AI-Based DC Motor Angle Prediction Process for Endpoint Position Control of a Klann Linkage

초록

Walking robots are finding growing applications across diverse industries, yet their control architectures often depend on costly brushless DC (BLDC) motors and multi-degree-of-freedom actuation, limiting commercial viability. This study introduces an artificial-intelligence approach that predicts joint angles of a single-DOF Klann linkage using only DC-motor current signatures. Load-induced current variations are captured by a low-cost current sensor and converted into time-series data synchronized with link-angle measurements for supervised learning. Four time series prediction models (LSTM, Attention-based LSTM, CNN-LSTM, and Neural ODE) with different approaches are used for learning, and Circular Mean Squared Error (Circular MSE) and Circular Mean Absolute Percentage Error (Circular MAPE) were used as key evaluation indicators to reflect the periodicity of angle data. Among all candidates, the ALSTM yields the lowest validation Circular MSE (98.4458) and the lowest Circular MAPE(0.02042), delivering the most accurate angle estimates. Experimental results confirm that precise endpoint position control of the Klann linkage can be achieved with minimal sensing hardware when paired with tailored deep-learning models. We present promising possibilities for the development of efficient link-based driving devices and walking robot systems as a novel approach to reduce the cost and mechanical complexity of the system by replacing high-cost encoders only with current sensors.

키워드