The rotary left ventricular assist device (LVAD) has been an effective option for end-stage heart failure. However, while clinically using the LVAD, patients are often at significant risk for ventricular collapse, called suction, mainly due to higher LVAD speeds required for adequate cardiac output. Some proposed suction detection algorithms required the external implantation of sensors, which were not reliable in long-term use due to baseline drift and short lifespan. Therefore, this study presents a new suction detection system only using the LVAD intrinsic blood pump parameter (pump speed) without using any external sensor. Three feature indices are derived from the pump speed and considered as the inputs to four different classifiers to classify the pumping states as no suction or suction. The in-silico results using a combined human circulatory system and LVAD model show that the proposed method can detect ventricular suction effectively, demonstrating that it has high classification accuracy, stability, and robustness. The proposed suction detection system could be an important part in the LVAD for detecting and avoiding suction, while at the same time making the LVAD meet the cardiac output demand for the patients. It could also provide theoretical basis and technology support for designing and optimizing the control system of the LVAD.