Negative selection algorithm for monitoring processes with large number of variables

Chemical processes are being more heavily instrumented leading to larger and larger number of variables to monitor. Developing models for process monitoring using this colossal amount of data in a high dimensional space of process variables is a difficult task. In the recent times, immune system inspired Negative Selection Algorithm (NSA) has been gaining much attention for fault detection. Generally, the entire set of process variables is provided as input without pre-selection and as the number of variables becomes large, the monitoring performance of NSA reduces drastically. In this paper we propose a metric based on Bhattacharyya distance to measure the extent of similarity of a particular fault operation w.r.t normal in the space of any subset of variables. Then using this similarity metric, a scheme is proposed to systematically identify a smaller set of key variables for a particular fault. We also demonstrate through the benchmark Tennessee Eastman challenge process that NSA performs significantly better in the fault specific space of key variables.