Wi-Fi and Bluetooth Low Energy-based (BLE) fingerprint software methods are being used as indoor positioning systems because of the advantage of recognizing position without a separate sensor, even in mobile devices. However, received signal strength indication (RSSI) may have an error of several with unstable values caused by surrounding obstacles or multipath effects. In this paper, to compensate for the shortcomings of the existing RSSI-based fingerprint positioning technique, we propose a method for generating Markov Transition Field (MTF) data by synthesizing noise with RSSI, time-series data, and a deep-learning-based indoor positioning model learned with the data. We use real-world collected datasets for model learning and performance comparisons. The proposed method learns RSSI and noise with variability in offline test environments by synthesizing noise in time series data by Convolutional Neural Network (CNN) models. Based on the test results, the positioning error of the model using the proposed method was 0.4639, and the positioning error of the model without synthesizing noise was 1.3861, indicating that the error distance was three times less.