Here is the abstract you requested from the wirebonding_2018 technical program page. This is the original abstract submitted by the author. Any changes to the technical content of the final manuscript published by IMAPS or the presentation that is given during the event is done by the author, not IMAPS.
|Automated Wire Bond Outlier Detection and Classification|
|Keywords: Automated Process Control, Outlier Classification, Wire Bond|
|We developed Advanced Process Diagnostics (APD) algorithm to automatically detect and classify outliers in wire bond production. APD is a feature in Kulicke and Soffa wedge bonders, whereas APD-algorithm is an analysis and diagnostics program within APD. APD helps bonder operators to detect process deviations and identify and locate outlier bonds. APD can command bonder to do actions, such as image capture, non-destructive pull test and halt, based on the detected outlier class and severity. APD works in fully automated manner, without a need for any operator actions, even at initial phase. Automation is enabled in APD-algorithm by subdividing bonds into groups that produce similar bonds. APD-algorithm uses process signals, such as deformation and ultrasonic current, to detect and classify outliers. Within each group, the APD-algorithm calculates and normalizes features that are fed into the classifiers. The output of the classifier is the estimated probability for the bond to fit into the outlier class. Current outlier classes are: No Wire, Short Tail, Contamination, Bond on Bond, Unstable Surface, High Deformation, Bouncing Deformation, Unstable Frequency and Generic Outlier. Each bond is classified in 1-10 milliseconds right after the bonding process when bond head moves to the next bond position. The timing varies depending on the length of the process traces but do not affect the bond cycle time. After each bond APD-algorithm sends the classification results to the APD for bonder action and data visualization. We tested APD-algorithm performance against test failure data and production data. The test data had roughly 1k bonds, whereas the production data sample size is over 10M bonds. The performance metrics, true and false failure detection rates for each failure class are reported for both test and production data.|
|Henri Seppänen , Process Engineer, PhD
Kulicke & Soffa Industries
Santa Ana, CA