Create Action From Data Analytics

Artificial intelligence, machine learning, and advanced analytics help address critical challenges in quality and productivity for manufacturing.

Manufacturers are finding ways to gather, analyze, and use data analytics to solve problems before they happen and drive upgrades in uptime and quality.

There was a time, not so long ago, when manufacturing companies had systems that were stuck in silos with little or no communication between humans and machines. Data was difficult to gather and shared only through spreadsheets and by word of mouth, making problems almost impossible to predict.

Today, not only are humans and machines communicating, but machines are talking to each other. Massive amounts of raw data are flying through the air at warp speed revealing all kinds of intelligence. It’s one thing to have the data. It’s another to be able to analyze the data. But the secret sauce, according to Nathan Oostendorp, co-founder and chief technical officer of Sight Machine, San Francisco (sightmachine.com), is dissecting the data, identifying the right data, and then using the knowledge gained in a meaningful way to create actionable results.

“The companies that are seeing results are those that are discovering ways to build a very connected digital ecosystem,” Oostendorp said. “Most new systems have a computer as part of their makeup, so there is some data available. There are more things being censored and inspected. Controls are getting tighter rather than looser. There are a lot of changing specifications in manufacturing, and there is a constant pressure to improve. We have realized a lot of gain from Toyota and Six Sigma practices that originated in the 1970s and kind of came of age in the 1980s and 1990s. But we are at the point now where doing a three-to-six-week workshop on every single problem is just much too slow and much too expensive to resource compared with the number of problems these manufacturers want to attack.”

Sight Machine was founded in 2011 with the goal of finding ways to apply big-data infrastructure manufacturing. Purpose-built for discrete and process operations, its platform uses artificial intelligence, machine learning, and advanced analytics to address critical quality and productivity challenges.

Solving Problems with Data

It is critical for data scientists to communicate closely with subject-matter experts, i.e., people who work directly with assets.

“A large amount of the effort required in data modeling is not just in the collection of data, but in transforming it into a useable standardized format,” Oostendorp said. “The data analyst must go back and forth with a subject-matter expert on that asset or on that process. The data must make sense to the person who is going to use it. Part of it is having an original data source against which to compare new data. Sometimes you find your system is right and other reports have been wrong for awhile because of a pass-around Excel template that hasn’t been updated in several years.”

Validation is a key element in finding the right data. Oostendorp said that when you are trying to solve problems, there are two avenues for using the data.

A basic statement and hypothesis. “You generally start with something like… ‘Our product loss is due to this particular type of defect. What are the costs for that?’” Oostendorp explained. “You can take your hypothesis and test it in the system and compare it with the population of products last month during a spike and the population of products during a previous month. You look at the properties of the machine or the properties of the material that passed and failed on this particular type of test.”

Focus on the data, not the hypothesis. “If you have all of your data modeled, you can let the computer do a lot of the basic imprints and ask it, ‘For these two populations what are the largest changes in absolute value or what are the largest changes in variability? What are the things that were out of control or that were anomalies during this timeframe?’” Oostendorp explained. “This way you focus more on the data rather than the hypothesis. You may involve a subject-matter expert. Often you discover things you wouldn’t have considered to check in that way. We have process data and discovery tools that are intended to look at the data first and not through a lens of what someone thinks might have happened. It focuses on things not associated with variation process like anomalies or things that are out of control. This is a better way to do exploratory analysis.”

Challenges with Data Analytics

Manufacturing data can be particularly difficult to gather, analyze, and utilize in meaningful ways, according to Oostendorp. “There are a couple of aspects that make it more difficult,” he said. “Problems have existed in various forms for companies in how to gather data, and how to compile the intelligence. Manufacturing is kind of unique in that the systems are just older and have been operating for a longer amount of time. It is very frequent to have systems that are deployed on the plant floor that just don’t ever get upgraded. Whether it’s IT resources or the ‘If it ain’t broke, don’t fix it’ mentality, these are just things that you have to cope with.”

Even when looking at the whole system, some facilities may have different versions of the same software or different operating systems running at the same time. This is a big challenge with helping plants use data analytics to solve problems.

“Most of these systems are not designed with data extraction in mind,” Oostendorp said. “Some have data-storage capabilities, but they tend to be oriented toward a black-box analysis where the data can be retrieved later to figure out what happened. If you need something when something goes wrong you can get to it, but there are very few systems built with data collection and data integration as a first-class priority in system design already built in.”

This is why using data-analysis platforms can help facilities.

“When I look at our platform and the field of manufacturing and data analytics in general, it drastically reduces the scope of time to fix these problems to prepare the data in a way that makes it always available for analysis,” Oostendorp said. “Then it is important to have tools that allow data analysis to be performed on the fly as opposed to with a great deal of investigation and preparation.”

As companies buy into using data analytics to solve problems, mistakes are often made. The biggest one, according to Oostendorp, is not reacting to the data with a sense of urgency.

“Companies may do a couple weeks of research on the Industrial Internet of Things, download a reference technology chart, and then ask their IT department to put together something in the next 9 to 12 months,” Oostendorp. “We see this happening over and over again. Everyone drastically underestimates the complexity in data acquisition, the complexity in the technologies even if it’s centralized in something like a data lake, and the amount of effort it takes to go from a raw-data product to extracting a meaningful outcome using analytics. A leadership team may decide to have a digital transformation, and then decide to build it themselves. It really does look a lot easier than it is. These are very complicated systems where you have a number of different issues with connectivity or validation issues. This is the most common mistake manufacturers are making today.”

Data-Modeling Tools Can Help

The Sight Machine manufacturing analytics platform, for example, gathers raw data from different sources and creates a universal model that applies across different manufacturing problems. It can reveal a discrete time on an asset and discrete transactions between materials and assets.

“With every step that a piece of material goes through the system, you want to be able to track every second of time that you are performing a value-add step to the materials you want to be able to track,” Oostendorp said.

He continued, “The output is kind of like a ledger. You now have a log of everything that happened on that asset and every piece of material. What we have found is that you get some useful information by just having the format for a single asset. The real power comes when you have an entire process or an entire factory modeled in that way. Now you can ask more interesting questions because you have all the data organized in a way that it can be connected and investigated.”

Creating Actionable Results

The key to a manufacturing-analytics platform, according to Oostendorp, requires more than just acquiring knowledge. You must also establish a workflow that accumulates the data, along with putting  together an action plan that will test your hypothesis and determine the outcomes.

“This is a significant part of the work,” he said. “Every engagement has its own complexities, but the technical part is pretty routine. We can take data, put it in data models, and build reports. Our goal is to get the data integrated into the system so the plant can use it as part of their morning meetings. They can identify something that doesn’t look right and then investigate it. Soon, the companies who use data in this way get to a point where teams are constantly building action plans where they identify things like: What did I observe? What do we need to do to correct it? Who is the owner of this? What are we going to do to look for the outcome? Answering these questions and then prioritizing the answers helps them to get in the habit of always realizing value.”

It’s not just about having the software. The mindset must become a part of the company’s culture. 

“This is where data analytics changes the practice of manufacturing,” Oostendorp said. “The improvements become data driven. There is a shift that happens where the first reaction is to ask questions about the data in order to get an actionable outcome. When I look at our successful customers, this is the difference. It requires a certain type of company culture. It is difficult to mandate this type of thing. It needs to happen naturally and evolve. It involves a certain amount of exploration and tinkering to get the routine right. But these are going to be the next-level competitive companies. They will produce better products, have better uptimes, and realize substantial growth.” EP

What is a Data Scientist?

Many companies are asking the question, “We have a lot of data. Now what do we do with it?”

To answer the question, manufacturers are using data scientists to work closely with their subject-matter experts.

A data scientist is generally described as a person employed to analyze and interpret complex digital data, such as the usage statistics of a website, to assist a business in its decision-making.

Data scientists are highly educated and have a deep depth of knowledge for programming. They know how to work with unstructured data, i.e., undefined content that doesn’t fit into database tables, and can translate complicated formulas into a format that is easy to understand for the layman. According to many online articles, the data scientist must possess a high level of intellectual curiosity, an aptitude for business acumen, and highly developed communication and teamwork skills.

Data scientists are difficult to find. An Aug. 2018 report, LinkedIn Workforce Report for US revealed that the demand for data scientists is “off the charts,” with data-science-skills shortages present in almost every large U.S. city. Nationally, there is a shortage of 151,717 people with data-science skills, with particularly acute shortages in New York City (34,032 people), the San Francisco Bay Area (31,798 people), and Los Angeles (12,251 people).

“When you think about people who use data, there is a spectrum of expertise, and there is a spectrum of readiness that the data needs to be in for that team to use it,” Nathan Oostendorp, co-founder and chief technical officer of Sight Machine, San Francisco (sightmachine.com), explained. “If you have someone who is really a data user and has a dashboard prepared for them, this is where most people who are stakeholders in the manufacturing process would prefer to live. They would prefer to have most of the information prepared for them so that they are armed with this data when they go into a meeting where decision making is happening.”

On the other extreme, the data scientist is someone who works with data tool kits rather than dashboards or Excel-type tools. These are tools that are program-language based and can be used to run single analyses to test a single hypothesis that involves some level of data preparation.

“In our case, the data scientist is also building repeatability into the line of inquiry,” Oostendorp said. “It’s not a matter of just running a report once. You are enabling people who are lower in the data sophistication spectrum to be able to run an analysis themselves and also comprehend the output.”

Having the data scientist work closely with the subject-matter expert is critical.

“There are some analyses you can do that are completely generalized—based on statistics and not process expertise,” Oostendorp explained. “Very frequently, we go to an engagement and see that the types of analytics that people really want are ways to optimize and check the actions of a human operator. If this happens, I go look at the alarms data, and I go look at the valve timing, and I go look at the time this valve opened and closed. There is a procedure. The data scientist wants to really dive into the core idea of what this person is trying to discover and then builds an analytics report that explores that connection. It is very important for data scientists and subject-matter experts to have face time. It is rare to find someone who is really good at the analysis and also really good at operating the equipment. It’s a partnership.”

As data scientists are used more frequently, their roles will make a significant shift, Oostendorp said.

“If you look at the data scientist’s work flow today, the classic formula is they spend 80% of their time doing data acquisition and preparation and 20% of their time actually doing the science part of their title where they are developing an algorithm or writing an analytic,” he explained. “What you will see as platforms like Sight Machine’s get out into the market, the data scientist will use the platforms and spend much more of their time working deeply with subject matter experts and spending more of their time on the science portion. It may seem like an incremental change at first, but when you look 10 years into the future, the role will change dramatically in terms of the raw skills of the good data scientist and how much a data scientist can produce.”

Michelle Segrest is president of Navigate Content Inc., and has been a professional journalist for almost three decades. She specializes in creating content for the processing industries. If you have an interesting efficiency, maintenance, and/or reliability story to tell, please contact her at michelle@navigatecontent.com.