Fault diagnosis has been a crucial task for industrial processes. Compared with the traditional contribution plot method, the reconstruction-based fault diagnosis has drawn special attention, especially for its superior performance in the fault data processing. However, the conventional principal component analysis method pays little attention to the fact that the fault data contains both the fault information and the normal operating information. So the fault subspace modelled by the conventional principal component analysis is not accuracy enough. In order to improve the accuracy of the fault subspace, a reconstructed fault subspace modelling method is proposed based on generalized principal component analysis. First, we analyze the correlation between the normal operation data and fault data to obtain the generalized principal components, and model the fault subspace using projection relationship. Then, the principal component analysis monitoring system is constructed, and the optimal combination of normal operating and fault subspaces are obtained by monitoring the reconstruction data to select the appropriate number of generalized principal components and fault directions. The method makes full use of the normal operating data and faulty operating data, thus the extracted fault subspace can reflect more fault information than the previous methods, which plays an important role to help improve the fault diagnosis accuracy. Finally, both Matlab numerical simulation and Tennessee Eastman(TE) industrial process verify the effectiveness of the proposed method.