Application of Bayesian Belief Networks to assess hydrogen gas retention hazards and equipment reliability in nuclear chemical plants

Many nuclear waste reprocessing and storage plant processes result in the generation of hydrogen gas. Radiolysis of radioactive liquors and corrosion of metallic magnesium waste are the main mechanisms for generating hydrogen in such facilities. Corrosion products such as magnesium hydroxide sludge are also formed which require storage in transportable vessels. Demonstration of sufficient reliability of systems such as purge air and ventilation extract is therefore required to protect against releases of hydrogen. Factors affecting hydrogen ignition and removal in nuclear environments as well as the identification of appropriate hazard management strategies have been the key areas of research for decommissioning and reprocessing plants. However, a knowledge gap has been identified in terms of assessing the likelihood of hydrogen retention within the sludge and waste matrix resulting in a sudden release of the gas into a vessel ullage. Hydrogen gas retention and the potential for a sudden release are affected by numerous factors such as faults leading to adverse waste disturbance. As such an appropriate technique must be applied to analyse the uncertainty from this gas behaviour. Bayesian Belief Networks (BBN) is an emerging statistical technique which allows uncertainty and dependencies between multiple variables to be taken into account in a quantified risk assessment. A BBN analysis has been undertaken to determine the key factors that would lead to disturbance of the sludge waste and the subsequent sudden release of hydrogen into the ullage space of a process vessel. The results show that the key sensitivities are adverse disturbance of the vessel sludge waste caused by faults leading to uncontrolled movements and clashes of the vessel. The benefits of applying the BBN technique to assess reliability of the purge and ventilation extract systems against radiolytic hydrogen release have also been explored. The BBN model has shown to be particularly advantageous, as it has allowed input of probability distributions of the key variables, instead of single point values, thus providing an enhanced understanding of uncertainty. Furthermore, the BBN technique has allowed updating of the probability of a known variable given a particular condition of the other variables. This updating function has enabled the key sensitivities to be determined.