A fuzzy logic based generative explainable AI for automated video auditing

Abstract Video auditing can provide vital insights into complex industrial operations where quality, safety, and compliance are critical. However, there is a need to develop explainable artificial intelligence frameworks capable of recognising workflow steps and producing interpretable reports from long, untrimmed video sequences. The key module enabling reliable auditing is action segmentation, which remains challenging due to ambiguous step boundaries, fine-grained action similarities, and uncertainties caused by environmental variability, multi-object interactions, and class imbalance. To address these challenges, we introduce a framework which integrates fuzzy logic systems, deep learning models, and large language models through retrieval-augmented generation (RAG). The fuzzy logic component operates at two stages: generating soft training labels at action boundaries and providing per-window decision confidence with linguistic interpretability at inference, while a hybrid VideoMAE–TimeSformer architecture with MS-TCN++ sequence modelling captures spatiotemporal dependencies for accurate step prediction. The large language model component enriches predictions with external knowledge and generates structured, human-readable audit reports aligned with operational standards through retrieval-augmented generation. Several experiments on real-world maintenance videos were conducted to compare the proposed system with conventional deep learning pipelines. The results demonstrate that the proposed fuzzy logic system enhances the interpretability of audit outcomes, providing human-understandable confidence reasoning at action boundaries and generating explainable audit reports aligned with operational standards.