Deep learning has proven to be effective for enhancing the accuracy and efficiency of attack detection through training with large sample sizes. However, when applied to cyber–physical systems (CPSs), it still encounters challenges such as scarcity of attack samples, the difficulty of selecting features for high-dimensional data, and weak model-generalization ability. In response, this paper proposes ResADM, a transfer-learning-based attack detection method for CPSs. Firstly, an intentional sampling method was employed to construct different sets of samples for each class, effectively balancing the distribution of CPS-attack samples. Secondly, a feature-selection method based on importance was designed to extract the meaningful features from attack behaviors. Finally, a transfer-learning network structure based on ResNet was constructed, and the training parameters of the source model were optimized to form the network-attack detection method. The experimental results demonstrated that ResADM effectively balanced the data classes and extracted 32-dimensional attack-behavior features. After pre-training on the UNSW-NB15 dataset, ResADM achieved a detection accuracy of up to 99.95% for attack behavior on the CICIDS2017 dataset, showcasing its strong practicality and feasibility.