🤖 AI

Towards Energy Impact on AI-Powered 6G IoT Networks: Centralized vs. Decentralized

This paper analyzes energy consumption in 6G IoT networks and demonstrates through a German railway testbed that decentralized machine learning architectures achieve competitive predictive accuracy while reducing overall electricity consumption by up to 70% compared to centralized approaches.

Original authors: Anjie Qiu, Donglin Wang, Sanket Partani, Andreas Weinand, Hans D. Schotten

Published 2026-04-22

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Anjie Qiu, Donglin Wang, Sanket Partani, Andreas Weinand, Hans D. Schotten

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a massive team of 100 detectives (your IoT sensors) scattered across a huge city (the railway network). Their job is to spot tiny cracks in the train tracks before they cause a disaster. They need to learn how to spot these cracks using Artificial Intelligence (AI).

The big question this paper asks is: What is the most energy-efficient way to teach these detectives?

Do we send every single detective's raw notes to a giant, super-powered "Brain HQ" (Centralized) to be analyzed? Or do we let the detectives learn locally and just send their "aha!" moments back to HQ (Decentralized)?

Here is the breakdown of the study, explained simply:

1. The Two Approaches: The "All-Hands" Meeting vs. The "Local Study Group"

The Old Way (Centralized Learning):
Imagine every detective has to run to the central office, hand over their entire notebook of raw observations, and wait for the super-computer at HQ to read them, figure out the pattern, and then send the final lesson back.

The Problem: Running back and forth with heavy notebooks takes a lot of energy (battery power). If you have thousands of detectives, the "running" (data transmission) burns more energy than the actual "thinking" (AI processing).

The New Way (Decentralized/Federated Learning):
Imagine the detectives stay in their local neighborhoods. They each learn from their own notebooks. Once a week, they don't send their notebooks; they just send a tiny summary of what they learned (the "weights" or "rules" of the AI) to HQ. HQ mixes all these summaries together to create a smarter "Master Guide," sends it back, and the detectives update their local knowledge.

The Benefit: They stop running back and forth with heavy notebooks. They only exchange tiny summaries. This saves a massive amount of energy.

2. The Real-World Test: The German Railway Experiment

The researchers didn't just do this on a computer; they actually tested it on the German railway system (Deutsche Bahn).

The Setup: They installed over 1,300 sensors on train tracks across four major cities. These sensors watch trains go by (at speeds up to 175 mph!) to predict when a track switch needs fixing.
The Data: They collected a mountain of data (14.3 Terabytes!) over a year.

3. The Results: Who Won?

They compared the two methods side-by-side:

Accuracy: Both methods were almost equally good at predicting track issues. They both got about 90% accuracy. The "Local Study Group" (Decentralized) was just as smart as the "Brain HQ" (Centralized).
Energy Consumption: This is where the Decentralized method crushed the competition.
- The Centralized method had to transmit all the raw data, which is like sending a truck full of bricks.
- The Decentralized method only transmitted the lessons learned, which is like sending a single postcard.
- The Result: The Decentralized approach saved up to 70% of the electricity used for transmission.

4. Why This Matters for the Future (6G)

We are moving toward 6G networks, which will connect billions of devices (smart fridges, self-driving cars, factory robots).

If we use the "Old Way" (Centralized), the energy bill to move all that data will be astronomical, and it will hurt the environment.
If we use the "New Way" (Decentralized), we can keep the devices smart and efficient without draining their batteries or the power grid.

The Bottom Line Analogy

Think of it like a classroom:

Centralized: Every student writes their entire essay, runs to the teacher's desk, and waits for the teacher to grade it and write a new essay for them. The students are exhausted from running.
Decentralized: Students write their essays at their desks. They only raise their hands to tell the teacher, "I learned that 'A' is important." The teacher writes a new rule on the board, and everyone updates their own notes. The students stay seated, save their energy, and still learn the same lesson.

Conclusion: By letting AI learn locally and only sharing the "wisdom" rather than the "raw data," we can build a smarter, greener, and more sustainable future for our connected world.

1. Problem Statement

The integration of Artificial Intelligence (AI) and Machine Learning (ML) into 6G-enabled Internet of Things (IoT) networks offers significant advancements in automation and predictive maintenance. However, this integration introduces a critical challenge: energy efficiency.

Centralized Bottlenecks: Traditional Centralized Learning (CL) architectures require transmitting massive amounts of raw sensor data from IoT devices to a central cloud server. This creates high communication overhead, consumes significant bandwidth, and leads to substantial energy consumption during data transmission and centralized processing.
Sustainability Gap: As IoT networks scale (e.g., smart cities, industrial rail), the cumulative energy cost of training models and transmitting data threatens the sustainability of 6G deployments. There is a lack of empirical data comparing the energy trade-offs between centralized and decentralized approaches in real-world, large-scale 6G scenarios.

2. Methodology

The authors propose a quantitative energy consumption model and validate it through a real-world testbed.

A. Energy Consumption Model

The study develops a generalized mathematical model to quantify energy costs ( $E$ ) for both architectures, breaking them down into Computing Costs ( $E_{Compute}$ ) and Communication Costs ( $E_{Trans}$ ).

Centralized Learning (CL):
- Process: Raw data ( $\epsilon_k$ ) is transmitted from $K$ devices to a central server via 5G Base Stations (BS). Training occurs entirely on high-performance server hardware (CPU/GPU/TPU).
- Energy Formula: $E_{CL} = E_{Compute} + E_{Trans}$ . The communication cost is dominated by the uplink transmission of large raw datasets ( $b(\epsilon_k)$ bits).
Decentralized/Federated Learning (FL):
- Process: Utilizes the Federated Averaging (FedAvg) algorithm. The global model is distributed to edge nodes. Local training occurs on edge devices (using embedded AI processors). Only model parameters/weights ( $b(W)$ bits), not raw data, are transmitted back to the server for aggregation.
- Energy Formula: $E_{FL} = E_{Compute} + E_{Trans}$ . While communication involves bidirectional transmission of model weights, the volume of data is significantly lower than raw sensor data. Computing is distributed across edge nodes.

B. Real-World Testbed (SLKI Project)

The theoretical model was validated using a deployment within the German railway infrastructure (Project: Second-Level Künstliche Intelligenz in Weichen - SLKI), in collaboration with Deutsche Bahn.

Scale: Over 1,300 integrated sensor units installed on 650 railway turnouts across four major German cities (Kassel, Mainz, Stuttgart, Ulm).
Data: The system operates 24/7, generating >5 million measurement records (14.3 TB) covering train speeds from 40 to 280 km/h.
Hardware:
- CL Setup: High-performance desktop for centralized training.
- FL Setup: Edge nodes using NXP i.MX 8M Plus embedded boards (featuring Cortex-A53 CPUs and NPUs) for local training.
Task: Predictive maintenance (predicting train speeds) using a Convolutional Neural Network (CNN).
Measurement Tools: The CodeCarbon library was used to estimate electricity usage and carbon emissions across CPU, GPU, NPU, and RAM components.

3. Key Contributions

Quantitative Energy Modeling: Proposed a unified framework to calculate energy costs for 6G IoT systems, explicitly separating computing and transmission costs for both CL and FL architectures.
Large-Scale Empirical Validation: Unlike many studies relying on simulations, this work presents results from a real-world, large-scale deployment involving thousands of sensors and terabytes of data in a critical infrastructure setting (railways).
Architecture Comparison: Provided a direct, apples-to-apples comparison of CL vs. FL, normalizing for predictive accuracy to ensure fair energy evaluation.
Sustainability Metrics: Integrated carbon emission estimation into the energy analysis, linking technical architecture choices to environmental impact.

4. Key Results

The study compared the two architectures based on prediction accuracy (RMSE) and total energy consumption (kWh).

Predictive Accuracy:
- Both CL and FL architectures achieved comparable performance, converging to approximately 90% accuracy after 20 training epochs (or federated rounds).
- FL showed a marginal decrease in accuracy compared to CL due to the limited computational power of edge devices, but the difference was negligible for the application.
Energy Consumption:
- Transmission: In the FL architecture, data transmission energy was higher than CL relative to the specific transmission events due to bidirectional model updates. However, the volume of data transmitted in FL (model weights) was drastically lower than the raw data in CL.
- Training: CL required significantly more energy for training due to the processing of massive raw datasets on central servers. FL distributed this load to energy-efficient edge devices.
- Total Savings: The decentralized (FL) approach resulted in a reduction of up to 70% in total energy consumption compared to the centralized approach.
- Breakdown: In the FL configuration, data transmission accounted for up to 70% of the total energy footprint, highlighting that while transmission is a major cost, the savings from avoiding raw data transmission and centralized heavy computing outweigh this.
City-Specific Data (Table I):
- Kassel: CL (144.6 kWh) vs. FL (41.7 kWh)
- Mainz: CL (1283.9 kWh) vs. FL (370.8 kWh)
- Stuttgart: CL (896.7 kWh) vs. FL (259.0 kWh)
- Ulm: CL (397.6 kWh) vs. FL (114.8 kWh)
- Observation: FL consistently reduced energy usage by roughly 70-71% across all locations.

5. Significance and Conclusion

Sustainable 6G Design: The findings prove that Decentralized/Federated Learning is a superior architectural choice for energy-constrained 6G IoT networks. It enables sustainable scaling without compromising predictive accuracy.
Mitigation of Transmission Costs: By shifting computation to the edge and transmitting only model parameters, the system mitigates the massive energy costs associated with transmitting raw sensor data over 5G/6G networks.
Industrial Impact: The successful deployment in the German railway sector demonstrates the viability of AI-driven predictive maintenance that is both operationally efficient and environmentally sustainable.
Future Directions: The authors suggest future research should focus on safety-critical systems where latency is paramount, and further assessment of the ROI of AI integration in national-scale IoT infrastructures regarding energy overhead vs. productivity gains.

In summary, the paper provides compelling evidence that transitioning from centralized to decentralized AI architectures is essential for the energy sustainability of future 6G IoT ecosystems.