The journey to a smart factory, what a smart factory looks like and how AI tools can be integrated into engineering were all questions that were answered in the latest episode of IoT Unplugged featuring Rachel Johnson, Principal Product Manager at MathWorks.
MathWorks’ own expertise in creating software to “help engineers and scientists do their best work,” as Johnson termed it, has provided them with a deep understanding of driving AI-predictive maintenance and anomaly detection solutions, which was a big part of the conversation.
What is a smart factory?
What defines a smart factory, she explained, is not necessarily clear-cut: “A lot of modern factories are on this spectrum of what we would consider a smart factory, and they’re getting smarter all the time,” said Johnson. “When we talk about a smart factory, we’re talking about combining traditional, physical manufacturing with advanced digital technologies … like AI, IoT, Cloud Computing, advanced analytics.” She added that the “goal” was to improve automation and efficiency.
IoT devices like sensors are being more commonly deployed in factories because of their ability to gather data and provide it for further analysis, which has enabled predictive maintenance.
Predictive maintenance represents one such area where factories have benefited from the application of smart technologies, as the days of old when a piece of equipment would break down, need fixing and in doing so, trigger production shutdown, are becoming less commonplace. Instead, engineers can collect the data, analyse it, and based on the information they are provided, decide whether or not machinery needs maintenance.
“This is all about preventing unexpected failures and being able to investigate problems earlier,” Johnson summarised.
Using data properly
The excitement around what new technologies like AI make possible may have a tendency to overlook particular challenges. In using AI to sort through data being collected in a smart factory, Johnson stressed that the challenges she sees in her line of work are data-related.
“There’s the issue of not having enough data, or not having collected it in the right way. [This] is easily the biggest blocker for companies on this path right now. Even if they have enough data, being able to understand what it means and how to define what anomalies look like is the problem they’re trying to solve,” she explained.
AI algorithms need to be trained on data in order to understand, for example, when a piece of equipment is showing signs of wear and may require predictive maintenance to prevent an unexpected failure. This is where historical data sets come in.
“One of the big challenges around anomaly detection is having the right type of data that represents the anomalies you want to predict,” said Johnson. “As an example, we have a pump … we don’t want it to fail. We’ve been performing regular maintenance on it throughout its entire lifespan. So we don’t collect a lot of anomalous data … What happens is we have this large data set of normal operating data and not a lot of examples of what the problems might look like.”
AI in engineering
Johnson said she saw engineers already beginning to consider the best way to set up these smart technologies in a factory so they get quality data, so to speak. “We’re seeing folks start to think about these questions … on the design side, where do I put IoT sensors on my equipment so that I can design an accurate AI algorithm to uncover the information I need?”
AI-driven anomaly detection is one of the projects Johnson sees people moving into, as “what looks different in my data” is a common question to ask and a general problem for companies.
Designing and deploying an AI-based anomaly detection system has “four parts”, according to Johnson: gathering data and defining the problem; exploring and pre-processing this data, “the goal here is to get that data into a clean slate so we can use it to train an AI model,” said Johnson; training the AI model and validating it; ad, finally, deploying and integrating the system.
There is no one best AI learning technique for anomaly detection, Johnson explained. “It’s important to experiment with different training approaches, to find the best fit for the type of data … The best technique is going to depend on the type of data you have.”
She concluded, “Engineers who understand the right features to extract and then experimenting with [the] different techniques is the key.”
The podcast episode featuring Rachel Johnson can be listened to on Spotify, Apple Podcasts, and on IoT Insider.
There’s plenty of other editorial on our sister site, Electronic Specifier! Or you can always join in the conversation by commenting below or visiting our LinkedIn page.
