A few weeks ago I wrote a myth-busting post to tell the truth about using AI in industrial
settings. But why is this coming up? What is different about industrial situations, and why can’t we use the cool AIs there?
The data itself is difficult
Industrial data is very lumpy and difficult to stitch together. AIs don’t operate well when the data come from many different types of sources. The need for a clean data domain in which all the predictors are accessible, normalized, and recognizable in the data practically rules out industrial settings. For example, can you still tell what device or vehicle or machine this data was extracted from, and what its location was at the time, let alone all the other information relevant to deciding its state (weather, temperature, history of repairs)? The output of assembling and stitching all that data together, along with an ontology that makes sense of it all, is what we call the digital thread
- and it’s critical to creating a solid data foundation.
Access to the data may be blocked
In the years since the first, second, and third industrial revolutions, the techniques, protocols, formats, and storage mechanisms of the data have changed dramatically. The systems and equipment are sitting behind firewalls, or in air-gapped buildings, or disconnected in combat zones, or in otherwise un-serviced areas. And if physical access is not the problem, then often getting security or permission to acquire the data is.
The predictions, if wrong, can be deadly
AI gets it wrong frequently (e.g., YouTube’s fact-checking mislabeled the Notre Dame fire as a possible terror attack
). When this happens on YouTube it can be inflammatory, insensitive, and potentially riot-inducing. When AI-aided automation performed in industrials goes wrong, it can be truly sensitive, dangerous, or deadly, as illustrated by the disastrous Boeing Air-Max situation
. Lengthy governance and certification processes beyond the development of automation is required. For this reason, our foray into AI for industrials must be measured, and careful.
What can you do now to pursue the promise of AI for Industrials?
Understandably, even Industrial companies want to pursue the value of pervasive data intelligence that is promised by AI. Here are a few steps and guidelines:
- Don’t skip the critical step of building your data foundation. AI cannot be done without having a solid set of data. Invest in your Digital Thread.
- Keep a lookout for specific AI applications that might be adopted from other industries.
- Understand what use cases are really feasible. AI/deep learning is being used most successfully to replace simple, repetitive human perception tasks that don’t involve sensitive or dangerous conclusions. Imagine these productive uses in industrial situations. For example:
- Test automation / test robotics: AI is beginning to be used to dramatically speed up the final test stage in consumer electronics assembly. Deep learning predicts the optimal way to navigate a test suite for the product by identifying tests that are likely to be redundant or unnecessary.
- Optical quality detection / classification: An application of image recognition that has been applied successfully to determine whether a batch of product should be classed as low quality, medium or high quality by detecting and characterizing scratches, smudges, and dents to surfaces. And it can do this more quickly and reliably than human inspectors.
- Safety inspection: Another optical detection application – railway operators are using vast suites of image recognition and classification processes to speed up safety inspections needed periodically on the many parts and components of rail cars.
- Use the least complicated method to start; move to other things only as the need becomes real. Unless you are just building your resume, pursuing AI and Deep Learning for the sake of it is not the first technique you should apply. Remember the Pareto Principle: simple math can reveal the low-hanging, high-value applications to pursue. Use the simplest technique you can to solve the problem.
- Use a collection, or “ensemble,” of models and a variety of techniques, including AI, that work together on various pieces of the problem:
- Automate simple data management tasks, like source-to-target probabilistic mapping.
- Perform pattern recognition that generate features that involve automating and improving human observation (e.g., visual, auditory, text).
- Feed AI-created features into more traditional statistical / machine learning models.
- Use “human-in-the-loop” automation to assist with creation of labelled datasets (my colleague will be explaining this concept in a future blogpost).
The Future is Here
It goes without saying that you must make sure to complete all these critical steps for success with analytics in Industrial
. But along the way, you may spot opportunities to use current AI tools to automate a portion of your industrial operation that help you learn and progress incrementally, without putting lives in danger or threatening the operation.
The future is already here… it’s just unevenly distributed.
Cheryl Wiebe is an Ecosystem Architect in Teradata’s Data and Analytics Strategy team in the Americas region, and works from her virtual office in Southern California. Her focus is on the business, data, and applications areas of analytic ecosystems. She has spent years working with customers to help create a digital strategy in which they can bring together sensor data and other machine interaction data, connect it with other enterprise and operational domain data for the betterment of the reliability and efficiency of large equipment, large machinery, and other large (and expensive) assets, as well as the supply chain and extended value chain processes around those assets.
View all posts by Cheryl Wiebe
Industry-spanning programs, such as Industry 4.0 and others that address enterprises in their goals to “go digital” in a journey to the cloud, are where Cheryl focuses. She helps companies leverage traditional and new IoT settings to organize and develop their business, data and analytic architectures. This prepares them to build analytics that can inform the digital enterprise, whether it’s in Connected Vehicle services, Smart Mobility, Connected Factories, and Connected Supply Chain, or specialized solutions such as Industrial Inspection / Vision AI solutions that address needs to replace tedious work with AI.
Cheryl’s background in Supply Chain, Manufacturing and Industrial Methods stem from her 12+years in management consulting, industrial/high tech, analytics companies, and Teradata, where she’s spent the last 18 years.