Neural methods in process monitoring, visualization and early fault detection

Abstract

This technical report is based on five our recent articles: ”Self-organizing map based visualization techniques and their assessment”, ”Combining neural methods and knowledge-based methods in accident management”, ”Abnormal process state detection by cluster center point monitoring in BWR nuclear power plant”, “Generated control limits as a basis of operator-friendly process monitoring”, and “Modelling power output at nuclear power plant by neural networks”. Neural methods are applied in process monitoring, visualization and early fault detection. We introduce decision support schemes based on Self-Organizing Map (SOM) combined with other methods. Visualizations based on various data-analysis methods are developed in large Finnish research project many Universities and industrial partners participating. In our subproject the industrial partner providing data into our practical examples is Teollisuuden Voima Oy, Olkiluoto Nuclear power plant. Measurement of the information value is one challenging issue. On long run our research has moved from Accident Management to more Failure Management. One interesting case example introduced is detecting pressure drift of the boiling water reactor by multivariate methods including innovative visualizations. We also present two different neural network approaches for industrial process signal forecasting. Preprosessing suitable input signals and delay analysis are important phases in modelling. Optimized number of delayed input signals and neurons in hidden-layer are found to make a possible prediction of an idle power process signal. Algorithms on input selection and finding the optimal model for one-step-ahead prediction are developed. We introduce a method to detect abnormal process state based on cluster center point monitoring in time. Typical statistical features are extracted, mapped to n-dimensional space, and clustered online for every step. The process signals in the constant time window are classified into two clusters by the K-means method. In addition to monitoring features of the process signals, signal trends and alarm lists, a tool is got that helps in early detection of the pre-stage of a process fault. We also introduce data generated control limits, where alarm balance feature clarifies the monitoring. This helps in early and accurate fault detection.