Risk analysis becomes very important especially with the increase of risk accidents in the industrial fields. In this context, we present in this paper a new approach based on belief functions theory for determining the safety integrity level of a safety instrumented system. This approach consists on collecting data from expert opinions by eliciting judgements using a qualitative method, dividing them in groups using the k-means algorithm and aggregating them by applying a hierarchical method. The output of the data collecting process will be integrated into a risk evaluation model in order to get the safety integrity level. As an evaluation method we proposed a new generalized risk graph named Evidential Risk Graph which is able to deal with imperfect data modeled with the belief functions theory.