Vibration-Based Structural Damage Identification Using P-CNN and Time-Frequency Hybrid Index under the Conditions of Uncertainties and Incomplete Measurements

Abstract

The accuracy of vibration-based structural damage identification is affected by uncertainties in vibration measurements and finite element modeling. Moreover, a sufficient number of sensor measurements are not practical in large-scale structures. In this regard, the frequency domain [frequencies and mode shapes (FMS)] and time domain [acceleration cross-correlation function (ACCF)] indexes are put into a parallel convolutional neural network (P-CNN, a new convolutional neural network architecture with dual-channel) to locate and quantify structural damage. First, this approach is verified by a numerical model of a simply-supported beam, and the performance is evaluated by comparing it with methods using FMS and ACCF indexes input to the conventional 2D-CNN, respectively. The comparative results demonstrate that the proposed method can identify the damage with the lowest error, and the errors are less than 5%. In addition, three sparse measurement conditions are adopted to investigate the effect of the number of measuring points on damage identification. It is found that under the influence of uncertainties, even if only four sensors are used, this approach can identify damage accurately. Second, an engineering example of a continuous rigid frame bridge is adopted to validate the feasibility of the proposed method. The results show that when the number of sensors accounts for 65% and 47% of test sensors, respectively, it performs well in locating and quantifying structural damage. Therefore, this method can not only reduce the number of sensors and eliminate uncertainties effects, but also improve the performance of the 2D-CNN through information fusion and complementarity, which is also beneficial to minimize the cost of sensor layout in actual structures.