Bolt Looseness Location Identification Using Band-limited Intrinsic Multiscale Entropy Analysis and Convolutional Neural Network

The multi-bolt joints has been widely used in many industries. A few bolts looseness may gradually lead to structure failures and catastrophic consequences if undetected. In this paper, a bolt looseness location identification method using band-limited intrinsic multiscale entropy analysis and convolutional neural network (CNN) is proposed. First, different from the traditional excitation signals, the chaotic ultrasonic signals are used as excitation signals to obtain high-frequency nonlinear responses of bolt joint. Then, the response signal received by the sensing piezoelectric patch is decomposed by variational mode decomposition to obtain band-limited intrinsic modal functions (BLIMFs). Each BLIMF is an amplitude-modulated-frequency-modulated signals, which carries the protentional bolt looseness information. The multiscale sample entropy values of each BLIMF are calculated to construct a feature matrix containing the looseness feature of each signal component in the multiscale. Finally, the looseness feature matrixes are transferred to CNN for training a classifier to identify which bolt is loose. To verify the proposed method, an experiment with piezoelectric active sensing is designed. The bolt looseness in different positions is controlled by loosening one or more M1 bolts which are mounted on an aluminum alloy plate. The experimental result shows that all loosened bolts at different locations are effectively identified, which verify the validity of the proposed method in this paper.