Support Vector Machines Combined to Observers for Fault Diagnosis in Chemical Reactors

A hybrid data/model-based approach is proposed for fault detection and isolation for chemical reactions in jacketed stirred vessels. Using data-based methods in high nonlinear systems requires training data to include a wide range of varying operations to ensure correct fault isolation. If such data is not available, a model-based approach can be used to enhance fault isolation. But, observers require a relatively precise process model, which is also not always available. In this work, we propose to combine an observer with statistical data-based methods (support vector machines, SVM) for fault detection in order to avoid at the time precise process modelling (necessary for model-based approach) and great number of training data (necessary for data-based approach). An interesting case study that falls in this category is a chemical stirred tank reactor, with high nonlinear reactions. Therefore, a simplified process model is used as a starting point to develop an observer for fault isolation. The used process model is corrected using information from SVM when no fault is detected. The methodology is validated experimentally in lab-scale and pilot-scale polymerisation reactors. For processes with linear dynamics, model-free SVM classification was found sufficient to detect and isolate sensor and actuator faults.