DATA-DRIVEN MODELING METHOD FOR VIBRATION OF ROTATING MACHINERY UNDER COMPLEX LOADS AND VARIABLE CONDITIONS

This paper aims to address the modeling challenges of rotating machinery under complex loads and variable conditions by proposing a data-driven modeling method based on recurrent neural networks (RNN). This method seeks to establish a time-series mapping relationship between vibrations and the operating parameters of rotating machinery. Monitoring the vibration of rotating machinery is crucial for ensuring its operational reliability and safety. However, vibrations exhibit significant variations under different loads and conditions, necessitating modeling and analysis. Traditional modeling methods rely on understanding the effects of internal and external environmental factors on the stiffness and damping characteristics of rotor systems. Yet, these methods are limited under complex loads and variable conditions due to difficulties in precisely determining model parameters, especially for rotating machinery like aircraft engines. In response to this challenge, data-driven modeling methods, particularly deep learning technologies, have shown great potential. The model proposed in this paper leverages the capability of RNNs to process time-series data, adapts to the characteristics of variable conditions with newly designed loss functions, and introduces adaptive initial hidden states considering the frequent and multiple startups of aircraft engines during their service period, aiming to overcome the limitations of traditional physical models and existing data-driven methods. The results demonstrate that the proposed data-driven modeling method achieves good accuracy in both training and testing sets, validating its effectiveness and practicality. The data-driven model established by this method can be used for lateral comparison of vibration characteristics under complex and variable conditions, offering a new perspective and tool for the health monitoring and assessment of rotating machinery.