A Continuous MLOps Lifecycle Framework for Automated Model Re-training and Drift Detection in Collaborative Distributed Deep Learning Systems

• Asher Noah

  covenant University

  Author

The deployment of deep learning models in collaborative distributed environments has introduced unprecedented challenges in maintaining model performance under non-stationary data conditions. While MLOps practices have matured for centralized systems, a significant research gap persists in frameworks capable of orchestrating drift detection and automated retraining across heterogeneous, privacy-sensitive distributed architectures. This study designed, implemented, and evaluated a continuous MLOps lifecycle framework integrating adaptive drift detection mechanisms with event-driven retraining workflows specifically tailored for collaborative distributed deep learning systems. The methodology employed a design-based research approach, implementing the framework on a federated learning testbed comprising eight edge nodes processing multi-source retail demand and IoT sensor data streams. The framework combined Population Stability Index and Adaptive Windowing for multivariate drift detection, with a reinforcement learning-based retraining policy that optimizes the trade-off between model freshness and computational cost. Experimental validation across 12-week deployment cycles demonstrated that the proposed framework achieved 89.4% accuracy retention compared to static baselines (76.2%), while reducing unnecessary retraining events by 47% and lowering overall computational overhead by 33%. The findings establish that continuous, drift-aware MLOps orchestration substantially extends model operational lifetime in distributed settings. Practically, the framework provides a replicable architecture for organizations deploying collaborative learning systems where data cannot be centralized, offering specific metrics for monitoring and retraining trigger calibration.

• pdf

Copyright (c) 2026 Asher Noah (Author)

This work is licensed under a Creative Commons Attribution 4.0 International License.