Performance evaluation of data stream-oriented approaches applied to fault detection of industrial processes
fault detection, data stream, RDE, TEDA, R-PCA.
In order to increase product quality and process performance, the degree of automation has grown significantly in industries. As a result, systems are increasingly complex and are accompanied by problems that are difficult to solve due to the high dimensionality of these systems and the large amount of information flow, as well as the randomness of faults and defects. An unexpected failure can lead to operational risks, so the importance of detecting and locating the fault, especially when the industrial plant is still operating in a controllable region and it is possible to act to bring the process back to normal, safe and operational. Thus, it is desirable for the fault detection system to provide fast and reliable responses with a computational effort appropriate for real-time processing, even though it requires handling large amounts of data. In this context, data stream-oriented algorithms to outlier detection may be promising candidates for fault detection of industrial process, because they work with sequences of temporarily ordered samples. In addition, they handle well with large amount of data because they are recursive and online algorithms that do not need to store past samples. Thus, in this dissertation two algorithms of this class are analyzed, named TEDA (Typicality and Eccentricity Data Analytics) and RDE (Recursive Density Estimation), when applied to fault detection of industrial processes. Their performances are compared to R-PCA (Recursive Principal Component Analysis) algorithm. The classic Tennessee Eastman Process benchmark was used as case study to evaluate these algorithms.