The current-voltage characteristics (I-V curves) of photovoltaic (PV) modules contain a lot of information about their health. In the literature, only partial information from the I-V curves is used for diagnosis. In this study, a methodology is developed to make full use of I-V curves for PV fault diagnosis. In the pre-processing step, the I-V curve is first corrected and resampled. Then fault features are extracted based on the direct use of the resampled vector of current or the transformation by Gramian angular difference field or recurrence plot. Six machine learning techniques, i.e., artificial neural network, support vector machine, decision tree, random forest, k-nearest neighbors, and naive Bayesian classifier are evaluated for the classification of the eight conditions (healthy and seven faulty conditions) of PV array. Special effort is paid to find out the best performance (accuracy and processing time) when using different input features combined with each of the classifier. Besides, the robustness to environmental noise and measurement errors is also addressed. It is found out that the best classifier achieves 100% classification accuracy with both simulation and field data. The dimension reduction of features, the robustness of classifiers to disturbance, and the impact of transformation are also analyzed.