Robust real-time algorithms for processing data from oil and gas facilities

2018-09-22Advancing sensor and data gathering technology has resulted in a substantial increase in the amount and frequency of process data collected, creating a valuable opportunity for data-driven modelling techniques in process monitoring. Since process facilities are large interconnected systems, data collected may also be strongly correlated. Therefore, this dissertation approaches the problem of data-driven modelling of process data from two perspectives, first for univariate, uncorrelated data sets, and second for multivariate, correlated data sets. Common methods for monitoring and analysis of univariate data include the use of cumulative sums, moving average charts, interpolation and other simple statistical methods. For correlated, multivariate data, methods such as principal component analysis, partial least squares, and contribution plots are conventionally used. ❧ In this dissertation, the application of online wavelet transforms (OWT) to several aspects of real-time analysis of univariate data was considered, namely, fault detection, data cleansing and data compression. In the first application, the OWT is used to build a model and detect faults in an uncorrelated tag (variable) which cannot use traditional data-driven modelling that relies on underlying relationships between variables. The OWT is successfully able to detect and identify various common sensor faults including missing and frozen values, linear drift and spikes. Recursive least squares (RLS) is then used to predict replacement values and is a significant improvement on current standards, particularly with highly variable data. The new algorithm also updates statistics with each new value, so decision making is accurate but does not require repeated calculations on the whole data set. This application directly affects real-time decision making as faults can be addressed immediately, and any predictive models built from the data will be representative of real operation. The OWT is also adapted to create a novel data compression algorithm to economically store large quantities of data in real-time. The industry standard swinging door algorithm requires extensive manual input, in addition to other drawbacks. The new method is an online filter that is completely automated, with calculations unique to each tag. ❧ There are numerous methods for building data-driven models for multivariate data sets, including the commonly used principal component analysis (PCA) and its extension, dynamic PCA (DPCA). However, there are several drawbacks to PCA and DPCA, including the inability to handle any missing or corrupted data. Two new algorithms are developed in this work by incorporating a robust PCA method called principal component pursuit (PCP) into DPCA to create a robust DPCA method. PCP uses a convex minimization problem to recover a fault-free matrix even when there are gross sparse errors present, and is therefore more robust to outliers than other methods. The PCP method is incorporated into DPCA in two ways. First, PCP and DPCA are combined sequentially where PCP is used to generate a ‘fault free’ training data set as a preprocessing step. In the second method, PCP is incorporated into DPCA to create an integrated robust DPCA method. The two methods are compared to each other and to traditional DPCA in two case studies using real process data. Both the methods present a significant improvement for fault detection over DPCA models trained on faulty data. Extension of this robust DPCA method to a recursive method is then discussed. In the final part of the dissertation, the extension of these methods from sensor fault detection to process fault detection is discussed and simulated data from the Tennessee-Eastman process is used to show the method’s effectiveness