Machine Learning Finds a Home in Process Manufacturing

Without a doubt, Machine Learning (ML) problem-solving tools have proven valuable for manufacturers, helping to drive everything from product improvements and identifying supply chain inefficiencies, to creating brand new revenue streams.

As the trend has gained momentum, predictive analytic tools and services have become more accessible, and ML skillsets and partners have become commonplace. Consequently, process industry manufacturers are increasingly leveraging ML to make data-driven decisions across the value chain. Some of the key areas include:

• Automating formulation models to predict output that is based on inputs to reduce manual and costly trial-and-error iterations;
• Minimizing inventory levels through optimal production scheduling and incorporating demand- and supply-side variables (e.g. weather, multi-node supply network, commodity pricing);
• Improving batch quality by predicting critical-to-quality parameters with sufficient time for adjustments to be made to the batch;
• Enabling customer customization and personalization tools through on-demand digital formulation matching;
• Improving retail performance with insight on how to target promotion strategies;
• Advancing environmental sustainability objectives, where ML and AI are cited as the single most important technology, per a 2021 IndustryWeek-Hitachi survey.

ML solutions typically benefit from experimental, iterative methods and systems during development but require certainty and supportability after deployment. While the lifecycle of a single ML solution is complex, as ML projects multiply within an organization, so too do the challenges.

For example, it’s increasingly difficult to ensure solutions are reliable, accurate and efficient in production, all while maintaining the spirit of adoption through “citizen developers,” who are devising, developing, and using the models developed. Without such assurances, ML development can stall, as well as associated business gains.

So, how can we address the apparent paradox of flexibility, accessibility and reliability desired for effective ML solutions? This where the practice of ML Operations (MLOps) comes in. Like other “Ops” categories, MLOps refers to methods and systems to manage ML models in a cohesive manner through consistent data integration, automation, governance, diagnostics and best practices.

Modern cloud technologies offer many advances to support the reliability, interoperability and maintainability of ML-based solutions.

• Standardize data integrations to existing data systems to reduce manual efforts and enforce data management policies.Semantic (or business) language can be applied to the data to ease downstream usage.
• Systematize routine tasks due to model or data change, or scheduled refresh. Automate data preparation, model training and model deployment to a user end point to ensure consistency and reduce valuable labor hours.
• Monitor performance of ML model execution with business-friendly dashboards and notifications on exceptions.Infrastructure (CPU, memory usage), ML Pipeline (health and statistics) and Application (usage, query performance) should all be monitored actively and centrally to assure satisfactory performance of all ML solutions.

Data scientists (and these days, that includes skilled data engineers) are trained in data analysis, ML algorithms, information visualization and focus on accuracy and usefulness of models developed. They understandably are not well-versed with (or in some cases even incentivized to care about) version control, commenting, pull requests, comprehensive unit testing and in some cases regulatory compliance issues. MLOps practices provides guidance and structure to enforce common use of data collection, semantic naming of data, ML software, coding, testing and documentation of best practices.

Smart application of MLOps tools and techniques can expand experimentation, engineering, adoption and use of industrial data science to drive business performance. Hitachi Vantara offers a practice that helps organizations set up and operate MLOps centers. A recent customer, for example, deployed a robust, but intuitive, push-of-the-button solution that automates ML training, testing and deployment. As part of the solution, we provided a playbook of best practices for the company’s data scientists and engineers to follow. On top of this, we created dashboards that provide the company a centralized diagnostic to assure that all ML executions are occurring flawlessly and accurately. In the end, however, the bigger result may be that the company now has an effective repeatable method to run models on-demand by an increased number of skilled users.

ML is here to stay in manufacturing as well as other industries. It’s imperative to start investigating, testing, and experimenting with projects as soon as possible. The value will be visible in both the short and long term.

Related

• Innovate with Data-driven OT
• AI & Insights Services
• Industrial Data & IoT Management