As a data collection and control terminal for the Internet of Things (IoT) in the power grid, the smart measurement terminal adopts functional software design and flexibly expands its functions through application program (APP) development. Intruding malicious software through data theft or tampering may cause information leakage and economic loss. This article combined the side channel data of smart terminals and deep learning to implement a non-invasive APP sensitive behavior monitoring method. It studied the collection method of multivariate time series data on system status and resource scheduling during APP runtime. Then a convolutional residual Encoder model was constructed to achieve classification of sensitive behaviors. Experiment results show that the proposed method is better than other baseline models. Within a monitoring time of 1 s, the recognition accuracy of individual behavior and overlapping behaviors is 90.05% and 83.85%, respectively.