As a common failure form of the bolted joint, bolt looseness can lead to relative slip between the connected surfaces. Under the action of shear external forces, the bolt experiences fatigue failure and fracture, which can affect the service performance and connection strength of the equipment, and even cause major safety accidents. At present, good results have been achieved in detecting the tightness of bolt connections through Structural Health Monitoring (SHM) technology. However, most of them are still carried out under laboratory conditions and cannot be applied in engineering. Therefore, this article proposes looseness recognition of bolts by knocking acoustic signals. This method converts the acoustic time series signal into a Mel-Frequency Cepstral Coefficient (MFCC) frequency spectrum and then it recognizes the bolt looseness status through the Mobilenetv3 deep neural network. The results show this method can identify loose bolts. Meanwhile, it has a strong engineering application background.