Recently, for bolt looseness detection, percussion-based methods have attracted more attention due to their advantages of eliminating contact sensors. The core issue of percussion-based methods is audio signal processing to characterize different bolt preloads, while current percussion-based methods all depend on machine learning–based techniques that require hand-crafted features and overlook bolt looseness at the incipient stage. Thus, in this article, the main contribution is that we propose a novel one-dimensional training interference capsule neural network (1D-TICapsNet) to process and classify percussion-induced sound signals, thus achieving bolt early looseness detection. First, compared to machine learning–based techniques, 1D-TICapsNet can fuse feature extraction and classification in one frame to achieve better performance. In addition, due to two tricks (i.e. training interference), including wider kernels in the first convolutional layer and the targeted dropout technique, our proposed 1D-TICapsNet outperforms several state-of-the-art deep learning techniques in terms of classification accuracy, computational costs, and the denoising capacity. We call these two tricks as “training interference” since they work during training procedure. Finally, we confirm the effectiveness and superiorities of 1D-TICapsNet via experiments. Considering the efficacy of 1D-TICapsNet, we can expect its real-world applications on bolt early looseness detection and other classification of one-dimensional signals.
Machine Learning Applied to Music/Audio Signal Processing
