A multitask Gaussian process regression method for crack prediction in composite structures based on multipath damage feature fusion

Abstract

To address the problem of crack propagation in composite structures used in rail transit, this article proposes a precise quantitative monitoring method for crack length based on multitask Gaussian process regression (MT-GPR). First, a network of piezoelectric sensors is deployed on the structure surface to collect ultrasonic-guided wave signals before and after damage occurs. Then, four typical damage indicators are extracted from both the time and frequency domains. Principal component analysis is applied to fuse these features across multiple propagation paths and excitation frequencies, thereby enhancing the representation capability of the damage features. To cope with the signal discrepancies caused by the anisotropy of composite materials and the irreversibility of crack growth, as well as the challenge of limited samples, a MT-GPR model capable of sharing information across tasks is constructed for high-precision crack length prediction. Experimental results demonstrate that, compared with conventional single-task GPR methods, the proposed approach achieves higher prediction accuracy under small-sample conditions and provides confidence intervals for effectively quantifying the uncertainty of the prediction results.