基于双向LSTM的Seq2Seq模型在加油站时序数据异常检测中的应用

加油时序数据包含加油行为的多维信息，但是指定加油站点数据较为稀疏，现有成熟的数据异常检测算法存在挖掘较多假性异常点以及遗漏较多真实异常点的缺陷，并不适用于挖掘加油站时序数据。提出一种基于深度学习的异常检测方法识别加油异常车辆，首先通过自动编码器对加油站点采集到的相关数据进行特征提取，然后采用嵌入双向长短期记忆（Bi-LSTM）的Seq2Seq模型对加油行为进行预测，最后通过比较预测值和原始值来定义异常点的阈值。通过在加油数据集以及信用卡欺诈数据集上的实验验证了该方法的有效性，并且相对于现有方法在加油数据集上均方根误差（RMSE）降低了21.1%，在信用卡欺诈数据集上检测异常的准确率提高了1.4%。因此，提出的模型可以有效应用于加油行为异常的车辆检测，从而提高加油站的管理和运营效率。

Abnormal time series data detection of gas station by Seq2Seq model based on bidirectional long short-term memory

Time series data of gas station contains multi-dimensional information of fueling behavior, but the data of specific gas station are sparse. The existing abnormal data detection algorithms are not suitable for gas station time series data, because many pseudo outliers are mined and many real abnormal points are missed. To solve the problems, an abnormal detection method based on deep learning was proposed to detect vehicles with abnormal fueling. Firstly, feature extraction was performed on data collected from the gas station through an automatic encoder. Then, a deep learning model Seq2Seq with embedding Bidirectional Long Short-Term Memory (Bi-LSTM) was used to predict the fueling behavior. Finally, the threshold of outliers was defined by comparing the predicted value and the original value. The experiments on a fueling dataset and a credit card fraud dataset verify the effectiveness of the proposed method. Compared with the existing methods, the Root Mean Squared Error (RMSE) of the proposed method is decreased by 21.1% on the fueling dataset, and abnormal detection accuracy of the proposed method is improved by 1.4% on the credit card fraud dataset. Therefore, the proposed method can be applied to detect vehicles with abnormal fueling behavior, improving the management and operational efficiency of gas station.