Enhanced Angle Estimation Using Optimized Artificial Neural Networks with Temporal Averaging in IMU-Based Motion Tracking

Accurate angle estimate is crucial for motion tracking systems, especially in biomedical applications like rehabilitation, prosthesis control, and wearable health monitoring. Traditional filters, such as the Kalman filter, frequently encounter difficulties with nonlinear noise and dynamic variations, hence constraining their resilience. This study presents feedforward artificial neural network (ANN) models as a highly accurate alternative by utilizing IMU sensor data from a gyroscope and accelerometer. The research contribution encompasses: (1) the creation of ANN architectures of diverse complexity, featuring an innovative (4×8) structure with temporal averaging for enhanced noise resilience; (2) a simulation-based assessment in comparison to the Kalman filter utilizing consistent performance metrics; and (3) an evaluation of execution-time viability for embedded applications. A dataset including 3,599 samples was acquired from an MPU6050 IMU and partitioned into 70% for training, 15% for validation, and 15% for testing. Model assessment was conducted utilizing mean absolute error (MAE) and root mean squared error (RMSE). The NN (4×8 + Averaging) model produced a minimum MAE of 0.2657 and an RMSE of 0.3691, indicating a 68% enhancement compared to the Kalman filter. Although compact models (2×4, 2×8) exhibited marginal improvements, deeper architectures demonstrated superior generalization and resilience, especially during dynamic motion phases. These results show that ANN-based estimators offer better accuracy and adaptability, making them a good choice for real-time biomedical uses. Future research will investigate hybrid ANN-Kalman designs and assess their performance across diverse motion types, including gait cycles and robotics.