This study proposes a new deep learning-based anomaly detection algorithm in a supervised framework for the Monte-Carlo (MC) simulation data of a missile control system. First, in a similar fashion to the famous BERT(Bidirectional Encoder Representations from Transformers) algorithm in NLP(Natural Language Programming), a pre-trained transformer-based encoder is utilized to provide an alternative representation of time-series data, replacing conventional data-preprocessing. For a more direct perception of differences in normal and abnormal data, a multi-head attention score is also considered as an input of a two-dimensional convolutional neural network classifier along with the output of transformer encoder. The training of the proposed framework is comprised of two steps: unsupervised pre-training of the transformer encoder with all generated MC simulation data and supervised training of the classifier with labelled data. To generate the MC simulation data, the high-fidelity missile 6DOF simulation program is developed with random variables that represent the uncertainties of realistic scenarios. The resulting anomaly detection performance from the generated MCsimulation dataset indicates that the proposed framework provides significantly better detection performance than other supervised machine/deep learning-based algorithms. Furthermore, using a multi-head attention score from the transformer encoder as an input of the classifier not only improves the detection performance but also contributes to reducing the necessity of labelled train data. Hence, the proposed method has the potential to greatly reduce the human effort in detecting anomalies in MC simulation data.