Artificial intelligence may be the force that finally brings predictive maintenance to shipping, but don’t expect a fleetwide adoption anytime soon
“Today it is all ‘artificial’ and no ‘intelligence’.” Not a description of the latest reality television offering, but the assessment of modern marine predictive maintenance technologies by Wärtsilä’s vice president of digital product development, Mikko Tepponen. Systems that offer predictive maintenance often seem to go little beyond fault detection, he argues. But it is what happens after detection that is far more interesting.
A current trial with Royal Caribbean Cruises offers a case in point. Under an agreement signed in October last year – itself the extension of a deal struck in 2013 – Wärtsilä will monitor 196 engines installed on 46 cruise ships. The 10-year deal provides the partners the opportunity to delve deeper into the role that artificial intelligence (AI) can play as shipowners target predictive maintenance. A small-scale test is combining machine learning to improve fault detection, with a collaborative application that allows the operator to review faults in context.
Wärtsilä’s work on AI began with its first condition monitoring systems around 15 years ago. It has accelerated this development over the past three years with the acquisition of performance monitoring and optimisation company Eniram and navigation, simulation and traffic management specialist Transas.
"Humans are better than computers when it comes to finding the reasons for complex issues. But for detecting anomalies, machines beat humans every day of the week"
Machine learning algorithms are the building blocks of AI. Using machine learning, anomalies can be detected based on analysing differences in data outputs – for example, discrepancies in readings from engine pressure sensors. The next step is for experts to investigate the anomalies and diagnose the impending fault. Using the raw, number-crunching power of AI in this way means that anomalies are detected much faster than with more traditional, rule-based condition monitoring, which uses a list of engineering principles that quickly becomes unmanageable as systems become more complex.
Mikko Tepponen (Wärtsilä): Aiming for an intelligent footing for predictive maintenance
As Mr Tepponen identifies, it is the step after anomaly detection that is critical and this means that human intelligence will likely remain a key element here for a long time. In the Royal Caribbean pilot, Wärtsilä is testing an application that will allow collaboration between equipment supplier and shipowner to diagnose faults and determine next steps. After the equipment expert has diagnosed a potential issue and included its recommendations, it becomes visible in the collaboration application. The expert then collaborates with the operator to resolve the issue.
So far, Wärtsilä reports that the machine learning algorithms and assigned analysts have been able to detect all potential issues in a timely manner, allowing the crew on board to take actions to prevent faults and optimise performance. That, said Mr Tepponen, is the result of cooperation between machine learning, equipment experts and the shipowner.
“Humans are better than computers when it comes to finding the reasons for complex issues,” he said. “But for detecting anomalies, machines beat humans every day of the week.”
According to Wärtsilä, AI can already be harnessed alongside human analysis to provide beneficial results. Mr Tepponen said that in one case – agreements bind him from disclosing the shipowner involved – the company was able to identify a fault four months before it occurred.
There are other advocates of AI who argue that even today, in what Mr Tepponen colourfully describes as the discipline’s “teenage years”, we can go further. Among them is industrial AI provider SparkCognition, which according to product marketing manager Gabe Prado boasts a “developing” portfolio in maritime.
For Texas-based SparkCognition, AI must be ‘explainable’ rather than a ‘black box’. In simple terms, it must show its workings, allowing operators to understand how it has arrived at a conclusion, instead of demanding they blindly accept its reasoning. This is the challenge which has led Wärtsilä to support the collaborative AI approach, with algorithms showing anomalies and humans analysing how they emerged and what should be done about it. SparkCognition believes these steps can be taken by AI today, but that the user interface is crucial in easing the transition to reliance on knowledge beyond that of ship crews.
“At this point it is still important that the human is in the loop and it is our aim to arm them with as much information as possible,” explained Mr Prado. “For every recommendation that our model makes, we expose indicators giving evidence why the AI made that decision. Say there is a pressure leak somewhere; our model will flag that as something we should look at and elevate the issue, with a list of the top reasons why it has decided this is a problem. That will include a direct reading perhaps and a lot of diagnostics to help operators understand why.”
SparkCognition’s software was recently tested by a shipowner interested in targeting failures that would either cost more than US$100,000 to repair or cause a vessel to lose time on a voyage. It provided the AI company with historical data from a class of vessels – comprising more than 600 data streams for each vessel – and asked it to spot failures in two areas (propulsion and electricity generation) at least two weeks before they occurred.
For propulsion motors, the company’s SparkPredict software predicted failures up to 10 months in advance and identified eight main features contributing to its prediction. For generators, SparkPredict predicted rotor pole failure six weeks in advance and identified three main features that contributed to this prediction. It also predicted a failure in an auxiliary machine for which it had not been given data.
Mr Prado believes that the lack of connectivity on many ships is a main obstacle to the uptake of AI, rather than any shortcoming in the technology itself. The company’s recently announced collaboration with Hewlett Packard Enterprise (HP) will help to overcome those difficulties. The partnership will enable the use of AI in ‘edge’ situations, where connectivity is not guaranteed and cloud-based solutions are therefore not appropriate. The aging bulk carrier could be a perfect example.
"By implementing predictive analytics right at the edge, we empower customers to make critical decisions fast, even in situations where they have slow, unreliable, or non-existent connectivity,” said HP vice president and general manager of converged servers, edge and IoT systems Tom Bradicich.
"For any machine learning model to work there needs to be a rich pipeline of data. And the reality is that a lot of the time those systems don’t exist already on the ship"
But operating on the edge is not the only obstacle, noted Mr Prado. Sometimes the actual lack of sensors installed on ship equipment is a limiting factor. There is, he believes, a kernel of truth in the perception of shipping as slow to adopt technologies that enable predictive analytics.
“The appetite is there but the big challenge is with the data itself. A lot of people we speak to are still taking readings manually, or the data is not centralised. For any machine learning model to work there needs to be a rich pipeline of data. And the reality is that a lot of the time those systems don’t exist already on the ship.”
That is not to say that shipping is averse to new technologies. As Wärtsilä’s Mr Tepponen explained, it is simply that shipping – like many heavy industries – features assets with a long lifecycle. For the next 10 years much of the fleet will be minimally digitally enabled at best. But for those with ships that are already connected and gathering data systematically, AI-driven predictive maintenance is one of the most exciting possibilities in shipping’s digital transformation.