DAMAGE IDENTIFICATION OF BOLT CONNECTIONS BASED ON WAVELET TIME-FREQUENCY DIAGRAMS AND LIGHTWEIGHT CONVOLUTIONAL NEURAL NETWORKS

In view of the difficulty of monitoring the states of bolt connections of large-scale structures, proposed a method for bolt looseness recognition by sound signals, which takes the advantages of the wavelet time-frequency analysis and the powerful image classification ability of the lightweight convolution neural network MobileNetv2. Continuous wavelet transform was carried out for the preprocessed sound signals to obtain the wavelet time-frequency diagrams. The lightweight convolutional neural network MobileNetv2 was trained using the wavelet time-frequency diagrams as samples. The trained model was used to identify the sound signals generated by loosen bolts. An outdoor test of a steel truss model showed that the proposed method could accurately recognize the sound signals of loosen bolts at different positions, with different numbers and different degrees of looseness, and with various environmental noise signals. This novel method has high identification accuracy and good stability, and requires low calculation cost and storage space. It can be carried out easily by mobile devices and embedded devices, providing a new idea for online damage recognition of large and complex structures under environmental excitation.