One-Class Convolutional Neural Networks for Water-Level Anomaly Detection.

Area of Science:

• Water resource management
• Artificial Intelligence
• Machine Learning

Background:

• Water systems face infrastructure issues like leaks, causing significant losses.
• Monitoring water systems is crucial for operational efficiency and preventing failures.
• Detecting water-level abnormalities is key to ensuring service sustainability and affordability.

Purpose of the Study:

• To investigate a 2D-CNN-based method for detecting water-level abnormalities.
• To explore the use of synthetic temporal data for training anomaly detection models.
• To evaluate the effectiveness of training CNN models from scratch versus using transfer learning for One-Class Classification.

Main Methods:

• Developed a 2D-Convolutional Neural Network (2D-CNN) for time-series anomaly pattern detection.
• Generated synthetic temporal data to augment scarce abnormal data for model training.
• Employed One-Class Classification (OCC) with binary classification using cross-entropy loss.
• Compared training CNNs from scratch with transfer learning from pre-trained models.

Main Results:

• The proposed 2D-CNN method significantly outperformed existing state-of-the-art approaches.
• Using synthetic data as a pseudo-class proved to be a promising strategy for anomaly detection.
• Transfer learning requires careful consideration to avoid underfitting due to model complexity.

Conclusions:

• The 2D-CNN-based approach offers a robust solution for detecting water-level abnormalities.
• Synthetic data generation is a viable and cost-effective method for improving anomaly detection models.
• Further research into transfer learning for OCC is warranted to address potential underfitting issues.