Before smart components, industry lived in a break/fix mechanical world where maintenance personnel practically looked at a machine and knew how to correct the issue. In those days, long runs of the same product led to fewer changeovers, meaning fewer adjustments. Today, machines are far more complex and difficult to analyze. Tight margins mean that downtime can be catastrophic to a company and undetected quality issues are equally bad.
Industry has developed from break/fix to preventative maintenance, to predictive maintenance, which relies on smart components to detect out-of-spec conditions. Machine learning (ML) and artificial intelligence (AI) are taking predictive maintenance to a new level of performance through early identification of out-of-spec conditions and being able to communicate with the crew either through trend’s dashboards or in easy-to-understand sentences presenting key facts of the situation. This leads to higher uptime, greater quality, and overall improved throughout. ML and AI systems for “advanced” predictive maintenance are practical tools for improving the bottom line. And they are available now.
These systems typically map data to learn a component, machine, product, or energy system’s healthy state. Once the map is in place, anomalies are easily identified by the AI software, which then prompts corrective actions.
ML and AI can map the big three — uptime, quality, and energy or any combination. ML and AI can be vendor agnostic, focusing not on the brand of component or type but on the data. And its presentation of actionable information fits comfortably within the end user’s workflow. The customer decides on where, when, and how to display information and is not required to adopt the vendor’s ecosystem.
While it is impossible to expect that these systems will perform similarly, it is not out of the ordinary for plants to improve process transparency by 100 percent, lower waste by over 50 percent and product rejection costs by over 45 percent. Machine availability can be improved by over 25 percent. Unplanned downtime can fall by more than 20 percent.
What data is involved in learning the healthy state of a system? The answer is basic operational data that may already be resident in the system. This includes speed, distance, pressure, flow, current, temperature, environmental factors such as humidity, torque, number of cycles in a component’s lifecycle, and product appearance, weight, and configuration. AI builds a digital bridge between this operational data and information technology.
Real-World Examples
In an automotive battery assembly cell, AI monitored motor currents, actuator position, the presence and absence of batteries, and battery weight. After mapping data for a healthy state, the system recognized anomalies. The AI system generated messages that were sent via an IoT gateway to the cloud to be displayed on a trend’s dashboard.
The data required for healthy versus unhealthy state already existed based on the machine’s motion. The development team layered AI over this data, mined it at the bottom where all the action occurred, and then provided information displayed graphically to operations personnel.
At a premium car production plant in Europe produces on average 1,000 cars per day. Welding is a core body-in-white process there. The plant has a total of 2,500 welding gun robots equipped with compact servo-pneumatic clamps. Using an AI system, the plant instituted continuous welding gun monitoring of cylinder and valve operating data and associated parameters.
AI aggregated the data on the 2,500 robotic welding guns and then, utilizing gateways, sent data to cloud-based dash-boards that provided easy-to-understand visualization of operating conditions. The plant now has, in effect, an asset management system on every robotic welding gun. Unplanned downtime fell by 25 percent.
At a silicon wafer plant, wafers are sawed from silicon cylinders in a process that takes more than eight hours. Poor cuts impact quality and lead to high rejection rates. The AI software mapped healthy cutting data utilizing more than 40 system parameters, environmental data such as humidity, and geometrical data. Anomalies led to early detection of defects. AI-derived maps were used to improve accuracy in quality assessment sampling. These led to a higher level of identification of true defects during quality observation. The wafer plant now saves up to $100,000 per machine per year through reductions in late defect detection.
This energy savings application for ice cream production differs from the previous examples because the data necessary to measure pneumatic air consumption, which was the goal of the project, was not available on the machine. To remedy the situation, the plant added hardware to digitally collect air pressure and flow data. The plant also installed additional sensors to increase data points for predictive maintenance. The AI solution monitored six pneumatic cylinders for predictive maintenance and monitored compressed air consumption for predictive energy savings. Anecdotal information indicates improved energy efficiency and uptime. Adding hardware and sensors to an existing machine enabled an AI-based asset management system.
Identification of Anomalies, Now What?
Early in the ML/AI development process, the idea was that all the data would be pushed into the cloud and displayed on vendor-created dashboards. The solution seemed reasonable — a direct line from machine to the cloud for anytime/anywhere monitoring. The hurdle was that customers had their own dashboards and own way of displaying data.
It became clear that customers wanted the AI healthy/not healthy data available in the manner most convenient for them, whether on their own dashboards, resident on the premises, on edge, or all three. Some customers wanted the data integrated with their maintenance management systems. Others wanted alerts sent to mobile devices. Many asked whether they could have easily understandable messages identifying what the problem was, where it was, and what corrective action should be taken. These requests led developers to understand that the system must support the end user’s workflow or the OEM’s HMI.
To be successful, as mentioned earlier, these systems must be vendor agnostic and must give the operations the ability to decide where the software platform can live directly on the system (on edge), on premises servers, or in the cloud. It must be flexible enough to enable connections to internal maintenance management software or spare parts management system to create an integrated end-to-end solution.
AI may appear futuristic, but this solution has advanced to the point where it is experienced by the user as a practical tool for improving the bottom line.
This article was written by Frank Latino, Product Management Electric Automation, Festo North America (Islandia, NY). For more information, visit here .