In the last two decades, soft sensors proved themselves as a valuable alternative to the physical sensor for gathering critical process information. A self-sensing technique for the magnetic bearing is considered as a soft sensor since the object position is estimated from the current signal of the electromagnet. Self-sensing techniques developed so far are the model-driven soft sensors. This paper presents a data-driven self-sensing technique to compensate for the nonlinear characteristic of the electromagnet. First, model-driven self-sensing techniques and their problems are reviewed. Then, data-driven self-sensing technique using recurrent neural network (RNN) is proposed to compensate for the nonlinear characteristics. Both the position control and self-sensing with the RNN are implemented in a single digital signal processor. The effectiveness of the proposed method is experimentally verified by comparison with the current slope method. Both estimation errors during initial levitation and jitter after levitation are reduced by 90% and 36%, respectively. Estimation error with 2 Hz sine wave is improved by 65.9%, while jitter during self-sensing levitation is cut down to 26.8%.

在过去的二十年中，软传感器证明了自己是收集关键过程信息的物理传感器的宝贵替代品。由于从电磁体的电流信号估计出物体的位置，因此将磁性轴承的自感应技术视为一种软传感器。迄今为止开发的自感应技术是模型驱动的软传感器。本文提出了一种数据驱动的自感应技术来补偿电磁体的非线性特性。首先，回顾了模型驱动的自感应技术及其存在的问题。然后，提出了一种基于递归神经网络（RNN）的数据驱动自感应技术来补偿非线性特性。RNN的位置控制和自感都在单个数字信号处理器中实现。通过与当前的斜率方法进行比较，实验验证了该方法的有效性。初始悬浮过程中的估计误差和悬浮后的抖动都分别减少了90％和36％。2 Hz正弦波的估计误差提高了65.9％，而自感悬浮过程中的抖动降低到26.8％。