Modular multilevel converters are attractive for medium- and high-voltage applications because of their modularity and high-power quality. Reliability is a critical issue for this converter because each submodule has a capacitor, and condition monitoring of these capacitors is highly challenging. There are some conventional techniques for monitoring the condition of submodule capacitors. However, the implementation of these techniques requires additional hardware or inclusion of a reference submodule. Therefore, this study proposes a machine learning-based condition monitoring technique for submodule capacitors in modular multilevel converters. A comprehensive dataset of submodule voltages and arm currents has been collected to extract statistical features and wavelet decomposition techniques are employed to remove the noise in the signals. Furthermore, a comparative analysis has been performed to show the efficacy of the different classifier models. The performance of the classifiers is examined through different indexes, such as accuracy, precision, \(F_1\) score, recall, kappa coefficient, and area under the curve. The results show that the random forest is the most reliable and robust classifier for detecting significant levels of degradation of submodule capacitors in modular multilevel converters.