Effective multilayer hybrid classification approach for automatic bridge health assessment on large-scale uncertain data

The health level of the bridge is critical to the safety and maintainability of the bridge. However, with the rapid increase of bridge data with complex information, manual evaluation of bridge health requires high labor and time costs and a lot of related knowledge. Meanwhile, due to the error of sensors and subjective judgment of experts, as well as the influence of the external environment of the bridge, there is great uncertainty in the evaluation of the bridge data. Thence, how to use a large amount of bridge data with large-scale uncertain labels to build a robust classification model for efficient and high-quality bridge health assessment has become an urgent task. In order to better assess the health of the bridge, we adopt a multi-layer hybrid method to iteratively determine the uncertain labels of the target dataset, evaluate the confidence of the large-scale uncertain labels, add high-confidence data to the training set, and correct the low-confidence data. Finally, we get an effective classification model with the optimized training dataset. This paper studies the learning problem of classification model on labeled data with large-scale uncertain labels, and proposes an effective hybrid classification model (HCM), which can establish a supervised classifier under the condition of detecting uncertain labels and realize error label correction. In order to measure the HCM label assignment problem, we introduce a new penalty function, which can evaluate the label consistency problem of two basic classifiers. Meanwhile, we apply the model to bridge data with uncertain labels for bridge health evaluation. Experiments conducted on synthetic data, benchmark data and real bridge datasets show that the proposed method is superior to other methods and provides an effective and convenient solution for bridge health assessment. At the same time, this method can also be used in other research fields where there are large-scale uncertain labels.