| Nowadays,the rapid development of sensor technology has widely used in various fields.For example,an industry is usually consisted of a large number of devices,running together with each other to produce productions for human society.These devices are deployed with tens of thousands of sensors to monitor the running status in real-time.Each sensor will collect sensor data reflecting an attribute of the device with a fixed frequency and transmit the sensor data to the data center.The large-scale,high-speed sensor data stream hides important information that can reflect the running status of a device.In fact,when a device runs in an abnormal way,the sensor data reflecting the related attributes will deviate the normal behavior of most data.These sensor data are called as anomalies.Anomaly detection from sensor data is one of the main-stream methods to detect device fault.At present,there are many works on anomaly detection to abnormal data points.The occurrence of abnormal data points is often instantaneous.But in a production environment,a device fault usually lasts a time period.Therefore,outlier detection cannot meet the requirements of real applications well.On the other hand,most of the traditional methods are proposed to deal with static data sets,which cannot handle sensor data streams in an industrial environment.Therefore,this paper presents an approach to detect the abnormal sequences from large-scale sensor data streams.This paper mainly has the following contributions.Firstly,we refine the formal definition of abnormal sequence based on sensor data streams.Secondly,this paper puts forward an approach to detect abnormal sequences from large-scale sensor data streams.We also do a lot of experiments based on a real dataset in a coal power plant to verify the effectiveness and efficiency of our approach.Thirdly,we design and implement a device fault detection system for a coal power plant based on our abnormal sequence detection approach. |
PDF Full Text Request