Cyber-physical systems are the response to the adaptability, scalability and accurate demands of the new era of manufacturing called Industry 4.0. They will become the core technology of control and monitoring in smart manufacturing processes. In this regard, the complexity of industrial systems implies a challenge for the implementation of monitoring and diagnosis schemes. Moreover, the challenges that is presented in technological aspects regarding connectivity, data management and computing are being resolved through different IT-OT (information technology and operational technology) convergence proposals. These solutions are making it possible to have large computing capacities and low response latency. However, regarding the logical part of information processing and analysis, this still requires additional studies to identify the options with a better complexity-performance trade-off. The emergence of techniques based on artificial intelligence, especially those based on deep-learning, has provided monitoring schemes with the capacity for characterization and recognition in front of complex electromechanical systems. However, most deep learning-based schemes suffer from critical lack of interpretability lying to low generalization capabilities and overfitted responses. This paper proposes a study of two of the main deep learning-based techniques applied to fault diagnosis in electromechanical systems. An analysis of the interpretability of the learning processes is carried out, and the approaches are evaluated under common performance metrics.