A GREENSTEAM WEBINAR SERIES

TURNING SHIP DATA
INTO INSIGHTS

In times of uncertainty,
it’s more important than ever to safeguard your future.

And with the volume of data in shipping growing at a fast rate, being able to turn it into actionable insights is becoming critical for anyone wanting to operate efficiently.

With this in mind, we have designed a short series of webinars to share some of our knowledge and experience – these will consist of brief but insightful presentations followed by questions from the audience.

Please join us in this unique opportunity to hear from our team of experts working on machine learning tools to help solve the shipping industry’s perennial challenge of reducing fuel wastage and how you can harness the latest insights from GreenSteam to improve your bottom line.

WHO ARE THEY FOR?

Operators

Owners

Charterers

Who want to improve their vessel operations and running costs.

WHAT DO THEY COVER?

3 WEBINARS TO WATCH

An introduction to
Machine Learning in shipping
In this webinar, GreenSteam’s Founder and CTO, Daniel J. Jacobsen will take you through an overview of machine learning and highlight why the shipping industry needs to harness this technology to gain real, attainable money saving insights.
WATCH THE WEBINAR
Focus on fouling –
finding the financial sweet spot
This webinar will explore the impact of fouling build-up, the benefits of optimising cleaning and how to identify the most efficient time to clean led by our Head of Performance Management, Jonas S. Frederiksen.
WATCH THE WEBINAR
Your data has value –
maximise its return!
Here we will explore the importance of gathering and analyzing the right data, transforming it into insights, and how we can help improve your data integrity to realize its full potential with Robert Corden, Sales Manager.
WATCH THE WEBINAR

QUESTIONS & ANSWERS

Topics raised during each of the webinar sessions

MACHINE LEARNING WEBINAR 8/4/20

1) How do you calculate Under Keel Clearance (UKC)?

The best source is the echosounder on the vessel – it measures UKC directly. If that is available, we use that. We offer autologging solutions that are easy to install that captures the echosounder measurements and we can work with echosounder data collected by 3rd party autologging equipment. If the echosounder is not available there are other ways of estimating UKC – contact us to learn more.

2) Once you have created a model, is that it? Or do you need to adapt your model for each vessel?

Each model learns to predict the performance of an individual vessel. But it never stops learning – as new data arrives, the model improves.

3) How long does it take to develop a model for a vessel?

That depends on the amount and quality of data available, and the trade of the vessel.

GreenSteam can work with historical and live data and, as you would expect, the more that is available, the quicker the performance model learns.If the data quality is high and the voyages generally short, the AI may be able to learn a performance model in less than a month.

However, if historical performance data is available, the model can process that in less than a day and be ready for operational use almost immediately.

4) If you use historical data, do noon reports have any value?

Yes, although the quality of these reports varies a lot.One of the beauties of machine learning is that in addition to obtaining the most accurate results, you also get estimates of the uncertainty of those predictions.

Thus the true quality of the noon reports is revealed. Based on this, one can see if the results are useful or if installation of autologging solutions and additional sensors, such as torque meters, is warranted.

5) How does Machine Learning accurately predict fouling without access to tank test data and the vessel’s optimal model?

While tank tests based on physical models can give a guide on how a vessel may perform in perfect sea conditions, they are by their nature divorced from real-life conditions, much as car performance figures are only a guide and cannot be replicated once on the road.

This questions should be turned around: why would you rely on tank test data to estimate fouling? That is bending reality to fit a poor model. Instead, you should use a machine learning performance model that has a built-in fouling component. That means it can learn the effects and level of fouling at the same time as it learns the effects of wind, wave, shallow water, wave making resistance, and so on. Such a model is realistic because the performance of a real vessel depends on all of these factors. This is what enables us to separate out the effects of fouling.

FOCUS ON FOULING WEBINAR 15/4/20

1) Can any hull fouling occur when the vessel is out of the water, e.g. dry docking?

It depends on the coating type, how long time the vessel is out of the water, and on the type of organism attached.

Most waterborne “fouling” organisms will die after a period and thus not develop further (the same applies to salt-water organisms being exposed to fresh-water and vice versa). The remains of the dead fouling will still be attached to the hull and need cleaning.

There is also the risk of contaminating the coating whilst in the dry dock, either when the coating is applied to the hull or after/if the vessel is not re-coated from exposure to heat and sunlight leading to oxidation, exhaust fumes, and other windblown contamination settling on the surface. In itself, this will not lead to fouling but can cause the coating to degrade, leading to higher fouling growth rates and inferior coating performance.

One final thing to keep mind is that a coating requires a short period to start working once immersed. A newly painted vessel, if launched and/or moored in an area where there is unusually high fouling pressure, may foul prematurely.

2) Is there a simple read across from the added resistance % to a % increase in fuel costs?

The fuel consumption and, therefore, costs associated with fouling depends ont he speed of the vessel, the loading condition, and the weather conditions.

Our prediction of added resistance due to fouling is a representation of the operational profile of the vessel, and therefore there is no direct translation from added resistance % to % increase in costs in our plots. However, you do get the same plot showing increased cost from fouling out of the total fuel consumed,which shows much more variation due to the dependency on speed, loading condition, and weather condition.

3) Does the model rely on historical data to optimize future maintenance?

Our Discover service is creating a machine learning enabled performance mode lof a vessel. The model is constantly updated when exposed to new data, and thus it is always a representation of the historical performance of the vessel.

As part of the onboarding of a customer and a vessel, we analyze historical data(from noon reports or high-frequency data logger) to establish this historical performance model. If historical data is not available, we commence data collection from noon reports or data logging equipment onboard to create a historical performance model.

4) Do you monitor Coating condition?

It’s an inherent part of the output of the Fouling Analyzer; we monitor the coating performance as an % added resistance from fouling and while that does not necessarily tell something about coating condition itself, the rate of change in added resistance between events usual gives a good indication on the coating performance.

In the future, once our data set is sufficiently big, we will be able to connect visual inspections to the vessel data and provide coating condition monitoring.

5) With small vessels on restricted coastal services here in Australia, upriver berthing seems to have the fouling damaged by fertilizers etc., hence 3 year fouling systems only last 1 year – can this be predicted?

It’s an interesting case, and it would be great to do an analysis. I would expect the decay in performance would be directly visible in the fuel consumption and the added resistance from fouling from the decay in performance and the properties of the waters the vessels operate in - so in that sense, it can be predicted.

6) Are we generalizing the sweet spot across various types of coatings, and do we see an increase in the ‘clean before trim’ or ‘sweet spot’ when biocides like Selektope is used?

Our model treats all coatings equally between cleaning and drydocking because we don’t have sufficient coating data yet. But no doubt, the sweet spot and cleaning advice will depend on the coating type and quality.

7) Are the results of modelling the same with noon data as opposed to sensor data?

We are offering the same types of modelling and services for both low and high frequency data. The results depend both on the frequency and the quality of the data. We see good noon reported data and bad high-frequency data due to low quality sensors, which is why we perform advanced outlier detection and data cleansing, and also output the confidence intervals on our models. With this, we can determine the quality of the models

8) Can the fouling between propeller and hull be distinguished to make targeted scrubbing decisions?

In the current implementation of our model, the relationship between shaft power and propeller power is not modeled. However, this is on our research roadmap and will be part of the model in the future. It is possible to separate the effects of the engine and the propeller when the vessel has shaft torque meters and fuel flow meters installed and combining these with other signals such as speed through water and RPM.

The current model is, however, able to differentiate between the effects of a hull cleaning and propeller cleaning because it will experience a different impact on the power of the two. For that reason, on a vessel where we know there is fouling, the added resistance due to fouling will not go to 0 % after a propeller cleaning.

9) How long does it take you to build your baseline model for a particular vessel,and what data do you need to build the said model?

Typically we would need 2-3 months of data to build models depending on the quality of the data and frequency. But with all machine learning; the more data you feed the model, the better outcome and advice you will get.

The minimum amount of data we need is some basic info on the vessel, such as IMO number, engine information, and event information. Then continuously, the model needs drafts, power or fuel, and the time period in a consistent format we can parse. We can add many other signals to build better models if high-frequency data is available.

We are developing a brand new pooled machine learning model that will make it possible to start providing advice from day one we get the vessel onboard the platform. Trials commencing next month, so stay tuned for more on that development.

10) Is it possible to identify fouling due to different operational modes, such as when slow steaming etc.?

Yes. Our fouling pressure model takes among other signals, the sea surface temperature and speed through water as direct inputs, and thus models all operational modes. For instance, in these times where many tankers are anchored and used as storage units, often in areas with a high fouling pressure from high temperatures and chlorophyll levels in the water, the fouling growth on these vessels will be high.

11) If we had a choice on where to deploy ships, is there a way to know which parts of the world have higher rates of fouling?

The fouling pressure world map is something we are planning on developing in the near future to be used in our models. In general, though, in areas with highsea surface temperatures and high chlorophyll levels such as tropical waters and near river beds with high concentrations of algae and nutrients level, there are higher fouling rates.

12) We note from your graph that above 10% added resistance, there appears to a noticeable jump in resistance after a vessel stopped period. Is it correct to assume the early life coating is more resistant to this jump after the vessel is stopped compared to an old coating?

After an idle period, fouling will typically take a jump, or the growth rate of fouling will increase. If the coating is fresh and its anti-fouling features are still working,the performance after the idle is expected to be better compared to an old coating. The same goes for different types and qualities of coatings.

13) Which are the main sources of data for creating the model, i.e. which data streams are absolutely mission-critical?

We can create our models with both high-frequency data and data from noon data and voyage reports. When using noon reports, we use something called Virtual Sensing, where we model the period between reports in 10 min windows to create an “artificial high-frequency signal”, where we combine the vessel data with AIS and hindcasted data.

14) How accurate will the model predict current level of fouling if the vessel is retrofitted with a completely different type of coating? (for which there is no previous “Average hull performance”)

If we have historical data before and after the dry dock, the model will predict the current level of fouling based on the performance of the vessel after the drydock.

So let’s say the new coating is significantly better performing (i.e. lower frictional resistance compared to the old coating when it was “fresh”), the model would predict the fouling level at 0%. If that is x% better than the best performance of the vessel in the past, the historical data is shifted up by x%. The rate of change of fouling would most likely be different and would be predicted by the model.

What is important to note is that the machine learning performance model of the vessel keeps improving with the new data, and thus the fouling levels will change bit by bit to become more accurate every time the model is trained on new data. If we don’t have data before the re-coating, the fouling level will be predicted at 0%.

15) Have you any experience with cleaning before entering an area of high risk of fouling for a prolonged period to reduce the overall fouling?

Not to my knowledge, but a customer might have done that. If our model predicted some fouling on a vessel and the coating is no longer brand new, I would probably advise a proactive cleaning.

Our general advice is to clean before the added resistance reached 10 % to protect the coating and remove fouling while it is still in its early stages, so if you stay idle for a longer period, you risk macro fouling will attach to the hull. If the idle period is longer than a month and cleaning at the location is possible, I would probably advise an inspection and or cleaning after the first 30 idle days.

16) Can the model identify when a bad scrubbing has been carried out - either too aggressive that causes damage to paint surface or mild scrubbing, which didn’t lead to an effective cleaning?

Yes, the model is very good at this. An ineffective cleaning or even partial cleanings often shows up as having smaller than expected effect. On the other hand, if the vessel was cleaned too aggressively, the improvement of the cleaning is most likely big, but then you will see that the performance quickly decays, and the fouling growth rate is very high. We have several examples of that, which is also why our mantra is to protect the coating because it makes financial sense.

YOUR DATA HAS VALUE WEBINAR 22/4/20

1) How long does it take to install your kit, and does it need a specialist team?

It depends, some of the simplest solutions are in an App format, so take no time at all, while others at the minimal Capex costs can be installed by the crew. And allows you to capture fuel flow and torque readings making a huge difference to the level and quality of data collected, resulting in a step change in the accuracy of models.

Some of the more complex pieces do require specialists to install, but GreenSteam have teams around the globe who can install when your vessel is in port.

2) Is the cost of the data collection options you showed an annual charge?

No, except for the App and the VDR unit it is a one-off cost for the kit. The analysis and modelling is separate and is charged on an annual basis, with prices depending on the services you choose and the size of your fleet.

3) Can your models work equally well with both noon reports and high frequency data?

We offer the same types of modelling and services for both low and high frequency data.

The results depend both on the frequency and the quality of the data. We see good noon reported data and bad high-frequency data due to low-quality sensors, which is why we perform advanced outlier detection and data cleansing and output the confidence intervals on our models.

As a starting point for a new vessel we can determine the quality of the data available and help identify solutions to improve the quality and quantity of data where appropriate.

4) How long does it take to create an accurate model?

That depends on the amount and quality of data available, and the trade of the vessel.

GreenSteam can work with historical and live data and, as you would expect, the faster the rate data is collected, the quicker our performance model learns. If the data quality is high and the voyages are short liner trades, the AI may be able to learn a performance model in less than a month.

However, if historical performance data is available, the model can process that in less than a day and be ready for operational use almost immediately.

5) Where do you get your weather information from?

To create sufficiently high resolution data we have our own in house Oceanographic specialists and marine forecasting system which access forecast and hind cast data from the following sources.

Type

Elements

Source

Meteorological

Wind / Air temperature

NOAA

Oceanographic

Sea-surface height,ocean current, sea water temperature, sea water salinity (hourly data @ spatial resolution of 9km)

EU CMEMS global ocean model

Tides

3-4 km horizontal resolution

OSU Topex / Poseidon Global inverse Solution TPXO Tidal Model

Wave

-

EU CMEMS global ocean and regional model