Wearable device based human activity recognition, as an important field of ubiquitous and mobile computing, is drawing more and more attention. Compared with simple human activity (SHA) recognition, complex human activity (CHA) recognition faces more challenges, e.g., various modalities of input and long sequential information. In this paper, we propose a deep learning model named DEBONAIR (Deep lEarning Based multimodal cOmplex humaN Activity Recognition) to address these problems, which is an end-to-end model extracting features systematically. We design specific sub-network architectures for different sensor data and merge the outputs of all sub-networks to extract fusion features. Then, a LSTM network is utilized to learn the sequential information of CHAs. We evaluate the model on two multimodal CHA datasets. The experiment results show that DEBONAIR is significantly better than the state-of-the-art CHA recognition models.