Comprehensive fault detection and isolation method applied to a recovery boiler

Early and accurate detection and isolation of industrial process faults are crucial to avoiding abnormal situations that cause productivity losses. Principal component analysis and reconstruction-based contribution (PCA-RBC) is a popular method used for such tasks. Unfortunately, this method does not guarantee correct fault isolation in cases where the faulty variables contribute little or do not contribute at all to the main principal components of the PCA model. This is the case, for example, of some pollutant emission levels that do not affect the global performance of a biomass boiler, but that should be maintained below certain thresholds. This paper proposes to adapt the PCA-RBC method to cope with such limitations. The idea is first to classify the data onto normal and abnormal conditions according to a selected parameter threshold, and then to build a PCA model using the normal dataset. The RBC approach is applied on the abnormal dataset to identify the variables that mostly contribute to the faulty situations. The proposed method is successfully demonstrated using real data from an industrial case. It is noted that an attempt to develop an accurate predictive model of the selected parameter using projection to latent structures (PLS) was unsuccessful.