The invention relates to the technical field of bolt fault detection, in particular to a bolt looseness detection method. The method comprises the following steps: acquiring a structure vibration video, and extracting synchronous vibration displacement signals of all selected pixel points from the video by using a
phase method to construct an original
signal sample
library with labels; constructing a preprocessing module, combining the synchronization signals, and expanding training samples; constructing a recognition model, connecting a preprocessing module with a main classification network in series, carrying out
supervised training by using an expanded
data set, and carrying out combined optimization on parameters of a preprocessing
convolution kernel and the classification network, so that the model can learn a mapping relation between a vibration
signal and a bolt state; during actual detection, vibration videos of a structure in an unknown state are collected under the same excitation, a
synchronizing signal is extracted, the
synchronizing signal is input into the trained classification model for
state prediction after being expanded by the trained preprocessing module, and the loosening position and degree of the bolt are comprehensively judged. According to the invention, non-contact and high-precision bolt looseness detection is realized.