Edge-enabled Federated Learning for Vision based Product Quality Inspection

Since the proliferation of Industry 4.0, manufacturing organisations are moving away from manual product quality inspection to a more digitized edge based inspection supported by sensor-actuator systems deployed at factory floor. Such edge based inspectors are equipped with trained machine learning (ML) models to infer the input data gathered from various sensor types and perform classification and/or detection of damaged products. In a dynamic scenario such as manufacturing, any unseen defects are marked as errors which are subsequently compiled at a back-end/cloud server for ML model re-training and re-deployed. Such model update and periodic release can prove to be costly in a time-critical manufacturing set-up. This paper proposes edge enabled federated learning (FL) based approach to enable visual inspectors recognize unseen defects by using on-device model fine-tuning and secure model exchange with each other. Preliminary results suggest proposed approach is able to improve edge based inspectors' accuracy to recognize previously unseen defects.