ORIGINAL ARTICLE

Optimisation process for robotic assembly of electronic components

K. T. Andrzejewski1 & M. P. Cooper1 & C. A. Griffiths1 & C. Giannetti1

Received: 21 February 2018 /Accepted: 26 August 2018 /Published online: 5 September 2018# The Author(s) 2018

AbstractAdoption of robots in the manufacturing environment is a way to improve productivity, and the assembly of electroniccomponents has benefited from the adoption of highly dedicated automation equipment. Traditionally, articulated 6-axisrobots have not been used in electronic surface mount assembly. However, the need for more flexible production systemsthat can be used for low to medium production builds means that these robots can be used due to their high degrees offlexibility, excellent repeatability and increasingly lower investment costs. This research investigated the application ofan articulated robot with six degrees of freedom to assemble a multi-component printed circuit board (PCB) for anelectronic product. A heuristic methodology using a genetic algorithm was used to plan the optimal sequence andidentify the best location of the parts to the assembly positions on the PCB. Using the optimised paths, a conditionmonitoring method for cycle time evaluation was conducted using a KUKA KR16 assembly cell together with fourdifferent robot path motions. The genetic algorithm approach together with different assembly position iterations iden-tified an optimisation method for improved production throughput using a non-traditional but highly flexible assemblymethod. The application of optimised articulated robots for PCB assembly can bridge the gap between manual assemblyand the high-throughput automation equipment.

Keywords Sequencing optimisation . Electronics assembly . KUKA robotics . Flexible manufacture . Genetic algorithm

NotationsPTP Point to pointSPTP Spline point to pointLIN LinearSLIN Spline linearGA Genetic algorithmSCARA Selective compliance assembly robot armIC Integrated circuitSIL Single in linePCB Printed circuit boardDOF Degree of freedomKRL KUKA Robot Languageα (alpha) Degree of rotationd Distanceh Height

1 Introduction

Electronics manufacturing has evolved over the past yearsfrom a labour-intensive activity to a highly automated one.In printed circuit board (PCB) assembly, the electronic com-ponents need to be placed into position prior to soldering, andthe use of automation provides accuracy, repeatability andefficiency to this process, when compared to manual assem-bly. Furthermore, the competition faced by electronic compo-nentmanufacturers causes a need for high-throughput rates forwhich automated assembly lines are a major asset [1].Implementation of robotics in assembly of electronics offerssome distinct advantages over manual methods due to its re-liability and flexibility. Assembly of electronic components isgenerally limited to two-dimensional horizontal x- and y-axismovements and usually comprise of several sub-systems, in-cluding part feeding systems, work holding and pick and placedevices. Depending on the application, various types of robotscan be used for pick and place operations, e.g., gantry/Cartesian robot, cylindrical robot, spherical robot, SCARA(Selective Compliance Assembly Robot Arm) and spider ro-botic arms [2, 3]. Despite its popularity, the articulated robot

* C. A. Griffithsc.a.griffiths@swansea.ac.uk

1 College of Engineering, Swansea University, Swansea, UK

The International Journal of Advanced Manufacturing Technology (2018) 99:2523–2535https://doi.org/10.1007/s00170-018-2645-y

type has beenwidely neglected for printed circuit board (PCB)assembly due to its relatively higher equipment prices.However, articulated robots with a much larger work envelopewhen compared to Cartesian and SCARA have the potentialto becomemore popular in the electronics industry due to highdegrees of flexibility, excellent repeatability and now also dueto lower investment cost and competitive prices.Understanding the costs and benefits of multiple degrees offreedom, especially those that create redundancy, is a funda-mental problem in the field of robotics [4]. Redundancy iswhen the robot has more degrees of freedom than needed forperforming the tasks [5]; therefore, a modern automation cellincorporating an articulated 6-axis robot needs has the free-dom to not only execute traditional automation activities butalso perform specialist assembly operations. In addition toimproved redundancy and space control, current technologyhas resulted in articulated robots being able to work in closeproximity to human workers. Collaborative robots allow thepossibility for assembly lines with close contact between hu-man and robot, this modern manufacturing development al-lows for the high performance of robots together with theunique capabilities of people.

The focus of this work is on low-volume assembly of elec-tronic parts. Low-volume production challenges are importantas new industrial strategies for high-value twenty-first centuryproducts need to consider flexible manufacturing methods.The test part in this research includes a front panel and twoprinted circuit boards connected through single in-line (SIL)headers. Each of the PCBs consists of numerous parts, such asswitches, jack ports, potentiometers and electronic compo-nents; and resistors, capacitors, transistors and fuses. Thisproduct lends itself to mass production systems that tradition-ally would have used production lines with either highmanuallabour costs or high investment specialised automation.However, in this automation research, a non-specialised sys-tem is used. In this way, the research will investigate the ap-plication of a robotic system that can be flexible in producingmultiple part and build options, while meeting the precision,efficiency, reliability and repeatability of a dedicated automa-tion system.

The finished product consists of multiple assembly opera-tions with several possible build routes; consequently, themain contribution of this paper is the application of a heuristicmethodology for planning the assembly operation using anarticulated KUKA robot with six degrees of freedom in apick-and-place operation. The introduction of robots into in-dustry seeks to upgrade not only the standards of quality butalso productivity, as working time is increased and idle orwasted time is reduced [6]. For this reason, path planningand the improvement of the autonomy is a fundamental issuein robotics [6, 7]. The common criterion for optimisation ismake-span minimisation or, in the context of repetitive assem-bly, cycle time minimisation [1, 8–11]. The preliminary study

prior to automation included a manual assembly of the prod-uct to identify fundamental issues involved in the process.This approach was necessary to provide a description of thegeneric steps involved in the assembly and to translate thoseinto the automated environment. It has been concluded that inprinciple, infinite joint configurations may result in executingthe same task [5] and that the production planning processoptimisation must determine the following sub-problems:

& The correct allocation of component feeders& Identify the optimum component placement sequence& The optimum orientation and work space arrangement

The paper is structured as follows. First, there will be areview of the feasibility of optimising the process using agenetic algorithm. This is followed by a description of theexperimental setup including the test part, robot and theworkspace arrangement. Then, the methodology adopted isprovided with a problem representation and description ofthe GA used for the feeder assignment and placement se-quence. A results section discusses the performance of theGA and the solutions for optimal motion type, orientation ofwork piece and the objective function for the feeder allocationand location of components. Finally, conclusions are made onthe proposed system.

2 Feasibility of the genetic algorithmin process optimisation

The genetic algorithm technique dates back to 1975 and wasfirst introduced by Holland [12]. In the literature, the geneticalgorithm has been widely used to solve optimisation prob-lems in many industrial applications such as job shop se-quencing and scheduling [13], assembly planning [14], selec-tion of machining parameters in turning operation [15], designof sheet-metal assembly and machining fixtures [16] and tol-erance design [17]. The processes of optimisation for thefeeder-slot allocation and component sequencing are the twomost important factors for improving the efficiency of thisassembly operation. In the literature, some of research focusedon only one of the two problems, while assuming the solutionto the other, is given in advance. Ball and Magazine [18]assumed that the feeder arrangement was fixed and solvedthe placement sequencing problem. Their solution to the prob-lem was modelled as a rural postman problem with a heuristicapproach used to solve it. Ahmadi and Mamer [19] modelledthe problem of sequencing the component types for placementand the problem of scheduling the movements between pointson the PCB as a collection of interdependent traveling sales-man problems. Klomp et al. [20] solved the component allo-cation problem for a turret type machine. Li et al. [21] consid-ered a pick-and-place machine with a revolving head. The

2524 Int J Adv Manuf Technol (2018) 99:2523–2535

approach to the placement sequencing was modelled as a trav-elling salesman problem. Then, in the second stage, theyemployed a genetic algorithm to solve the feeder arrangementproblem. Other research work has focused on tackling bothproblems (the feeder arrangement problem and the placementsequencing problem) by iterating their solutions. Grunowet al. [22] considered a pick-and-place manipulator with arevolving head and solved both problems iteratively. Theyused simple heuristics to obtain the initial feeder arrangementand then solved the placement sequence as a simple vehiclerouting problem.

Finally, many researchers proposed heuristic ormetaheuristic methods for solving the feeder arrangementand placement sequencing problems simultaneously. Sohnand Park [23], which considered a turret-type machine,adopted an integrated approach in their solution. They solvedthe feeder arrangement and placement sequencing problemsintegrally for a sequential pick-and-place machine. A simplepairwise exchange heuristic was used to solve the feeder ar-rangement, while the evaluation of each feeder arrangementsolution was performed by finding the placement sequenceusing a modified farthest insertion heuristic. An integratedInteger Programming method was established by Broadet al. [24]. Their model was solved by a binary integer pro-gramming package. Deo et al. [25] proposed a genetic algo-rithm for a complex problem of component sequence optimi-sation in multiple setups, which were necessary with limitedfeeder holding capacity. Ho and Ji [26] hybridised the geneticalgorithm with the nearest neighbour heuristic, the 2-opt heu-ristic and an iterated swap procedure. Ellis et al. [27] devel-oped a heuristic for solving the feeder arrangement and theplacement sequencing problems simultaneously for a turret-type machine. A construction procedure was used with a set ofrules to generate an initial placement sequence and feederallocation. Their model also contained an improvement pro-cedure to the initial solution. Magyar et al. [28] proposed ahierarchical solution approach to solve the problem of deter-mining the placement sequence and feeder arrangement. Sunet al. [29] studied a dual-gantry collect-and-place machine.They used a genetic algorithm to decide the component allo-cation between the two revolving. They proposed a greedyheuristic for the work cycle formation and pickup sequencingdecisions achieved by maximising the number of simulta-neously picked up components for each access of a multi-head module, or equivalently minimising the number ofpickups, and balancing the workload between the two gan-tries. Kulak et al. [30] proposed genetic algorithms for bothsingle-gantry and dual-gantry collect-and-place machineswith revolver-type placement heads. The feeder arrangementproblem and the placement sequencing problem were solvedby a novel genetic algorithm approach, which integrates aclustering algorithm for generating sub-sections of the PCBand grouping the corresponding placement operations.

Garcia-Naijera and Brizuela [31] proposed an efficient geneticalgorithm to solve the problem of component sequencing withfeeder arrangement. Their computational experiments showthat the algorithm they developed improves the state-of-the-art result on a benchmark for the problem.

It can be concluded from the above survey that research hasbeen conducted to investigate printed circuit board assemblyoptimisation problems. However, no GA optimisation re-search has been performed using a single 6-axis robot, andno theoretical results have been validated in a real-life setting.The need for this is clear as the 6-axis robot allows for a wholeproduction cell that allows for the widest possible autonomyandmanufacturing capability. This non-traditional approach isof particular importance when considering low-volume pro-duction and prototype builds with a wide range a productvariation.

3 Experimental setup

3.1 KUKA robot

The study was conducted using a KUKA KR16 6-axis indus-trial robot. The KUKA industrial robots are highly accuratewith a position repeatability of ± 0.04 mm [31]. This is aprecision suitable for the test part assembly. Once the se-quence and feeder slot designation had been produced fromthe genetic algorithm, the path had to be programmed into theKUKA robot. The locations of all the feeders and componentpositions were saved as way points using the teach methodand within the sequence, timers commands were added to startrecording at the first point and stop recording at the end. TheKUKA robot offers six different types of path motion. Two ofthese will be disregarded as they are circular motions, whichare not relevant to the proposed problem. The motion typesavailable for the electronics assembly are as follows:

& Point to point (PTP)—This motion type involves follow-ing the quickest path between two points and is not nec-essarily a straight path.

& Linear (LIN)—The linear motion type follows a straightpath and uses more joints in constant motion to trace thestraight path.

& Spline point to point (SPTP)—This is similar to the PTPmotion; however, it allows for continuous spline motionswhere points are estimated and a smoother motion isavailable.

& Spline linear (SLIN)—As with the SPTP, this motion typeuses splines between linear motions.

The effects on the cycle time for each motion type weretested using the same sequence and feeder arrangement. Bymaking this comparison, a solution can be obtained that not

Int J Adv Manuf Technol (2018) 99:2523–2535 2525

only reduces the cycle time through good placement sequenc-ing, but also through direct optimisation of movement usingthe best motion type.

3.2 Demonstrator test part

The Eurorack Serge filter was chosen as the focus of the studydue to its suitable complexity and variety of components. Theassembly operation of the printed circuit board can beachieved using a variety of placement technologies; however,the main focus of this paper is to present the procedure offinding the best solution for a 6-axis robot. The EurorackSerge filter consists of two PCBs (Fig. 1) with 100 compo-nents between them that have to be placed and soldered. Thefirst PCB consists mostly of resistors, capacitors and diodes,while the second has larger components such as jack ports andpotentiometers. The components vary greatly in size andshape and thus cannot be placed with the same end effector.Twelve component feeders are required, and the product pro-vides a good representation of the PCBs that would beprototyped or produced in small batches.

3.3 Workspace arrangement

The PCBs were positioned on a 58 cm × 58 cm worktable at aheight of 80 cm; using the centre of the table as a body frame,the initial distance from the robot axis 1 global frame was100 cm. With a fixed position for the component feeders, thePCBs are arranged with the long and short edges adjacent tothe feeders at an initial distance of 10 cm (Fig. 2). The twoarrangements are expected to produce different results due tothe total distance between the feeder of a particular componenttype and the designated place on the board being different forboth cases. By means of a simple experiment in which the

proposed genetic algorithm is run with both orientations, thebest option can be chosen based on the fastest solution. Thearrangement of the entire work piece (feeders and PCB) inrelation to the base of the KUKA robot can also be optimised.By varying the angle (alpha) (α) by which the PCB andfeeders are arranged to the KUKA robot, the optimal orienta-tion can be determined. Similarly, the proximity of the workpiece to the base also has an effect which can be determined.Both considerations were investigated using an experimentalmethod of trialing various positions using the KUKA robot.The angle problemwas also investigated using a computation-al method to relate the angle of rotation to the total distancetravelled in a cycle.

4 Methodology

4.1 Problem representation

In order to design an efficient assembly system for a particularprinted circuit board, two distinct but related problems need tobe solved. First is the reduced distance allocation of the com-ponent types to the feeder positions. The distance between theindividual component pick up location and its destination onthe PCB depends on the feeder rack arrangement, the distancebetween the feeder and PCB and the orientation of the circuitboards on the holding table [24]. The PCB assembly problemconsists of placing a number of electronic components of pre-specified types at set locations on a PCB. The core decisionvariables in the feeder allocation problem refer to the alloca-tion of individual component feeders to positions in the mag-azine [31]. The goal is to minimise the total distance and thusreduce the corresponding cycle time. The next problem is todetermine the sequence in which the components are placedinto the boards. The assembly operation works as follows: Theend effector traverses to the feeder rack, from which it picksup a component and travels to place the component on thePCB. After completing the placement tour, the steps are re-peated for a new placement tour [18]. There are, however,some major assumptions and operation specifications of theassembly system, upon which this investigation is based:

& Each component type is setup only once in the feedermagazine for simplification, as this approach eliminatesa component retrieval problem

& The robotic arm is equipped with the set of end effectorsneeded to pick up all types of components one componentat a time.

& Each component type can be picked up with a subset oftools, that is, one head with a specific tool can only pickupcomponents from a limited set of component types [20]. Ithas been concluded that five sub-groups made out ofFig. 1 Eurorack Serge filter PCBs

2526 Int J Adv Manuf Technol (2018) 99:2523–2535

similar design features of the component types can bedistinguished.

& The assembly operation starts with the smallest compo-nents due to basic assembly rules established in the man-ual build.

4.2 The genetic algorithm

Due to the combinatorial nature of the problem studied in thispaper, an approach based on the genetic algorithm was devel-oped to solve both the feeder slot assignment and componentsequencing problem. In this method, a candidate solution isrepresented as an individual with a set of properties calledchromosomes, and the group of individuals is called a popu-lation. The population evolves to the next generation throughcrossover and mutation with the two-point crossover methodadopted, the mutation operates on the offspring created in thecrossover step based on set probability [32]. After the repro-ductive operations, the fitness of the offspring is assessed andcompared against the parents. Based upon survival of the fit-test rule, the best individuals of the offspring are selected forthe next generation. The genetic algorithm runs until termina-tion criteria are satisfied or no improvement is observed dur-ing a number of generations. It is the associated cost thatdetermines the fitness of the individual in population. Thelower the cost, the higher the fitness is. In the optimisationof the robotic assembly, the cost is often cycle time [1, 8–11]and the associated total distance the robot manipulator travelsduring an assembly operation [25, 33]. In the model adopted,the best individuals are chosen to minimise the associated costand provide the shortest path between two points wasemployed. The cost is evaluated as the sum of all distancesbetween the feeder and component locations on the PCB forthe feeder slot assignment problem. In the placement sequenceproblem, the objective function is a sum of all the distances

between each component in the sequence and the feeder loca-tion of the next component in the sequence. The objectivefunction for the first problem can be expressed as follows:

Min∑ni¼1∑

mj¼1

ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi

xi−x j� �2 þ yi−y j

� �2þ zi−z j� �2

r

ð1Þ

where

n length of individual (number of feeder slotsin feeder rack)

m total number of all componentsxi, yi, zi Cartesian coordinates of the position of ith

feederxj, yj, zj Cartesian coordinates of the destination of

jth component

The objective function for the placement sequence problemcan be expressed as:

Min d f ;c þ dc; fþ1

� � ð2Þ

d f ;c ¼ ∑ni¼1∑

mj¼1

ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi

xi−x j� �2 þ yi−y j

� �2þ zi−z j� �2

r

ð3Þ

dc; f ¼ ∑ni¼1∑

mj¼1

ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi

xi−x j� �2 þ yi−y j

� �2þ zi−z j� �2

r

ð4Þ

where

n vector of integers describing sequence ofcomponent placement destinations

m vector of integers describing consequentfeeders from which component are picked

xi, yi, zi Cartesian coordinates of the position of ithfeeder

Fig. 2 (a) Orientation of the workpiece in 3D space in relation to therobot. (b) Orientation of assemblyprocess when rotated by an angleof α (alpha)

Int J Adv Manuf Technol (2018) 99:2523–2535 2527

xj, yj, zj Cartesian coordinates of the destinationof jth component

df, c sum of distances describing feeder tocomponent path

dc, f sum of distances describing component tofeeder path

The developed genetic algorithm (Fig. 3) takes all the in-puts from the coordinates of all components and feeders. Thealgorithm calculation returns the optimal feeder and sequencesolutions, and the output is then optimised in accordance withthe pre-set performance parameters. Those include initial pop-ulation size, number of generations, crossover and mutationprobabilities, and the type of the objective function (minimumdistance the end-effector travels) corresponding to it total as-sembly time.

4.3 Feeder slot assignment

The core variable in the assembly optimisation refers to thearrangement of feeder slot locations to the component types.As mentioned, the goal is to find suitable allocation of thecomponent types to feeder slots which minimises the totaldistance the pick and place manipulator travels and thus thecorresponding cycle time [28]. The algorithm analyses theneighbourhood relations between the different types of com-ponents and the corresponding placement locations on theboards. The idea behind this heuristic approach is to arrangecomponent types, which are characterised by strong

neighbourhood relations, adjacent to each other in the compo-nent magazine [22]. In the example device, 12 componenttypes were distinguished. Therefore, each gene within thechromosome represents a number corresponding to the feederfor a specific set of components, and the order of numberswithin the chromosome corresponds to the feeder locationarrangement.

4.4 Placement sequence

After the feeder arrangement has been determined, the nextstep is to find the sequence of placing the individual compo-nents on the circuit boards. The objective again is to minimisethe total distance travelled by the placement tool. For thisparticular problem, each gene was assigned a component IDnumber. To ease the computational time, several constraintsand assumptions were employed. Each component ID wasassigned with a tool number, which corresponds to the endeffector suitable for the pick-and-place operation. End effectorchangeover time negatively affects cycle time; therefore, toimprove productivity, it is essential to minimise tool change-over for different components. To facilitate this constraint, thealgorithm groups the components based on their tool numberand searches for optimal sequencing solutions within eachgroup. As the total number of components to be placed on aboard is 100, there are actually 105 genes in each chromo-some, which drives the number of possible combinations to anextremely large number. The first 100 numbers correspond tothe component IDs, with the last 5 corresponding to the tool

Fig. 3 Genetic algorithm inputs and outputs

2528 Int J Adv Manuf Technol (2018) 99:2523–2535

sequence. Once the solution is obtained, the heuristic calcu-lates the neighbourhood relationship between marginal com-ponents. The suitable tool sequence is defined based onminimising the distances between the last component in eachsub-group and location of the magazine from which the firstcomponent in the next tool-group is being collected. Figure 4illustrates the chromosome structure based on the cluster ap-proach described above.

5 Results

5.1 Genetic algorithm reliability

The genetic algorithm has proven to be a reliable tool in find-ing good quality solutions to two combinatorial problems. Thefeeder slot assignment problem has been solved, and for thequickest solution, the feeder slots are positioned well to avoidunnecessary travel. The placement order of the componentshas also been optimised to reduce the total path distance trav-elled. The genetic algorithm has been used to find the shortestand conversely longest paths. These represent the “best” and“worst” solutions. Between these solutions, there is over a 2-simprovement per cycle for the SLIN motion type, which iscommonly used in this sort of operation (Table 1).

5.1.1 GA results—feeder slot assignment

The results obtained using the genetic algorithm code for thefeeder slot assignment problem showed that the initial popu-lation indicates the group of the six fittest individuals gener-ated through random permutations. A relatively high distancevalue was observed. The “distance” gives indication of thevalue produced by the cost function, and its reciprocal is pre-sented as “fitness.” With each evolution of the population,new feeder slot arrangements of the higher value of fitnessare found. The 14th population yielded the chromosome withthe highest fitness value and therefore the suitable solution tothe feeder allocation problem. Repeatability analysis of theresults obtained for the feeder slot arrangement problem ispresented in Fig. 5. The upper graph shows the collectiveresult of the evolution of the population shown as fitnessagainst generations. The performance analysis was based ondata collected from a batch of 100 runs of the genetic algo-rithm with population size 200, crossover probability 0.8 andmutation probability 0.6. The algorithm has shown

convergence toward the optimal solution with a probabilityof 80%, and the standard deviation calculated for results oftotal distance travelled was 0.01425 m which accounts for0.16% of the total distance.

Parameter sensitivity analysis was carried out on the pro-grammed genetic algorithm using the placement sequenceproblem. The initial population size for this analysis was setto 200, and the number of generations/iterations was set to1000 to ensure quality of the results. Figure 6 presents thefindings of the crossover probability investigation. The parentsolution with high fitness is crossed over more often whichincreases the algorithms ability to explore a larger space ofsolutions, and large space search reduces chances of settingfor local optimum. It can be concluded that a higher crossoverprobability improves the quality of the results and produces abetter solution. The mutation probability plot, presented inFig. 7, also shows benefits of applying higher rates. The plotindicates that a search with lower mutation rates, of the datafor mutation probability of 0.2 and 0.4, weakens the usefulperturbation among the potentially promising genes. In a re-sult, many solutions are never tried out. However, setting themutation rate too high increases the risk of offspring losingresemblance to their parent chromosomes, resulting in rapidconvergence and significant loss of the algorithm’s ability tolearn from the history search (data points for mutation proba-bility of 0.6 and 0.8).

On the basis of the examinations that were performed onthe initial population size, it can be concluded that for thefeeder slot assignment, the genetic algorithm remains unre-sponsive to increase in the population size. However, for theplacement sequencing problem, the initial population size wasproven to be a significant factor in finding a quality solution.In Fig. 8, it can be seen that the fitness of the population dropsrapidly at the first 100 iterations. When the population size issmall (100), the algorithm finds poor quality chromosomesduring the later stage. On the other hand, when populationsize is set to 200, the algorithm generates good offspring

Fig. 4 Placement sequenceproblem—chromosome structurebased on cluster approach

Table 1 Comparison of experimentally recorded times for best andworst solutions for all motion types

Motion type LIN SPTP SLIN PTP

Best cycle time (ms) 107,796 94,284 93,264 56,340

Worst cycle time (ms) 119,604 96,936 95,460 60,108

Difference 11,808 2652 2196 3768

Int J Adv Manuf Technol (2018) 99:2523–2535 2529

quickly from comparably highly fit parents. When the algo-rithm finds the best solution, the curve reaches the plateau,indicated by a horizontal stretch in a learning curve, as the bestsolution is already the optimal one.

5.2 Optimal motion type

Using the solutions created with the genetic algorithm, a timedifference between the best and worst has been found.Depending on the motion type (LIN, PTP, SLIN, SPTP), the

improvement has been found as between 2 and 11 s (Table 1).The motion type has a large effect on the speed of the pathfollowed. As can be seen from the results, the PTP motiontype is far quicker than the others; however, while slowest, theLINmotion type offers the best improvement between the bestand worst. This is however negated by the quicker motiontypes. There are drawbacks to the faster motion types; at fasterspeeds (particularly PTPmotion), vibrations occur in the robotstructure. This could cause wear in the robot and reduce rep-etition accuracy within the process. The nature of the assem-bly process requires high levels of control and accuracy, and it

Fig. 5 Repeatability analysis:fitness vs generations plot (top);probability vs fitness histogram(bottom)

Fig. 6 Parameter sensitivity: effect of crossover probability on algorithmperformance

Fig. 7 Parameter sensitivity: effect of mutation probability on algorithmperformance

2530 Int J Adv Manuf Technol (2018) 99:2523–2535

would be potentially damaging to the robot to run at full speedwith PTP motion. Because of this, the PTP motion type wasretested at lower speeds to reduce vibration, and, at 50% ve-locity, the time was 79,956 ms, which is still quicker than thenext best motion type (SLIN). The SPTP motion type per-formed worse than expected, and it is theorised that this isdue to the increasingly complex calculations performed bythe robot and a notable “thinking” time was observed betweenmovements.

5.3 Orientation of work piece

During the experiments, it was found that some paths, whilelonger, showed a reduced assembly time. Upon further inves-tigation, it was found that cycle time is not exclusively

affected by path length. The orientation of the work piece tothe KUKA co-ordinate system affected the time it takes toperform a motion between two points. This was investigatedfor angles between 0° and 90°, as other angles could simply beinferred from these results. By rotating the angle at which thefeeder and PCBs are arrangedwithin the workspace, the X andY components of motion can be reduced. Between two points,the distance travelled is the same; however, it takes less time toperform (Fig. 9). In the figure, it can be seen that by rotatingthe entire process, the Y component of motion in particular isreduced and, for most components, this translates to a shorterprojection and faster movement time. Variation in assemblytime was seen for angles between 0° and 90° rotations. Anglescloser to 45° showed improvement in time which is due to thisbeing the shortest projection (Table 2) (Fig. 10). These resultswere found by assigning a midpoint around which the wholelayout was rotated. It can be seen that the numerical results ofprojected distance against angle results closely match the ex-perimentally tested time against angle results. This shows thatby shortening the projected distance, the time increases, whichconfirms the beneficial effect of rotating the assembly processby an angle of α close to 45°.

Further to this, the distance (d) from the KUKA base atwhich the assembly process was performed was investigatedas this influenced the assembly cycle time. The entire assem-bly operation was translated to various locations within theworkspace. The closer to the base the operation was per-formed, the faster the assembly time (Table 3). This is likelydue to a reduction of moment within the robot arm allowingfor faster acceleration and deceleration and a higher travelspeed. Therefore, the optimised operation should be per-formed as close to the base as possible in the d direction,

Fig. 8 Parameter sensitivity: effect of population size on solution quality

Fig. 9 Components of motion between feeders and component placement points

Int J Adv Manuf Technol (2018) 99:2523–2535 2531

without reaching the axis limits (Fig. 2). The height (h) atwhich the assembly is performed showed slight variation butnot enough to conclude that anything other than the availablework table should be used.

5.4 Objective function for feeder allocationand location of components

During the experiments, it was found that some componentswere less affected by the feeder assignment. This was due toan unanticipated but logical phenomena occurring. It wasfound that the time it takes for a diagonal line to be tracedby the robot is dictated by the longest X or Y component of themotion, as both happen simultaneously. Therefore, changingthe objective function tominimise projection rather than linear

distance improves the solution and speeds up the process byshortening the dictating X or Y component. This led to thediscovery that certain areas of the PCB are not affected byfeeder placement, as the distance from the feeder is greaterthan the gap between feeders 1 and 12. This is demonstratedby timing the end effector as it travels to one such point fromevery feeder where it can be observed and the time is identical.This led to a section of the PCB that is not important in feederplacement and theoretically could be prioritised below thecomponents closer to the feeders (Fig. 11). The shaded zoneshows the components that can have any feeder allocation andstill take the same time for the end effector to move to them.This is demonstrated by comparing the time in which the endeffector can move between a labelled component and eachfeeder (F1–12) (Table 4). The motion times for componentJK5 (Fig. 11) show variation; therefore, it is affected by feederallocation and is outside of the shaded zone. Component JK8shows small variation and is on the edge of the zone and canbe assigned any feeder apart from 12. The motion times forcomponent JK9 show no variation and the component is in-side the zone; therefore, any feeder could be assigned to thispart.

6 Conclusions

This research investigated the application of an articulated 6-axis robot to assemble a multi-component PCB for an elec-tronic product. The approach considers the need for a highlyflexible automation system for medium- and low-volume

Table 2 Experimentallyrecorded angle ofrotation and cycle time

Angle (°) Time (ms)

0 92,448

10 92,520

20 92,004

30 90,936

40 89,724

50 89,280

60 89,988

70 91,404

80 92,436

90 93,072

Fig. 10 Plot comparingexperimental and numericalresults of rotation angle effect

2532 Int J Adv Manuf Technol (2018) 99:2523–2535

manufacture. This non-traditional assembly robot for electron-ic assembly can be used to bridge the gap between time-consuming and costly manual builds and high-volume pro-duction lines using dedicated high investment automationequipment. To increase the potential of using a highly flexiblerobot with six degrees of freedom, a heuristic methodologyusing a genetic algorithm for planning the assembly operationwas used. The investigation utilised a KUKA KR16 roboticassembly cell to validate the optimised programs and usedcondition monitoring to identify which of the path planningprograms yielded the best results. The method was success-fully implemented for determining the optimal location offeeders and parts in relation to the robot and the best sequence

of robot movements in order to maximise production through-put. The main findings are as follows:

& The genetic algorithm showed an 80% probability of con-vergence toward the optimal solution and identified solu-tions to the combinatorial problems in this investigation.The feeder slot assignment problem was solved to find thebest allocation of the various component types based ontheir position on the circuit board. The order sequence forcomponent placement has also been optimised. Both so-lutions result in a significant reduction the total distancetravelled by the KUKA robot leading to an important im-provement to the assembly cycle.

& When considering the relationship between the KUKArobot global position and the possible angles at whichthe feeder and PCBs are arranged within the workspace,the x- and y-axis components of motion can be reduced.By arranging the PCB and feeders at a 45° angle to therobot primary axis, the assembly operation can be per-formed with a reduction in cycle time. Also by movingthe work piece closer to the robot base, the cycle time isfurther reduced. This is due to a reduction in the momentrequired to move the arm and improved acceleration con-trol within the robot system.

Table 3 Experimentally recorded translation and cycle time results

Direction Translation (mm) Time (ms)

D − 300 87,540

D 0 92,448

D 250 98,844

H 100 92,616

H − 200 92,952

Fig. 11 Zone on PCB unaffected by feeder allocation and location of components JK5, JK8 and JK9

Int J Adv Manuf Technol (2018) 99:2523–2535 2533

& It was found that some components on the PCB wereunaffected by the feeder positioning assignment. This isdue to them being in a position where the projected dis-tance of the robot primary world axis is not reduced bychanging the feeder position. This led to an area of thePCB that did not need to be prioritised in the feeder as-signment problem. However, by changing the objectivefunction to accommodate this finding, the total cycle timewas not reduced. For the number of components tested, achange was not observed, but on a larger scale, changingthe objective function in such a manner could provideimprovements to cycle time.

& The experimental results were tested using the KUKArobot to validate the identified best solution. Using condi-tionmonitoring, the four different robot motion types werecompared. Point-to-point motion provided the fastest as-sembly time; however, this did increase the vibration with-in the kinematic structure when compared to other pathmotions. Therefore, to avoid wear within the robot andunwanted vibrations, a reduction in operation velocitywould be necessary when using this motion type.

& Flexible manufacturing is important in modern indus-try, and current 6-axis robot systems lend themselvesto a wide range of automation activities. They exceedCartesian and SCARA robots in manipulation capabil-ity and lend themselves to automation activities be-yond basic pick and place. This research has shownthat genetic algorithm methodologies can be used tosolve a combination of problems and provides the ba-sis for increased use of 6-axis robots in PCB assembly.It also showed that the relationship between the robotmotion types and the position of the robot to the PCBassembly area needs to be considered when trying tosolve the optimum sequence order.

Funding information This work was supported by the AdvancedSustainableManufacturing Technologies (ASTUTE 2022) project, whichis partly funded from the EU’s European Regional Development Fundthrough the Welsh European Funding Office, in enabling the researchupon which this paper is based. Further information on ASTUTE canbe found at www.astutewales.com.

Open Access This article is distributed under the terms of the CreativeCommons At t r ibut ion 4 .0 In te rna t ional License (h t tp : / /

creativecommons.org/licenses/by/4.0/), which permits unrestricted use,distribution, and reproduction in any medium, provided you give appro-priate credit to the original author(s) and the source, provide a link to theCreative Commons license, and indicate if changes were made.

Publisher’s Note Springer Nature remains neutral with regard to jurisdic-tional claims in published maps and institutional affiliations.

References

1. Crama Y, Flippo OE, van de Klundert J, Spieksma F (1997) Theassembly of printed circuit boards: a case with multiple machinesand multiple board types. Eur J Oper Res 98:457–472

2. Crama Y, van de Klundert J, Spieksma F (2002) Production plan-ning problems in printed circuit board assembly. Discret Appl Math123:339–361

3. Moghaddam M, Nof SY (2016) Parallelism of pick-and-place op-erations by multi-gripper robotic arms. Robot Comput IntegrManuf 42:135–146

4. MilutinovićD, Dejan JR (2013) Redundancy in robot manipulatorsand multi-robot systems. Springer-Verlag, Berlin. https://doi.org/10.1007/978-3-642-33971-4

5. Erdos G, Kovacs A, Vancza J (2016) Optimized joint motion plan-ning for redundant industrial robots. CIRPAnn Manuf Technol 65:451–454

6. Rubio F, Llopis-Albert C, Valero a F, Suñera JL (2016) Industrialrobot efficient trajectory generation without collision through theevolution of the optimal trajectory. Robot Auton Syst 86:106–112

7. Fu B, Chen L, Zhou Y, Zheng D, Wei Z, Dai J, Pan H (2018) Animproved A* algorithm for the industrial robot path planning withhigh success rate and short length. Robot Auton Syst 106:26–37

8. Wang C, Ho L-S, J. Cannon D (1998) Heuristics for assemblysequencing and relative magazine assignment for robotic assembly.Comput Ind Eng 34(2):423–431

9. Mukund Nilakantan J, Huang GQ, Ponnambalam SG (2015) Aninvestigation on minimizing cycle time and total energy consump-tion in robotic assembly line systems. J Clean Prod 90:311–325

10. Ji P, Wan YF (2008) Minimizing cycle time for PCB assemblylines: an integer programming model and a branchand-bound ap-proach. Inf Manag Sci 19(2):237–243

11. Kodek DM, Krisper M (2007) Optimal algorithm for minimizingproduction cycle time of a printed circuit board assembly line. Int JProd Res 42(23):5031–5048

12. Holland JH (1975) Adaptation in natural and artificial systems: anintroductory analysis with applications to biology, control, and ar-tificial intelligence. University of Michigan Press, Ann Arbor

13. Wang W, Brunn P (2000) An effective genetic algorithm for jobshop scheduling. Proc Inst Mech Eng B J Eng Manuf 214(4):293–300

Table 4 Component to feedermotion time in milliseconds Component Time (ms)

F1 F2 F3 F4 F5 F6 F7 F8 F9 F10 F11 F12

JK5 516 516 504 504 504 504 504 504 504 516 516 528

JK8 528 528 528 528 528 528 528 528 528 528 528 540

JK9 528 528 528 528 528 528 528 528 528 528 528 528

2534 Int J Adv Manuf Technol (2018) 99:2523–2535

14. Lu C, Wong YS, Fuh JYH (2006) An enhanced assembly planningapproach using a multi-objective genetic algorithm. Proc Inst MechEng B J Eng Manuf 220(2):255–272

15. Yildiz AR, Ozturk F (2006) Hybrid enhanced genetic algorithm toselect optimal machining parameters in turning operation. Proc InstMech Eng B J Eng Manuf 220(12):2041–2053

16. Liao YG (2003) A genetic algorithm-based fixture locating posi-tions and clamping schemes optimization. Proc Inst Mech Eng B JEng Manuf 217(8):1075–1083

17. Sing PK, Jain SC, Jain PK (2005) Comparative study of geneticalgorithm and simulated annealing for optimal tolerance designformulated with discrete and continuous variables. Proc Inst MechEng B J Eng Manuf 219(10):735–758

18. Ball MO, Magazine MJ (1988) Sequencing of insertions in printedcircuit board assembly. Oper Res 36(2):192–201

19. Ahmadi RH, Mamer JW (1999) Routing heuristics for automatedpick and place machines. Eur J Oper Res 117:533–552

20. Klomp C, van de Klundert J, Spieksma F, Voogt S (2000) Thefeeder rack assignment problem in PCB assembly: a case study.Int J Prod Econ 64:399–407

21. Li S, Hu C, Tian F (2008) Enhancing optimal feeder assignment ofthe multi-head surface mounting machine using genetic algorithms.Appl Soft Comput 8:522–529

22. Grunow M, Günther HO, Schleusener M, Yilmaz IO (2004)Operations planning for collect-and-place machines in PCB assem-bly. Comput Ind Eng 47:409–429

23. Sohn J, Park S (1996) Efficient operation of a surface mount-ing machine with a multihead turret. Int J Prod Res 34(4):1131–1143

24. Broad K, Mason A, Ronnqvist M, Frater M (1996) Optimal roboticcomponent placement. J Oper Res Soc 47:1343–1354

25. Deo S, Javadpour R, Knapp GM (2002) Multiple setup PCB as-sembly planning using genetic algorithms. Comput Ind Eng 42:1–16

26. Ho W, Ji P PCB assembly line assignment: a genetic algorithmapproach. J Manuf Technol Manag 16(6):682–692

27. Ellis KP, Vittes FJ, Kobza JE (2001) Optimizing the performance ofa surface mount placement machine. IEEE Trans Electron PackagManuf 24(3):160–170

28. Magyar G, Johnsson M, Nevalainen O (1999) On solving singlemachine optimization problems in electronics assembly. J ElectronManuf 9(4):249–267

29. Suna D, Lee T, Kim KH (2005) Component allocation and feederarrangement for a dual-gantry multi-head surface mounting place-ment tool. Int J Prod Econ 95:245–264

30. KulakO,Yilmaz IO, Günther HO (2007) PCB assembly schedulingfor collect-and-place machines using genetic algorithms. Int J ProdRes 45(17):3949–3969

31. Garcia-Naijera A, Brizuela CA (2005) PCB assembly: an efficientgenetic algorithm for slot assignment and component pick and placesequence problems. 2005 IEEE Congress on Evol Comput 2:1485–1492

32. Hong DS, Cho HS (1999) A genetic-algorithm-based approach tothe generation of robotic assembly sequences. Control Eng Pract 7:151–159

33. Taha Z, Tahriri F (2010) Optimizing the robot traveling time(ORTT) in robot assembly line balancing problem (RALBP), The11th Asia Pacific Industrial Engineering and Management SystemsConference

Int J Adv Manuf Technol (2018) 99:2523–2535 2535