汽车制动力曲线异常检测
Anomaly detection of automobile braking curves

研究了制动力曲线异常检测方法,分析了回踩异常特性,考虑了制动力检测工况和制动力曲线变化趋势,基于余弦相似度与相对误差,对制动力数据进行聚类,建立了制动力曲线分段算法; 将制动力曲线分为阻滞段、上升段、持续段和释放段,提取出相应的数据子集; 对3家检验机构的9 120条制动力曲线进行人工筛选和分析,归纳了制动超前、回踩、增长滞后3种异常特征,给出了相应异常检测算法; 对于较难识别的回踩异常,根据动态规划思想,找出上升段最长连续趋势下降子序列,计算了该子序列占制动力曲线上升段的行程比,并结合经验值来判定该子序列是否异常。研究结果表明:对于维度不大于32的低维制动力数据,通过余弦相似度可聚类制动力曲线的阻滞段、上升段、持续段和释放段; 对于维度大于32的高维数据,因为维数较高,行程比较小,分界点对整个序列相似度影响较小,在这种情况下,必须在考虑相似度的情况下,通过分界点的相对误差进一步约束聚类结果,可以确定制动力曲线的阻滞段、上升段、持续段和释放段; 由于采集的回踩子序列占制动力曲线的行程比为9.8%,大于行程比的经验阈值8.2%,因此,该制动力曲线具有回踩异常,判断结果正确,方法可靠。
The anomaly detection methods of automobile braking curves were studied. The features of the depressing-releasing-depressing anomaly were analyzed. The detection conditions and changing trends of the braking curves were considered. The braking data were clustered based on the cosine similarity and relative error. The segmentation algorithm of the braking curves was established. The braking curves were divided into blocking segment, rising segment, continuous segment, and releasing segment, and the corresponding data subsets were extracted. 9 120 braking force curves from 3 inspection institutions were manually screened and analyzed, the three anomalous features including braking advance, depressing-releasing-depressing and braking growth lag were summarized, and the corresponding anomaly detection algorithms were given. For more difficult to identify the depressing-releasing-depressing anomaly, the subsequence with the longest continuous and descending trend in the ascending segment was searched according to the dynamic programming idea. The distance ratio of the subsequence in the ascending segment was calculated, and compared with the corresponding experiential value to determine whether the subsequence was anomalous or not. Research result shows that the low-dimensional braking data can be clustered into blocking segment, rising segment, continuous segment, and releasing segment according to the cosine similarity only when the dimensions are not higher than 32. For the high-dimensional data, because their dimensions are higher than 32, the demarcation point has less influence on the overall sequence similarity. In this case, the relative error of the demarcation point must be combined with the similarity to constrain the clustering result for determining the blocking segment, rising segment, continuous segment, and releasing segment of the braking curve. Since the distance ratio of the depressing-releasing-depressing subsequence in the ascending segment is 9.8% and larger than the empirical value of 8.2%, the braking curve has