FCGNet: A Deep Learning Methodology for Real-Time Prediction of Fatigue Crack Growth Rate Using Acoustic Emission Signals

The acoustic emission (AE) has gained popularity in real-time prediction of fatigue crack growth rate (CGR). However, the past methods rely significantly on the fatigue cycle and threshold-based AE parameters, such as count rate, event rate, and energy rate, making them less appealing for real-time applications. Therefore, we introduced a novel methodology for fatigue CGR prediction that eliminates the need for fatigue cycles and threshold-based AE parameters, making it more suitable for real-time situations. This study uses in situ AE signals acquired during the propagation of fatigue cracks in ship steel DMR 249A. The fatigue experiments were conducted for three different load ratios of 0.1, 0.25, and 0.5 with the in situ AE sensors. These load ratios correspond to the sea-state IV, where waves’ height reaches 1.25 to 2.5 m, and the wind is moderate. The detected AE signals are utilized to predict the fatigue CGR by using a hybrid deep learning model consisting of long short-term memory (LSTM) and convolutional neural network (CNN). Prior to prediction, raw AE data undergoes meticulous preprocessing, involving the application of B-spline basis functions and first-order derivatives. The projected CGR attains strong correlations of 79.64%, 86.87%, and 94.42% for load ratios 0.1, 0.25, and 0.5, respectively, in comparison with the actual CGR values. The same is also validated through the microstructural characterization. Subsequently, the proposed hybrid DNN model is deployed on the Nvidia Jetson Nano platform. This implementation demonstrates the model’s adaptability to real-world industrial challenges, when used on edge devices, showcasing the potential of AE signals for real-time CGR prediction.