The present invention relates to a radar signal railway roadbed disease detection method based on a radar signal of a deep convolution neural network. According to the method, through the automatic scanning and sampling of a vehicle-borne roadbed detection radar, roadbed original radar data is obtained, then a radar reflection signal sequence of an interface on each scanning way track sleeper is extracted and spectrum analysis is carried out, and spectrum grayscale maps of sleeper and rail strong reflection signals are obtained. Through marking a disease in a collected partial signal spectrogram, the deep convolution neural network is designed to extract a signal analysis characteristic pattern, a candidate area network and a full connection layer are employed, a railway roadbed disease detection model is built by multiple iterations, and roadbed disease classification and detection box prediction are obtained. According to the method, the deep convolution neural network is used to analyze the sleeper and rail strong reflection signals in railway roadbed ground penetrating radar signals for the first time, the rapid detection and disease identification of ballasted railway roadbeddiseases are realized, technical support is provided for the rapid remediation and treatment of the roadbed diseases, the requirements of automatic detection and rapid maintenance of a line in the future are satisfied, and the safety of railway operation is ensured.