Smarter Predictive Maintenance

Posted by Steve Sickler on 16 Feb, 2018

While managing a large portion of SAP’s Big Data and Analytics business, I ran into a big problem. We created solutions for "smart factory" predictive maintenance by feeding humongous amounts of data into SAP’s HANA in-memory database for near instant analysis and recommended corrective action to prevent problems.  It worked. It was amazing.  Yet nearly none of our large manufacturing clients could justify the expense of data transmission, storage and powerful CPU cycles.

Luckily, the super-smart math geeks amongst us have found a less complex and costly approach using “Edge Anomaly Detection.”  At Tend.ai, we are implementing this system in our Smart Robot Predictive Maintenance module.  It goes like this: Tend.ai places a simple device connected to each robot that captures all of the data generated by the robot.   These “Edge” devices can be our hardware or the client’s hardware running a virtual machine. But rather than send ALL of this data to our Smart Cloud Robotic center, these devices only send the unusual data: the anomalies.

This approach solves another common problem: most of the current “predictive” maintenance systems are based soley on problems that have been seen before.  Like TSA making us all take off our shoes when going through security; it is based on a past problem (a terrorist making a bomb that fit in his shoe) being an indicator of a future problem (anyone wearing shoes could be carrying a bomb).  Most of today’s predictive maintenance solutions for factory robots is based on the same basic premise: A past issue is the best indicator of a future issue.  Problem is, like TSA and our shoes, this solution often over-states the statistical probability that it will actually happen again, while missing a new problem that has not yet been identified. These systems tell you to fix a motor based on hours used even if the motor is running great. But will miss a vibration indicating the motor bearing is failing on a relatively new motor until it has been reported before.

TSA Shoes.jpg

However, anomaly detection solutions are often plagued by the “chicken little” problem: lots of false alarms; too many false positives.  The first System Management Systems (SMS) like those from Tivoli/IBM had this problem; so many of the alerts were false that the engineers managing the systems began ignoring them.  The correct refinement of the edge algorithms is critical, but context is as important: the robot engineer might be aware of information that is not apparent to the edge algorithm.  For example, a machine-learning algorithm tasked with profiling a robotic arm that has never moved more than 8 inches to the left will raise an alert the first time it sees the arm move 8.1 inches left.  

Therefore, by giving the edge algorithms some definition of what is “normal and allowed” behavior, it is possible to greatly reduce the number of false alarms.  During set-up of a robot, the edge algorithm is given a set of normal and acceptable behaviors.

In our research, we have found that adding some context-based logic to our anomaly detection edge algorithms raises the number of accurate or true alarms to 99%.  In other words, only 1% of the alarms are false and the rest deserve some attention.  This also allows alarms to be sent and reacted to locally – without waiting to process them through our Smart Cloud Robotics AI.  Additionally, since only the data associated with “true” alarms is sent to the AI for improved future predictive maintenance and edge algorithm performance, we see and exponential reduction in the amount of data transmitted from the edge devices, dramatically reducing data transmission costs and CPU power needed.

This solution also solves the “TSA” problem:  A context-enabled anomaly detection system will send an alarm even when the problem has not been seen before while ignoring an “acceptable” anomaly that has been previously programmed.

Net – Tend.ai’s Predictive Maintenance Module will greatly reduce the cost and complexity of factory robots while improving the reporting of actual problems.  Further ongoing improvements will also be achieved over time as our Smart Cloud Robotics engine “learns” how to correlate new alarms with previous results.  A practical and cost-effective Preventative Maintenance solution for robots is finally within reach.  As we say, Smart Robots run Tend.  To participate in our Beta or Early Access program for this solution, let us know here:

Sign-Up for a Free Trial

Topics: predictive maintenance

Written by Steve Sickler

Steve was a founding member of the MSPAlliance, the largest and oldest vendor neutral organization for cloud and managed service providers. He has held senior management positions with some of the SaaS pioneers and leaders such as SiteLite (one of the first MSPs), Tivoli/IBM and SAP. Currently, Steve is President of Tend.ai, disrupting how robots are managed via their robot-agnostic monitoring, alerting and predictive-maintenance platform
Find me on:

Smart Robots Run Tend.ai

Smart: Big thoughts, Algorithms, Results

Robotics: They're here, more are coming

This blog is dedicated to bringing new thoughts on the goal of keeping factories up and running.  Not just giant ones building planes and cars, but factories using robots to change the world with lower costs making cooler stuff, like craft beer.

Subscribe to Email Updates

Recent Posts