Fuzzy Logic Gives AI Video Auditors the Confidence to Explain Themselves

A study published in the Journal of Ambient Intelligence and Humanized Computing proposes a framework for AI systems that can audit long industrial videos in a way humans can understand and trust. Led by Yahia Mady and Hani Hagras of the University of Essex with researchers from British Telecom, the system combines deep video transformers, fuzzy logic, and large language models tied to a knowledge base via retrieval-augmented generation. It focuses on action segmentation—labeling every moment in unedited recordings—rather than recognizing short, trimmed clips. The approach uses overlapping 64-frame sliding windows, two parallel frozen encoders (TimeSformer and VideoMAE), and an MS-TCN++ temporal convolution network to refine boundaries across up to 1024 windows, while fuzzy logic supports confidence reporting in plain language.