No-Code Intelligent Monitoring at the Edge
Using Vibration Data to Predict Asset Failure with Elipsa and Losant
Intelligent Monitoring/Predictive Maintenance
Automobile manufacturers today typically suggest routine maintenance, such as changing the engine’s oil every ~5,000 miles. This suggested mileage threshold is based on studies of how the average engine performs over time. However, not all engines are operated equally. Based on a variety of usage factors, Engine A may need an oil change at 4,000 miles, while Engine B could stretch it out past 5,000. What if your car could dynamically tell you, based on data, when it needs to be maintained?
This is the promise of predictive maintenance and intelligent monitoring. Your machines are equipped with sensors collecting an array of data points in real-time. With the use of machine learning, you can proactively monitor these sensors to determine exactly when maintenance needs to be performed.
The result leads to reduced maintenance costs, better parts inventory management, scheduling optimization, and extending the useful life of your equipment.
With predictive maintenance, using machine learning to monitor multiple data points allows the computer to find patterns indicative of future failures. Combining multiple data points such as temperature and humidity of an HVAC for example, or RPMs and voltage of an electric motor can help to monitor the general health of a machine.
Machine data points generated by sensors can prove to be extremely important in terms of successfully predicting future faults. The best data points will vary based on the type of machine in question, but studies have shown that the most consistently important data point for intelligently monitoring the health of a machine is vibration.
In fact, in the use case we are exploring in more detail below, we were able to monitor vibration from a single accelerometer to accurately predict the degradation of a bearing on a rotating motor.
In our particular example, we explore a dataset where four bearings were installed on a shaft. The rotation speed was kept constant at 2000 RPM by an AC motor coupled to the shaft via rub belts. A radial load of 6000 lbs is applied onto the shaft and bearing by a spring mechanism. All bearings are force lubricated. Rexnord ZA-2115 double row bearings were installed on the shaft as shown in Figure 1. PCB 353B33 High Sensitivity Quartz ICP accelerometers were installed on the bearing housing. Sensor placement is also shown in Figure 1. All failures occurred after exceeding the designed lifetime of the bearing which is more than 100 million revolutions.
Building an AI Model with Clicks Not Code
Using the Elipsa no-code analytics platform, we ingested the historical vibration data from the accelerometer to build an outlier detection model. The goal was to create a model that learns from historical data what is considered “normal” operating behavior for this particular bearing in order to monitor future data for abnormalities indicative of the bearing degradation.
With access to multiple accelerometers, we could build one model to find patterns across all four sensors, monitoring the overall health of the motor. In our case, we decided to build a model monitoring a single accelerometer, allowing us to more accurately look for issues with the individual bearings themselves.
Using Elipsa, we built the predictive model in minutes. Running the model against our run till failure test data, we were able to provide an accurate prediction of abnormal behavior leading up to the eventual bearing failure. This advanced warning could have triggered actions such as a technician performing maintenance to applying additional lubrication helping to extend the useful life of the bearing, or it could have triggered purchasing to procure a new bearing for better inventory management.
Deploying for Edge
This new model can be automatically saved and deployed to the Elipsa cloud, enabling a company to intelligently monitor the system on new data going forward. These predictions can be made by calling the Elipsa cloud-based APIs.
In addition, models can be deployed closer to the edge device, in our case, the motor and accelerometer. In the Elipsa platform, models are selected for Edge deployment at which point the model and the Elipsa inference engine, are packaged together in a container for remote delivery.
These Elipsa containers can then be deployed directly onto a customer’s edge controllers. (Note: as of the writing of this post, Elipsa Edge containers are not compatible with 32-Bit operating systems or ARM chipsets)
Benefits of the Edge
Cloud computing has been instrumental in the resurgence and development of artificial intelligence as it creates a centralized location of data for AI to thrive off of. However, as AI matures, so has the need to push its capabilities away from the cloud and closer to the connected devices themselves. Computing on the devices, or on “the edge” allows companies to take advantage of multiple benefits as they look to scale AI across the organization.
Latency in computing can be defined as data transfer time or more appropriately, delays in transfer time. Large files, or a large number of files, tend to have greater latency as they take longer to transmit. Many use cases such as computer vision are therefore not well suited for the cloud because of the amount of time to transfer images to the models. With edge computing, the models reside on the device itself and are therefore capable of processing new images against the model for predictions in much more of a real-time fashion.
In addition to the latency involved in larger or more frequent files, you also need to consider the costs of transferring this data. Large amounts of bandwidth, and even 5G for remote devices, are required to transmit the streaming data to the model for predictions. This connectivity can be costly and thus companies often need to limit the number of predictions they make against a model. With the AI model residing on the edge, a company could make additional predictions at no extra cost. So, in the case of monitoring an asset for faults, you could make predictions every 5 seconds (or more even more frequently) instead of every 5 minutes. This ability to make predictions closer to real-time could prove very beneficial in catching a problem, helping to eliminate unnecessary downtime and reduce costs.
In addition to the size and transfer rate of the data, there are also considerations for the data itself and more importantly the privacy requirements around it. With AI models deployed to the edge, your data remains under your control without having to transfer it to a third-party cloud.
Pushing to Edge with Workflow through Losant
With the custom model built and deployed, Elipsa leverages the technologies of our partners, such as Losant Edge Compute to create an end-to-end solution. Losant allows users to define workflows, enabling the device to interact with the local model in real-time.
The Elipsa model container can be pushed directly to Edge devices using standard orchestration tools such as Kubernetes. On the Losant IoT Platform, we are able to build a workflow to take real-time streaming data from the device to call the Elipsa model. With Losant Edge Compute, this workflow can be pushed to the edge device enabling an end-to-end predictive workflow on the edge.
To get started with Elipsa’s Approachable AI for the cloud or edge, book a demo @ https://www.elipsa.ai/demo. To schedule a walkthrough of our partner’s platform visit https://www.losant.com/request-a-platform-walkthrough
See Elipsa and Losant together live at IoT World, November 2021!