SailGP pilots AI for preventive maintenance and high availability

With a loud splash, the 15-meter catamaran smacks the waves as it makes a fast turn during a SailGP practice run off the coast of Australia. As it maneuvers, the hydraulics aboard the F50 sailboat work hard to stabilize control surfaces, including rudders, foils, wings, and more, to maintain speeds approaching 60 mph or 100 kph. Electric motors powered by lithium-ion batteries drive the hydraulics needed by the F50 to complete its practice run and move on to the next race in the international sailing competition.

SailGP, founded in 2019, is a global organization where teams from 10 countries compete using identical boats. Racing events span two days, and all technical information about each catamaran is shared with all participating teams. That ensures that race outcomes are determined by the sailors’ skills, not mechanical differences.

Despite the adrenaline-pumping action that hooks sailors, team owners, and fans, this isn’t a story about racing. Rather, this is a business tale about how SailGP boats use high-tech tools, including artificial intelligence (AI), to keep availability high and maintenance costs low.

After all, without working hydraulics, there’s no way to sail the boat—and no possibility of winning the trophy, $1 million prize, and global bragging rights.

The business problem

While individual SailGP teams are driven to win, the broader objective of the global racing league is fan engagement. SailGP is made up of national teams—Australia, Canada, Denmark, France, Germany, New Zealand, Spain, Switzerland, United Kingdom, and United States—each with passionate followers who watch events live or online, engage through virtual reality, purchase clothing and memorabilia, post on social media, and otherwise help the relatively new league flourish and grow.

The most recent complete season, which ended in May 2023 after 11 races, saw Australia as the overall winner, with Denmark named the team judged to have the lowest environmental footprint; sustainability is important to the SailGP organization.

If mechanical problems keep a team from competing in a race, disappointment could lead to fan disengagement. And if a boat fails mid-competition, there’s always the prospect of crew injury or race disruption. Plus, nobody wants to see that sort of news coverage in the sporting press.

With so much at stake, the league’s business executives want to ensure that catamarans are on the water as much as possible for both training and racing, while minimizing the chances of failure during an event. That means expending resources on monitoring the boats’ infrastructure and mechanical operations and using the results of that analysis to fine-tune the extensive maintenance performed after each sail.

Those boats on the water? I call them ‘extreme IoT devices.’ We get a huge amount of information off the boat and then turn that into metrics that we can understand and make better decisions with.

Warren Jones CTO, SailGP

That’s where AI-based anomaly detection comes in. By analyzing billions of data points collected from sensors on the boats, the engineering team can identify parts that might be about to fail and preemptively replace them, just in case.

While parts for an F50 catamaran aren’t cheap, the opportunity cost of an in-use failure is significantly higher.

The technology platform

The F50 racing catamarans, constructed in Warkworth, New Zealand, are so technologically sophisticated that they might be confused for data centers with sails. With only a few exceptions, hydraulics control critical components, such as pitch controls and the rudder.

“Those boats on the water? I call them ‘extreme IoT devices,’” says Warren Jones, CTO of SailGP, pointing out that there are 125 sensors on each boat. “We get a huge amount of information off the boat and then turn that into metrics that we can understand and make better decisions with.”

Indeed, hundreds of sensors and cameras generate telemetry every second. All that data is logged locally, and select information is beamed in real time to SailGP’s UK operations center, where it is used by performance analysts, systems engineers, and operations staff, and to inform fans around the world of conditions on the boat.

Despite the sophisticated computers on each catamaran, the crew is completely in charge. Digital sensors are passive, and sailors operate the boat through switches, knobs, wheels, and other controls—all race-day plan execution is performed by hand. That makes each competition a true test of skill for the six-person crew.

After a race, all data is downloaded from each boat and transmitted to Oracle Cloud. There, the data is shared with each team’s managers and with the league’s officials. SailGP practices full transparency; each team receives all the same information from each sailing—even from competitors’ boats. That helps constantly improve performance and allows for each team to understand exactly what happened on the water.

“We have an open data policy,” says Jones. “Instead of the traditional racing circuit, where data is kept secret, SailGP uses that to go faster. Team Australia can see what Team Britain is doing, and Team Britain can see data from Team USA.”

That’s another reason data collection is so essential.

“This information about what’s going on in the boat is highly critical to how the teams perform and how they strategize,” he said. “It’s a constant: more data, more information, more people.”

In addition, the engineering team uses telemetry data to detect faults within each boat, going beyond routine post-sailing maintenance to address unexpected issues or unusual readings. That’s where AI-powered anomaly detection, which was initially deployed to monitor each craft’s numerous hydraulic actuators, comes in.

Hydraulics under pressure

On watercraft, as on land vehicles, hydraulic systems work by pumping specialized fluids into cylindrical pistons. When more hydraulic fluid is pumped into the piston, the piston is pushed out, moving a control surface in one direction. When fluid is released from the cylinder, the piston retracts, and the control surface moves in the opposite direction.

For those who aren’t familiar with sailboats, you can see the same activity when hydraulic actuators on an airplane move the flaps to control the plane’s lift.

When too much hydraulic pressure is applied, however, bad things happen. Pressure seals inside the piston may begin to leak. The flexible hoses delivering hydraulic fluid stretch or break. Hose couplings and fittings can weaken. A piston rod might deform or jam. Best case, performance suffers. In other cases, that hydraulic component stops functioning. In the worst case, catastrophic failure leads to a spray of slick hydraulic fluid all over the boat and the crew, which can not only be dangerous in itself on a fast-moving vessel that might be on rough seas but might lead to cascading failures.

To maintain the control positions set by the crew, hydraulic systems are constantly increasing or decreasing in pressure.

“We use pressure sensors to manage the load within the design constraints of the boat,” says Scott Babbage, head data analyst for SailGP. “From a system safety point of view, we have a pressure vessel, and the pressure that we’re storing in there needs to be measured. We have a pressure sensor, so when the pressure accumulated drops to a certain level, we recharge it with the pump, and when it reaches its upper limit, we stop charging.”

Sounds easy, Babbage says, but that’s deceptive.

“If a pressure sensor is failing, we’re getting incorrect readings on those sensors,” he said. “We don’t know exactly when to fill and when to stop filling. We potentially could overfill, over pressurize an accumulator, and have one fail.”

That could be catastrophic and, given that sensor faults are caused by hard use and a hostile environment, staying ahead of the problem is an ongoing battle.

“There’s corrosion with the salts in the water and the g-forces when the boats hit the water,” says Jones. “And sometimes, some parts are a little lower quality than others—we call them ‘Friday afternoon’ parts.”

In other words, there are many opportunities for hardware to fail.

In all but the simplest cases, crews are unable to repair the problem, which means the catamaran will be harder to control at speed. There are manual overrides that, in case of hydraulic failures, should enable the boat to get back to shore. But even so, a catastrophic failure creates a safety hazard for that boat and crew and potentially for other boats maneuvering nearby.

From a business viewpoint, the boat is out of the race and out of the running for prize money, fans are disappointed, and bad publicity might affect the popularity of the sport. In addition, repairs tend to be costly and time-consuming, which can affect the team’s ability to train and compete.

As with all major sports leagues, SailGP operates on a tight schedule.

Patterns of predicted failure

Sensors on each F50 catamaran measure the pressure on each hydraulic line and record the position of each piston rod. Each hydraulic device also has a switch that tells the system to turn the pump on and off; readings from switches are captured and logged.

In a hydraulic system, pressure is not constant. Take a control surface that’s slapping against the sea. When that surface is pushed harder, the piston rod is compressed, and pressure increases slightly. For any given amount of pressure in the hydraulic cylinder, the position of the piston rod will normally jitter slightly.

F50 sailboat telemetry systems capture billions of readings. If telemetry reveals a pattern that doesn’t match expected ranges, there’s a good chance something is amiss. Perhaps a switch is beginning to malfunction. Maybe a hose or coupling needs to be replaced, or a pressure sensor is miscalibrated. Anomalous readings are early predictors of failure, if not in tomorrow’s race, perhaps in the next few weeks. It’s less costly—in every way—to replace or repair a part before it breaks.

“It’s been quite useful to see when pressure sensors are failing,” said Jones. “Often they don't fail completely, but they'll start to measure erratically.”

Another challenge: It may not be that a particular reading is off, but that relationships between associated sensors are slightly abnormal.

How will you find such a needle in a haystack? For humans, having too much data is a problem. Fortunately, for AI-based anomaly detection algorithms, a seemingly overwhelming quantity of data is a gift that keeps on giving, especially for this small engineering team.

“Scale is the benefit here,” says Babbage. “We have a small team of engineers servicing a lot of boats. When you think of Formula One racing, they have probably 20 engineers working on each race car. We have a team of four engineers who are looking after all our boats.”

Enter anomaly detection

After each sailing, data is extracted from the F50, often via a compressed binary file, containing billions of data points, including each sensor’s identification tag, the sensor’s reading value, and a timestamp.

A simple extract, transform, and load (ETL) operation converts the binary data into a comma separated values (CSV) format and then uploads the data into Oracle Autonomous Data Warehouse. Once data is in the cloud, Oracle Cloud Infrastructure (OCI) Anomaly Detection takes over, scanning data associated with the hydraulic systems—nearly 400 million readings per race.

The AI service has been trained, using machine learning, on the relationships among the various components of each hydraulic actuator. It generates a report detailing potential problems with each boat, including the specific hydraulic actuator, the anomaly detected by the service, and the potential severity of that anomaly. This data allows the engineering crew to prioritize maintenance activities.

“Anomaly detection takes around six or seven hours to do a deep dive into that data,” says Jones. “We wake up in the morning, and it tells what our technical team needs to swap out. It’s very efficient.”

The results

Does all this high-tech analysis work? In a word, yes. According to SailGP, OCI Anomaly Detection finds one fault on average after two days of sailing—which includes shakedown cruises and crew training days as well as the races themselves. Since adopting cloud-based anomaly detection, SailGP reports that no hydraulic malfunction has impacted a boat or affected any race activities.

Fans are happy, crews are safe, and the SailGP organization can operate its full fleet of boats, in part due to behind-the-scenes technology in the cloud.

“I’m really, really proud of what we do with Oracle,” says Jones. “Every challenge that I've ever brought to them, Oracle has said, Hey, we’ve got this.”

Photographs: Courtesy of SailGP

View more Oracle Connect articles