Existing mine tunnel field strength prediction models are often complex, have low accuracy, and lack strong generalization. To overcome these challenges, this paper presents a CNNSE-based path loss prediction model for mine tunnels. The model considers factors such as antenna frequency, tunnel wall roughness, and humidity as input features, with path loss as the output, allowing for accurate path loss predictions. The approach incorporates the SE module and multi-scale dilated convolution to improve the model’s ability to capture electromagnetic wave propagation patterns in complex environments, enhancing its attention to crucial features. PGD adversarial training is also used to boost the model's generalization, ensuring robust and accurate predictions in mining conditions. Experimental results indicate that the model achieves an average absolute error of less than 0.5 dB and a correlation coefficient of 0.9969, demonstrating significant improvements in both prediction accuracy and generalization.