In this paper, a crack diagnosis framework is proposed that combines a new signal-to-imaging technique and transfer learning-aided deep learning framework to automate the diagnostic process. The objective of the signal-to-imaging technique is to convert one-dimensional (1D) acoustic emission (AE) signals from multiple sensors into a two-dimensional (2D) image to capture information under variable operating conditions. In this process, a short-time Fourier transform (STFT) is first applied to the AE signal of each sensor, and the STFT results from the different sensors are then fused to obtain a condition-invariant 2D image of cracks; this scheme is denoted as Multi-Sensors Fusion-based Time-Frequency Imaging (MSFTFI). The MSFTFI images are subsequently fed to the fine-tuned transfer learning (FTL) model built on a convolutional neural network (CNN) framework for diagnosing crack types. The proposed diagnostic scheme (MSFTFI+FTL) is tested with a standard AE dataset collected from a self-designed spherical tank to validate the performance under variable pressure conditions. The results suggest that the proposed strategy significantly outperformed classical methods with average performance improvements of 2.36 - 20.26%.

本文提出了一种裂缝诊断框架，该框架结合了新的信号到成像技术和转移学习辅助的深度学习框架，以使诊断过程自动化。信号到成像技术的目标是将来自多个传感器的一维（1D）声发射（AE）信号转换为二维（2D）图像，以捕获可变操作条件下的信息。在此过程中，首先将短时傅立叶变换（STFT）应用于每个传感器的AE信号，然后融合来自不同传感器的STFT结果，以获得裂纹的条件不变2D图像；这种方案称为基于多传感器融合的时频成像（MSFTFI）。MSFTFI图像随后被馈送到基于卷积神经网络（CNN）框架构建的用于诊断裂纹类型的微调传递学习（FTL）模型。建议的诊断方案（MSFTFI + FTL）用从自行设计的球形储罐收集的标准AE数据集进行测试，以验证在可变压力条件下的性能。结果表明，所提出的策略明显优于经典方法，平均性能提高了2.36-20.26％。