Advances in machine learning and the Internet of Things have made predictive maintenance an indispensable component of industrial systems. However, the effectiveness of these approaches is directly dependent on the efficiency of data preprocessing techniques. In particular, data imbalance, feature selection, and normalization methods are among the key factors that determine model performance. In this study, the classification of machine failures and four different failure modes is investigated using the AI4I 2020 dataset. The performance of shallow machine learning algorithms, including Decision Trees (DT), Support Vector Machines (SVM), and k-Nearest Neighbors (k-NN), is comprehensively analyzed. To address data imbalance, random sampling and the Synthetic Minority Over-sampling Technique (SMOTE) are applied, with SMOTE demonstrating superior performance. In addition, min-max and z-score normalization techniques are compared, and z-score normalization is found to enhance classification performance. Failure modes are further divided into binary classes to determine the most suitable feature set, after which five features are sequentially eliminated using the backward elimination method and their effects on classification performance are examined. The results indicate that the proposed approach significantly improves classification performance. Notably, the k-NN algorithm achieves the highest accuracy of 99.2% on the SMOTE-balanced dataset. This study provides an original contribution toward improving the reliability of machine learning-based predictive maintenance applications.