Professional Documents
Culture Documents
ABSTRACT
A methodology for the inversion of a scaled charge weight superposition model to predict peak
particle velocity (PPV) has been examined. The inversion involves two steps: first simulated
annealing is used to get an initial estimate of the model parameters; and second, these parameters
are used as input to a quasi-Newton non-linear optimisation method, used to refine the value
of the parameters. The methodology recovers the three required parameters (two charge weight
scaling parameters and a time window width) and does this directly from a single production
blast for a given point of interest. It has been tested on a synthetic example and recovered the input
model parameters successfully in the presence of introduced noise with a coefficient of variation
ranging up to ten percent. Another test used the vibration measurements of three tunnel blasts
monitored during the construction of the Citybanan tunnel in Stockholm. In this case, the three
model parameters are found from one blast and for a given point of interest, and the PPV at the
same point of interest is predicted for the second and third blast fired in the same vicinity as the first
blast. The three blasts used different firing sequences. The predicted PPV limits use a confidence
interval determined from the scatter in the peak levels obtained from the single calibration blast.
The methodology was able to predict an upper bound for the PPV in 21 of the 26 tunnel blast
vibrations studied. These predictions improve on a previous study involving the same model but
which did not use the inversion process above. Given this reasonable success, further testing of the
method on other data sets and alternative forward prediction models is recommended.
INTRODUCTION
The prediction of peak blast vibration levels is difficult due to where:
the observed variations in measured data for similar blasts. PPV is the peak particle velocity
Spathis (2009) reviews various approaches used for such x is the distance to the point of interest
predictions. The models range from empirical curve fits to
W is the charge weight fired instantaneously or within
measured data such as charge weight scaling laws (Dowding,
an 8 ms window
1985), waveform superposition models that use measured
seed waveforms (Blair, 1999), numerical analyses based on A and B are the constants found by fitting the curve to the
finite elements or analytical solutions (Blair and Minchinton, measurements
2006). One model tested recently (Spathis and Wheatley, A prediction for the maximum peak particle velocity (PPV)
2012) is somewhat between a charge weight scaling law and for a different scaled distance uses Equation 1 directly.
waveform superposition. It uses the primary factors that Similarly, waveform superposition approaches to prediction
affect blast vibrations: charge weight, distance to the point rely on a charge weight scaling law or similar and the details
of interest, initiation time, and travel time to the point of of the blast design such as charge weight, distance to a point
interest. It also requires the standard charge weight scaling of interest, initiation time and travel time. A simulated blast
law parameters plus one other parameter, a time window vibration is predicted by adding suitably scaled and delayed
that tries to capture the interaction and cooperation between waveforms at the point of interest. As implemented by Blair
the vibrations from charges fired within a given time period. (1999), the model can include delay and amplitude scatter
This model is referred to here as RaVE for rapid vibration in a rational manner to produce a distribution of possible
estimation and is described by Spathis (2006). vibration outcomes.
An interesting feature of the models described above is that An alternative approach is to use a set of measured data to
they are what may be termed forward models the input invert for the model parameters at a given point of interest by
parameters of the model are given or determined outside minimising an error term or objective function and obtain an
it and these are used to make a prediction. For example, in objective value subject to relevant constraints. In essence this
charge weight scaling a common approach is to use:
approach uses a forward model to predict the given vibration
-B output features and iterates over the model parameters until
PPV = A c x m
W (1) a minimum error is achieved between the measured and
1. Vacation Scholar, Orica Australia Pty Ltd, PO Box 196, Kurri Kurri NSW 2327. Email: mitchell.lawlor-oneill@uon.edu.au
2. MAusIMM, Senior Research Associate, Orica Australia Pty Ltd, PO Box 196, Kurri Kurri NSW 2327. Email: alex.spathis@orica.com
11TH INTERNATIONAL SYMPOSIUM ON ROCK FRAGMENTATION BY BLASTING / SYDNEY, NSW, 2426 AUGUST 2015 97
M L LAWLOR-ONEILL AND A T SPATHIS
predicted data. The method of finding the model parameters The inversion process we adopted first uses simulated
that result in the minimum is known to be sensitive to the annealing (Bertsimas and Tsitsiklis, 1993) to obtain an initial
nature of the forward model, including the sensitivity of the estimate of the model parameters, and these are used as
model output to changes in the model parameters, and to the starting values for a quasi-Newton method, both implemented
starting values used for the model parameters. in Scilab, version 5.3 (by Scilab Enterprises) (Baudin, Couvert
While the approach described above is quite general, in and Steer, 2010). The advantage of simulated annealing is that
the present work we use RaVE as the forward model that is it searches the parameter space quite widely and hence has a
inverted using a single measured blast vibration. Firstly, we good prospect of finding the vicinity of the global minimum.
describe briefly the formal concept of inversion and the RaVE Once we are near the global minimum the quasi-Newton
model for completeness. We proceed to demonstrate the method commences from that location to rapidly find the
efficacy of the approach using ideal data and extend that to minimum. A description of the two methods may be found in
include data with various levels of noise to mimic some of the Gershenfeld (1999).
variability observed in field measurements of vibrations from The objective function was chosen as the root mean square
blasting. We use the data of Spathis and Wheatley (2012) that (RMS) error between the measured value and the current
involved the prediction of the vibrations from three tunnel inversion estimate at each time step. The RMS error is
blasts fired using electronic delay detonators and recorded at expressed as:
several points of interest. In that work, 26 separate predictions
were made in a round robin approach where the vibration Obj = 1
N / iN= 1 (PPVi - PPV Ri) 2 (3)
measured from one blast at a given point of interest was used
to predict the vibration at the same point of interest for the where:
other two blasts. A calibration method was developed by
Obj is the objective value
scaling average peak levels for the clearly separated holes fired
early in the blast compared to the levels observed from the N is the total number of time steps evaluated
later firing holes. Finally, we conduct inversions for the data PPVi is the measured PPV at time step i
of Spathis and Wheatley (2012) and discuss the effectiveness PPVRi is the RaVE PPV estimate at time step i
of the associated predictions.
The vibration model
PARAMETER ESTIMATION, INVERSION AND The vibration model used here is described by Spathis (2006).
It is a simple extension of the traditional charge weight scaling
THE VIBRATION MODEL law and attempts to account for the influence of the timing
sequence in a blast. It retains the usual two parameters, A
Parameter estimation and inversion process and B, in Equation 1 and introduces a sliding time window
Aster, Borchers and Thurber (2013) provide a concise parameter T. The time window determines the cooperative
description for estimating parameters of a model via inversion. effect of multiple charges firing close in time as seen at a given
We follow their lead and explain briefly the concept of point of interest. The peak vibration level is given by the
parameter estimation or of process inversion used in this paper. maximum value of the following expression involving a
While the methods used for inversion can be challenging, the single convolution (denoted as , )
inversion process may be abstracted as follows where we are
dealing with a so-called discrete inverse problem. We assume PPV = max {A [P (t) ,Q (t)] B} (4)
that a process G acts through a number (vector) of model
parameters m to produce a (vector) data set d: where:
G(m) = d (2) (t) is a unit rectangular function of width T
Q(t) is the chronological sequence of scaled charge weights
In an ideal situation, there are no errors or noise in the data assuming square root scaling that are given by:
and the estimation of the model parameters by inverting Wi
the process is likely to deliver a reasonable estimate of the (5)
xi2
original model parameters. In some circumstances there may
be multiple processes that produce the same data and hence
where:
there is a need to use a good understanding of the process in
order to reduce such ambiguity. That is, it is assumed that the Wi is the charge weight at a distance xi from the point of
process G is understood sufficiently and is well represented. interest.
Of course, real data includes noise, and the inversion may be The exponent B is applied term by term on the resulting
less robust so that the estimation of the model parameters is convolution. An alternative expression for Equation 3 (where
more likely to be in error. the sliding time window process is implied) is:
The so-called forward problem is the application of the B
PPV = max * A f / i x 2i p 4
W
process G given m to find d, usually in the presence of noise. The (6)
inverse problem is to have a known process G and measured i
data d to determine the model parameters m. In our case, the
measured data d is a vibration waveform, or more specifically, The summation in Equation 6 is taken over the rectangular
the vector peak particle velocity versus time measured at a window width of duration T, with the subscript i referring to
point of interest. The process G is the RaVE vibration model. each charge within the sliding window. Using Equations5or6,
The model has just three model parameters m described but excluding the maximum operator, produces a prediction
in the next section. The inversion process is thus a means to across the whole charge sequence and so its prediction can
obtain these three parameters by minimising an error term or be compared across a full vector peak particle velocity signal.
objective function that is a measure of the difference between That is, there is an array of summations implied by Equation6
the prediction and the measured vibration data. and the maximum value is the reported PPV.
98 11TH INTERNATIONAL SYMPOSIUM ON ROCK FRAGMENTATION BY BLASTING / SYDNEY, NSW, 2426 AUGUST 2015
AN INVERSION METHOD FOR THE PREDICTION OF PEAK PARTICLE VELOCITY IN BLASTING
The range of the default values for the model parameters used R,c is the mean peak value greater than the cut-off
in the inversions were: A between 100 and 10000, B between Cpercent of the maximum peak in the RaVE prediction
zero and four, and T between two and 100 (milliseconds). for that blast
The parameter ranges for A and B are based on the typical A predicted maximum PPV upper bound was defined as:
maximum range noted in other charge weight scaling data. The
range for T is based on data from Spathis and Wheatley (2012). PPVmax = s1 n predicted + 3v measured (8)
where:
e,c is the mean peak value for peaks greater than the cut-
off C percent of the maximum peak in the envelope of
the measured vibration data for a given blast FIG 1 Hole timing (in milliseconds) of simulated blast with timing contours.
11TH INTERNATIONAL SYMPOSIUM ON ROCK FRAGMENTATION BY BLASTING / SYDNEY, NSW, 2426 AUGUST 2015 99
M L LAWLOR-ONEILL AND A T SPATHIS
Inversion trials were conducted at five different noise correctT. For the remaining 69 trials, the number of sample
levels, including the case of no noise. For these trials, known points processed was increased to 1000. This provided
parameters A and B were set at 1100 and 1.60 respectively, more accurate results at ten per cent noise, with 65 out of
and T was randomly selected from the range of two to 100 the 69 trials being successful. However, the percentage of
for each trial. Early during this current work it was observed unsuccessful cases remained the same when the number of
that the value of T has a large impact on the shape of the plot, sample points was increased to 1000 at zero percent noise.
hence finding the correct value of T is critical to producing At zero per cent noise, all unsuccessful cases occurred
an accurate fit. The values of A and B had less impact on the for high values of T (between 82100 ms), and the value
shape of the plot, especially at lower values of T, but instead T found by the inversion process was off by 1 ms in each
were important in providing the correct peak scaling. unsuccessful case. In some cases, repeat trials at a given
Successful trials for zero per cent noise were defined as unsuccessful value of T produced the same unsuccessful
those that returned the known values of each of A, B and T. result. The most likely explanation for this is poor sampling
Successful trials for one percent, two percent, five percent in the simulated annealing process. It is noted that high
and ten per cent noise were defined as those that returned values of T result in a larger number of charges involved
the known value of T, and returned A and B such that the in estimating the peak vibration within the blast window,
objective value was within five percent of that found using leading to less noticeable variation in the particle velocity
the known parameter values. It was observed that the optimal versus time plot when compared to lower values of T. In a
solution found by the inversion process provided a better fit real waveform, significant noise is expected, so an error of
to noisy data than the fit provided for known A, B and T 1 ms in the value of T may be considered acceptable. Errors
(designated the no noise function) in many cases. The results in T greater than 1 ms were observed only at ten per cent
of the inversion trials are summarised in Table 1. noise with a lower amount of simulated annealing sampling.
This issue was resolved when the number of sample points
Trials at one per cent, two per cent and five per cent was increased. Sample plots showing a comparison of the
noise were all run with the simulated annealing inversion inversion output, the no noise function using known A, B
processing 1000 sample points for each of the 20 times the and T and the measured (noise added) function for each
temperature was cooled. The initial 31 trials run at both noise level are presented in Figure 2 to Figure 7. Note that
zero per cent and ten per cent noise were run with the Tinv refers to the T found by the inversion process. Also, the
simulated annealing inversion processing 500 sample points Measured data refers to the ideal generated synthetic data
(A, B, T combinations) for each of the 20 times the temperature that would be measured from an ideal process that follows
was cooled. This did not provide successful results at ten Equations 4 or 5.
percent noise of these 31 trials, 18 failed to reproduce the
Phase two Citybanan tunnel, Stockholm
TABLE 1 From the available data for the three tunnel blasts,
Result of phase one inversion trials. 26predictions were possible. These are described by Spathis
and Wheatley (2012). An example of a vector particle velocity
Noise level ( vn ) Total number of trials Successful trials versus time plot is presented in Figure 8. The inversion was
conducted by comparing model predictions and measured
0% 100 (69)
a
81 (56)
data within region E as shown in Figure 8. This region was
1% 100 88 chosen as it contained the PPV in 24 of the 26 trials. Some
2% 100 95 preliminary testing was conducted: inclusion of region D in
the inversion skewed the results by forcing the optimal value
5% 100 100
of T lower to accommodate the individual holes firing; the
10% 100a (69) 78 (65) inversion constantly over predicted the effects of rapid firing
a. One hundred trials were run; however, the first 31 trials were run with fewer simulated annealing in region F, which then forced A and B towards the lowest
iterations. The 69 remaining trials (results shown in parentheses) were run at increased values allowed by the model to produce essentially a single
simulated annealing iterations. horizontal line fit as the best fit. Following these preliminary
A B
FIG 2 Trials at zero percent noise. (A) Trial 72 T = 31 ms, Tinv = 31 ms; (B) Trial 66 T = 99 ms, Tinv = 100 ms.
100 11TH INTERNATIONAL SYMPOSIUM ON ROCK FRAGMENTATION BY BLASTING / SYDNEY, NSW, 2426 AUGUST 2015
AN INVERSION METHOD FOR THE PREDICTION OF PEAK PARTICLE VELOCITY IN BLASTING
A B
FIG 3 Trials at one percent noise. (A) Trial 6 T = 21 ms, Tinv = 21 ms; (B) Trial 32 T = 89 ms, Tinv = 90 ms.
A B
FIG 4 Trials at two percent noise. (A) Trial 18 T = 23 ms, Tinv = 23 ms; (B) Trial 38 T = 63 ms, Tinv = 63 ms.
A B
FIG 5 Trials at five percent noise. (A) Trial 79 T = 38 ms, Tinv = 38 ms; (B) Trial 91 T = 58 ms, Tinv = 58 ms.
tests, the inversion for each blast was conducted in regionE bound was predicted in 21 of the 26 cases. The measure of
and the estimated model parameters were applied to the success is that the envelope of the measured PPV was below
whole blast including all three regions D, E and F. the value given by Equation 8; that is, below the scaled mean
Initial trials presented within this report were run plus three standard deviations. A sample set of predictions
with the scaling peak cut-off parameter C at 60 per cent (Echo2) are shown in Figure 9 to Figure 11. The horizontal
(Equations 7 and 8). For this cut-off level, the PPV upper lines on each plot represent the 68.2percent, 95.4 percent
11TH INTERNATIONAL SYMPOSIUM ON ROCK FRAGMENTATION BY BLASTING / SYDNEY, NSW, 2426 AUGUST 2015 101
M L LAWLOR-ONEILL AND A T SPATHIS
A B
FIG 6 Initial trials at ten percent noise with fewer simulated annealing iterations. (A) Trial 1 T = 44 ms, Tinv = 42 ms; (B) Trial 11 T = 97 ms, Tinv = 94 ms.
A B
FIG 7 Trials at ten percent noise with increased number of simulated annealing iterations.
(A) Trial 40 T = 40 ms, Tinv =40 ms; (B) Trial 50 T = 76 ms, Tinv = 75 ms.
FIG 8 Sample vector particle velocity versus time plot for Blast 3148 at Echo 1. Region D blastholes fired
individually, Region E multiple blastholes fired and Region F rapid blasthole firing.
102 11TH INTERNATIONAL SYMPOSIUM ON ROCK FRAGMENTATION BY BLASTING / SYDNEY, NSW, 2426 AUGUST 2015
AN INVERSION METHOD FOR THE PREDICTION OF PEAK PARTICLE VELOCITY IN BLASTING
A A
B B
C C
FIG 9 Echo 2 predictions using Blast 3148. (A) Blast 3148 initial FIG 10 Echo 2 predictions using Blast 3152. (A) Blast 3152 initial
fit; (B) prediction of Blast 3152; (C) prediction of Blast 3156. fit; (B) prediction of Blast 3148; (C) prediction of Blast 3156.
and 99.7 per cent confidence limits for peak levels, that A comparison of the measured PPV (prior to the envelope
is, one, two and three standard deviations from the mean being applied) and the predicted PPV are shown in Table 3.
predicted level. Two predicted PPVs are given: one that includes the scatter
As noted previously the inversion method was successful in given in Equation 8, and the other that excludes that scatter
predicting the PPV upper bound in 21 of the 26 studied cases, component. Scaling was used in both cases using Equation 7.
with the other five cases falling just outside of the 99.7 percent Figure 12 shows the results for upper bound predictions
confidence level (or three standard deviations). The PPVs that used scaling and scatter. The relative error between the
of the envelope data are all larger than the corresponding measured peaks and predicted upper bounds shows that
measured peak value, although the range of differences in the majority of the predictions lie above the measured data.
the two peaks is non-uniform, lying between approximately Limit lines are shown for -5 percent and 25 percent (a span
threepercent and 36 percent (Table 2). of 30 per cent) and most of the larger errors occur for the
11TH INTERNATIONAL SYMPOSIUM ON ROCK FRAGMENTATION BY BLASTING / SYDNEY, NSW, 2426 AUGUST 2015 103
M L LAWLOR-ONEILL AND A T SPATHIS
TABLE 2
A Difference in measured and envelope peak particle velocity.
40
FIG 11 Echo 2 predictions using Blast 3156. (A) Blast 3156 initial
fit; (B) prediction of Blast 3148; (C) prediction of Blast 3152.
20
between zero percent and -30 percent, the same range as the
dashed limit lines used in Figure 12. Figure 14 reproduces the 0 5 10 15 20 25
104 11TH INTERNATIONAL SYMPOSIUM ON ROCK FRAGMENTATION BY BLASTING / SYDNEY, NSW, 2426 AUGUST 2015
AN INVERSION METHOD FOR THE PREDICTION OF PEAK PARTICLE VELOCITY IN BLASTING
TABLE 3
Comparison of field measured peak particle velocity (PPV) and predicted PPV using both the three standard deviations upper bound and the mean of the scaled peaks.
Trial Monitor Basis blast Predicted blast Measured PPV Predicted PPV level Predicted PPV
(upper bound) (scaled peaks mean)
1 Echo 1 3148 3152 1.14 1.40 1.10
2 Echo 1 3152 3148 0.89 1.30 0.87
3 Echo 2 3148 3152 1.75 2.11 1.69
4 Echo 2 3148 3156 1.94 2.26 1.84
5 Echo 2 3152 3148 1.74 1.82 1.21
6 Echo 2 3152 3156 1.94 2.26 1.64
7 Echo 2 3156 3148 1.74 1.85 1.25
8 Echo 2 3156 3152 1.75 2.00 1.39
9 Echo 3 3148 3152 7.36 8.49 6.07
10 Echo 3 3148 3156 8.79 8.54 6.13
11 Echo 3 3152 3148 6.73 8.80 6.38
12 Echo 3 3152 3156 8.79 8.55 6.14
13 Echo 3 3156 3148 6.73 7.82 5.42
14 Echo 3 3156 3152 7.36 8.82 6.42
15 Echo 4 3148 3152 6.48 8.29 6.10
16 Echo 4 3148 3156 8.28 8.81 6.63
17 Echo 4 3152 3148 6.54 7.30 4.90
18 Echo 4 3152 3156 8.28 8.26 5.87
19 Echo 4 3156 3148 6.54 8.96 6.18
20 Echo 4 3156 3152 6.48 9.67 6.89
21 Echo 7 3148 3152 22.18 26.93 20.60
22 Echo 7 3148 3156 33.58 30.69 24.36
23 Echo 7 3152 3148 20.19 27.56 18.68
24 Echo 7 3152 3156 33.58 32.50 23.62
25 Echo 7 3156 3148 20.19 30.77 20.41
26 Echo 7 3156 3152 22.18 33.69 23.33
PPV peak particle velocity.
40
20
A B
Measured scaled excluding scatter
Predicted (scaled excluding scatter)
(Predicted - Measured)/Measured (%)
10
30
Peak Particle Velocity (mm/s)
0
20
-10
10
-20
-30
0
0 5 10 15 20 25 0 5 10 15 20 25
FIG 13 Measured peaks versus predicted peaks using envelopes without scatter. (A) Raw data; (B) percentage
differences with limits shown as dashed lines at -30 percent and zero percent.
11TH INTERNATIONAL SYMPOSIUM ON ROCK FRAGMENTATION BY BLASTING / SYDNEY, NSW, 2426 AUGUST 2015 105
M L LAWLOR-ONEILL AND A T SPATHIS
80
50
A B
Measured Scaling for cut holes
Predicted
60
(Predicted - Measured)/Measured (%)
40
Peak Particle Velocity (mm/s)
40
30
20
20
0
-20
10
-40
0
0 5 10 15 20 25 0 5 10 15 20 25
FIG 14 Measured peaks (envelopes not raw data as in Figures 12 and 13) versus predicted peaks using forward RaVE model
with T = 50 ms and scaling for cut holes that fired singly without overlap (after Spathis and Wheatley, 2012). (A) Raw data;
(B) percentage differences. Note the data is not plotted in the same order as in used in Figures 12 and 13.
the outlier data occurs at the smaller measured values (note are similar in range, with the Spathis and Wheatley (2012)
the reversed order for the data plotted). predictions having the largest box and overall range.
Figure 15 shows box plots of the relative errors for the data As seen in Figure 12, the predictions with scatter provide
shown in Figures 12 to 14. The box contains 50 percent (two a reasonable upper bound prediction of the majority of
middle quartiles) of the data and the whiskers span the the measured PPVs. Significant deviation is observed in
rest of the data, except any assessed outliers of which there those trials corresponding to Echo 7, which was the closest
is just one that comes from the Spathis and Wheatley (2012) point of interest to the blast and thus had the largest PPVs.
predictions. The solid vertical line in the box is the median In two cases, an under-prediction of less than 3 mm/s was
observed (less than ten percent error), and in the other four
value of the given data. The predictions from Spathis and
cases an over-prediction was observed. Both of these may be
Wheatley (2012) are centred near one per cent whereas, the
considered acceptable.
data with scatter is centred near 16 percent, and that without
scatter near -14 per cent. The boxes for all three data sets In the majority of cases, the inversion fit converged to the
minimum allowable value for A. This is thought to be due to
the majority of the raw data points used in the RMS objective
function being low compared to the relatively few data points
associated with each peak. Having a low value of A limits the
using scatter
106 11TH INTERNATIONAL SYMPOSIUM ON ROCK FRAGMENTATION BY BLASTING / SYDNEY, NSW, 2426 AUGUST 2015
AN INVERSION METHOD FOR THE PREDICTION OF PEAK PARTICLE VELOCITY IN BLASTING
100
levels of normally distributed noise. Phase two involved
Echo 1
Echo 2 applying the inversion technique to measured data from
Echo 3 tunnel blasts to recover values of A, B and T that best fit the
80
approach presented.
The phase one results for a simulated blast design with
different time window parameters show that the inversion
20
11TH INTERNATIONAL SYMPOSIUM ON ROCK FRAGMENTATION BY BLASTING / SYDNEY, NSW, 2426 AUGUST 2015 107
M L LAWLOR-ONEILL AND A T SPATHIS
108 11TH INTERNATIONAL SYMPOSIUM ON ROCK FRAGMENTATION BY BLASTING / SYDNEY, NSW, 2426 AUGUST 2015