(371p) A Continuous and Smooth Function for Similarity Measure in Quantitative Process Trend Classification

Process-history based approaches for process monitoring and fault diagnosis are popular in the chemical process industries. Process trends described in terms of qualitative features present in the measured data can be used for both manual and automated interpretation of the state of the process. It has been shown in many different ways that the use of several data points from the temporal evolution of measured variables improves the diagnostic resolution. As a fact, an approach known as qualitative trend analysis or its variants exploit this concept to nearly the fullest extent.

During manual interpretation, an operator may use the qualitative shapes such as whether the measured values is increasing or decreasing at a constant rate, increasing or decreasing rate. Effective time-constant related information is also useful to the operator as they can relate it to characteristic time-scale of different unit-operations. During automated interpretation, one would like to use both qualitative and quantitative information contained in the signal. In the context of qualitative or semi-quantitative trend analysis, the existing approaches use discrete functions (such as average of shape-based similarity or their time-weighted average). This effectively makes the mathematical function that captures the similarity between two trends a discontinuous or at least a non-smooth function of the measurements. Since the similarity measure also evolves as the process/signal evolves, such a behavior of similarity measure makes it less reliable for use in automated fault identification of existing faults. More precisely, accurate diagnosis may be delayed. For novel faults, it becomes a source of confusion for the operator during manual diagnosis.

With the above problems in mind, an improved approach for computing the similarity measure between two trends is presented. The proposed mathematical function is continuous and smooth and is able to include the desired features of existing formulations for similarly measure. We apply the methodology on the Tennessee Eastman benchmark process and discuss the effectiveness of addressing the above issues with similarity measure.