Abstract: Bolt loosening of bolted connection structures can easily lead to structural failure. How to monitor the loosening state of bolts of a structure is a hot spot of current research. This paper uses the correlation analysis of structural vibration responses under environmental excitation and the deep learning technology, and studies a neural network model that uses both inner product matrix (IPM) and deep convolutional autoencoder (CAE), named inner product matrix and convolutional autoencoder (IPM-CAE) based deep learning model. This paper verifies the feasibility and effectiveness of the method through an experimental study on the bolt loosening state monitoring of a bolt connected plate. Compared with the convolutional neural network (CNN), stack autoencoder (SAE) and capsule network (CapsNet) using IPM, the proposed IPM-CAE method shows better network training convergence speed and recognition accuracy.