Abstract:

In order to explore the important factors affecting the structural safety of houses in urban waterlogged areas, 21 attributes, such as house year, floor and area, were collected and selected to construct an early warning index system, and the discretization and imbalance of samples were solved by over-sampling and unique heat coding. Secondly, four different integrated algorithms and six machine learning models were used to build an early warning model to learn and test the safety data of housing structures. Then, the performance of the early warning model was compared comprehensively by applying the harmonic average of accuracy, accuracy and recall, average accuracy and area under the curve (AUC), and the correlation analysis and importance ranking of each warning index were carried out. Finally, 2 215 houses in 35 waterlogging areas in 11 counties and cities of Fujian Province were taken as examples to verify the scientific and validity of the model. The results show that: whether the house belongs to the key inspection, whether the construction team is professional, house area, year and the number of the ground floor are all more than 150, which are the most important five indicators for building safety warnings in vulnerable waterlogging areas. The early-warning model based on the lifting method strategy has the best early-warning accuracy, and the overall prediction accuracy is 99.10%. The model can detect the structural safety of houses in vulnerable waterlogging areas more accurately and efficiently.

Key words: vulnerable to waterlogging, building structure safety, integrated algorithm, machine learning, early warning model