Wind pressure data reconstruction using neural network techniques: A comparison between BPNN and GRNN

• Wind pressure monitoring data reconstruction in sensor fault is addressed.
• Bayesian-regularized neural network (NN) model performs best in three NN models.
• Reconstruction capability degrades as faulty sensor away from sensor array center.

Structural health monitoring (SHM) technique is increasingly used in civil engineering structures, with which the authentic environmental and structural response data can be obtained directly. To get accurate structural condition assessment and damage detection, it is important to make sure the monitoring system is robust and the sensors are functioning properly. When sensor fault occurs, data cannot be correctly acquired at the faulty sensor(s). In such situations, approaches are needed to help reconstruct the missing data. This paper presents an investigation on wind pressure monitoring of a super-tall structure of 600 m high during a strong typhoon, aiming to compare the performance of data reconstruction using two different neural network (NN) techniques: back-propagation neural network (BPNN) and generalized regression neural network (GRNN). The early stopping technique and the Bayesian regularization technique are introduced to enhance the generalization capability of the BPNN. The field monitoring data of wind pressure collected during the typhoon are used to formulate the models. In the verification, wind pressure time series at faulty sensor location are reconstructed by using the monitoring data acquired at the adjacent sensor locations. It is found that the NN models perform satisfactorily in reconstructing the missing data, among which the BPNN model adopting Bayesian regularization (BR-BPNN) performs best. The reconstructed wind pressure dataset has maximum root mean square error about 23.4 Pa and minimum correlation coefficient about 0.81 in reference to the field monitoring data. It is also shown that the reconstruction capability of the NN models decreases as the faulty sensor location moves from center to corner of the sensor array. While the BR-BPNN model performs best in reconstructing the missing data, it takes the longest computational time in model formulation.