Graph neural networks have advanced unsupervised fault detection in industrial multivariate time series by modeling inter-sensor dependencies. However, existing methods operate exclusively in the time domain, relying on a single dependency graph that neither exploits complementary frequency-domain signatures nor captures structural disagreement across representation spaces. This paper proposes DualSTGF, an information fusion framework that jointly learns temporal and spectral dependency graphs for incipient fault detection. Jensen-Shannon divergence between dual-view attention distributions serves as both a structural training regularizer and an inference-time anomaly indicator.

This paper presents a dual-graph neural network architecture for robust structural damage detection. The approach leverages both spatial and spectral graph representations to detect anomalies in structural health monitoring data, demonstrating improved robustness to noise and sensor failures compared to existing methods.

This work evaluates the robustness of counterfactual explanation methods designed for tree-based ensemble models. We propose metrics and benchmarks to assess how stable and actionable generated counterfactuals remain under perturbations and model updates.

Counterfactual explanations are an effective method of explaining the decisions made by machine learning models to end-users. However, it is crucial to ensure that certain qualities of counterfactuals such as plausibility and proximity are satisfied. This paper proposes a novel objective function using conic optimization to generate high-quality counterfactual explanations that balance plausibility and proximity.

An expert evaluation system for automated pothole defect detection using image analysis and machine learning. The system integrates domain expert knowledge with deep learning models to provide reliable assessment of road surface conditions.

This study presents a precision leak detection framework for supermarket CO2 refrigeration systems using categorical gradient boosting. The method leverages multi-sensor time-series data to detect refrigerant leaks early, reducing environmental impact and maintenance costs.