Abstract
Motivated by the need to improve the performance of badminton players, various motion monitoring systems have been developed to assist coaches in badminton technique instruction. While traditional video or optical methods are limited to fixed scenarios and require tedious postprocessing, IMUs, with their compact size and lightweight design, offer a portable solution for badminton motion monitoring. This study develops a wearable sensing network consisting of two IMUs to capture wrist motions during badminton strokes. The measurements are fed into a one-dimensional convolutional neural network (1D-CNN) to classify six stroke actions and fifteen badminton movement trajectories. The network includes two convolutional layers with 8 and 16 filters (3 × 1 kernel size), followed by batch normalization and ReLU activation. It also features a max-pooling layer, a fully connected layer, and a Softmax classifier. The network was trained with cross-entropy loss and the Adam optimizer (learning rate = 0.005, batch size = 64 iterating over 200 epochs). Experiments were conducted with six national-level badminton athletes to collect motion data for network training and prediction and stratified 5-fold cross-validation was employed to divide the training and testing datasets, ensuring the stability of model evaluation and the reliability of the results. Feature differences between the patterns were intuitively displayed using heatmaps and t-SNE approaches. The results show that the proposed method achieves classification accuracies of 97.16 ± 0.6% for six stroke actions and 86.07 ± 1.1% for fifteen badminton trajectories, outperforming traditional machine learning methods such as KNN (94.94% / 78.71%), SVM (94.32% / 75.06%), and DTA (92.81% / 72.93%). By combining a simple sensor configuration (two sensors) with a lightweight 1D-CNN, this study achieves accurate badminton motion monitoring, which is expected to contribute to sports training and enhance athletes' skills.