Sensor Anomaly Detection And Application

Anomaly detection is to detect and discover those abnormal data patterns not conforming to normal(expected)behavior in observed data.As one of the most active topics in data mining,occupying a leading position and it has a wide range of applications,such as preventing Network Intrusion detection,making sure the safety of industrial production,monitoring equipment failure etc..With the improvement of technology,many sensor devices are installed on industrial equipment,which can monitor working state of the equipment through data output.Considering about wide application requirement,anomaly detection of sensor data has received a lot attention.The main task of anomaly detection is to find and detect rare or unknown data patterns,which are consistent with expected behavior in data patterns.In general,the two kinds of observation samples are very uneven: anomaly samples are much less than target ones,which makes detection great difficulty.With the acceleration of technology,the number of sensors installed on device is more and more.With the accumulation of time,sensor data are increasing,while the number of abnormal data is still rare.When dimension of data is lower,traditional methods can be effective.However,in practical application,we should consider both size of data and efficiency of detection,which is a new challenge.Based on practical application,we have done exploratory researches in anomaly detection.The main work of the paper is:1.We briefly introduced basic knowledge,challenges and wide applications of anomaly detection.We also described in detail many existing methods and analyzed the strengths and weaknesses,as well as compared with related researches in recently.2.We described the basic idea and concept of methods of on-line Least Square Support Vector Machine,extended Kalman filter and regularization sparse encoding.We also analyzed their theoretical basis,and introduced steps of anomaly detection.3.We applied above three methods to thermal power plant for anomaly detection.Comparison between predicted value and actual is used to anomaly detection.Experimental analysis shows that such three methods are effective.