Approachable AI Applied: Monitoring Bearing Degradation
Machine data through sensors and connected devices are generating a wealth of data that can be converted into value-added knowledge for your organization. Traditional approaches to data can allow organizations to monitor for defined conditions but these approaches are naturally backward-looking. The introduction of machine learning to your data can enable predictive analytics and forward-looking analysis to further unlock insights.
Traditionally, it has been difficult to implement and scale the usage of machine learning primarily due to technical complexities and a shortage of AI talent across industries.
This series aims to feature a collection of use cases showing how Elipsa enables Approachable AI to allow an organization's existing talent to become AI talent and apply advanced analytics without needing to write code.
Rotating machines are utilized across a vast number of industries and use cases. These machines are often deployed in instances of heavy use requiring constant uptime. As a result, machine parts such as bearings are susceptible to wear over time even if properly maintained.
Preventative maintenance can be employed to schedule routine maintenance and help to extend the useful life of bearings, but failures still occur leading to downtime, excessive costs, and lost revenue.
Manually monitoring bearings is time-consuming and inconsistent. Accelerometer sensors can be deployed to monitor the vibration or acceleration of motion of an individual bearing. These readings can then be used by machine learning as a proxy for the health of the bearing and the overall equipment.
Being able to predict how a bearing is degrading can enable organizations to proactively manage parts inventory, plan for proactive maintenance and replacement, and reduce downtime helping to save on costs while increasing revenue.
In our example, we will be exploring the data from a NASA run till failure test.
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 the figure below. High Sensitivity Quartz ICP accelerometers were installed on the bearing housing. Sensor placement is also shown in the figure.
The data is collected every 10 minutes, where each data point is the average reading from the sensor for the prior 10 minute period.
Building a Predictive Model with Elipsa
The elipsa platform is a no-code solution that enables users to apply advanced analytics to their data without the need for a data scientist.
As a first step, we need to select what type of model we are looking to build. In our case, we are going to monitor the accelerometer for outliers. In other words, we will build a model that learns what the normal behavior of this accelerometer is in order to monitor streaming data for abnormal behavior indicative of a degrading bearing.
Once the model type is selected, we need to import the data that we will use to build the model. In our case, it is the dataset detailed above. With the dataset uploaded, we set the sensitivity of the model. The sensitivity will dictate how far from normal a data point needs to be in order to be considered an outlier. With a lower sensitivity, you will be alerted to fewer outliers with only the more extreme outliers being detected. The sensitivity setting is primarily dependent on your tolerance for false positives as well as your knowledge of how “normal” your training data is.
Once you have selected the sensitivity, you simply select the columns in your data file that you want to monitor. If multiple columns are selected, the Elipsa AI Engine will find patterns between the different data points to detect outliers. In our example, we could monitor all four bearings together to monitor the overall health of the system. However, because the wear of the bearings are often independent of each other, we are going to build a model to monitor a single bearing.
Once you click to ‘Create My Model’ the elipsa engine will automatically build a predictive model for you to easily deploy to the cloud or directly to your edge device. In our example, we have deployed the model to the Elipsa Cloud to view the results.
As detailed above, the machine was run till failure. Thus, we know that the last data point was taken just prior to the bearing reaching an unacceptable level of wear. Our goal is to be able to run the data through Elipsa and see if we would have predicted abnormal behavior ahead of the known failure.
You can see from the results that the Elipsa model considered the accelerometer data to be normal for the start of the test run and then a series of outliers were detected leading up to machine failure.
In the results above, the blue lines indicate the confidence level of the model predicting the data point as an outlier. As you can see, the confidence appears to increase as the system gets closer to the known failure.
In our case, the machine was running for a 1 week period under high-stress conditions. After 163 hours, the bearings wore down to the point of failure.
Monitoring the single reading from the accelerometer allowed us to build a predictive model alerting of failure nearly 12 hours in advance under a high-stress scenario. Under less stressful use, the bearings would have likely lasted longer but also would have likely given more advanced notice of failure as the bearings would more slowly degrade over time.
This advanced notice could allow an organization to better manage their inventory of parts, and schedule required maintenance such as applying additional lubrication. Predictive analytics on a single sensor could prevent downtime and keep machines running longer helping to cut costs, increase revenue, and allow for more efficient use of employee time.
Please continue to checkout additional uses cases featured in our Approachable AI Applied series and if you are interested in exploring Elipsa’s capabilities book a demo at elipsa.ai/demo