Sensor drift is an important problem for biologically inspired Electronic Nose (E-Nose) in industrial cyber-physical systems and their related applications, as it will deteriorate the sensing performance and lower system accuracy. Motivated by the observation that the regular and drift sensing data are oriented to the same high-level decision-making task, but show different low-level data distributions due to domain mismatch caused by sensor drift, this paper seeks to solve the challenging problem by learning a middle-level domain-invariant subspace. To achieve this, a Domain Adaptive Subspace Transfer (DAST) model is developed to transfer the key knowledge of the regular source domain and the drift target domain into an intermediate shared domain for both domain consistency and drift compensation. To be more specific, a transformation matrix is learned to transform the samples from the two domains to the intermediate shared subspace wherein each drift target domain sample can be well reconstructed by a sparse combination of some valuable and informative regular source domain samples, such that knowledge from the two domains is adaptively matched. In addition, the Laplacian manifold regularizations are incorporated to maintain the local affinity manifold of the drift target domain data, and enhance the discriminative structure of the regular source domain data in the shared subspace. The quantitative experiment results on two benchmark gas sensor drift datasets show that the developed DAST model performs well compared to different representative sensor drift compensation methods.