Subspace learning by transfer domain reconstructing on an extreme learning machine for enhancing E-nose system performance

Electronic nose (E-nose) systems have a wide range of application scenarios. However, the output deviation of chemical sensors is inevitable, which greatly limits the development of E-noses. In this paper, a subspace learning method of the transfer domain reconstruction extreme learning machine (TDRELM) is proposed to compensate for the performance degradation of the E-nose system caused by sensor drift. It constructs a unified feature representation space based on an extreme learning machine under various constraints. In projection space, variance information after data projection and discriminant information between samples are considered, while position information between samples is retained. The distribution difference between the source domain and target domain after projection is further reduced through reconstruction. Finally, the l2,1-norm and Frobenius norm constraints are applied to the reconstruction coefficient matrix and projection matrix to build a robust subspace learning model. We use alternate iteration to solve the model and the quantum particle swarm optimization (QPSO) algorithm to find the optimal parameters. Finally, we perform a large number of experiments on three common E-nose datasets, and the results demonstrate the effectiveness of the proposed method.