New conference contributions at the "International Joint Conference on Neural Networks (IJCNN) 2024"

"Multi-Task Representation Learning with Temporal Attention for Zero-Shot Time Series Anomaly Detection"

• Authors: Chandana Priya Nivarthi, Zhixin Huang, Christian Gruhl, Bernhard Sick

• Abstract: This paper introduces MTL-LATAM, a Multi-task Learning framework for time series anomaly detection. It combines LSTM autoencoder with temporal attention, offering adaptability, reduced runtime, and better performance in zero-shot scenarios. Results show MTL-LATAM's effectiveness across various datasets, achieving 95% and 97% task synergy, outperforming single-task models.

Spatial-Temporal Attention Graph Neural Network with Uncertainty Estimation for Remaining Useful Life Prediction

• Authors: Zhixin Huang, Yujiang He, Chandana Priya Nivarthi, Christian Gruhl, Bernhard Sick

• Abstract: This paper focuses on improving the accuracy of predicting remaining useful life in complex industrial systems. It proposes a new model called USTAGNN, which combines graph neural networks and temporal convolutional networks. USTAGNN addresses issues like uncertainty estimation and sensor graph construction. Results show it outperforms existing methods, achieving up to a 35.9% improvement in prediction score on the C-MAPSS dataset.