Scaling machine learning for the manufacturing masses
Enabling access to machine learning algorithms in easy-to-use advanced analytics applications accelerates insights in process data. A specialty chemical manufacturer used ML tools to predict over 90% of quality deviations and save more than $500,000 per year in quality downgrades.
Manufacturers in process manufacturing have heavily invested in the people and technologies required for effective digital transformation including machine learning (ML) software. As a result, data scientists are in high demand, data storage and migration strategies are rapidly evolving and cloud-native analytics offerings are flooding the market.
A specialty chemical manufacturer used ML tools to predict over 90% of quality deviations and adopt a model-based control scheme, saving more than $500,000 per year in quality downgrades.
To bolster the return on investment (ROI) in data science resources, organizations need to scale the efforts of data scientists. Data science software must serve data science teams and the employee base of frontline process engineers, technicians and specialists.
One way to achieve this is publishing algorithms generated by data scientists, partners and third parties into easy-to-use tools and experiences, which can be used by process engineers to scale digital innovation efforts throughout a company. With these and other advancements, the limits of advanced analytics applications are fading, with new extensible interfaces to support widespread ML innovations in process manufacturing organizations.
Democratizing ML: Machine learning for all
Democratization is one of the most broadly used buzzwords of the fourth industrial revolution, taking on many flavors during the process industries’ journey to digital transformation.
First there was the democratization of data. Aided by the COVID-19 pandemic, traditionally slow-moving manufacturing companies accelerated investments in data strategies. The underlying objective of data democratization was making the right data available to the right people from anywhere, whether within the plant firewall, a home office or a remote monitoring facility. A critical component of this was the data access, integrity and alignment provided by one advanced analytics application with a live connection to source data systems.
Subject-matter experts solve high-value business problems
Democratization of data gave way to democratization of analytics and insights. Given access to the relevant process and contextual data sources – along with the user-friendly experience of advanced analytics – subject matter experts (SMEs) could now solve high-value business problems. While the insights gleaned from these analyses have historically been siloed, knowledge capture and collaboration capabilities of advanced analytics applications enabled results to be captured and shared throughout the organization.
The democratization of ML algorithms is poised to become the next big wave. One of the most effective ways to scale the investment in algorithms created by corporate data science teams or third parties is to put the algorithms into a form accessible to a broader employee base. This requires advanced analytics solutions with a live data source connection to support continuously learning models, along with integration to existing data analytics interfaces, to ensure successful deployment and adoption.
Extensible applications, artificial intelligence
As with the extensibility of advanced analytics, it’s possible to democratize ML and artificial intelligence (AI) by making the application user interface (UI) extensible to access customer-created software modules and algorithms. Areas of application extensibility include add-on functions, add-on display panes and add-on tools.
Such add-on functions provide organizations with the capability to incorporate their intellectual property and methodologies into an existing software suite of time series-specific functions. Opportunities for this approach include standardizing equipment monitoring by creating add-on functions for parameters like heat exchanger U-value, control loop error, or boiler efficiency.
In the U-value example, an add-on function empowers users to apply the first principles equations describing heat transfer by referencing a function defined as “U_Value” and specifying input variables of temperatures, flow rates, density and specific heat. Add-on functions can range in complexity from linear equations for heat transfer to differential equations describing fluid dynamics and reaction kinetics.
A growing number of data visualization options are available through business intelligence tools, statistical software and open-source programming libraries. There is significant value to be gained by making these visualization options available in advanced analytics applications with a live connection to all the relevant source data systems.
Add-on display pane visualizations enable end users to interact with live data by leveraging unique visualizations. These displays can be configured by customers, partners, or third-party vendors and “plugged in” to a display dropdown accessible by process engineers using available software (Figure 1).
Removing data silos with user personas
To break down the silos between data science teams and SMEs, advanced analytics applications have introduced experiences targeting different user personas. The application experiences for each persona must interact to foster collaborative model development, eliminating multiple iterations of feedback and rework.
Once a model or algorithm has been finely tuned, the challenge becomes scaling it from a single process at a single site to the rest of the manufacturing organization. This challenge can be addressed with Add-on tools, offering point-and-click UIs as a front end for complex ML and AI algorithms, making them accessible to process engineers and other SMEs.
An application mode is used to display interactive front-end designs for the algorithms running behind the scenes. Scheduled execution of data lab notebooks enables models to be continuously learning, and recalculating.
Manufacturers are seeking flexible analytics software to incorporate their internal intellectual property (iP) contained in algorithms built by partners or third parties or purchased in public marketplaces. An extensible analytics application addresses each of these needs.
Honoring existing data governance
Configuring permissions on multiple levels, from data sources to algorithms to individual calculated items, is a key component of a successful solution to ensure support for organization data governance and access requirements. Deploying custom functions, algorithms and displays in an advanced analytics application with support for data governance and access thus empowers companies to take advantage of existing secure data source connections.
Another critical consideration when deploying machine learning algorithms is ensuring the solution architecture maintains the full integrity of the source data system with no down-sampling, aggregation, or compression. Rapid machinery failures, for instance, become much more challenging to detect when critical data features have been smoothed out by hourly averages.
Two use cases demonstrate algorithms at scale
Environmental stewardship: An oil and gas company wanted to progress towards sustainability milestones using advanced analytics. A centralized data science team worked with site engineers to develop a neural network algorithm which estimated NOx emissions based on current state operations. The algorithm was operationalized as an add-on tool and made accessible to site process engineers for continuous monitoring. This near-real-time insight into process emissions empowered engineers to make proactive process adjustments for reducing overall greenhouse gas emissions and improving site environmental performance. The add-on tool was made available to engineering resources at each of the company’s other refineries and widely adopted as a best practice for monitoring environmental performance.
Product quality excellence: A specialty chemical manufacturer was looking to build an accurate forecast of product quality disposition but was unsure which measured and manipulated variables had the greatest impact on the target signal. A correlation algorithm deployed as an add-on tool identified the input signals with greatest effect on product quality (Figure 2).
The algorithm also automatically calculated the process dynamic lags between the upstream signals and the target variable. A reduced number of signals, with appropriate time delays, was pushed back to workbench software, where a predictive model was deployed. The model was then validated against historical data and found to accurately predict more than 90% of quality deviations. This new model-based control scheme was adopted, saving more than $500,000 per year in quality downgrades.
Advanced analytics make ML accessible
ML is a necessary innovation to increase efficiency in manufacturing organizations dealing with more data and pressure to improve outcomes. Widespread access has been hindered by the need for constant interaction between data scientists and SMEs when applying ML algorithms to solve process control problems.
Advanced analytics applications address these and other issues by providing functionality to incorporate ML algorithms in their software. This empowers SMEs and other frontline employees to directly employ and interact with these algorithms, democratizing this powerful tool for widespread use across the enterprise.
KEYWORDS: Data analytics, machine learning, artificial intelligence
Are subject-matter experts around you able to apply machine learning? Are competitors?