4.1. Response Surface Analysis (RSA)
The impact of change in firing patterns on rock fragmentation during blasting can be evaluated properly if other variables of blasting are kept constant. However, it is important to note that in bench blasting there are variations in design pattern due to drilling, charging, and measuring errors. Moreover, there are conflicts in some factors and variables that make perfect modeling difficult. To develop an easy-to-use model for rock fragmentation prediction, response surface analysis (RSA) of the data was carried out while using the variables that most influence the fragmentation. The results of the RSA evaluation (
Table 8) obtained through multivariate non-linear ANOVA method (
Table 9), were finally used in developing the model. A back propagation algorithm being robust in nature was deployed to evaluate the variables over
p-value,
R2, Akaike information criterion (AiCC) and Bayesian information criterion (BiCC), to eliminate insignificant and redundant terms in a quadratic model suggested by the initial analysis.
The Model F-value of 106.42 implies that the model is significant. There is only a 0.01% chance that an F-value this large could occur due to noise and p-values less than 0.0500 indicating that the model terms are significant. In this case, A, B, C, D, BC, and D
2 (see
Table 8 for terms) are significant model terms. Values of
p > 0.1000 indicate the model terms are not significant. The modeling results are presented in
Table 8.
The Predicted R
2 of 0.88 is in reasonable agreement with the Adjusted
R2 of 0.89, i.e., the difference is less than 0.2. Adequate precision, a measure of the signal to noise ratio, should be greater than four. In our case, the ratio of 38.70 indicates an adequate signal. This model can be used to navigate the design space. Several diagnostics were deployed (
Figure 3) before accepting the final equation.
The plot of normal probability of externally studentized residuals follows a straight line (
Figure 3a) indicating a proper transformation of the output, and that the residuals are within the expected ranges (
Figure 3b) with no outliers. The Box-Cox plot for transformation (
Figure 3c) confirms the transformation applied to the output, and all the data in the Cook’s distance (
Figure 3d) are quite well within the limits. The diagnosis thus points to the well-behaved structure of the design and analysis. The predicted vs. observed plot of the mean fragmentation size (
k50) shows that the prediction is quite significant with an adjusted
R2 of 0.89 and predicted
R2 of 0.88. Accordingly, the final equation for prediction of mean fragmentation size (
k50) in terms of the independent variables is provided in Equation (1).
where
Int. is the intercept and equals 0.471 for L, 0.389 for D, and 0.327 for the V-type firing patterns (for other symbols please see abbreviations at the end of this paper).
A comparison of mean fragmentation size predicted by the RSA model given in Equation (1) and measured value is shown in
Figure 4 and confirms that the model can be used for mean fragmentation size (
k50) prediction.
To ascertain the surface response through the model and the interactions between the variables over space, several iterations were conducted while keeping two variables constant and varying one at a time. The results of such simulations are presented in
Figure 5a–i.
Figure 5a–c show the relationship with
RF and
ρee versus mean fragmentation size (k
50) with average (
B × S). Similar trends are observed in all three firing patterns. However, there is significant reduction in fragment size despite the constant variables at average (
B × S) in case of V-type of firing pattern.
Figure 5d–f shows the influence of RF and (
B × S) on the mean fragmentation size (
k50) with an average value of
ρee and have similar trends of fragment size (
k50) irrespective of the type of firing pattern. The figure confirms that the relationship of burden and spacing is not linear with fragmentation and an optimum value is evident for achieving the best possible fragment size. However, a significant change in fragment size is observed in the case of the V-type firing pattern.
The influence of
ρee and (
B × S) on the mean fragmentation size (
k50) with average value of RF is shown in
Figure 5g–i. The trends in all the cases of firing patterns are similar, except the size of fragmentation that is varying over the three firing patterns tested. These figures also provide an optimum value of (
B × S) at which we can achieve smaller fragmentation with the same value of
ρee. There is a marked change in fragment size in the case of diagonal and V-type firing patterns. Distinct trends in the change in fragmentation with variation in (
B × S) and
ρee are, however, apparent from the figures.
4.2. Fragmentation Prediction Using Artificial Neural Network (ANN)
Artificial neural network (ANN) is a computational method consisting of several processing elements that receive inputs and deliver outputs based on their predefined activation functions. ANN consists of three layers, viz., input layer, the hidden layer, and the output layer. The input layer picks up the input signals and transfers them to the next layer and, finally, the output layer gives the prediction. The neural networks have to be trained with some training data to obtain a solution to a complex process output. The ANN and related methods have a capability to solve complicated problems, especially when the process and results are not fully understood [
20]. The case is similar in blasting where the design variables present a complex relationship with rockmass, which in turn has several inconsistencies such as inhomogeneity and anisotropy.
Various algorithms have been suggested for training of the neural network, but the backpropagation algorithm is the most versatile and robust technique and provides the most efficient learning procedure for multy layer perceptron (MLP) networks. An experimental database including enough datasets is required to train the ANN model. Once the training process is completed, prediction can be made for a new input dataset.
Accordingly, to predict rock fragmentation by blasting, a back propagation ANN model was deployed for the data acquired and analyzed earlier by ANOVA. Several iterations were made to find the best possible network and hidden layers. The model that trained well and presented the best results is given in
Figure 6. The plot of training progression thus obtained during the process is given in
Figure 7.
In the above training process, the network is presented with a pair of patterns: an input pattern and the corresponding desired output pattern. The firing patterns can be treated as a string in the ANN training and therefore it is possible to estimate the mean fragment size from the trained network. The network can be queried for such results and hence compared.
Table 10 and
Table 11 show the input parameters and output parameters with their symbols and range, respectively, considered for developing the neural network. For introducing fragmentation to the network, an image analysis method, i.e., “Fragalyst” software was employed to determine muck pile size distribution. The process of fragmentation assessment by Fragalyst software is illustrated in
Table 6. The 50% passing size (k
50) was chosen to determine the fragmentation quality. Out of a total of 92 datasets, 73 datasets were used to train the ANN model and 19 separate datasets (not used in training) were utilized for the purpose of validation thus representing the standard 80:20 ratio.
The training results of the ANN model and the validation results are presented in
Figure 8a,b respectively. The results of the analysis show
R2 and RMSE of 0.96 and 0.040 in the case of training and 0.884 and 0.049 in the case of validation tests. The results point to the fact that the ANN method can be well used for the prediction of fragmentation.
In addition to the above, the sum of the absolute weights of the connections from the input node to all the nodes in the first hidden layer defines the importance of the input variables. The relative importance can thus be worked out from such results. The results of such analysis are presented in
Figure 9.
There are variables in the above list that can be grouped together or represent explosive distribution in a blasthole such as charge length, stemming length, decking length, and sub-drill length that has a significant contribution to fragmentation. This is probably the reason that some of the variables such as B and S assume less importance in ANN but have retained their importance in RSA when several such variables were combined in a single factor ρee.