Contribution plots based fault diagnosis of a multiphase flow facility with PCA-enhancec canonical variate analysis

• Published in: 2017 23rd International Conference on Automation and Computing (ICAC) pp. 1 - 6
• Main Authors: Ruomu Tan, Yi Cao
• Format: Conference Proceeding
• Language: English
• Published: Chinese Automation and Computing Society in the UK - CACSUK 01.09.2017
• Subjects: contribution plots; Covariance matrices; Data models; Fault detection; fault detection and diagnosis; Fault diagnosis; Feature extraction; Monitoring; multivariate process monitoring; Principal component analysis
• DOI: 10.23919/IConAC.2017.8081992

Process monitoring plays a vital role in order to sustain optimal operation and maintenance of the plant in process industry. As an essential stage in process monitoring, datadriven fault detection and diagnosis techniques have evolved quickly owing to the prosperity of multivariate feature extraction methods. In addition to the application of basic feature extraction methods, hybrid algorithms combining different methods have also been invented for better monitoring performance. In the meantime, little study has been done towards the fault diagnosis techniques under this 2-stage feature extraction framework. To deal with complex faults which will have impact on multiple process variables and the relationships among them, the Principal Component Analysis (PCA) enhanced Canonical Variate Analysis (CVA) based fault detection and diagnosis algorithm is investigated in this paper. PCA is used to pre-process the raw measurements and extracts the principal components as better indicators of process condition; CVA is conducted sequentially to further project the principal components to canonical variate space and the detection statistics are calculated based on these canonical variates. When a fault has been detected, the contributions of original process variables in monitoring statistics are derived to identify influential variables and locate the fault. To validate, along with other multivariate statistical monitoring techniques, this PCA-enhanced CVA algorithm is applied to a benchmark data set collected from an industrial scale multiphase flow facility in Cranfield University for performance evaluation.