Regional safety management using bi-directional building networks based on LSTM with time and frequency domain data

This study presents a bi-directional building network for regional safety management. In the network, the relationship between the structural responses of multiple buildings in a region is bi-directionally established via long short-term memory (LSTM). To build a network, the buildings’ frequency and time domain response were used to consider the characteristics of buildings. In the model, the structural response of the building, which was assumed to suffer data loss, was intentionally excluded, and the responses of the remaining buildings were set as inputs. The excluded response was set as the output of the model. Even when data loss occurs in any of the target buildings, the model trained using datasets enabled the prediction of lost data using the responses of the remaining buildings. The proposed method was applied to the safety management of a specific region to which multiple buildings were assumed to be adjacent.