Causal continual unlearning with disentangled anomaly representations for private industrial vision

Continual Learning and Private Unlearning (CLPU) in industrial anomaly detection (IAD) demands lifelong adaptation to new product-defect pairs while enabling GDPR-compliant erasure of proprietary client data. Existing methods fail to reconcile three critical limitations: (i) disentanglement gap, inability to share universal anomaly primitives while isolating IP-sensitive semantics; (ii) causal leakage, unmodeled transitive privacy violations in multi-stage inspection pipelines; and (iii) memory inefficiency, prohibitive storage costs for high-resolution vision. We redefine unlearning not as parameter deletion, but as structured information control: a theoretically grounded process that selectively removes sensitive knowledge while preserving transferable representations. We introduce Causal Disentangled Industrial Unlearning (CaDIU), which establishes (1) privacy-preserving disentanglement via causally factorized latent spaces, (2) system-level causal indistinguishability through Defect Propagation Graphs with ancestral restoration, and (3) information-theoretic memory bounds enabling exact unlearning with (Formula presented) storage. Evaluated on four industrial benchmarks, VisA, MVTec-AD, Real-IAD, and BTAD, CaDIU achieves state-of-the-art performance: Anomaly Primitive Fidelity  ≥  0.96, IP Reconstruction Error  ≤  0.12, Causal Leakage Score  ≤  0.03, and Memory Efficiency Ratio = 0.08. Our framework is the first to enable efficient, private, and certifiable lifelong industrial vision, providing a foundation for GDPR-compliant, edge-deployable AI in multi-client manufacturing environments.