基于多域多尺度深度特征自适应融合的焊缝缺陷检测研究

Weld defect detection based on adaptive fusion of multi-domain and multi-scale deep features

针对焊缝缺陷检测信号信息丰富度低、深度网络架构人工依赖性强等问题,开展基于多域多尺度深度特征自适应融合的焊缝缺陷检测研究。构建时域数据集并衍生至实数域与复数域中,丰富检测信号的特征表达;设计多域信息融合模型,充分融合特征域信息;提出面向卷积神经网络多维超参数自寻优的模型优化策略,提高模型的效率和性能。试验表明,所提方法对五类焊缝缺陷识别准确率为96.54%,能够在提升识别准确率同时保持较少的参数量和计算消耗,具有较强的实用性和泛化性。

Here,aiming at problems of low information richness of weld defect detection signals and strong manual dependence on deep network architecture,weld defect detection studies were performed based on adaptive fusion of multi-domain and multi-scale deep features.Firstly,a time-domain dataset was constructed and derived into real and complex domains to enrich feature expression of detection signals.Secondly,a multi-domain information fusion model was designed to fully integrate feature domain information.Finally,a model optimization strategy oriented to multi-dimensional hyperparametric self-optimization of convolutional neural network(CNN)was proposed to improve the model’s efficiency and performance.Tests showed that the proposed method has an accuracy of 96.54%for identifying 5 types of weld defects;it can improve the recognition accuracy while maintaining a smaller number of parameters and computational consumption;it has stronger practicality and generalization.