MARITIME MACHINE LEARNING SYSTEMS
Bureau Veritas’ guidance note on machine learning systems is helping the maritime industry understand this rapidly evolving technology.
Machine learning systems are a subset of artificial intelligence that focus on pattern recognition. Machine learning systems “learn” patterns that exist in training data and are able to make inferences about new data. This enables them to make predictions, or even decisions, without explicit instructions.
In a maritime context, machine learning systems have several promising applications. Installing these systems on vessels can improve efficiency, security and cost optimization, among other uses. Vessels around the world already use machine learning systems and uptake is expected to grow as more of these tools become available. However, guidance is needed to ensure that shipowners deploy these systems responsibly.
Understanding the risks of machine learning systems
There are a number of risks associated with machine learning systems installed on vessels, including biased data, misinformation and cybersecurity vulnerabilities. In general, machine learning risk pertains to the data on which the systems are trained, and how the systems use it. Mitigating risk is essential to ensuring that the machine learning system does not demonstrate unexpected or unexplainable behavior. Non-addressed risks can also create cybersecurity concerns for maritime assets.
Bureau Veritas brings clarity to maritime machine learning
Bureau Veritas Marine & Offshore is working to ensure that machine learning systems in the maritime industry are safe, secure and transparent. That’s why we recently published our new NI 692 guidelines on the use of machine learning systems in a maritime context. These guidelines seek to provide a structured framework for the responsible and effective use of machine learning systems on vessels and in other maritime contexts. They align with existing international and EU standards.
Supporting the development of safe machine learning
Bureau Veritas Marine & Offshore is also supporting the safe development of new applications for AI and machine learning systems.
Our Approvals in Principle ensure that maritime software applications are secure, transparent and in compliance with relevant legislation. We look at how systems work, what data they use and how they process it to uncover any potential bias and recommend mitigation strategies. In addition, we work to ensure the reliability and robustness of these machine learning systems.
Bureau Veritas Marine & Offshore is also currently developing a recommendation on data quality.
As AI and machine learning systems continue to evolve, Bureau Veritas will continue to stay at the forefront of this technology to ensure their responsible deployment in the maritime industry. We are by your side for a more digitalized maritime future.
Technical Advisor – Data & AI
Bureau Veritas Marine & Offshore
“The maritime world is rapidly digitalizing and machine learning systems can potentially revolutionize vessel operations. Bureau Veritas is focused on ensuring the safety, security and efficacy of these systems, to support a more connected industry.”
-
What are the current use cases for maritime machine learning systems?
Currently, machine learning systems are mostly being deployed on marine assets, although applications for offshore assets are also under development. Existing use cases include:
Object identification, by helping a vessel distinguish whether there are objects near it and how to navigate around them, through the use of cameras or sensors
Route planning and optimization, including “just-in-time” arrivals o Gas consumption monitoring and other energy efficiency measures
Mine detection via drones for military operations.
- Remote operation centers can also be equipped with AI systems to help optimize operations at the fleet level.
-
How can machine learning data create risks?
- The quality of the data set on which machine learning systems are trained is of the utmost importance.
- One issue is “overfitting.” If the data set is too small, the system will not properly “learn” the relationship between the data points. This means that if the system comes across data that is different from its training set when deployed, it will perform badly.
- “Underfitting” is also an issue. This is a scenario where the system is unable to learn the relationship between data points, due to an overly simple model. This generally results in a high error rate for the system in both training and actual deployment.
- Insufficient cleaning of the data set can also cause problems for the system. If the data set contains too much irrelevant information, it can prevent the system from learning what truly matters for the task at hand.
- Some systems are equipped with “reinforcement learning” where actions are optimized to earn rewards (generally in the form of positive numerical values). Bias can occur if the system prioritizes accumulating rewards over achieving the intended goal, which may lead to bad incentives and undesirable outcomes.
-
Why is bias a problem in machine learning systems?
- Bias is a key issue that needs to be addressed in machine learning systems, as biased data can have serious performance and even safety impacts.
- For example, if a computer vision system is trained exclusively on images captured in ideal weather with high visibility, it may perform poorly in real-world scenarios such as fog, storms or night-time conditions.
- Or a system might optimize a vessel’s route toward a specific destination simply because the destination was more present in the data set, not because it’s actually the optimal journey.
- Bias needs to be addressed at every step of a machine learning system’s development process.
-
What is the current regulatory landscape for machine learning?
- Since this technology is still relatively new, relevant regulations are largely lacking.
- The EU Artificial Intelligence (AI) Act recently entered into force and potentially impacts EU vessels. As its scope does not comprehensively cover the maritime sector, it provides only limited clarity.
- Several gaps in the regulatory environment remain to be addressed.
-
What is the EU AI Act and how does it impact the maritime industry?
- The EU AI Act is the first-ever comprehensive legal framework on AI worldwide. The Act lays out risk-based rules for the development and deployment of AI in the European Union. The rules aim to guarantee trustworthy, secure and human-centric AI.
- The key point of the Act is how to ensure trustworthiness in AI systems. As a result, there is a focus on adding many layers of security, risk assessments and ways of monitoring the systems using different evaluation criteria. Human supervision remains paramount in guaranteeing trustworthiness.
- The Act explicitly addresses the maritime industry in only a few places. Its main direct relevance to the industry pertains to its rules for equipment with integrated AI in high-risk scenarios. These rules do impact the deployment of AI in maritime assets, but considerable regulatory gaps remain.
-
What impact could machine learning systems have on crews?
- Bureau Veritas’ new guidelines (NI692) stress the role and importance of human oversight in ensuring the proper functioning of these systems.
- Crews should be given sufficient resources and training so they understand machine learning systems on board, how to interact with them and how to recognize bias. Crews need to know what these systems can do and what they cannot.
