Based on the idea that multidimensional data is better summarized as a shell rather than a cloud, we have developed a surveillance approach that can detect with high sensitivity behaviour changes in a monitored process and alert the operator. Our methodology uses the time series of a high number of monitored indicators which we cluster together dynamically as a function of operating conditions. These clusters represent groups of similar realizations used to characterize a multidimensional manifold that can be interpolated to assess each new realization of the process behaviour. We evaluated the methodology on the data from a hydroelectric turbine. The event of interest was the loss of the turbine propeller runner cone. The results are good and the approach is able to detect the abnormal behaviour months before the event happened. We are currently looking at larger scale deployment to benchmark the approach's performance.