An end-to-end framework combining time–frequency expert knowledge and modified transformer networks for vibration signal classification

Vibration signal classification plays an important role in operation monitoring of mechanical structures. In this paper, a novel end-to-end framework is proposed for intelligent vibration signal classification including data preprocessing, time–frequency feature extraction, modified Transformer network as well as integral optimization. Two case studies in different engineering fields are conducted including the health monitoring of a bearing component through long-term running-to-failure experiments under constant loading conditions and the flight state identification of a novel self-sensing wing structure through a series of wind tunnel experiments under varying angles of attack and airspeeds. Multi-sensor fusion experiments are further conducted to enhance the classification accuracy. Results from both case studies demonstrate that the proposed method can not only extract distinguished high-level features but also be optimized jointly to achieve the best performance over convolutional neural network and recurrent neural network based methods in signal classification.