PSIG 0301

Leak Detection and Locating – A Survey

Gerhard Geiger, University of Applied Sciences Gelsenkirchen, Faculty of Electrical Engineering, Neidenburger Str. 10,
45877 Gelsenkirchen, Germany, gerhard.geiger@fh-gelsenkirchen.de

• API 1130 2nd Ed. [2] (USA), and

• “Technische Regeln für Fernleitungen” (TRFL) (Techni-

ABSTRACT cal Rules for Pipelines) [3] (Germany).
Many fluids transported by pipelines are in some sense haz-
ardous. It is therefore often necessary to install leak detection API 1130 2nd Edition (USA)
(and locating) systems (LDS), especially due to legal regula-
tions like the “Code for Federal Regulations (CFR) Title 49 The 2nd Edition of API 1130 “Computational Pipeline Moni-
Part 195” [1], API 1130 2nd Ed. [2], both for the USA, or the toring for Liquid Pipelines” was published from the American
“Technische Regeln für Fernleitungen” (TRFL) (Technical Petroleum Institute in November, 2002, [4]. Other regulations
Rules for Pipelines) in Germany [3]. like the “Code for Federal Regulations (CFR) Title 49 Part
195” [1] refer to API 1130. The API 1130 focuses on the de-
This paper gives a survey of methodologies, methods and sign, implementation, testing and operation of CPM systems;
techniques for leak detection and locating, 4] and [5] are two it is limited to single-phase liquid pipelines.
other interesting sources giving an overview of the topic.
API 1130 defines a CPM-system as an “algorithmic approach
The survey starts with some remarks concerning (legal) regu- to detect hydraulic anomalies in pipeline operating parame-
lations both for the USA and for Germany. Some few words ters”. The technical overview section introduces to method-
about externally based systems (due to API 1130 2nd Ed.) fol- ologies of CPM-systems, classifying them into
low next. A significant part of the paper deals with internally
based systems (also due to API 1130 2nd Ed.) like balancing • externally based leak detection systems, and
systems (line balance, volume balance, compensated mass
balance etc.), Real Time Transient Model LDS (RTTM-LDS), • internally based CPM systems,
pressure/flow monitoring and statistical analysis LDS. Differ-
ent methods for leak locating (gradient intersection method, see Fig. 1.
wave propagation analysis etc.) will also be shown. Externally based systems. Externally based systems use local
The presentation of an Extended RTTM approach (E-RTTM) sensors, generating a leak alarm. System costs and complexity
of installation usually are high; applications therefore are lim-
combining advantages of conventional RTTM LDS and statis-
tical analysis follows next, together with the demonstration of ited to special high-risk areas, e.g. near rivers or nature protec-
tion areas. Examples for such a type of LDS are acoustic
applicability by means of two examples, a liquid multi-batch
pipeline, and a gas pipeline. emission detectors monitoring noise levels and location and
vapor sensing cables, sensing gas or hydrocarbon vapor near a
Sketching future work and the conclusion conclude the survey. leak.

Internally based systems. Internally based systems utilize

REGULATORY FRAMEWORK field sensors (e.g. for flow, pressure and fluid temperature) to
monitor internal pipeline parameters. These field signals are
Companies operating pipelines transporting hazardous fluids used for inferring a leak. The classical line balance method
(e.g. liquids or gases) often have to consider a dedicated regu- balancing inlet and outlet volume flow is an example.
latory framework. Examples are
Other sections of the API 1130 cover topics like
• Code for Federal Regulations (CFR) Title 49 Part 195 [1]
(USA) • field instrumentation, SCADA/Communication, data pres-
2 GERHARD GEIGER, THOMAS WERNER, DRAGO MATKO PSIG 0301

presentation,
EXTERNALLY BASED SYSTEMS
• operation, maintenance and testing, and
Local leak sensors of externally based systems generate a leak
• descriptions of types of internally based CPM systems. alarm which e.g. can be evaluated by SCADA-systems. This
kind of LDS is characterized by a very good sensitivity to
leaks and is very accurate with respect to the leak locating. On
TRFL (Germany) the other side, system costs and complexity of installation
usually are high; applications therefore are limited to special
TRFL is the abbreviation of “Technische Regel für Fernlei- high-risk areas, e.g. near rivers or nature protection areas.
tungsanlagen” (Technical Rule for Pipeline Systems); it is the
successor of the TRbF 301, “Technische Richtlinie für
brennbare Flüssigkeiten” (Technical Guideline for flammable Acoustic emission detectors
Liquids). Up to now, the TRFL is not officially published; the
actual proposal dates from June, 2002. Nevertheless, all new Escaping liquids creates an acoustic signal as it passes through
pipeline systems must comply with the TRFL, which therefore a perforation in the pipe. Acoustic sensors affixed to the out-
in fact is valid. side of the pipe monitor internal noise levels and location,
creating a baseline acoustic “fingerprint” of the line. When a
The TRFL summarizes requirements for pipelines being sub- leak occurs, the resulting low frequency acoustic signal is de-
ject of official regulations. It covers tected and analyzed. Deviation from the baseline “fingerprint”
would signal an alarm [6]. The received signal is stronger near
• pipelines transporting flammable liquids, the leak site thus enabling leak locating, see below.

• pipelines transporting liquids being dangerous for water, Fiber optic sensing cables
and
The fiber optic sensing leak detection method involves the
• many pipelines transporting gas. installation of a fiber optic cable along the entire length of the
pipeline, Fig. 3. The substances to be measured come into con-
The TRFL is divided into the two parts “Operation” and tact with the cable in case of a leak occurrence, changing the
“Constitution” of pipeline systems. Only a small part is con- temperature of the cable. The distributed fiber optical tempera-
cerned with leak detection and locating, focusing to the speci- ture sensing technique offers the possibility to measure
fication of measures necessary in order to detect and locate temperature along the pipeline.
leaks. Five different LDS and LDS functions are required1, see
Fig. 2: This technique is based on the Raman Effect and the so-called
Optical Time Domain Reflectometry (OTDR). A pulsed laser
• Two independent LDS for continuously operating leak is coupled into the optical fiber which is the sensing element.
detection during steady state operation. One of these sys- In the fiber the photons interact with the molecules of the fiber
tems or an additional one must also be able to detect leaks material. Therefore, the laser light is scattered as the laser
during transient operation, e.g. during start-up of the pipe- pulse propagates through the fiber owing to changes in density
line. and composition of the fiber as well as to molecular and bulk
vibrations. Some of the photons are scattered backwards. The
• One LDS for leak detection during stand-still operation. spectrum of the backscattered will be analyzed. The spectral
analysis is combined with measuring propagation time of the
• One LDS for creeping leakages. laser pulses along the fiber (radar principle) because the veloc-
ity of the light in the fiber is known. Scanning the entire length
• One LDS for fast leak locating. of the fiber by short intervals (e.g. 3ft) the temperature profile
along the fiber is determined, leading to leak detection and
The TRFL is focused to general requirements necessary to
locating [7].
detect and locate leaks; it gives not much information about
technical aspects like design or implementation of LDS.
Vapor sensing cables
There are many possibilities to classify externally and inter-
nally based systems; we want follow the API classification The vapor sensing tube leak detection method involves the
scheme [4]. installation of a tube along the entire length of the pipeline,
Fig. 4. This tube is highly permeable to the substances to be
detected in the particular application. If a leak occurs, the sub-
stances to be measured come into contact with the tube in the
1
It is partially possible to combine different LDS functions form of vapor, gas or dissolved in water. The tube is pressure-
into one LDS device. tight and is filled with air (at atmospheric pressure) when in-
PSIG 0301 Leak Detection and Locating – A Survey 3

stalled. In the event of a leak, some of the leaking substance with mass flow M i and M O at inlet (I) and outlet (O), re-
diffuses into the tube due to the concentration gradient. After a
spectively, and the mass M L stored in the line. In general, line
certain period of time, the inside of the tube produces an accu-
rate image of the substances surrounding the tube, regardless fill2 M L for a pipeline of length L changes over time due to
whether the tube is installed in air, water or in the ground. changes of the fluid density ρ and cross-sectional area A ac-
cording to
In order to analyze the concentration distribution present in the
sensor tube, a pump pushes the column of air in the tube past a L L
dM L d d
detection unit at a constant speed, thus recording the measured
level as a function of the pumping time. The concentration dt
=
dt ∫ ρ ( x) A( x) dx = ∫
0 0
dt
ρ ( x) A( x) dx (2)

profile is not affected by the pumping action. The detector unit

at the end of the sensor tube is equipped with gas sensors. with position coordinate x, 0 ≤ x ≤ L . ρ changes according
Every increase in gas concentration results in a pronounced to the relation ρ ( x) = ρ ( p ( x), T ( x) ) (corresponding to the
"leak peak". The height of the peak is proportional to the con-
centration of the substance and is therefore an indication of the thermodynamic state equation p = p ( ρ , T ) ), whereas A
size of the leak (a small leak produces a small peak and a large changes according to the relation A( x) = A( p ( x), T ( x) ) as a
leak produces a large peak). result of the pipeline deformation.
An electrolytic cell at the end of the detected line is used to Line balance. If ρ and A are assumed to be constant,
inject a specific volume of test gas prior to each pumping ac-
tion. This gas is transported through the entire length of the
ρ = ρ ( p, T ) = ρ L and A = A( p, T ) = AL , then dM L dt = 0
sensor tube together with the air. When the test gas passes and eq. (1) can be rewritten into
through the detector unit, it generates a marking peak or end
peak. Its arrival serves as a control marker to indicate that the M I (t ) − M O (t ) = 0 (3)
entire air column contained in the sensor tube has passed
through the measuring station. The end peak is thus an indica- Additionally assuming equal and constant densities
tion of the overall length of the sensor tube. Based on the ratio ρ I (t ) = ρ I = ρ L and ρ O (t ) = ρ O = ρ L for inlet and outlet
of the travel time of the leak peak to that of the end peak, the
leak location can be accurately calculated [8]. mass flow, respectively, and introducing volume flow V with
M = ρV , we finally get
Liquid sensing cables
V I (t ) − VO (t ) = 0 . (4)
Liquid sensing cables are buried beneath or adjacent to a pipe-
line and are specifically designed to reflect changes in trans-
The estimated imbalance3 R(t ) ≡ V I (t ) − VO (t ) can be com-
mitted energy pulses as a result of impedance differentials
included by contact with hydrocarbon liquids. Safe energy pared with
pulses are continuously sent through the cable. The pulses are
reflected and a baseline reflection “fingerprint” is measured. < A ⇒ no leak
R  (5)
≥ A ⇒ leak
When a leak occurs, the cable is saturated with fluid, altering
the impedance of the sensing cable, which in turn alters the against a threshold A to evaluate the leak alarm. This is the
reflection pattern returned. Deviation from the baseline “fin- classical line balance method.
gerprint” would signal an alarm. Measuring time delay be-
tween input pulse and reflected pulse enables leak locating. Volume balance. Using a representative bulk modulus of
Specific cable types are chosen for each application based on elasticity K, assuming a pipeline without deformation and us-
the specific fluid being monitored [5]. ing average line pressure p L as well as an average line fluid
temperature T L allows for a simple line fill calculation
INTERNALLY BASED SYSTEMS M L = M L ( p L , T L ) and V L = M L ρ L with some kind of av-
erage density ρ L of the fluid along the pipeline. The esti-
Balancing systems mated imbalance R can be enhanced to
These kinds of LDS use the principle of mass conservation:
Mass is conserved if there is no leak:

dM L
M I (t ) − M O (t ) = (1) 2
Terms like „line pack“ or „pipeline inventory“ are also used.
dt 3
Also referred as „residual“.
4 GERHARD GEIGER, THOMAS WERNER, DRAGO MATKO PSIG 0301

dV L momentum equation
R (t ) ≡ V I (t ) − VO (t ) − (6)
dt
dv 1 ∂p
+ ⋅ + fD = 0 , (11)
Volume balance permits very simple line pack elimination. dt ρ ∂x

Compensated mass balance. Returning to eq. (1), a more and energy equation
rigorous approach leads to the residual definition
dh 1 dp
dM L − ⋅ − lL = 0 . (12)
R(t ) ≡ M I (t ) − M O (t ) − (7) dt ρ dt
dt
Remarks:
with dM L dt according to eq. (2). Dividing the pipeline into
n segments where p, T and ρ can be regarded as uniform re- • These equations describe (for simplicity) one-dimensional
(location x) transient (time t) single-phase fluid (liquid
sults in the line fill calculation
and gas) flow in a single pipeline segment without diffu-
L n sion. It is a hyperbolic type PDE system.
M L = ∫ ρ ( x) A( x) dx ≈ ∑ ρ i Ai . (8)
0 1 • Appropriate thermodynamic state equations p = p ( ρ , T )
and h = h( ρ , T ) are required to eliminate enthalpy h ,
with uniform segment density ρ i . For liquids (especially oil leading to three equations with three unknowns ρ , v and
products), ρ i can be calculates using the thermodynamic rela- p.
tion
• Drag force f D per unit mass includes force of gravity per
ρ = ρ ( p, T ) = C p ( p − p 0 , T − T0 ) ⋅ C T (T − T0 ) ⋅ ρ 0 (9)
unit mass and friction force per unit mass: f D ≡ f G + f F .
with volume correction factors for pressure Cp and CT for tem-
• Losses l L per unit mass includes heat flow per unit mass
perature [9]. ρ 0 denotes the liquid density at some reference
point p0, T0. Using eq. (9) together with the cross-sectional and dissipative losses per unit mass: l L = lQ + v ⋅ f F .
area A = A( p, T ) (see [9]), line fill eq. (8) can be computed
numerically providing segment variables pi, Ti, e.g using addi- • The computation of heat flow per unit mass lQ requires
tional measurements or a steady-state model of the pipeline. an additional thermal model.

Model compensated mass balance. Using Real-time Tran- • Special model extensions could be required for multi-
sient Models (RTTM) allow for the computation of density phase conditions and other non-standard conditions (slag-
ρ ( x, t ) along the pipeline, see below [4]; model inputs are a line, usage of drag reducing agents, DRA).
choice of variables pI, M I ( V I ), TI at inlet and pO, M I ( V I ), Up to now, there is no known analytical solution. Therefore,
TO at outlet. Therefore, line fill ML in eq. (8) can be computed numerical algorithms have to be used instead [11]. Examples
using model output ρ ( x, t ) . From a theoretical point of view, are method of characteristics (MOC) and finite difference,
this is the most rigorous balancing approach. finite volume and finite element methods.

The resulting numerical algorithm is of boundary value prob-

RTTM lem type with three boundary conditions required.
Using the increasing computing power of modern digital com- Friction force. Using Darcy-Weisbach equation, friction force
puters, it is possible to calculate in real time the profiles for per unit mass fF for a pipeline with diameter D is given by
flow v4, pressure p and density ρ (or temperature T) along the
pipeline. This requires solving a partial differential equation vv
(PDE) system as a result of applying continuity equation [10] fF = f (13)
2D
dρ ∂v
+ ρ ⋅ = 0, (10) with Darcy-Weisbach friction factor f [12], often calculated
dt ∂x for turbulent flow using Colebrook-White formula [13]

4
Relation to volume flow is given by V = A ⋅ v .
PSIG 0301 Leak Detection and Locating – A Survey 5

 1 ε  using friction force per unit mass fF according to eq. (13). κ is

1 2.523
= −2 ⋅ log +  (14) the effectiveness of the DRA, which depends on the DRA
f  3.7 D R ⋅ f  concentration.
 

with Reynolds number R and roughness height ε ; see [13] Alternative approaches. Transfer function models for the
and [14] for alternative turbulent flow formulas. For laminar PDE system eq. (10-12) are obtained by linearizing these
flow, eq. (14) simplifies to equations and carrying out a Laplace transformation. The re-
sulting transfer function is transcendent. Simple models of the
64 pipeline in the form of a lumped parameter system can be ob-
f = . (15) tained by a Taylor series expansion of transcendent transfer
R
functions. The resulting algorithms are less time-consuming
State equations. It is interesting to see, that the thermody- and hence better suited for critical real time applications [18].
Use of Neural Nets (NN) presents another possibility for sys-
namic state equations p = p ( ρ , T ) and h = h( ρ , T ) are the
tem modeling using a black box approach: trained by field
only fluid specific equations in eq. (10-12). The simplest ap- data, NN are able to describe the pipeline behavior without
proach for an ideal gas results in [15] any knowledge about pipeline physics. NN are of special in-
terest [19]
R
p = p( ρ , T ) = 1⋅ ⋅ ρ ⋅T (16)
M • for pipelines with complex physical behavior, where a
physical description is time consuming (or maybe not
with gas constant R and molecular weight M, and possible to find), or

 ∂h  • as an addendum to conventional RTTM approaches ac-

dh =   dT = c p dT (17)
 ∂T  p cording to eq. (10-12)

LDS using RTTM. Using and solving eq. (10-12) in real

with constant-pressure specific heat cp; see [16] for a further
time, it is possible to eliminate transient effects introduced by
discussion of improved state equations p = p ( ρ , T ) .
• fluid compressibility and pipe wall elasticity, and
Simplifications. Simpler mathematical models can be derived
using additional assumptions like adiabatic or isothermal flow • temperature dependence of the density.
[17], sometimes reducing the order of the PDE system eq. (10-
12). For liquids, neglecting heat transfer and the conversion of Early proposals using RTTM for LDS purposes can be found
frictional work into thermal energy leads the water hammer in [20] and [21]. Corresponding LDS are called real time tran-
equations [12], a PDE 2nd order PDE system. sient model (RTTM) systems, [2]. RTTM-LDS can also be
used during transient pipeline operation, e.g. during start-up of
Multi-phase flow, slack line. Condensation from gas into a pipeline; this is especially useful for gas pipelines, where
liquid often can be observed in gas pipelines from off-shore large compressibility results in severe transients. Two possi-
wells, resulting in a two-phase gas-liquid flow. Condensate bilities for using mathematical model information are given
pipelines that are liquid only due to pressurization or crude here:
pipelines that run through mountainous terrain exhibit the op-
posite problem: slack line, introducing some volume of gas in Deviation analysis: Only three boundary conditions are re-
the pipeline. Both phenomena have a significant impact on the quired to drive the numerical solution algorithm, e.g. p I (t )
hydraulic operation of the pipeline [11]. The modeling of and pO (t ) for inlet and outlet and ρ I (t ) (or TI (t ) ) for inlet.
multi-phase flow requires the introduction of multi-component
transport and the model being capable of performing the indi- If more measurements are available (e.g. flows v I (t ) and
vidual boiling or condensation of each of the individual com- vO (t ) or additional pressure measurements along the pipe-
ponents. line), these measurements can be compared with the simulated
values. If there is a significant deviation, leak alarm will be
Drag reducing agents, DRA. DRAs improve the delivery given. See the new RTTM LDS approach given later for fur-
capability and reduce the cost of pressurization by reducing ther details.
the pressure drop per unit length of pipeline. The impact of
DRAs can be modeled by [11] Model Compensated Mass Balance: The RTTM can be used
to calculate the line fill
fvv
f F′ = κ ⋅ f F = κ ⋅ (18) L
2D
M L = ∫ ρ ( x) A( x) dx . (19)
0
6 GERHARD GEIGER, THOMAS WERNER, DRAGO MATKO PSIG 0301

according to eq. (2) in real-time. The imbalance subsequently tively. Pressure drop ∆p is estimated using
can be compared with
∆pˆ (k ) = µˆ 0 (k ) − µˆ 1 (k ) . (26)
< A ⇒ no leak
R  (20)
≥ A ⇒ leak Appropriate alarm thresholds α arise from the fact, that

against a threshold A to evaluate the leak alarm. ∆pˆ

t≡ . (27)
N 0 − N1 σ
⋅
Pressure/flow monitoring N 0 −1 N1
Sudden leaks result in sudden changes of pressure and flow at
inlet and outlet, respectively. Usually the pressure drops as a is Student-t-distributed, as described in [22].
result of a leak.
Statistical analysis
Wave alert method. Applying pressure transducer along the
pipeline, the negative pressure drop ∆p can be observed as a The degree of statistical involvement varies widely with the
wave propagating with wave speed a through the pipeline, different methods in the API classification of internally based
downstream and upstream with respect to the point of the leak. systems. In a previous section, we investigated the Pressure
Assuming isentropic flow without friction, the pressure wave Point Analysis (PPA), which has been assigned to pres-
amplitude is given by [12] sure/flow monitoring methods; an alternative assignment to
statistical analysis methods would also be possible.
∆p = − ρ ⋅ a ⋅ ∆v (21)
Another example is the LDS described in [23], basing on the
where ρ denotes fluid density and ∆v describes the flow Wald’s Sequential Probability Ratio Test (SPRT). The imbal-
ance
amplitude caused by a sudden leak. Leaks will be detected
computing
R(k ) ≡ M I (k ) − M O (k − τ ) (28)
dp ∆p  ≤ − A ⇒ no leak
≈  . (22) will be used and checked for two hypothesizes:
dt ∆t > − A ⇒ leak
H 0: R ( k ) ∝ N ( µ 0 ; σ )
The wave alert permits leak locating as shown below. . (29)
H 1: R (k ) ∝ N ( µ 0 + ∆µ ; σ )
Pressure point analysis. The Pressure Point Analysis (PPA)
[22] utilizes the pressure drop produced by a leak, see Fig. 5.
Inlet mass flow M I (k ) and outlet mass flow M O (k ) are sam-
Two time windows
pled, with discrete-valued time k ≡ k ⋅ T0 for sampling time T0.
p 0 (k ) ≡ [ p(k − N 0 + 1) … p (k )] The time shift τ ≡ τ ⋅ T0 considers the time for a (flow) wave
(23)
p 1 (k ) ≡ [ p(k − N 1 + 1) … p(k )] needed to propagate with wave speed a through a pipeline
with length L:
with window length N0 and N1 define two populations,
N 0 > N 1 . p(k) is the actual sampled pressure for discrete- L
τ= . (30)
valued time k ≡ k ⋅ T0 for sampling time T0. Each population is a
assumed to be Gaussian distributed, N ( µ 0 ; σ ) and N ( µ 1 ; σ ) ; Wald’s SPRT results in a sequential computation of the likeli-
hood ratio
p(k ) − p (k − N 0 + 1)
µˆ 0 (k ) = µˆ 0 (k − 1) + . (24)
N0  N ( µ 0 + ∆µ ; σ ) 
λ = ln 

 N (µ 0 ; σ )  . (31)
and
∆µ  ∆µ 
λ (k ) = λ (k − 1) + 2  R(k ) − µ 0 − 
p(k ) − p(k − N 1 + 1) σ  2 
µˆ 1 (k ) = µˆ 1 (k − 1) + . (25)
N1
This variable will be compared with
provide estimates for unknown means µ 0 and µ 1 , respec-
PSIG 0301 Leak Detection and Locating – A Survey 7

λ
< B ⇒ no leak
 (32) dp ρv v (
M − M leak M − M leak )( )
= f = f . (35)
≥ A ⇒ leak dx 2D 2 DA 2 ρ
against two thresholds A and B with A>B to evaluate the leak Eq. (34) and eq. (35) define two straight lines, which intersect
alarm. No decision is mad if B ≤ λ < A . Mean µ 0 and stan- at the leak location, Fig. 6. Measuring the straight line gradi-
dard deviation σ are not known; therefore estimates are re- ents (assuming that the friction factor f is known and estimates
quired. A simple approach is to use running estimates like eq. of the leak rate M leak are available) and calculation of the
(24), leading to a sequential generalized likelihood ratio test. It
point of intersection therefore allow for leak location estima-
is also possible to use different pre-assigned values for σ ac-
tion [21].
cording to the actual operating condition, e.g. during transient
operation, hereby decreasing the false alarm probability at the
Wave propagation analysis. If a sudden leak M leak occurs
expense of increasing the lowest detectable leak rate.
at time tleak, a negative pressure wave with wave front ampli-
tude ∆p can be observed propagating with wave speed a
LEAK LOCATING through the pipeline with cross-sectional area A, downstream
and upstream with respect to the leak location xleak. Applying
Externally based LDS eq. (21) yields6
Local leak sensors of externally based systems often provide 1
very accurate leak locating at the expense of high system costs ∆p = − ρ ⋅ a ⋅ ∆v = − ⋅ a ⋅ M leak . (36)
A
and complexity of installation.
Using pressure transducers installed downstream and upstream
Internally based LDS of the leak location, this wave can be detected at time

From a statistical point of view, leak detection is a detection x leak − x up

problem, whereas leak location (and rate determination) is an t up = t leak + (37)
estimation problem: Given the field data, the location (and the a
rate) of the leak has to be established.
passing the pressure transducer upstream position xup and
Gradient intersection method. For a simple horizontal pipe-
line of diameter D (and cross-sectional area A), steady state x down − xleak
analysis basing on eq. (10-12) for pipelines transporting liq- t down = t leak + (38)
uids with constant line density ρ L yields
a
passing the pressure transducer downstream position xdown.
dp ρv v M M The leak location then yields
= f = f (33)
dx 2D 2 DA 2 ρ L
⋅ ((x down − x up ) + a ⋅ (t up − t down )),
1
x leak = (39)
for pressure drop per unit length dp dx . f denotes the Darcy- 2
Weisbach friction factor according to eq. (14) and (15) for
turbulent and laminar flow, respectively. Pressure drops line- see Fig. 7; time difference determination ∆t ≡ t up − t down (as-
arly if f and D are constant along the pipeline and in absence suming pressure transducer locations are known) gives the
of a leak. If a leak of leak rate M leak occurs at leak location leak location estimate.
xleak, upstream pressure drop is given by5

dp ρv v M M
= f = f , (34)
dx 2D 2 DA 2 ρ

whereas downstream pressure drop decreases to

6
In real applications, the wave amplitude ∆p decreases due to
5
Unchanged mass flow of inlet had been assumed for simplic- heat transfer and friction losses, leaving the basic principle
ity but without loss for generality. untouched.
8 GERHARD GEIGER, THOMAS WERNER, DRAGO MATKO PSIG 0301

Optimization methods. It is possible to extend the mathe- solvable by well known statistical methods, e.g. multiple
matical model eq. (10-12) with a leak described by leak rate hypothesis testing.
M leak at leak location xleak. This can be used using a RTTM-
Exponents of this kind of thinking come to the point: leak
based deviation analysis approach. If a leak occurs, simulated detection is a decision problem, and they can apply well
and measured values from redundant variables differ at the known statistical methods like SPRT. But they tend to ne-
first moment. After leak detection, model parameters glect the process variables are not stationary at all, hereby
M leak and xleak will be tuned automatically using an optimiza- reducing the possibility of include a-priori-knowledge.
tion method, e.g as proposed by Levenberg and Marquard.;
goal of the optimization is to minimize the difference between In order to combine the advantages and eliminates the disad-
vantages of both points of view, the now presented extended
measured and simulated variable. M leak and xleak correspond to
RTTM (E-RTTM) approach [21, 24, 25] combines both meth-
the minimum of the difference. odologies:
Multiple hypothesis testing. If an extended mathematical • A “classical” RTTM generates residuals basing on the
model with model parameters M leak and xleak is available, mul- RTTM deviation analysis. This module will be called
tiple mathematical models can be computed in parallel. Each subsequently “pipeline observer”.
model corresponds to a specific hypothesis, associated to spe-
cific values of the model parameters M leak and xleak. Leak lo- • A statistical classifier uses the residuals as input. It as-
signs the pipeline to one of two classes: class “no leak”
cation estimation is equivalent with finding the best-matching and class “leak”. This module will be called subsequently
mathematical model by means of a statistical hypothesis test. “pipeline classifier”.

AN EXTENDED RTTM APPROACH This forms an online pattern recognition scheme [26] with a
feature extraction module (pipeline observer) and a classifica-
The internally based LDS methodologies presented previously tion module (pipeline classifier), Fig. 8.
point up two different points of views:
Pipeline observer. The pipeline observer computes residuals
• Simulation point of view. From this point of view, detect-
ing a leak mainly requires the knowledge of the transient x(t ) ≡ v I (t ) − vˆ I (t )
(40)
behavior of a pipeline, either without or with leak. Model y (t ) ≡ v O (t ) − vˆ O (t )
compensated mass balance techniques are an example:
Assuming that the mathematical model (e.g. eq. (10-12)) with measured flow v I (t ) and v O (t ) at inlet an outlet, respec-
is correct, the computation of the corrected imbalance (re-
sidual) eq. (1) leads to the straightforward leak detection tively8. vˆ I (t ) and vˆ O (t ) denote estimated flows at inlet and
eq. (20); not much statistics are required. outlet, assuming that there is no leak.; calculation is computer
based solving the PDE system eq. (10-12) for given boundary
Exponents of this kind of thinking are able to include a- conditions. Up to that point, the extended LDS just is a RTTM
priori-knowledge of the pipeline in a very sophisticated LDS using the deviation principle as described before.
way. But they tend to neglect the decision nature of leak
detection: at the end, beside all sophisticated computa- Pipeline classifier. The residuals are inputs to the pipeline
tions, there must be the simple binary decision: “no leak” classifier, which uses them as features in the context of a pat-
or “leak”. tern recognition scheme. Transient effects are compensated to
some extend by the pipeline observer; so the classifier doesn’t
• Statistical point of view. Statisticians strongly emphasize need to consider them. From a theoretical point of view, ne-
the decision aspect of leak detection: at the end, the LDS glecting model errors and assuming appropriate measurement
has automatically to decide if there is a leak or not. This noise characteristics, the residuals are stationary normal dis-
leads to decision theory [27]. The LDS basing on Wald’s tributed process variables. It is then straightforward to use
Sequential Probability Ratio Test (SPRT) described in well known statistical classification tools to perform leak de-
[23] and the Pressure Point Analysis PPA [22] are previ- tection; a rich set of them is available [26]. An appropriate
ously discussed examples: Assuming that process vari- technique is described in [27]: to perform leak detection, the
ables like uncompensated mass imbalance (SPRT) or residual sequence will be analyzed to discover changes in the
pressure (PPA) behave in a statistical “good” manner7, statistical model behavior. If a significant change of parame-
leak detection reduces to a statistical detection problem, ters describing the mathematical model will be detected, leak

8
7
Usually meaning stationary normal distributed process vari- Volume flow V and mass flow M could also be used instead
ables. of v ; F in Fig. 8 therefore denotes one of these three variables.
PSIG 0301 Leak Detection and Locating – A Survey 9

alarm is raised. Leak persistence is tested using a simple out- • If a leak occurs during start-up with severe transients,
lier removal algorithm: during a test phase, the leak test must mean value increases to 164l (1.040bbl, 43.7 US gal).
steadily indicate a leak; otherwise, the leak alarm will be dis-
carded. Accuracy of leak locating. Near the middle of the pipeline, a
test quantity had been removed from the line; Fig. 11 summa-
Leak locating. The pipeline classifier additionally estimates rizes the results of the leak locating algorithm.
leak location (and rate) using two methods described previ-
ously: gradient intersection method [22] and wave propagation Conclusion. The E-RTTM LDS according to Fig. 8 has
analysis. The algorithm automatically uses the method best proven his applicability and very good performance for a liq-
suited for the specific application and operation condition. uid application, a multi-batch-pipeline transporting up to 7
different batches.
Field Test 1: Liquid Pipeline, Multi-Batch
Field Test 2: Gas Pipeline
The first application is al liquid pipeline of length L=9,854m
(6.123miles), see Table 1. The pipeline is operated in multi- The second application [28] for testing the E-RTTM method is
batch-mode with 7 different batches. E-RTTM applicability concerned with a gas pipeline of length L=10,625m
had been tested performing extensive tests including leak tests. (6.602miles), see Table 2. Gas is CO (carbon monoxide),
Three different operating conditions were defined: which has in the present case nearly ideal gas properties; state
equation eq. (16) can therefore be used.
• Normal operation, steady state. The pipeline is pressur-
ized with nominal pressures, nominal flow is present; we Leak tests. Leak tests had been accomplished by flaring gas at
have (more of less) steady state. the inlet section. Three different kinds of leaks had been intro-
duced: 0.5%, 1% and 2.5% with respect to the nominal flow of
• Normal operation, small transients. The pipeline is 4000m3/h (~40ft3/s) at standard conditions 0°C, 1.013bar
pressurized with nominal pressures, nominal flow is (32°F, 14.7psi).
present. Controlling a valve behind the outlet area, small
transients were generated. Fig. 12 compares measured flows v I (t ) and v O (t ) and esti-
mated flows vˆ I (t ) and vˆ O (t ) for inlet out let, respectively.
• Start-up, severe transients. Staring the pumps, severe
The differences are the residuals according to eq. (40). Due to
transients were generated.
the deviation principle, significant differences indicate a leak
Smallest detectable leak rate. The smallest detectable leak in the pipeline, which occurs for the application at time
rate was determined using a statistical analysis [24]; results are t=8900s; please notice the significant reactions of the pipeline
shown by Fig. 9. This quantity depends on operating condition observer. Results of the pipeline classification are shown by
and kind of batch.. Fig. 13. A test variable T is tested against a alarm threshold A.
T exceeds A shortly after leak occurrence; now, testing leak
• If a leak occurs during normal operation, steady state, the persistence within the scope of outlier removal procedure
mean for the smallest detectable leak rate is about 0.5%. starts. Leak alarm is raised after leak persistence has been con-
firmed.
• If a leak occurs during normal operation with small tran-
sients, mean value increases to 0.9% Lost gas mass. Like for liquids, the lost gas mass up to the
leak alarm is a more comprehensive measure of environmental
• If a leak occurs during start-up with severe transients, pollution; Table 3 summarizes the results fro different leak
mean value increases to 1.2%. detection time parameters. Finally, leak detection time pa-
rameter had been chosen to 320s; corresponding lost gas
Lost batch volume. The lost batch volume up to the leak masses are between 2.64kg (5.82lb) and 11.38kg (25.09).
alarm is a more comprehensive measure of environmental
pollution; Fig. 10 summarizes the results. Zero false alarm warranty. Alarm threshold A has been cho-
sen to A=0.041m/s (0.135ft/s) to ensure no false alarm for the
• If a leak occurs during normal operation, steady state, the whole test time period of more then 4 months. Detection of
mean value for the lost batch volume up to leak alarm is leak according to Fig. 8 is therefore possible without any false
about 69l (0.434bbl, 18.2 US gal). alarm.

• If a leak occurs during normal operation with small tran- Conclusion. The E-RTTM LDS according to Fig. 8 has
sients, mean value increases to 120l (0.752bbl, 31.6 US proven his applicability and very good performance for a gas
gal). application, a pipeline transporting carbon monoxide (CO).
10 GERHARD GEIGER, THOMAS WERNER, DRAGO MATKO PSIG 0301

7. Großwig, S. et al. (2001): Distributed Fiber Optical Tem-

FUTURE WORK perature Sensing Technique – A Variable Tool for Moni-
toring Tasks. Proceedings of the 8th International Sympo-
Future work is concerned with following topics:
sium on Temperature and Thermal Measurements in In-
• Extensions of the pipeline observer. The pipeline observer dustry and Science, June 19 – 21, 2001.
Fig. 8 compensates fluid compressibility and pipeline 8. Framatome ANP GmbH (1998): LEOS – Leak Detection
elasticity. Complex flow situations introduced by multi- and Location System.
phase flow [11] due to the use of DRAs or slack line con- 9. Liou, C. P. (1993): Pipeline Leak Detection Based on
ditions [29] can be handled by an extension of the Mass Balance. Proceedings of the 1993 International Con-
mathematical model eq. (10-12) through a deeper physical ference on Pipeline Infrastructure II, ASCE, 1993.
analysis [30]. Another possibility is to use black box 10. Shapiro, A. H. (1953): The dynamics and Thermodynam-
modules like Neural Nets (NN) as an addendum [32], in- ics of Compressible Fluid Flow. The Ronald Press Com-
troducing the concept of knowledge based leak monitor- pany.
ing. 11. Hanks, K. H.; Stanley, G. T. (1998): Requirements of
Pipeline Simulation: Accurately Modeling Transient Pipe-
• Test of other classification methods. There is a rich set of line Operation. Proceedings of the 1998 International
alternatives for the pipeline classifier [26]; at the moment, Pipeline Conference, ASME, 1998.
a systematic evaluation of different approaches like NN 12. Wylie, E. B.; Streeter, V. L.; Suo, L. (1993): Fluid Tran-
classifier is missing. sients in Systems. Prentice Hall.
13. Wylie, E. B.; Streeter, V. L.; Bedford, K. W. (1998):
Fluid Mechanics. McGraw-Hill.
CONCLUSIONS 14. Schroeder, D. W. (2001): A Tutorial on Pipe Flow Equa-
tions. PSIG.
This paper presented a survey about leak detection and locat- 15. Schmidt, F. W.; Henderson, R. E.; Wolgemuth, C. H.
ing methodologies, methods, techniques and systems. LDS (1993): Introduction to Thermal Sciences. John Wiley &
were subdivided according to API 1130 2nd Ed. [2] into inter- Sons.
nally based and externally based systems; for both categories, 16. Modisette, J. L. (2000): Equation of State Tutorial. PSIG.
examples had been presented. 17. Osiadacz, A; Chaczykowski, M. (1998): Comparison of
Isothermal and Non-isothermal Transient Models. PSIG.
Furthermore, an Extended RTTM approach was presented
18. Matko, D.; Geiger, G. (2002): Models of Pipelines in
combining the advantages of RTTM LDS and statistically
Transient Mode. Mathematical and Computer Modelling
based LDS. The applicability of this method was proven for a
of Dynamical Systems Vol. 8 Nr. 1, p. 117-135. Swets &
liquid multi-batch pipeline and a gas pipeline.
Zeitlinger Publishers.
The E-RTTM method is now ready for commercial usage. The 19. Geiger, G.; Matko, D.; Werner, T. (2003): Extending
GALILEO LDS from KROHNE Duisburg bases on this tech- classical pipeline models with Neural Nets. Proceedings
nique. of the 11th Mediterranean Conference on Control and
Automation, June 18-20, 2003, Rhodes (Greece).
20. US Patent 4,308,746 (1982): Liquid Pipeline Leak Detec-
REFERENCES tion.
1. 49 CFR Part 195 (2002): Transportation of Hazardous 21. Billmann, L. (1985): Methoden zur Lecküberwachung
Liquids by Pipeline. Code for Federal Regulations Title und Regelung von Gasfernleitungen. Fortschrittsberichte
49 (January, 2002). VDI Reihe 8, VDI-Verlag.
2. API 1130 (2002): Computational Pipeline Monitoring for 22. US Patent 4,796,466 (1989): System for Monitoring Pipe-
Liquid Pipelines. 2nd Edition (November, 2002). Ameri- lines.
can Petroleum Institute. 23. Zhang, X. J. (1993): Statistical leak detection in gas and
3. TRFL (2003): Technische Regeln für Fernleitungen liquid pipelines. Pipes & Pipelines International, July-
(Technical Rules For Pipelines). August 1993, p. 26-29.
4. Whaley R. S., et al (1992): Tutorial on Software Based 24. Geiger, G. (1998): Anwendung eines modellgestützten
Leak Detection Techniques. PSIG Pipeline Simulation In- Verfahrens zur Leckerkennung und –ortung. GMA-
terest Group, 1992. Kongress 1998 Meß- und Automatisierungstechnik,
5. Alaska Department of Environmental Conservation 18./19. Juni 1998, Ludwigsburg.
(1999): Technical Review of Leak Detection Technolo- 25. Geiger, G., Gregoritza, W., Matko, D. (2000): Leak De-
gies, Volume 1. tection and Localization in Pipes and Pipelines. Proceed-
6. Fuchs, H. V. (1991): One Year of Experience eith Leak ings of the 33rd European Symposium on Computer
Detection by Acoustic Signal Analysis. Applied Acous- Aided Process Engineering ESCAPE-10, May 7-10, 2000,
tics 33 (1991), p. 1-19. Florence (Italy).
26. Duda, R. O.; Hart, P. E.; Stork, G. E. (2000): Pattern
PSIG 0301 Leak Detection and Locating – A Survey 11

Classification, 2nd Edition. Wiley-Interscience.

27. Key, S. M. (1998): Fundamentals of Statistical Signal
Processing – Detection Theory. Prentice Hall.
28. Bollermann, B. (2003): Modellgestützte Lecküberwa-
chung an einer CO-Pipeline. HDT-Seminar Lecküberwa-
chuhg an Rohrleitungen und Pipelines, Essen, 18. März
2003.
29. Nicholas, R. E. (1995): Simulation of Slack Line Flow –
A Tutorial. PSIG.
30. Rygg, O. B.; Ellul, I. R. (1993): The Engineering of Off-
shore Pipelines: A Dynamic Simulation Approach. Pipe-
line Engineering, ASME 1993.
31. Shou, G. (1999): Solid-liquid Flow System Simulation
and Validation. PSIG.
32. Geiger, G.; Werner, T.; Matko, D. (2001): Knowledge-
based Leak Monitoring for Pipelines. 4th IFAC Workshop
on On-Line Fault Identification and Supervision in the
Chemical Process Industries, June 7-8, 2001, Jejudo Is-
lands (Korea).

ACKNOWLEDGEMENTS
The author thanks KROHNE Duisburg for (financial) support
and all staff members of both (unnamed) companies from the
application area for their patience and enthusiasm.
12 GERHARD GEIGER, THOMAS WERNER, DRAGO MATKO PSIG 0301

TABLES

Parameter Symbol SI Imperial

Length of pipeline L 9,854 m 6.123 miles
Diameter D 0.2065 m 8.13 inch
Number of batches 7 7
Density range ρ 590 – 904 kg/m (15°C, 1bar)3
25 – 108 °API (59°F, 14.5psi)
Viscosity range ν -7
5.2·10 – 9.5·10 m /s -7 2
52-95 cSt
Nominal flow V 200 - 350 m3/h 1,260 – 2,200 bbl/h

Table 1 – Parameters for a multi-batch pipeline transporting liquids.

Parameter Symbol SI Imperial

Length of pipeline L 10,625 m 6.602 miles
Diameter D 0.157 m 6.18 inch
Type of gas CO CO
3 3
Nominal flow (Standard) V 4000 m /h (0°C, 1.013bar) ~40 ft /s (32°F, 14.7psi)

Table 2 – Parameters for a gas pipeline.

Leak detection time MLost (2,5% leak) MLost (1 % leak) MLost (0,5 % leak)
80s 2.84 kg (6.26 lb) 0.94 kg (2.07 lb) 0.66 kg (1.46 lb)
140s 4.97 kg (10.96 lb) 1.65 kg (3.64 lb) 1.16 kg (2.56 lb)
200s 7.11 kg (15.67 lb) 2.35 kg (5.18 lb) 1.65 kg (3.64 lb)
260s 9.24 kg (20.37 lb) 3.06 kg (6.75 lb) 2.15 kg (4.74 lb)
320s 11.38 kg (25.09 lb) 3.76 kg (8.29 lb) 2.64 kg (5.82 lb)
380s 13.51 kg (29.78 lb) 4.47 kg (9.86 lb) 3.14 kg (6.92 lb)
440 s 15.64 kg (34.48 lb) 5.17 kg (11.40 lb) 3.64 kg (8.02 lb)

Table 3 – Results using E-RTTM for gas pipeline.

PSIG 0301 Leak Detection and Locating – A Survey 13

FIGURES

Figure 1 – Leak Detection Systems (LDS) according to API 1130 2nd Edition (USA).
14 GERHARD GEIGER, THOMAS WERNER, DRAGO MATKO PSIG 0301

Figure 2 – Requirements concerning LDS according to TRFL (Germany).

Figure 3 – Externally based leak detection and locating using optical fiber. ©GESO GmbH, Jena (Germany).
PSIG 0301 Leak Detection and Locating – A Survey 15

Figure 4 – Externally based leak detection and locating using vapor sensing tube. ©Framatome ANP, Erlangen (Germany).

Figure 5 – Pressure Point Analysis (PPA) principle. ©EFA Technologies, Inc., Sacramento (USA).
16 GERHARD GEIGER, THOMAS WERNER, DRAGO MATKO PSIG 0301

Figure 6 – Leak locating with gradient intersection method.

Figure 7 – Leak locating with wave propagation analysis.

PSIG 0301 Leak Detection and Locating – A Survey 17

Figure 8 – Extended RTTM-LDS. ©KROHNE Duisburg (Germany).

Figure 9 – Smallest detectable leak rate using E-RTTM (liquid application).

18 GERHARD GEIGER, THOMAS WERNER, DRAGO MATKO PSIG 0301

Figure 10 – Lost batch volume up to a leak alarm using E-RTTM (liquid application).

Figure 11 – Results for leak locating using E-RTTM (liquid application).

PSIG 0301 Leak Detection and Locating – A Survey 19

Figure 12 – Measured and estimated flow using E-RTTM, leak rate 1% (gas application).

Figure 13 – Leak test using E-RTTM, leak rate 1% (gas application).

20 GERHARD GEIGER, THOMAS WERNER, DRAGO MATKO PSIG 0301

BIOGRAPHY
Gerhard Geiger (author and presenter)
Born in Kaufbeuren, Germany on October 6, 1954. 1974 –
1979 student at the Faculty of Electrical Engineering in Darm-
stadt, Germany. Master degree 1979, Doctor degree 1985,
both on the Faculty of Electrical Engineering in Darmstadt and
in the field of Control Systems. Founder of the MAGNUM
GmbH (process automation) in 1987. MAGNUM CEO from
1987 – 1995. Since 1995 full professor at the University of
Applied Sciences in Gelsenkirchen, Germany. Lectures: In-
dustrial Instrumentation and Communication Technology,
Applied Computer Science. Main focus of research: Leak
monitoring systems for pipelines. 31 scientific papers and con-
ference papers.

Thomas Werner (author)

Born in Gelsenkirchen, Germany on Mai 11, 1973. 1995 –
1999 student at the Department of Electrical and Electronic
Engineering at the University of Applied Sciences Gelsen-
kirchen. Bachelor Degree 1999 in Electrical Engineering at the
University of Applied Sciences Gelsenkirchen. Since 1999
scientific assistant at the University of Applied Sciences in
Gelsenkirchen: Area of research: Development and applica-
tion of leak detection systems for pipelines

Drago Matko (author)

Born in Trbovlje, Slovenia on July 8, 1947. 1966 – 1971 stu-
dent at the Faculty of Electrical Engineering in Ljubljana, Slo-
venia. Master degree 1973, Doctor degree 1976, both on the
Faculty of Electrical Engineering in Ljubljana and in the field
of Adaptive Control Systems. Assistant, assistant professor
and associate professor on the Faculty of electrical engineer-
ing in Ljubljana 1973 – 77, 1977 – 82 and 1982 – 88. Since
1988 Full professor on the Faculty of Electrical Engineering in
Ljubljana. Lectures: Discrete Control Systems, Computer
Controlled Systems, Identification and Computer Aided De-
sign of Control Systems. Award of Research Council of Slo-
venia for achievements in Computer aided design of control
systems in 1988. Founder of the Automatic Control Society of
Slovenia in 1989. Co-author of the book: R. Isermann, K.H.
Lachmann and D. Matko: Adaptive Control Systems (Prentice
Hall, 1992). This book was awarded as the Book of the year
1992 of the System Research Foundation. Author of the book:
D. Matko, R. Karba and B.Zupančič): Simulation and Model-
ing of Continuous Systems (Prentice Hall, 1992). Editor of
the book: Space technology applications (in Slovene, pub-
lished by Didakta in 1996). 30 scientific papers, 23 books, 44
projects, 208 conference papers, 1 translation,1 technical im-
provement, 1 popular article.