Failure diagnosis represents an important task for operational maintenance teams, focusing on the task of identifying the causes of equipment problems that can lead to deviations in expected behavior, as well as reducing expected efficiency. The application of detection and diagnostic techniques associated with predictive methods, commonly known as prognosis, enables a more accurate and adequate planning to deal with unexpected events that may put the system under study at risk. Through an early and detailed identification of possible causes and threats, maintenance teams can mobilize themselves in a more appropriate, planned and assertive manner, to deal with unwanted situations, before they actually occur, favoring greater system reliability and, consequently, avoiding unexpected service interruptions, reducing the possibility of material and human losses. Several techniques have been suggested in the literature to address questions about failure prognosis, with great emphasis on methods based on deep learning. Such methods are considered to have characteristics called "black box", as they do not provide the means to explain the results obtained, making it difficult to adopt informed and reliable decisions. Thus, this research proposes an unsupervised method for the diagnosis and prognosis of failures, based on deep learning techniques, which provide means to attest and contribute to the robustness of the results obtained, promoting greater confidence and assertiveness in the prediction and identification of possible problems. In the case study, real physical measurements of parameters of wheels and bearings of railroad cars carrying heavy loads are considered, captured from multiple track sensors coupled at specific points in a Brazilian railroad. Means are proposed that address restrictions in modeling with data and that attest to the reasonableness of the predictions made.