Resilient Prognostics via Multimodal Alignment: A Self-supervised Computer Vision Approach for Time-Series Analysis on Critical Infrastructure

In the context of Industry 5.0, predictive maintenance requires models that operate effectively in uncertain and partially observable environments. Traditional deep learning methods for time-series forecasting, such as LSTMs and 1D-CNNs, often struggle with limited interpretability and reduced reliability when dealing with sensor faults. To tackle these issues, we present VisuAlign-RUL, a new multimodal method applied to the NASA C-MAPSS dataset to estimate Remaining Useful Life (RUL). We convert multivariate time-series data into visual forms using Gramian Angular Fields (GAF) and a grid-based spatial fusion technique. This approach allows us to use advanced Computer Vision architectures. The main contribution is a Dual-Stream Contrastive Prognostic Network that aligns raw sensor signals and visual representations in a shared latent space through self-supervised learning. Experimental results show that VisuAlign-RUL achieves an RMSE of 14.05 and an S-Score of 289.63. This performance surpasses supervised Transformers on safety metrics while remaining competitive in accuracy. Additionally, Explainable AI analysis using Grad-CAM reveals that the model independently focuses on thermodynamically linked variables, such as compressor pressures and turbine temperatures. This confirms the physical consistency of the learned features without relying on specific domain rules.