In the shipbuilding industry, pre-welding defect detection is an essential phase. In the process of intelligentizing this phase, due to the influence of natural and unnatural factors in the data collection step, some data are polluted by noise or have not obvious features. The existence of such data leads to poor training effects of neural networks, which in turn results in poor performance of the trained models during the prediction process. To address this issue, this paper, inspired by the divide-and-conquer philosophy, proposes a method for defect detection prior to ship hull welding based on distribution adaptive. In this method, a brand-new training strategy is adopted, and a perturbation scoring mechanism is used to design a discriminator that can identify data polluted by noise or have not obvious features. By searching for the optimal perturbation, the feature scoring difference between data polluted by noise or have not obvious features and other data is maximized. The optimal perturbations are added to the data as input to the model, and the output results of the neural network are scored used, this score in combination with the calculated threshold, it is possible to determine the distribution to which the input data belongs is judged. Finally, according to the proposed training strategy, the data with different distributions are predicted separately. Experiments have proven that the method proposed in this paper significantly improves model performance.