The advent of digital transformation in manufacturing has spawned numerous projects, initiatives and pilots. Though there are similarities with what's been done for business analytics (BA) and business intelligence (BI), there are many important differences. One is the nature of the data. BA/BI projects invariably involve extensive data cleansing and an extract, transform, load (ETL) process before applying analytics. This is largely due to the nature of “business” data, which is often collected for different purposes using multiple processes with different goals. Manufacturing data found in the process industries is different and requires a different approach.
One aspect of process data that makes it easier to clean and prepare for analytics and modeling is that it's more deterministic than other types of information, according to Peter Guilfoyle, CEO at Northwest Analytics. "Process data is most often found in historians, and is the same data used to control the process—so there's already an element of trust when used for analytics," says Guilfoyle. "Laboratory data, most often found in laboratory information management systems (LIMS) are the result of laboratory tests and measurements using defined methods and quality assurance (QA)/quality control (QC) procedures. These analyses are also performed for specific purposes, such as product quality, lot clearance, raw material assessment and in-process samples used to make adjustments, which also contributes to a level of trust. This makes manufacturing information better than business data, which is often pulled for purposes that aren't predefined, and are harder to prepare and pull into analytics and modeling applications. The problem is that enterprise users often don't understand process measurements."
Northwest Analytics monitors real-time operations for equally well-defined purposes, and Louis Halvorsen, CTO at Northwest Analytics, explains the traditional patchwork of plant-floor equipment and processes is no longer a barrier to analytics because most data is stored in historians like OSIsoft's PI System, which is made possible by wide adoption of communication protocols like OPC and OPC UA. "A large facility may have as many as 50,000 point sources of data being recorded, but the patchwork usually isn't an issue because all their performance data for the past few hours and days is in the historian," says Halvorsen. "We can get a temperature reading from eight years ago if it's needed. Sometimes, we have numbers, but we're unable to associate a device, process or even product with it, so data marshalling is still a challenge. However, it's a recognized challenge because most if not all required data goes into historians, and they can be queried by users, manufacturing execution systems (MES) or other databases to identify processes, times and dates, and other events and results."
Halvorsen reports many plants and users are investing in digital transformation to get data where it's needed most, but Guilfoyle adds this doesn't means they're getting rid of the process control infrastructures they rely on. "They're usually adding more sensors and other data sources, often employing IIoT, but they either need to expand the existing data system (usually by adding a historian) or need a new place for the added data that's often cloud-based,” says Guilfoyle. "Many of these data storage methods are the same as before or support the same data integration standards. The important thing for analytics software like ours is that the data be in what we call a 'managed state,' which implies stability, security and consistency. In the life sciences, you can add 'validated,' which is a regulatory requirement. In any case, as an analytics vendor, we stay on our side of the databases, and leave the real-time data acquisition and tapping data streams to applications designed specifically for those purposes.
"Beyond adding new data sources, another key driver of digital transformation is one that's not always recognized—process industry companies that collect mountains of data for years without fully leveraging it to improve their processes. Our part of this transformation is to help identify data and analytics methods that build on what our customers already have, extract value, and put the results in the hands of the people who can use it to make a difference. Some industries like pharmaceuticals and food and beverage aren't as far along on managing and leveraging manufacturing data as process industries like and oil and gas and chemicals. Sometimes they're just at different stages of development, but there are also systemic and regulatory barriers."
Culture vs. digital transformation
Guilfoyle reports that many Northwest Analytics customers think the main issue is applying analytics where they hadn’t before, but realize their first issue is culture, such as addressing entrenched habits, perceived added workload and trust. "The first step in digital transformation turns out to be transforming the workforce to engage with analytics as part of their normal work, and understanding and trusting what their analytics are telling them,” says Guilfoyle. "Plant workforces are notorious for discontinuing work that doesn’t seem helpful, and stopping the use of tools they don’t fully understand or trust—and they usually don’t ask first. Process industry firms typically have huge deployments of data collection and management software across multiple plants with a wide range of analytics applications and approaches. They may take pride in being 'data-centric,' but they still lack ways to leverage that data and incorporate analytics into how their staff is working day-to-day. As a result, their applied analytics tools are of limited use because of what they're not doing with them."
Guilfoyle adds that what works best is involving people at the plants in the process of executing a digital transformation initiative and making sure it's relevant to their jobs. "The goal is to provide tools and methods that, as one customer said, 'makes their lives easier.' What they meant was that their jobs become more rational, they become more effective, and they're able to do the job their were hired for," he says. "Instead of putting out fires, they focus on fire prevention (early problem detection, issue resolution, risk reduction), and can spend more time on process improvements that reduce improve quality and yields while reducing costs. The result can be called 'democratization of analytics.' ”
Appropriate analytics
Another key to making digital transformation stick is starting with an appropriate application of analytics by choosing methods that are readily understood and that develop a high level of trust at the plant level. Guilfoyle reports that engineers and operators won’t act on signals they don’t understand or trust. "Too often, projects start with the new shiny thing, such as advanced machine learning (ML) and artificial intelligence (AI) that have challenges in delivering understandable and reliable information, and often aren't sustainable at the plant level," he says. "Starting with univariate methods, such as applied statistical process control (SPC) to individual variables (process parameters and laboratory data), provides easy-to-understand content, robust statistical methods and high levels of trust (low false-positives). Marshalling the data for univariate methods starts a project on a highly effective path (so far, more than 90% of the significant gains our customers have seen from their initiatives came from univariate analytics). This also lays the groundwork for successfully introducing multivariate methods (machine learning, etc.) that can be targeted to situations they're better at than univariate methods, such as detecting issues and solving problems that univariate methods struggle with. This combination of well-implemented univariate and targeted model-based analytics provides a real platform for success."
Guilfoyle explains, "If we begin by imposing models created by experts, there will be arguments over the models and disconnects with plant personnel. We must start with that shared view of the process, build trust and acceptance first, and then add appropriate model-based analytics later. We have to join what's in people's heads with the signals and data we're gathering, so they'll be able and willing to use the models they create and produce a big return on investment (ROI) and achieve sustainability.
"We have customers that are good examples of the combination of paying attention to culture and employing appropriate analytics where they started by involving the people that run the plants early in the process and introducing analytics in stages. This avoids the temptation of management to rush into the 'next shiny thing' without establishing the bases for adoption and sustainability. Once the plants saw success with the initial phases and understood that these initiatives were actually going to help them do their jobs, introducing more elements to the deployed analytics was much easier. People actually looked forward to doing more. These companies did take different approaches overall, but a particularly good example started with getting it right at one facility, which was selected for being strategically important and also having a technically strong workforce. Culture and appropriate analytics, and attention to building trust were the primary guides. That effort allowed the company to plan the rollout to other plants based on proven success, rather than learning as they went and risking failure plant-by-plant."
