Fall detection based on embedded sensor is a practical and popular research direction in recent years. In terms of a specific application: fall detection methods based upon physics sensors such as [gyroscope and accelerator] have been exploited using traditional hand crafted features and feed them in machine learning models like Markov chain or just threshold based classification methods. In this paper, we build a complete system named TSFallDetect including data receiving device based on embedded sensor, mobile deep-learning model deploying platform, and a simple server, which will be used to gather models and data for future expansion. On the other hand, we exploit the sequential deep-learning methods to address this falling motion prediction problem based on data collected by inertial and film pressure sensors. We make a empirical study based on existing datasets and our datasets collected from our system separately, which shows that the deep-learning model has more potential advantage than other traditional methods, and we proposed a new deep-learning model based on the time series data to predict the fall, and it may be superior to other sequential models in this particular field.