Baskı Devre Kartlarında Lehim Hatası Denetimi

Abstract

In the assembly process of printed circuit boards (PCBs), most of the errors are caused by solder joints in surface mount devices (SMDs). We propose using beta-Variational Autoencoder (beta-VAE) that learns from normal solder joints and disentangles factor of variations in data for anomaly detection of solder joints. Latterly, skewed Jensen-Shannon divergence (JSD) is proposed to maximize entropy of the probability distribution over data to classify fine-grained solder joints. In this thesis, two different methods are proposed to carry out solder inspection under different conditions: anomaly detection with an unsupervised deep generative model and fine-grained image classification with entropy regularization. At first, in order to solve the optical inspection problem in unrestricted environments with no special lighting and without the existence of error-free reference boards, we propose a new beta-VAE architecture for anomaly detection that can work on both integrated circuit (IC) and non-IC components. We show that the proposed model learns disentangled representation of data, leading to more independent features and improved latent space representations. We compare the activation and gradient-based representations that are used to characterize anomalies and observe the effect of different beta parameters on accuracy and untwining the feature representations in beta-VAE. Finally, we show that anomalies on solder joints can be detected with high accuracy via a model trained directly on normal samples without designated hardware or feature engineering. The next proposed method for solder joint inspection is fine-grained image classification. Entropy-regularization based methods are compared in fine-grained classification of solder joints and it is shown that the proposed skewed JSD outperforms others on different model architectures. Besides, regularizing entropy with skewed JSD enables the model to focus on more class discriminative regions. While unsupervised deep generative models do not require defective solder joints that are rarely found, they can not achieve as high accuracy as the supervised fine-grained methods. Consequently, there is a trade-off between the proposed methods.