IEEE TRANSACTIONS ON AUTOMATION SCIENCE AND ENGINEERING, VOL. 11, NO. 4, OCTOBER 2014 1155 A Bayesian Approach to Automated Optical Inspection for Solder Jet Ball Joint Defects in the Head Gimbal Assembly Process Chee Wai Mak, Nitin V. Afzulpurkar, Matthew N. Dailey, and Philip B. Saram Abstract—Automation or selective automation is adopted as a solution to most productivity problems in the hard disk drive (HDD) industry as the industry continues to grow at a 40% compounded annual growth rate. An automated production line for manufacturing the head gimbal assembly (HGA) has been developed as part of the automation solution. In the automated HGA production line, a solder jet ball (SJB) soldering station connects the suspension circuit to the slider body. We propose a Bayesian approach to automated optical inspection (AOI) of the SJB joint in the HGA process, implementing Tree Augmented Naïve Bayes Network (TAN-BN) plus check classifier in-situ using GeNIe/SMILE within the inspection software. The system is fur- ther enhanced with a result checker, achieving an overall accuracy of 91.52% with 660 production parts in a blind test. Note to Practitioners—This paper was motivated by the problem of inspecting for defective solder joints in linear, automated pro- duction line for hard disk drive parts. The size and placement of the part in the tool presented a challenge to capturing a full view of the object under inspection. Existing approaches manipulate parts of the image under different conditions. This paper suggests a method that associates the likelihood of a measured feature of the image to the quality of the solder joint produced. In this paper, we charac- terized the features mathematically and established a probabilistic relationship between the features and the quality of the solder joint. We then showed how the relationship can be used in real-time de- termination of the quality of a solder joint presented to the inspec- tion system. We showed that the system achieved reasonable accu- racy when applied to production. Index Terms—Automated optical inspection (AOI), Bayesian networks, Peter–Clark Bayesian network (PC-BN), solder-joint defect, solder-joint inspection, tree-augmented Naïve Bayesian network (TAN-BN). I. INTRODUCTION AND MOTIVATION T HE CONTINUOUS growth of digital content creation, con- sumption, and preservation is fueling demand for hard disk drives (HDDs). The amount of digital content created and consumed surpassed 1.8 zettabytes (1.8 1021 bytes) in 2011, according to Gantz and Reinsel in [1]. About 52% of digital content are stored in HDDs, according to Chenery in [2]. This translates to a need for a large number of high Manuscript received July 25, 2013; revised November 16, 2013; accepted January 11, 2014. Date of publication February 27, 2014; date of current ver- sion October 02, 2014. This paper was recommended for publication by Asso- ciate Editor S. Shah and Editor M. C. Zhou upon evaluation of the reviewers’ comments. C. W. Mak is with Western Digital (Thailand), School of Engineering and Technology, Asian Institute of Technology, Klong Luang, Pathumthani 12120, Thailand (e-mail: st113357@ait.ac.th; mak.cheewai@wdc.com). N. V. Afzulpurkar and M. N. Dailey are with the School of Engineering and Technology, Asian Institute of Technology, Bangkok, Pathumthani 12120, Thai- land (e-mail: nitin@ait.ac.th; mdailey@ait.ac.th). P. B. Saram is with Back-End Engineering, Western Digital Malaysia, Petailing Jaya 47300, Selangor, Malaysia (e-mail: Philip.Bernard@wdc.com). Digital Object Identifier 10.1109/TASE.2014.2305654 Fig. 1. Graphical representation of a typical hard disk drive manufacturing process. capacity HDDs as data storage devices required by the market. Re- searchers and analysts have put the growth of demand for HDDs at a 40% compound annual growth rate, a trend that resembles the well- known Moore’s Law. This HDD growth trend is known as Kryder’s Law [3]. In less than 60 years since the HDD was invented, the amount of data stored on a single-medium disk has increased more than 1 000 000-fold while mechanical structures have shrunk 10 000 times, using a 2.5-inch form-factor (width: 69.85 mm by length: 100 mm by height: 9.5 mm typical) HDD as comparison. The sizes of components have thus shrunk significantly, making the HDD a precision device to manufacture. As the HDD market grows, production output must increase to meet demand. The output problem is an exponential one—manufac- turing time increases along with HDD capacity, compounded by the increasing number of units demanded. Full or selective automation in the HDD manufacturing process solves this by improving productivity and yield, thus increasing output to complement manufacturing ex- pansion. First introduced in HDD assembly lines, automation is now a “must-have” in component subassembly manufacturing processes. HDD manufacturing involves several complicated subprocesses to manufacture components. The components are manufactured in their respective subprocesses. The parts are then assembled and tested to form the HDD. The primary interest of this paper resides in the head gimbal as- sembly (HGA) subprocess at the head stack assembly component end. Specifically, it relates to the solder jet ball (SJB) soldering station in the automated HGA production line to achieve an electrical connection be- tween the suspension and slider circuit. A typical HDD manufacturing process is shown in Fig. 1. A typical automated HGA production line is schematically represented in Fig. 2(a), while a typical SJB process flow is shown in Fig. 2(b). In our automated HGA line problem, the quality of the solder joint (output of SJB process) is determined in an offline manual inspection station. The parts are examined at magnification (through a mi- croscope) by operators. Fatigue, lighting condition, operator experi- ence/speed, and subjective interpretation of results are some of the fac- tors affecting the accuracy of judgment. The offline inspection process also “breaks” the automated flow and negates the benefit of constant units-per-hour (UPH) of an automated process. We seek to implement an automated system to inspect the quality of the solder joint in-situ the SJB process. 1545-5955 © 2014 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.