在智能化制造技术蓬勃发展的大背景下，通过工业大数据与人工智能为现代无损检测技术提供有力支撑，正成为未来结构安全评估领域发展的必然趋势。采用无损检测方法对铝材进行质量控制是铝材在制造、仪器、航天等行业中得以广泛应用的必要保障。其中，电磁超声无损检测的智能化应用正成为未来相关工程领域发展的重要趋势，深度神经网络与工业数据的有机融合将为铝材缺陷识别与量化提供新的解决途径。本文以缺陷高精度智能量化为导向，提出了基于超声导波缺陷量化物理知识的物理启发式深度神经网络，构建了以聚焦换能器设计为基础、量化算法为支撑、深度神经网络为手段的智能超声无损检测研究框架。以聚焦换能器结构设计为基础，提出了基于洛伦兹力原理的SH导波聚焦电磁超声换能器以实现对铝板缺陷的高灵敏度检测，并从理论基础、结构设计、实验验证、参数分析与优化四个方面对所设计的聚焦换能器进行全面研究。研发了聚焦性能更为优异的单向斜聚焦型换能器，在消除了对称位置额外焦点影响的同时也大幅提高了导波聚焦强度。以缺陷量化算法为支撑，提出了一种物理启发式导波缺陷量化混合模型，并采用该模型以及导波量化特征、逻辑与理论搭建了基于超声导波缺陷量化物理知识的物理启发式深度神经网络。在网络训练中，采用了输入预处理、网络结构优化和损失函数重定义三种方法以提高模型的物理可解释性和小数据集适用性。以物理启发式深度神经网络为手段，通过分析实体层次性和守恒性假设下的混合模型适用范围，分别提出了面向体积型缺陷低阶聚焦导波量化的物理启发式递进残差深度卷积网络ProResNet，以及面向裂纹缺陷高阶聚焦导波量化的物理启发式门控循环单元网络GuwNet。前者分析了缺陷量化特征与逻辑，提出了输入预处理与网络结构设计方法，实现了对体积型缺陷的高精度量化；而后者将缺陷量化逻辑与理论相结合，从网络结构与损失函数着手实现了对裂纹缺陷的智能量化。本文面向铝材电磁超声无损检测场景提出了一套缺陷智能检测与量化的完整解决方案，首次实现了无损检测理论和工业数据的全面深度融合，从而在提高网络可解释性的同时大幅降低了深度神经网络对大规模数据集的依赖。希望本文工作可为无损检测和智能制造领域相关从业人员提供有效助益。

With the vigorous development of intelligent manufacturing technology, providing robust support for modern nondestructive testing (NDT) technology through big industrial data and artificial intelligence is becoming an inevitable trend in the future development of structural safety assessment. Aluminum material has been widely used in manufacturing, instrumentation, aerospace, and other industries, so it is necessary to use NDT methods to control the quality of aluminum materials. In addition, the intelligent application of electromagnetic acoustic NDT is becoming an essential trend in the development of related engineering fields. The organic fusion of deep neural networks and industrial data will provide a new solution for identifying and quantifying aluminum defects. In this paper, guided by the high-precision and intelligent quantification of defects, a physics-informed deep neural network based on the physical knowledge of ultrasonic guided wave defects quantification is proposed. The research framework of intelligent ultrasonic NDT is constructed including the structure design of a focusing transducer, defect quantification algorithm, and deep neural network.Based on the structural design of the transducer, a shear horizontal (SH) guided wave focusing electromagnetic acoustic transducer based on the principle of Lorentz force is proposed to achieve high sensitivity defects detection to aluminum plates. In addition, this paper conducts comprehensive research on the designed focusing transducer from four aspects: theoretical basis, structural design, experimental verification, and parameter analysis and optimization. A unidirectional oblique focus EMAT with better focusing performance is designed: it eliminates the influence of extra symmetrical focus and greatly improves the signal intensity of the guided wave.Based on the defect quantification algorithm, this paper proposes a physics- informed guided wave hybrid model to achieve defect quantification. It builds a deep neural network using the physical knowledge of ultrasonic guided wave testing. In the network training, three methods of input data preprocessing, network structure optimization and loss function redefinition methods are used to improve the model’s physical interpretability and the applicability of small industrial data sets. By analyzing the suitability of hybrid models under the assumption of entity hierarchy and conservation，this paper proposes a physics-informed deep neural network and a progressive residual deep convolutional network ProResNet for the low-order focusing guided wave volumetric defects quantification respectively. Moreover, a gated recurrent unit network GuwNet for the high-order focusing guided waves quantification of crack defects is also proposed. The former analyzes the characteristics and logic of defect quantification, proposes input preprocessing and network structure design methods, and realizes high-precision quantification of volume defects. The latter combines defect quantification logic and theory, and realizes intelligent quantification of crack defects from the network structure and loss function.This paper proposes a complete solution for the intelligent detection and quantification of defects in the electromagnetic acoustic NDT scene of aluminum materials. This solution achieves the comprehensive and in-depth integration of NDT theory and industrial data for the first time, thereby significantly reducing the dependence of deep neural networks on large-scale data sets while improving the interpretability of the network. The author hopes this work can provide practical assistance to relevant practitioners in NDT and intelligent manufacturing and provide a valuable reference for advancing related research on the focusing guided wave defect detection.