Ballistocardiagram (BCG) is a non-contact and non-invasive technique to obtain physiological information with the potential to monitor Cardio Vascular Disease (CVD) at home. Accurate detection of J-peak is the key to get critical indicators from BCG signals. With the development of deep learning methods, many researches have applied convolution neural network (CNN) and recurrent neural network (RNN) based models in J-peak detection. However, these deep learning methods have limitations in inference speed and model complexity. To improve the computational efficiency and memory utilization, we propose a robust lightweight neural network model, called JwaveNet. Moreover, in the preprocessing stage, J-peaks are re-modeled by a new transformation method based on their physiological meaning, which has been proven to increase performance. In our experiment, BCG signals, including four different sleeping positions, were collected from 24 subjects with synchronous electrocardiogram (ECG) signals. The experiment results have shown that our lightweight model greatly reduces latency and model size compared to other baseline models with high detecting accuracy.