Procedural Mistake Detection (PMD) aims to identify deviations from an expected multi-step workflow while a user is performing a task. In egocentric settings, this requires strictly causal inference, low latency, and robustness to visual noise such as occlusions and fine-grained hand–object interactions. In this thesis, PMD is formulated as a One-Class Classification problem: models are trained only on correct executions and must detect any deviation at inference time. We adopt a dual-branch online pipeline where (i) a step recognition branch performs Online Action Detection (OAD) and (ii) a step anticipation branch predicts the next expected step from the recognised history; a mistake is signalled when executed and expected steps disagree.