Percussion-Assisted Laser Speckle Approach for Bolt Looseness Classification Using Deep Learning

Abstract The existing bolt looseness detection methods in structural health monitoring suffer from various challenges, including noise and lighting conditions. To overcome the existing limitations, a percussion-assisted speckle-learned analysis is proposed for efficient monitoring of bolt looseness. Initially, the percussion approach was adopted to excite the bolted joint, and the laser speckle-learned analysis was carried out to extract speckle images simultaneously. The extracted speckle images were investigated using the autocorrelation and difference of maps method to establish a relation between the variations in speckle images respective to the bolt torque conditions. An increase in bolt torque led to a 19.11% reduction in speckle size, as observed from autocorrelation analysis, and a noticeable decrease in high-intensity speckle spots based on the difference in maps analysis. This indicates that tighter bolts generate more uniform surface vibrations, resulting in finer and less varied speckle patterns, an effect that can be reliably used to distinguish between different torque levels. Subsequently, a custom-designed convolutional neural network distinguished the bolt torque conditions with 97.7% classification accuracy, indicating the effectiveness of speckle pattern images in classifying bolt torque conditions. Overall, the proposed percussion-assisted laser speckle-learned approach offers a promising, high-sensitivity solution for structural health monitoring, enabling accurate and efficient onsite bolt looseness detection.