Diagnosis of fatigue crack growth with recursive random weight networks

• The fatigue crack growth process is treated as a neural network system.
• The recursive random weight networks are developed to diagnose fatigue crack growth.
• The input weights of networks are uniformly randomly selected.
• The output weights of networks are optimized with the batch learning of least squares.

Recursive random weight networks (RRWNs) have been developed to diagnose fatigue crack growth in ductile alloys under variable amplitude loading. The fatigue crack growth process is considered as a recursive network system. RRWNs are constructed by taking the current loading, crack opening stress, and the previous computed crack length as inputs of the network system. The input weights of conventional single-layer feed-forward neural networks are uniformly and randomly selected. The output weights of RRWNs are globally optimized with the batch learning type of least squares. The trained RRWNs are capable of determining the dynamics of crack development. The proposed model is validated with fatigue test data for different types of variable amplitude loading in alloys. Compared with other experimental diagnosis models, RRWNs show excellent performance in predicting crack length growth.

Figure optionsDownload as PowerPoint slide