Modern cars generate a significant amount of data while running. When recast into relevant information, such data can improve several aspects, such as increasing the safety and useful life, reducing production costs, fuel consumption, and emissions. In this paper, we analyze misfire, which happens when combustion does not happen correctly, impacting the overall functioning of the engine. After an exhaustive feature selection process related to the fault, we collected data directly from a vehicle’s ECU to train and evaluate machine learning algorithms to detect misfires without using synthetic data or public datasets. The results achieved up to 94.24% of F1-Score, 92.40% of precision, and 96.16% recall, which are better than related works. In addition, our approach acquires data in real-time and sends it to a server in the cloud that detects the failure with the vehicle running, applying computationally less expensive algorithms.