The intelligent defect detection algorithm based on electromagnetic ultrasonic technology can be used to monitor the quality status of important parts and ensure the safe and reliable operation of equipment. In the actual detection process, on the one hand, the collected signals are often polluted by noise, which interferes with the detection results. On the other hand, the defect signals of important parts often have less data and cannot meet the needs of neural network training. Therefore, this paper proposes a noise reduction algorithm based on variational mode decomposition to pre-process detected signals to improve signal quality, proposes an improved virtual sample generation technology to expand the sample set, and uses transfer learning technology to reduce the number of parameters in neural network training to solve the problem of insufficient sample number. The average prediction accuracy of this method is 97.2% in the depth detection example of aluminum plate surface defects. Therefore, this method has certain reference significance for the surface defect detection of non-ferromagnetic materials.