Convolutional neural networks for the detection of damaged fasteners in engineering structures

Locating and classifying damaged fasteners, such as bolts, in large engineering structures is vital in many health monitoring applications. Whilst traditional signal processing methods are often used to identify the presence of such fasteners, accurately estimating their location remains an ongoing challenge. In recent years, image detection (or the location of objects within images) using deep learning algorithms, such as convolutional neural networks (CNNs), has seen substantial improvements. This is largely due to the abundant database of images provided by internet search engines, as well as significant advances in computing power. Moreover, advances in digital imaging technology mean that affordable computer vision systems are now more readily available than ever before.

In this paper, a CNN architecture is proposed for the task of detecting damaged bolts in engineering structures. The new architecture forms part of a regional convolutional neural network (R-CNN), which applies a bounding box regression algorithm for bolt location alongside a softmax classifier for damage classification. A dedicated training set is also developed, which combines internet search engine data with images of a specifically-designed bolt rig. The new images extend the current dataset with the purpose of developing a bolt detector that is invariant to camera angle and location, as well as environmental factors such as lighting and shadows.