Temperature-induced deflection separation based on bridge deflection data using the TVFEMD-PE-KLD method

The bridge deflection data measured in field are greatly affected by temperature. In some situations, temperature can be the dominant factor comparing with loads and other factors. Generally, dynamic deflection contains information related to the health condition of bridge. It is wildly used in dynamic deflection warning system. The temperature-induced deflection might lead to the inaccuracy of such system. Therefore, it is important to separate the temperature-induced deflection when implement bridge health condition assessment using dynamic deflection data. However, the accuracy of existing methods of temperature-induced deflection separation is low. This low accuracy is normally caused by mode aliasing and excessive elimination of feature components, etc. To this end, this study proposed an innovative method of temperature-induced deflection separation based on the time-varying filtered empirical mode decomposition (TVFEMD), permutation entropy (PE) and Kullback-Leibler divergence (KLD). The proposed method is used to extract temperature-induced deflection from bridge deflection data. It overcomes the problems of mode aliasing and end-point effects by reconstructing the signals with the same composition. The separating accuracy is also improved by the non-linear feature extracting ability of the proposed method. First, TVFEMD was introduced to decompose the original bridge deflection data into a number of intrinsic mode functions (IMFs). Then, PE was adopted to measure the complexity of each IMF, by which all IMFs can be reconstructed into several new subseries. Furthermore, KLD was employed to calculate the divergence values between these reconstructed subseries and the measured ambient temperatures. The subseries with the smallest divergence value can be considered to be the temperature-induced deflection. Finally, numerical study and locale measurement were utilized to comprehensively validate the effectiveness of the proposed method. The results show that the identification accuracy of the proposed method exceeds 90%. It outperforms the other involved models in terms of separating daily and annual temperature-induced deflection.