Real-time torque prediction for ultrasonic motors using an attention-based BiLSTM model and improved differential evolution algorithm

Abstract

Ultrasonic motors (USMs), characterized by their miniaturization, high precision, and low noise, are widely utilized in robotics, medical devices, and aerospace applications. However, existing torque control methods are heavily dependent on sensors, which not only increase system cost and complexity but also restrict the deployment of USMs in space-constrained environments, thereby undermining their miniaturization advantages. Furthermore, the complex nonlinear torque characteristics and significant temperature effects of USMs have made traditional torque prediction methods based on physical models inadequate to meet the high-precision requirements of practical applications. To address these challenges, a real-time torque prediction method based on a hybrid attention mechanism, Hodrick-Prescott (HP) decomposition, and bidirectional long short-term memory (BiLSTM) network is proposed in this study. HP decomposition is employed to effectively capture both long-term trends and short-term fluctuations in time series data. The hybrid attention mechanism further highlights key input variables by distributing weights across time steps and feature dimensions. Finally, an improved differential evolution algorithm is applied to optimize the attention weights, enhancing model performance and reducing manual tuning effort. The proposed method’s superiority is confirmed by experimental results, which demonstrate high prediction accuracy and rapid response under various operating conditions. These qualities make the method highly suitable for real-time, high-precision, and miniaturized applications such as small robotic joints driven by USMs and precise medical machines.