An intelligent laser ultrasonic imaging and classification method for cracks in thermal barrier coating substrates...

Thermal barrier coating (TBC) has long served in extreme thermal-mechanical coupled environments, rendering it highly susceptible to damage such as substrate cracking. Due to the multilayered, heterogeneous nature of TBC, conventional ultrasonic testing faces critical bottlenecks, including severe acoustic field scattering and difficulty in extracting weak damage features. To address these challenges, this paper proposes an intelligent laser ultrasonic imaging and classification method based on a physics-constrained dual-channel one-dimensional convolutional neural network (DC-1DCNN). First, a high-fidelity three-dimensional (3D) Multiphysics numerical model of TBC is constructed. A dual-driven dataset is generated through both numerical simulations and limited experimental data. Furthermore, a transformation matrix is introduced to establish a feature-interaction mapping mechanism between the virtual numerical space and the physical-entity space, thereby overcoming the bottleneck of poor generalization for damage samples in multi-layered heterogeneous structures. Meanwhile, a constraint matrix is constructed based on surface wave amplitudes, and an innovative threshold sparse self-attention mechanism (TSSAM) is proposed. This mechanism strongly constrains the depth of feature extraction in non-stationary ultrasonic signals, significantly enhancing the network's adaptive capability to capture weak crack features at heterogeneous interfaces. Experimental results demonstrate that the model achieves a test accuracy of 96.72 %, with the sparse self-attention mechanism improving it by 4.22 %. This method achieves an end-to-end mapping from single-point time-domain A-scan signals to high-performance C-scan imaging and enables precise crack classification, providing a new paradigm for the efficient and reliable non-destructive evaluation of multi-layered heterogeneous structures.