The increasing demand for high-precision location-based services is driving the development of advanced positioning technologies, essential for applications like navigation, emergency response, and asset tracking. While satellite networks like GPS is effective in open outdoor environments, urban canyons and indoor scenarios present significant challenges. Traditional technologies such as WiFi and Bluetooth face limitations in range and incur high deployment costs. However, the 3GPP Release 16 standard for 5G technology offers integrated sensing and communication capabilities, enabling six positioning schemes that can achieve meter or sub-meter level accuracy. Therefore, this work proposes a deep learning-based approach to enhance 5G New Radio (NR) positioning by utilizing a fine-grained dataset generated through a multilevel feature synthesis method, compliant with 5G NR specifications. By condensing position-related features into channel frequency response matrices, we increase information density and improve feature representation. Data augmentation techniques are employed to enhance noise robustness. Our deep and convolutional neural networks demonstrate significant improvements in indoor positioning accuracy and reliability. Extensive experiments in typical 5G indoor scenarios show superior performance compared to state-of-the-art methods, particularly in terms of accuracy and noise resilience.