This work focuses on the development of a new approach based on deep learning to implement an efficient and flexible intrusion detection system using the behavioral approach and mainly intended for critical infrastructures and industrial control systems. Based on the assumption that modeling the normal network behavior of industrial control systems is feasible and reliable, because the operations performed in these systems are quite stationary and repetitive, Convolutional Neural Networks (CNN), a deep learning technique, are used on the NSL-KDD dataset, a reference dataset used for the implementation of intrusion detection systems. The performance of the approach is presented and compared to some previous works. The metrics used include the percentage of correct classification, accuracy and false positives show that the proposed approach improves the performance of previous systems.