Reconstruction-based contribution for process monitoring

This thesis presents several methods to diagnose faults that occur in a process. The diagnosis is based on plots of the contributions of the sensors to the statistical index that is monitored. The statistical indices to be monitored are the SPE, T2 and ' indices. The contributions are determined for single and multiple sensor faults. Also, contributions of sensor blocks are examined using multiblock PCA. The contributions are determined based on the reconstruction of the aforementioned statistical indices. Control limits are calculated for each of the reconstruction-based contributions to the indices. Furthermore, in the case of multiple sensor faults, a sequential reconstruction method is proposed. Three examples are shown where the reconstruction based contribution methods are applied. The first example shows a process with a single sensor fault; in the second example there is a process with a two sensor faults and the third example is a polyester film process where data are grouped in blocks. The results show the reconstruction-based methods to be more decisive for diagnosing the causes of faults; specifically the ' index is more accurate for the diagnosis of the faults than the SPE and T2 indices.