Detection of Abnormal Oil Data Based on Feature Selection

Abstract

To achieve effective oil management, it is critical to disclose the laws of oil supply, consumption, and natural loss through data analysis. However, the accuracy of data analysis is often suppressed by the mistakes and irrelevance of the input data, which are inevitable due to the large size and diversity of the data collected from the oil depots. To solve the problem, this paper proposes an abnormal oil data detection approach based on feature selection (AODDFS). In the AODDFS, the format of the input data was preprocessed to satisfy the requirements of feature selection; the fisher score was then employed to compute the relevance of each entry with normal features; finally, the abnormal entries were located based on the relevance values. Then, the AODDFS results were analyzed with boxplot and standard deviation. Finally, the AODDFS was verified through a case study on the data collected from several large oil depots. The results show that the AODDFS can effectively detect abnormal oil data with a precision of 85.00% and a recall of 80.94%.