Classification of Test Pads from Clustered PCB images

A robotic flying probe tester can be programmed to move the robotic probe to access all possible locations of test pads in a printed circuit board (PCB), and to record all connection test results like open or short circuits between all possible pairs of test pads in the board. For this purpose, Tan and Kit performed a clustering-based image cluster analysis on the photo image data of printed circuit boards to recover all test pad locations on the board and reported successful results. Their clustered data has been open to the public since 2024. So in this paper, several classification techniques for human comprehension were applied to give the robotic flying probe tester the location of test pads. As the final results of clustering were reviewed and corrected by experts in the original paper, we created machine learning results of classification that are easy for humans to understand, so that it could be easier to review the machine learning results before giving them to the robotic flying probe tester as input. For the classification task, we focused on knowledge discovery methods that can give the coordinates of the grey or test pad to a robot and are readable by humans. Decision trees and rules have the advantage of being relatively easy to understand because the knowledge models are expressed in a single tree structure or a set of rules, so they are widely accepted in the fields where the interpretation of trained knowledge models is important. Three different decision trees and two kinds of rule sets were constructed - J48, Random tree, REP tree (Reduced Error Pruning tree) for the decision trees, and JRIP and PART (PARtial decision Tree) for the rule sets. The accuracy of all four generated knowledge models is 100% except that of the REP tree which is 99.9997%. The size of the generated decision trees was relatively very small compared to the size of the data, 723,552 records, and the generated rule set by JRIP has only two rules. Therefore, we can conclude that the decision trees and the sets of rules for determining the test pads in the PCB have produced very successful results in terms of comprehensibility and accuracy.