The telemetry data for missile mass production assembly is large, with multiple parameters, significant waveform differences, and few fault samples. Manual interpretation is time-consuming and laborious, and it is also difficult to automatically analyze features through code description. Therefore, research has been conducted on the rapid analysis and fault diagnosis of complex time-series telemetry data. By aligning and calibrating the recorded telemetry time-series data, the method of converting the data into waveform images and inputting them into a deep learning model for fault diagnosis model training has been adopted. Generative adversarial network technology has been used to augment small sample fault data, and a VGG16 transfer learning model has been established. Based on historical missile fault data, intelligent fault diagnosis simulation experiments were conducted for a certain missile test parameter. The model loss function was reduced to 0.04, and the accuracy of the validation set reached 99%. The experimental results verified the correctness and effectiveness of the proposed fault diagnosis model. Applying the model to the batch production and final assembly of actual missile models can significantly improve product production efficiency, enhance batch quality, and reduce risks such as post production maintenance and batch repairs.