Purpose This article reports on an ongoing research project, which is aimed at implementing advanced probabilistic models for real-time identification of hazardous events at construction sites. The model has intelligent capabilities for near real-time automated recognition of hazardous events during the execution phase. To achieve this, features of Bayesian Networks have been exploited. In addition, inputs to the model are assumed to be provided by a pervasive monitoring system deployed on the site. The need for this kind of intelligent tool is determined by the complexity inherent in construction sites, due to a variety of reasons, such as heterogeneity of the actors, the simultaneous nature of operations, harsh contextual conditions, and the only partially efficient current approach based on health and safety plans. Hence, this model is proposed as a support tool for health and safety coordinators for supervision of sites as they cannot guarantee a continuous physical presence. Method Given that there are no long-time series on past occurrences of hazardous events in all the potential contextual combinations presently available, the probabilistic models cannot be learned just through datasets. For that reason, the available data have been integrated with expert opinions. In particular, the conditional probabilities of the Bayesian networks are estimated by an elicitation process of subjective knowledge from the opinions of experts. The complexity of the phenomena under analysis are modelled as a tree structure with several levels (corresponding to the work-breakdown structure hierarchy), which itself is based on the top-down technique; it provides therefore a clear view of the global picture. The built-hierarchical tree allows the expert to weigh more easily causal relationships involved and also to define the qualitative structure of the net. Furthermore, the article describes and tests how conditional probabilities of the variables in the networks can be estimated, through gathering and interviewing groups of stakeholders and experts. Results & Discussion Our research has led to the definition of a probabilistic model using elicitation techniques for subjective knowledge. Furthermore, the development of such a model is part of a wider system relying on the implementation of a real-time monitoring network.

This article reports on the current state of an ongoing research project which is aimed at implementing intelligent models for hardly predictable hazard scenarios identification in construction sites.As past evidences showed that no programmatic action can deal with the unpredictable nature of many risk dynamics, we tried to survey on how the current approach for safety management in the construction industry could be improved. In our previous research the use of Bayesian networks elicited from subjective knowledge were preliminarily tested. Those networks might be meant as a reliable knowledge map about accident dynamics and they showed that a relevant ratio of occurrences fall in "hardly predictable hazards" class, which cannot be warded off by programmatic safety measures.This paper reports the second outcome of our research project, which focused on the development of first elementary fragments, regarding the occurrence of a possible "hardly predictable scenario". Instead of experts' contributions (who, over their carrier, seldom incurred in accidents), we used "legal cases" as an accurate source of information. They suggested which categories of "hidden hazard scenarios" are more likely to happen. We found that the most frequent hidden hazard scenarios are linked to operator's negligence and abnormal behavior, e.g. irregular removal of scaffolding's components, unprotected openings, improper use of PPE, etc. Every pattern determined by legal cases has been formalized by a fragment (i.e. elementary network) of the overall Bayesian network.Finally, all the elementary networks were integrated into a comprehensive intelligent tool for realtime hardly predictable hazards prevention. The final setup, asked for interfacing these intelligent models to a low-level sensor network and used to feed them with inputs about the current state of the context, is discussed too.

One of the critical aspects in health and safety is the control of fine particle emissions from demolition and construction activities. Such exposure is very often the cause of professional illnesses causing a relevant economic burden for welfare and insurance institutions, besides harming workers. Hence this paper performs a feasibility study of a realtime control system of fine particle concentration on construction sites. It was conceived as a Zigbee TM based wireless, pervasive and non-invasive system, which is easy to deploy over the site and relatively cheap. Dust sensors were interfaced with the system and calibrated in the laboratory. The prototype is described in detail and tested under controlled and real conditions, in order to determine its potential for application. The prototype was shown to be an excellent tool to support health and safety inspectors, to provide in real-time a broad map of dust concentration over the whole extension of the site, provided that calibration coefficients are worked out for the various types of dust which can be encountered on the site.

No abstract