Loose bolt detection method based on wavelet packet decomposition with CNN

Abstract

Loose bolts in PV mounts can affect the smooth operation of the system. If the loose bolts are not detected in time, the PV modules may fall off and cause the mounts to collapse in serious cases. The ultrasonic wavelet packet decomposition (WPD) combined with convolutional neural network (CNN) is proposed as a detection method to address the problem that the ultrasonic signal collected when the bolt is loose indistinctively. First, the ultrasonic signal with bolt loosening information is decomposed into four layers of wavelet packets to extract the energy composition of the feature vectors of each sub-band signal; second, the CNN model is designed and the network is trained with the feature vectors as samples; finally, the discrimination of bolt loosening conditions is achieved. The feasibility of the method is verified through experiments, and the experimental results show that the accuracy of the model is improved from 95.64% to 99.92% compared with the original data training model, and from 86.96% to 99.92% compared with the support vector machine (SVM).