Modular Digital Twin for Human Activity Simulation based on Finite-State Machines

Human Activity Recognition (HAR) is becoming a key component in contemporary settings like Industry 5.0 and advanced smart home systems. In this study, we propose the use of a time-triggered, probabilistic Extended Finite-State Machine (EFSM) to build a modular Digital Twin (DT) of the system made by a moving person and the corresponding environment - including the sensors for their monitoring - to realistically reproduce the daily activities of the person and the signals generated by the sensors. The use of an EFSM allows to model the details of user's behaviors and to easily address the trade-off between accuracy and complexity of the model. In particular, the probabilistic nature of the EFSM allows to introduce variability in the simulations while maintaining the model simple. Simulations performed using the DT generate accurate extended data that can be used to feed and train HAR algorithms, while the corresponding ground truth is used to label the data for the evaluation of the algorithms. The empirical analysis of the generated patterns shows that closely capture the behavior of the occupants in a simulated indoor environment. Moreover, a simple model based on a Long-Short Term Memory (LSTM) neural network was devised to show the usage of the synthetic dataset in the inference of a person's position based on motion sensor signals.