Abstract: Aiming at the problem that the composite mechanical connection structure is easy to be damaged, but it is difficult to monitor the strain at the connection location, a neural network-based strain field reconstruction method for composite hole-containing feature parts is proposed. Firstly, the feature samples with different pore diameters are designed according to the composite mechanical joint structure, and the multilevel loading spectrum is determined. Secondly, the data set is constructed based on numerical simulation, and the reconstruction accuracy of the strain field by support vector machine, extreme learning machine, random forest and BP neural network are compared to each other. It is determined that the reconstruction model based on BP neural network has a higher accuracy, so that the preview of the reconstructed strain field is accomplished. Finally, the multistage loading test is conducted on the 5 mm aperture sample, and the proposed measurement points are obtained by correcting the numerical simulation results. The strain reconstruction at the location of the largest mechanical response is completed by taking the strain information of the measured and proposed points as input, and the average reconstruction error is 6.4%, which verifies the feasibility of the method.

Key words: Composite, Stress concentration, Neural network, Strain reconstruction