Accurate depth assessment of metal surface defects is critical for industrial safety. Traditional eddy current testing methods rely on manual expertise, resulting in low efficiency and high subjectivity. To address the limitations of existing approaches—namely the scarcity of high-quality datasets and the restricted feature extraction capabilities of models—this paper constructs a large-scale, multivariate dataset of metal pipeline defects. This dataset encompasses 20 depth levels and multiple real-world operational variables. simultaneously introducing the ReSeNet eddy current detection approach. This method incorporates a channel attention mechanism within ResNet to achieve adaptive calibration of eddy current signal features. Experimental results demonstrate that ReSeNet-34 achieves a classification accuracy of 93.71% on the test set, representing a 4.78% improvement over the baseline ResNet-34. This exhibits outstanding model performance, significantly enhancing the automation level and evaluation accuracy of eddy current detection.