The intent of an inspection program of offshore platforms is to make them safer. The hope is that problems will be caught during the inspection process and thereby avoid a major incident. The inspection process serves the purpose of risk reduction in two ways: 1) operators must manage platforms in a conscientious manner or face a platform shutdown, or civil or criminal penalties, 2) the inspection process gathers information about a platform's operational history that might be useful in predicting the likelihood of future problems.
The Minerals Management Service (MMS)1 has acquired a large amount of information regarding the inspection and operational history of production platforms. The agency wanted to analyze this information to see if a model could be constructed to predict which platforms are more likely to have an accident or spill. Much of the work involved in this paper was simply manipulating the data to make the construction of models possible. Therefore, the hypotheses that are checked are rather modest. Hypothesis 1 – There is a relationship between a platform's operational and physical characteristics and the likelihood of an accident or spill. Hypothesis 2 – A model can be constructed which will enable the ranking of platforms in terms of risk.
This paper is organized into four parts: 1) a discussion of risk, the data, and some assumptions that are relevant to how the models are constructed, 2) a discussion of neural network models and model predictions for 1993, 1994 and 1995, 3) a discussion of how statistical methods can be used to check individual risk factor significance and how a logistic regression model can be used to rank platforms, 4) a discussion of how to construct an expert model.
There are two sources of information that are used to construct the models: 1) platform data for 1986 to 1995 contained in databases maintained by the MMS, 2) three surveys of platform inspectors and personnel at the MMS Headquarters in Herndon, VA. The surveys were used to perform an initial screening of the over 2,600 potential risk factors. An abbreviated set of model inputs was chosen in accordance with the results of the surveys.
A neural network model is presented which shows that "incidents" (accidents and or spills) can be predicted based on a very limited set of inputs. The neural network model contains 3 independent variables: 1) the slot count of a platform2, 2) the cumulative number of INCs3 received in the prior 5 years, 3) the cumulative number of accidents that occurred in the prior 5 years. Three prediction years were chosen: 1993, 1994 and 1995.
The neural network model consistently predicts 50% of the accidents or spills that will occur in the prediction year in the top 15% of ranked platforms. Additionally, the relative importance of the 3 independent variables changes over time. This indicates that the neural network model needs periodic re-training to take into account the variation in independent variable importance over time.
Next, the development of expert and logistic models is discussed. An example of how an expert model might be developed is presented. Also, a logistic regression of INCs to predict accidents and platform age to predict accidents is presented which shows that age alone is not a good predictor of which platforms will have accidents.