How Physics-Informed AI Transforms Real-Time Risk Assessment in Safety Systems

Global, Source:

In the high-stakes world of safety-critical sensing and decision-support systems, the ability to accurately assess risk in real-time is paramount. Modern remote-sensing pipelines, which integrate radar measurements, distributed tracking, and auxiliary signals, face significant challenges such as spatial registration errors, asynchronous fusion, clutter, and latency-sensitive decision windows. A new framework titled TA-DE-ELM addresses these complexities by combining expert-rule-guided synthetic data generation with a Two-Stage Adaptive Differential Evolution algorithm optimized for Extreme Learning Machines (ELM). This approach offers a reproducible, transparent method for ordered-risk assessment that outperforms traditional model-driven methods and standard data-driven models.

Traditional risk assessment often relies on expert-knowledge models like entropy weighting or fuzzy-rule systems. While interpretable, these methods struggle with nonlinear multi-source patterns and are sensitive to subjective weighting designs. Conversely, purely data-driven approaches lack transparency and often fail when reliable labeled data is scarce. The TA-DE-ELM framework bridges this gap by using a controlled synthetic benchmark where risk scores are generated from explicit, auditable expert rules rather than learned causal laws. This ensures that the underlying logic of risk prioritization remains clear and consistent, even when training data is limited.

The core innovation lies in the optimization mechanism. Standard differential evolution can converge prematurely when optimizing ELM input parameters. TA-DE solves this by separating the search process into two stages: an initial exploration phase using classical DE/rand/1 mutation to cover broad search spaces, followed by a refinement phase that applies current-to-best style mutation and elite local search. Additionally, a stagnation-triggered restart mechanism reinitializes part of the population if progress stalls, ensuring diversity and preventing premature convergence. This structured approach allows the model to learn complex, nonlinear mappings between heterogeneous observations and six distinct ordered-risk levels efficiently.

Experimental results demonstrate that TA-DE-ELM significantly outperforms baseline ELM variants, including PSO-ELM, DE-ELM, and GA-ELM, under a unified finite computational budget. It achieved an accuracy of 0.9404 and a macro-F1 score of 0.9396, closing nearly 29% of the gap to the theoretical Oracle ceiling. Furthermore, it surpassed traditional scoring references like RF-GRA and Entropy-GRA in both discrimination and risk-order consistency. The framework also exhibited robustness against feature degradation, showing slower performance decay under continuous noise compared to other optimizers. By isolating the contributions of refinement and restart mechanisms through ablation studies, the research confirms that these specific enhancements are critical for achieving superior classification performance without increasing inference latency. This work provides a vital tool for early-stage model screening and methodological comparison in environments where real-world labels are difficult to obtain, paving the way for more resilient and transparent automated decision-support systems.