桥梁等基础设施的长期稳定运行攸关社会民生安全。桥梁结构健康监测传感网可长期获取桥梁动力学参数和环境参数的观测数据，是桥梁安全评估、桥梁寿命预测、车辆荷载监测等关键应用的基础。但在长期的监测过程中，传感网不可避免地会受传感器漂移和输入信号分布漂移的影响。在桥梁监测场景下，传统的离线校准成本高、时效性差，破坏数据一致性，难以开展，严重影响数据的准确性和后续应用的可靠性。针对该挑战，本文利用信号的时空冗余度，在传感网原位部署且不依赖参考真值的情况下，研究在线盲校准算法及其在车辆荷载监测中的应用。首先，本文提出了基于时空回归的传感器漂移在线盲校准算法。该算法基于观测信号真值的连续性和低秩性，将其在低维子空间内表示，并训练高斯过程回归模型；该算法进一步利用传感器漂移的缓变性和稀疏性，用稀疏向量加权的高斯过程对其进行回归，结合贝叶斯LASSO识别漂移传感器，解决了时间冗余度建模和漂移传感器识别的难题。实验结果表明，该算法可同时校准的漂移传感器数量相比现有算法提升了30%。接下来，本文提出了基于子空间投影的输入信号分布漂移在线盲校准算法。该算法基于输入信号的低维性和其分布漂移的连续性，使用变分贝叶斯方法在输入信号所在的低维子空间内校准其协方差矩阵，解决了空间冗余度建模的难题。实验结果表明，在单点推断场景下，该算法推断位移状态信号的精度相比现有算法提升了30%；在整体推断场景下，现有算法难收敛甚至无法校准，而该算法仍能成功校准；除此之外，该算法还能应用于楼宇的结构健康监测，有广阔的应用范围。最后，本文提出了基于车轴定位和承载信号分离的车辆荷载监测系统。该系统从单目交通监控视频中定位车轴在桥梁上的坐标，从桥梁位移状态信号中分离承载信号，标定分段多项式承载模型，解决了实际交通状况下多车轴、多车过桥和车辆变速的难题。实验结果表明，该系统应用于短跨径和长跨径桥梁均能将车辆荷载监测精度相比现有系统提升至少10%，且现有系统仅能适用于单一桥梁类型。本文面向桥梁结构健康监测的需求，在传感器和输入信号分布漂移在线盲校准方面的研究成果不仅可提升传感网的时效性，延长工作时间和寿命，还可实现更精确、更实用的车辆荷载监测，为治理超载顽疾、保障交通安全提供新的技术手段。本文的研究成果已应用于多个实际的桥梁结构健康监测系统。

The long-term service of infrastructures, including bridges, has been vital for the security of society and livelihood. The bridge structural health monitoring sensor networks can gather measurements of the dynamical and environmental parameters of bridges on the long-term basis, which lay the foundation of various applications, including bridge safety assessment, bridge remaining useful life prediction, and vehicle load monitoring. However, over years of operation, the sensor networks are prone to sensory drift and input signal distribution drift inevitably. In the bridge structural health monitoring scenarios, the traditional offline calibration methods exhibit high cost, poor timeliness, and can destruct the consistence of data. Therefore, they are impractical to be carried out, which will severely affect the accuracy of data and the reliability of various proceeding applications. To address the above challenges, by utilizing the spatial-temporal redundancy of the signal, this thesis studies the online blind calibration algorithms and their application in the field of vehicle load monitoring, assuming that the sensor network is deployed in place and without the help of ground truth data.First, this thesis proposes an online blind calibration algorithm of sensory drift based on spatial-temporal regression. The proposed algorithm represents the ground truth of the measurement signal in a low-dimensional subspace, and trains a Gaussian process regression model of it based on its low-rank structure and continuousness. Furthermore, the algorithm regresses the gradual-varying sensory drift based on its sparsity using Gaussian processes that are weighted by a sparse vector, and identifies the drifted sensor using Bayesian LASSO. In doing so, the proposed algorithm solves the problem of temporal redundancy modeling and drifted sensor identification. The experiment results show that the proposed algorithm can successfully calibrate 30% more drifted sensors than previous algorithms.Second, this thesis proposes an online blind calibration algorithm of input signal distribution drift based on subspace projection. Based on the assumption that the input signal is low-dimensional and the distribution drift of it is continuous, the proposed algorithm calibrates the covariance matrix of the input signal in its subspace using the variational Bayes method. In doing so, the proposed algorithm solves the problem of spatial redundancy modeling. The experiment results show that in the single-point inference scenario, the proposed algorithm estimates the displacement state signal at an accuracy of 30% better than previous algorithms; in the general inference scenario, previous algorithms diverge and fail to calibrate, while the proposed algorithm can still calibrate successfully. In addition, further experiment results show that the proposed algorithm can also be applied to the health monitoring of buildings, and has wide application prospects.At last, this thesis proposes a vehicle load monitoring system based on axle locating and carrying signal extraction. The system locates the coordinates of the vehicle axles on the bridge from the monocular traffic surveillance camera recordings, extracts the carrying signal from the displacement state signal of the bridge, and calibrates the piecewise polynomial carrying model of it. In doing so, the proposed system solves the problem of multiple vehicle axles, multiple passing vehicles, and time-varying vehicle speed induced by actual traffic conditions. The experiment results show that the proposed system can monitor the vehicle loads at an accuracy 10% better than previous systems, on both short span and long span bridges, while previous systems can only apply to one type of bridges.Targeting at the need of bridge structural health monitoring, the researches of this thesis on the online blind calibration of both sensory and input signal distribution drift not only can improve the timeliness of the sensor networks, and extend the working time and the useful life of them, but also provides a more accurate and applicable vehicle load monitoring solution, which will help to regulate overweigh vehicles and to ensure the traffic security. The researches of this thesis have been applied to several bridge structural health monitoring systems.