Diagnosis method for abnormal data in bridge health monitoring using BMA-assisted EasyEnsemble-AlexNet

Abnormal data diagnosis in the bridge health monitoring (BHM) system is of great significance to improve the credibility of structural safety assessment. However, a notable imbalance exists between the number of abnormal and normal samples in deep learning-based methods, resulting in a lower accuracy when identifying limited abnormal categories. To this end, this study proposes an abnormal data diagnosis method for BHM using EasyEnsemble, AlexNet and Bayesian model averaging (BMA). First, multiple balanced EasyEnsemble subsets are constructed from the training set by random sampling. Subsequently, each subset is used for the training of a single AlexNet model, and the posterior probability of each model is determined by calculating their log-likelihood on the training subset. Then, the prediction results of all models are combined by BMA according to their posterior probabilities. Finally, the effectiveness of this method is verified using data from a BHM system. The results show that the proposed method significantly strengthens the diagnosis performance of abnormal data, achieving an overall accuracy of 96.66%. Compared to traditional methods, the proposed method offers higher training efficiency and diagnosis accuracy with fewer samples, especially for similar abnormal categories.