Mining Probability Frequent Patterns To Recover Uncertain RFID Data Stream

RFID is a non-contact and automatic identification technology and has been widely used in many fields, owing to the properties of small size, low-cost, non-contact, automatic identification and so on. Because of the susceptible to the external environment and the instability of RF signals, the original data detected by RFID readers is always unreliable, incomplete and noise such as missing-reads and cross-reads. Therefore, how to recovery these data from RFID data steams has become an urgent and important topic in RFID middleware area.Firstly, the paper analyzes the characteristics of RFID data streams in details, and indicates that RFID data is simple, relevance, temporal, spatial and uncertain. Then, it introduces and analyses related technologies including window models of data streams, mining technologies on data streams, recovery methods on RFID data streams, and so on.In order to recover the uncertain RFID data more accurately, the paper proposes a new RFID data streams recovery method which takes several various factors into consideration comprehensively. The method mainly includes three steps:The first step is to make use of an uncertain data streams mining method based on PFP-tree to mine probability frequent patterns. By setting a transaction-index table and an item-index table, the method can find the frequent patterns quickly and efficiently. The method uses a damping window model to distinguish the contributions of the data generated by new transactions and the data by historic transactions, to guarantees high recall rate of frequent patterns.The second step is to determine the maximum relevance item using the mining results from PFP-tree, and the maximal similar trajectories item through computing similar trajectories for a specific item. Based on these information, a discriminate multivariate statistical analysis method—RR method is used to recovery RFID data, using a vector including the item's location information in current window and the information in previous window, the location information of the maximal relevance item, and the information of the maximal similar trajectories item. On the basis of these factors proposed, the true place of the tagged item can be found by computing the Euclidean distance of the vector.The third step is to use a Bayesian-based correction method, using the ratios of missing-reads and cross-reads, to correct the probability of items in the reading ranges of readers, to provide more accurate posterior probability for next windows. Finally, by experiment analysis and comparison, it has been proved that the performance of PFP-tree is better than that of SUF-growth, and RR recovery method can get higher recovery accurate rate than SIS method.