With the prevalence of mobile e-commerce, fraudulent transactions conducted by robots are becoming increasingly common in mobile payments, severely undermining market fairness and resulting in huge financial losses for the industry. Mobile payment applications are facing the tough problem of identifying robotic automation accurately and efficiently in massive transactions. Current research does not propose any effective methods or engineering implementations. In this paper a novel deep seeded-clustering model is proposed to detect fraudulent payment transactions from robotic automation based on user behavior patterns. First, the model filters payment transactions to obtain suspicious transactions and prepossesses the transactions to convert the user behavior into two-dimensional images, so that the complex and random user behaviors are represented in standardized and simplified formats. Second, a deep seeded-clustering method is proposed by integrating the seeded-clustering algorithm with the convolutional neural network, in which clustering assignments of the former act as pseudo-labels for optimization of the latter. Third, the model is trained in semi-supervised mode using a small number of labeled samples and a large number of unlabeled samples in two steps: pretraining using the former in supervised mode and then training using the latter in unsupervised mode. Based on the experimental results, the model can classify the major fraud patterns based on a small number of labeled samples with precision, recall rate, and F1 score over 95% on the business dataset, is robust to low-level noise in labeled samples, adapts well to unbalanced-sample scenarios, and shows good generalization capabilities.