Anomaly Detection in Discrete Manufacturing Systems using Event Relationship Tables

    •  Laftchiev, E., Sun, X., Dau, H.-A., Nikovski, D.N., "Anomaly Detection in Discrete Manufacturing Systems using Event Relationship Tables", International Workshop on Principle of Diagnosis, August 2018.
      BibTeX TR2018-128 PDF
      • @inproceedings{Laftchiev2018aug,
      • author = {Laftchiev, Emil and Sun, Xinmaio and Dau, Hoang-Anh and Nikovski, Daniel N.},
      • title = {Anomaly Detection in Discrete Manufacturing Systems using Event Relationship Tables},
      • booktitle = {International Workshop on Principle of Diagnosis},
      • year = 2018,
      • month = aug,
      • url = {}
      • }
  • MERL Contact:
  • Research Areas:

    Artificial Intelligence, Data Analytics, Optimization


Anomalies in discrete manufacturing processes (DMPs) can result in reduced product quality, production delays, and physical danger to employees. It is difficult to detect anomalies in DMPs, because sequencing devices such as programmable logic controllers (PLCs) usually do not allow a process engineer to easily determine which sequences of operations are observed and checking against each sequence becomes computationally difficult. This paper proposes a new anomaly detection approach for discrete manufacturing systems. The approach models the normal behavior of the DMP from the PLC output as an event relationship tables. These models are then used to determine if new sequences of PLC outputs could be generated by the system. Outputs that do not fit the learned model are labeled anomalous. This method is tested in simulation for DMPs that contain concurrent sub-process with unique or repeated events. The results are compared to a baseline method proposed in prior publications. Experiments show that the proposed algorithms are capable of achieving a higher F-score with less than 10% of the data required by the baseline method. Furthermore, this modeling approach has a linear space complexity in the length of the observed event sequences, as compared to polynomial complexity of prior work.