Monitoring of weld damage based on deep learning and acoustic emission methods

This study employs acoustic emission (AE) technology to monitor the tensile tests of stainless steel weld specimens, analyzing and comparing the mechanical properties of the welds and base materials, as well as the AE characteristic differences during crack propagation. The results indicate that rise time (RT) is introduced for the first time as a quantitative indicator for identifying the initiation and propagation of early microcracks in welds. By integrating RT with the RA and AF parameters, the tensile process of welds can be divided into three distinct stages: elastic, strengthening, and fracture. Furthermore, a novel hypothesis for weld damage stages is proposed based on AE waveform characteristics. Additionally, the maximum load ratio (L/Lmax) is introduced as a quantitative indicator for the first time, revealing that the load-bearing capacity of welds during early microcrack formation is significantly lower than that of the base material. To further enhance damage characterization, wavelet transform is employed to extract the time-frequency and amplitude-frequency features of AE signals. An image dataset is constructed and utilized in the GoogLeNet model for intelligent identification of weld damage states. The model exhibits excellent convergence and high accuracy, thereby optimizing the completeness of feature extraction and overcoming the limitations of traditional methods in analyzing the damage evolution process.