Towards Self-Healing Machine Learning Systems: Continual Learning and Drift-Aware MLOps Pipelines for Supply Chain Operations

Abstract

Supply-chain machine learning operates in inherently non-stationary settings, promotions, assortment churn, shocks, and sensor drift routinely shift data and labels, so static models degrade in accuracy, service levels, and latency. This review consolidates evidence on drift-aware, continual-learning pipelines and proposes a governed, self-healing MLOps architecture to support reliable, auditable operations. This review conducted a structured search (2015–2025) across Scopus, Web of Science, IEEE, ACM, and INFORMS, retaining studies using operational tabular/time-series data with business or reliability metrics and excluding theoretical, vision-only, and non-operational work. Extracted studies were thematically synthesised into drift landscape and adaptation, operational effectiveness and risk, and self-healing MLOps design. Fifteen applied studies spanning retail, logistics, manufacturing, pharma, and cold-chains reported reduced costs and errors, improved service stability, fairness-aware dispatch, better ETA/promise accuracy, and fewer false alerts. However, explicit drift detectors, reliability KPIs (time-to-detect, time-to-recover), and governance artefacts were inconsistently reported. The review recommended standardising reliability reporting and audits; creating temporal benchmarks that couple forecasting and policy tasks with drift injections and label latency; evaluating composite monitors and rollback/canary policies; strengthening data validation, feature/label lineage, and edge–cloud designs; and extending to federated, privacy-preserving settings with energy and carbon accounting.