Sensor drift is a critical issue in the research of Electronic Nose (EN) for industrial cyber physical systems due to its uncertainty, which deteriorates the sensing performance and reduces the odor recognition accuracy. This paper aims to address the challenge via domain adaptation, and proposes a novel Domain Adaptive Subspace Transfer (DAST) model to connect the regular domain (source domain) and the drifted domain (target domain) in order that drift compensation based on domain consistency can be achieved. Specifically, a projection matrix is utilized to transfer both the regular and drifted domain samples to an intermediate shared subspace wherein each drifted sample can be well reconstructed via a spare superposition of the regular samples such that the samples of different domains can be adaptively interlaced. Meanwhile, the manifold regularizations with Laplacian graphs are introduced to enhance the locality affinity manifold and the discriminant structure of data in the shared subspace. The quantitative experimental results on benchmark gas sensor dataset show that the proposed method can yield promising performance and the average improvement in odor recognition accuracy is about 7.98% higher than existing sensor drift compensation methods.