By Rafael Gomez, Director Product Strategy, Bright Machines
July 31, 2019
The ability to predict the future is a popular theme in science fiction movies. When presented with the foresight of an imminent accident or catastrophe, a film’s hero will take necessary actions to prevent the undesirable outcome and the benefactors of this action predictably applaud this superhuman power. In the real world, however, the ability to predict the future is met with justified skepticism. From fortune tellers, to sportscasters, to financial investors, many have tried to profit from predictions for the future. The accuracy of these predictions is usually 50% at best as they’re based on a subjective interpretation of simple historical patterns.
Application of prediction models in the manufacturing industry is met with this same skepticism. In general, factories have been hesitant to invest in innovative solutions, such as using data science to foretell manufacturing outcomes. However, the recent buzz around Industry 4.0 and a constant push for cost optimization is driving many companies to revisit the possibility of using factory data to become more efficient.
Most factories collect machine data: the manufacturing eco-system consists of thousands of physical actions that are applied for the single purpose of producing a finished product. Real-time insight gained from applying science to these physical data points can be used to optimize assembly or avoid negative outcomes. Unfortunately, most manufacturers don’t know how or where to start to operationalize that information. The following thought framework is necessary to begin the journey to a successful predictive project deployment:
You must first identify what problems you’re trying to avoid or resolve:
- Diagnostics: What processes have the most significant contribution to my manufacturing defects?
- Scrap & Yield: How do I predict defects minutes or hours before they occur and take immediate action?
- Performance: How do I proactively improve productivity & OEE by avoiding non-value actions, such as machine maintenance and assembly ‘stop line’ situations?
- Operating costs: Can I reduce inspection and test operations and assets by using analytics to determine inspection and test sample rate?
- Headcount: How can I reduce non-value activity such as repair and re-testing of products so that I can optimize headcount?
- Smarter production: How can I create a closed loop feedback process so that I can manufacture products with less variability?
Once you’ve identified the areas where your insight can solve business-critical issues, you can start collecting and analyzing data:
Step 1: Form a committed cross-functional project team that includes Manufacturing, Operations, Data Science and SW Development.
Step 2: Create use case process map for the problem you are trying to solve, including prescriptive action model.
Step 3: Create a list of key metrics, including financial metrics, to be monitored during deployment to quantitatively measure project success.
Step 4: Select machine data parameters and implement an IOT solution to acquire this data – be sure to collect linking provenance data such as product serial number.
Step 5: Determine product clustering (grouping) schemes to improve analytic success
Step 6: Employ an architecture for reiterative data modeling and start with simple analytic. methods – assess the required computational overhead required for your solution.
Step 7: Determine whether analytic confidence levels meet your deployment and financial requirements. Deploy a predictive solution to production once targets are met. Monitor daily.
It is fundamentally important to utilize the expertise of a cross-functional team to create a deployable and financially viable solution with predictive and prescriptive elements. This level of collaboration will allow factories to collect and analyze data and empower them to take action with it. Cloud-based analytic tools have enabled greater access to cost-effective computing power required to iterate the data models until a successful scenario is achieved, so there’s no reason traditional factories shouldn’t use data science to monetize their process data. Effective use of this data provides true cause-effect explanations and enables focused proactive action, helping to avoid negative outcomes and waste while optimizing for output and cost efficiency.
While deployment of successful prediction outcome solutions into a factory may not be a compelling as our favorite science fiction movies, the real-world impact is profound. With dramatically greater profitability, quality and innovation, there is no doubt that prediction solutions will – in the very least – cast the analytics project team as the superheroes of the factory.
About the author
Rafael Gomez is a Director of Product Strategy, focused on developing advanced analytic strategy and deployment models for Bright Machine customers. With over 20 years of experience in automation software and hardware development, he has managed the deployment of innovative solutions to factories in Asia, North America, South America and Europe. Prior to joining Bright Machines, he held engineering roles at Flex and Intel. Rafael has a Mechanical Engineering degree from Texas A&M University. Based in Austin, Texas, when not helping Bright Machines customers automate their factory lines, he enjoys traveling with his family, performing music with his band and playing tennis.