Utilizing Timed Petri Nets to Guide Data-Driven Fault Diagnosis of PLC-Timed Event Systems

Highly synchronized timed event systems managed by programmable logic controllers are ubiquitous in manufacturing. Discrete event system methods such as finite state machines and Petri Nets have been useful for diagnosing known faults in such systems. However, existing methods do not scale as well for practical applications where faulty behavior is not fully known and event timing may vary. With the advent of big data, a new methodology is presented that strengthens fault diagnosis capabilities utilizing machine learning and Timed Petri Nets. The hybrid approach consists of building a Timed Petri Net model of the normal process to help select discriminating features based on timed event sequences. This work focuses on modeling Timed Petri Nets to serve as knowledge representations of the nominal process behavior to guide data-driven fault diagnosis of timed event systems. A methodology based on identification of observable events is introduced, consisting of three different and complementary heuristics including demarcating periods, modeling concurrencies, and imposing synchronizations. Then, using features selected by using the Timed Petri Net, a nonlinear SVM is implemented for multiclass fault classification. The proposed framework is applied to an existing semiconductor manufacturing process with events timed via binary programmable logic controller signals. With over 97% validation accuracy achieved for fault diagnosis, the hybrid modeling approach shows promise for smart manufacturing applications.