Drift Modeling and Compensation for Force Sensor Signal Using a Dual-Stage Attention-Based Transformation-Gated LSTM Network

In minimally invasive surgical robot system, high-precision force perception will improve the efficiency of surgery. The sampling signal of multidimensional force sensor usually appears baseline drift and noise. In this article we propose a method that tries to solve this issue. The method combines wavelet decomposition and neural network to model and compensate the drift. This paper uses wavelet decomposition to construct features, and uses a dual-stage attention-based transformation-gated LSTM (DA-TGLSTM) network to fit the drift. Then this paper uses the signal acquired subtract the drift predicted by the network to correct the drift. The verification experiments based on the actual drift data show that this method can track the drift over better than several other basic methods on multiple indicator criteria.