Abstarct: One of the most important features a smart home should have is predicting and detecting unusual and abnormal actions. The rapid development of sensor technologies on the one hand, and the significant increase in the number of elderly or ill living, on the other hand, have made this necessary. The problem is that the multiplicity and variety of sensors and the high volume of information received from them make the dimensions of the problem closer to the big data boundaries and the need for data fusion makes it more difficult. For this reason, the processing of this data, especially the extraction of appropriate features, involves considerable computational (temporal and spatial) complexity. In this thesis, we present a method for detecting and predicting unusual and abnormal actions in a smart home using deep learning. The data of this home, according to the prevailing tradition of smart homes, is collected through non-intrusive wireless sensors such as PIR and Door sensors. The output data of these sensors is a set of discrete asynchronous events occurring on a continuous time basis. The intervals of these events are not the same, and the interval between the two events may indicate important behavioral characteristics. For this reason, we have used point process theory instead of the usual methods of time series analysis. This theory is a powerful mathematical frxamework that provides the appropriate statistical language for formulating and analyzing asynchronous events. Common models in the Classic theory of point processes often take specific and limiting assumptions about the way events are produced. For this reason, we have used a deep neural network to surround the behavior of a process and extract normative patterns. This network, as a nonlinear mapping, predicts the type and timing of the next event, with event logs. We extend this prediction by employing a local beam search to a set of probable events to determine the possible range of future actions. The range of actual events within that range will indicate that the sequence is abnormal. The results of experiments show that this method enables the detection of abnormal activities with acceptable accuracy.