Improving Trust in Safety-Critical AI Systems: Explainable AI and Anomaly Detection Frameworks for human safety in Smart Industries

The growing use of deep learning systems in safety-critical applications, such as smart industries, collaborative robots environment, and healthcare, raises basic questions of trust, transparency, and accountability. While modern propositions have shown incredible predictive performance but their black-box nature of decision-making poses a challenge in applications where trust and readability are of paramount importance. This thesis focuses on the advancement in the field of Explainable Artificial Intelligence (XAI) through methodological contributions and applications in industry and safety-critical environments, especially the field of explainable anomaly detection. First, the thesis focuses on the issue of instability of local surrogate explanations in LIME-Image. It presents a new sampling strategy called LIME Stratified. The proposed approach implements the unbiased stratified sampling with respect to the superpixel masks instead of the random perturbations, which are based on Monte Carlo sampling. By explicitly managing the distribution of the feature coalitions the proposed strategy lessens the variance in the synthetic neighborhood, improves the coverage of relevant feature combinations and gives explanations that are significantly more stable and faithful. Extensive empirical analysis shows state-of-the-art improvements over several common LIME, on ImageNet data sets using resnet CNN model. Second, the thesis presents ShapBPT, which is a hierarchical feature attribution algorithm that uses Binary Partition Trees and the Owen approximation of the Shapley value. By using the hierarchy of image regions, ShapBPT generates semantically meaningful attribution results with a more efficient use of computational budgets. Comprehensive experiments on convolutional and transformer-based architectures demonstrate that ShapBPT is better at localization accuracy and efficiency compared to existing cutting-edge attribution methods and thus can be applied on a large scale and in real time. Third, the thesis studies the dependability of anomaly detection systems by tightly integrating generative models and XAI techniques. Variational Autoencoder- Generative Adversarial Networks (VAE-GANs) are trained only on data without any defect to produce the pixel-level anomaly maps, which are explained through LIME and SHAP. By comparison of explanation maps and ground truth defect regions, the contributions of analyzing votes for anomalies to analyze and interpret those anomalies alert to detect reasons that are not correct. This study makes it clear how performance-only evaluation has limitations and drives the introduction of explanation-based validation as a necessary complementary aspect for trustworthy anomaly detection. Finally, a new zoned anomaly detection and visual intelligence system for industrial safety monitoring, called ADVIS, is introduced in this thesis. ADVIS combines v area-specific VAE-GAN-based detectors, anomaly scoring and saliency-based explanations in order to produce accurate, interpretable, and real-time safety evaluations. Validation on realistic synthetic palletizing scenarios, like human-robot interaction data, highlights high precision and recall and provides interpretable outputs that can be acted on by the engineers and safety operators. Overall, this thesis makes methodical contributions to boost the robustness, faithfulness and practicality of XAI methods and verify their effectiveness in real industrial scenarios. The work concludes by identifying open challenges and future research directions, including multimodal explainability, integration of XAI with vision-language models, and scalable deployment of trustworthy AI in dynamic industrial environments.