Online method for Learning Process States in Cyber-Physical Production Systems

Cyber-Physical Production Systems (CPPS) require robust techniques for automation, monitoring, and diagnosis. Model-based approach provides such techniques based on capturing spatio-temporal features of system’s behavior in a dynamic process model. Typically, the model must be learned from system observations, however, in certain CPPS cases constructing a robust model remains a challenging task while existing models perform poorly. In this study, the problem of state-based dynamic modelling in such CPPS is addressed. The hypothesis is that the main limitation of existing models comes from low informativeness of learned process states. To remove this limitation, a method for learning states from dynamic observations is proposed that employs incremental model-based clustering (IMCF). This method is evaluated on CPPS data where existing modeling techniques have proven ineffective. It is argued that obtained states have higher informativeness, comparing to existing research, while the method has higher computational effectiveness and can be applied for stream processing. The latter extends potential applications of state-based modelling in major tasks of CPPS.