| dc.description.abstract | In model-based fault detection, processed input and output time-series data are used to generate models and then perform on-line predictions. Many practical considerations in model-based fault detection systems rule out most of the available approaches in generating on-line predictive models. One approach that satisfies the needs of on-line predictive model generation is the piece-wise linear robust regression (PWLRR) method. The PWLRR requires too much data for initial training, it has limited ability to perform accurate predictions beyond the initial training region (extrapolation), it does not perform accurate predictions without significant follow-on or on-line training, and it constantly runs into blind spots losing track of the asset condition. The performance of a fault detection system based on the PWLRR suffers in many instances, mostly in the form of delayed detection, missed faults and/or false alarms.
This work presents an alternative learning method to model generation that overcomes many of the existing disadvantages of the PWLRR approach. Gaussian process (GP) based regression, which exhibits good approximation properties with minimal training data points and has very good extrapolation properties, is selected as an alternate learning method in the fault detection system. The fraction of blind spots, number of data points in the training and validation sets, extent of extrapolation, average prediction error, true negatives (missed faults), false positives (false alarms), and detection time are all considered performance indicators when testing and comparing the model generation methods. Five (5) cases with artificial data and ten (10) cases with real world data from both staged experiments in the laboratory and fielded production sites are used to benchmark the performance of the GP approach, and compare it against the PWLRR approach.
Empirical research results comparing GP to PWLRR demonstrate that for comparable training levels, less data is required to train a GP model and with fewer blind spots. Extrapolation by the GP model improves significantly. Among the real test cases with faults, GP results in 100% detection rate, while PWLRR results in 50 % detection rate. For the test cases where both approaches result in true positives, detection time is improved by roughly 2.5 times when using the GP. In test cases without faults, GP results in no false positives, while PWLRR results in three (3) false positives. The proposed method for training based on GP can generate models with less computational resources than the PWLRR and requires less human intervention. Further testing is needed to verify the performance of the proposed approach on data sets with a wider variety of statistical properties. | en |
