(339u) A Data-Driven Process Monitoring and Fault Diagnosis Method for Multi-State Industrial Process Operations

Data-driven methods for process monitoring and fault diagnosis has been an active field in recent years. The benchmark Tennessee Eastman process is usually used to evaluate the performance of proposed methods in literature. However, there are still many challenges remaining in real industrial cases, especially in multi-state process operations, as industrial operation is frequently adjusted according to the market situation and administrative regulations, and most data-based monitoring method is developed for a given steady operating condition. Once a new operating condition emerges, it is more likely considered as an abnormal deviation, which could result in a large number of false alarms and therefore, significantly limit the application of data-driven process monitoring methods. Tennessee Eastman process is a well-developed benchmark process, and has been widely adopted in literature to evaluate the performance of process monitoring and fault diagnosis. Most fault diagnosis methods applied to the Tennessee Eastman process are supervised classification methods based on the data sets with clear labels, but abnormal deviations could be much more diversified and impacted by random disturbance in real industrial process.

In this work, an unsupervised data-driven process monitoring and fault diagnosis method based on mutual information is proposed. Mutual information value is a measure of the correlation between a pair of variables, which characterizes the correlation from the respect of information theory, in other words, calculates the relative deviation between a pair of variables respect to given operating conditions. Once a system reaches a steady state, the mutual information between two variables will only depend on their randomness. For a specific process, there is no significant difference on randomness once a steady state is reached, as it will be only decided by signal noise from the system. Therefore, the mutual information should display a unique statistical feature for most steady-states of a system. When an abnormal deviation occurs, corresponding changes in mutual information statistics can be observed, by which process monitoring can be implemented. In order to determine the causal relation between a pair of variables, a time lag parameter can be introduced for fault propagation path identification. The proposed method is applied to both Tennessee Eastman process and a practical industrial process, and the results show a good performance on process monitoring and root cause recognition.