Unsupervised deep anomaly detection in a recirculating aquaculture system
| dc.contributor.author | Robinson, William | |
| dc.contributor.examiningcommittee | McNeill, Dean (Electrical and Computer Engineering) McLeod, Robert (Electrical and Computer Engineering) | en_US |
| dc.contributor.supervisor | Shafai, Cyrus (Electrical and Computer Engineering) | en_US |
| dc.date.accessioned | 2020-03-23T21:42:55Z | |
| dc.date.available | 2020-03-23T21:42:55Z | |
| dc.date.copyright | 2020-03-20 | |
| dc.date.issued | 2020 | en_US |
| dc.date.submitted | 2020-03-20T20:33:07Z | en_US |
| dc.date.submitted | 2020-03-20T21:26:25Z | en_US |
| dc.degree.discipline | Electrical and Computer Engineering | en_US |
| dc.degree.level | Master of Science (M.Sc.) | en_US |
| dc.description.abstract | An unsupervised deep anomaly detection system is implemented to augment the water quality monitoring system used at a recirculating aquaculture system (RAS) facility. Its purpose is to increase the system’s anomaly detection capabilities by improving its accuracy and decreasing the timeframe in which anomalies can be detected. Quick and precise detection of abnormalities leads to earlier action to reduce mortalities within the fish population, or prevent them altogether. The machine learning model introduced in this work, given the name aMSCRED or adaptive Multi-Scale Convolutional Recurrent Encoder-Decoder, is an expansion of the MSCRED model featured in previous work by Zhang et al.[1] This model is a spatio-temporal network (STN) composed of stacked CNNs and RNNs, structured in an autoencoder architecture. This configuration is capable of learning what characterizes normal behaviour within a multivariate timeseries dataset, which can thereafter be leveraged to detect abnormal behaviour, which may indicate a problem in the system. Using data obtained from the monitoring system at the RAS facility, aMSCRED is able to outperform its predecessor in terms of anomaly detection performance (measured in terms of Recall, Precision and F1 score). Recall scores of up to 97% were achieved, as well as F1 scores of up to 94%. It also outperforms its predecessor in root cause identification (RCI), achieving accurate prediction rates of ∼ 70%, compared to ∼ 50% using the model from Zhang et al. The improved results are made possible due to modifications which enable the model to adaptively select, on a per-dataset basis, different signature matrix generating strategies, model structure parameters, anomaly scoring methodologies, and root cause scoring methodologies. | en_US |
| dc.description.note | May 2020 | en_US |
| dc.identifier.uri | http://hdl.handle.net/1993/34579 | |
| dc.language.iso | eng | en_US |
| dc.rights | open access | en_US |
| dc.subject | Machine learning | en_US |
| dc.subject | Deep anomaly detection | en_US |
| dc.subject | Aquaculture monitoring system | en_US |
| dc.subject | aMSCRED | en_US |
| dc.title | Unsupervised deep anomaly detection in a recirculating aquaculture system | en_US |
| dc.type | master thesis | en_US |