Adaptive Transfer Learning Based Anomaly Detection for Industrial IoT System

Abstract

This paper proposes a transfer learning-based anomaly detection framework coupled with adaptive learning for predictive maintenance in Industrial Internet of Things (IIoT) systems. The proposed framework is designed and validated using a small-scale prototype to demonstrate its feasibility and effectiveness. The proposed architecture addresses the computational limitations of edge devices by employing a cloud-assisted/remote processing framework, where sensor data are streamed through Apache Kafka for real-time analysis. A lightweight Conditional Autoencoder is used to detect anomalies from streaming sensor data, making the framework suitable for resource-constrained industrial environments. To overcome the scarcity of machine-specific training data, a global model is trained using data collected from machines belonging to the same machine category. The trained global model is subsequently adapted to individual machines using transfer learning whenever machine-specific data are available. The framework continuously monitors incoming sensor data for concept drift and performs adaptive retraining when persistent drift is detected, enabling the model to learn changes in machine behavior caused by aging and wear. The system processes incoming data through parallel anomaly detection and drift detection pipelines while storing labeled operational data for future model adaptation and retraining. The proposed framework provides an adaptive and scalable approach for real-time anomaly detection in Industrial IoT predictive maintenance systems. The experimental implementation is carried out using a remote processing framework. The proposed architecture is cloud-ready and is intended to support cloud-assisted deployment for large-scale Industrial IoT environments in future work.

Introduction

The text presents a cloud-assisted, adaptive anomaly-detection framework for Industrial Internet of Things (IIoT) predictive maintenance. The main objective is to detect machine anomalies in real time while addressing two major industrial challenges: limited computational resources at edge devices and concept drift caused by changing machine conditions over time.

The proposed system uses Apache Kafka to continuously stream sensor data from industrial machines to a remote/cloud processing environment. Sensor measurements such as vibration, temperature, acoustic signals, current, and signal features are transmitted in JSON format. This separates data collection from computationally intensive processing and enables scalable real-time monitoring.

The framework consists of two main layers:

• Data Collection Layer: Collects sensor data from industrial machines and publishes it to Kafka topics. This provides reliable and asynchronous communication between machines and processing components.
• Data Processing Layer: Processes the incoming stream through anomaly detection, concept-drift detection, and adaptive model retraining modules.

For anomaly detection, the system uses a lightweight Conditional Autoencoder. Instead of training a separate model for every machine—which is difficult because machine-specific data are often limited—a global model is initially trained using data from machines of the same category. The model and its metadata are stored separately, with metadata managed through PostgreSQL.

When a new machine is introduced, transfer learning adapts the global model using available machine-specific data. This allows the model to retain generalized knowledge while learning the unique characteristics of the new machine. If insufficient local data are available, the global model can be used directly.

During real-time operation, sensor data are normalized and passed to the Conditional Autoencoder. The model reconstructs the input, and the reconstruction error is compared with a predefined threshold. Data exceeding the threshold are classified as anomalies, while normal observations are stored in the database for future learning.

The framework also addresses concept drift, which occurs when machine behavior changes because of wear, aging, environmental conditions, or changing operating conditions. A separate drift-detection pipeline operates alongside anomaly detection. Importantly, only normal observations are used for drift analysis so that abnormal events do not incorrectly trigger drift detection. The system uses ADWIN to identify persistent changes. When drift continues for a predefined number of detections, the system considers it genuine concept drift and retrains the machine-specific model using stored normal operating data.

The implementation follows a modular producer–consumer architecture with separate components for database management, streaming, anomaly detection, and adaptive training. Multiple threads are used to handle database connections, anomaly detection, and adaptive training.

For experimentation, real IoT sensors were not directly connected. Instead, a Python simulation script generated sensor data based on the Smart Manufacturing IoT-Cloud Monitoring Dataset and streamed it to the Kafka topic machine-data. Three machine instances—M01, M02, and M03—were analyzed. Clustering showed that M01 and M03 had similar operating behavior. M01 was therefore used to create the global model, and transfer learning was applied to adapt this model to M03.

The transfer-learning experiment compared four situations: using only M03 data before transfer learning, combined M01–M03 data before transfer learning, only M03 data after transfer learning, and combined data after transfer learning.

Overall conclusion

The proposed framework combines Kafka-based real-time streaming, Conditional Autoencoders, transfer learning, concept-drift detection, and adaptive retraining to create a continuously improving predictive-maintenance system. Its key advantage is that it reduces dependence on large machine-specific datasets while allowing models to adapt as machines change over time. The approach therefore addresses both data scarcity and evolving machine behavior, making it suitable for scalable cloud-assisted IIoT anomaly detection.

Conclusion

Adaptive transfer-learning-based anomaly detection has been implemented, and various test cases were performed to evaluate the generalization of the model. The model was better able to identify machine-specific anomalies after adaptive transfer learning. The effectiveness of the proposed transfer learning approach was evaluated using machine-specific data. After applying transfer learning, the model achieved an accuracy of 92.08%, precision of 90.40%, recall of 94.17%, and an F1-score of 92.24%. Among these metrics, recall showed the most significant improvement, increasing from 88.3% to 94.1%. This demonstrates that the adaptive transfer-learned model became more effective in identifying anomalies specific to a particular machine, while requiring only a limited amount of machine-specific training data. Since predictive maintenance prioritizes detecting actual faults over minimizing false alarms, the improved recall indicates that the proposed adaptive transfer learning approach enhances machine-specific anomaly detection and reduces the likelihood of missing critical anomalies.

References

[1] R. Saleh, K. P. Dinh, B. Villányi, and T.-S. Hy, “Self-Evolving Multi-Agent Network for Industrial IoT Predictive Maintenance,” arXiv preprint arXiv:2602.16738, 2026. [2] H. Choi, D. Kim, J. Kim, and P. Kang, “Explainable anomaly detection framework for predictive maintenance in manufacturing systems,” Applied Soft Computing, vol. 125, p. 109147, 2022. [3] L. Bowden, Q. Chu, B. Cena, K. Ohno, B. Parney, D. Sharma, and M. Takeori, “Machine Failure Detection Based on Projected Quantum Models,” arXiv preprint arXiv:2601.15641, 2026. [4] A. Giannoulidis and A. Gounaris, “A context-aware unsupervised predictive maintenance solution for fleet management,” Journal of Intelligent Information Systems, vol. 60, pp. 521–547, 2023. [5] S. O. Alhuqayl, A. T. Alenazi, H. A. Alabduljabbar, and M. A. Haq, “Adaptive machine learning models for predictive maintenance in industrial internet of things systems,” Scientific Reports, 2026. [6] S. F. Chevtchenko, M. C. M. dos Santos, D. M. Vieira, R. L. Mota, E. Rocha, B. V. Cruz, D. Araújo, and E. Andrade, “Predictive Maintenance Model Based on Anomaly Detection in Induction Motors: A Machine Learning Approach Using Real-Time IoT Data,” arXiv preprint arXiv:2310.14949, 2023. [7] H. Zheng, A. R. Paiva, and C. S. Gurciullo, “Advancing from Predictive Maintenance to Intelligent Maintenance with AI and IIoT,” arXiv preprint arXiv:2009.00351, 2020. [8] W. Wang, H. Zhang, D. Yang, and Y. Xiang, “Deep-learning-enabled predictive maintenance in industrial internet of things: Methods, applications, and challenges,” IEEE Systems Journal, vol. 17, no. 2, pp. 2602–2615, 2022. [9] A. L. Alfeo, S. Cimino, and G. Vaglini, “Using an autoencoder in the design of an anomaly detector for smart manufacturing,” Pattern Recognition Letters, vol. 136, pp. 271–278, 2020. [10] B. Steenwinckel, J. De Paepe, S. De Turck, and F. Ongenae, “FLAGS: A methodology for adaptive anomaly detection and root cause analysis on sensor data streams by fusing expert knowledge with machine learning,” Future Generation Computer Systems, vol. 123, pp. 30–45, 2021. [11] M. Canizo, I. Triguero, A. Conde, and E. Onieva, “Multi-head CNN-RNN for multivariate time series anomaly detection: An industrial case study,” Neurocomputing, vol. 363, pp. 246–260, 2019. [12] Z. Ding and M. Fei, “An anomaly detection approach based on isolation forest algorithm for streaming data using sliding window,” IFAC Proceedings Volumes, vol. 46, no. 20, pp. 12–17, 2013. [13] U. W. Hore and D. G. Wakde, “Intelligent Predictive Maintenance for Industrial Internet of Things Using Machine Learning Approach,” in Proc. Int. Conf. Intelligent Cyber Physical Systems and IoT, Springer, 2022. [14] R. Anandan, S. Gopalakrishnan, S. Pal, and N. Zaman, Industrial Internet of Things (IIoT): Intelligent Analytics for Predictive Maintenance. Wiley, 2022. [15] D. Fernández-Francos, B. Martínez-Rego, O. Fontenla-Romero, and A. Alonso-Betanzos, “Automatic bearing fault diagnosis based on one-class classification using vibration signals,” Computers in Industry, vol. 125, 2021. [16] D. Galatro, M. Shroff, and C. H. Amon, “Adaptive Transfer Learning Strategy for Predicting Battery Aging in Electric Vehicles,” MDPI, Jan. 2025. [17] Kaggle, “Industrial IoT Dataset (Synthetic).” [Online]. Available: https://www.kaggle.com/datasets/canozensoy/industrial-iot-dataset-synthetic

Copyright

Copyright © 2026 Abinya S, Dr. S. Sasirekha. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.