A Multiple Neural Network System to Classify Solder Joints on Integrated Circuits

The following paper introduces a diagnostic process to detect solder joint defects on Printed Circuit Boards assembled in Surface Mounting Technology. The diagnosis is accomplished by a Neural Network System which processes the images of the solder joints of the integrated circuits mounted on the board. The board images are acquired and then preprocessed to extract the regions of interest for the diagnosis which are the solder joints of the integrated circuits. Five different levels of solder quality in respect to the amount of solder paste have been defined. Two feature vectors have been extracted from each region of interest, the “geometric” feature vector and the “wavelet” feature vector. Both vectors feed the neural network system constituted by two Multi Layer Perceptron neural networks and a Linear Vector Quantization network for the classification. The experimental results are devoted to comparing the performances of a Multi Layer Perceptron network, of a Linear Vector Quantization network, and of the overall neural network system, considering both geometric and wavelet features. The results prove that the overall classifier is the best compromise in terms of recognition rate and time required for the diagnosis in respect to the single classifiers.