Abstract: Exploratory data analysis (EDA) statistical graphs have been widely used for dynamic data quality assessment in bridge health monitoring. In order to reduce the subjectivity of manually observing statistical graphs, the quantitative analysis of statistical graphs was achieved through an approximation measurement method, and multiple indicators were obtained for rapid quality evaluation of monitoring data. The dynamic response of bridge structures exhibited short-term linear stationary characteristics and approximately followed a normal distribution under the stimulation of the operating environment. Taking the vibration data of a large-span cable-stayed bridge as the research object, sample data histograms and Q-Q plots were firstly plotted. By observing the data distribution characteristics, the data quality was prejudged and determined to be excellent, good, or poor. Subsequently, the sample data histograms and Q-Q plots were quantitatively analyzed by two approximation measure methods, namely KL divergence and cosine similarity, to obtain indicators of the degree to which the data followed a normal distribution. By detecting the global outliers of the sample data through box plots, the proportion of normal data was obtained. The corresponding relationship between the quantified values and the prior quality levels was obtained by statistical analysis. And the data quality evaluation standards were determined by using histogram KL divergence and cosine similarity as the main indicators and box plot normal data proportion as the secondary indicator. Finally, a portion of the data was taken as the validation set to further validate the rationality of each indicator of the proposed method, and the application results of the proposed method in practical engineering were given.

Key words: bridge engineering; bridge structural health monitoring (BSHM); data quality assessment; exploratory data analysis (EDA); Kullback-Leibler divergence (KL divergence); cosine similarity; box plot

摘要： 探索性数据分析统计图在桥梁健康监测动态数据质量评估中已有广泛应用。为了减少人工观察统计图的主观性，通过近似度量方法实现统计图的量化分析，得到多个指标对监测数据进行快速质量评估。在运营环境激励作用下，桥梁结构动力响应具有短时线性平稳性，近似服从正态分布。以某大跨斜拉桥振动数据为研究对象，首先，绘制样本数据直方图和Q-Q图，通过观察数据分布特征预先判断数据质量，确定优、良和差3个等级。然后，分别通过KL散度和余弦相似度2种近似度量方法对样本数据直方图和Q-Q图进行量化，得到数据服从正态分布程度的指标；通过箱线图检测样本数据全局异常点，得到正常数据占比；统计分析得到量化值和先验质量等级的对应关系，确定以直方图KL散度和余弦相似度为主、以箱线图正常数据占比为辅的数据质量评估标准。最后，取部分数据为验证集，进一步验证所提方法各个指标的合理性，并给出该方法在实际工程上的应用结果。

关键词: 桥梁工程；桥梁结构健康监测；数据质量评估；探索性数据分析；KL散度；余弦相似度；箱线图