Principal component analysis (PCA) is an effective approach to process monitoring, and substantial works in this field have been reported. However, the fault detection behavior and performance of PCA are still equivocal and frequently lead to incorrect understanding of the detection results. This issue is addressed in this paper from two directions simultaneously. First, the expectation formulas of T-2 and squared prediction error statistics are presented and their relations to the statistical parameters of process data are discussed. Based on these relationships, the process disturbances and faults can be differentiated, which makes further fault diagnosis more reliable. Second, detectability conditions of different faults both in the principal subspace and in residual subspace are given. A new conception of critical fault magnitude was introduced to provide a definite description about the fault detection performance of PCA. The acquired results were illustrated and verified by monitoring a simulated double-effective evaporator process.