Professional Documents
Culture Documents
net/publication/244364108
CITATIONS READS
20 54
2 authors:
19 PUBLICATIONS 569 CITATIONS
National Technical University of Athens
158 PUBLICATIONS 2,036 CITATIONS
SEE PROFILE
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
All content following this page was uploaded by Fotis Rigas on 27 February 2021.
Received 2 June 2004; received in revised form 21 July 2004; accepted 27 July 2004
Abstract
Generation and transmission of blast waves in real terrains is of major importance for risk analysis procedures involving acci-
dental explosion scenarios. The problem arises from the impact of overpressure wave on people and structures that may be lethal
or catastrophic under certain conditions. In this paper, a CFD simulation of shock wave propagation in obstructed terrain is
attempted. Overpressure histories as well as a series of critical parameters, namely the positive and negative peak overpressure,
the arrival time, and the positive and negative phase duration at specific points within the domain were obtained during the simu-
lation. Their comparison with experimental measurements from field-scale high explosive blast tests performed by HSE showed a
reasonably good agreement indicating that CFD computer programs provide reliable tools for estimating explosive shocks in
complex terrains.
# 2004 Elsevier Ltd. All rights reserved.
Keywords: Shockwave; Detonation; Explosion simulation; Overpressure; CFD; Explosives; Complex terrain
Nomenclature
C constant (dimensionless)
Ce1 k–e turbulence model unitless constant equal to 1.44
Ce2 k–e turbulence model unitless constant equal to 1.92
Cl k–e turbulence model unitless constant equal to 0.09
Er energy release rate (J/s)
h specific static (thermodynamic) enthalpy (m2 s2)
htot specific total enthalpy (m2 s2)
k turbulence kinetic energy per unit mass (m2 s2)
p static (thermodynamic) pressure (kg m1 s2)
p0 modified pressure (kg m1 s2)
P total pressure (kg m1 s2)
Pex power of explosion (J/s)
Pk production rate of turbulence (kg m1 s3)
Pkb buoyancy production term (kg m1 s3)
Prt turbulent Prandtl number (dimensionless)
SE energy source (kg m1 s3)
SM momentum source (kg m2 s2)
T static thermodynamic temperature (K)
t time (s)
t0 time of release initiation equal to 0 s
t1 detonation phase duration (s)
tc time constant equal to 1 s
u fluctuating velocity component in turbulent flow (m s1)
u near-wall velocity (m s1)
us friction velocity (m s1)
U velocity vector (m s1)
Ut velocity tangent to the wall (m s1)
V volume of integrated cell (m3)
X, Y horizontal coordinates (m)
y reduced distance from the wall (dimensionless)
Z vertical coordinate (m)
Greek letters
e turbulence eddy dissipation rate (m2 s3)
d the identity matrix or Kronecker delta function (dimensionless)
j the von Karman constant
k dimensionless parameter
leff effective viscosity accounting for turbulence, l þ lt (kg m1 s1)
l molecular (dynamic) viscosity (kg m1 s1)
lt turbulence dynamic viscosity (kg m1 s1)
q density (kg m3)
rj dimensionless turbulence model constant for the k equation equal to 1.0
re dimensionless constant in the j–e turbulence model equal to 1.3
sx the wall shear stress (kg m1 s2)
/ general scalar variable
/0 solution field in the first order Backward Euler scheme
/00 solution field in the second order Backward Euler scheme
Superscripts
T transpose of matrix
S. Sklavounos, F. Rigas / Journal of Loss Prevention in the Process Industries 17 (2004) 407–417 409
2. Numerical methods
2.2. Governing equations
2.1. Building the geometry
2.2.1. Transport equations
This is the first step in a CFD simulation project Empirical models consist of correlations of process
including the definition of the current region geometry, variables and the output to describe a physical
namely the computational domain, in addition to the phenomenon and they are usually derived by para-
sub-division of the domain into a number of smaller meterization of the investigated process on the basis of
control volumes (cells). The geometry comprises a uni- experimental observations. On the contrary, CFD tools
fied set of parametric surfaces built in an appropriate follow a more general deterministic procedure to
CAD interface. approximate a problem: they consider the fundamental
Mesh generation is a crucial process during the prob- Navier–Stokes equations for mass, momentum, and
lem setup, since it affects substantially the final solution heat transfer processes ((2.1)–(2.3), respectively), in
accuracy. In general, the larger the number of cells the conjunction with other partial differential equations for
410 S. Sklavounos, F. Rigas / Journal of Loss Prevention in the Process Industries 17 (2004) 407–417
describing further processes such as turbulence. In The role of the k–e model is to introduce two new
CFX 5.6, the unsteady Navier–Stokes equations in variables into the transport equations: the turbulence
their conservation form are implemented (Versteeg & kinetic energy (k) and the turbulence eddy dissipation
Malalasekera, 1995): (e), leading to the following forms:
@q=@t þ rðqUÞ ¼ 0 ð2:1Þ
Continuity equation
@ðqUÞ=@t þ rðqU UÞ
@q=@t þ rðqUÞ ¼ 0 ð2:4Þ
¼ r½ pd þ lðrU þ ðrUÞT Þ þ SM ð2:2Þ
Energy equation
@ðqhtot Þ=@t @p=@t þ rðqUhtot Þ
@ðqhtot Þ=@t @P=@t þ rðqUhtot Þ
¼ rðkrðTÞÞ þ SE ð2:3Þ
¼ r½krðTÞ þ ðlt =Prt ÞrðhÞ þ SE ð2:5Þ
The main characteristic of explosions is the sudden The mean total enthalpy (htot) is given by:
release of substantial energy amounts heating the sur-
rounding air, which expands and propagates in three htot ¼ h þ 1=2ðU 2 Þ þ k ð2:6Þ
dimensions in wave form. Therefore, a heat transfer where the contribution from the turbulent kinetic
model needs to be used for predicting heat transfer by energy (k) is provided by:
conduction, convection and turbulent mixing. For this
purpose, two models are currently available: the ther- k ¼ 1=2ðuÞ2 ð2:7Þ
mal energy and the total energy models (ANSYS, Momentum equation
2003). It should be noted that, in front of high-speed
@qU=@t þ rðqU UÞ rðleff rUÞ
gas flows (Mach number > 0:2) normally presented in
explosion events, kinetic energy effects on the transport ¼ rp0 þ rðleff rUÞT þ B ð2:8Þ
of enthalpy through the fluid domain become signifi- where B is the sum of body forces and (p0 ) is the
cant and cannot be neglected as for low-speed flows. modified pressure given by
This feature, which constitutes the determinant
between the two models, led to the adoption of the p0 ¼ p þ 2=3ðqkÞ ð2:9Þ
total energy model. Besides kinetic energy effects, this and
model takes into account buoyancy effects arisen by
hot spots in the domain (i.e. due to thermal sources) leff ¼ l þ lt ð2:10Þ
and hence large local density variations.
The k–e model assumes that the turbulence viscosity
is linked to the turbulence kinetic energy and dissi-
2.2.2. Turbulence modeling
pation via the relation:
Generally speaking, turbulence is the random and
chaotic way with which the properties of a general fluid lt ¼ Cl qj2 =e ð2:11Þ
change in three dimensions. Turbulence consists of The values of k and e are directly calculated from the
fluctuations in the flow field in time and space; it is a differential transport equations for the turbulence kin-
complex, actually three-dimensional dynamic physical etic energy and turbulence dissipation rate:
process with a wide range of length scales, which may
have significant effect on the characteristics of the flow. @ðqkÞ=@t þ rðqUkÞ ¼ r½ðl þ ls =rj Þrk þ Pj qe
Turbulence occurs when the inertia forces in the fluid ð2:12Þ
become significant compared to viscous forces and is @ðqeÞ=@t þ rðqUeÞ ¼ r½ðl þ ls =re Þre
usually characterized by high Reynolds numbers.
þ e=kðCe1 Pj Ce2 qeÞ ð2:13Þ
To enable the effects of turbulence to be predicted, a
large variety of statistical models have been developed where Pj is the turbulence production due to viscous
seeking to solve a modified set of transport equations and buoyancy forces, which is modeled using:
by introducing averaged and time varying components
into the original transport equations. In this work, the Pj ¼ lt rUðrU þ rU T Þ 2=3ðrUÞð3lt rU
k–e eddy viscosity turbulence model was utilized. This þ qkÞ þ Pkb ð2:14Þ
family of models suggests that turbulence consists of The k–e model as well as the rest of two-equation
small eddies which are continuously forming and dis- family models offers a good compromise between
sipating, and in which the Reynolds stresses are numerical effort and computational accuracy.
assumed to be proportional to mean velocity gradients. Especially, the k–e model has been implemented in
Both the velocity and length scale are solved using sep- most general purpose CFD codes being one of the
arate transport equations, hence they are named two- most prominent models in turbulence simulation field.
equation models. It has proven to be stable and numerically robust able
S. Sklavounos, F. Rigas / Journal of Loss Prevention in the Process Industries 17 (2004) 407–417 411
to provide good turbulence estimations in explosion integrated equation representing flow processes such as
cases (Pritchard et al., 1996). convection, diffusion and sources. The latter are solved
The k–e model uses the gradient diffusion hypothesis iteratively at nodal points inside each cell aiming at
to relate the Reynolds stresses to the mean velocity minimization of the root mean square (RMS) residuals
gradients and the turbulent viscosity. Within CFX–5.6, (iteration loop error) until the desired convergence be
it utilizes the scalable wall-function approach to satisfied.
improve robustness and accuracy when the near-wall For non-steady state, the transient terms are
mesh is very fine. approximated using the first and second Backward
The wall-function approach employs empirical for- Euler schemes, applicable for constant and variable
mulas that provide near-wall boundary conditions for step sizes. The second order Backward Euler method is
the mean flow and turbulence transport equations. a robust, implicit timestepping scheme and, in contrast
Thus, the following logarithmic correlation applies: with that of first order, is second order accurate in
ðUt =us Þ ¼ ð1=jÞlogðy Þ þ C ð2:15Þ time. However, it is not monotonic and therefore inap-
propriate for some quantities that must remain boun-
The value of the wall shear stress (sx) is obtained from:
ded, such as turbulence quantities. Consequently, the
sx ¼ qus u ð2:16Þ first order Backward Euler scheme applies in turbu-
lence equations. It is robust, as well as the second
where
order scheme, fully implicit, bounded, conservative in
y ¼ ðqu DyÞ=l ð2:17Þ time, and does not create a timestep limitation. The
transient term has no bearing on the steady state sol-
and
ution, but it is only first order accurate in time and
u ¼ ðCl Þ1=4 ðkÞ1=2 ð2:18Þ therefore may induce numerical diffusion in time.
The first order Backward Euler scheme approximates
In the scalable wall-function approach, the (y ) value the transient term as:
used in the logarithmic correlation (2.15) is limited by a ð
lower value of yl ¼ maxðy ;11:06Þ, where 11.06 repre- @ q/ dv =@t ¼ qV ðu u0 Þ=Dt ð2:19Þ
sents the intersection between the logarithmic and the v
linear near-wall profile. As a result, the computed yl
and the second order scheme as:
value is not allowed to fall below this limit and there-
ð
fore, all mesh points are outside the viscous sublayer.
@ q/ dv =@t ¼ qV =Dtð1:5u 2u0 þ 0:5u00 Þ ð2:20Þ
The scalable wall functions allow solution on arbi- v
trarily fine near-wall grids, which is a significant
The numerical technique utilized for the completion
improvement over standard wall functions (ANSYS,
of the advection term discretization was the first order
2002).
upwind differencing scheme (UDS) (Chung, 2002). Dif-
ference schemes in their majority are based on series
2.3. Discretization and solution of the governing expansion approximations (such as the Taylor series)
equations for continuous functions. UDS is very robust, namely
The solution process goes on by discretizing the numerically stable, and is guaranteed to not introduce
three-dimensional spatial domain into finite control non-physical overshoots or undershoots in the solution.
volumes using the mesh constructed in the pre- In contrast with central differencing scheme, UDS
processor stage. The numerical technique employed takes into account the flow direction when determining
in this work was the finite volume method (FVM) the value at a cell face being able to produce reason-
(Versteeg & Malalasekera, 1995), which has been able results even though strongly convective flows
successfully applied in two-dimensional explosion cal- occur in the domain.
culations (Huld, Peter, & Stadtke, 1996). The govern- Several solvers employ a solution strategy according
ing equations ((2.4), (2.5), (2.8)) are integrated over to which the momentum equations are first solved for a
each control volume, such that the relevant quantity guessed pressure and an equation for a pressure correc-
(mass, momentum or energy) is conserved in a discrete tion is obtained leading to a guess-correction process
sense for each cell. Each control volume surrounds a that typically requires a large number of iterations. On
unique node, in which all solution variables and fluid the contrary, CFX 5.6 employs a coupled solver which
properties are stored. Afterwards, the integral equa- solves the hydrodynamic equations for pressure and
tions obtained are converted to a system of algebraic velocity components as a single system. This solution
equations via the substitution of a variety of finite- approach uses a fully implicit discretization of the
difference-type approximations for the terms in the equations at any given timestep and, as a result, redu-
412 S. Sklavounos, F. Rigas / Journal of Loss Prevention in the Process Industries 17 (2004) 407–417
ces the number of iterations required for calculating positioned above them. These conclusions are sub-
the transient solutions in a time-dependent analysis. stantiated computationally as discussed hereinafter.
The simulated experiment belongs to a series of field- The computational domain was built in ANSYS
scale explosion tests performed four years ago, on the Design Modeler CAD tool that accompanies the CFX
Blast Range of the Health and Safety Laboratory at software, with dimensions 80 m 45 m 5 m. The
Buxton, Derbyshire. The main purpose was to study relatively large extent that domain occupies was deliber-
the sheltering effects of solid obstacles presented on ately decided to ensure that the flow profile is not chan-
ground surface during high explosive (HE) detona- ging across the boundaries, but stays far away from any
tions, in association with a subsequent modification to zone of influence, which could be for example a recircu-
TNT methodology establishing it able to account for lation area. Such a situation will probably entail conver-
the presence of obstructions. The obstacles in the trials gence problems and hence solver failure. On the other
were identical equidistant concrete blocks of square hand, excessive dimensions will entail large computa-
cross-section, while the energetic material utilized was tional time for a certain mesh refinement, so the user
the plastic explosive PE4 with a TNT equivalence of should be careful while setting the computational field.
approximately 1.3. The explosive mass had been As mentioned above, mesh generator provide the
formed into a hemisphere and activated remotely ability of mesh adaptation at certain domain regimes.
through a detonator inserted into the top of the HE Because the interest is focalized in the street canyons
load. among the obstacles, a very refined mesh was con-
Overpressure time histories were obtained using structed in this region. The grid consisted of 195 433
piezo-electric gauges, which were positioned in several volume elements (from which 57 768 were prisms) con-
locations among the obstructions along or offset from stituting a total number of 57 146 nodes. The gener-
the central axis. The exact dimensions of the experi- ation of prismatic elements was prescribed for ground
mental arrangement illustrated in Fig. 2 were: and obstacles surfaces, where sharp changes in velocity
L1 ¼ 1:7 m, A ¼ H ¼ 0:6 m, L2 ¼ 1:2 m, B ¼ 8:5 m vectors are expected.
and L3 ¼ 1:8 m. Gauges 1, 2 and 4 were positioned 0.3 The explosion energy release (Er) within the domain
m above ground along central axis; gauge 3 was found was given through a properly adapted step function
along central axis but at 0.9 m height, while gauge 5 at under the form:
0.3 m height and 2.125 m offset.
Er ¼ Pex step½ ðt t0 Þ ðt t1 Þ=t2c ð4:1Þ
Further parameters studied during the experiments
were the positive and negative phase duration, the posi- where Er is the energy release rate (J/s), t the time vari-
tive and negative peak overpressure and the shock able (s), t1 the detonation phase duration (s), t0 ¼ 0 s, tc
front arrival time. Generally, the experimental data ¼ 1 s and Pex the power of the explosion (J/s). Experi-
gave complicated pressure wave forms at measurement mental data for explosive material properties are found
points indicating that complex interactions occur in the literature and can be used for the calculation of
between the air shockwave and the obstructions. More- Eq. (4.1) parameters as shown in Table 1. The assump-
over, gauges positioned at the half height of the blocks tion of the stepwise energy production is a good approxi-
recorded notably lower pressure values than gauges mation of the real process, since dense explosives possess
Table 1
PE4 properties and calculated parameters entered into the step function (4.1)
the attribute of exploding at constant rate, namely with peak values, but for arrival times, too. The code, in
stable detonation velocity (Baker et al., 1983). accordance with the experiment, predicted a re-shock
Before performing the execution of the problem in that follows the initial positive phase yielding a con-
transient form, a steady state run was executed with siderable overpressure. Consequently, once an
zero energy source term for obtaining an initial values explosion event takes place, material structures sur-
file. Afterwards, the transient problem was solved for a rounded by others are subject to more than one shock
number of very small timesteps as the rapid progress of impact, which may result in extensive damages.
the phenomenon impose. The timesteps sequence used On the other hand, the computational wave form
was initially 9 108 s, followed by 9 107 and seems to be smooth compared to that obtained from
9 106 s to include the energy release phase, and then the experiments. Thus, the experimental shape, show-
ing a sharp increase of pressure in the shockwaves, is
2999 105 s until the final time of 30 ms was com-
not reproduced sufficiently well by the numerical pro-
pleted. The convergence criterion was the RMS resi-
duals to be less than 104, while two iterations per
timestep were demanded by the solver to achieve that.
The complete execution required 11 days and 19 h
clock time (CPU time: 11 days and 5 h) on an 800
MHz Intel1 Celeron processor with 512 MB RAM.
Fig. 3. Computational versus experimental overpressure plots at Fig. 5. Computational versus experimental overpressure plots at
gauge 1. gauge 3.
414 S. Sklavounos, F. Rigas / Journal of Loss Prevention in the Process Industries 17 (2004) 407–417
Table 2
Relative errors between computational and experimental positive
peak overpressures
Table 3
Fig. 7. Computational versus experimental overpressure plots at
Relative errors between computational and experimental negative
gauge 5.
peak overpressures
cedure, due probably to the shortcomings of the mod- Gauges Experimental Computational Relative
value (kPa) value (kPa) error (%)
eling or the time-averaging process performed through-
out the calculation procedure. It might be possible that Gauge 1 12.7 12.5 1.6
Gauge 2 6.7 7.1 +5.9
a much more refined mesh would yield a more precise
Gauge 3 2.4 2.6 +8.3
solution to the problem, but this would lead to a sig- Gauge 4 2.5 3.4 +36.0
nificant increase of the computational time. The small Gauge 5 3.4 2.5 23.5
perturbations in the experimental curves appear rather
due to sensor noise.
Moreover, it seems that there is a discrepancy
between computed and experimentally acquired peak Table 4
overpressure referring at monitor point 3. It is possible Relative errors between computational and experimental arrival times
that the code failed to approximate the overpressure Gauges Experimental Computational Relative
rise, still imponderable factors may misquoted the value (ms) value (ms) error (%)
measurement. The presence of signals before the arrival Gauge 1 3.8 3.6 5.3
of the air blast clearly shown on the experimental curve Gauge 2 8.8 8.0 9.1
of Fig. 6, indicates that the shockwaves are also being Gauge 3 17.3 16.0 7.5
transmitted through the ground indeed at a faster rate Gauge 4 17.4 16.3 6.3
Gauge 5 18.0 16.8 6.7
than shockwave propagation in air as expected.
S. Sklavounos, F. Rigas / Journal of Loss Prevention in the Process Industries 17 (2004) 407–417 415
Fig. 8. Overpressure isosurfaces at 5.8 ms (a), and 11.2 ms (b) Fig. 9. Normalized velocity vector plots on the centerline plane at 5
representing the values of 15 500 and 7000 Pa, respectively. ms (a), 10 ms (b), and 16 ms (c).
416 S. Sklavounos, F. Rigas / Journal of Loss Prevention in the Process Industries 17 (2004) 407–417
Fig. 10. Overpressure contour plots on the ground surface at 2 ms (a), 5 ms (b), 7 ms (c), and 13 ms (d).
front surface are liable to damage potential of the same initial positive phase is followed by a strong secondary
order with the centerline position. shock in the regions between the obstacles, which will
likely enhance the effects on nearby structures in acci-
dental explosions. The discrepancies ascertained by the
6. Conclusions comparison of computed results with the experimental
ones, concerned the underestimation of peak over-
In this work, a CFD approach of shockwave propa-
pressures and overestimation of the pulse positive
gation in complex terrain was attempted. The purpose
was to validate computed shockwave parameters with phase durations. Yet, it seems that CFD numerical
experimental data from field-scale dense explosive approach can be effectively used in explosion hazard
trials, in addition to studying the pressure wave spread- assessment procedures providing reliable engineering
ing in the obstructed flow field. simulations for real environments.
The visualization of the results allowed some major
observations to be drawn. First, the percussion of the
pressure wave to the front sides of the obstacles does Acknowledgements
not only imply the reflection of the pressure wave
backwards, but in addition, it spreads alongside preser- This work was supported by the Ministry of
ving at the same time high overpressure levels. Thus, National Education and Religious Affairs (Community
the offset regions of the surface are liable to damage Support Framework 2000–2006) under the HER-
potential similar to that corresponding to the centerline ACHITUS research program.
location. Moreover, the rear faces of the intervening
obstacles are subject to high overpressures, although
they are not directly exposed to the blast wave. This References
may be attributed to the way a pressure wave seems to
transmit in presence of obstacles, specifically, in a ANSYS Corporation. (2002). CFX-5 solver theory manual. Oxford-
direction parallel to the upper obstacle surface as well shire: CFX Ltd.
as alongside the perimeter of the obstacles near the ANSYS Corporation (2003). Heat transfer predictions using
ground surface. advanced two-equation turbulence models. Report CFX-VAL 10/
0404, ANSYS Ltd, Canonsburg.
Overpressure traces were computed in several points Baker, W. E., Cox, P. A., Westine, P. S., Kulesz, J. J., & Strehlow,
within the domain showing good agreement with the R. A. (1983). Explosion hazards and evaluation. Amsterdam:
experimental records. Both of them revealed that the Elsevier.
S. Sklavounos, F. Rigas / Journal of Loss Prevention in the Process Industries 17 (2004) 407–417 417
Baker, W. E., & Tang, M. J. (1991). Gas, dust and hybrid explosions. Mader, C. L. (1996). Numerical modeling of explosives and pro-
Amsterdam: Elsevier. pellants. (2nd ed.). New York: CRC press.
CCPS. (1996). Guidelines for evaluating process plant buildings for Miura, A., Mizukaki, T., Shiraishi, T., Matsuo, A., Takayama, K.,
external explosions and fires. New York: AIChE. & Nojiri, I. (2004). Spread behavior of explosion in closed
CCPS. (2000). Guidelines for chemical process quantitative risk analy- spaces. Journal of Loss Prevention in the Process Industries, 17(1),
sis. (2nd ed.). New York: AIChE. 81–86.
Chung, T. J. (2002). Computational fluid dynamics. Cambridge: Cam- Owens, K. A., & Hazeldean, J. A. (1995). Fires, explosions and
bridge University Press. related incidents at work in 1992–1993. Journal of Loss Prevention
Crowl, D. A. (2003). Understanding explosions. New York: AIChE. in the Process Industries, 8(5), 291–297.
Formby, S. A., & Wharton, R. K. (1996). Blast characteristics and Phillips, H. (1994). Explosions in the process industries. (2nd ed.).
TNT equivalence values for some commercial explosives detonated Warwickshire: IChemE.
at ground level. Journal of Hazardous Materials, 50(2–3), 183–198. Powers, J. M., & Krier, H. (1986). Attenuation of blast waves when
Guoshun, Z. (2000). Causes and lessons of five explosion accidents. detonating explosives inside barriers. Journal of Hazardous Mate-
Journal of Loss Prevention in the Process Industries, 13(3–5), 439–442. rials, 13(2), 121–133.
HSE (2001). Explosion hazard assessment: a study of the feasibility Pritchard, D. K., Freeman, D. J., & Guilbert, P. W. (1996). Predic-
and benefits of extending current HSE methodology to take tion of explosion pressures in confined spaces. Journal of Loss
account of blast sheltering. Report HSL/2001/04, Sheffield. Prevention in the Process Industries, 9(3), 205–215.
Huld, T., Peter, G., & Stadtke, H. (1996). Numerical simulation of TNO (1989). Methods for the determination of possible damage.
explosion phenomena in industrial environments. Journal of Haz- Report CPR 16E, Voorburg.
ardous Materials, 46(2–3), 185–195. Versteeg, H. K., & Malalasekera, W. (1995). An introduction to com-
Lees, F. P. (1996). Loss prevention in the process industries. (2nd ed.). putational fluid dynamics—the finite volume method. New York:
Oxford: Butterworth-Heinemann. Longman.