Although machine learning is being employed in other cases at Konecranes, its application to the Agilon is one of the first where it has been taken into practical use.
The ability to predict failures before they take place would bring a major advantage in increasing a solution’s availability. Because Konecranes Agilon, an award-winning automated material handling solution, includes many moving parts, it will be highly beneficial to prepare for any possible failures or even to prevent them entirely by replacing faulty parts before any breakdowns occur.
As there was a clear need to predict breakdowns in the Agilon, the team behind the solution had already been gathering data that was related to failures. Furthermore, they saw a clear need and opportunity to apply machine learning in order to streamline maintenance processes so that failures could be prevented.
“For Konecranes’ part, quite a lot of manual monitoring goes into keeping Agilon systems in good running order,” explains Mikael Björkbom, Manager, Crane Intelligence and Analytics at Konecranes.
“With the help of data analysis, predictions could be made automatically – this means that a larger number of systems could be monitored,” Björkbom says.
Preventing failures before they occur
Konecranes’ data analysis team examined data from Agilon units’ motor torques to detect the occurrence of faulty motors, increased friction or mechanical problems. Their hypothesis was that data from motor controllers could be used to detect mechanical malfunctions that increase friction, thereby helping to predict motor failures. These malfunctions are caused by, for example, normal wear and tear on bearings and other similar parts.
The team analyzed data from five types of motors in Agilons’ robots. These motors have different functions; namely, moving the robot along its rail, lifting and lowering it, docking, and grappling the items – an action that involves gripping and back-and-forth motions.
“We were excited to see how clearly we could observe a picture emerging just from the data we had from the motors,” says Sami Terho, Research Engineer at Konecranes.
“The data we used to observe how the Agilon units’ devices function was anonymous data from Agilons in customer use – we had no information on what they were carrying or the parts or tools that were stored in the systems,” elaborates Terho.
In addition to the data from the electric motors, the team initially analyzed information from other sources, including usage statistics and error logs, as well as service or maintenance data that described actual failures through incident descriptions, service tickets and failure analyses.
The data analysis team employed a whole set of analytics software in the study. They used an approach where they initially prepared the data and defined the target, transferred the data, and attempted to interpret it and find cases of faults or failures.
Next the data was preprocessed by combining and cleaning it. This was followed by feature extraction, calculating features and smoothing time-related variations, and then actual machine learning wherein the model was trained on known failures and no failure data. The next steps were validation and testing.
A step-by-step, cyclic approach
The team proceeded with the development in phases. First, they developed the algorithm for one motor, then they tested it before refining and finally generalized the algorithm to work on all motors.
In the test phase, the data analysis team coordinates with the Agilon service team to see how well the predictions work. At the same time, more data is being collected on a greater number of failure cases.
“The model becomes increasingly accurate as more data is gathered. So basically, we just need to retrain the model with more data. It should then become a bit better every time we get more failure cases to learn from and find new failure modes that could not have been predicted before,” relates Terho.
“The next steps involve gathering even more detailed data from Agilons, analyzing it and trying to see if we can make better models. But these will be minor steps. The biggest was to start from scratch and have the first model to predict the failures,” Björkbom says.
Terho suggests that similar methods are being developed that will bring intelligence and machine learning to other types of Konecranes equipment: “There is a lot of potential concerning machine learning with different kinds of applications, such as predicting and optimizing procurement,” he says.
Less downtime and greater availability
Predictions on needed component changes can be used to plan maintenance tasks.
Breakdowns can be prevented more efficiently as the model is refined further.
About the experts
Mikael Björkbom, Senior Chief Engineer; and Sami Terho, Senior Specialist in Crane Intelligence, are in charge of research on crane systems and data analytics for Konecranes’ Industrial Internet initiative. Both have backgrounds as postdoctoral researchers at Aalto University, in Helsinki, Finland.
Text: Patricia Ongpin Steffa