BiLSTM-MLP based hysteresis modeling for soft pneumatic joint actuator

Abstract

Soft pneumatic joint actuators show great potential in applications such as medical machinery, wearable devices, and soft grippers due to their inherent compliance. However, due to the internal friction and periodic relaxation of the elastic materials, the output force of soft pneumatic joint actuators has complex hysteresis characteristics, which seriously affects their control accuracy. To address the asymmetry hysteresis of actuator output force, a neural network model combining a bidirectional long-short-term memory network (BiLSTM) and a multilayer perceptron (MLP) is proposed. The MLP serves as the output layer of the BiLSTM, which is utilized to capture the time dependency of the actuator output force. By applying further nonlinear transformations, the feature representation of the BiLSTM output is retrieved and merged, improving the prediction and generalization capabilities of the model. The experimental results show that the BiLSTM-MLP model has a maximum output force error of only 0.306 N, an average error of less than 0.08 N, and a goodness of fit more than 0.999. Compared with the MLP, LSTM, BiLSTM, and improved PI models, the BiLSTM-MLP model can have higher prediction accuracy, better characterize the hysteresis properties of soft joint actuators, and provides a promising approach for hysteresis modeling of various elastomer actuators.