面向非规范化数据源的动物体温异常识别方法

在动物体温异常识别中，红外测温等方式容易产生系统偏差使得判断结果不可靠。基于深度学习的方法在不同测温设备上的鲁棒性与泛化性能较差，且难以应用于数据量少、随机性强、标准不一致等非规范化的测温场景。因此，本文提出了一种面向非规范化数据源的动物体温异常识别方法，通过衡量体温时序数据间的相似度即可完成异常识别。针对常用的相似性度量算法在序列匹配、序列间距度量上效果不佳的问题，提出了一种改进的动态时间规整算法(Improved dynamic time warping, iDTW)。在点间度量方式上，综合欧氏距离和一阶导数，改善了序列过度对齐问题。使用序列交并比表示序列整体特征，提升了序列间距度量效果。针对不等长序列及过长序列的异常检测问题，提出了基于滑动窗口和序列等分的异常检测方法。以较短序列为滑动窗口遍历较长序列得到一组序列间距，根据训练和检测的不同阶段分别选择其中的最小值或最大值作为相似度衡量结果，以解决不等长序列匹配问题。将过长的样本数据序列等分为多个子序列，取子序列的间距和为样本间距，以解决过长序列导致的正常样本间距过大和异常漏检问题。在公开数据集UCR上的实验分析表明，相比于欧氏距离...

Anomaly Recognition for Animal Body Temperature Based on Non-standardized Data Source

In the anomaly recognition of animal body temperature, methods such as infrared temperature measurement are prone to system bias, making the results unreliable. Deep learning based anomaly detection algorithms has poor robustness and generalization performance on different temperature measurement devices, and is difficult to apply to non-standardized temperature measurement scenarios with low data volume, strong randomness, and inconsistent standards. Therefore, a method of animal body temperature anomaly recognition for non-normalized data sources was proposed. The abnormal animal body temperature detection could be completed by measuring the similarity between body temperature time series data. An improved dynamic time warping (iDTW) algorithm was proposed to solve the problem that the commonly used similarity measurement algorithms were not effective in sequence matching and sequence distance measurement. The Euclidean distance and the first derivative were integrated in the measurement between data points, which effectively solved the problem of sequence over-alignment. The sequence intersection ratio was used to represent the overall characteristics of the sequence, which improved the effect of sequence distance measurement. Aiming at the problem of anomaly detection of unequal length sequence based on similarity measure, an anomaly detection method based on sliding window and sequence equal division was proposed. The shorter sequence was used as the sliding window to traverse the longer sequence to obtain a set of sequence distance. According to the different stages of training and detection, the maximum or the minimum value was selected as the similarity measurement result to solve the problem of unequal length sequence matching. To solve the problem of excessive distance between normal samples and undetected anomaly caused by the long sequence, the long data sequence was equally divided into multiple sub-sequences, and the sum of the sub-sequence distance would be taken as the final similarity measurement result. Experimental results on the public dataset UCR showed that the iDTW algorithm outperformed Euclidean distance, dynamic time warping, derivative dynamic time warping and weighted dynamic time warping by an average of 6.0, 3.0, 5.2 and 2.5 percentage points on 10 time series datasets, respectively. Compared with the classical anomaly detection algorithms, the F1 score of the anomaly detection method based on sliding window and sequence equal division on three animal body temperature datasets were increased obviously.