GRU neural network-based method for box girder crack damage detection

  Objectives   With the development of intelligent ships, it has been difficult for traditional crack damage detection methods to meet the detection requirements. This paper proposes a real-time crack damage detection method for box girders based on a gated recurrent unit (GRU) neural network.

  Methods  Using the secondary development technology of Abaqus based on the Python language, a box girder crack damage model is built, and its acceleration response under dynamic Gaussian white noise excitation is calculated. A dataset is generated by expanding the original data using the data cropping method, and the influence of noise is considered. A box girder crack damage detection model based on GRU is established, the acceleration response dataset is directly used as input and the minimum loss function value is used as a target to train the model. This method is then compared to the wavelet packet transform-based multi-layer perceptron (WPT-MLP) model.

  Results  The comparison shows that the GRU model proposed in this paper has higher detection accuracy than the WPT-MLP model in damage location and extent detection. It is less sensitive to noise and has higher accuracy in approximate prediction.

  Conclusions  The results of this study verify the applicability of GRU neural networks in the crack damage detection of box girders containing multiple plates.