Data Science Primer: Reducing Downtime with Machine Learning
Downtime is one of the biggest financial risks in manufacturing. Asset-heavy companies are exposed to a lot of risk, because every time a machine breaks they lose labor costs, suffer from decreased production, and possibly even miss sales targets. But typical maintenance schedules are somewhat arbitrary and tend to be expensive. Companies pay technicians to look at machines in no danger of breaking down even as actual problems get overlooked. Machine learning allows manufacturers to combine sensor data with the power of the cloud to catch problems just before they happen, all while spending less on routine maintenance.
Typical Maintenance Schedules are Arbitrary and Inefficient
With all the risks associated with downtime, it’s no wonder systems like TPM emphasize maintenance. A company would rather be careful than sorry, so they’ll pay top dollar to have their maintenance people come more often. To do this, they’ll often use the maintenance schedule recommended by the manufacturer. But there’s an inherent conflict here. The manufacturer is often paid to provide the maintenance, so they err on the side of more visits. This is arbitrary and highly inefficient because companies end up paying someone to drive every month and look at machines that may not have a real danger of failing. At the same time, machines which really are at risk of failure might not be caught until the next maintenance cycle.
Failing to use predictive algorithms to perform maintenance adds substantial expense to an organization and leaves them more vulnerable to market downturns. It also leaves them at risk of downtime or having to pay for repairs that could have been prevented using predictive analytics.
Creating a Perfect Schedule for Each Machine
Prior to cloud computing, there wasn’t a good way to reliably capture the data necessary to leverage predictive maintenance computational models. Historically, sensors to test variables like engine speed or internal temperature had to be close to the hard drive the sensor data was writing to. And it was difficult to study this data for correlations and develop predictive algorithms.
With the Internet of Things, all that has changed. Inexpensive, interchangeable sensors allow companies to automatically send thousands of measurements to the cloud every second. Companies can track environmental factors like moisture, air pressure, and temperature across all their factories and track internal conditions like engine temperature, oil temperature, and RPM time history for every machine.
Then machine learning algorithms and software are used to optimize maintenance schedules or only provide maintenance when necessary. An AI takes historic data and turns it into a model of future performance. In this way, it determines what machines are most likely to break down. Then it back-tests this model against a different set of historic data to see how accurate it would have been. It refines the model until it finds one that is mathematically optimized for accuracy. Using this information, executives can make an informed decision on when to apply maintenance. This reduces risk even as it decreases costs.
IoT in the Field
Rockwell Automation is using this IoT technology across the supply chain for the gas and oil industry. Oil pumps off the Kenai Peninsula work 24 hours a day year-round, and a single pump failure can cost $100,000 to $300,000 per day in lost production. Rockwell connected the pumps’ variable speed drives to IoT sensors, allowing their engineers to monitor them from Cleveland and diagnose problems immediately. The company says the next step is developing a predictive analytics model that will allow them to diagnose potential problems before they occur.
Rolls-Royce is already at that point. The company has 13,000 engines in commercial aircraft around the world and offers comprehensive maintenance services. Their services are based on a model where customers are charged per hours flown. Using Microsoft Azure, Rolls-Royce took the thousands of signals each engine generates and created a model for maintenance needs. In an interview with Microsoft, Rolls-Royce’s Product Manager Data Services Michael Chester explained that with gigabytes of data every hour coming from large fleets, there was plenty of room for an AI to work.
Chester said, “By looking at wider sets of operating data and using machine learning and analytics to spot subtle correlations, we can optimize our models and provide insight that might improve a flight schedule or a maintenance plan and help reduce disruption for our customers.”
Components like fuel pumps have recommended service intervals based on the amount of hours they’ve been in operation. But those averages don’t take into account individual variations in performance. When a pump’s performance starts to fall behind what would be expected for others in its fleet, the AI will flag it for earlier maintenance. Conversely, a pump outperforming expectations may get a clean bill of health to stay in service a little longer. Multiplied by every component, on every plane, on every fleet, the potential gains in efficiency are enormous.
Beyond Arbitrary Scheduling, Beyond Human Intuition
Human intuition is error prone when statistics or a lot of variables are involved. Concurrency’s machine learning solutions remove that element as much as possible. Sensors on the factory floor provide real time insight into machine performance, and predictive analytics show when problems are most likely to arise. This lets companies prioritize their maintenance schedules.
Using a scientific approach to the problem of downtime, manufacturers can save money and improving performance at the same time.