Blink detection is crucial in various practical application scenarios, such as eye disease detection, human-computer interaction, fatigued driving prevention, etc. To address the serious effect on the extraction of blink signal from the accompanying interference induced of the human body′s micro-scale movement, we propose a blink detection system, mmBlinkSEN, which can overcome the effects of accompanying interference and recover the blink waveform effectively. Inspired by the fact that blink and accompanying interference are mixed in a non-linear manner, a self-supervised deep contrastive learning method with a non-linear independent component analysis framework is proposed to separate blink and accompanying interference. A separation network ES-Net1 is designed, which is based on temporal correlation. The network takes two positive and negative sample sequences with temporal correlation and temporal uncorrelation as input to the network. The internal feature extractor inside the ES-Net1 is utilized to recover the temporal structure of the blink and the accompanying interference signal. Thus, the separation of the non-linear mixed signal is achieved. This article implements the mmBlinkSEN prototype system based on TI′s AWR1642 millimeter wave radar platform and validates the effectiveness of mmBlinkSEN with 14,000 sets of data. Experimental results show that mmBlinkSEN detects blink frequency with up to 88% accuracy in the presence of accompanying human interference.