Intelligent optimization of pressure sensor arrangement and prediction of wind loading time series on a square cylinder

This paper proposes a novel methodology for classifying and arranging key pressure sensors on structural surfaces by integrating Proper Orthogonal Decomposition (POD) and K-means++ cluster analysis. For the wind tunnel experimental time series data obtained from multiple pressure sensors on a square cylinder, POD-K-means++ quickly groups the sensors into several clusters and produces the silhouette scores. Using the silhouette score distributions at various incidences, the arrangement scheme of key sensors is proposed, considering various sensor amounts and an axisymmetric condition. The POD-BPNN (Back Propagation Neural Network) method is used to predict the pressure time series on any circumferential location using data from limited sensors. By comparing the overall prediction precision of POD-BPNN using 10 different axisymmetric sensor layouts and the same sensor amount of 20, the case using the sensor layout identified by POD-K-means++ is proven to have superior precision in all incidences. POD-BPNN is further used to study the sensitivity of prediction errors to training sensor amounts (4–28 within a total sensor amount of 44). 20 training sensors achieve an optimal balance between accuracy and efficiency.