Given the advent of the 5.5G era, the widespread application of Integrated Sensing and Communication (ISAC) in a variety of industries, including wireless communication, intelligent transportation, and smart homes, highlights the critical importance of secure signal transmission. Chaos communication, with its non-periodicity, noise-like nature and sensitivity to initial conditions, offers notable security and reliability benefits. However, traditional chaos systems face limitations, such as the direct transmission of reference signals that weaken security and the impact of channel noise and interference on reliability. Furthermore, existing deep learning-based systems often suffer from high bit error rates and limited generalization in complex scenarios. In this paper, we propose an advanced OFDM-DCSK chaos communication model incorporating dual TCN, self-attention, Transformer, and DRQN, namely DTSAT-DRQN. First, based on the classic deep learning OFDM-DCSK system, we present dual TCN to improve feature extraction and fully exploit the features of time series data. Second, we use the self-attention mechanism and the Transformer model to adaptively capture long-distance interdependence in time series. Finally, we use DRQN reinforcement learning in model training to improve the model’s decision-making abilities and help it acquire the best behavioral tactics for complicated circumstances. Extensive simulations across various channels validate the method, showing superior performance in inference accuracy, robustness, and generalization compared to existing algorithms.