This is an approach to anomaly detection using 1D CNN layers as the backbone of a variational autoencoder on a wastewater system.
How is it great compared to the related works?
CNN's have shown exceptional performance on image classification tasks. For example, previous techniques like LSTM networks and RNN's known to handle time-series were outperformed by Temporal CNN's.
What are the key technical differentiators?
The approach uses CNN's to represent time dynamics as stacked layers of deep autoencoders for unsupervised anomaly detection tasks. The effectiveness of this approach is illustrated over a case study using time-series data from insewer process monitoring.
How did they validate the advantages?
By trying to minimise the reconstruction error the approach was validated on unseen testing data and achieved 35% accuracy.
Are there any discussions around the proposal?
The approach did not prove very useful in part due to limited training data. Many false positives were characteristic of the approach as a result. This method needs further investigation and more work is needed.
Stefania Russo, Andy Disch, Frank Blumensaat, Kris Villez
Keywords: [anomaly detection]; machine learning; wastewater systems
Paper PDF
This is an approach to anomaly detection using 1D CNN layers as the backbone of a variational autoencoder on a wastewater system.
CNN's have shown exceptional performance on image classification tasks. For example, previous techniques like LSTM networks and RNN's known to handle time-series were outperformed by Temporal CNN's.
The approach uses CNN's to represent time dynamics as stacked layers of deep autoencoders for unsupervised anomaly detection tasks. The effectiveness of this approach is illustrated over a case study using time-series data from insewer process monitoring.
By trying to minimise the reconstruction error the approach was validated on unseen testing data and achieved 35% accuracy.
The approach did not prove very useful in part due to limited training data. Many false positives were characteristic of the approach as a result. This method needs further investigation and more work is needed.
Anomaly detection using Reinforcement Learning.