Volume 9, Issue 8, August – 2024 International Journal of Innovative Science and Research Technology

ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24AUG1039

Optimized Selective Assembly using

Hungarian Algorithm
Amal P R1; Anjumol K S1; Unnikrishnan S1, Dilip V2; George Oommen2
1
Scientist/Engineer, CSC, LPSC-V, ISRO
2
Divisional Head, CSC, LPSC-V, ISRO

Abstract:- Assembly of discrete parts guided by I. INTRODUCTION

hardware design specification constitutes the final phase
in product manufacturing. In the course of mass A mechanical design comes to life after the assembly
production of components, mating parts with geometric of its constituent parts according to the engineering design
or dimensional deviation from their intended design can drawings. Assembly of parts are convenient, provided all the
be made acceptable by the identification of suitable pairs part geometry and dimensions conform to the original
after analysing design fit and tolerance limits. By design dimensions. In the case of a non-conforming
transforming this application-specific problem into a geometry or dimension, an acceptance study has to be
unified mathematical model, an optimal solution can be carried out to determine the effect of deviation in the
achieved that minimizes the rejection of non-conforming product design. If the part is not acceptable, salvage actions
fabricated parts. Regardless of the type and range of a have to be performed which includes rework, if possible.
design fit, the problem can be mapped into a matrix But if the non-conformance is reported for a mating
using a ranking function defined by the user. The geometry or dimension, suiting or finding a matching
ranking function is modifiable as per the user counterpart to achieve the required tolerance fit as per the
requirements and may vary based on the selection assembly design is a viable option. This matchmaking is
criteria for an assembly. Based on the type of ranking termed as suiting of parts. Figure 1 shows the various fits in
function used, the tabulated matrix is solved using the a mechanical assembly.
Hungarian minimization/maximization algorithm, which
is a powerful combinatorial optimization algorithm that For deviation from the design dimensions, rework or
solves the classical assignment problem in mathematics. rejection are the practical solutions manufacturers adopt.
This approach ensures maximum number of suiting However, rework is not always a feasible option in scenarios
pairs as well as nominal suiting of parts with each other involving mass production of components. Instead, after
resulting in high-quality products and maximum completion of dimensional inspections, the deviated parts
utilization of fabricated resources. (referred to as parts of type A) can be scrutinized for
compatibility with the corresponding deviated mating parts,
Keywords:- Optimisation, Part Suiting, Hungarian denoted as parts of type B. The dimensions of each deviated
Algorithm, Fit, Tolerance, Assembly. part A are systematically assessed against the dimensions of
part B to ascertain a proper fit. For n number of part A
fabricated, an appropriate pair can be identified from the set
of part B to have a Selective Assembly (SA).

Fig 1 Various Types of Fits in a Hole-Shaft Assembly

The conventional approach of creating such a suiting number of fabricated parts are made use of. Achieving an
list involves inspecting mating pairs to determine whether optimal solution demands iterative checking of the parts
the fit falls within an acceptable range. However, this through various combinations. Additionally, when dealing
method, especially in mass production scenario, is time with a significant number of deviated parts, the task
consuming and a complex exercise. The resultant solution becomes even more cumbersome.
may not necessarily be the optimal match where the most

IJISRT24AUG1039 www.ijisrt.com 1272

Volume 9, Issue 8, August – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24AUG1039

In this study, the problem of identifying suitable pairs computationally intelligent, rule-based machine learning
is transformed into a minimization assignment problem by system to solve the SA by taking Taguchi’s Loss function as
assigning ranks to fit values. Subsequently, the Hungarian criteria of acceptance along with the achieved tolerance.
algorithm, also known as the Munkres algorithm, is
employed to solve this assignment problem. As seen above, for component manufacturing with
bigger batch size, preparation of suiting pairs for a set of
The fit value, denoting the compatibility of a suiting deviated parts is time consuming and the suiting list so
pair, is calculated based on measured dimensions of mating arrived does not guarantee an optimal solution. All the
surfaces. The acceptability of a fit value is contingent upon present mathematical models and work existing guarantee
both the type of fit (Refer Figure 1) and the predetermined acceptance but not in its optimum. Here in this work, an
range of acceptable levels specified for the given assembly. optimal solution of part level suiting with product quality
adherence is ensured.
II. LITERATURE REVIEW
III. ASSSIGNEMENT PROBLEM
Over the last 40 years, a lot of fundamental research
effort has been dedicated to explore the mathematical basis The assignment problem represents a fundamental
of tolerance analysis. Due to the ever increasing combinatorial optimization challenge, commonly depicted
requirements on high precision manufacturing technology, through a graphical model known as a complete bipartite
SA has got attraction over the past years. The idea to deal graph, as illustrated in Figure 2.
and reduce the dimensional and geometric deviations in
product manufacturing buds from 1960s. Since this paper is In graph theory, a bipartite graph, also known as a
aimed to apply optimization to a SA problem of 2 parts, the bigraph, is defined as a graph in which the set of vertices
related works are only presented here. Mansoor (1961)[1] in can be partitioned into two distinct and independent sets,
his work, has explored the relation of manufacturing denoted as U and V[3][10].In this partition, every edge of the
machine tolerance to achievable part tolerances. He used an graph connects a vertex in set U to a vertex in set V. In a
example of the nozzle unit of a fuel injection pump and weighted bipartite graph, the edges between the two disjoint
categorized the components based on bore diameter. sets of vertices (U and V) have associated weights. Each
Segregation of mating parts of an assembly into selective edge is assigned a numerical value, indicating a certain
groups based on their deviations in dimensions were measure or cost associated with the connection between the
previously discussed by Mansoor 1961[1]; Fang and Zhang, corresponding vertices in sets U and V. Optimal assignment
1995[2]; Kannan et al., 2003[3]. Fang and Zang (1995)[2] entails matchmaking of a given size, where the sum of edge
firstly elaborated establishing dimensional parameter weights is either minimized or maximized in a weighted
relations and the possibility of SA to minimize rejections bipartite graph.
and avoid loss. The concept of Process Capability Indices
(PCIs) is studied and proposed as an intermediary to ensure
quality and statistical process control (SPC) parameters by
Zhang and Fang, 1999[4]. Kannan et al. (2003)[5] has
attempted the problem of SA using an optimization
technique of genetic algorithm. This is one among
evolutionary algorithms in operational research which starts
with a set of random matches and subsequently crossover
and mutates to find derived matches to reach near acceptable
criteria. Kannan et al. (2008)[6] applied the concept of
Taguchi’s quality loss function in hole shaft assembly and Fig 2 Complete Bipartite Graph
developed the mathematical models for clearance range in
terms of the quality loss function. They used a genetic If the number of vertices in set U equals the number of
algorithm to obtain the best combination in the SA. In the vertices in set V, the problem is termed a balanced
work by Tan and Wu, 2012[7], SA for multiple parts making assignment problem. However, if the counts differ, resulting
a single assembly is discussed. The problem of direct SA is in an unbalanced scenario, a conversion to a balanced form
studied to be of two variants: Direct SA (DSA) and Fixed becomes necessary before solving.
Bin SA (FBSA). The former is SA using information from
measurements on component characteristics directly, IV. FORMULATION OF SUITING PROBLEM
whereas the latter is SA of components sorted into bins. The
component matching problem for DSA is found as an axial The suiting problem is formulated into a matrix form to
multi-index assignment problem, whereas for FBSA, is an facilitate the application of optimization methods, allowing
axial multi-index transportation problem. Dantan et al., 2012 users the flexibility to choose their preferred optimization
[8]
in his review work, gives an overview of available approach. Consider a scenario with n instances each of part
mathematical models to solve SA problems. He also A and part B. The fitness values for all conceivable pairings
elaborates the limitations to extend the problem to a solvable can be systematically organized into a matrix, as
form. Babu and Asha (2015)[9] in their work has employed exemplified in Table 1, demonstrating the concept using
an artificial immune system algorithm which is a class of

IJISRT24AUG1039 www.ijisrt.com 1273

Volume 9, Issue 8, August – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24AUG1039

three instances of both part types. Subsequently, using the During a dimensional inspection, various locations of
available inspection reports for an array of parts, these the same dimension in a given part undergo measurement,
fitness values are populated within the matrix, constituting a and the resulting minimum and maximum values are
valuable dataset for optimization. documented in the inspection reports. Consequently, a single
dimension for a particular part is represented by two distinct
To elevate this matrix into a mathematically optimized values, denoting its minimum and maximum extents. In the
problem, a ranking system is introduced based on assessment of fits, particularly for interferences, both of
predetermined selection criteria. This ranking process, an these values demand consideration.
integral precursor to optimization, allows for the
categorization of data such that optimal values assume low For instance, in the case of evaluating interference
ranks in the case of minimization objectives, while between two parts, denoted as A (the shaft) and B (the hole),
unacceptable fits are assigned elevated ranks. The converse the minimum possible interference is calculated as 𝑩𝒎𝒂𝒙 −
holds for maximization algorithms, where superior fits 𝑨𝒎𝒊𝒏 , while the maximum possible interference is
garner higher ranks, and undesirable fits are relegated to determined as 𝑩𝒎𝒊𝒏 − 𝑨𝒎𝒂𝒙 .
lower ranks.
As each fit encompasses two distinct values arising
In scenarios where the quantity of parts available for from the minimum and maximum measurements of a given
pairing is uneven, dummy parts are introduced to convert the dimension, two rank matrices are constituted to address the
problem into a square matrix. Notably, the rank values minimum & maximum acceptability range. The effective
assigned to dummy parts mimic those of unacceptable fits. ranking matrix which is used for optimization is determined
Following the establishment of a balanced ranked matrix, by the summation of both rank matrices. To ensure the
optimization procedures can be instituted. positivity of the cumulative value and prevent nullification
due to cancellation, the magnitude of each value is
considered for the summation process.

Table 1 Matrix Formulation Based on Fit Values for a Two-Part Suiting

Parts B1 B2 B3
A1 Fit for A1 with B1 Fit for A1 with B2 Fit for A1 with B3
A2 Fit for A2 with B1 Fit for A2 with B2 Fit for A2 with B3
A3 Fit for A3 with B1 Fit for A3 with B2 Fit for A3 with B3

V. RANKING FUNCTION 𝑅 - Rank

Any selection problem can be reformulated into an 𝐹 - Fit obtained for given pair
optimization assignment problem through a systematic
conversion to its optimizable matrix form. This 𝑅𝑎𝑣𝑔 - Theaverage of accepted fit band limits
transformation holds significant import, as the criteria for
selection or assignment may vary from one problem to 𝑀𝑖𝑛𝑙𝑖𝑚𝑖𝑡 is the minimum limit in the accepted fit band
another. The nuanced variations in these criteria must be
effectively captured through a mapping function. 𝑀𝑎𝑥𝑙𝑖𝑚𝑖𝑡 is the maximum limit in the accepted fit band

In the context of the current part suiting problem, a The above function is similar to the famous Taguchi’s
specific range of acceptable fits must be allotted a lower Loss function, 𝐿 = 𝑘(𝑦 − 𝑚)2 , where m is the theoretical
rank, while unacceptable fit ranges are assigned a higher 'target value' and y is the actual size of the product, k is a
rank. Given the adoption of a minimization algorithm, the constant and L is the loss[6]. So in equation (1) loss is similar
chosen ranking function should map the most optimal fit to to Rank (R), theoretical target value is similar to average of
the minimum rank. Within the acceptable fit band, the accepted fit band limits (𝑅𝑎𝑣𝑔 ), and the actual size of the
optimal fit is defined as the mid-value, and the rank product is similar to the fit obtained for the given pair (F).
incrementally increases towards both sides within the band Hence on minimization of rank, we are essentially
limits. Conversely, the unacceptable fit ranges are assigned maximizing the quality of the fit. Once the minimum and
an exceptionally high rank. The adopted mapping for this maximum fit values are processed through the mapping
particular problem is illustrated in Figure 3 and is function, the resulting rank matrices for both the minimum
mathematically expressed by the following equation: and maximum fits are combined using summation to
determine the cumulative effect of the fit combination's
|𝑭 − 𝑹𝒂𝒗𝒈 |, acceptability. In the consolidated matrix, all elements
𝑅= 𝑓𝑜𝑟 𝐹 Є [𝑀𝑖𝑛𝑙𝑖𝑚𝑖𝑡 , 𝑀𝑎𝑥𝑙𝑖𝑚𝑖𝑡 ] …..(1)
exceeding or equal to 1000 times the maximum limit are
𝟏𝟎𝟎𝟎 ∗ 𝑴𝒂𝒙𝒍𝒊𝒎𝒊𝒕 , preserved to uphold uniformity in rejection criteria. The
conclusive rank matrix is subsequently subjected to the
{𝑓𝑜𝑟 𝐹 Є (−∞, 𝑀𝑖𝑛𝑙𝑖𝑚𝑖𝑡 ) ∪ (𝑀𝑎𝑥𝑙𝑖𝑚𝑖𝑡 , +∞)}
Hungarian optimization technique.

IJISRT24AUG1039 www.ijisrt.com 1274

 Volume 9, Issue 8, August – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24AUG1039

to its most reduced form, ultimately leading to the

determination of the optimal solution [11]. Subsequently, the
assignment operation is performed.

Hungarian optimization algorithm ensures the gradual

refinement of the assignment until an optimal solution is
achieved, covering the nuances of the assignment problem
systematically.

VII. A SAMPLE SUITING PROBLEM:

FORMULATION AND RESULTS

Consider a sample suiting having five numbers of each

component within a mating pair, as illustrated in Figure 4.
The acceptable interference fit range for the parts is 10-35μ.
Φ10.8 is the mating dimension for the parts.

Fig 3 Minimization Mapping (Piece-Wise Linear Graph) The available dataset comprises the inspected values of
mating dimensions, as outlined in Table 2.
Typical minimization mapping using a piece-wise
linear graph is illustrated in Figure 3. The mapping process
assigns optimal and acceptable values to a lower rank,
resembling a dip, while unfavourable or unacceptable values
are allocated to a higher rank. The best fit (Ravg) receives the
lowest rank in this ranking scheme.

In cases for 2 parts having multiple mating interfaces

(Eg: pitch circle diameter holes) ranking can be extended by
adding ranks of individual mating dimensions.

VI. HUNGARIAN OPTIMIZATION ALGORITHM

The Hungarian method is a polynomial-time

optimization algorithm designed to solve assignment
problems efficiently. Widely applicable to various primal- Fig 4 Typical Example of Mating Pairs with an Interference
dual optimization problems, the algorithm revolves around a Fit
series of matrix operations aimed at simplifying the matrix

Table 2 Inspected Dimensional Values of Mating Surface for Parts A & B (5 nos. each)
Part P Part Q
ø10.8 +0.035/+0.025 ø10.8 +0.015
Idn.
Idn. No. Min Max Min Max
No.
Pmin P max Q min Q max
P1 10.822 10.825 Q1 10.793 10.801
P2 10.823 10.826 Q2 10.798 10.811
P3 10.825 10.827 Q3 10.791 10.807
P4 10.824 10.828 Q4 10.792 10.805
P5 10.826 10.829 Q5 10.790 10.804

For this particular problem, all the possible pairing combinations of the part P and part Q are listed as in Table 3. Finally 5
optimal pairs have to be identified in such a way that each part P is matched with a unique part Q. In the below combination
matrix, the optimization is to be done with respect to their fit value.

Table 3 Possible Suiting Combinations

Q1 Q2 Q3 Q4 Q5
P1 P1-Q1 P1-Q2 P1-Q3 P1-Q4 P1-Q5
P2 P2-Q1 P2-Q2 P2-Q3 P2-Q4 P2-Q5
P3 P3-Q1 P3-Q2 P3-Q3 P3-Q4 P3-Q5
P4 P4-Q1 P4-Q2 P4-Q3 P4-Q4 P4-Q5
P5 P5-Q1 P5-Q2 P5-Q3 P5-Q4 P5-Q5

IJISRT24AUG1039 www.ijisrt.com 1275

Volume 9, Issue 8, August – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24AUG1039

For an interference fit problem, the minimum interference is given by 𝑷𝒎𝒊𝒏 − 𝑸𝒎𝒂𝒙 and maximum interference is given by
𝑷𝒎𝒂𝒙 − 𝑸𝒎𝒊𝒏 . For all the combinations as in Table 3, a minimum interference matrix is tabulated and ranked using equation (1) to
generate M1 matrix (Table 4). Similarly, maximum interference matrix is tabulated and ranked to generate M2 matrix (Table 5).

Fig 5 Minimization Mapping using Ranking Function

Table 4 Matrix M1
Q1 Q2 Q3 Q4 Q5
P1 2 12 8 6 5
P2 1 11 7 5 4
P3 2 9 5 3 2
P4 1 10 6 4 3
P5 3 8 4 2 1

Table 5 Matrix M2
Q1 Q2 Q3 Q4 Q5
P1 10 5 12 11 13
P2 11 6 13 12 35000
P3 12 7 35000 13 35000
P4 13 8 35000 35000 35000
P5 35000 9 35000 35000 35000

These matrices are added to give the net ranks. The elements greater than 1000 ∗ Maxlimit are changed to 1000 ∗ Maxlimit
itself to avoid unnecessary iterations in computation as they all represent unacceptable pairs and are hence similar. The final M
matrix in Table 6 gives the rank of fits against all the combinations.

Table 6 M1+M2=M
Q1 Q2 Q3 Q4 Q5
P1 12 17 20 17 18
P2 12 17 20 17 35000
P3 14 16 35000 16 35000
P4 14 18 35000 35000 35000
P5 35000 17 35000 35000 35000

Further employing Hungarian algorithm as explicated in Section 5, the M matrix is solved and is depicted in Table 7.

IJISRT24AUG1039 www.ijisrt.com 1276

Volume 9, Issue 8, August – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24AUG1039

Table 7 Optimised Solution

Q1 Q2 Q3 Q4 Q5
P1 0 5 0 3 0
P2 0 5 0 3 34982
P3 0 2 34978 0 34980
P4 0 4 34978 34984 34980
P5 34983 0 34975 34981 34977

Hence, the SA identified is tabulated in Table 8 with minimum and maximum interference value.

Table 8 Selective Assembly of the Sample Problem

Sl. No. Part 1 Part 2 Min. Int. Max. Int.
1 P1 Q5 18 35
2 P2 Q3 16 35
3 P3 Q4 20 35
4 P4 Q1 23 35
5 P5 Q2 15 31

VIII. CONCLUSIONS FUTURE SCOPE

In this study, the two-part suiting problem within an The suiting problem addressed in this study, on mating
assembly is transformed into a minimization assignment pairs, holds the potential for extension to encompass an
problem by assigning a rank to the design fit value. entire assembly, beginning with a minimum of three parts.
Subsequently, the problem is resolved utilizing the As the 2-part suiting problem unfolds into a 2-dimensional
Hungarian algorithm. The key observations derived from the matrix assignment problem, scaling up to an n-part selection
implemented method of suiting are outlined below: problem introduces an assignment problem encapsulated
within an n-dimensional matrix. The formulation of
 The developed method demonstrates a capacity to yield solutions and the algorithms implicated in this expanded
optimal solutions, encompassing a maximal number of context become subjects of heightened interest.
viable suit pairs and achieving optimal matches.
 The computational formulation of the entire suiting The prospect of devising a computational method
problem, transitioning from manual processes to tailored to handle assembly challenges involving multiple
computational algorithms, notably diminishes the parts holds significant promise. With comprehensive
requisite time compared to traditional manual suiting inspection data available for all groups of parts, the
methodologies. envisioned methodology allows for the optimal solution and
 The method allows in experimenting with the fit values assembly of any given pairs in a singular execution. Such a
giving an ability to be selectively stringent or lenient computational approach not only ensures optimization but
with the tolerance values as demanded by the also facilitates the refinement and perfection of assembly
application. procedures. As elaborated by Tan and Wu, 2012[7] Fixed Bin
 Designs with micron level tolerance fit, say 2µ to 3µ as Selective Assembly of n parts, is an axial multi-index
in fuel pump component parts of automobiles is achieved transportation problem requiring further research.
by suitably machining with the achieved dimensions of
its mating part. Introduction of this suiting software REFERENCES
liberates such production processes from the tedious in-
situ suiting methods to cost-effective batch production [1]. Mansoor, E.M. (1961) “Selective assembly – its
processes. analysis and applications“, International Journal of
 The mapping function used for the acceptable tolerance Production Research, Vol. 1, No. 1, pp.13–24,
range is similar to Taguchi’s Loss function, 𝐿 = doi:10.1080/00207546108943070.
𝑘(𝑦 − 𝑚)2 . Thus the presented method ensures SA is of [2]. Fang, X.D. and Zhang, Y. (1995) “A new algorithm
tolerance fit with minimum quality loss, ie. maximum for minimising the surplus parts in selective
quality and maximum suiting pairs compared to existing assembly“, Computers and Industrial Engineering,
mathematical algorithms implemented. Vol. 28, No. 2, pp.341–50.
 The devised formulation delivers solutions through the [3]. Bondy, Jhon Adrian; Murty, U.S.R (1976 ), “Graph
utilization of the Hungarian algorithm. Given the Theory with Applications”, ISBN 0-444-19451-7,
polynomial time complexity of the algorithm it is page 5.
imperative to acknowledge that the computational time [4]. Zhang, Y. and Fang, X.D. (1999) “Predict and
escalates proportionally with the number of parts slated assure the matchable degree in selective assembly
for pairing [12]. via PCI-based tolerance“, Journal of Manufacturing
Science and Engineering, Vol. 121, No. 3, pp.494–
500.

IJISRT24AUG1039 www.ijisrt.com 1277

Volume 9, Issue 8, August – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24AUG1039

[5]. Kannan, S., Jayabalan, V. and Jeevanantham, K.

(2003) “Genetic algorithm for minimizing
assembly variation in selective assembly“,
International Journal of Production Research,Vol.
41, No. 14, pp.3301–13, doi:10.1080/
0020754031000109143.
[6]. Kannan, S.M., Jeevanantham, A.K. and Jayabalan, V.
(2008) “Modelling and analysis of selective assembly
using Taguchi’s loss function“, International Journal
of Production Research,Vol. 46, No. 15, pp.4309–30,
doi: 10.1080/00207540701241891.
[7]. Tan, M.H.Y. and Wu, C.F.J. (2012) “Generalized
selective assembly“, IIE Transactions, Vol. 44, No. 1,
pp.27–42, doi:10.1080/0740817X.2010.551649.
[8]. Dantan, J-Y., Gayton, N., Dumas, A., Etienne, A. and
Qureshi, A.J. (2012) “Mathematical issues in
mechanical tolerance analysis“, Proceedings of 13th
National AIP Primeca Conference, No. 1, pp.1–12.
[9]. Babu, J.R. and Asha, A. (2015) “Minimising
assembly loss for a complex assembly using
Taguchi’s concept in selective assembly“,
International Journal of Productivity and Quality
Management, Vol. 15, No. 3, pp.335–56.
[10]. Diestel, Reinhard (2005), “Graph Theory” (3rd ed.),
ISBN 3-540-26182-6. Electronic edition, page 17.
[11]. H. W. Kuhn. “The Hungarian Method for the
Assignment Problem”.
[12]. James Munkres. “Algorithms for the Assignment and
Transportation Problems”.

IJISRT24AUG1039 www.ijisrt.com 1278