Machine learning touches almost every part of modern life. From internet search engines to recommendations from your favorite online streaming service, machine learning uses algorithms to analyze patterns of behavior, predict future events and propose useful insights.
Importantly, the outputs from machine learning tools improve as the amount of data inputted grows, making machine learning an increasingly useful risk-assessment tool for aviation safety management systems, which examine vast amounts of data to promote safety and reduce risk.
“Machine learning is a work saver and an efficiency tool, it reduces the effort and is more accurate. This is the true value of machine learning.”
Scott Macpherson Founder, TrainingPort.net
“Aviation safety systems rapidly become data-intensive, and with so many millions of data points being produced, this quickly can become overwhelming,” says Scott Macpherson, founder of TrainingPort.net, a company that uses machine learning to create customized risk-based training tools. “This volume of data has been an issue since the very early days of safety management, when we quickly recognized that while SMS was great in theory, it only worked if you could handle the data, both in terms of storage and in terms of interpretation and analysis,” added Macpherson, who was a business aviation operations manager and chief pilot before founding TrainingPort.net.
Greater Data Management Capability
Machine learning’s ability to thrive in a data-intensive environment enables it to excel in risk assessment, says Myles Boone, a developer at Baldwin Aviation, which is creating its own suite of machine learning tools.
“Think of machine learning as an implementation strategy for solving specific problems and quantifying the risk,” Boone explains. “Machine learning, when implemented correctly, can deliver a more data-driven assessment of how much risk you are taking. Machine learning can assess data and make connections and notice trends and patterns in ways a human cannot and provide a better prediction of the potential risk you are exposed to.”
“Machine learning, when implemented correctly, can deliver a more data-driven assessment of how much risk you are taking.”
Myles Boone Developer, Baldwin Aviation
This capacity for comprehensive risk-based, training-needs assessment eliminates uncertainty, too, says Macpherson.
“Machine learning, in an instant, can provide you with recommendations tailored specifically for your operation,” he explained. “This enables operators to justify not just adding topics or increasing their frequency but, equally importantly, allows operators to reduce frequency or even eliminate a topic when the risk is effectively mitigated.”
“I definitely would have adopted machine learning when I was managing a corporate flight department because it would have reduced my workload and made my risk assessment more accurate and more defensible,” continued Macpherson. “Machine learning gives you a certainty that you are not going off track while also improving safety by reducing or removing measures which are no longer necessary. You can choose to accept these recommendations or not, but you don’t have to sit there and wonder if a particular risk is applicable to you.”
For now, machine learning tools are used to address specific queries, but as the amount of aviation safety data grows and innovations like natural-language processing become more common, machine learning’s position in aviation safety could grow in prominence, says Boone.
“Machine learning could potentially find hazards without people specifically asking the tool to consider that problem. As machine learning tools improve, not only could they discover previously unidentified hazards, they also could assess the actual risk of those hazards and determine proper mitigation without human input,” noted Boone.
The evolution of natural-language processing is an important development in machine learning, says Macpherson.
“Aviation has a vast array of national attitudes toward definitions of an accident or incident, what’s reportable and non-reportable, and what’s an injury and not an injury, so taxonomy and nomenclature have been real stumbling blocks for adopting machine learning in aviation safety management systems. But, as natural-language processing has progressed, we’ve seen machine learning reach a more functional level that is able to take phrases, general attitudes and patterns and see in them and extract from them analytical value,” he explained.
“If we can introduce a concept like machine learning that takes in data to help make more exact and unbiased decisions, we can help organizations determine the most appropriate actions needed to meet their risk tolerance.”
Jason Starke Director of Standards, Baldwin Aviation
Eliminating Bias in Risk Assessment
Machine learning also eliminates the conscious and unconscious biases that plague the field of risk assessment, says Jason Starke, director of standards at Baldwin Aviation.
“One of the biggest concerns in safety management right now is the introduction of bias into risk assessment,” said Starke. “Risk assessment is a purely qualitative exercise, and if we can introduce a concept like machine learning that takes in data to help make more exact and unbiased decisions, we can help organizations determine the most appropriate actions needed to meet their risk tolerance.”
However, the benefits of machine learning extend beyond data analysis and risk assessment, says Macpherson.
“I would say that any application of the ability for a machine to emulate human learning and human anticipation is going to be an advantage to you,” said Macpherson. “Machine learning is a work saver and an efficiency tool; it reduces the effort and is more accurate. This is the true value of machine learning.”
Machine Learning is Not a Safety Management Panacea
Machine learning, however, should not be misconstrued as a ready-made solution to safety management, Macpherson adds.
“Machine learning in and of itself is no panacea; it has to be a targeted tool and introducing machine learning to a safety management system without a full appreciation of its purpose could overwhelm the SMS.”
While most machine learning tools will be developed by software-as-a-service vendors, operators still play a key role in creating tools that address their specific needs, says Boone.
“Machine learning is not necessarily a cookie-cutter solution,” notes Boone. “You have to be intelligent about the scenario you are designing the tool for, and you have to be careful about the training data set inputted into the tool to help it learn.
“For every machine learning algorithm that gives you useful information, there’s a million that could give you useless information,” continued Boone. “Or even worse, [they could] steer you in the wrong direction, so operators must be prepared to commit continual attention to a machine learning tool,” he said.
Using Safety Data to the Best Advantage
For Macpherson, an SMS requires some user-culture readiness and maturity before an operator can benefit from machine learning.
“It takes some time for an operator to collect a sufficient amount to start using it effectively,” explained Macpherson. “At that point, the operator will start to contemplate if they are using that data to the best advantage, and this is when the operator should start to look at machine learning to see if it benefits their operation. Tools only work when we perceive a need for them, and until an operator has used its SMS enough to see the need for machine learning, it will be meaningless to them.”
Machine learning tools will only be as robust as the safety culture that supports an organization’s SMS, says Boone.
“If your safety culture has little interest in recording data and if you’re trying to use data to predict outcomes, then a machine learning tool will simply not work for you,” said Boone.
“Conversely, a strong safety culture that gathers the appropriate data and appreciates the usefulness of data sharing will benefit from the addition of a machine learning tool,” Boone added.
Artificial Intelligence vs. Machine Learning
The terms “artificial intelligence” (AI) and “machine learning” are often, and mistakenly, used interchangeably. While they are often employed together, there are important distinctions between these two scientific innovations.
AI is a broad technological concept to give machines and computers the ability to perform tasks that mimic human intelligence. A good example of AI is an automotive manufacturing robot.
Machine learning, meanwhile, is the use of algorithms to process and analyze data, not only to learn, but also to continually improve as more data is processed.
Google’s ubiquitous search engine and its personalized website suggestions are everyday examples of machine learning.
Netflix is another service provider that uses machine learning to generate recommendations, and the success of this approach has prompted the company to use the technology to determine its future programming and billions of dollars on investment “by learning characteristics that make content successful.”