This paper addresses the challenge to design an effective method for managers to efficiently process hazardous states via recorded historical data by developing a stochastic state sequence model to predict discrete safety states – represent the hazardous level of a project or individual person over a period of time through a Real-Time Location System (RTLS) on construction sites. This involves a mathematical model for state prediction that is suitable for the big-data environment of modern complex construction projects. Firstly, an algorithm is constructed for extracting incidents from pre-analysis of the walk-paths of site workers based on RTLS. The algorithm builds three categories of hazardous region distribution – certain static, uncertain static and uncertain dynamic – and employs a frequency and duration filter to remove noise and misreads. Key regions are identified as either ‘hazardous’, ‘risky’, ‘admonitory’ or ‘safe’ depending on the extent of the hazard zone from the object’s boundary, and state recognition is established by measuring incidents occurring per day and classifies personal and project states into ‘normal’, ‘incident’, ‘near-miss’ and ‘accident’. A Discrete-Time Markov Chain (DTMC) mathematical model, focusing on the interrelationship between states, is developed to predict states on construction sites. Finally, a case study is provided to demonstrate how the system can assist in monitoring discrete states and which indicates it is feasible for the construction industry.