Anomaly Detection System by Statistical Analysis of Incorrect Signal About Monitoring System of Railway(Machine Elements, Design and Manufacturing)
Naoto Makino, Atsushi IWASAKI, Shinsuke Sakai, Junji Sugimoto, Hirotatsu Yamazaki · TRANSACTIONS OF THE JAPAN SOCIETY OF MECHANICAL ENGINEERS Series C · 2010
In this paper, the method of extracting true abnormal information from the statistics analysis of the occurrence probability of a binary output of an abnormality diagnosis system is examined. In general, since a fluctuation occurs in a sensor measuring quantity in the normal state by various factors, such as a noise and environment, an incorrect diagnostics arises in the abnormality diagnosis system which use threshold value. Therefore, abnormal detection will detect more than the actually caused abnormalities significantly. In this case, the judgement to all the abnormal information is required, the treatment to an individual phenomenon becomes shortly, and it causes hindrance of accurate judgment. Such a situation causes a delay of the detection of the true accident and this delay also causes a delay of a maintenance. And at the worst case, it may also causes a further accident. Therefore, for quick action and the maintenance, reduction of the incorrect information is desired. Then, in this paper, the method of extracting true abnormal information from the statistics analysis of the occurrence probability of a binary output of an abnormality diagnosis system is examined. By the abnormality diagnosis system, exis ence or not of an abnormality is shown by the binary parameter (1 and 0, for example.). In this research, a diagnostics is performed from parameters, such as duration and interval of the alarm. By using these parameters, threshold value between the normal and the abnormal condition is decided from the approximation of occurrence probability of the parameters using exponential distribution. Finally, this method was applied to the actual rail crossing abnormality diagnosis system, and it succeeded in reducing almost 98% of incorrect information using the duration of the alarm and reducing almost 94% using the interval of the alarm.