A Unified Framework for Fault and Performance Prediction Using Spatio-Temporal Geometric Features Based on STSFE

This paper proposes a unified deep-learning framework for fault and performance prediction in communication equipment by utilizing spatiotemporal geometric features. The core methodology, Spatio-Temporal Slope Feature Extraction (STSFE), transforms irregular time-series data into slope-, area-, and volume-based representations, capturing both temporal dynamics and spatial correlations. We develop three distinct yet structurally aligned prediction models: (1) passive MUX fault classification, (2) SFP port-level fault detection, and (3) regression-based forecasting of Rx signal degradation. All models employ a multi-branch neural architecture that integrates MLP, CNN, and attention mechanisms, along with customized loss functions designed to enhance sensitivity to tail-zone deviations. To evaluate the generalization capability of the proposed framework, we conduct a comparative analysis using the NASA Battery dataset, which is reshaped via STSFE to emulate industrial signal characteristics. Experimental results demonstrate that our models outperform existing approaches in terms of classification accuracy, mean absolute error, and tail prediction performance. This research provides a flexible and robust methodology for predictive maintenance across diverse time-series domains in industrial communication networks.