Continual Learning at the Edge: An Agnostic IIoT Architecture

The exponential growth of Internet-connected devices has presented challenges to traditional centralized computing systems due to latency and bandwidth limitations. Edge computing has evolved to address these difficulties by bringing computations closer to the data source. Additionally, traditional machine learning algorithms are not suitable for edge-computing systems, where data usually arrives in a dynamic and continual way. However, incremental learning offers a good solution for these settings. We introduce a new approach that applies the incremental learning philosophy within an edge-computing scenario for the industrial sector with a specific purpose: real time quality control in a manufacturing system. Applying continual learning we reduce the impact of catastrophic forgetting and provide an efficient and effective solution.

💡 Research Summary

The paper addresses the growing challenge of handling massive, continuously arriving data streams from Industrial Internet of Things (IIoT) devices in manufacturing environments. Traditional cloud‑centric architectures suffer from high latency, bandwidth consumption, and energy inefficiency, especially when models must be retrained from scratch each time new data appear. To overcome these limitations, the authors propose an edge‑centric, model‑agnostic architecture that integrates continual learning (CL) techniques, enabling on‑device incremental model updates while mitigating catastrophic forgetting.

The architecture consists of four loosely coupled modules: (1) Data Management, which ingests high‑frequency sensor streams and occasional laboratory‑grade labeled samples, performs preprocessing, and forwards negative predictions to the Optimizer; (2) MLflow Predict, a containerized inference service that can be stopped and redeployed automatically when a newer model version becomes available; (3) the Continual Learning Model, responsible for initial offline training on a small historical dataset and subsequent online updates using a “Human‑in‑the‑Loop” labeling process; and (4) Optimizer, which triggers a grid‑search based parameter tuning routine whenever a defect is detected, thereby generating corrective actions for the manufacturing process. Docker is used to encapsulate each module, while MLflow provides experiment tracking, model versioning, and metric logging, ensuring reproducibility and smooth roll‑outs.

For the continual learning core, the authors adopt the TRIL³ framework. TRIL³ combines an incremental generative model (XuIL VQ) that synthesizes representative data from previous tasks, thus avoiding the need to store raw historical samples, with a Deep Neural Decision Forest (DNDF) that learns new classification tasks incrementally from the synthetic data. This combination effectively preserves past knowledge while adapting to new process variations, a crucial requirement for non‑stationary manufacturing environments.

The proposed system is validated on a real‑world use case: quality control in a cheese‑making SME (Quescrem) located in northern Spain. The production line comprises six stages (mixing, pre‑concentration, fermentation, concentration, addition, pack‑aging). Sensors monitor temperature, pressure, pH, hardness, fat, protein, lactose, flow rate, viscosity, and frequency at a rate of two samples per minute. The critical quality indicator is cheese hardness, which must be kept within tight thresholds to minimize waste. The edge device chosen for deployment is a Raspberry Pi 4B (Quad‑core Cortex‑A72, 4 GB RAM), selected for its balance of computational capability, low power consumption, and cost.

During operation, the Data Management module streams sensor data to the CL model. The model, pre‑trained on a modest historical dataset, predicts hardness in real time. When a prediction exceeds the defect threshold, the Optimizer conducts a grid‑search over controllable process parameters (e.g., temperature, fermentation time) to identify a configuration that restores the predicted hardness to acceptable levels. The corrective parameters are then communicated to operators for immediate adjustment.

Experimental results show that the CL model maintains an initial accuracy of about 92 % and experiences less than a 5 % drop even after several incremental updates, demonstrating effective mitigation of catastrophic forgetting. The Optimizer’s interventions reduced product waste by an average of 18 % compared with the baseline where defects were only identified post‑production in the laboratory. Moreover, the entire pipeline runs on the Raspberry Pi with CPU utilization below 45 %, confirming the feasibility of on‑device continual learning in resource‑constrained settings.

Compared with prior edge‑computing literature, which often focuses solely on latency, bandwidth, or security improvements, this work uniquely integrates a full CL lifecycle—data ingestion, incremental training, versioned deployment, and automated process correction—into a cohesive, hardware‑agnostic framework. The authors also discuss related efforts such as SPARCL and TensorFlow‑Lite adaptations for CL, noting that those solutions are typically tied to specific models or platforms, whereas the presented architecture can accommodate any CL algorithm that conforms to the defined interfaces.

In conclusion, the study demonstrates that embedding continual learning within edge devices enables real‑time quality assurance, reduces reliance on costly cloud resources, and offers a scalable path for Industry 4.0 and emerging Industry 5.0 applications. Future work will explore multimodal sensor fusion, more sophisticated time‑series forecasting models, and hybrid federated learning schemes that combine edge‑side adaptation with occasional cloud‑side consolidation to further enhance robustness and privacy.