Modeling Vessel Shaft Power from Noon Reports: Data Fusion Strategies for Deep Learning under varying Data Availability

The global economy moves on the water. Roughly eighty percent of all international trade travels by ship, making maritime transport the invisible backbone of modern life. To keep these massive vessels running efficiently, operators need to understand exactly how much power their engines are using at any given moment. This power, known as shaft power, is directly tied to fuel consumption and the greenhouse gases ships emit. The more efficiently a ship moves through the water, the less fuel it burns and the cleaner the air remains. However, knowing this power in real time is difficult. While some modern ships are equipped with high-tech sensors that record engine performance every fifteen minutes, many vessels operate without them. Instead, they rely on "noon reports," a traditional logbook entry made once a day by the crew. These daily notes capture a snapshot of the ship's journey, including its speed, the weather, and the engine's output, but they lack the fine detail of continuous monitoring. For decades, this gap in data has made it hard to predict how a ship will perform or to optimize its route for fuel savings, leaving a significant blind spot in the effort to make shipping greener.

A team of researchers at Simula Research Laboratory in Norway has developed a new way to fill this gap. They asked whether it is possible to use the sparse, daily noon reports, combined with public data on weather and ocean currents, to accurately predict the engine power of a ship, even when high-frequency sensor data is missing. To test this, they built three different deep learning models, a type of artificial intelligence that learns patterns from data. Each model was designed for a different situation: one for ships with very little historical data, one for ships that have a few months of logs, and one for ships with years of records. The researchers tested these models on two very different groups of ships: eight large bulk carriers and seven oil tankers. They found that by fusing the daily crew logs with external information about wind, waves, and currents, they could significantly improve the accuracy of power predictions. More importantly, they discovered that the best method to use depends entirely on how much data is available for a specific ship, rather than on how similar that ship is to others.

The core of the challenge lies in the nature of the data. The noon reports provide a daily average, but the ocean is constantly changing. A ship might encounter a storm in the morning and calm seas in the afternoon, yet the logbook only records one number for the whole day. The researchers realized that to predict the engine power accurately, they needed to reconstruct the missing details of the day. They did this by taking the daily report and "upsampling" it, filling in the gaps with data from public sources like satellite weather maps and ocean current monitors. This created a much richer picture of the ship's journey, effectively turning a single daily note into a detailed timeline of events. However, simply feeding this new data into a standard computer model did not work well because the model still lacked the true, minute-by-minute power readings to learn from. The researchers had to invent new ways to teach the models using only the daily summaries.

Their first approach, designed for ships with very little data, used a technique called transfer learning. Imagine a student who has already learned a subject in great detail and then tries to teach a new student who has only a few notes. The researchers took a model that had been trained on the high-quality sensor data of one ship and used it as a starting point for a different ship that only had daily logs. By freezing the early layers of the model that understood general physics and only retraining the final layer with the new ship's limited logs, they could transfer knowledge effectively. This method proved to be the most robust for ships with almost no history, allowing them to make reliable predictions with as little as one month of data. It worked well enough to demonstrate cross-fleet transferability between the specific bulk carrier and oil tanker fleets tested, though the study noted that this was evaluated using a single source vessel and generalization to other vessel categories was not assessed.

For ships with a moderate amount of data, specifically between two and four months of logs, the researchers found that a different strategy worked best. This approach treated the daily log as a puzzle. The model was asked to do two things at once: predict the engine power and try to reconstruct the missing high-frequency details of the day based on the daily average. By forcing the model to understand the underlying patterns of the day to solve the reconstruction task, it learned a better representation of the ship's behavior. This method, which the researchers called reconstruction-guided learning, provided a stable middle ground. It was more reliable than the transfer learning method for ships with this specific amount of history, offering a balance of performance and stability without needing a "sister ship" to learn from.

When ships had a long history of logs, spanning six months or more, a third approach became the clear winner. This method, known as aggregate loss supervision, changed the way the model learned. Instead of trying to predict the power for every single moment of the day, the model was trained to ensure that the average of its predictions for the day matched the actual daily log entry. This allowed the model to learn the overall trends and patterns of the ship's operation without getting confused by the noise of missing data points. With enough historical data, this method produced the most accurate predictions of all, outperforming the other two approaches. It demonstrated that with sufficient data, the model could learn to generalize the ship's behavior so well that it could predict power at a high frequency, matching the detail of the expensive sensor data, even though it was only trained on the daily logs.

A surprising finding emerged when the researchers looked at how to choose the right method for a new ship. They had expected that the physical similarity between ships—such as their size, engine type, or speed—would be the key factor. They built a complex index to measure how similar two ships were, hoping to use it to guide the selection of the best model. However, the data showed no significant link between this similarity and the model's success. Instead, the most reliable predictor of success was simply how well the ship's own daily logs could be used to predict its power before any advanced modeling was applied. If a ship's own logs were already somewhat predictable, the transfer learning method worked best. If the logs were chaotic, the other methods were needed. This suggests that operators do not need to search for a "sister ship" to get a good model; they just need to assess the quality of the data their own ship is already producing.

The study also addressed the practical question of how much data is needed to get started. For the transfer learning method, the researchers found that a ship could achieve reliable predictions with just one month of data if it was being trained on a similar vessel's history. For the reconstruction method, about four months of data were needed to reach a stable state. The aggregate loss method, while the most accurate, required at least six months of data to converge on a reliable solution. These thresholds provide a clear roadmap for shipping companies: if a ship is brand new, use the transfer learning approach; if it has a few months of logs, use the reconstruction method; and if it has a year of logs, switch to the aggregate loss method for the highest accuracy.

In the end, the research demonstrates that the limitations of sparse, daily data can be overcome by fusing it with the right external information and the right learning strategy. The study ruled out the idea that a single model could work for every ship or that physical similarity was the deciding factor. Instead, it showed that the amount of historical data available is the most critical variable. By matching the modeling approach to the data availability, it is possible to turn simple daily logs into a powerful tool for monitoring ship performance. This capability allows operators to optimize routes and reduce fuel consumption without needing to install expensive sensors on every vessel in the fleet. The findings offer a practical path forward for the maritime industry to improve efficiency and reduce its environmental footprint, proving that even the oldest data sources can be revitalized with modern techniques.