Understand the benefits of artificial intelligence in modern manufacturing today.

Understanding Traditional AI vs. Adaptive AI

Most AI today isn’t really intelligent or adaptive at all… it’s actually a form of machine learning (ML). That’s because all AI solutions center around models, software designed to achieve the desired results, which are then trained with large amounts of data until they are ready for real-world responsibilities.

For example, you can input thousands of pictures of what a component looks like, and, eventually, AI will develop a pretty good idea of what that component is supposed to look like. Then, the model can predict when a component is not as expected, which is when it becomes truly useful in a production setting.

However, as mentioned above, model drift is inevitable. There are four ways that data can change, making models obsolete:

  • Sudden drift: The data changes, with old values suddenly replaced by new values. COVID provided many great examples of this, like predictions about air travel, which suddenly hit zero, and mask and toilet paper sales, which suddenly spiked.
  • Gradual drift: The data begins changing, with new values becoming more and more frequent over an extended period. This is a common phenomenon in economies, like when a competing product launches and becomes more and more well-established over time.
  • Recurring drift: The data changes for a while, then returns to the original values. Seasonal fluctuations, like wrapping paper, ornaments, and eggnog, are great examples here.
  • Incremental drift: As the name suggests, this type of drift occurs when values change incrementally over time, such as raw material prices or employee salaries.

Due to all these factors, even the best model breaks—and when that happens, it can’t fix itself. Realistically, you may be able to use a model for months—or only for weeks, or days. Then you’ll need to shut things down and start all over with the time-consuming process of R&D and model building.

Why Drifts Happen

It’s inevitable that drifts will happen. Here are the top 3 reasons we see daily.

  1. An additions/removal/editing of different tests in the production line test scheme
  2. Changes in the material itself
  3. Changes in how the machine/process behaves (e.g. the machine got destabilized, thus changing the measurements distribution)

Adaptive AI

The Challenges ​

The manufacturing sector has an unfair reputation for being behind the times—but that’s not exactly true. In fact, as cited by a recent Google study, spurred on by global pandemic conditions over the last couple of years, 76% of manufacturing senior executives have turned to AI and other digital-enablement solutions.

The promise for AI in manufacturing is clear. When margins are tight, as they are in most sectors, and production quantities are high, every little bit adds up to huge savings. AI promises to help out in manufacturing by optimizing the use of raw materials, cutting down on human effort, eliminating faulty units, and more.

But for too many companies, the business benefits of AI remain elusive. It simply has not yet solved the problems it’s supposed to and often fails to deliver the promised ROI given the costs and challenges of getting up and running with traditional AI.

Let’s look at the most common reasons AI models fail in manufacturing settings and then explore options to help you achieve the gains promised by AI with less effort, less cost, and zero downtime.

According to the Google study above, AI is finding new uses in such diverse sectors as automotive OEM, automotive supply, and heavy machinery. What kinds of use cases are being adopted?

  • 39% are using AI for quality inspection. This includes production line inspection, quality checks, and visual product inspection.
  • 35% are using AI for production line quality checks, including early detection of equipment failure and predictive maintenance.
  • 36% are using AI for supply chain management, including supply chain optimization, inventory management, demand forecasting, and more.

Analyzing Data

There’s one essential basic fact about AI that must be considered: All of these applications are very data- and compute-intensive. You’ve probably heard the often-cited statistic that 90% of the data that’s ever existed has been created in the last two years. In fact, only about 2% of data created is actually retained long-term, but data volume will probably grow by about 19% per year between now and 2025. And according to a Deloitte report, manufacturing produces more data than other large industries, like communications, finance, and retail.

Certainly, AI wouldn’t be possible without cloud, which has brought down the cost of the compute power necessary to drive business data analytics. That enables organizations to build great models for a range of business applications.

AI Talent Shortage

While more and more manufacturing businesses are investing in AI, the entire field is plagued by a talent shortage. Few AI solutions work out of the box for every industry, so hiring data scientists to customize a solution for your manufacturing setup and product line can add to the cost and demand more talent than the business value warrants; it can also take tech teams away from other tasks.

Model Drift

Usually, an even bigger challenge is that models break once they’re in production. You can’t just build a model and deploy it once. Inevitably, it will break.

This is known as model drift. Something changes in the real world, and suddenly, the predictive value of the model is useless. Or, in some cases, worse than useless: wasting your teams’ time with false positives or not flagging problems until it’s too late.

Given all these challenges, it’s easy to see why companies might fail to see ROI from their investment in AI.

Manufacturing AI

A few stopgap solutions can help address the problem of drift and keep models in operation longer:

  • Drift detection. This stops false negatives, meaning at least you know the model isn’t working. But then you need to shut everything down to fix the model.
  • Model retraining. Every once in a while, you can proactively shut everything down to retune the model. But this doesn’t stop problems from occurring at random and still requires downtime.
  • Model updating. This entails constantly collecting new data and training new models, periodically deploying to production. But this requires ongoing R&D and data science effort, both of which add to ongoing costs.
  • Data weighting. If you give less weight to new data, then the model could remain relevant longer. But this depends heavily on the nature of your data and the drift, and it could fail to alert you to problems in time to actually do something about them.

Despite these temporary fixes, keeping AI in production usually comes down to additional R&D effort and expense, often above and beyond your initial predictions. Which leaves everybody scratching their head and wondering… “Why can’t AI think more like WE do?”

If AI could learn the way people do, not only taking in information but also adapting its responses based on changes to that information over time, it would work so much better. And that’s where adaptive AI comes in.

An Adaptive AI Solution that Works Better

Adaptive AI
Source: Leuven.AI

When real-world conditions change—as they always will—then the model is no longer relevant. It needs to be taken down and fixed, sometimes by changing the model, sometimes by retraining the model with new data.

But in an adaptive AI model, AI can perceive changes in the data and adapt accordingly, incorporating new data to create a model that is stronger and more resilient. Instead of breaking, the model is continuously learning and improving—with zero downtime.

Vanti’s adaptive AI platform, which is production-ready for industrial and manufacturing applications, makes AI simple. Here’s how:

1. Prepare the model

Using data you already have, Vanti’s predictive model gets started automatically. Your model is ready in minutes instead of days or weeks, with no complicated data science or personnel required.

2. Monitor the model

Once in production, Vanti keeps tabs on your model to make sure it’s working at its best. Vanti will alert you to any problems and can disable predictions if accuracy drops.

3. Data adaptation

When drift is detected, as when a feature is dropped or added, Vanti begins backtesting. It attempts to adjust the data structure to account for the new change, thus improving the model without any intervention.

4. Ongoing model optimization

Even when everything is working normally, Vanti is constantly monitoring your data. It uses your live production data to create new models with better predictive performance.

Adaptive AI
Source: Leuven.AI

Vanti Makes AI Hassle-Free

Born in the manufacturing sector, Vanti gives you adaptive AI that understands your unique needs and challenges. Vanti is the only SaaS tool for engineers that provide real-time optimization of your models based on drifts.

Whether you’re looking for early fault prediction, anomaly detection, or simply to ease the stress of manufacturing bottlenecks like quality control stations, get in touch with Vanti to find out how a smart, adaptive AI solution can help.

Click here to schedule a live demo today.