Predicting blasting propagation velocity and vibration frequency using artificial neural networks

https://doi.org/10.1016/j.ijrmms.2012.05.002Get rights and content

Abstract

We describe artificial neural networks used to predict the velocity and frequency of ground vibrations caused by blasting in an open-pit mine. The aim was to predict peak particle velocity and frequency of ground vibrations from information on the physical and mechanical properties of the rock mass, the characteristics of the explosive and blasting design. Some the parameters that could possibly have a bearing on the prediction were considered. A distinction was drawn between two kinds of parameters: those defining the surroundings in which the wave is propagated (rock type, rock mass, distance to be covered by the wave and significant subsoil discontinuities) and those defining the energy of the wave (the kind of explosive, explosive charge and blasting geometry and sequence).

Vibrations were monitored using seismographs capable of capturing vibration data and transforming them into acceleration and frequency terms. To validate this methodology, the predictions obtained were compared with those obtained using conventional statistical methods. The correlation coefficients obtained for our methodology was 0.98 for peak particle velocity and 0.95 for frequency, compared to 0.50 and 0.15, respectively, for Multiple Linear Regression.

Section snippets

Introduction and background

The growth in open-pit mining, motivated by a growing demand for minerals (especially aggregates, fuel and ornamental stone), has led to a considerable increase in the use of explosives for blasting purposes. Explosives are an efficient source of energy for breaking up and excavating rock. An explosive detonated inside a blasthole immediately releases a massive amount of energy in the form of pressure and temperature. Although there have been important advances in explosive technology [1], [2],

Geological description of the area

The study zone is located close to the town of Langreo, near the city of Oviedo in northern Spain, in a quarry operated by the Bahoto Mining Company. The mine is accessed from kilometre 9 of the AS-243 (Oviedo–Frieres) road. Fig. 1 shows the location of the study zone, where the highest elevation is Peña Villa (591 m).

Bahoto quarry is over a limestone mountain in the Valdeteja formation. Fig. 2 is a detailed geological map of the western part of the formation where the open pit is located. A red

Objectives and measuring equipment

Our aim was to study how the vibrations transmitted by blasting in an open-pit mine were transmitted in different directions, so as to determine, in turn, how the vibrations affected the different structures located in the quarry, immediately outside the quarry limits and in nearby villages. We selected four zones, depicted in Fig. 3 and labelled ‘Old installations’, ‘Offices’, ‘Quarry’ and ‘Pilot mine’. As can be observed, the equipment was located so as to account for the main blast

Neural network architecture

The nature and intensity of the vibrations and the induced frequency in the blasting terrain depend on many factors [13].

One group defines the characteristics of the surroundings. (a) The type of rock and rock mass (rock mass rating, RMR), as each terrain has a dominant transmission velocity and frequency that favours wave propagation for that frequency [14]. The presence of families of discontinuities and their characteristics (essentially, aperture, fill and water) also affect transmission.

Network training and learning

ANNs need to be trained before they interpret information. Of the many neural network training algorithms available, the most versatile and robust algorithm is the FFBP algorithm, which is especially effective for networks with several layers. FFBP algorithms are especially capable of resolving prediction problems, for which reason they are particularly useful for identification problems. The FFBP neural network is always formed of an input layer, a hidden layer and an output layer [16].

Each

Network calibration and training validation

To test and validate the ANN methodology, we selected sets of new input–output data. These data were not used in training the network. They thus validate use of the ANN methodology in a more versatile way. The set of twenty data items used were from the five validation tests summarized in Table 6. Fig 9.

Our results demonstrate the quality of prediction using the ANN methodology [17]. The percentage of mean relative error and the correlation coefficient between the predicted and observed values

Conclusions

Using Bayesian interpolation and an optimum number of hidden layer neurons, the mean relative errors for the prediction of PPV and F were 0.65 and 2.77, respectively. The corresponding correlation coefficients were 0.98 and 0.95, respectively, indicating a satisfactory prediction for PPV and F using the ANN methodology.

The prediction made using multiple linear regression produced mean relative errors of 2.77 for PPV and 3.27 for F. Furthermore, the correlation coefficients obtained were very

References (17)

There are more references available in the full text version of this article.

Cited by (0)

View full text
Copyright © 2012 Elsevier Ltd. All rights reserved.