Deep learning and probabilistic mixture model powered monitoring of abnormal vibration for dynamic system

Dynamic systems in the real world will occur abnormal vibration under self-state deterioration or complex external excitation, which potentially threats the stability and safety of systems. The sudden increase in vibration amplitude is a direct manifestation of abnormal vibration, but machines cannot intelligently recognize this change like humans, especially when it exists various modes of vibration for real dynamic systems. While some networks of deep learning excel in modeling time series data, they struggle in learning and modeling high frequency vibration signal. To address this limitation, this paper proposes an intelligent method of abnormal vibration detection and state monitoring for dynamic systems, which utilizes the bidirectional long and short-term memory (Bi-LSTM) network for time series prediction of vibrate amplitude feature, and then applies the Gaussian mixture model (GMM) to fit the probability density of the predicted results. Based on the distribution characteristics of GMM of predicted vibrate amplitude feature, abnormal vibrations can be effectively identified. Experimental results demonstrate that the proposed method can achieve accurate detection of abnormal vibration, providing valuable support for the structural health monitoring of dynamic systems.