----------------------------------------------------Business Index & ASAP------ AUTHOR(s): Brinkley, Paul A. TITLE(s): Northern Telecom achieves improved quality by combining DOE and SPC. (statistical process control)(includes bibliography) illustration graph Summary: Digital switching system manufacturer Northern Telecom Ltd uses wave soldering techniques that were designed with an approach that incorporates two types of process optimization. The first optimization technique seeks to reduce process variability by breaking process design into a large number of highly specialized experiments. The second technique is known as response surface methodology, and uses the smallest number of design experiments possible to model processes. Northern Telecom blames the majority of manufacturing defects on imperfect wave soldering, a process that has many parameters, including the speed of the product conveyor, the heat of both board and solder, a resinous flux treatment application, and the shape and action of the solder wave. The company used statistical methods to measure the effects on production of changes in each of these factors, and the interactions of each parameter. The statistics that resulted from experiments on each parameter in various situations were analyzed using generalized linear modeling techniques. Industrial Engineering p63(3) May 1993 v25 n5 DESCRIPTORS: Switches Production Process Control Soldering Switching circuits_Design and construction FILM NUMBER: 71R 1100 There has been a large amount of discussion in industrial literature regarding the use of experimental design and quality engineering techniques to achieve optimal quality in manufacturing processes. This work is characterized by highly fractionated factorial experiments coupled with optimization techniques that seek to reduce process variability. Many Japanese quality engineers and statisticians, in particular Dr. Genichi Taguchi, have advocated such methods to improve product quality "off-line." Another less widely acknowledged method of process optimization, known as response surface methodology, is endorsed by many statisticians as a technique for modeling complex manufacturing processes with a minimal number of experiments. This method receives less attention in industrial literature due to the more complex analytical techniques required for process optimization. This article discusses Northern Telecom's approach to optimizing complex wave soldering processes using response surface methodology combined with useful concepts advocated by Taguchi. The role of real-time statistical process control in maintaining optimized quality levels is also presented. Northern Telecom is a leading manufacturer of digital switching systems for use in both public and private telecommunications networks. The Research Triangle Park (RTP) facility is responsible for the production of central office switching systems used by regional and international telecommunications companies as well as independent suppliers of communications services. The Operations Research Department at Northern Telecom seeks to develop modeling methodologies for improving manufacturing productivity. The work discussed in this report is the result of a joint effort between NT Operations Research and the Department of Statistics at North Carolina State University. This joint effort was supported through funding by the Northern Telecom University Interaction Program. Wave soldering Wave soldering typically generates a greater number of manufacturing defects than any other production process in a printed circuit board manufacturing operation. This fact is primarily due to the large number of units (solder joints) produced each time a single board passes through the wave solder process. The manufacturing defects range from shorts or bridges between leads to solder skips where no solder joint is created. While these defects are easy to repair, usually requiring quick manual touch-up by a qualified technician, they also are difficult to detect by visual inspection, and frequently are not discovered until the finished board undergoes functional testing. This increases the possibility that a defect will go undetected and that a defective product will be installed at a customer site. The wave soldering process is complex due to the large number of adjustable parameters that directly affect solder joint quality. During the soldering process, boards travel down a conveyor, during which time a resinous flux is applied to remove oxidation from leads, allowing proper solder adherence. Following application of flux, the boards are heated to activate the flux and to prepare the board for the high temperature solder, thus eliminating the effects of thermal shock. The board then passes over the molten solder wave, at which time all exposed leads are attached to the board circuitry. The finished board then exits the soldering processes, continuing on to the next stage of manufacturing. The adjustable control parameters for the wave soldering process include the speed of the conveyor, the speed at which the spray drum for the flux applicator rotates, the temperature settings for board preheating ovens, the temperature of the solder, and the settings that dictate the shape and action of the solder wave. Northern Telecom's dedication to a CFC-free manufacturing environment eliminated the post-wave flux cleaning process. The elimination of board cleaning required a reduction in the amount of flux applied to the board, increasing the sensitivity of solder quality to the adjustable machine settings. Obtaining high quality solder joints requires the determination of the optimal combination of control parameters with respect to the physical and practical limitations of the wave solder process. As one might expect, many of these parameters interact significantly. For example, the temperature that the board achieves prior to passing over the solder wave is a function of both the temperature settings of the preheat ovens as well as the speed of the conveyor. The conveyor also interacts with the flux applicator, as does the flux applicator with the preheat ovens. Other interactions also frequently prove significant, depending upon the architecture of the circuit board and its solderability characteristics. Another factor increasing the importance of machine parameters is Northern Talcum's environmentally conscious CFC-free manufacturing policy which began in 1991. The elimination of CFCs combined with the inherent complexity of the soldering process made determination of optimal machine settings necessary to maintain acceptable board quality levels. Process optimization To improve soldering processes at Northern Telecom RTP, efforts began to determine the combination of machine settings that provided the highest possible solder joint quality. The method used for achieving optimal quality involved the design of statistical experiments to measure the individual effects of factors (machine settings) as well as the significant interactions of those factors. The use of either central composite or three-level designs allowed both the linear and the quadratic effects of factors to be estimated. The size of the experiments performed ranged from 36 to 52 runs, depending upon the number of machine settings to be studied as well as the number of quadratic effects to be estimated. Every experimental run was repeated multiple times to allow the estimation of variance for each run. For most of the experiments performed, defect count data for each run served as the response to be modeled and optimized. Response surfaces representing defect count data were then constructed using generalized linear modeling techniques. Generalized linear models single model. They are particularly effective when modeling processes with non-constant variance, as they allow models to be constructed without transformation of the response data. Frequently, different models were constructed for different categories of defect data collected from a single experiment. For example, separate models representing the occurrence of solder shorts and solder skips would often be constructed from data collected from a particular experiment. This was necessary when multiple classes of defects occurred on the same circuit board. In any case, the models constructed represented the response variable, product quality, as a function of the factors and interactions found to be significant. Following construction of the response surfaces, constrained optimization was performed using GAMS-MINOS non-linear optimization software to determine the optimal factor settings to implement on the shop floor. GAMSMINOS was chosen due to its effectiveness in handling non-linearity in both the objective and the constraint equations. For the case when a single generalized linear model representing defects was constructed, the model was optimized subject to the constraint of the factor settings falling within a feasible operating range. When multiple generalized linear models were constructed, the model representing the critical defect (usually solder skips) was minimized, subject to the models of the other defects equaling the smallest value possible (ideally zero defects). The factor settings were also constrained to a feasible operating range. In each case, the primary goal of optimization was to minimize defect levels and variation for each solder process. Following optimization of the response surfaces, confirmation experiments were undertaken to verify the fit of the models, and to assist in the final choice of settings to implement on the solder process studied. Optimization frequently yielded numerous alternative optimal solutions, each of which provided low defect levels. Other considerations then entered into the process of selecting the "globally" optimal settings. Those considerations included throughput effects of a particular conveyor speed, sensitivity of the solutions to minor deviations in machine settings, and the general robust behavior of the process for each optimal solution tested. Example and results One of the first areas chosen for implementation of the methodology described above was a multi-layer circuit board manufacturing cell. This cell had experienced high levels of solder defects due to the mix of product passing through the wave solder process. Factors chosen for study included solder pot temperature, conveyor speed, the temperature settings of three preheat ovens, flux drum rotation speed, flux applicator air pressure, solder wave height (lambda wave setting), and solder wave vibration (omega wave setting). An experiment was performed on each of three different board types chosen to represent the majority of product manufactured in the cell. Generalized linear models were developed representing board solder quality as a function of the factors studied. Numerous second order interactions and quadratic factor effects were included in these models. Constrained optimization of these surfaces yielded several alternative optima. Confirmation experiments were then undertaken to determine which of these optima performed best overall for the three boards studied, and therefore for all of the boards produced in the cell. Implementation of the final optimal solution yielded dramatic reductions in defective levels for the process studied. Figure 1 illustrates the improvement in quality realized as a result of process optimization. Solder defect levels prior to optimization were over 200 percent higher than those measured following implementation of optimal machine settings. More importantly, the soldering process became stable as compared to its previous state, where large swings in process performance from week to week were commonplace. Upon determination of the optimal settings for the process studied, difficulty was encountered obtaining cooperation from wave solder machine operators in maintaining the optimal settings. This lack of cooperation was due to the nature of the wave operators, job requirements prior to process optimization. Although assisted by engineering staff, operators typically were held responsible for short term declines in process quality. To improve quality, operators would routinely adjust the solder machine settings until a more desirable quality level was achieved. This constant adjustment, while necessary for a non-optimized process, contributed greatly to the variation present in the quality of the solder joints produced. The challenge, therefore, was to shift the focus of the wave operator's job from constant adjustment of the soldering process to maintenance of the process at its optimal level of performance through aggressive monitoring of solder levels, flux quality, machine performance, etc. The SPC solution Prior to process optimization, attribute defect data were collected at a post-wave solder touch-up station. These data were then plotted on a run chart at the end of every shift. Weekly Pareto analysis of these defectivity data was performed by cell personnel, with Causal Teams consisting of manufacturing engineers and machine operators attempting to address the causes of leading defects. While the weekly Pareto analysis proved effective at addressing long-term issues affecting solder defectivity, the run chart provided no feedback to the operator regarding the short term performance of the soldering process. To alleviate this problem, a real-time system of attribute statistical process control was implemented. A u-chart format was selected as the method of recording data, with control limits based upon the post-optimization solder quality levels established. Points were plotted on the chart following inspection of every ten packs. Vigorous action plans were put into place, providing touch-up inspectors and wave operators with a specific framework for communication in the event that the soldering process assumed an out-of-control condition. The most important component of these action plans was the requirement that the soldering process be shut down until the cause of the condition was alleviated. Providing production personnel and cell supervisors with real-time SPC as a tool for process improvement resulted in further reductions in defect levels in the manufacturing cell. maintaining the optimal machine settings. Adjusting the solder machine settings resulted in an immediate increase in solder defects at the inspection station. Action plans then mandated that soldering be temporarily halted until the process returned to its optimal state. This indirect method of maintaining optimal machine settings also ensured that proper machine maintenance procedures were carried out. The results described in the case study provided have been replicated on several soldering processes at Northern Telecom's RTP manufacturing plant. Combining experimental design and response surface optimization with aggressive SPC is now a standard method of improving soldering quality at Northern Telecom. Implementation of this combined methodology has resulted in substantial savings due to reduced repair and improved throughput. The process of improving soldering quality at Northern Telecom is ongoing, however. When new soldering technologies, new products, or product upgrades are introduced into the plant, processes must he re-optimized. The investment of time and effort is well worth the return as the company seeks to maintain its position as a leader in world-class quality manufacturing. The author wishes to acknowledge the contributions of Peter Mesenbrink, Kevin Meyer, and Dr. J.C. Lu, Department of Statistics, North Carolina State University, and Richard McKenzie, quality engineer, NTI, to the development of the combined response surface modeling - nonlinear optimization methodology described in this report. Thanks also go to Dennis Ellis, NTI manufacturing engineer, for his assistance in developing an understanding of the wave solder process; and to Bob Masini, manufacturing cell manager; Javad Taheri, operations research manager; and Andy Artola, NTI-RTP plant manager, for their continued dedication to world class quality initiatives. For further reading Box, G.E.P., and Norman E. Draper, Empirical Model-Building and Response Surfaces, John Wiley & Sons, New York, 1987. Box. G.E.P., Hunter, William G., and J. Stuart Hunter, Statistics for Experimenters, John Wiley & Sons, New York, 1978. McCullagh, P and J.A. Nelder, Generalized Linear Models, Chapman & Hall, London, 1989. Montgomery, Douglas C., Introduction to Statistical Quality Control, John Wiley & Sons, New York, 1991. Nair, Vijayan N., Editor,"Taguchi's Parameter Design: A Panel Discussion," Technometrics, Volume 34, No. 2, 1992. Vining, G. Geoffrey and Raymond H. Myers, "Combining Taguchi and Response Surface Philosophies: A Dual Response Approach," Journal of Quality Technology, Vol. 22, No. 1. 1990. Paul A. Brinkley is an operations research analyst with Northern Telecom, Research Triangle Park, N.C., where he is involved in developing and implementing statistical methodologies for quality and productivity improvement. He has a B.S. and M.S. in industrial engineering, specializing in operations research from Texas A&M University. He is a member of IIE.