Skip to Main Content
To solve problems in data-driven fault prognostic study such as prediction uncertainty management, multiple fault features and on-line prognostics, an algorithm based on multivariate relevance vector machine (MRVM) is presented. It extends the existing time series iterative multi-step prediction to the application with multiple fault features by matrix partitioning technique. For on-line application, it divides prognostics into a three-phase process to handle differently, namely on-line learning, short term and long term prediction, which properly satisfies the accuracy and execution time requirements at the same time. In on-line learning phase, the algorithm greatly reduces the time cost by means of sliding window technique. For on-line prediction, it comes up with a solution by increasing useful information for prediction decision making. In short term prediction phase, it adopts particle filter technique, which updates prediction results through the way of introducing new observations. It also employs fusion technique which results a weight sum of several previous predictions weighted by a certain forget factor. In long term prediction phase, MRVM is retrained by adding the short term prediction results to training data set. A simulation experiment is adopted which demonstrate the effectiveness of the proposed algorithm.