen sulphide, or as a result of technological processes, including carbon monoxide and nitrogen oxides. In recent years, scientists have been analysing the impact of nitrogen oxides (NOx) in underground mines, as they are one of the most harmful gases present. In general, NOx are composed of nitrogen oxide (NO) and nitrogen dioxide (NO
) [
20,
21].
All NOx gas hazards in underground mining operations can be grouped into those that appear naturally or those that appear due to technological processes. The natural sources of NOx in an underground mine are the emission of gas from the rock mass or the oxidation of nitrogen from the atmosphere. However, it should be noted that the largest volumes of NOx are generated in processes related to the technological cycle of the mine: welding work, blasting work, and, above all, diesel-powered machinery [
2]. The primary source of NOx in the mine atmosphere is technological processes, specifically the operation of diesel machinery [
22,
23]. One of the main sources of harmful compounds that are released into the mine atmosphere is diesel-powered machinery. Diesel engines generate many chemical compounds in liquid, solid, and, most importantly, gaseous states. The most important of these are sulphur dioxide (SO
), carbon monoxide (CO), carbon dioxide (CO
), hydrocarbon compounds (CH
), nitrogen oxide (NO), nitrogen dioxide (NO
), and solid particles [
24,
25,
26,
27,
28,
29]. According to Stavert and Lehnert [
30] and Kampa [
31], NOx is a harmful air pollutant that can have significant impacts on both the environment and human health. Both monoxide NO and dioxide NO
are odourless gases. Nitrogen monoxide is a colourless gas, while nitrogen dioxide at a certain concentration can take on a brown colour [
32,
33]. In the exhaust of a diesel engine, the share of nitrogen oxides is about 90% NO and about 10% NO
[
34]. NOx can be a danger to human health, particularly for people with respiratory problems or heart disease. Exposure to high levels of NOx can cause irritation of the eyes and throat, coughing, shortness of breath, and reduced lung function. Long-term exposure to NOx can increase the risk of respiratory infections and chronic lung diseases [
35,
36,
37,
38]. The more toxic gas is nitrogen dioxide [
39], which at 1.5 ppm already causes respiratory problems, while at 5 ppm it leads to a drop in blood pressure. When the NO
concentration reaches about 200 ppm, it can result in human death [
40]. Due to the negative influence of NOx on humans, it is important to monitor the concentrations of nitrogen oxides in underground mines. However, it is even more important to predict the NOx emissions from LHD vehicles and heavy transport trucks, depending on individual working modes. In the case of accurate NOx emissions prediction tool development (a soft sensor), the overall performance of the ventilation system can be optimized and the working atmosphere improved in the underground mine.
Due to the characteristics of NOx and its negative effects on human health, a lot of research is being conducted to improve occupational safety by reducing emissions. Currently, many researchers are working to develop an innovative method to control and predict dangerous gases from diesel engine emissions in underground mines [
41,
42].
To predict the concentration of dangerous gases in underground mines, air monitoring systems can be installed to continuously monitor the concentrations of various gases in the mine atmosphere. Real-time monitoring of gas concentrations can help to identify potential hazards and enable mine operators to take appropriate measures to control the release of dangerous gases [
43,
44,
45]. In addition, mathematical models can be developed to predict the concentration of dangerous gases in the mine atmosphere. These models take into account various factors such as ventilation rates, engine emissions, and the physical characteristics of the mine and can help mine operators anticipate dangerous levels of gases and take appropriate measures to mitigate risks [
46].
The operation of LHD vehicles is characterized by reverse motion with frequent changes from acceleration to deceleration and repeating cycles, as well as quick transitions from one level of engine load to another [
47]. Although the powertrain of these vehicles usually contains a hydraulic torque converter, which reduces torque peaks [
48] and smooths torque demand to the engine, there is a big challenge to meet the environmental legislation because of inertia in the flow rate and temperature in exhaust gas recirculation (EGR) systems [
49]. To tackle this problem, different models for NOx prediction have been developed and tested both in the lab and real conditions.
A simplified physics-based model of the engine is proposed in [
50] to predict NOx emissions by narrow-range inputs. However, fuel combustion in diesel engines is a significantly nonlinear process. Physics-based models cannot account for all influencing parameters due to certain assumptions and simplifications. Therefore, data-driven approaches are also widely used in this domain. The Hammerstein–Wiener model is applied in [
51] but for a small region of operation parameters. In work [
52], NOx emissions from a diesel engine are modelled with a nonlinear autoregressive with exogenous input (NARX) model. Experimental results show that NOx emissions can be estimated with a reduced set of regressors in order to be more stable and robust.
In [
53], the authors proposed a combined grey-box modelling approach with numerical identification of static maps, while the main factors are accounted for by physical assumptions. This model showed a wide range of validity and high accuracy, but the fitting performance in highly dynamic conditions is insufficient.
To capture memory effects, Volterra polynomials were employed in [
54] for identification of nonlinear models of diesel engine emissions. An increasing number of inputs and the degree of polynomials increases the set of estimated parameters and makes this approach difficult for practical application. In work [
55], the authors give a model for NOx and soot emissions in the form of local linear regression models where the parameters are represented in tables. Then, using the
-spline function, they find the parameters of a globally optimal model by solving a linear least-squares problem. However, this work is evaluated only for steady-state engine operations.
The authors of work [
56] estimated NOx emissions of a heavy-duty diesel engine with engine speed and pressure measurements. Principal component analysis (PCA) and L2 regularization techniques are used to derive a simple and reliable estimator. The developed estimator shows sufficient performance in steady-state regimes but improvements are required for transient cycles of engine loading.
An ANN is employed in [
57] to estimate emissions of
, NOx, and PM of a Common-Rail Diesel Injection (CRDI) engine. It was shown that increasing the number of hidden layers and neurons causes over-fitting and decreases the generalization of the model.
In the work [
58], to solve the engine optimization problem, a multi-layer perception (MLP) neural network followed by multi-objective optimization including a non-dominated sorting genetic algorithm II (NSGA-II) and strength Pareto evolutionary algorithm (SPEA2) were used. This study allowed the authors to decide which algorithm is preferable for optimizing engine emissions and fuel consumption. As an alternative to complicated physics-based models, a multidimensional data-driven approach is proposed in [
59] to estimate NOx emissions. Using Deep Neural Networks (DNNs), separate models were developed: engine-out NOx and tailpipe NOx emissions. Two datasets were used from the onboard diagnostic system, namely, an engine dynamometer and a chassis dynamometer. Both the cold/hot Federal Test Procedure (FTP) and the Ramped Mode Cycle (RMC) were applied. The authors proved that high precision of the DNN models (R
= 0.92–0.95 up to 0.99) can be achieved by utilizing minimal engine and exhaust gas after-treatment parameters.
Another factor affecting diesel engine emissions in articulated heavy-duty underground loaders is the depth of the mine. In [
60], engine emissions were determined by a portable gas analyser at various depths from the surface (up to 7000 feet below sea level). Based on the measurement results, the authors concluded that carbon monoxide (CO) and diesel particulate matter (DPM) emissions decrease with depth because of the higher air density and air/fuel ratio for the same parameter set in the vehicle Electronic Control Unit (ECU). Instead, NOx emissions increase with depth. The authors related this to the effect of pressure on in-cylinder NO formation. The influence of ambient air temperature and humidity on NOx emissions is investigated in [
61], which noted that an increase in intake air humidity (in the range of 31–80%) causes a 3–14% reduction in the NOx emissions at a constant temperature of 26 °C. The influence of intake air temperature on engine torque and emissions is analysed in [
62]. They obtained accurate regression models (RMSE 72.38 and accuracy 99.2%) and discovered that the ambient temperature in the range 5–30 °C has a great influence on both the torque and the prediction of NOx. Nevertheless, since the temperature and humidity at a certain depth and geological conditions of underground mines are approximately constant, these factors can be neglected in the prediction model.
3. Measurements
NOx emissions measurements were carried out on a KGHM ZANAM vehicle—more precisely, the LHD LKP-1701 operated in the underground copper ore mine of KGHM Polska Miedź (Poland). This vehicle is depicted in
Figure 1, and its general technical specifications are given in
Table 1.
The equipment is specifically designed to operate in the confined spaces of low tunnels and is equipped with a DEUTZ TCD 12.0 V6, which is a turbocharged diesel engine and SCR system, as shown in
Figure 2. The relationship between diesel engine power and torque can be found in
Figure 3, while a comprehensive list of engine parameters is provided in
Table 2. It is important to note that the parameters listed are taken directly from the manuals provided by the diesel engine manufacturer, and the best point consumption refers to diesel fuel with a density of 0.835 kg/dm
at 15 °C.
Diesel engines are usually operated with an overstoichiometric air-to-fuel ratio to provide the full combustion of soot and to restrict exhausting unburnt fuel. Excess air leads to high NOx emissions. This engine produces 390 kW of power and contains a BlueTec system for exhaust gas after-treatment. This process of NOx emissions reduction is carried out by selective catalytic reduction (SCR) with an ammonia slip catalytic converter and diesel oxidation catalytic converter (DOC). SCR uses Diesel Exhaust Fluid (DEF), or AUS 32 (Aqueous Urea Solution 32%) by ISO 22241. DEF from a special tank is injected into the exhaust pipeline to decompose it to ammonia by the exhaust heat. Inside the SCR catalyst, the ammonia reduces NOx into non-polluting water and nitrogen, which is then released into the atmosphere. The exhaust gas after-treatment unit reads signals from the sensors and transmits them via CAN bus to the engine management control unit: temperature upstream of the SCR catalytic converter; temperature downstream of the SCR; NOx downstream of the SCR; pressure (fluid level) and temperature in the AdBlue/DEF tank; and intake air humidity and temperature.
Table 3 presents the standard parameters of the NOx sensor. However, under current regulations, NOx mass measurements must have a minimum accuracy of ±20% or ±0.1 g/bhp-h [
65], which most NOx sensors cannot achieve under transient load conditions. The causes of this include cross-sensitivity to ammonia (NH
), exhaust gas flow rate, mass air flow (MAF), or sensor position. Additionally, many sensors have noise levels as low as 10 ppm due to residual amounts of NOx in the exhaust system even when the NOx concentration is zero [
66].
Development and validation of a vibration-based virtual sensor are conducted in [
67] for real-time monitoring of NOx emissions from a diesel engine. The virtual NOx sensor is validated on a single-cylinder diesel engine bench. The prediction error was less than ±10% for the steady-state mode and below ±20% for transient conditions. The NOx prediction model is based on principal component regression (PCR). Unfortunately, in this approach, an additional sensor is required to reconstruct the in-cylinder pressure from the vibration signal. The application of this virtual NOx sensor for multi-cylinder engines probably requires more sensors or advanced signal processing techniques.
The onboard monitoring system obtains parameters of LHD vehicle operation and the diesel engine via CAN bus, stores them locally, and uploads data to the enterprise server via a wireless connection once per working shift (about 6 h). Almost all monitored parameters are sampled with a time interval of 1 s and are given in
Table 4.
6. Discussion
The accuracies of the NOx emissions prediction achieved in the training, validation, and testing datasets are given in
Table 5. To check the robustness of the developed ANN-based model, different statistical metrics are used.
The cumulative error of the NOx prediction over time is shown in
Figure 16 for all data. The coefficient of determination seems high—up to 0.86757—and the total deviation of NOx emissions is moderate during half of the working shift (about 3 h).
Using the ANN-based MLP model for the prediction of NOx emissions allows the estimation of the environmental impacts of LHD vehicles working in the underground mine and equipped with the monitoring system. Since not every LHD has a NOx sensor, the results obtained can easily be implemented on a large number of working machines.
The main problem in NOx prediction is to account for the transient modes of LHD bucket filling, acceleration, and deceleration when the engine is subjected to maximum loading and fuel combustion. However, following the analysed dataset, which represents many working cycles, the duration of several outliers in NOx emission is very short (several seconds) compared to the entire time; therefore, the amount of NOx is less than 1%. Therefore, before ANN training, those samples can be rejected, which greatly improves prediction accuracy (up to 85%). The criterion for outlier rejection from the dataset can be a condition of NOx > 3 × STD (standard deviation), i.e., beyond the Gaussian distribution range.
For longer distances of delivery, i.e., fewer LHD cycles per working shift, the accuracy of predictions is expected to be higher due to more-stable modes of diesel engine loading with fewer transient periods. Pieces of blasted material with bigger sizes create a greater load on the engine when the LHD bucket penetrates the hill; hence, heterogeneous material will reduce the accuracy of NOx prediction.
Future research is directed toward improving the prediction of NOx emissions in transient modes of operation of LHD vehicles (e.g., excavation of bulk material and reverse motion). This ANN-based “soft-sensor” can be utilized in ventilation power demand estimation in certain geological conditions of underground mining, which accounts for material granulation after blasting, road inclination, and distance of delivery to dumping points. In addition, trend analysis in NOx emission of certain LHD vehicles can be used for the assessment of individual operator driving qualifications. Further, trends of NOx emissions in certain LHD vehicles over all operators can demonstrate malfunctions in fuel injection or after-treatment systems as well as supplied fuel quality.