Thermal signature for solder defect detection using a neural network approach

This paper describes a neural network approach to detecting solder defects on printed circuit boards when using thermal signatures. Solder defects such as open and insufficient solder was investigated. A multi-layer neural network with multiple inputs and a single output was utilized. A back-propagation algorithm was utilized within the network. Computer mouse printed circuit boards with known introduced solder defects and amounts of solder were used for experiments. Thermal images were acquired as the boards were powered up. A Visual Basic program was written to retrieve temperature data from an encoded image file format. Afterwards, MATLAB neural network routines were applied to analyze the data. The neural network was able to diagnose solder defects on two of five resistors with 91.1% accuracy, and on three of five resistors with 61.1% accuracy.