The identification of the cleaning mechanism is a necessary step to decide, which physical subprocesses are relevant for modeling the cleaning processes. In this work, a new approach is proposed employing machine learning based on neural networks to identify the prevailing cleaning mechanism. Existing grayscale image data from cleaning experiments with dried starch, ketchup and petroleum jelly is prepared as training data for the neural networks. First, an offline approach is proposed which determines the dominating cleaning mechanism along the whole cleaning process by supervised learning. The trained networks achieve accuracies up to 95% for unknown test data. Second, the offline approach is extended to an online technique, which aims for a time resolved determination of the cleaning mechanism. The online approach achieves over 80% accuracy on unseen test data. An advanced application test on a new soil with spatially and temporally varying cleaning mechanism shows good qualitative agreement of the predicted cleaning mechanism. This proves the present approach to be a useful tool for analysis of experimental cleaning data and understanding cleaning processes as a whole.