Estimation of Normal Ground Reaction Forces in Multiple Treadmill Skiing Movements Using IMU Sensors With Optimized Locations

Normal ground reaction forces (NGRFs) are key biomechanical parameters that determine the speed, stability, technique, and injury probability of skiers. However, most NGRF measurement devices are unsuitable for skiing due to limitations such as inconvenient portability, susceptibility to damage, and interference with human movement. Recently, it has been suggested that estimating NGRFs from data measured by inertial measurement units (IMUs) is feasible. However, most studies have only estimated NGRFs for simple movements such as walking or running with low estimation accuracy, and there are few studies for skiing. In addition, this article requires a large number of IMUs and has not optimized the number and layout of IMUs. Based on these issues, this article develops a new gated recurrent unit (GRU)-convolutional neural network (CNN)-particle swarm optimization (PSO)- bootstrap aggregating (bagging)-network (GCPB-Net) that utilizes IMUs to estimate NGRFs in multiple skiing movements. To improve the accuracy of the estimation model, the PSO-based multimodel fusion method and bagging ensemble learning are applied. To avoid limiting skiers’ mobility and increasing sensor costs, a dynamic IMU location optimization method based on maximum relevance and minimum redundancy (Dynamic-MRMR) is presented. With this optimization method, the optimal numbers and locations of IMUs for NGRF estimation in different skiing movements are given. This study is the first time to use machine learning and IMU layout optimization methods to estimate NGRFs in skiing. Based on the ablation and comparison experimental results, the GCPB-Net outperforms the base learners and most existing models.