Standard partial least square (PLS) is a useful tool for process monitoring; however, it still encounters some problems for the diagnosis of key performance indicator (KPI) faults. One of its recent modifications, improved PLS (IPLS), decomposes the process measurements into KPI‐related and KPI‐unrelated parts according to the correlation matrix obtained from the standard PLS. The entire residual space of PLS is categorized as the IPLS's KPI‐unrelated part. However, the residual space still involves some information related to KPI, and hence IPLS's decomposition may be inappropriate. In this study, a new modified PLS is proposed, which also decomposes the residual space according to the KPI. The loadings of input data are first decomposed to obtain a projection model. Next, the input data are more appropriately decomposed into KPI‐related and KPI‐unrelated parts. Correspondingly, two statistic indices can be designed for fault diagnosis. A numerical example and the Tennessee‐Eastman (TE) benchmark process are utilized to demonstrate the effectiveness and advantages of the proposed approach.