Intelligent Energy Forecasting: Empowering Underwater Wireless Sensor Nodes

The maritime domain is currently undergoing a significant transformation, driven by the digitization of its infrastructure, communication networks, and the modernization of marine services. The emergence of Smart Shipping (SMS) has introduced a wide range of smart services and applications that can be provided on-demand to smart vessels, smart ports, and various stakeholders, such as transportation companies and public authorities.

These smart applications, however, rely on a complex interconnected mesh of marine entities, including ports, ships, vessels, and a growing Internet of Maritime Things (IoMT) nodes. To effectively implement these smart applications, machine learning (ML) approaches have become essential, as they can leverage vast amounts of historical data to provide accurate predictions and optimize critical parameters, such as vessel fuel consumption, speed, routing, and port resource utilization.

While the introduction of ML has partially addressed the need for energy-efficient solutions in the maritime sector, the asymmetrical scaling of the number of IoMT devices that need to participate in the marine environment as active nodes poses a significant challenge. These heterogeneous IoMT devices, including fishing vessels, underwater devices, and ferries, not only need to periodically provide their collected data to a cloud server but also actively communicate with each other, participating in learning frameworks.

To overcome these challenges, the integration of Federated Learning (FL) and Over-the-Air Computation (AirComp) in the context of 6G-enabled Maritime Communication Networks (MCNs) has emerged as a promising solution. FL enables collaborative, privacy-preserving, and communication-efficient learning, while AirComp enhances the spectrum utilization and energy efficiency of the model sharing operations required by FL.

Federated Learning in Maritime Communication Networks

Conventional centralized learning schemes, where data from all IoMT devices are transmitted to a cloud-based centralized server, are often unsuitable for maritime applications due to several limitations:

1. Data Communication Overhead: The transmission of data from all IoMT devices to the centralized server can be energy-inefficient and introduce significant communication overhead.
2. Stringent Energy Constraints: IoMT devices, especially those deployed in remote or harsh maritime environments, often have limited battery resources, making energy-efficient communication a critical requirement.
3. Increased Transmission Failures: The harsh propagation environment of the maritime domain can lead to higher rates of transmission failures, further exacerbating the communication challenges.
4. Data Privacy Concerns: The data generated by IoMT devices may belong to multiple stakeholders, such as private companies and public authorities, requiring a privacy-preserving solution.

To address these limitations, the Federated Learning (FL) approach has been adopted in the context of MCNs. FL enables collaborative learning across multiple IoMT nodes without the need to centralize the data. Instead, each IoMT node trains a local ML model using its own data and then shares the model parameters with a central server, typically the shore-based Base Station (BS). The central server then aggregates the received model parameters, updates the global model, and sends the updated model back to the participating IoMT nodes.

This decentralized approach offers several advantages:

1. Data Privacy Preservation: The data generated by IoMT devices remain at the edge nodes, and only the model parameters are shared, ensuring data privacy.
2. Reduced Communication Overhead: By sharing only the model parameters, instead of the raw data, the communication overhead is significantly reduced.
3. Improved Energy Efficiency: The energy consumption associated with data transmission is minimized, as the IoMT nodes only need to transmit their model parameters, which are typically much smaller in size compared to the raw data.
4. Enhanced Generalizability: The global model, derived from the aggregation of local models, exhibits better generalization capabilities compared to models trained solely on local data, as it captures the collective intelligence from various IoMT nodes.

Over-the-Air Computation for Efficient Model Sharing

While FL addresses several limitations of centralized learning, the resource and energy constraints of the edge IoMT nodes can still pose challenges. Running complex ML models on these resource-constrained devices can quickly deplete their batteries, and the frequent model parameter transmissions required by the FL process can further exacerbate this issue.

To enhance the efficiency of the model sharing operation in FL, the Over-the-Air Computation (AirComp) technique has been integrated. AirComp is a physical-layer technique that enables multiple IoMT nodes to transmit their model parameters simultaneously over a common multi-access channel, leveraging the superposition property of the wireless medium.

In an AirComp-enabled system, all IoMT nodes transmit their pre-processed model parameters using the same bandwidth. The shore-based BS, acting as the fusion center, then applies a scaling function to the combined received signal, directly obtaining an estimation of the desired aggregated model parameters.

The integration of AirComp with FL offers several benefits:

1. Improved Computational Efficiency: By offloading the computation tasks to the more powerful shore-based BS, the burden on the resource-constrained IoMT nodes is reduced, improving their energy efficiency and battery lifetime.
2. Enhanced Spectrum Utilization: The use of a common multi-access channel for model parameter transmission, as opposed to individual channels, leads to more efficient spectrum utilization.
3. Scalability: The AirComp-FL approach can scale to a massive number of IoMT devices without significantly increasing the communication and computational overhead.
4. Continued Data Privacy: The FL process ensures that the data remains at the edge nodes, while the shared model parameters can be further protected through encryption.

Maritime Communication Network Architecture

The 6G-enabled Maritime Communication Network (MCN) considered in this context consists of three interconnected layers:

1. Underwater IoMT (UWI) Layer: This layer comprises Unmanned Underwater Vehicles (UUVs) responsible for collecting underwater data, such as images, temperature, currents, and pollution levels. The UUVs can either transmit data horizontally to other UUVs or vertically to the managing Unmanned Surface Vehicles (USVs).
2. Sea-Surface IoMT (SSI) Layer: This middle layer includes intelligent IoMT entities, such as ships, vessels, and buoys, capable of training their own local ML models based on locally collected data.
3. Aerial-Relay IoMT (ARI) Layer: The upper layer consists of Unmanned Aerial Vehicles (UAVs) that can act as relays, enabling communication between distant SSI entities and the shore-based BS. The UAVs can also support mission-critical services and disaster relief operations in remote locations.

The shore-based BS, located in close proximity to the sea level, serves as the central hub for the MCN, providing backhaul connectivity to the mobile network and the internet. Alternatively, the connectivity of the MCN nodes can be achieved through satellite communications for entities located beyond the reach of the shore-based BS.

The integration of FL and AirComp in this MCN architecture addresses several key challenges:

1. Communication Efficiency: The AirComp technique enhances the spectrum utilization and energy efficiency of the model sharing operations required by the FL process, reducing the communication overhead.
2. Privacy Preservation: The FL framework ensures that the data generated by IoMT devices remains at the edge nodes, while the shared model parameters can be further protected through encryption.
3. Scalability: The combination of FL and AirComp enables the scalable participation of a massive number of IoMT devices in the learning process, without significantly increasing the communication and computational burden.

Intelligent Maritime Transportation Systems

One of the key applications of the proposed FL-AirComp framework in the maritime domain is the prediction of Cargo Ship Propulsion Power (CSPP), which is a crucial parameter for intelligent maritime transportation systems.

Accurate prediction of CSPP can enable proactive strategies, such as predictive maintenance and fuel consumption optimization, ultimately reducing the operational costs and environmental impact of maritime transportation.

In the experimental scenario, the MCN monitors the CSPP of six cargo ships through a connected pair of shore-based BS and MCN Agent. Each cargo ship is equipped with the capability to locally train an LSTM-based ML model using its own collected data, such as ground speed, water speed, wind intensity, wind angle, and ship trim.

The key steps of the proposed FL-AirComp framework for CSPP prediction are as follows:

1. Local Model Training: Each cargo ship trains its own LSTM model using the locally collected data, optimizing the model’s hyperparameters, such as learning rate and depth.
2. Model Parameter Sharing: The cargo ships share their locally trained model parameters with the shore-based BS using the AirComp technique, which efficiently aggregates the model parameters over the air.
3. Global Model Aggregation: The shore-based BS applies a scaling function to the received superposed signal to obtain the average of the local model parameters, effectively creating a federated global model.
4. Global Model Broadcast: The updated global model is then broadcast back to the participating cargo ships, enabling them to update their local models and benefit from the collective intelligence.

The experimental results demonstrate that the proposed FL-AirComp approach outperforms a baseline Ensemble Learning (EL) technique by a factor of 3.04 in terms of the goodness of fit between the model-predicted and actual CSPP curves.

Furthermore, the impact of key parameters, such as the past-values window and forecast moment, on the prediction accuracy was investigated. The results suggest that a past-values window of 5 and a forecast moment of 15 minutes offer the best trade-off between computational complexity, proactive capabilities, and prediction accuracy.

Conclusion and Future Directions

The integration of Federated Learning and Over-the-Air Computation in the context of 6G-enabled Maritime Communication Networks presents a promising solution for addressing the challenges faced by intelligent maritime transportation systems. By leveraging the collective intelligence of distributed IoMT nodes while preserving data privacy and enhancing communication efficiency, the proposed framework can enable accurate predictions, optimize critical parameters, and ultimately reduce the environmental impact of the maritime sector.

As the maritime domain continues to evolve, several future research directions can be explored:

1. Multi-Hop Relaying: Extending the proposed framework to support multi-hop relaying of local models, optimizing for parameters such as total hopping delay and power consumption.
2. Energy-Efficiency Optimization: Reformulating the objective function to jointly minimize the computation distortion and the energy-efficiency of the MCN, further enhancing the sustainability of the system.
3. AirComp Error Analysis: Quantifying the impact of the AirComp-induced error on the overall performance and accuracy of the FL training process, incorporating real-world channel state information.
4. Partitioned AirComp Systems: Investigating the partitioning of the MCN into multiple AirComp subsystems to further improve the spectral and energy efficiency of the model sharing operations.
5. Intelligent Reflecting Surfaces: Exploring the integration of intelligent reflecting surfaces on the UAVs to enable fast and reliable model aggregation in unfavorable maritime wireless environments.

By continuously adapting and improving the proposed framework, the maritime sector can leverage the power of distributed intelligence and efficient communication to unlock new opportunities for sustainable, cost-effective, and data-driven decision-making in the era of Smart Shipping.