Cascading failure process in interdependent networks has always been an important field of network cascading analysis. Different from the previous studies, we take people's demand for minimizing travel costs into consideration in this article and propose a network dynamics model based on the cost constraint. On this basis, we pay attention to the characteristics of different layers in the interdependent network, taking the real-world traffic network as an example, we define different load propagation modes for different layers. Then, we carry out the simulation experiment of cascade failure in the artificial network. By changing the structure of the network and the parameters in the model, such as the capability value of the network side and the connectivity of the network, we were able to focus on the effects of traditional protection strategies during the simulation and get some interesting conclusions. It is generally believed that increasing the quantity of connections in the network or improving the quality of edges will enhance the network robustness effectively. However, our experimental results show that these methods may actually reduce network robustness in some cases. On one hand, we find that the resurrection of some special edges in the network is the main reason for the capacity paradox, as these edges will destroy the stable structure of the original network. On the other hand, neither improving the internal connectivity of a single-layer network nor enhancing the coupling strength between interdependent networks will effectively improve network robustness. This is because some critical edges may appear in the network while the number of edges increasing and a large amount of the network load will be attracted by them leading to the decrease of network robustness. These conclusions warn us that blindly investing resources in network construction cannot achieve the best protection effect. Only by scientifically designing network structures and reasonably allocating network resources can network robustness be effectively improved.