Forget whatever you’ve seen in science-fiction movies. Artificial intelligence, usually known as AI, is an umbrella term for computer programs that give machines “human-like” intelligence. As far as we’re concerned, it falls into two broad categories:
- Narrow AI
- General AI
Narrow AI is what we have today. A narrow AI works well for specific tasks, for example identifying cat breeds in photographs, but it’s useless in all other areas. Just as you can’t use the camera app on your phone to order something from Amazon, an AI designed to diagnose skin cancer from photographs of moles is completely useless for steering a self-driving car or recommending which movie to watch next.
In the future, we expect to have general AI. General AI will work across a range of areas, rather than being confined to one specific task. We’re not there yet, but in a 2019 survey, 45% of technologists believed we would have it by 2060.
How does AI work?
At the moment, the main technology under the AI umbrella is machine learning (ML). In machine learning, we provide structured and labelled training data, for example 1000 photographs of tugs and 1000 photographs of container ships. The computer analyses the data and learns to tell the difference between a photograph of a tug and a photograph of a container ship.
The main problem with machine learning is that, in most cases, we need carefully labelled training data. Unlabelled data is useless for standard machine learning. Converting thousands of entries in a database to the correct format then manually labelling them is expensive and time-consuming. In addition, machine learning systems usually need several smaller programs, known as models, to solve a problem. For example, you could build a system to look at photographs of oncoming ships and decide what action to take to avoid collision. In this case, one model could locate ships in a photograph and feed that information into the next model. The next model might identify the heading of the other vessel, while a third model would take that data and determine what action to take. You couldn’t use machine learning to build a single model to look at the photograph and recommend a course of action.
Deep learning is a type of machine learning that uses artificial neural networks. The neural network is arranged in layers. Each layer processes the unstructured data, then inputs it into the next layer. Through this process, the system finds patterns in the data and eventually develops a model.
Neural networks accept unstructured and unlabelled data, and they resolve problems end-to-end rather than one part at a time. The downside is that they need a lot more training data and computing power, and they take longer to train than standard machine learning models.
What challenges does AI face in the maritime industry?
Barriers to AI adoption range from fear of the unknown and laws not designed for AI, to a lack of appropriate training data and a shortage of data scientists.
More digitalised companies adopt AI at higher rates than less digitalised companies. This suggests that the digitalisation trend in the maritime industry could lead to wider adoption of AI systems.
What are some examples of AI in the maritime industry?
Even without general AI, AI is creeping into all aspects of the maritime industry. Any repetitive, structured task has the potential to be carried out by a narrow AI model. Marine insurance, Fire detection from CCTV systems, AI-operated tugs, predictive maintenance, and fuel efficiency improvements are all moving towards AI-driven systems.
A study by the National Cargo Bureau found 6.5% of containers carrying dangerous goods had mis-declared cargo. To address this, Maersk is among the companies using AI screening tools to detect undeclared and mis-declared dangerous goods. HazCheck Detect, a new AI cargo screening tool, scans all booking details and highlights suspicious bookings. In the future, the same tool could screen cargoes to identify, for example, wildlife smuggling.
After demonstrating the world’s first fully-autonomous ferry in Finland in 2018, Rolls Royce is now using an AI system to provide deeper insight into the performance of installed ship equipment. This will lead to increased efficiency and reduced emissions.
Every year, 20% of vessels are diverted due to crew illness, and human error (including fatigue) accounts for around 75% to 90% of marine accidents. Communications provider KVH foresees the use of AI for seafarer health monitoring, to reduce accidents and diversions for crew illness or injury.
But illness and injury aren’t the only causes of human error: fatigue, intoxication, excitement and stress also lead to mistakes. Senseye uses high-resolution images of the iris to identify fatigue and intoxication, while Sensing Feeling uses real-time video to identify early signs of stress and fatigue.
What’s next for AI in maritime?
As with any new technology, adoption of AI will be slow until it reaches a tipping point. As adoption of AI becomes widespread, many of the cultural barriers to AI are likely to disappear. For the last decade, the rate of AI adoption across all industries has been accelerating. Just as we’ve become accustomed to email and the internet, we’ll soon take AI systems for granted too.
The bigger question is what impact AI will have on the industry. Maritime legislation, vessel manning, and much more are predicated on having a human in the loop. As autonomous ships become commonplace, we need to ensure that AI works for us.