With the increasing demand for ore in the world, the complexity, size and price of mining equipment have increased considerably. As these machines have embedded monitoring technology, the use of such data to increase the reliability and availability of the equipment has become essential in order to reduce maintenance costs. The objective of this work is developing a model that supports the decision of stopping an equipment, based on its actual state, using pattern recognition by neural networks. The proposed model aims to identify the state of equipment failure or pre-failure based on the data stored in the equipment and on the records of failure, so as to assess the risk of failure of equipment and to decide whether it should be stopped or wait for a new programmed shutdown. This dissertation was developed in four parts: study of the main models currently used for maintenance; design and implementation of the model to address this problem, based on artificial neural networks; performance evaluation of the proposed model; and simulation of equipment downtime using the proposed model. In the study of the main models a research was made about the evolution of maintenance techniques, through models of corrective maintenance, preventive maintenance and, finally, reaching the maintenance model based on condition monitoring. For the last two types of maintenance, it is presented the main models used in addressing the problem, its benefits and shortcomings. The development of the model was segmented into three main stages: processing of databases, from the data obtained directly from the equipment to the base of record of equipment failure; variable selection, based on the calculation of the influence of each equipment sensor to determine its failure state, as well as the definition of the ideal range of group data, and definition of the topology of networks. In the stage of assessing the performance of the proposed model we used data from corrective failures more often of two specific components of off-highway trucks: engine and transmission. To compare the performance between the different models evaluated, two factors were more important: classification performance and the longest interval between the identification of a pre-failure pattern and the occurrence of the failure. The results of classification of pre-failure patterns were quite satisfactory for most case studies, with hit rates ranging between 85% and 96%. From the classification model given in the previous step, we moved on to simulate different failure scenarios, calculating the equipment downtime that would have been avoided if the interventions defined by the model had been implemented, thus analyzing the increased availability provided by the use of the proposed model.