Good morning, everyone. My name is Julien, and I'm a tech evangelist focusing on AI and machine learning at AWS. Today, I'd like to discuss predictive maintenance with machine learning. The objective of maintenance is, of course, to minimize downtime and unscheduled maintenance, ensuring that production facilities keep running as smoothly as possible. Unfortunately, this is a real challenge for many companies, and a lot of them struggle with defining and executing maintenance schedules. As we can imagine, this has a significant impact on the efficiency of production facilities, costing tens of billions of dollars due to machine breakdowns and temporary facility shutdowns. This is a problem every industry faces, and we're going to explore how we can improve it with machine learning.
Maintenance is nothing new; it has been around since the dawn of industry. The first strategy is the reactive approach: wait until something breaks down, diagnose the issue, and fix it. The problem here is that things break down frequently, and Murphy's law ensures that these breakdowns often occur at the worst possible times, interrupting production repeatedly. A slightly better strategy is planned maintenance, where you have a well-defined schedule to proactively replace parts that might fail. This is better because you can manage production interruptions more effectively, minimizing their impact. However, you may still replace parts that don't need replacing and periodically interrupt production for maintenance.
More and more companies are moving to predictive maintenance, using data and analytics to determine when to replace a part. If it doesn't need replacing, you won't replace it. If early signs of failure indicate an impending issue, you can act quickly, interrupting production at a specific time to fix the problem and then resume operations. Predictive maintenance is an increasingly popular technique to ensure production facilities run smoothly. It involves several aspects: anomaly detection, identifying when something unusual is happening; root cause analysis, understanding the actual problem; and predicting the remaining useful life of a part, determining how long it can be used before it is likely to fail. Anomaly detection is already very valuable because it alerts you to potential issues, allowing you to take action.
Predictive maintenance is expected to save industrial companies a significant amount of money in the coming years by preventing failures, minimizing risks, improving quality, and reducing repair and maintenance costs. By acting only on parts or equipment that are about to fail, you avoid unnecessary maintenance on equipment that is functioning correctly. There are many positive aspects to this approach.
Building a predictive maintenance platform involves three major steps. First, you need sensors attached to your equipment to monitor physical quantities like vibration, temperature, and pressure, which can serve as early warning signs. You can either build these sensors yourself or purchase them. The second step is to capture the data from the sensors and funnel it into a database or storage area for further cleaning and preparation for analytics. Building a secure and scalable pipeline to process sensor data is not easy. Finally, once the data is in a central repository, you need to process it and train predictive models. This is where machine learning comes into play, presenting a new set of challenges.
Let's consider a simple example: a two-motor pump. We capture RPMs and flow rate. Based on experience and manufacturer specifications, we know certain value ranges are normal, indicating the pump is operating as expected. If we see low RPMs and low flow rate, the pump is idle or not running fast, which is fine. If we see high RPMs and high flow rate, the pump is working as intended. However, issues can arise, such as a clogged pump, leading to high RPMs and low flow rate. This is an abnormal condition that machine learning should detect. Machine learning should understand normal operation and identify abnormal operation, ideally catching issues early to allow ample time for repairs.
This example might seem simple, but it involves unsupervised learning, where the data points are raw and unlabeled. You need to determine which data points are normal and which are not. Machine learning can help with this, but it also presents challenges, such as identifying the right applications and use cases and understanding the data's quality. Real-life data is complex, and different equipment and sensors can have varying ages, failure histories, and variabilities. Data cleaning and normalization are critical tasks, often taking up to 80% of a data scientist's time.
To address these challenges, you need skilled people to define the machine learning problem, understand and clean the data, build and train models, and manage the necessary IT infrastructure. A typical machine learning workflow for predictive maintenance includes capturing data, aligning timestamps, imputing missing values, selecting algorithms, optimizing parameters, and training models. Each step is complex and requires expertise, which can be hard to find, especially for smaller companies.
To simplify this process for our customers, we built Amazon Monitron, a turnkey system for companies interested in predictive maintenance. It includes everything you need: Monitron sensors, a gateway, and a mobile app. The sensors are small, battery-powered, and wirelessly connect to the Monitron Gateway via Bluetooth. The gateway sends data to the AWS cloud over Wi-Fi, and you receive alerts on your mobile app if the models detect unusual activity. Technicians can then inspect the equipment and confirm whether an alarm is valid. The setup is straightforward, using NFC on your phone, and can be completed in minutes.
Monitron is suitable for monitoring various rotating industrial equipment, such as pumps, fans, and compressors. Customers like Fender, General Electric, and RS are already using Monitron, and Amazon itself uses it to monitor conveyor belts in fulfillment centers. According to Mr. Ackeson, the CIO of GE Gas and Power Manufacturing, Monitron is quick to deploy, works with existing equipment, and requires no technical skills. Within minutes, data flows to the cloud, and alerts appear once the models are trained.
The benefits of Monitron include its simplicity, cost-effectiveness, and security. The starter kit, which includes one gateway and five sensors, costs $715, and you pay $4.17 per sensor per month, totaling $50 per sensor per year. This is a low cost compared to the value of the equipment being monitored. The system is secure end-to-end, and it improves over time with over-the-air updates and better models as more data becomes available.
If you already have sensors and historical data, we offer Amazon Lookout for Equipment. This service allows you to use your existing sensors and data, uploading it to Amazon S3 for analysis. Lookout for Equipment uses AutoML to train models, and you can predict new data on a schedule. It integrates easily with your existing IT systems, sensors, and maintenance workflows, making it a non-disruptive addition. Companies like GS EPS in Korea have successfully used Lookout for Equipment to diagnose equipment issues, optimize production, and reduce costs without needing ML expertise.
The benefits of Lookout for Equipment are similar to Monitron: ML-fueled automation, no need to change your factory or warehouse setup, and easy data preparation. You pay as you go, with costs based on data ingestion, training, and inference. The service is scalable, integrating seamlessly into existing workflows, and allows for quick proof-of-concept testing.
To get started with Lookout for Equipment, sign up for the preview. The service was announced at AWS re:Invent last December and is currently in preview. You can find code on GitHub showing how to prepare and upload data, train models, and schedule inference on new data. We have a collection of Jupyter notebooks to help you get started.
That's what I wanted to share today. We explored predictive maintenance with machine learning on AWS, focusing on Amazon Monitron and Amazon Lookout for Equipment. For more information on these services and our other AI and ML offerings, visit ml.aws. Feel free to connect with me on Twitter or elsewhere if you have any questions. Thank you for inviting me, and thank you for listening. Have a great day.