| In an environment where most process maneuvers are automated, algorithms to detect and classify abnormal trends in process measurements are of critical importance. The petrochemical industry in the United States loses billions of dollars annually due to improper abnormal situation management, and a staggering one in 16 plant accidents results in a fatality. Hence, Statistical Process Control and Monitoring (SPC) has been an active area of research for many decades and a variety of statistical and machine learning-based methods have been developed. However, most existing methods for process monitoring learn the signal characteristics at a fixed scale, and are best for detecting changes at that single scale. In contrast, data from most industrial processes are inherently multiscale in nature due to events occurring with different localization in time, space, and frequency. Unfortunately, existing techniques are unable to adapt automatically to the scale of these features. Many existing methods also require the measurements to be uncorrelated, whereas, in practice, autocorrelated measurements are very common in industrial processes.; In this work, we have investigated the use of multiscale techniques to improve upon these shortcomings of existing single-scale approaches. Because of fundamental functional relationships such as process chemistry, energy and mass balances, measurements in multivariate processes are correlated. Our approach learns these correlations and clustering behaviors in the wavelet space using machine learning methods such as Adaptive Resonance Theory (ART-2) and Principal Component Analysis (PCA), resulting in higher detection accuracy coupled with noise reduction. The performance of our method, named Multi-Scale Statistical Process Control and Monitoring (MSSPC), is compared with existing methods based on the average detection delays for detecting shifts of different sizes. Our ART-2 based MSSPC detector is currently deployed in a large scale petrochemical plant to detect process anomalies in real time by incrementally learning normal process operation in the wavelet domain. Several case studies for the detection of real process malfunctions, including the comparison with the performance of human operators, are also presented in this work. These results indicate that MSSPC is a good method for monitoring of measurements with unknown and different types of changes. |
