Bridge data anomaly detection based on improved WGAN-GP and CBAM-Dense-DANN

Abstract

Data anomaly detection based on machine learning models has been widely applied in bridge health monitoring systems to automatically identify anomalous monitoring data. However, its performance is often constrained by the randomness, scarcity, and limited distributional representativeness of known anomalous data. This paper proposes a framework for data anomaly detection based on an improved Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) and a Domain Adversarial Neural Network incorporating a Convolutional Block Attention Module and dense block structures (CBAM-Dense-DANN). Firstly, traditional techniques are employed to augment the initial dataset. Subsequently, the augmented data are transformed into two-dimensional time-frequency matrices via the Short-Time Fourier Transform. Then, the two-dimensional time-frequency matrices of known anomalous data are input into the improved WGAN-GP to alleviate the scarcity of anomalous data. Finally, the CBAM-Dense-DANN is constructed by combining Entropy Minimization (EM) loss and Local Maximum Mean Discrepancy (LMMD) loss to address the limited distributional representativeness of anomalous data. Distribution feature alignment between the known anomalous data in source domain dataset and real anomaly distribution data in target domain dataset is achieved by the CBAM-Dense-DANN. This mechanism not only facilitates the efficient identification of massive unlabeled data within the target domain and significant reduction of manual annotation costs, but also realizes the deep mining and utilization of unlabeled bridge data. The effectiveness of the proposed hybrid model is validated through numerical simulations and monitoring data from a real bridge. The results demonstrate that classification accuracies of 97.56% and 98.16% are achieved on the numerical and field datasets, which exceed the classical DenseNet by 11.40% and 10.30%, respectively.