With the advancement of array signal processing and its widespread application across various fields, signal sources enumeration remains a fundamental problem. The traditional methods are widely used in many fields but significantly degrade performance when the signal-to-noise ratio (SNR) is low. In recent years, various deep neural networks (DNNs) have been proposed to estimate the number of sources, and have strong performance even under low SNR. However, while DNNs effectively overcome the limitations of low SNR, their generalizability is lower than traditional methods. Traditional can be applied across scenarios with different numbers of sensors. DNN-based enumeration for the number of sources typically requires separate models tailored to a specific number of sensors, thereby increasing application costs. To enhance the performance of DNNs in the sources enumeration, this study proposes two eigenvalue padding methods and the natural logarithmic (log) transform as preprocessing steps. Simulation results confirm that the two proposed padding methods significantly improve model generalizability, while the log transform further enhances enumeration accuracy.