路面异常会给驾驶员与行人带来不便，更可能引发交通事故。提出了一种通过传感器时序信号数据进行路面异常检测的算法。针对行驶过程中采集的不同传感信号之间具有较强的高阶时序相关性的问题，通过构建高阶动态贝叶斯网络分类器来实现异常检测。首先，通过相关性分析和 Granger 因果分析分别构建初始时刻网络和转移网络的初始网络; 然后，将传感信号进行小波分解，通过卷积神经元网络实现网络嵌入学习;最后，利用网络嵌入进行链路预测，结合 MDL 评分实现网络修正学习算法。实验结果表明，该检测算法相对于传统的时间序列分类方法，在分析时序相关性较大的信号数据时，具有更低的误检率和漏检率、更高的 F1 值，并且更加鲁棒。

Pavement anomalies can bring inconvenience to drivers and passengers，and even cause traffic acci-dents． A pavement anomaly detection algorithm based on sensor time series data was proposed． Considering the problem that sensing signals collected during driving are strongly high-order sequential correlative，high-order dy-namic Bayesian network classifier was constructed to realize the anomalies detection． Firstly，correlation analysis and Granger causality test were used to initialize the network structure． Secondly，the sensing signals were decom-posed by wavelet transform，and convolution neural network was used to realize network embedding． Finally，link prediction with minimal description length was used to optimize the network structure． The results show that，com-pared with the traditional method of time series classification，the proposed method can reduce fallout rate and missing rate，and increase F1 score on the sequential correlative signals，and thus is more robust．