Manufacturers are facing four key challenges when implementing AI into their product lines. Here’s how you can eliminate (or prevent) them using a knowledge-1st approach.
Unfortunately, it's a reality for manufacturers across the world. In the U.S., aging equipment is the leading cause of unexpected downtime. And it's costing industrial companies roughly $50 billion a year (or $250,000 per hour).
But it's not just a major financial gamble—it potentially risks the lives of employees or the end-users relying on your equipment.
This is the reason you adopted AI predictive maintenance technology is because it can improve production by 25% and eliminate breakdowns by up to 75%.
However, now that you're using artificial intelligence and machine learning, it's not driving the results you anticipated. Maybe you're getting indications of potential faults, but no way of identifying the actual part needing replacing.
This delays the maintenance process and potentially leads to wasting time and money fixing the wrong part. But don't blame the AI—it's your approach to "teaching" it.
Data alone isn't enough—let's explore how using a knowledge-first approach can enhance your machine learning solutions.
Large manufacturers are collecting mountainous volumes of data daily. It’s too much for a human to process, which is why AI and machine learning are on the rise. With it, manufacturers can process mounds of information faster than ever.
Not only does it provide 20-20 hindsight vision—it gives a glimpse into the foreseeable future. This empowers manufacturers to enhance existing maintenance operations using predictive maintenance. And when done right, it results in millions in savings.
Now, there are two ways manufacturers are pulling this off:
Take, for example, transportation companies using fleet maintenance software like Cetaris to maintain mobile assets. According to Dr. Guang Sung, AI Engineer at Cetaris, it's common for fleet managers to complete a reliability, maintainability, and availability (RMA) analysis of their operations using AI. Common KPIs to be tracked:
By tracking the above KPIs and benchmarks, it reduces lifecycle costs by:
While this is great news for manufacturers looking to decrease costly repairs and downtimes, it’s not feasible for every business. Particularly, those with insufficient data.
“Christopher-san, I think we at Panasonic have a small-data problem, not a big-data problem.”—Kazuhiro Tsuga, CEO Panasonic (former)
Small data is troublesome for many manufacturers adopting AI solutions into production lines today. That’s because data scarcity presents an immense challenge when training ML models to predict future issues: it’s ineffective.
A World Manufacturing Forum report shows 58% of industrial businesses state a lack of data resources as the most prominent barrier to deploying AI solutions. And this even holds true for global giant OEM companies like Panasonic.
Because most OEM suppliers aim to have three to four defects per million parts. This is too rare to gather sufficient data to train ML models for visual inspections.
The lack of human intellect and bias of AI predictive maintenance platforms is evident when there's no human engagement.
One thing our customers learned through developing their AI capabilities is that they can't solve industrial problems with just a few ML models. They have all concluded that to create commercial value out of AI in the industrial world, they need AI applications that integrate ML and human knowledge and also have human interactions.
— Vinh Luong, Head of Product Management at Aitomatic.
Implementing AI and machine learning into a predictive maintenance tool isn't the challenge. It's how manufacturers build predictive models to enhance efficiency.
Here's a breakdown of the leading problems we see with AI fault prediction in manufacturing.
To make AI and ML work, it needs sufficient data to learn patterns and behaviors. Without this, manufacturers are stuck with an inefficient platform. So they have to spend many cycles (years) to get the data needed to train the AI to operate effectively.
Manufacturers with equipment containing sensors are actively gathering data. Over the years, this amounts to mountainous volumes of information. Unfortunately, this, too, presents a problem with machine learning modeling.
Imagine force-feeding a machine with thousands of interactions and variables (with no insights). This makes it impossible for it to offer definitive remedies to follow.
So some resort to combing through and scrubbing data manually to polish the system. This process may include writing algorithms, seeing the results, and then changing what's necessary.
Rinse and repeat.
It's a tedious method that leads to manufacturers forgoing the use of AI and machine learning technology.
You have field workers, domain experts, and data scientists. All play a critical role in maintaining your equipment. AI makes it easier for these teams to work together, but it doesn't prevent data silos.
Some manufacturing plants have this problem because their AI system doesn't allow cross-department collaboration.
Knowing exactly what the problem is with your machinery is vital to eliminating the threat early on. But this is difficult to do when your AI platform only detects "anomalies."
An anomaly is nothing more than a blip in the system the AI detects as abnormal. It doesn't point out possible root causes. And if it does, the accuracy is subpar. This wastes your field crew's time troubleshooting various components to identify the culprit.
But there's a solution to these common issues: A knowledge-first approach to encode human subject matter expertise and in-the-field information from your teams, then coupling it with machine learning and automate it all.
Let's explore this further.
To eradicate data walls due to insufficient data, manufacturers must use their SME's knowledge to model algorithms for AI. Here's a real-world example:
Panasonic, the electronics company many of us know (and own products from), has a heating and cooling division. And in this division, the brand faced issues with significant unexpected failures with their commercial cold chain equipment.
These refrigerator units sit inside grocery stores, keeping food cold and fresh for consumer consumption. Unfortunately, these downtimes lead to spoiled products, increased maintenance costs, and nightmare logistics scenarios.
At scale, there was a significant loss in money, time, and resources. It was especially problematic for its experts in IT operations, R&D, and engineering to align and resolve the issues. All they had then was an anomaly detection system using their data to flag abnormalities four to six weeks in advance.
It was accurate 80% of the time, but this wasn't helpful knowing that something may happen. Without specifics, they were clueless about the next steps to take to resolve the problems.
So they came to Aitomatic for a better solution.
The Panasonic AI team used Aitomatic knowledge-first app engine to leverage their field operators' expertise with machine learning to build an AI fault prediction system. This project covered over 100,000 units of equipment, running 24-7, leading to better efficiency and $25 million in savings just within the first year.
Here’s a peek at what’s happening behind the scenes:
You can even use varying experts with differing opinions. The experts can "vote" on the best inputs based on various indicators from the AI and visual inspections by field workers. Over time, AI will learn which experts are more accurate and will resort to their input more frequently.
Models themselves don't create value. They need to be a part of a full application that's delivered to multiple stakeholders in the organization—on both the builder and user side.
And they need to be dynamic instead of static artifacts. Aitomatic Knowledge-First App Engine allows AI teams (the builder) to codify the domain expertise, combine it with ML models and automate everything into an application that interacts with operations teams and equipment, manufacturing experts (the user).
This way, the system continuously takes in algorithms and data, and insights from human experts. Plus, receives feedback from the operations team regarding case alerts in the system about the health of the equipment unit. This is a proven formula for success in commercializing industrial AI.
Not only does it makes AI possible, it empowers manufacturers to go to market quicker and puts predictive maintenance systems in operation within days or weeks. Not months.
Put knowledge first in your AI predictive maintenance program
Industry 4.0 isn't the future of manufacturing—it's the present. But you won't see results if you rely solely on machine learning. By incorporating your employees' expertise, you can build a predictive maintenance system that's highly accurate within weeks instead of years.
If you’re dealing with a small data problem, knowledge-first AI is the answer.
Unfortunately, it's a reality for manufacturers across the world. In the U.S., aging equipment is the leading cause of unexpected downtime. And it's costing industrial companies roughly $50 billion a year (or $250,000 per hour).
But it's not just a major financial gamble—it potentially risks the lives of employees or the end-users relying on your equipment.
This is the reason you adopted AI predictive maintenance technology is because it can improve production by 25% and eliminate breakdowns by up to 75%.
However, now that you're using artificial intelligence and machine learning, it's not driving the results you anticipated. Maybe you're getting indications of potential faults, but no way of identifying the actual part needing replacing.
This delays the maintenance process and potentially leads to wasting time and money fixing the wrong part. But don't blame the AI—it's your approach to "teaching" it.
Data alone isn't enough—let's explore how using a knowledge-first approach can enhance your machine learning solutions.
Large manufacturers are collecting mountainous volumes of data daily. It’s too much for a human to process, which is why AI and machine learning are on the rise. With it, manufacturers can process mounds of information faster than ever.
Not only does it provide 20-20 hindsight vision—it gives a glimpse into the foreseeable future. This empowers manufacturers to enhance existing maintenance operations using predictive maintenance. And when done right, it results in millions in savings.
Now, there are two ways manufacturers are pulling this off:
Take, for example, transportation companies using fleet maintenance software like Cetaris to maintain mobile assets. According to Dr. Guang Sung, AI Engineer at Cetaris, it's common for fleet managers to complete a reliability, maintainability, and availability (RMA) analysis of their operations using AI. Common KPIs to be tracked:
By tracking the above KPIs and benchmarks, it reduces lifecycle costs by:
While this is great news for manufacturers looking to decrease costly repairs and downtimes, it’s not feasible for every business. Particularly, those with insufficient data.
“Christopher-san, I think we at Panasonic have a small-data problem, not a big-data problem.”—Kazuhiro Tsuga, CEO Panasonic (former)
Small data is troublesome for many manufacturers adopting AI solutions into production lines today. That’s because data scarcity presents an immense challenge when training ML models to predict future issues: it’s ineffective.
A World Manufacturing Forum report shows 58% of industrial businesses state a lack of data resources as the most prominent barrier to deploying AI solutions. And this even holds true for global giant OEM companies like Panasonic.
Because most OEM suppliers aim to have three to four defects per million parts. This is too rare to gather sufficient data to train ML models for visual inspections.
The lack of human intellect and bias of AI predictive maintenance platforms is evident when there's no human engagement.
One thing our customers learned through developing their AI capabilities is that they can't solve industrial problems with just a few ML models. They have all concluded that to create commercial value out of AI in the industrial world, they need AI applications that integrate ML and human knowledge and also have human interactions.
— Vinh Luong, Head of Product Management at Aitomatic.
Implementing AI and machine learning into a predictive maintenance tool isn't the challenge. It's how manufacturers build predictive models to enhance efficiency.
Here's a breakdown of the leading problems we see with AI fault prediction in manufacturing.
To make AI and ML work, it needs sufficient data to learn patterns and behaviors. Without this, manufacturers are stuck with an inefficient platform. So they have to spend many cycles (years) to get the data needed to train the AI to operate effectively.
Manufacturers with equipment containing sensors are actively gathering data. Over the years, this amounts to mountainous volumes of information. Unfortunately, this, too, presents a problem with machine learning modeling.
Imagine force-feeding a machine with thousands of interactions and variables (with no insights). This makes it impossible for it to offer definitive remedies to follow.
So some resort to combing through and scrubbing data manually to polish the system. This process may include writing algorithms, seeing the results, and then changing what's necessary.
Rinse and repeat.
It's a tedious method that leads to manufacturers forgoing the use of AI and machine learning technology.
You have field workers, domain experts, and data scientists. All play a critical role in maintaining your equipment. AI makes it easier for these teams to work together, but it doesn't prevent data silos.
Some manufacturing plants have this problem because their AI system doesn't allow cross-department collaboration.
Knowing exactly what the problem is with your machinery is vital to eliminating the threat early on. But this is difficult to do when your AI platform only detects "anomalies."
An anomaly is nothing more than a blip in the system the AI detects as abnormal. It doesn't point out possible root causes. And if it does, the accuracy is subpar. This wastes your field crew's time troubleshooting various components to identify the culprit.
But there's a solution to these common issues: A knowledge-first approach to encode human subject matter expertise and in-the-field information from your teams, then coupling it with machine learning and automate it all.
Let's explore this further.
To eradicate data walls due to insufficient data, manufacturers must use their SME's knowledge to model algorithms for AI. Here's a real-world example:
Panasonic, the electronics company many of us know (and own products from), has a heating and cooling division. And in this division, the brand faced issues with significant unexpected failures with their commercial cold chain equipment.
These refrigerator units sit inside grocery stores, keeping food cold and fresh for consumer consumption. Unfortunately, these downtimes lead to spoiled products, increased maintenance costs, and nightmare logistics scenarios.
At scale, there was a significant loss in money, time, and resources. It was especially problematic for its experts in IT operations, R&D, and engineering to align and resolve the issues. All they had then was an anomaly detection system using their data to flag abnormalities four to six weeks in advance.
It was accurate 80% of the time, but this wasn't helpful knowing that something may happen. Without specifics, they were clueless about the next steps to take to resolve the problems.
So they came to Aitomatic for a better solution.
The Panasonic AI team used Aitomatic knowledge-first app engine to leverage their field operators' expertise with machine learning to build an AI fault prediction system. This project covered over 100,000 units of equipment, running 24-7, leading to better efficiency and $25 million in savings just within the first year.
Here’s a peek at what’s happening behind the scenes:
You can even use varying experts with differing opinions. The experts can "vote" on the best inputs based on various indicators from the AI and visual inspections by field workers. Over time, AI will learn which experts are more accurate and will resort to their input more frequently.
Models themselves don't create value. They need to be a part of a full application that's delivered to multiple stakeholders in the organization—on both the builder and user side.
And they need to be dynamic instead of static artifacts. Aitomatic Knowledge-First App Engine allows AI teams (the builder) to codify the domain expertise, combine it with ML models and automate everything into an application that interacts with operations teams and equipment, manufacturing experts (the user).
This way, the system continuously takes in algorithms and data, and insights from human experts. Plus, receives feedback from the operations team regarding case alerts in the system about the health of the equipment unit. This is a proven formula for success in commercializing industrial AI.
Not only does it makes AI possible, it empowers manufacturers to go to market quicker and puts predictive maintenance systems in operation within days or weeks. Not months.
Put knowledge first in your AI predictive maintenance program
Industry 4.0 isn't the future of manufacturing—it's the present. But you won't see results if you rely solely on machine learning. By incorporating your employees' expertise, you can build a predictive maintenance system that's highly accurate within weeks instead of years.
If you’re dealing with a small data problem, knowledge-first AI is the answer.
