This study presents a comprehensive evaluation of predictive maintenance models for vehicle fleets, detailing a sequence of systematic steps to ensure model performance and address real-world variability. The process begins with database creation and data preprocessing, where relevant maintenance records are filtered, and datetime columns are converted to facilitate time-based calculations. Grouping and aggregation techniques are then applied to count occurrences of specific maintenance activities and identify common failure types. For model training, we define a neural network architecture comprising dense and dropout layers to mitigate overfitting, compile the model with suitable loss functions and optimizers, and train it using the prepared data. The trained model, along with the scaler and encoder, is saved for future use. To augment the dataset, synthetic data is generated using the Faker library and random distributions, with added noise to mimic real-world variability. Preprocessing steps are reapplied to this synthetic data to ensure consistency. By implementing this neural network, we achieved a sensitivity of 0.93 and an ROC of 0.71. Following these detailed steps, we develop a robust predictive maintenance model that effectively identifies failures and non-failures, ultimately enhancing the reliability and efficiency of vehicle fleet management.