Cascading commutation failures (CFs) pose severe risks in multi-infeed high voltage direct current (HVDC) systems. Different from the single or concurrent CF, not only the time-relevance of signals but also the spatio coupling and even control correlation of HVDCs will attribute to the cascading CFs. The conventional approaches to identify them tend to fall into a dilemma due to their complicated dynamics, wide-area coupling and vague threshold of judgement. In this paper, a deep-learning method based on the data-driven idea is proposed to recognize the cascading CFs. It analyzes the crucial factors leading to the cascading relationship of multiple HVDCs, while classifying them into time and space signals. To extract the inherent correlation between HVDCs as well as the time relevance in question, a spatio-temporal convolutional network (STCN) is formulated. The data generated in case of faults with diverse severity are applied to train STCN. Finally, the proposed framework and STCN method are validated by a customized IEEE 39 bus system and a practical power grid.