Development of Elite Soccer Player Artificial Intelligence RPE Prediction Model using Wearable GPS Devices

Abstract

In professional soccer, the workload is a major cause of fatigue and injuries, so the managing workload has been important. In this study, we tried to achieve effective workload management by analyzing the correlation between external workload and internal workload. The objective of this study is to predict Rating of Perceived Exertion (RPE), one of the internal workload metrics, using Global Positioning System (GPS) data, an external workload metric. GPS was collected using wearable Electronic Performance and Tracking System (EPTS) devices, and RPE data was acquired by surveying professional soccer teams. A total of 4,987 GPS datasets and RPE datasets were obtained from 162 training sessions and 47 match sessions from 35 players of K-league professional soccer teams. The GPS data was converted into effective GPS metrics through preprocessing. Effective GPS metrics are defined as metrics extracted by adding the researcher’s insight into the raw data obtained through GPS devices. A prediction model RPE-NET was constructed using Convolutional Neural Network (CNN), a deep learning algorithm proficient in time-series data processing. In addition, RPE-NET is combined with a Regression Activation Map (RAM) that visualizes discriminative part of the feature used for prediction contributed significantly to the result. To construct RPE-NET, a Global Average Pooling (GAP) layer and a single node of fully connected layer were added at the end of the CNN model. The Adam optimizer and mean squared loss function were adopted for RPE-NET. The accuracy of the model was 0.90 and its F1 score was 0.89. The error metrics MAE was 0.90, RMSE was 1.28. The importance map for the metrics was drawn through RAM, and the decisive parts of the features used for RPE prediction were visualized. Through this study, the effective GPS metrics have higher prediction accuracy than the commonly used collective level metrics, and it was possible to explain when fatigue was accumulated with RAM.