Real-Time Vibration Estimation and Compensation With Long Short-Term Memory Recurrent Neural Network

Abstract

Vehicles, as the moving platforms of various activities, have played important roles in modern society. However, the mechanical vibration due to various sources greatly degrades the performance of on-board devices that require high precision. To compensate the vibration, the technical challenges include: 1) the vibration possesses multiple time-varying dominant frequencies; 2) the broad bandwidth; 3) the phase difference between compensating movement and vibration; and 4) realizing real-time (RT) operation. In this article, we propose an AI-aided RT estimation and compensation method to address these challenges. The proposed method consists of two recursive least square-based filters to remove the gyroscope noise and drift, and a long short-term memory-based recursive neural network to remove the phase shift. Applied techniques are all implemented in RT. The method is validated by simulations and RT experiments using vibration data sampled from a real vehicle and achieves a 75% compensation rate, which outperforms existing methods.