Construction Research Congress 2012 © ASCE 2012 495

Pipe spool fabrication sequencing by automated planning

Di HU1 and Dr. Yasser MOHAMED2

1. PhD candidate, Civil and Envir. Engrg. Department, 1-047 Markin/CNRL Natural
Resources Engineering Facility, Univ. of Alberta, Edmonton, AB, T6G 2W2; PH
(780) 492-8093; [email protected]

2. Assistant Professor, Civil and Envir. Engrg. Department, 3-011 Markin/CNRL

Downloaded from ascelibrary.org by New York University on 08/01/15. Copyright ASCE. For personal use only; all rights reserved.

Natural Resources Engineering Facility, Univ. of Alberta, Edmonton, AB, T6G 2W2;
PH (780) 492-1321; FAX (780) 492-0249; [email protected]

ABSTRACT

Construction of Industrial facilities involves a substantial amount of piping. Pipe

spools are usually pre-fabricated from a number of raw pipes and pipe fittings (e.g.
elbows, flanges, tees, etc.) in fabrication shops. Pipe spool fabrication is often
affected by various disruptions from within or outside the shops. Previous research
mainly focuses on shop layouts, dispatching rules, buffer location and standardized
products. Another critical factor, the sequencing of pipe spool fabrication, is usually
overlooked. A pipe spool can be fabricated in several alternative sequences that are
often decided by shop foremen based on experience. It is rare that these alternative
sequences get compared and evaluated. A simulation experiment shows that shop
productivity can be improved by varying spool fabrication sequence. This paper
presents an investigation of Artificial Intelligence (AI) planning approach that
automatically identifies the optimal fabrication sequence for pipe spools while
considering various fabrication logics. Experiments are conducted with different AI
planners to evaluate their capabilities. The results indicate that one of the planners is
more suitable for solving the sequencing problem than others. However, it requires
special pre-processing of the input that may be prohibiting for practical use.
Directions of future research to overcome these limitations are discussed.

INTRODUCTION

Industrial construction includes a wide range of projects, such as petroleum

refineries and chemical and power plants. This type of construction involves intensive
piping, which connects a variety of equipments and conveys process fluid and gas.
Due to compressed schedule and limited space on site, industrial construction projects
rely heavily on the offsite fabrication and assembly. As such, piping work is divided
into three stages: pipe spool fabrication, module assembly, and site installation.

Timely supply of pipe spools to module yard and installation site is the key to
the success of whole project. However, studies showed that pipe spool fabrication

Construction Research Congress 2012

 Construction Research Congress 2012 © ASCE 2012 496

shops are faced with various interruptions (e.g. out-of-sequence deliveries, change
orders) and often not operating at an optimal productivity (Howell and Ballard 1996,
Tommelein 1998, and Wang et al. 2009). Another major challenge faced with
fabrication shops is that most of pipe spools are unique (Wyss 2009). Pipe spools can
be unique in material, configuration, type of joints and many other properties. As
such, pipe spools cannot be entirely or partially fabricated in advance, which means
fabrication shops are unable to use on-hand inventory to buffer against variability
Downloaded from ascelibrary.org by New York University on 08/01/15. Copyright ASCE. For personal use only; all rights reserved.

from within or outside the shop.

Unique design and configuration means these pipe spools need to be custom
built. More specifically, the fabrication process usually varies from one pipe spool to
another. Since most fabrication operations (e.g. cutting, fitting and welding) involved
are similar, the variation mainly lies in the sequence of these operations. The
fabrication sequence determines steps that pipe spools go through from raw materials
to the final product. In reality, pipe spool fabrication sequence is determined by shop
foremen in heuristic manner. Given the enormous number of pipe spools involved in
an industrial project and the fast-tracking nature of the project, it is quite challenging
for human planners to come up with fabrication sequence with both efficiency and
quality. It is not a surprise to find that fabrication sequence for the same pipe spool
varies with human planners, because there is no standard way of sequencing in the
industry. Moreover, a pipe spool can be fabricated in several alternative sequences.
However, it is rare for these alternative sequences to get compared and evaluated.

A study by Hu and Mohamed (2011) shows that different fabrication

sequences can lead to different shop performance. This motivates us to develop a
solution to automate fabrication sequencing process where human skills are
continuously needed and the result affects operation performance in significant ways.
This paper explores the use of artificial intelligence (AI) planning technique to solve
the spool fabrication sequence planning problem. The remainder of the paper starts
with brief introduction of pipe spool fabrication and fabrication sequence problem,
followed by a simulation experiment (Hu and Mohamed 2011) that shows the impact
of fabrication sequence on the operation performance. The paper then continues with
explaining reasons why choosing AI planning over other techniques. The current
research progress is described by experimenting with pipe spools with increasing
complexity. A tentative conclusion about capabilities and limitations of AI planning
is provided at the end.

PIPE SPOOL FABRICATION

Pipe spools are fabricated from a number of raw pipes and pipe fittings (e.g.
elbows, flanges, tees, etc.) in fabrication shops. Raw pipes are cut to the required
sizes and moved with pipe fittings to a fitting table, where some of the components

Construction Research Congress 2012

 Construction Research Congress 2012 © ASCE 2012 497

are fitted together (i.e. temporarily connected). The resulting sub-assembly (part of
the final pipe spool) continues with welding operations (i.e. permanent connected)
before it comes back to the fitting table and gets fitted with other spool components.
Spool fitting and welding can be grouped into two types: (1) roll fitting and welding
and; (2) position fitting and welding. Roll fitting and welding means the main pipe
can be turned by a rolling machine and the fitter or the welder does not have to
change his or her position to perform the operation, whereas position fitting and
Downloaded from ascelibrary.org by New York University on 08/01/15. Copyright ASCE. For personal use only; all rights reserved.

welding occur when one or more branches of the main pipe exceed the clearance limit
(see figure 1). In such case, the fitter or the welder has to move around the main pipe
run to accomplish fitting or welding. As a result, position fitting and welding usually
takes more time to finish than roll fitting and welding. To minimize number of
position fitting and welding is one of the goals of pipe spool fabrication sequencing.

Figure 1 Roll Welding VS. Position Welding

PIPE SPOOL FABRICATION SEQUENCE

The fabrication sequence defines the process of how a pipe spool will be
fabricated gradually from raw materials (e.g. pipes and fittings), to intermediate spool
components, and eventually to the final product. As mentioned before, a pipe spool,
in many cases, can be fabricated through a number of alternative sequences. Figure 2
shows an example of a pipe spool with relatively simple configuration. It shows that
the pipe spool can at least be fabricated by two different sequences from the same raw
materials. Fabrication sequence1 requires three operations to produce the final
product while sequence2 only needs two. Moreover, handling is needed between
these operations, which further deteriorate the shop performance. In reality, sequence
is determined by shop foremen in very heuristic manner and these alternative
sequences seldom have a chance to be compared and evaluated. As a result,
opportunities of productivity improvement slip away. The spool fabrication sequence
problem was identified when several industry professionals were interviewed. The

Construction Research Congress 2012

 Construction Research Congress 2012 © ASCE 2012 498

roo
ot cause is atttributed to th
he fact that currently
c theere is no stanndard, structuured way to
idenntify sequennce for pipe spools in th he industry, and no acaademic reseaarch on this
speecific topic has
h yet been found.

Raw Step 1 Step 2 Step 3

Material
Fabrication Seequence 1
1 Flange
Downloaded from ascelibrary.org by New York University on 08/01/15. Copyright ASCE. For personal use only; all rights reserved.

2 Pipe

1 + 2 + 3 = 7 5 + 6 = 8 7 + 8 + 4
Final Produ
uct
3 Elbow

Fabrication Seequence 2

4 Flange

5 Pipe

3 + 4 + 5 + 6 1 + 2 + 7
6 Plate
= 7

Figure 2 Alternative
A pipe
p spool fab
brication seqquences (adaapted from H
Hu and
Moh hamed 2011))
SIM
MULATION
N EXPERIM
MENT

A simulation experriment is con nducted to ttest the hyppothesis that fabrication

seq
quence has im mpacts on pipe
p spool performance.
p . 22 pipe sppools are sellected. Two
alteernative fabrrication sequ
uences are prepared for each pipe sppool. They aare inputted
intoo the simulaation model that
t represennts the real sshop operatiions and thee make span
for fabricating all 22 pipe spools is coollected and compared. T The result (in figure 3)
shoows 10.09% reduction in i the make span and 116.88% decrrease in the number off
hanndlings (i.e. a type of non
n-value addiing activity)..

Figurre 3 Simulattion Experim

ment Result ((Hu and Mohhamed 20111)

Construction Research Congress 2012

 Construction Research Congress 2012 © ASCE 2012 499

SEQUENCING OF CONSTRUCTION PROCESSES

Previous research on sequencing construction processes mostly focuses on

scheduling and prioritizing processes under limited resource availability. Relatively
less attention has been paid to the area where planning is more focused on logic
dependency between processes by considering the geometric and technological
requirements. This paper is more related to the latter type of sequencing. A review of
the previous relevant researches reveals that two major topics are: (1) identify and
Downloaded from ascelibrary.org by New York University on 08/01/15. Copyright ASCE. For personal use only; all rights reserved.

formalize construction sequence rationales (2) automate generation of construction

sequences. Gray (1986) identify sequence rationales such as “covered by” or
“weather protected by other components” that are generalized from different
contractors’ schedules. Navinchandra et al. (1988) identify similar dependencies as
“supported by” and “connected to”. Most of these sequence rationales are derived
from the physical relationships between building components. Echeverry et al (1991)
enrich the body of knowledge by adding three more types of factors that might govern
the sequence of construction. Basically, they identify, in addition to “physical
relationships”, “trade interaction”, “path interference”, and “code regulations”. Under
each factor type, they comprehensively enumerated more specific sequencing
constraints. As an extension to this, many researchers attempt to develop an Artificial
Intelligent (AI) planner that not only store all the relevant sequencing rationales but
also automatically apply this knowledge to identify construction sequence, e.g.
OARPLAN (Darwiche et al. 1988), GHOST(Navinchandra et al. 1988), BUILDER
(Cherneff et al. 1991), and MDA (Jägbeck 1994). Aalami et al. (1998) categorize
these systems into two major groups: (1) process-based reasoning system, e.g.
GHOST; (2) component-based reasoning system, e.g. OARPLAN. A recent study by
Koo et al. (2007) pointed out that many research on domain specific AI planning
systems is more focused on identifying a correct construction sequence rather than
discovering a number of possible sequence alternatives. They introduce a prototype
system named "CLCPM", that make use of a constraint ontology and a classification
mechanism to automatically assign "role" and "status" to relevant activities in CPM.
A decision will be made with regard to whether or not the sequence of the target
activity could be changed.

An observation from the previous construction sequencing research is that

most of them are focused on building construction projects. Many sequencing
constraints are derived from the physical relationships between building components
(e.g. columns, beams, walls and slabs). This makes it difficult to apply existing
planning systems to industrial construction projects, i.e. the building blocks are pipe
spools, equipments, and modules which most likely need to be pre-fabricated or pre-
assembled before the final installation on site. The sequence constraints between
these components are significantly different from those applied in building

Construction Research Congress 2012

 Construction Research Congress 2012 © ASCE 2012 500

construction. In addition, existing planning systems are mostly knowledge-based and

depend on enumeration of sequencing logic and, sometimes, require an existing
schedule to infer re-sequence options.

AUTOMATED PLANNING

Artificial Intelligence (AI) Planning is process of selecting and sequencing a

set of actions that can change the system from an initial state to a desired goal state. If
Downloaded from ascelibrary.org by New York University on 08/01/15. Copyright ASCE. For personal use only; all rights reserved.

evaluation criteria are provided, some of AI planning techniques is able to identify

optimal sequence with respect to the predefined objectives (e.g. minimum cost).
Research into the AI planning has been of major interest in recent years. AI planning
systems evolve from classical planning tools to more complicated planners that can
address issues encountered in real-life settings such as numeric-valued variables, time
constraints, or a non-deterministic environment. Planning Domain Description
Language (PDDL) is the standard language for domain-independent planners. It has
been refined and extended since it was first proposed by Drew McDermott (1998). AI
planning has been successfully applied in several areas, such as robot navigation,
manufacturability of machined parts, and emergency evacuation (Ghallab et al. 2004).

Planning problems studied in AI usually involve: (1) a dynamic system where objects
of interest are interacting, (2) finite system states, (3) a set of actions that can change
the system state, (4) an initial state and a goal state of the system, and, sometimes, (5)
an evaluation criteria. Pipe spool fabrication sequencing problem is a good candidate
for using AI planning. Pipe spool fabrication can be considered as the dynamic system,
where focus is placed on objects of interest such as welds, raw pipes and piping
fittings or sub-assemblies (part of the final pipe spool). Fitting and welding are
actions available to change the state of raw materials or in-progress assemblies. The
initial state of the pipe spool starts with a group of raw materials and the goal state is
the final product. The output is a sequence of fitting and welding operations which
assemble all raw materials to the final product. Evaluation criteria could be the least
number of handlings or the least number of position fitting and welding.

Using PDDL to solve a planning problem entails a modeling process which

results in two pieces of descriptions. First, there is a general description that
represents the system and the system dynamics under study. It is stored in a Domain
file. Another piece of description is specific to each planning problem (e.g. specific
objects, their initial states and goal states). It is called Problem file. Usually, the same
domain file can be used to solve different planning problems as long as the same
system representation stays valid for these problems, whereas the problem file varies
with each planning problem (i.e. different initial or goal states). Both these
descriptions are fed to an AI planner which in turn uses a heuristic search method to
find an optimum or near-optimum solution.

Construction Research Congress 2012

 Construction Research Congress 2012 © ASCE 2012 501

APPLICATION OF PDDL TO PIPE SPOOL FABRICATION SEQUENCE

A series of experiments were conducted to test the capability of PDDL to

model and plan pipe spool fabrication sequencing problems. Three popular planners
(domain-independent) were used in the experiments, namely Metric-FF (Hoffmann
2002), LPRPG (Coles et al. 2008) and LPG (Gerevini and Serina 2002). Metric-FF
searches on a state space while the other two search on a plan space. The experiments
begin with very simple pipe spool and then gradually move to more complex and
Downloaded from ascelibrary.org by New York University on 08/01/15. Copyright ASCE. For personal use only; all rights reserved.

more realistic configuration. Each experiment includes two steps: (1) using PDDL
language to model a pipe spool system (e.g. Domain file) and a specific case of pipe
spool (e.g. Problem file), and (2) using each of these AI planners to generate the
fabrication sequence. The following are the brief description of each experiment.

Experiments

In experiment1, a simple pipe spool is designed (figure 4). No numerical

values are considered. It simply tests if the AI planers can handle the logic aspect of
pipe spool sequencing problem. The Domain file and Problem file are shown in figure
5 and figure 6. Experiment2 uses a pipe spool with more complex configuration
(figure 4).Again, no numerical values are considered in the experiment2. For space
limit, its domain file and problem file is not shown here. Although the pipe spool in
experiment3 seems to have a simpler configuration than that in experiment2, it
involves the major challenge that dimensions of assembly parts are considered. The
motivation is that one of goals of fabrication sequencing is to minimize the number of
position fitting and welding. As mentioned previously (figure 1), the way to
distinguish roll welding and position welding is to see if the length of major branches
exceed the clearance limit of the rolling machine. Dimensions of assembly parts have
to be considered and need to be updated after each fitting or welding operation. All
resulting plans are compiled in Table 1.

Figure 4 Pipe spools respectively for experiment 1, 2 and 3

Construction Research Congress 2012

 Construction Research Congress 2012 © ASCE 2012 502
Downloaded from ascelibrary.org by New York University on 08/01/15. Copyright ASCE. For personal use only; all rights reserved.

Figure 5 Domain Deefinition file of experimeent 1

Figure 6 problem deefinition file of experimeent 1

Table 1 resulting plans
p from eaach AI plannner

Experiment1 Experiment2
t2 Experiiment3
Step1: ROOLL-X-P1-
P22-1
LP
PG Unsolvable Unsolvablee (P1 P2 W1 W3 W2 C)
LP
PRPG Unsolvable Unsolvablee Unsollvable
Step1: RO
OLL-FITTINGG-X P1 P7 W1
Sttep1:
Step2: RO
OLL-FITTINGG-Y P1 P2 W2
ROOLL-FITTING
G
Step3: RO
OLL-FITTINGG-Y P1 P8 W3 Step1: ROOLL-X-P1-
P1
1 P2 W1
Meetric-FF Step4: RO
OLL-FITTINGG-Z P1 P3 W4 P22-4
Sttep2:
Step5: RO
OLL-FITTINGG-Z P1 P6 W5 (P1 P2 W1 W1 W1 C)
ROOLL-FITTING
G
Step6: RO
OLL-FITTINGG-Z P1 P4 W6
P1
1 P3 W2
Step7: RO
OLL-FITTINGG-Z P1 P5 W7

The ressults obtaineed from all three experriments indiicate that M

Metric-FF is
more capable than the otther two AI A planners in terms oof handling pipe spool
fabrication logiic. Howeverr, Metric-FF
F has its lim
mitation too. In experimeent 3 where
merical calculation and assignmentt is involvedd Metric-FF
num F could not handle the

Construction Research Congress 2012

 Construction Research Congress 2012 © ASCE 2012 503

combination of conditional effects and numerical calculations. One of the ways to get
around this is to break down the conditional effect by moving the condition part to the
preconditions of the whole action and keep effect in the effect part of the action
formulation. This requires making actions more specific and enumerating all possible
situations with respect to which weld belong to which assembly parts. In PDDL or
Lisp language, this is called a “grounding” process. After converting the conditional
effect, LPG planner is now able to return a solution which shows in Table1. Metric-
FF seems to be able to do the same but a closer check finds that it returns an illogical
solution. The challenge regarding the grounding process is that the number of actions
Downloaded from ascelibrary.org by New York University on 08/01/15. Copyright ASCE. For personal use only; all rights reserved.

in domain file will grow exponentially with the number of welds in the pipe spool. If
a pipe spool has N welds, then 2N-1 actions need to be explicitly formulated in domain
file (e.g. a pipe spool with 13 welds requires 4096 actions defined). For extremely
complicated pipe spools, it could be computationally prohibitive to find a solution.
CONCLUSION
This study represents an in-progress research which explores the feasibility of
use of AI planning technique to solve pipe spool fabrication sequencing problems.
The use of PDDL as standard AI planning language seems to capture all necessary
logic of the domain. The challenge however lies in finding a suitable domain-
independent planner that is capable of processing pipe spool fabrication logic and
performing numeric calculations and assignments as specified in a PDDL domain file.
Experiments with three most commonly used AI planners—LPG, LPRPG and
Metric-FF shows that each planner has its own limits and that LPG seems to be a
promising planning algorithm. To use LPG planner, a grounding process is required,
which poses a challenge to sequence complex pipe spools. Future work include
developing a program that can automatically perform the grounding process (i.e.
generate actions without any conditional effect) and experimenting with LPG for
complex pipe spools to test its planning efficiency. Future research will also
investigate the applicability of domain-specific planners and dynamic programming
approaches to generate optimized fabrication sequences.
REFERENCE
Aalami, F., Kunz, J., and Fischer, M. (1998). “Model-based sequencing mechanisms
used to automate activity sequencing.” Working Paper No. 50, CIFE, Stanford Univ.,
Stanford, Calif.
Cherneff, J., Logcher, R., and Sriram, D. (1991). "Integrating CAD with
Construction-Schedule Generation." J. Computing in Civ. Engrg., ASCE, 5(1), 64-84.
Coles, A.I., Fox, M., Long, D. and Smith, A.J. (2008). "A Hybrid Relaxed Planning
Graph-LP Heuristic for Numeric Planning Domains." Proc., Eighteenth Int. Conf. on
Automated Planning and Scheduling (ICAPS 08), Sydney, Australia, September.
Darwiche, A., Levitt, R., and Hayes-Roth, B. (1988). "OARPLAN: Generating
Project Plans by Reasoning about Objects, Actions and Resources." AI EDAM, 2(3),
169-181.

Construction Research Congress 2012

 Construction Research Congress 2012 © ASCE 2012 504

Echeverry, D., Ibbs, C. W., and Kim, S. (1991). “Sequencing knowledge for
construction scheduling.” J. Constr. Engrg. and Mgmt., ASCE, 117(1), 118–130.
Gerevini, A. and Serina, I. (2002). "LPG: a Planner based on Local Search for
Planning Graphs." Proc., Sixth Int. Conf. on Artificial Intelligence Planning and
Scheduling (AIPS'02), AAAI Press, Toulouse, France.
Ghallab, M., Nau, D., and Traverso, P. (2004). Automated Planning: Theory and
Practice, Elsevier Inc. San Francisco.
Downloaded from ascelibrary.org by New York University on 08/01/15. Copyright ASCE. For personal use only; all rights reserved.

Gray, C. (1986). “Intelligent construction time and cost analysis.” Journal of

Construction Mgmt and Economics, 4(2), 135–150.
Hoffmann, J. (2002). "Extending FF to Numerical State Variables." Proc., 15th
European Conf. on Artificial Intelligence, Lyon, France, July.
Howell, G.A. and Ballard, H.G. (1996). “Managing Uncertainty in the Piping
Process.” RR 47-13, Construction Industry Institute, University of Texas, Austin, TX,
103 pp.
Hu, D and Mohamed, Y. (2011). “Effect of pipe spool sequence in industrial
construction processes.” Proc., 3rd International/9th Construction Specialty Conf.,
CSCE, Ottawa, Ont., June.
Jägbeck, A. (1994). "MDA Planner: Interactive Planning Tool Using Product Models
and Construction Methods." J. Computing in Civ. Engrg., ASCE, 8(4), 536-554.
Koo, B., Fischer, M., and Kunz, J. (2007). “Formalization of construction sequencing
rationale and classification mechanism to support rapid generation of sequencing
alternatives.” J. Computing in Civ. Engrg., 21(6), 423–433.
McDermott, D., and the AIPS-98 Planning Competition Committee (1998)."PDDL–
the planning domain definition language." Technical report, <
http://www.cs.washington.edu/education/courses/cse473/06sp/pddl.pdf > (Nov.,
2011).
Navinchandra, D., Sriram, D., and Logcher, R. D. (1988). "GHOST: Project Network
Generator." J. Computing in Civ. Engrg., ASCE, 2(3), 239-254.
Tommelein, I.D. (1998). “Pull-driven Scheduling for Pipe-Spool Installation:
Simulation of Lean Construction Technique.” J. Constr. Engrg. and Mgmt., ASCE,
124 (4), 279-288.
Wang, P., Mohamed, Y., AbouRizk, S. M., and Rawa, A. R. T. (2009). “Flow
Production of Pipe Spool Fabrication: Simulation to Support Implementation of Lean
Technique.” J. Constr. Engrg. and Mgmt., ASCE, 135(10), 1027-1038.
Wyss, Stephen. (2009). “Active Management of Pipe spool Fabricators.” Chemical
Engineering, 116(1): 40-45.

Construction Research Congress 2012