基于多通道小波卷积神经网络的路面异常检测算法

Pavement anomaly detection algorithm based on multi-channel wavelet convolutional neural networks

为了保证路面质量和行人与驾驶员的安全,提出了一种利用传感器时序多变量信号数据进行路面异常检测的算法.针对行驶过程中需要结合多种传感器信号在不同尺度对路面特征进行分析的问题,提出结合小波卷积网络和多通道网络技术,实现路面异常检测.首先,在多级小波变换间加入卷积神经元网络,从多个尺度分析单个传感器信号的局部连续性;然后,构建多通道神经网络,将多个传感器信号分别作为不同通道的输入,计算多个信号相结合的特征向量;最后,使用多层感知机根据多通道小波网络的输出实现路面异常检测.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,同时考虑了多尺度分析、信号局部连续性和多变量信号的结合,在分析多变量时序信号数据时,具有更低的误检率和漏检率,更高的F1值.

To ensure road surface quality and safety for pedestrian and drivers,apavement anomalies detection algorithm based on multi-variate sensor time series data analysis is proposed.In order to realize multi-scale analysis on multi-variate sensor signals collected during driving,wavelet convolutional neural networks and multi-channel networks are used for the anomalies detection.Firstly,convolutional neuron networks are inserted between the multilevel wavelet transforms to analyze the local continuity of a single sensor signal on multiscale.Secondly,the multi-channel neural network is constructed,of which multi-sensor signals are used as the input of different channels respectively,and the feature vectors of multisignal are calculated.Finally,multi-layer perceptron is used to detect abnormal road surface according to the output of multi-channel wavelet network.It is shown that the proposed method utilizes the combination of multi-scale analysis,local continuity of signals and combination of multi-variable signals,reduces the false detection rate and missing detection rate,and increases F1 score on the sequential correlative signals.