Machine Learning Assisted Signal Analysis in Acoustic Microscopy for Non-Destructive Defect Identification

Abstract

Signal processing and data interpretation in scanning acoustic microscopy is often challenging and based on the subjective decisions of the operator, making the defect classification results prone to human error. The aim of this work was to combine unsupervised and supervised machine learning techniques for feature extraction and image segmentation that allows automated classification and predictive failure analysis on scanning acoustic microscopy (SAM) data. In the first part, conspicuous signal components of the time-domain echo signals and their weighting matrices are extracted using independent component analysis. The applicability was shown by the assisted separation of signal patterns to intact and defective bumps from a dataset of a CPU-device manufactured in flip-chip technology. The high success-rate was verified by physical cross-sectioning and high-resolution imaging. In the second part, the before mentioned signal separation was employed to generate a labeled dataset for training and finetuning of a classification model based on a one-dimensional convolutional neural network. The learning model was sensitive to critical features of the given task without human intervention for classification between intact bumps, defective bumps and background. This approach was evaluated on two individual test samples that contained multiple defects in the solder bumps and has been verified by physical inspection. The verification of the classification model reached an accuracy of more than 97% and was successfully applied to an unknown sample which demonstrates the high potential of machine learning concepts for further developments in assisted failure analysis.

Using convolutional neural network for intelligent SAM inspection of flip chips

“…In opposite to those machine learning algorithms, deep learning does not require a manual extraction of features, it rather uses the network to learn features from the data and implements an end-to-end model from the data to the result. Convolutional neural networks (CNNs) are essential tools for deep learning, which are especially suitable for images as inputs [13,14] but also used for other applications such as analyzing time-domain signals [15][16][17]. Lin et al [18] designed a CNN based defect inspector named LEDNet to classify between normal and defective chips (with line blemishes or scratch marks).…”