<p>Transfer learning is to use the existing knowledge to learn new knowledge. Knowledge of the data can be transferred through domain adaptation (DA). However, the subspace obtained by the existing DA method is non-orthogonal, which increases the difficulty and error of data reconstruction. In this paper, an uncorrelated subdomain alignment and global information preservation (USAGIP) method is proposed to implement fault detection and identification of blast furnace. The proposed method is able to distinguish local and global features of the data and extract the orthogonal transfer features with minimal redundancy. This is achieved by (1) aligning subdomains of source and target domains to preserve local structures of both domains, (2) maximizing the covariance of the source and target domains to preserve global structures of both domains, and (3) adding uncorrelated constraint to obtain orthogonal projection matrix. Furthermore, the proposed USAGIP is able to transfer knowledge in homogeneous or heterogeneous domains through domain-specific projections. To verify the feature transfer ability of USAGIP, blast furnace fault detection and identification experiments are conducted on homogeneous and heterogeneous datasets, and the results of the experiments show that USAGIP outperforms other existing DA methods.</p>