TY  - JOUR
AU  - L Sindhu Priyanka
AU  - Y Shravani
AU  - A Lakshmi Narayana
PY  - 2026
DA  - 2026/07/25
TI  - An Adaptive Continual Learning Framework with Dynamic Memory Replay for Incremental Classification of Non-Stationary Data Streams
JO  - Global Journal of Engineering Innovations and Interdisciplinary Research
VL  - 6
IS  - 5
AB  - The rapid growth of real-time data generated from Internet of Things (IoT) devices, sensor networks,
financial systems, cybersecurity infrastructures, autonomous vehicles, healthcare monitoring systems,
and social media platforms has significantly increased the demand for intelligent machine learning
models capable of continuously adapting to evolving data distributions. Unlike traditional machine
learning models that assume stationary data distributions and offline training, real-world streaming
environments are characterized by concept drift, evolving class distributions, changing feature spaces,
and continuously arriving data instances. Conventional deep learning models suffer from catastrophic
forgetting when incrementally trained on new data, resulting in significant degradation of previously
acquired knowledge and poor long-term learning performance. Continual Learning (CL) has emerged as
an effective paradigm for enabling artificial intelligence systems to learn continuously while preserving
historical knowledge. This paper proposes an Adaptive Continual Learning Framework with Dynamic
Memory Replay for Incremental Classification of Non-Stationary Data Streams, integrating adaptive
concept drift detection, dynamic memory replay, transformer-based feature learning, uncertainty-
aware sample selection, adaptive knowledge distillation, experience replay optimization, and continual
representation learning into a unified incremental learning architecture. Initially, streaming data
are continuously monitored for distributional changes using adaptive drift detection mechanisms.
Representative historical samples are dynamically maintained within an optimized replay memory to
preserve previously learned knowledge while minimizing storage requirements. The proposed framework
establishes a scalable, adaptive, and trustworthy continual learning system suitable for intelligent
decision-making across IoT analytics, cybersecurity, financial forecasting, healthcare monitoring,
industrial automation, autonomous systems, smart cities, and future real-time artificial intelligence
applications.
SN  - 3066-1226
UR  - https://dx.doi.org/10.33425/3066-1226.1313
DO  - 10.33425/3066-1226.1313