USING FEDERATED LEARNING WITH FIDUCIAL MARKERS FOR ARE MODELING OF THERMODYNAMIC PROCESSES WITHIN INDUSTRIAL PIPELINE NETWORKS

Abstract

The described work focuses on the development of an intelligent predictive thermal control system for industrial pipelines based on federated learning. The main emphasis is placed on implementing a decentralized architecture in which edge devices (local computing modules such as NVIDIA Jetson Nano) independently train models without transmitting raw sensor data to a central server. To process time series of sensor information-including temperature differentials, flow rate, and solute concentration-Temporal Convolutional Networks (TCNs) are employed, enabling parallel convolutional forecasting without recurrence. For local anomaly detection, an ensemble XGBoost model is used, allowing the computation of a numerical deviation score for thermodynamic profiles. A node’s participation in the global federated learning cycle is initiated based on its own temporal anomaly score, which determines the degree of deviation between the current model and local observations. If this score exceeds a threshold, the node generates a model update, encrypts it, and transmits it to the aggregator. Post-aggregation updates occur through weighted blending of the global model with the local one, followed by a brief phase of localized retraining. Spatial binding is achieved via AR markers containing a unique 40-bit identifier encoded in the center of an 8 × 8 binary matrix. To validate the architecture, a digital twin was implemented in Unity3D, simulating thermodynamic processes using approximated Navier-Stokes equations. Each pipeline segment is linked to a corresponding edge node, which exchanges data in real time via gRPC. The efficiency of marker detection algorithms was tested under conditions of simulated blur and varying illumination. AR visualization was realized based on thermal trajectory mapping, enabling intuitive display of risk zones and dynamic anomalies within the pipeline. Future research may focus on scaling the system for multi-level pipeline networks, expanding the range of monitored parameters, and implementing adaptive optimization of thermal regimes through multi-agent coordination.