Using Pattern Recognition to Improve Industrial Operations
Machine learning through pattern recognition can be a faster, less-expensive approach to monitoring manufacturing.
Lawrence S. Gould
Patterns exist everywhere. In pictures. In sounds. In data. Particularly manufacturing data. Isolating a particular pattern in the glut of manufacturing data can point to the deviations and anomalies that lead to unstable operations and inferior product quality—and do so quickly and accurately.
That’s exactly what Falkonry LRS aims to do. From Sunnyvale-based Falkonry, Inc. (falkonry.com), LRS is a “ready-to-use machine learning system that finds hidden patterns in your data to help solve your hardest operational problems,” says Nikunj Mehta, founder and CEO of Falkonry. “It applies pattern discovery, condition recognition, and predictive analytics— analyzes multivariate time-series data—to provide the early warnings that prevent expensive failures and improve operations, that find the invisible warning signs you need to see to improve quality, yield, performance, or safety.
However, unlike people visually processing patterns, LRS never gets eye strain.
Conventional operations monitoring
Manufacturers use a variety of methodologies to monitor operations. Foremost is statistical process control (SPC), except that SPC parameters need to be known before they can be monitored and acted upon. In any plant, which has thousands of such parameters, narrowing down that list to a more-manageable hundred most-important parameters is daunting. Then there’s interpreting SPC charts and understanding which signal (parameter) deviated, the cause of that deviation, and what action to take. That’s a lot of manual effort, says Mehta. Data science, he posits, is better at dealing with the complexity in ordinary industrial operations. However, data science projects don’t scale well, and data scientists need a lot of time to learn what matters in industrial operations monitoring and how data behaves in those operations.
Falkonry’s approach “prepackages” and automates data science as industrial predictive analytic software. This software, LRS, isolates the data in the form of patterns that identify unstable operational behaviors. “Any operational floor has digital signatures that can be extracted from a myriad of sources, such as sensors, machines, events and inspection logs, along with workflow knowledge that is codified into processes,” explains Mehta. “LRS can ingest all this information with the goal of finding patterns—hidden or not, causal or unrelated—to ensure that subject-matter experts [SME] closest to the operational floor can understand the health, quality, and status of their assets, and hence the operational output.”
Much of the operations data is simply not visible to conventional monitoring and control equipment, let along people. Time-series data is complex, explains Greg Olsen, senior vice president, products. “Rarely does the behavior of a single signal tell the story of what is going on in a complex system. Only by examining multiple signals over a period of time can we deduce the future state of a system. The biggest challenge with time-series data is that temporal patterns in data appear over windows in time—not in a single snapshot.”
Setup is fast
LRS generates “assessments” (patterns) from operations data. These patterns are meaningless; at first, LRS merely displays them in color to differentiate them from the background noise of patterns that arise from normal, stable operations. An SME, such as an industrial engineer, annotates these colored patterns just like one tags a person in a photograph. Annotation is based on operations experience; it’s to capture the expert knowledge regarding the causal reasons for the patterns. For example, a specific set of patterns might precede the variation in quality of robotic welds. The SME’s context completes the predictive model of the operation. LRS can then continue monitoring the real-time data stream for similar patterns that predict a failure about to occur.
When an SME annotates patterns with facts about manufacturing operations, LRS maps the pattern(s) with operational behaviors. Annotating a consistent pattern as a precursor to some behavior can provide an early warning or prediction about a critical event occurring in the near future. But there’s also machine learning. The pattern can be the basis for LRS to monitor other patterns that might indicate potential critical events. The last step in implementing LRS is for the SME to link specific patterns to specific actions, such as trigger alerts to the appropriate machine operators and maintenance staff.
“It is that simple,” says Mehta. “There’s no programming or math to learn. Users don’t have to know SPC or machine learning or programming or anything like that. What’s more, Falkonry automatically handles noise, missing data, and other issues so that you can focus on contextualizing the data and bringing that operational knowledge together.”
Falkonry LRS runs on various versions of Linux, requires Docker version 1.12, and has an installation file weighing in at 2 GB. LRS can be deployed both on-premises or in the cloud.
The price of benefits
The price of Falkonry LRS, which typically starts at $125,000 annually, is based on usage—the number of outputs or correlations running. A correlation might require two dozen parameters for analysis, and a customer might run dozens of such correlations. For all that, claims Mehta, customers often see a fivefold or higher ROI.
Manufacturing operations are already full of “old” tech that accomplish the same things the “new technology”—machine learning and pattern recognition—in Falkonry LRS provides manufacturers: improving manufacturing performance, asset safety, lowering production latency, increasing machine uptime and throughput, improving yield, reducing scrap, and enhancing the overall quality of production.
So on the one hand, LRS is “merely” another “tool in the toolbox” for managing operations. On the other hand, LRS provides those benefits and efficiencies at a relatively lower cost in terms of setup, ease of use, and ongoing monitoring. According to Mehta, SMEs can learn to use Falkonry within an hour, and the software system can “deliver value in as little as three weeks with minimal integration overhead.”
All this simply by “keeping an eye on patterns in your operational data to help you improve uptime, quality, and throughput.”
You talk with Pete Rander, president of Argo AI, a company that is working with Ford on developing Level 4 autonomous vehicle technology, and as the details become clearer, the enormity of the challenge seems daunting—yet Rander, who has been working on this for some 30 years, seems delighted he and his team are doing it.
AI will usher in new jobs, according to a book from the consulting firm Accenture. It’s a quick and worthwhile read.
Out on the production floor, nobody talks about AI or expert systems, even though the technology does exist and it's very much a part of automotive production.