(365h) System-Wide Anomaly Detection By Single Value Using Mscred in MEG Regeneration Pilot Plant

Process plant data have many different sensor measurements such as temperature, pressure, level. In general these inputs are correlated with each other and also correlated with its own past time series data. However, in a field of anomaly detection of a process, cannot put these all in concern. Here we suggest system wide spatio-temporal anomaly detection using the modified MSCRED(Multi Scale Convolutional Recurrent Encoder -Decoder) framework. MSCRED is data driven, especially using deep learning models. This framework detects anomalies resulting in anomaly scores, which is modified as univariate time series data, which allows real-time monitoring of the system. Anomaly score increases when many sensor’s data are out of normal temporal behavior and decreases when sensors show normal behavior. Anomaly scores were clustered with time series correlation coefficients to be classified by its fault cases. Data used for validating the algorithm was gathered by a MEG regeneration pilot plant with 42 sensor data, with 5 normal cases, 4 issue cases and 1 start-up case. This framework has been trained with different subset of normal data and compared to find optimal training set, and anomaly detection was carried out for all 4 issue cases and start up case for many window sizes, each resulting in a time series anomaly score. These results were clearly clustered and classified for each fault case, showing that the anomaly score can classify fault cases and indicate the faults. In this study we suggest the single time series variable obtained from a data-driven framework to monitor the plant wide sensor data to detect the system anomalies and identify the fault characteristics.