The rapid development of sensor and network technology promotes the development and application of cyber physical system, while the intrusion detection technology of the traditional network system has matured. The cyber physical system (CPS) can be improved in combination with its own characteristics based on the traditional intrusion network technology. The geographic location of CPS is complex and the network transmission is unreliable, which lead to that the detection robustness is not high. Aiming at this problem, an intrusion detection algorithm based on sparse denoising auto-encoders (SDAE) is proposed. What's more, CPS requires models to be adaptive and generalized, so difference brain storm optimization (DBSO) based optimization of SDAE (DBSO-based optimization of SDAE, DBSO-SDAE) detection algorithm is formed by combining DBSO with improved auto-encoders. The algorithm can automatically extract the optimal feature representation of intrusion data and greatly enhance the adaptability of the model while further improving the robustness of the model. The simulation results show that the DBSO-SDAE model proposed in this paper has higher robustness, adaptability and better real-time detection than other models, which greatly satisfies the high demand of CPS for detection algorithms.