Authors:
Ana Cravidão Pereira
1
;
Marília Barandas
1
and
Hugo Gamboa
1
;
2
Affiliations:
1
Fraunhofer Portugal AICOS, Rua Alfredo Allen 455/461, 4200-135 Porto, Portugal
;
2
Laboratório de Instrumentação, Engenharia Biomédica e Física da Radiação (LIBPhys-UNL), Departamento de Física, Faculdade de Ciências e Tecnologia da Universidade Nova de Lisboa, Monte da Caparica, 2829-516 Caparica, Portugal
Keyword(s):
Pattern Learning, Anomaly Detection, Topic Modeling, Incremental Learning, Lone Workers.
Abstract:
Learning routines and identifying anomalous behaviors play a critical role in worker safety. Identifying deviations from normal patterns helps prevent accidents, ensuring enhanced safety in complex environments. Topic modeling is frequently used to discover hidden semantics patterns and is well-suited to the complexity of routines in human behavior. However, its utility in complex time-series analysis and as a baseline for anomaly detection has not been widely explored. This work proposes a novel solution to accurately model complex routines using topic modeling, enabling the identification of anomalies through a statistical approach. A dataset of human movement recordings was collected over up to seven consecutive months, capturing the routines of three lone workers, with each accumulating between 414 and 955 hours of recording time. This dataset served as the basis for a comprehensive analysis of the results, showing strong alignment between visually observed patterns in routines a
nd the outcomes of the proposed method. Additionally, detecting anomalies across models with varying training days confirms that online learning enhances the accuracy and adaptability of routine modeling. Topic modeling allows for in-depth learning of routines, capturing latent patterns undetectable to humans. This capability prevents abnormal events, leading to a proactive approach to predictive risk assessment.
(More)