Abstract: In order to fully mine the temporal information of the measurement data in the transient process of a power system, and further improve the accuracy of power system transient stability assessment (TSA), a TSA model based on the improved one-dimensional convolutional neural network (1D-CNN) is proposed. This model directly employs the underlying measurements as the input feature. By using a multi-size convolutional kernel to replace the traditional single-size convolutional kernel, the multi-grained temporal information of the measurement data is extracted effectively, realizing the end-to-end TSA. On the other hand, the focal loss function is introduced to guide the model training, which effectively discovers the difficult samples and alleviates the imbalanced classes of the samples, improveing the identification performance of the model. In addition, by applying the Guided Grad-CAM method the class activation map of the TSA model is visually analyzed, improving the interpretability and transparency of the model. The simulation results performed on the New England 39-bus test system demonstrate that compared with the TSA methods based on the traditional machine learning and deep learning, the proposed method has better evaluation performance, and that it is more robust to those "contaminated" data.