Anomalous sensing data recovery with mutual information and Relevance Vector Machine

Sensing data are the basic input for system prognostics and health management. However, sensing data may become anomalous due to the sensor fault and failure, malfunction of connectors, etc. The sensing anomaly data could bring wrong prognostics result and unreasonable maintenance schedule. The problem becomes more challenging when the sensing data are sparse, i.e., only a few sensors are available for utilization. To deal this problem, the sensing anomaly data recovery approach is proposed in this article. The relationship among sensing data is analyzed by mutual information to select the maximal relevant training data for anomaly data recovery. One dimensional training data to recover anomaly data is achieved by Relevance Vector machine which has the feature of sparsity. The effectiveness of the proposed approach is evaluated by utilizing the Prognostics and Health Management 2008 Challenge data.