Loosening detection method for hexagonal screws based on error compensation and deep learning

Screws are extensively utilised across various domains. However, the problem of screw loosening not only impacts the stability and safety of engineering structures but also poses the risk of severe accidents and losses. The hollow design of hexagonal screw heads limits the effectiveness of conventional methods for detecting screw loosening. Additionally, traditional visual detection methods are influenced by the surrounding environment, affecting detection accuracy. To tackle these challenges, this study introduces a novel approach based on deep learning for detecting loosened hexagonal screws. This method primarily relies on the you only look once version 8 (YOLOv8) algorithm to accurately detect the coordinates of the four key points of the hexagonal screw. By applying geometric imaging theory, equations are derived for calculating the loosening angle and length of hexagonal screws. By utilising these equations along with bolt parameters, the degree of loosening is determined. Furthermore, photos captured in low-light environments are enhanced using an improved automatic colour enhancement (ACE) algorithm, which saturates image colours to improve environmental adaptability. This enhancement facilitates better recognition of hexagonal screws even in dark environments. Four colour enhancement methods are also evaluated based on four criteria. Moreover, by employing a back-propagation (BP) neural network for error compensation, the proposed method brings predicted values closer to actual values. The experimental results demonstrate an angular identification error of less than 1° and a length identification error of less than 1 mm for loose hexagonal screws.