The AI4Robotics pilot experiment aims at demonstrating that the AI4EU Experiment function can be exploited in industrial robotics for testing robot's wear.
The pilot consists of two different setups: 1) one setup is related to measuring wearing of the wrist on robots, and 2) another setup is related to measuring wearing of paint pumps used in relation to robotized painting. In both setups, one or more vibration sensors are attached to strategic places and vibration data is measured. The measured values from the vibration sensors are then uploaded on the AI4EU platform. Both setups are set up to run in a 24/7 interval time and will continuously upload vibration data on the platform. There will also be uploaded metadata such as measurement of play in gears, leakage in pumps, service intervals and maintenance logs. Based on the vibration data and the metadata, the goal of the pilot experiment is to develop predictive maintenance models which can estimate when running maintenance service is needed, and what kind of maintenance.
AI4EU platform exploitation
The pilot is planned to be integrated with the features of the AI4EU platform that are dedicated to AI models, that is, the resources catalog and AI4EU Experiments.
Cooperation of the organisations
ABB Robotics, as the industrial partner of the pilot, builds and operates the experimental robot platform at its premises in Bryne, Norway. Simula Research Laboratory (SRL) is charged of creating the central infrastructure allowing data sharing with the other research partners and allowing technical collaboration. This central infrastructure for data sharing and computational experiments has been set up (see description below).
Outside the partner of the pilot, there is an ongoing collaboration with Aristotle University of Thessaloniki that applies Explainable AI techniques on predictive maintenance models.
Currently, the AI4Robotics pilot uses a private workspace on the Teralab platform, that is available to all partners in the pilot. The workspace is set up for automated data collection from the experimental platform at ABB Robotics and allows the research partners to access and inspect this data, as well as to run experiments and develop the predictive maintenance model.
Once the AI4EU Experiments subsystem with the Design Studio is available, our goal is to utilize this functionality to model the training pipeline using our own data storage and model architecture.
Timeline and achievements
The AI4EU Experiments subsystem is central to the modeling and deployment strategy for the AI4Robotics pilot.
We consider three deployment scenarios for the final predictive maintenance model. First, the model is deployed directly on the robot, respectively its controller unit. In this case the model is dedicated to a single robot unit, which is useful if only a single robot of this type is deployed or the response time for the model is critical. Second, the model is deployed centrally on-site and can be used with multiple robots, which is useful in an industrial manufacturing setting, e.g. production lanes. Third, the model is deployed in a cloud environment and can be used with multiple robots at different locations, which can be useful in case of multiple production locations or if the predictive maintenance is offered as a central service to multiple customers. Which deployment scenario is chosen is depending on the business case and the actual use case, but all of the scenarios can be realized with the AI4EU Experiments subsystem.
To achieve this flexibility, we bundle the trained predictive maintenance model as a dockerized model with a protobuf interface. This model can then be onboarded on the AI4EU Experiments from where deployment options to both public cloud providers, e.g. Microsoft Azure, and a local cloud with Kubernetes are available.
It is also planned to publish the collected data set and the final predictive maintenance models in the AI4EU resources catalog, once the data collection and model development is completed.