Explainable artificial intelligence (XAI) is increasingly required for anomaly detection in high-dimensional sensor systems operating in safety-critical and resource-constrained environments. While existing post-hoc explanation methods provide useful insights, they often suffer from high computational cost, unstable attributions, and limited applicability in unlabeled or unsupervised settings. This paper proposes KFASL, a variance-stable and computationally efficient XAI framework that approximates Shapley-based feature attributions using a variance-optimized weighting strategy. The framework integrates local and global explanations with causality-aware regularization to improve attribution stability and interpretability under limited labeling conditions. The proposed approach reduces the computational complexity of Shapley approximation from exponential to polynomial time, enabling scalable deployment for high-dimensional telemetry data. KFASL is evaluated using a combination of real and synthetic datasets, with spacecraft telemetry used as a representative safety-critical case study. Experimental results demonstrate improved attribution stability, explanation fidelity, and runtime efficiency compared to existing XAI techniques, including SHAP, Kernel SHAP, LIME, and Anchors. These results indicate that KFASL provides a general and practical solution for explainable anomaly detection in complex sensor-driven systems.
Tahir et al. (Mon,) studied this question.