@article{WANG2023, 
author = {Lei WANG and Xiaoling WANG and Jun ZHANG and Jia YU and Jiajun WANG},
title = {Comprehensive evaluation method for dam safety considering fusion weight optimization and conflicting information sources},
year = {2023},
journal = {Journal of Tsinghua University (Science and Technology)},
volume = {63},
number = {10},
pages = {1566-1575},
keywords = {dam safety comprehensive diagnosis, evidence conflict, distance evidence, time series decomposition, isolation forest (iForest)},
url = {https://www.sciopen.com/article/10.16511/j.cnki.qhdxxb.2023.22.039},
doi = {10.16511/j.cnki.qhdxxb.2023.22.039},
abstract = {ObjectiveEvidence theory, an effective multisource information fusion method, has been widely applied in the field of dam safety monitoring and health diagnosis. However, two limitations must be addressed: the accuracy of evaluation results of existing conflicting evidence-processing methods is low, and the existing fusion weight calculation methods may result in unreasonable fusion weights when the differences in the contribution of each monitoring type to the dam safety evaluation results are ignored. In this paper, we propose a comprehensive evaluation method for dam safety considering fusion weight optimization and conflicting information sources. The negative impact of conflicting evidence on the accuracy of information fusion is significantly eliminated by means of both decreasing conflicting information generation sources and preprocessing evidence bodies. Furthermore, reasonable fusion weights for each monitoring type can be obtained through subjective assignment.MethodsIn the data preprocessing stage, we propose the STL (seasonal-trend decomposition procedure based on loess)-iForest (isolation forest algorithm) method for temporal outlier detection. First, the original dam safety monitoring sequence data are decomposed into trend, seasonal, and residual terms using the STL temporal decomposition method, which eliminates the influence of data trend and periodicity on the total error rejection accuracy. Then, the iForest algorithm is used to remove outliers from the data of the residual term. Finally, the data are reconstructed using the Hermite interpolation method. Using the above methods, we suppress the generation of conflicting evidence at the source as much as possible. In the multisource information fusion stage, we calculate the evidence fusion weight of low-level information using the χ2 divergence and Shannon entropy, which can overcome the shortcomings of traditional distance metrics, such as the Euclidean distance, which have poor reliability of fusion results. For high-level information, we use the scaling method to determine the corresponding fusion weight to ensure that the weight reflects the difference in the contribution of different monitoring types to the dam safety evaluation results.ResultsThe results with data preprocessing show that the STL temporal decomposition method separates the trend term and seasonal term of the data well and keeps the outliers in the residual terms, providing a good basis for outlier detection. Then, the iForest algorithm can detect all obvious outliers in the data, which shows that the proposed outlier detection method can ensure the high reliability of the multisource monitoring data by eliminating conflicting evidence from the sources. In the multisource information fusion method proposed in this paper, the discrete degree and uncertainty of the evidence are fully considered to ensure the accuracy of the evaluation results, especially in the case of conflicting evidence, which is consistent with the actual situation. Compared with evidence theory based on the Euclidean distance and the case of unexcluded data coarseness, the proposed method improves the probability corresponding to the basic normal state to 0.842, indicating that the method has better reliability.ConclusionsThe proposed method shows high-accuracy fusion results and high-reliability qualitative evaluation results. In addition, it can accurately quantify the differences in the contributions of each monitoring type to the comprehensive dam safety evaluation results. The proposed method has considerable potential in the diagnosis of dam safety, given its high accuracy and good robustness.}
}