An Autoencoder with Convolutional Neural Network for Surface Defect
Detection on Cast Components
Abstract
There is unrealized potential in using automation to alleviate the
visual inspection associated with non-destructive testing in
manufacturing facilities. The identification of defects during the
production can help avoid substantial manufacturing errors by indicating
that preventative maintenance should be introduced. The use of an
autoencoder for this application reduces the need to generate datasets
for various defect types, instead only one training dataset would be
needed. To address this, this paper proposes a Convolution Neural
Network (CNN) autoencoder approach to detect surface defects on cast
components during the production. The proposed method categorizes the
data into damaged and undamaged components by clustering based on the
loss associated with the reconstructed image. The average F1-score and
accuracy from retraining the model 10 times was 89.14% and 88.52%
respectively. Although previous studies have obtained higher metrics,
they have focused their efforts on supervised training techniques where
as this research proposes an unsupervised training method with results
comparable to the previous studies.